content
stringlengths 1
15.9M
|
|---|
\section{Conclusion}
In this paper, novel NN-based methods for effective mutual interference mitigation and denoising in the context of automotive radar sensors have been presented. Noteworthy is the usage of (complex-valued) spectrograms from different steps in the range-Doppler signal processing as network inputs and their suitability to be processed by CNNs.
An extensive simulation framework was used for data generation, training and evaluation. The most promising model architecture was then compared to a small selection of well-known conventional interference mitigation techniques. It was shown that the CNN-based model is capable of preserving the object peaks, while suppressing noise and interference by several orders of magnitude compared to conventional methods. However, its performance may be less robust, especially when considering the distortion of object peak values. Furthermore, the achieved amount of suppression indicates an implicit peak detection capability, which is of course attributed to the use of clean training data.
The most important issue in the future is to analyze the generalization capability of the architectures to real-world data. In addition, we would like to investigate the potential of temporal information in range-Doppler processing using NNs.
\section{Introduction}
Automotive radar sensors are key elements of current driving assistance systems as well as of autonomous driving applications. Nowadays, frequency modulated continuous wave (FMCW)/chirp sequence (CS) radars are prevalent. They share a non-regulated spectrum, transmitting sequences of linear chirp signals. Requirements for fine range resolution demand ever larger radio frequency (RF) transmit bandwidths, while the number of sensors deployed is also rising. Hence, mutual interference between radar sensors is becoming increasingly likely. The most common form of mutual interference is non-coherent interference~\cite{TOT18}, in which radar sensors with non-identical transmit signal parameters interfere. This leads to time-limited broadband disturbances in the baseband signal, whose primary effect is a reduced object detection sensitivity~\cite{BRO07}. Therefore, interference mitigation is a crucial part of current and future radar sensors used in a safety context.
Several conventional signal processing algorithms have been proposed in order to mitigate mutual interference. The most basic method is to zero out detected interference samples. In \cite{WAG18}, nonlinear filtering in slow-time is performed to remove interference. A different method is proposed in \cite{Bechter2017a}, where the useful signal is iteratively reconstructed using Fourier transforms and thresholding \cite{MAR12}. The interference component of the signal may be estimated and subtracted, such as in~\cite{BEC17}. Furthermore, beamforming can be used to reduce the impact of interference from particular directions \cite{Bechter2016}. Some machine learning techniques were discussed in the context of interference detection and classification in~\cite{Zhang2018a}. However, no explicit machine learning approach has been proposed in the context of interference mitigation.
In this paper, we use neural networks (NNs) as a powerful machine learning method to mitigate interference. In particular, convolutional NNs (CNNs) are employed. They are capable of learning local patterns, by considering inputs that are located close-by, and recognize them throughout the whole data signal. This structure can also be advantageous for spectrogram representations. Additionally, CNNs require a relatively small amount of learnable parameters compared to fully connected NNs, which makes them more appropriate for deployment on resource-constrained systems such as integration on chip level.
We will show how a two-channel representation of complex spectrogram data~\cite{fu2017complex} can be used as network input at two different points in the processing chain.
Since automotive radar is a safety-critical application, certain requirements must be fulfilled for interference mitigation and signal denoising algorithms. Besides an adequate noise suppression, no artifacts may be generated by the processing that can lead to spurious detections (ghost objects). We address these issues by using a detailed performance comparison that evaluates different application-relevant measures in a Monte-Carlo simulation~\cite{performance-comparison-interf-mitigation}.
Main contributions of this paper are:
\begin{itemize}
\item We show specific CNN structures capable of denoising radar signals.
\item We present numerical results using application-related performance metrics in a comparison with the state of the art.
\item We show that an excellent level of noise reduction and hence an improvement of detection sensitivity can be achieved.
\end{itemize}
\subsection{Comparative Analysis with Other Techniques}
The algorithms described in Section~\ref{subsec:setupcomp} were implemented in the simulation framework and evaluated using performance metrics from Section~\ref{subsec:perfmeas}. The results were statistically compared with the best-performing CNN-based method (RD RIS Model D) from Section~\ref{sec:performance-cnn}.
\begin{figure}[tb]
\centering
\resizebox {0.9\columnwidth} {!} {
\input{rd-sinr_cdf.tex}
}
\caption{CDF comparison of RD SINR with other techniques.}
\label{fig:cdf-rd-sinr}
\end{figure}
Fig.~\ref{fig:cdf-rd-sinr} shows the range-Doppler SINR performance of the selected techniques. Zeroing already considerably increases the SINR, since it removes interference completely (due to the perfect detection assumption), although it also removes parts of the object signal. IMAT is a natural improvement of zeroing, while ramp filtering can achieve an even larger noise suppression due to the principle of its non-linear operation. RD RIS denoising results in a superior average SINR compared to all other approaches. In fact, it appears to implicitly detect object peaks, enabling it to maximally attenuate the surrounding noise. It is only in severely interfered scenarios that object peaks are not recognized, and thus suppressed. This is shown by the long tail of the SINR CDF.
\begin{figure}[tb]
\centering
\resizebox {0.9\columnwidth} {!} {
\input{rd-evm_cdf.tex}
}
\caption{CDF comparison of RD EVM with other techniques.}
\label{fig:cdf-rd-evm}
\end{figure}
Fig.~\ref{fig:cdf-rd-evm} shows the corresponding EVM performance. Due to its non-linear nature, ramp filtering performs the worst among conventional methods. IMAT performs the best, reducing the bias in object peak values introduced by zeroing. RD RIS denoising on the other hand yields much higher EVMs. This indicates that object value preservation cannot be reliably guaranteed by such a denoising method.
\begin{table}
\centering
\caption{Performance comparison with state-of-the-art interference mitigation methods.}
\begin{tabular}{l >{\bfseries}r r >{\bfseries}r r}
Signal & SINR (RD) & EVM (RD) \\
\hline
Noisy & 46.16 & 0.01 \\
Interfered & 26.67 & 0.10 \\
Zeroing & 40.42 & 0.08 \\
Ramp filtering & 46.50 & 0.15 \\
IMAT & 43.20 & 0.03 \\
RD RIS Model D & 98.84 & 0.78 \\
\end{tabular}
\label{tab:performance-comparison}
\end{table}
Table~\ref{tab:performance-comparison} shows the mean values of performance metrics over the simulated test scenarios. For the conventional methods, ramp filtering achieves the highest SINR at the cost of an elevated EVM, while IMAT can improve on the performance of zeroing and lower the EVM. RD RIS denoising outperforms all the conventional methods in terms of SINR, while the average EVM is considerably higher.
\begin{figure}
\centering
\resizebox {0.9\columnwidth} {!} {
\input{range_cut.tex}
}
\caption{Range cut at velocity $v=5.5\textrm{m/s}$, shown up to $40\textrm{m}$. The object is located at a distance $d=7.9 \textrm{m}$ as indicated by the vertical black marker.}
\label{fig:range-cut}
\end{figure}
\begin{figure}
\centering
\resizebox {0.9\columnwidth} {!} {
\input{velocity_cut.tex}
}
\caption{Velocity cut at distance $d=7.9 \textrm{m}$. The object is located at a velocity $v=5.5\textrm{m/s}$ as indicated by the vertical black marker.}
\label{fig:velocity-cut}
\end{figure}
In order to illustrate the effects of the different mitigation methods, Figs.~\ref{fig:range-cut} and~\ref{fig:velocity-cut} show range- and velocity cuts of a RD map. The magnitude-normalized RD map of the first receive channel is plotted as a log-magnitude spectrum at a distance $d=7.9 \textrm{m}$ and velocity $v=5.5 \textrm{m/s}$ respectively. It can be noted that ramp filtering strongly suppresses noise, though mainly on the range axis. Zeroing and IMAT have almost the same effect for the visualized scenario. However, the previous statistical analysis shows that IMAT is superior, especially regarding object peak value preservation. The RD RIS model enhances object peaks, while it strongly reduces the noise floor to a \emph{constant} level. This reassures our presumption that the CNN-based denoising has an implicit thresholding effect.
In summary, CNN-based methods have superior noise and interference suppression capabilities compared to conventional algorithms. However, some exhibit a considerably high EVM which may lead to distortions in object peak values. Such distortions may have negative effects on further radar processing, such as on angular estimation or object classification. Due to these properties, the CNN-denoised signal is very well suited for object detection on the RD map while further processing can alternatively be performed using the non-mitigated object peak values.
\section{Methodology}
\label{sec:methodology}
The proposed denoising and interference mitigation network architecture is based on CNNs. We investigate two different denoising approaches, RPD and RDD, as shown in Fig.~\ref{fig:spchain_block}. Furthermore, we use either one input channel for the log-magnitude spectrogram (LMS) or two channels, i.e. the real and imaginary parts of the complex-valued spectrogram (RIS). The underlying goal is the same, namely, to reduce the impact of both noise and interference in order to enable a reliable detection of object parameters at a large sensitivity.
\subsection{Model architecture}
{
\thinmuskip=0mu
\medmuskip=0mu
\tiny
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{cnn_arch_bd}
\caption{Proposed CNN architecture for radar signal denoising. It uses \emph{ReLu}, \emph{Batch Normalization (BN)} and the convolution operation $\textrm{conv}(i,o,s_1\times s_2)$; see Fig.~\ref{fig:cnn_arch_conv} for details on the convolution operation.}
\label{fig:cnn_arch}
\end{figure*}
}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{conv}
\caption{Schematic representation of the convolution operation $\textrm{conv}(i,o,s_1\times s_2)$, where $i$ is the number of input channels, $o$ is the number of kernels, and thus output channels, and $s_1\times s_2$ is the kernel size. The convolution with a kernel is indicated by $*$.}
\label{fig:cnn_arch_conv}
\end{figure}
The proposed CNN architecture (see Figs.~\ref{fig:cnn_arch} and ~\ref{fig:cnn_arch_conv}) consists entirely of convolutional layers. The first layer uses the convolution operation and a \emph{ReLu} \cite{journals/jmlr/GlorotBB11} activation function, while subsequent layers include \emph{Batch Normalization} \cite{DBLP:journals/corr/IoffeS15}, the convolution operation and the \emph{ReLu} non-linearity, except the last layer which uses a linear activation function instead.
The model architecture differs slightly for the two approaches RPD and RDD. RPD is applied to one-dimensional data ($1 \times N$) and therefore uses one-dimensional kernels. RDD has two-dimensional input samples, i.e. $N \times M$ patches, and uses square kernels. In both approaches, zero-padding is used for the values at the outer boundaries, such that the inputs and outputs for each layer have the same dimensionality.
\subsection{LMS versus RIS Denoising}
We compare denoising with LMS and RIS inputs with focus on performance, memory requirements and application relevance. Denoising LMS can be used for object detection tasks but not for further processing based on the complex spectral values, because the phase information is lost. In RD LMS denoising, the original data, which was not processed by the NN, can be used for further AS calculations. With RIS inputs on the other hand, the denoised spectra can also be directly used for further processing, thus for the RD calculation in RP denoising or the AS calculations in RD denoising.
RDD with e.g. six layers and sixteen kernels with size $3 \times 3$ requires 10002 parameters for RIS and 9713 parameters for LMS inputs, this corresponds to a reduction of only 2.8 \%. A comparable architecture for RPD, i.e. eight layers and eight kernels with size $1 \times 41$, requires 17210 parameters for RIS denoising. See Section \ref{sec:experimental-results} for a more detailed performance comparison.
\subsection{Data Preprocessing}
Prior to model training and evaluation the data samples are standardized, to increase learning capability and model robustness. Two methods are investigated: \emph{Zero-Mean Unit-Variance Scaling (ZMUVS)} and \emph{Complex Standard Scaling (CSS)} \cite{DBLP:journals/corr/TrabelsiBSSSMRB17}.
In both approaches the complex data points are translated to zero-mean, while for ZMUVS the data points are then scaled to unit-variance and for CSS the data points are scaled to the standard normal complex distribution using the inverse square root of the covariance matrix of real and imaginary values.
\subsection{Loss functions}
\label{sec:learn-params}
The loss function defines the similarity of NN-outputs to the NN-targets, thus represents the learning goal. We evaluate the following measures:
\begin{itemize}
\item The \emph{Mean Squared Error (MSE)} is calculated from the real and imaginary parts of the values of the complex spectrogram.
\item The \emph{Signal-to-interference-plus-noise-ratio (SINR)} is the proportion of signal power compared to the noise floor, where the latter is given by both, noise and interference.
\item The \emph{Weighted MSE} is determined as convex combination of the MSE of the complex spectrum, and the magnitude and phase of object peaks.
\end{itemize}
\subsection{Training Setup}
The \emph{Adam} \cite{DBLP:journals/corr/KingmaB14} algorithm is used for training with a learning rate of $0.00005$ and two input samples per batch.
\section{Experimental Setup}
\label{sec:experimental-setup}
In our experiments, we use simulated FMCW/CS radar signals, which gives us access to interfered data and their corresponding clean equivalent. The basic receive IF signal is generated according to~\eqref{eq:signal-model} and processed as described in Section~\ref{sec:sigmod}. The resulting signals depend on the parameters of the random scenarios, which are generated according to uniform distributions $\mathcal{U}(min,max)$ in the respective domains. Among these parameters are the number of objects $\mathcal{U}(1,20)$ and for each object the distance $\mathcal{U}(0\mathrm{m},153\mathrm{m})$ and velocity $\mathcal{U}(-20\mathrm{m/s},20\mathrm{m/s})$ relative to the radar, such that all object parameters lie within the radar's limits.
The interferer parameters are uniformly sampled within the ranges depicted in Table \ref{tab:interferer-param}. The SIR and SNR are used to scale the interference and noise powers relative to the object signal power respectively, when generating the interfered and noisy time domain signal $s_{IF}[n,m]$. The victim radar parameters are kept constant and chosen as shown in Table~\ref{tab:victim-radar-param}. Fig.~\ref{fig:rd-interfered-and-target} shows a RD map processed from simulated data from a scenario with eight objects, where Fig.~\ref{fig:rd-interfered} shows an interfered signal and Fig.~\ref{fig:rd-target} shows the corresponding clean data.
\begin{table}
\centering
\vspace{0.5cm}
\caption{Ranges of interference and noise parameters.}
\begin{tabular}{l l r r}
Parameter & & Lower limit & Upper limit \\
\hline
$N_\textrm{I}$ & Number of interferers & 1 & 3 \\
$f_{\textrm{0,I}}$ & Sweep start frequency & $75.8 \textrm{GHz}$ & $76.2 \textrm{GHz}$ \\
$B_\textrm{I}$ & Sweep bandwidth & $0.6 \textrm{GHz}$ & $1.4 \textrm{GHz}$ \\
$T_\textrm{I}$ & Sweep duration & $40 \mu \textrm{s}$ & $46 \mu \textrm{s}$ \\
$\textrm{SIR}$ & Signal-to-interference-ratio & $-20 \textrm{dB}$ & $-60 \textrm{dB}$ \\
$\textrm{SNR}$ & Signal-to-noise-ratio & $-10 \textrm{dB}$ & $+10 \textrm{dB}$ \\
\end{tabular}
\label{tab:interferer-param}
\end{table}
\begin{table}
\centering
\vspace{0.5cm}
\caption{Victim radar and signal processing parameters.}
\begin{tabular}{l l r r}
Parameter & & Value \\
\hline
$f_{\textrm{0,V}} $ & Sweep start frequency & $76 \textrm{GHz}$ \\
$B_\textrm{V}$ & Sweep bandwidth & $1 \textrm{GHz}$ \\
$T_\textrm{V}$ & Sweep duration & $48 \mu \textrm{s}$ \\
$B_{\textrm{IF,V}}$ & IF bandwidth & $20 \textrm{MHz}$ \\
$N$ & Number of fast-time samples & $1024$ \\
$M$ & Number of slow-time samples/ ramps & $128$ \\
$A$ & Number of antennas & $8$ \\
$w$ & Window type & $\textrm{Hann}$ \\
\end{tabular}
\label{tab:victim-radar-param}
\end{table}
\begin{figure}
\footnotesize
\centering
\subfigure[Interfered]{
\input{eval_evaluation_test_rd_doppler-range_matrix_interfered_p8_id0.tex}
\label{fig:rd-interfered}
}
\hspace{-0.6cm}
\subfigure[Clean]{
\input{eval_evaluation_test_rd_doppler-range_matrix_targets_p8_id0.tex}
\label{fig:rd-target}
}
\caption{Exemplary range-Doppler magnitude spectra in dB of a scenario with eight objects.}
\label{fig:rd-interfered-and-target}
\end{figure}
\subsection{Data Sets}
Three separate data sets are used for training, validation and testing the models. The data sets contain samples of 2000 scenarios, i.e. RPs or RD maps, for training, and 250 scenarios for validation and testing each. Data from a single scenario are exclusively contained either in the training, validation or test set.
\subsection{Performance Measures}
\label{subsec:perfmeas}
For performance evaluation, we examine different quantitative and qualitative measures, which cover two fundamental aspects of object detection in chirp sequence radar processing \cite{performance-comparison-interf-mitigation}:
\begin{itemize}
\item The \emph{detection probability} gives the chance that an object is detected on the RD map.
\item The \emph{determination of detected object properties} is defined through the correctness of object location on the RD map as well as object resolution and peak distortion, which provide information about the object's radar cross section and thus its physical characteristics.
\end{itemize}
The goal of interference mitigation is to maximize the detection probability while avoiding modifications in object properties, i.e. the object peak's magnitude and phase.
\subsubsection{Quantitative measures}
The \emph{signal-to-interference-plus-noise ratio (SINR)} directly relates to the detection probability. It is defined through the ratio of signal power at the object peaks compared to the noise floor. In the two-dimensional case, i.e. in the range-Doppler domain, for multi-object scenarios the SINR is defined as:
\begin{equation}
\textrm{SINR} = 10 \log \bigg(\frac{
\frac{1}{N_O} \sum_{\{n,m\} \in \mathcal{O}}^{} {\mid \tilde{S}_\textrm{RD}[n,m]\mid}^{2}
} {
\frac{1}{N_N} \sum_{\{n,m\} \in \mathcal{N}}^{} {\mid \tilde{S}_\textrm{RD}[n,m]\mid}^{2}
}\bigg),
\label{sinr}
\end{equation}
where $n$ and $m$ are row and column indices of the RD matrix, $\mathcal{O}$ is the set of object peaks and $\mathcal{N}$ is the set of $N_N$ noise cells. Noise cells are defined to have a minimum distance to each object peak depending on the bin width in distance and velocity domain as well as the physical resolution of the radar. In the one-dimensional case, thus for AS evaluation, the SINR is defined analogously.
The \emph{error vector magnitude (EVM)} gives information about the detected object properties. It is defined as the magnitude of the error vector between the clean RD map $S_\textrm{RD,clean}$ and the denoised signal $\tilde{S}_\textrm{RD}$, i.e. in a multi-object scenario:
\begin{equation}
\textrm{EVM} = \frac{1}{N_O} \sum_{\{n,m\} \in \mathcal{O}}^{}\frac{\mid S_\textrm{RD,clean}[n,m] - \tilde{S}_\textrm{RD}[n,m]\mid}{\mid S_\textrm{RD,clean}[n,m] \mid}.
\end{equation}
\subsubsection{Qualitative measures}
During visual inspection of the RD map and the AS, we consider criteria such as object peak and noise floor intensity, object peak location, resolution and distortion as well as artifact appearances.
\input{setup_comparison.tex}
\section{Experimental Results}
\label{sec:experimental-results}
The proposed NN-based architecture is evaluated in several steps: First, the optimal network architectures for RIS denoising are analyzed using the MSE loss function and ZMUVS as ``basic" setup. The best performing architectures in terms of overall-performance and performance-complexity are further evaluated with respect to scaling methods, loss functions and according to their generalization capabilities. Second, the different proposed CNN-based approaches, i.e. RPD and RDD with LMS and RIS inputs, are evaluated and analyzed. Third, the best-performing CNN-based model is compared to state-of-the-art mitigation algorithms.
\subsection{Analysis of Optimal CNN-Architecture}
\label{sec:exp-results-architecture}
We used grid search to systematically find the best CNN architecture for RD RIS denoising, i.e. the number of layers ($4$, $6$, $8$), kernel sizes ($1\times1$, $3\times3$, $5\times5$, $7\times 7$) and number of kernels per layer ($2$, $8$, $16$, $32$). The choice of the kernel size is a trade-off between object resolution in the denoised spectrogram and noise suppression. Larger kernels enable better denoising performance, but result in distortion of the peak shapes and thus a possible loss of resolution. We use ZMUVS and the MSE loss function for training the CNNs.
Fig.~\ref{fig:arch-comparison} shows the SINR and EVM based performance comparison of all evaluated RD architectures for RIS inputs using MSE loss and ZMUVS. The SINR and EVM are illustrated in blue and red respectively, while the x-axis indicates the number of parameters of the NN-model. The best performing models are marked with A to F and listed in detail in Table \ref{tab:arch-comparison-top-results}. The best performance for RD RIS denoising (Model D) is obtained using a model with $6$ layers, a kernel size of $3\times 3$ and $16$ kernels per layer. Model A offers the best performance-complexity trade-off with only four layers and two kernels with a size of $3\times 3$. With an average SINR loss of only 3.8 dB, there is a parameter reduction of 98.4 \% compared to Model D. However, the small model size comes with the cost of a high average EVM. This indicates a large distortion of the complex values of object peaks, which causes a notable decrease of SINR in the AS. For the other architectures listed in Table~\ref{tab:arch-comparison-top-results} the EVM values appear to be small enough such that other effects, e.g. the main- to side-lobe ratio of the AS peak, dominate the AS SINR.
Using similar parameter ranges except for one-dimensional kernels ($5$, $13$, $21$, $25$, $31$, $41$, $43$, $47$, $51$, $55$, $61$), the most suitable architecture for RP denoising has $6$ layers, a kernel size of $1\times 41$ and $16$ kernels per layer.
\begin{figure}
\centering
\resizebox {\columnwidth} {!} {
\input{arch_comparison.tex}
}
\caption{CNN architecture performance comparison for RD RIS denoising using MSE loss and ZMUVS. Labels A-F refer to model parameters given in Table~\ref{tab:arch-comparison-top-results}.}
\label{fig:arch-comparison}
\end{figure}
\begin{table*}
\centering
\vspace{0.5cm}
\caption{Best performing architectures for RD RIS denoising.}
\begin{tabular}{l r r r r >{\bfseries}r r >{\bfseries}r r}
Model & Layers & Kernels & Kernel Size & Parameters & SINR (RD) & EVM (RD) & SINR (AS) \\
\hline
A & 4 & 2 & ($3 \times 3$) & 160 & 73.67 & 0.90 & 7.50 \\
B & 8 & 8 & ($3 \times 3$) & 3898 & 73.60 & 0.40 & 9.47 \\
C & 4 & 16 & ($3 \times 3$) & 5298 & 74.57 & 0.30 & 9.39 \\
D & 6 & 16 & ($3 \times 3$) & 10002 & 77.47 & 0.58 & 9.56 \\
E & 8 & 16 & ($3 \times 3$) & 14706 & 77.78 & 0.47 & 9.54 \\
F & 6 & 32 & ($3 \times 3$) & 38434 & 72.20 & 0.51 & 9.43 \\
\end{tabular}
\label{tab:arch-comparison-top-results}
\end{table*}
Data scaling has a strong impact on the training progress in terms of duration and stability. While CSS statistically results in stronger denoising performance when using MSE loss (on average additional 11.41 dB SINR), it does not seem to have any positive effects when using SINR loss. CSS has a slightly negative effect on the object peak values which results in an increased average EVM. ZMUVS on the other hand produces results with a lower average SINR, but it also has a lower average EVM. Additionally, CSS leads to a smoother and more stable learning improvement.
When analyzing the loss functions introduced in Section \ref{sec:learn-params}, we can see that the SINR loss produces the highest denoising performance for most models, but it fails to preserve phase information for further processing. For example when evaluating Model D with ZMUVS, it increases the SINR performance metric by an average of 17.58 dB compared to the MSE loss function. On the contrary the MSE loss takes the relation of imaginary and real values of the inputs into account, and thus better preserves object peak values. The \emph{Weighted MSE} performs worse than the other loss functions, both in terms of SINR and EVM.
We use a second pair of training and validation data sets, that contain only up to two interferer, in order to investigate generalization capabilities. Thus we want to show how the trained network performs on a test set coming from a slightly different distribution, i.e. with three interferers, than the data seen during the training process. With an average SINR loss of 5.6 dB (Model A) and 23.89 dB (Model D) RD RIS denoising seems to generalize to similar data as seen during training. Even with a decreased SINR by 23.89 dB RD RIS denoising outperforms the state of the art. However, the performance loss between the two models suggests that stronger regularization is required for bigger architectures.
\subsection{Performance Analysis of CNN-based Approaches}
\label{sec:performance-cnn}
The different CNN-based approaches as described in Section~\ref{sec:methodology} are analyzed using performance metrics as introduced in Section~\ref{subsec:perfmeas}. For RD LMS denoising and RD RIS denoising, we use Models A and D from Section~\ref{sec:exp-results-architecture}, thus the best architectures in performance-complexity and overall-performance respectively. RPD is performed using the model described in Section~\ref{sec:exp-results-architecture}. We use CCS and the MSE loss function for training the CNNs.
The performance is illustrated using the cumulative distribution function (CDF) of the respective metric. The values are computed from the test set in a Monte Carlo simulation as introduced in Section~\ref{sec:experimental-setup}. The interfered signal without mitigation (\emph{interfered}) and the signal with only AWGN (\emph{noisy}) are included as references.
\begin{figure}[tb]
\centering
\resizebox {0.9\columnwidth} {!} {
\input{rd-sinr_cdf_cnns.tex}
}
\caption{CDF of RD SINR of different CNN-based models.}
\label{fig:cdf-rd-sinr-cnn}
\end{figure}
Fig.~\ref{fig:cdf-rd-sinr-cnn} shows the range-Doppler SINR performance of the different models. RD LMS denoising generally requires larger models in order to perform well. Model A with LMS fails to learn the denoising task, while Model D with LMS seems to perform quite well on average. However, the SINR CDF and visual inspections suggest that RD LMS denoising performs well on data with weak interference while the denoising performance drastically decreases on data with stronger interference. In severe cases, it completely fails to detect object peaks contained in the RD map. Thus, this approach is rather unreliable and performs worse compared to the other CNN-based approaches.
RPD has a lower average SINR than the other approaches, but it shows remarkably low variance. This suggests, that RPD is a more stable approach, with a solid performance also on strongly interfered data.
RD RIS denoising results in a superior average SINR compared to the other approaches. Model A performs well on around 75 percent of the data samples despite its small model size. The SINR drops notably for the other 25 percent tough, which suggests that Model A is not capable of denoising a broad variety of interference. Model D on the other hand has an even stronger and also more reliable denoising performance than Model A, which becomes apparent especially for scenarios with stronger interference.
\input{results_comparison.tex}
\input{conclusion.tex}
\bibliographystyle{ieeetr}
\subsection{Mitigation Methods Selected for Comparison}\label{subsec:setupcomp}
A small number of the most well-known and promising state-of-the-art signal processing algorithms have been chosen for a comparative analysis. This also allows for a discussion of the properties of the NN-based approaches in a broader context of interference mitigation. A short summary of these methods is presented below.
\subsubsection{Zeroing} Zeroing is selected as a baseline, since it is a simple and well-known method. Time domain samples of the IF signal, that are determined to be dominated by interference, are simply set to a value of zero. Its properties have been discussed in e.g.~\cite{Fischer}.
\subsubsection{Iterative method with adaptive thresholding (IMAT)} IMAT~\cite{Bechter2017a} is based on an initial zeroing step, eliminating interference. The resulting ``gaps" in the signal are then interpolated in an attempt to fully reconstruct the object signal. This reconstruction is done by an iterative thresholding method making use of the theory of sparse sampling.
\subsubsection{Ramp filtering (RFmin)} Ramp filtering~\cite{WAG18}, as opposed to previously mentioned techniques, processes the signal after the first DFT. It exploits the sparsity and diversity of interference over the slow-time domain, using a non-linear filtering operation to achieve considerable interference as well as noise suppression. Several choices of filtering can be considered. In this work, a simple \emph{minimum operator} is implemented.
Note that both zeroing and IMAT require the detection of interfered IF signal samples. In this paper, it will be assumed that this operation works perfectly. However, in general, errors in interference detection may have a strong impact on the performance of mitigation algorithms~\cite{performance-comparison-interf-mitigation}. Ramp filtering, as well as the proposed novel approaches, are not directly based on an interference detection step.
\section{Signal Model}
\label{sec:sigmod}
\begin{figure}[tb]
\centering
\input{spchain.tex}
\caption{Block diagram of a basic FMCW/CS radar processing chain. Dashed boxes indicate the locations of optional interference mitigation steps, including our proposed methods. The signal at every point in the chain is labeled according to the variable names used in this paper.}
\label{fig:spchain_block}
\end{figure}
Fig.~\ref{fig:spchain_block} illustrates the \emph{range-Doppler (RD) processing} chain in a conventional FMCW/CS radar. First, the radar sensor performs a measurement by transmitting a sequence of linearly RF modulated chirp signals, also called \emph{ramps}. For each ramp, the received object reflections are then mixed, i.e. multiplied, with the transmit signal. This way the according time delays, and hence range information, are translated into corresponding constant frequency sinusoidals. The signal after mixing is limited by the \emph{intermediate frequency (IF)} bandwidth of the receiver and therefore termed the \emph{IF signal}. The velocity (Doppler) information is estimated by evaluating the rate of ramp-wise linear phase change of the received IF signal. More detailed descriptions as well as mathematical derivations of these principles can be found in~\cite{STO92,WIN07}.
From a data processing point of view, the received IF signal consists of $N$ time domain samples for each of the $M$ transmitted ramps. Hence, it can be interpreted as a two-dimensional data matrix $s_{\mathrm{IF}}[n,m]$ with the corresponding indices $n$ and $m$, also called \emph{fast-} and \emph{slow-time}, respectively. Essential processing steps include discrete Fourier transforms (DFTs) over the fast and slow-time, to reveal distance and velocity information accordingly. The resulting two-dimensional spectrum $S_{\mathrm{RD}}[n,m]$ ideally contains peaks at the objects' corresponding distances and velocities, which are then to be detected. In a system with multiple receive channels (antennas), additional information about the angle of arrival of object reflections can be extracted. This is done by evaluating the phase change of each object peak value over the receive channels by another DFT, yielding the so-called \emph{angular spectrum (AS)}. Further processing steps are then performed in higher layers of the application and may include sensor fusion, tracking, or classification.
However, in addition to the object reflections, IF signals in real radar systems also contain disturbances in the form of receiver noise and (mutual) interference. Other radar sensors in the radio range act as \emph{interferers} when transmitting inside the receiver IF bandwidth of the \emph{victim radar}. Accounting for this, the model of the IF signal $s_{\mathrm{IF}}[n,m]$ can be written as
\begin{equation}
s_{\mathrm{IF}}[n,m]=\sum_{o=1}^{N_{\mathrm{O}}} s_{\mathrm{O},o}[n,m] + \sum_{i=1}^{N_{\mathrm{I}}} s_{\mathrm{I},i}[n,m] + \upsilon[n,m] \, ,
\label{eq:signal-model}
\end{equation}
where $s_{\mathrm{O},o}[n,m]$ is the signal component of the $o^{th}$ object reflection, $N_{\mathrm{O}}$ denotes the number of objects, $s_{\mathrm{I},i}[n,m]$ is the signal component of the $i^{th}$ interferer assuming $N_{\mathrm{I}}$ interferers, and $\upsilon[n,m]$ is a receiver noise term. Receiver noise is modeled as AWGN, while mutual interference generally causes burst-like disturbances in time domain corresponding to broadband disturbances in frequency domain~\cite{TOT18,Kim2018}.
State-of-the-art (``classical") interference mitigation methods are mostly signal processing algorithms that are applied either on the time domain signal $s_{\mathrm{IF}}[n,m]$ or on the frequency domain signal $S_{\mathrm{R}}[n,m]$ after the first DFT. The two NN-based methods presented in this paper are applied at two different steps in the radar signal processing, i.e.,
\begin{enumerate}
\item Range-Profile Denoising (RPD)\label{rp-denoising}: Denoising of range-profiles after the first DFT.
\item Range-Doppler Denoising (RDD)\label{rd-denoising}: Denoising of range-Doppler maps after the second DFT.
\end{enumerate}
|
\section{Introduction}
We study numerical approximation of the stochastic total variation flow (STVF)
\begin{align}\label{TVF}
\d X&= \div\left(\frac{\nabla X}{\be{\nabla X}}\right) \d t -\lambda (X - g) \d t +X\d W, &&\text{in } (0,T)\times \O, \nonumber\\
X & = 0 && \text{on } (0,T)\times \partial \O, \\
X(0)&=x_0 &&\text{in } \O, \nonumber
\end{align}
where $\O \subset \mathbb{R}^d$, $d\geq 1$ is a bounded, convex domain with a piecewise $C^{2}$-smooth boundary $\partial \O$,
and $\lambda \geq 0$, $T>0$ are constants. We assume that $x_0,\, g \in \L$ and consider
a one dimensional real-valued Wiener process $W$, for simplicity; generalization for a sufficiently regular trace-class noise is straightforward.
Equation (\ref{TVF}) can be interpreted as a stochastically perturbed gradient flow of the penalized total variation energy functional
\begin{align}\label{jfunc}
\mathcal{J}_{\lambda}(u):= \int\limits_{\mathcal{O}} \be{\nabla u} \d x +\frac{\lambda}{2} \int\limits_{\mathcal{O}} \bes{u-g} \d x.
\end{align}
The minimization of above functional, so-called ROF-method, is
a prototypical approach for image denoising, cf. \cite{ROF};
in this context the function $g$ represents a given noisy image and $\lambda$ serves as a penalization parameter.
Further applications of the functional include, for instance, elastoplasticity and the modeling of damage and fracture,
for more details see for instance \cite{bm16} and the references therein.
The use of stochastically perturbed gradient flows has proven useful in image processing.
Stochastic numerical methods for models with nonconvex energy functionals
are able to avoid local energy minima and thus achieve faster convergence and/or
more accurate results than their deterministic counterparts;
see \cite{kp06} which applies stochastic level-set method in image segmentation,
and \cite{swp14} which uses stochastic gradient flow of a modified (non-convex) total variation energy functional for binary tomography.
Due to the singular character of total variation flow (\ref{TVF}), it is convenient to perform numerical simulations using a regularized problem
\begin{align}\label{reg.TVF}
\d X^{\epsilon}&= \div\left(\frac{\nabla X^{\epsilon}}{\sqrt{|\nabla X^{\epsilon} |^2+\epsilon^2}}\right)\d t-\lambda(X^{\epsilon}-g)\d t+X^{\epsilon}\d W &&\text{in } (0,T)\times \O, \nonumber\\
X^{\epsilon} & = 0 && \text{on } (0,T)\times \partial \O, \\
X^{\epsilon}(0)&=x_0 &&\text{in } \O\,, \nonumber
\end{align}
with a regularization parameter $\epsilon >0$. In the deterministic setting ($W\equiv 0$) equation \eqref{reg.TVF} corresponds to the gradient flow of the regularized energy functional
\begin{align}\label{def_jepslam}
\mathcal{J}_{\epsilon,\lambda}(u):= \int\limits_{\mathcal{O}} \sqrt{\bes{\nabla u} +\epsilon^2} \d x + \frac{\lambda}{2} \int\limits_{\mathcal{O}} \bes{u-g} \d x.
\end{align}
It is well-known that the minimizers of the above regularized energy functional converge
to the minimizers of (\ref{jfunc}) for $\epsilon \rightarrow 0$, cf. \cite{Prohl_TVF_numerics} and the references therein.
{
Owing to the singular character of the diffusion term in (\ref{TVF})
the classical variational approach for the analysis of stochastic partial differential equations (SPDEs), see e.g. \cite{kr_07}, \cite{Roeckner_book}, is not applicable to this problem.
To study well-posedeness of singular gradient flow problems it is convenient to apply the solution framework developed in \cite{Roeckner_TVF_paper}
which characterizes the solutions of (\ref{TVF}) as stochastic variational inequalities (SVIs).
In this paper, we show the well posedness of SVI solutions using the practically relevant regularization procedure (\ref{reg.TVF})
which, in the regularization limit, yields a SVI solution in the sense of \cite{Roeckner_TVF_paper}.
Throughout the paper, we will refer to the solutions which satisfy a stochastic variational inequality
as SVI solutions, and to the classical SPDE solutions as variational solutions.
Convergence of numerical approximation of (\ref{reg.TVF}) in the deterministic setting ($W\equiv 0$)
has been shown in \cite{Prohl_TVF_numerics}.
Analogically to the deterministic setting, we construct an implementable finite element approximation of the problem (\ref{TVF})
via the numerical discretization of the regularized problem (\ref{reg.TVF}). The scheme is implicit in time
and preserves the gradient structure of the problem, i.e., it satisfies a discrete energy inequality.
The deterministic variational inequality framework
used in the the numerical analysis of \cite{Prohl_TVF_numerics} is not directly transferable to the stochastic setting.
{Instead, we show the convergence of the proposed numerical approximation of (\ref{reg.TVF}) to the SVI solution of (\ref{TVF}) via an additional
regularization step on the discrete level. The convergence analysis of the discrete approximation is inspired by the analytical approach of
\cite{Gess_Stability} where the SVI solution concept was applied to the stochastic $p$-Laplace equation.}
As far as we are aware, the present work is the first to show convergence of implementable numerical approximation for singular stochastic gradient flows
in the framework of stochastic variational inequalities.
The paper is organized as follows. In Section~\ref{sec_not} we introduce the notation and state some auxiliary results.
The existence of a unique SVI solution of the regularized stochastic TV flow (\ref{reg.TVF})
and its convergence towards a unique SVI solution of (\ref{TVF}) for $\epsilon\rightarrow 0$
is shown in Section~\ref{sec_exist}.
In Section~\ref{sec_num} we introduce a fully discrete finite element scheme for the regularizared problem (\ref{reg.TVF})
and show its convergence to the SVI solution of (\ref{TVF}).
Numerical experiments are presented in Section~\ref{sec_sim}.
\section{Notation and preliminaries}\label{sec_not}
Throughout the paper we denote by $C$ a generic positive constant that may change from line to line.
For $1\leq p \leq \infty $, we denote by $(\mathbb{L}^p,\no{\cdot}_{\mathbb{L}^p})$ the standard spaces of $p$-th order integrable functions on $\O$,
and use $\no{\cdot}:= \no{\cdot}_{\L}$ and $(\cdot,\cdot):=(\cdot,\cdot)_{\L}$ for the $\L$-inner product.
For $k \in \mathbb{N}$ we denote the usual Sobolev space on $\O$ as $(\mathbb{H}^k,\no{\cdot}_{\mathbb{H}^k})$,
and $(\mathbb{H}^1_0,\no{\cdot}_{\mathbb{H}^1_0})$ stands for the $\mathbb{H}^1$ space with zero trace on $\partial \O$ with its dual space $(\mathbb{H}^{-1},\no{\cdot}_{\mathbb{H}^{-1}})$.
Furthermore, we set $\langle \cdot ,\cdot \rangle:=\langle \cdot ,\cdot \rangle_{\mathbb{H}^{-1} \times \mathbb{H}^1_0}$, where $\langle \cdot ,\cdot \rangle_{\mathbb{H}^{-1} \times \mathbb{H}^1_0}$
is the duality pairing between $\mathbb{H}^1_0$ and $\mathbb{H}^{-1}$. The functional (\ref{def_jepslam}) with
$\lambda=0$ will be denoted as $\mathcal{J}_\epsilon := \mathcal{J}_{\epsilon,0}$.
We say that a function $X \in L^1(\Omega\times(0,T);\L)$
is $\mathcal{F}_t$-progressively measurable if $X \mathbbm{1}_{[0,t]}$ is $\mathcal{F}_t\otimes \mathcal{B}([0,t])$-measurable for all $t \in [0,T]$.
For the convenience of the reader we state some basic definitions below.
\begin{defs}\label{resolvent and Yosida approximation}
Let $\mathbb{H}$ be a real Banach space, $A :D(A)\rightarrow \mathbb{H}$ a linear operator and $\rho(A)$ its resolvent set.
For a real number $\xi \in \rho(A)$ we define the resolvent $R_{\xi}: \mathbb{H} \rightarrow \mathbb{H}$ of $A$ as
\begin{align*}
R_{\xi}(x):=( I-\xi A)^{-1}x\,.
\end{align*}
Furthermore we define the Yosida approximation of $A$ as
\begin{align*}
T_{\xi}(x):=AR_{\xi}=\frac{1}{\xi}(I- R_{\xi})x.
\end{align*}
\end{defs}
\begin{defs}\label{orth.Proj}
The mapping $\mathcal{P}_m : \L \rightarrow \mathbb{V}_m\subset\L$ which satisfies
\begin{align*}
\ska{w-\mathcal{P}_m w,v_m}=0 ~~ \forall v_m \in \mathbb{V}_m.
\end{align*}
defines the $\L$-orthogonal projection onto $\mathbb{V}_m$.
\end{defs}
\begin{defs}\label{Bounded Variation}
A function $u \in L^1(\O)$ is called a function of bounded variation, if its total variation
\begin{align}\label{total variation}
\int\limits_{\mathcal{O}} \be{\nabla u}\dx := \sup\left\{-\int\limits_{\mathcal{O}} u\, \div\, \mathbf{v} \dx;~ \mathbf{v} \in C^{\infty}_0(\O,\mathbb{R}^d), ~\no{\mathbf{v}}_{L^{\infty}}\leq 1\right\},
\end{align}
is finite. The space of functions of bounded variations is denoted by $BV(\O)$.
For $u \in BV(\O)$ we denote
$$
\int\limits_{\mathcal{O}} \sqrt{\be{\nabla u}^2 + \epsilon^2}\dx := \sup\left\{\int\limits_{\mathcal{O}} \Big(-u\, \div\, \mathbf{v} + \epsilon\sqrt{1-|\mathbf{v}(x)|^2}\Big)\dx;~ \mathbf{v} \in C^{\infty}_0(\O,\mathbb{R}^d), ~\no{\mathbf{v}}_{L^{\infty}}\leq 1\right\}\,.
$$
\end{defs}
The following proposition plays an important role in the analysis below; the proposition
holds for convex domains with piecewise smooth boundary, which includes the case of practically relevant
polygonal domains, cf. \cite[Proposition 8.2 and Remark 8.1]{Roeckner_TVF_paper}.
\begin{props}\label{Resolvent_estimate}
Let $\mathcal{O}\subset\mathbb{R}^d$, $d\geq 1$ be a bounded domain with a piecewise $C^2$-smooth and convex boundary.
Let $g :[0,\infty) \rightarrow [0,\infty)$ be a continuous and convex function of at most quadratic growth such that $g(0)=0$,
then it holds
\begin{align}
\int\limits_{\mathcal{O}} g (\be{\nabla R_{\xi}(y)}) \d x \leq \int\limits_{\mathcal{O}} g (\be{\nabla y}) \d x, ~~\forall y \in \mathbb{H}^1_0.
\end{align}
\end{props}
\section{Well posedness of STVF}\label{sec_exist}
In this section we show existence and uniques of the SVI solution of (\ref{TVF}) (see below for a precise definition)
via a two-level regularization procedure.
To be able to treat problems with $\L$-regular data, i.e., $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$, $g\in \L$
we consider a $\mathbb{H}^1_0$-approximating sequence {$\{x^n_0\}_{n\in \mathbb{N}} \subset L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$
s.t. $x^n_0 \rightarrow x_0$
in $L^2(\Omega,\mathcal{F}_0; \L)$ for $n\rightarrow \infty$ and $\{g^n\}_{n\in \mathbb{N}} \subset \mathbb{H}^1_0$
s.t. $g^n \rightarrow g$
in $\L$ for $n\rightarrow \infty$.} We
introduce a regularization of (\ref{reg.TVF}) as
\begin{align}\label{vis.TVF}
\d X^{\epsilon,\delta}_{n}=&\delta \Delta X^{\epsilon,\delta}_{n} \d t +\div\left(\frac{\nabla X^{\epsilon,\delta}_{n}}{\sqrt{|\nabla X^{\epsilon,\delta}_{n} |^2+\epsilon^2}}\right)\d t \nonumber\\
&-\lambda(X^{\epsilon,\delta}_{n}-g^n)\d t+X^{\epsilon,\delta}_{n}\d W(t) &&\text{in }(0,T)\times \O,\\
X^{\epsilon,\delta}_{n}(0)=&x^n_0 &&\text{in } \O,\nonumber
\end{align}
where $\delta>0$ is an additional regularization parameter.
We define the operator $A^{\epsilon,\delta} : \mathbb{H}^1_0 \rightarrow \mathbb{H}^{-1}$ as
\begin{align}\label{Operator}
{\langleA^{\epsilon,\delta} u, v\rangle_{\mathbb{H}^{-1} \times \mathbb{H}^1_0}=\int\limits_{\mathcal{O}} \delta \nabla u \nabla v +\fe{u}\nabla v +\lambda (u- g^n)v \dx ~~\forall u,v \in \mathbb{H}^1_0 },
\end{align}
and note that \eqref{vis.TVF} is equivalent to
\begin{align}
\d X^{\epsilon,\delta}_{n} +A^{\epsilon,\delta} X^{\epsilon,\delta}_{n} \d t&=X^{\epsilon,\delta}_{n} \d W(t)\,,\\
X^{\epsilon,\delta}_{n}(0)&=x^n_0. \nonumber
\end{align}
The operator $A^{\epsilon,\delta}$ is coercive, demicontinuos and satisfies (cf. \cite[Remark 4.1.1]{Roeckner_book})
\begin{align}
&\langleA^{\epsilon,\delta}(u)-A^{\epsilon,\delta}(v),u-v\rangle_{\mathbb{H}^{-1} \times \mathbb{H}^1_0}\geq \delta \|\nabla (u-v)\|^2 + \lambda\nos{u-v}, && \forall u,v \in \mathbb{H}^1_0,\label{Monotonicity}\\
& \no{A^{\epsilon,\delta}(u)}_{\mathbb{H}^{-1}} \leq C(\delta,\lambda,\|g^n\|)(\no{u}_{\mathbb{H}^1_0}+1), && \forall u \in \mathbb{H}^1_0.\label{a_bnd}
\end{align}
The following monotonicity property,
which follows from the convexity of the function $\sqrt{\bes{\cdot}+\epsilon^2}$, will be used frequently in the subsequent arguments
\begin{align}\label{eps.convexity.inequality}
&\ska{\fe{X}-\fe{Y},\nabla(X-Y)}
\nonumber\\
=&\ska{\fe{X},\nabla(X-Y)}+\ska{\fe{Y},\nabla(Y-X)}
\\
&\geq \mathcal{J}_\eps(X)-\mathcal{J}_\eps(Y)+\mathcal{J}_\eps(Y)-\mathcal{J}_\eps(X)=0.
\nonumber
\end{align}
The existence and uniqueness of a variational solution of {(\ref{vis.TVF})}
is established in the next lemma; we note that the result only requires $\L$-regularity
of data.
\begin{lems}\label{lem_visc_exist}
For any $\epsilon, \delta>0$ and $x^n_0 \in L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$, $g^n\in \mathbb{H}^1_0$
there exists a unique variational solution $X^{\epsilon,\delta}_{n} \in L^2(\Omega;C([0,T];\L))$
of \eqref{vis.TVF}. Furthermore, there exists a $C\equiv C(T)>0$ such that the following estimate holds
\begin{align*}
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon,\delta}_{n}(t)}} \leq C (\mathbb{E}{\nos{x_0^n}} + \|g^n\|^2).
\end{align*}
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{lem_visc_exist}}]
On noting the properties (\ref{Monotonicity})-(\ref{a_bnd}) of the operator $A^{\epsilon,\delta}$ for $\epsilon,\delta>0$
the classical theory, cf. \cite{Roeckner_book},
implies that for any given data $x^n_0 \in L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$, $g^n\in \mathbb{H}^1_0$ there exists
a unique variational solution $X^{\epsilon,\delta}_{n} \in L^2(\Omega;C([0,T];\L))$
of (\ref{vis.TVF}) which satisfies the stability estimate.
\end{proof}
In next step, we show a priori estimate for the solution of (\ref{vis.TVF}) in stronger norms;
the estimate requires $\mathbb{H}^1_0$-regularity of the data.
\begin{lems}\label{laplace_energy_estimate}
Let $x_0^n\, \in L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$, $g^n \in \mathbb{H}^1_0$.
There exists a constant $C\equiv C(T)$
such that for any $\epsilon,\, \delta>0$ the corresponding variational solution $X^{\epsilon,\delta}_{n}$ of \eqref{vis.TVF} satisfies
\begin{align}
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{\nabla X^{\epsilon,\delta}_{n}(t)}}+\delta \mathbb{E}{\int_0^t \nos{\Delta X^{\epsilon,\delta}_{n}(s)} \d s} \leq C\left(\mathbb{E}{\nos{x_0^n}_{\mathbb{H}^1_0}}+\nos{g^n}_{\mathbb{H}^1_0}\right).
\end{align}
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{laplace_energy_estimate}}]
Let $\{e_i\}_{i=0}^{\infty}$ be an orthonormal basis of eigenfunctions of the Dirichlet Laplacian $-\Delta$ on $\L$
and $\mathbb{V}_m := \text{span}\{e_0,\ldots, e_m\}$. Let $\mathcal{P}_m :\L \rightarrow \mathbb{V}_m$
be the $\L$-orthogonal projection onto $\mathbb{V}_m$.
For fixed $\epsilon,\,\delta,\,n$ the Galerkin approximation $X^{\epsilon,\delta}_{n,m} \in \mathbb{V}_m$ of $X^{\epsilon,\delta}_{n}$ satisfies
\begin{align}\label{gal.vis.TVF}
\d X^{\epsilon,\delta}_{n,m} =& \delta \Delta X^{\epsilon,\delta}_{n,m} \d t +\mathcal{P}_m \div\left(\frac{\nabla X^{\epsilon,\delta}_{n,m}}{\sqrt{|\nabla X^{\epsilon,\delta}_{n,m} |^2+\epsilon^2}}\right)\d t
\nonumber \\
& -\lambda (X^{\epsilon,\delta}_{n,m}-{\mathcal{P}_m g^n})\d t+ X^{\epsilon,\delta}_{n,m} \d W(t),\\
X^{\epsilon,\delta}_{n,m}(0)=&\mathcal{P}_m x^n_0. \nonumber
\end{align}
By standard arguments, cf. \cite[{Theorem 5.2.6}]{Roeckner_book},
there exists a $X^{\epsilon,\delta}_{n} \in L^2(\Omega;C([0,T];\L))$ such that $X^{\epsilon,\delta}_{n,m} \rightharpoonup X^{\epsilon,\delta}_{n} $ in $L^2(\Omega\times(0,T);\L)$
for $m\rightarrow \infty$.
We use It\^o's formula for $\nos{\nabla X^{\epsilon,\delta}_{n,m}(t)}$
to obtain
\begin{align}\label{ito_nabla}
\nonumber
\frac{1}{2} \nos{\nabla X^{\epsilon,\delta}_{n,m}(t)}=& \frac{1}{2}\nos{\nabla \mathcal{P}_m x_0^n}-\delta \int_0^t \nos{\Delta X^{\epsilon,\delta}_{n,m}(s)}\d s
\nonumber
\\
\nonumber
&-\int_0^t \ska{\div \fe{X^{\epsilon,\delta}_{n,m}(s)},\Delta X^{\epsilon,\delta}_{n,m}(s)}\d s
\\
&-\lambda\int_0^t \ska{(X^{\epsilon,\delta}_{n,m}(s)-{g^n}),\Delta X^{\epsilon,\delta}_{n,m}(s)}\d s
\\ \nonumber
&-\int_0^t \ska{\Delta X^{\epsilon,\delta}_{n,m}(s), X^{\epsilon,\delta}_{n,m}(s)\d W(s)}\d s
\\
\nonumber &+{\frac{1}{2}}\int_0^t \nos{ X^{\epsilon,\delta}_{n,m}(s)}_{\mathbb{H}^1_0}\d s.
\end{align}
Let $T_\xi$ be the Yosida-approximation
and $R_{\xi}$ the resolvent of the Dirichlet Laplacian $-\Delta$ on $L^2$, respectively; see Definition~\ref{resolvent and Yosida approximation}.
By the convexity, cf. (\ref{eps.convexity.inequality}), we get
\begin{align*}
&\ska{-\Delta X^{\epsilon,\delta}_{n,m}(s),\div \fe{X^{\epsilon,\delta}_{n,m}(s)}}\\
&=\lim\limits_{\xi \rightarrow \infty}\ska{T_\xi X^{\epsilon,\delta}_{n,m}(s),\div \fe{X^{\epsilon,\delta}_{n,m}(s)}}\\
&=\lim\limits_{\xi \rightarrow \infty} \frac{1}{\xi}\ska{X^{\epsilon,\delta}_{n,m}(s)- R_{\xi}X^{\epsilon,\delta}_{n,m}(s),\div \fe{X^{\epsilon,\delta}_{n,m}(s)}}\\
&=\lim\limits_{\xi\rightarrow \infty} \frac{1}{\xi}\ska{\nabla R_{\xi}X^{\epsilon,\delta}_{n,m}(s)-\nabla X^{\epsilon,\delta}_{n,m}(s),\fe{X^{\epsilon,\delta}_{n,m}(s)}}\\
&\leq \lim\limits_{\xi \rightarrow \infty} \frac{1}{\xi} \left( \mathcal{J}_{\epsilon}( R_{\xi}X^{\epsilon,\delta}_{n,m}(s)) -\mathcal{J}_{\epsilon}( X^{\epsilon,\delta}_{n,m}(s))\right)\\
&\leq 0,
\end{align*}
where we used Proposition \ref{Resolvent_estimate} in the last step above.
{The Burkholder-Davis-Gundy inequaltiy for $p=1$ implies that
\begin{align}\label{BDG nabla}
\mathbb{E}{\sup\limits_{t \in [0,T]} \int_0^t \nos{\nabla X^{\epsilon,\delta}_{n,m}(s)} \d W(s)} \leq& C\mathbb{E}{ \left(\int_0^T \no{\nabla X^{\epsilon,\delta}_{n,m}(s)}^4 \d s\right)^{\frac{1}{2}}} \nonumber \\
\leq& C\mathbb{E}{ \sup\limits_{t \in [0,T]}\no{\nabla X^{\epsilon,\delta}_{n,m}(t)}\left(\int_0^T \no{\nabla X^{\epsilon,\delta}_{n,m}(s)}^2 \d s\right)^{\frac{1}{2}} } \\
\leq& \frac{1}{4} \mathbb{E}{\sup\limits_{t \in [0,T]}\nos{\nabla X^{\epsilon,\delta}_{n,m}(t)}} +C\mathbb{E}{\int_0^T \nos{\nabla X^{\epsilon,\delta}_{n,m}(s)} \d s}\,. \nonumber
\end{align}
After taking supremum over $t$ and expectation in \eqref{ito_nabla}, using \eqref{BDG nabla} along with the Tonelli and Gronwall lemmas we obtain}
\begin{align*}
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{\nabla X^{\epsilon,\delta}_{n,m}(t)} +\delta \int_0^T \nos{\Delta X^{\epsilon,\delta}_{n,m}(s)}\d s} \leq C(\mathbb{E}{\nos{x_0^n}_{\mathbb{H}^1_0}} + \nos{g^n}_{\mathbb{H}^1_0}).
\end{align*}
Hence, from the sequence $\{X^{\epsilon,\delta}_{n,m}\}_{m\in\mathbb{N}}$ we can extract a subsequence (not relabeled), s.t. for $m \rightarrow \infty$
\begin{align*}
&X^{\epsilon,\delta}_{n,m} \rightharpoonup X^{\epsilon,\delta}_{n} ~\text{in}~ L^2(\Omega;L^2((0,T);\mathbb{H}^2)\\
&X^{\epsilon,\delta}_{n,m} \rightharpoonup^* X^{\epsilon,\delta}_{n} ~\text{in}~ L^2(\Omega;L^{\infty}((0,T);\mathbb{H}^1_0)).
\end{align*}
By lower-semicontinuity of the norms, we get
\begin{align*}
\mathbb{E}{\sup\limits_{t \in [0,T]}\frac{1}{2} \nos{\nabla X^{\epsilon,\delta}_{n}(t)} +\delta \int_0^T \nos{\Delta X^{\epsilon,\delta}_{n}(s)}\d s} \leq C(\mathbb{E}{\nos{x_0^n}_{\mathbb{H}^1_0}}+\nos{g^n}_{\mathbb{H}^1_0}) .
\end{align*}
\end{proof}
{We define the functions
\begin{align*}
\bar{ \mathcal{J}}_{\epsilon,\lambda}(x)=
\begin{cases}
\mathcal{J}_{\epsilon,\lambda}(x) + \int_{\partial \O} \be{\gamma_0(x)} \d \mathcal{H}^{n-1} ,~~ \text{for}~ x \in BV(\O)\cap L^2(\O),\\
+\infty ,~~ \text{for}~ x \in BV(\O)\setminus L^2(\O)
\end{cases}
\end{align*}
and
\begin{align*}
\bar{ \mathcal{J}}_\lambda (x)=
\begin{cases}
\mathcal{J}_\lambda (x) + \int_{\partial \O} \be{\gamma_0(x)} \d \mathcal{H}^{n-1} , ~~\text{for}~ x \in BV(\O)\cap L^2(\O),\\
+\infty , ~~\text{for}~ x \in BV(\O)\setminus L^2(\O),
\end{cases}
\end{align*}
where $\gamma_0(x) $ is the trace of $x$ on the boundary and $\d \mathcal{H}^{n-1}$
is the Hausdorff measure.
$\bar{\mathcal{J}}_{\epsilon,\lambda}$ and $\bar{\mathcal{J}}_{\lambda}$ are both convex and lower semicontinuous over $L^2(\O)$ and the lower semicontinuous hulls of $\bar{\mathcal{J}}_{\epsilon,\lambda}\vert_{\mathbb{H}^1_0}$ or $\bar{\mathcal{J}}_{\lambda}\vert_{\mathbb{H}^1_0}$ respectively, cf. \cite[Proposition 11.3.2]{book_attouch}.}
We define the SVI solution of (\ref{reg.TVF}) and (\ref{TVF})
analogically to \cite[Definition~3.1]{Roeckner_TVF_paper} as a stochastic variational inequality.
\begin{defs}\label{def_varsoleps}
Let $0 < T < \infty$, {$\varepsilon \in [0,1]$} and {$x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$ and $g \in \L$}.
Then a $\mathcal{F}_t$-progressively measurable map {$X^{\epsilon} \in L^2(\Omega; C([0,T];\L))\cap L^1(\Omega; L^1((0,T);BV(\O)))$
(denoted by $X \in L^2(\Omega; C([0,T];\L))\cap L^1(\Omega; L^1((0,T);BV(\O)))$ for $\epsilon=0$)}
is called a {SVI solution} of (\ref{reg.TVF}) (or (\ref{TVF}) if $\varepsilon=0$) if $X^{\epsilon}(0)=x_0$ ($X(0)=x_0$), and
for each $(\mathcal{F}_t)$-progressively measurable process $G\in L^2(\Omega \times (0,T),\L) $ and for each $(\mathcal{F}_t)$-adapted $\L$-valued process
$Z \in L^2(\Omega \times (0,T);\mathbb{H}^1_0)$ with $\P$-a.s. continuous sample paths which satisfy the equation
\begin{align}\label{test}
\d Z(t)= -G(t) \d t +Z(t)\d W(t), ~ t\in[0,T],
\end{align}
it holds for {$\epsilon \in (0,1]$} that
\begin{align}\label{reg.SVI}
\frac{1}{2}& \mathbb{E}{\nos{X^{\epsilon}(t)-Z(t)}}+\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\epsilon,\lambda}}(X^{\epsilon}(s)) \d s} \nonumber\\
&\leq \frac{1}{2} \mathbb{E}{\nos{x_0-Z(0)}}+\mathbb{E}{\int_0^t { \bar{\mathcal{J}}_{\epsilon,\lambda}}(Z(s)) \d s} \\
&+ \mathbb{E}{\int_0^t \nos{X^{\epsilon}(s)-Z(s)} \d s}
+\mathbb{E}{\int_0^t \ska{X^{\epsilon}(s)-Z(s),G} \d s}\,,\nonumber
\end{align}
and analogically for $\epsilon=0$ it holds that
\begin{align}\label{SVIeps0}
\frac{1}{2}& \mathbb{E}{\nos{X(t)-Z(t)}}+\mathbb{E}{\int_0^t { \bar{\mathcal{J}}_{\lambda}}(X(s)) \d s} \nonumber\\
&\leq \frac{1}{2} \mathbb{E}{\nos{x_0-Z(0)}}+\mathbb{E}{\int_0^t { \bar{\mathcal{J}}_{\lambda}}(Z(s)) \d s} \\
&+ \mathbb{E}{\int_0^t \nos{X(s)-Z(s)} \d s}
+\mathbb{E}{\int_0^t \ska{X(s)-Z(s),G} \d s}\nonumber.
\end{align}
\end{defs}
In the next theorem we show the existence and uniqueness of a SVI solution to \eqref{reg.TVF} for $\varepsilon > 0$
in the sense of the Definition~\ref{def_varsoleps}.
\begin{thms}\label{Thm_reg.SVI}
Let $0 < T < \infty$ and $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$, $ g \in \L$.
For each $\epsilon\in (0,1]$ there exists a unique SVI solution $X^{\epsilon}$ of (\ref{reg.TVF}).
Moreover, any two SVI solutions $X^{\epsilon}_1,X^{\epsilon}_2$ with $x_0\equiv x^1_0$, $g\equiv g^1$ and $x_0\equiv x^2_0$, $g\equiv g^2$ satisfy
\begin{align}\label{reg_stability_inequality}
&\mathbb{E}{\nos{X^{\epsilon}_1(t)-X^{\epsilon}_2(t)}}\leq C\left(\mathbb{E}{\nos{x^1_0-x^2_0}}+\nos{g^1-g^2} \right )\, ,
\end{align}
for all $t \in [0,T]$.
\end{thms}
\begin{proof}[\textbf{Proof of Theorem \ref{Thm_reg.SVI}}]\hfill \\
We show that for fixed $\epsilon>0$ the sequence $\{X^{\epsilon,\delta}_{n}\}_{\delta,n}$ of variational solutions of (\ref{vis.TVF})
is a Cauchy-sequence w.r.t. $\delta$ for any fixed $n\in \mathbb{N}$,
and then show that it is a Cauchy-sequence w.r.t. $n$ for $\delta\equiv0$.
We denote by $X^{\epsilon,\delta_1}_{n_1},X^{\epsilon,\delta_2}_{n_2}$ the solutions of \eqref{vis.TVF}
for $\delta\equiv\delta_1$, $\delta\equiv \delta_2$
and {$x_0\equiv x^{n_1}_0\in L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$, $x_0\equiv x^{n_2}_0 \in L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$}, respectively,
where $x^{n_1}_0$, $x^{n_2}_0$ belong to the $\mathbb{H}^1_0$-approximating sequence of $x_0\in L^2(\Omega,\mathcal{F}_0;\L)$.
By It\^o's formula it follows that
\begin{align*}
\frac{1}{2}& \nos{X^{\epsilon,\delta_1}_{n_1}(t)-X^{\epsilon,\delta_2}_{n_2}(t)}\qquad\\
& = \frac{1}{2}\nos{x^{n_1}_0-x^{n_2}_0} +\int_0^t \ska{\delta_1\Delta X^{\epsilon,\delta_1}_{n_1}(s)-\delta_2\DeltaX^{\epsilon,\delta_2}_{n_2}(s),X^{\epsilon,\delta_1}_{n_1}(s)-X^{\epsilon,\delta_2}_{n_2}(s)}\d s
\\
&\qquad -\int_0^t \ska{\frac{\nabla X^{\epsilon,\delta_1}_{n_1}(s)}{ \sqrt{ \vert \nabla X^{\epsilon,\delta_2}_{n_2}(s) \vert^2 +\epsilon^2}}-\frac{\nabla X^{\epsilon,\delta_2}_{n_2}(s)}{ \sqrt{ \vert \nabla X^{\epsilon,\delta_2}_{n_2}(s) \vert^2 +\epsilon^2}},\nabla (X^{\epsilon,\delta_1}_{n_1}(s)-X^{\epsilon,\delta_2}_{n_2}(s))}\d s \\
&\qquad -\lambda\int_0^t \nos{(X^{\epsilon,\delta_1}_{n_1}(s)-X^{\epsilon,\delta_2}_{n_2}(s)}\d s +\int_0^t \nos{X^{\epsilon,\delta_1}_{n_1}(s)-X^{\epsilon,\delta_2}_{n_2}(s)}\d W(s)\\
&\qquad +\int_0^t \nos{X^{\epsilon,\delta_1}_{n_1}(s)-X^{\epsilon,\delta_2}_{n_2}(s)}\d s.
\end{align*}
We note that
\begin{align*}
&\ska{\delta_1\Delta X^{\epsilon,\delta_1}_{n_1}(s)-\delta_2\DeltaX^{\epsilon,\delta_2}_{n_2}(s),X^{\epsilon,\delta_1}_{n_1}(s)-X^{\epsilon,\delta_2}_{n_2}(s)}\\
&=-\ska{\delta_1 \nabla X^{\epsilon,\delta_1}_{n_1}(s)-\delta_2\nablaX^{\epsilon,\delta_2}_{n_2}(s),\nablaX^{\epsilon,\delta_1}_{n_1}(s)-\nablaX^{\epsilon,\delta_2}_{n_2}(s)}\\
&\leq C(\delta_1+\delta_2)(\nos{\nabla X^{\epsilon,\delta_1}_{n_1}(s)}+\nos{\nabla X^{\epsilon,\delta_2}_{n_2}(s)}).
\end{align*}
Hence by using the convexity \eqref{eps.convexity.inequality}, Lemma \ref{laplace_energy_estimate}, the Burkholder-Davis-Gundy inequality for $p=1$,
the Tonelli and Gronwall lemmas we obtain
{
\begin{align}\label{delta_datum_estimate}
& \mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon,\delta_1}_{n_1}(t)-X^{\epsilon,\delta_2}_{n_2}(t)}}
\leq C\mathbb{E}{\nos{x_0^{n_1}-x_0^{n_2}}}
\nonumber \\
&\qquad +C\Big(\mathbb{E}{\nos{x_0^{n_1}}_{\mathbb{H}^1_0}},\mathbb{E}{\nos{x_0^{n_2}}_{\mathbb{H}^1_0}},\nos{g^n}_{\mathbb{H}^1_0}\Big)(\delta_1+\delta_2).
\end{align}
}
Inequality \eqref{delta_datum_estimate} implies for $x_0^{n_1} \equiv x_0^{n_2}\equiv x_0^n$ that
\begin{align*}
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon,\delta_1}_{n}(t)-X^{\epsilon,\delta_2}_{n}(t)}}
\leq C\left(\mathbb{E}{\nos{x_0^n}_{\mathbb{H}^1_0}},\nos{g^n}_{\mathbb{H}^1_0}\right)(\delta_1+\delta_2).
\end{align*}
Hence for any fixed $n$, $\epsilon$ there exists a $\{\mathcal{F}_t \}$-adapted process $X^{\epsilon}_n \in L^2(\Omega,C([0,T];\L))$, s.t.
\begin{align}\label{delta_limit}
\lim_{\delta \rightarrow 0} \mathbb{E}{\sup\limits_{t \in [0,T]}\nos{X^{\epsilon,\delta}_{n}(t)-X^{\epsilon}_n(t)}} \rightarrow 0.
\end{align}
For fixed $n_1$, $n_2$, $\epsilon$
we get from \eqref{delta_datum_estimate} using (\ref{delta_limit})
by the lower-semicontinuity of norms that
\begin{align}\label{lim_n}
&\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon}_{n_1}(t)-X^{\epsilon}_{n_2}}}
\leq
\liminf_{\delta\rightarrow 0}\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon,\delta}_{n_1}(t)-X^{\epsilon,\delta}_{n_2}}}
\\ \nonumber & \qquad
\leq\frac{1}{2}\mathbb{E}{\nos{x_0^{n_1}-x_0^{n_2}}}.
\end{align}
Since $x_0^{n_1},\,x_0^{n_2}\rightarrow x_0$ for $n_1,n_2\rightarrow \infty$
we deduce from (\ref{lim_n}) that for any fixed $\epsilon$
there exists an $\{\mathcal{F}_t \}$-adapted process $X^{\epsilon} \in L^2(\Omega;C([0,T];\L))$ such that
\begin{align}\label{lim_nnew}
\lim_{n \rightarrow \infty} \mathbb{E}{\sup\limits_{t \in [0,T]}\nos{X^{\epsilon}(t)-X^{\epsilon}_n(t)}} \rightarrow 0.
\end{align}
In the next step, we show that the limiting process $X^{\epsilon}$ is a SVI solution of \eqref{reg.TVF}.
We subtract the process
\begin{align*}
\d Z(t)= -G(t)\d t + Z(t)\d W(t)\,,
\end{align*}
with $Z(t)=z_0$ from (\ref{vis.TVF}) and obtain
\begin{align*}
\d \left(X^{\epsilon,\delta}_{n}(t) -Z(t) \right)= \left( -A^{\epsilon,\delta} X^{\epsilon,\delta}_{n}(t)+G(t) \right)\d t + \left(X^{\epsilon,\delta}_{n}(t)-Z(t)\right)\d W(t).
\end{align*}
The It\^o formula implies
\begin{align}\label{ito_eps}
\frac{1}{2}&\mathbb{E}{\nos{X^{\epsilon,\delta}_{n}(t)-Z(t)}}
\nonumber \\
=&\frac{1}{2}\mathbb{E}{\nos{X^{\epsilon,\delta}_{n}(0)-z_0}}-\mathbb{E}{\int_0^t \langleA^{\epsilon,\delta} X^{\epsilon,\delta}_{n}(s),X^{\epsilon,\delta}_{n}(s)-Z(s) \rangle \d s}
\\
&+\mathbb{E}{\int_0^t \ska{G(s),X^{\epsilon,\delta}_{n}(s)-Z(s)}\d s}
+ \mathbb{E}{\int_0^t \nos{X^{\epsilon,\delta}_{n}(s)-Z(s)} \d s}.
\nonumber
\end{align}
We rewrite the second term on the right-hand side in above inequality as
\begin{align*}
&\mathbb{E}{\int_0^t \langleA^{\epsilon,\delta} X^{\epsilon,\delta}_{n}(s),X^{\epsilon,\delta}_{n}(s)-Z(s)\rangle \d s}\\
&=\mathbb{E}{\int_0^t \delta(\nabla X^{\epsilon,\delta}_{n}(s),\nabla (X^{\epsilon,\delta}_{n}(s) - Z(s))) \d s}
\\&\quad +\mathbb{E}{\int_0^t \left(\fe{X^{\epsilon,\delta}_{n}(s)},\nabla \big(X^{\epsilon,\delta}_{n}(s)-Z(s)\big)\right)\d s}
\\ &\quad + \mathbb{E}{\int_0^t \lambda(X^{\epsilon,\delta}_{n}(s)-{g^n},X^{\epsilon,\delta}_{n}(s)-Z(s)) \d s}.
\end{align*}
The convexity of $\mathcal{J}_\epsilon$ along with the Cauchy-Schwarz and Young's inequalities imply that
\begin{align*}
&\mathbb{E}{\int_0^t (\fe{X^{\epsilon,\delta}_{n}(s)},\nabla (X^{\epsilon,\delta}_{n}(s)-Z(s)))\d s}
\\ & +\mathbb{E}{\int_0^t \lambda(X^{\epsilon,\delta}_{n}(s)-{g^n},X^{\epsilon,\delta}_{n}(s)-Z(s)) \d s}
\\& \geq \mathbb{E}{\int_0^t \mathcal{J}_{\epsilon}(X^{\epsilon,\delta}_{n}(s))-\mathcal{J}_{\epsilon}(Z(s))\d s}
\\ & +\mathbb{E}{\int_0^t \frac{\lambda}{2}\nos{X^{\epsilon,\delta}_{n}(s)-{g^n}}-\frac{\lambda}{2}\nos{Z(s)-{g^n}} \d s} .
\end{align*}
By combining two inequalities above with (\ref{ito_eps}) we get
\begin{align}\label{est_svidelta1}
&\frac{1}{2}\mathbb{E}{\nos{X^{\epsilon,\delta}_{n}(t)-Z(t)}}+\mathbb{E}{\int_0^t {\mathcal{J}_{\epsilon}(X^{\epsilon,\delta}_{n}(s))\d s}+\frac{\lambda}{2}\nos{X^{\epsilon,\delta}_{n}(s)-{g^n}} \d s}
\nonumber \\
&+ \frac{\delta}{2}\mathbb{E}{\int_0^t \nos{\nabla X^{\epsilon,\delta}_{n}(s)})\d s}
\nonumber \\
& \leq \frac{1}{2}\mathbb{E}{\nos{X^{\epsilon,\delta}_{n}(0)-Z(0)}} + \mathbb{E}{\int_0^t {\mathcal{J}_{\epsilon}(Z(s))\d s}+\frac{\lambda}{2}\nos{Z(s)-{g^n}}\d s}
\\
&\quad + \frac{\delta}{2}\mathbb{E}{\int_0^t \nos{\nabla Z(s)})\d s}+\mathbb{E}{\int_0^t \ska{G(s),X^{\epsilon,\delta}_{n}(s)-Z(s)}\d s}
\nonumber \\
&\quad +\mathbb{E}{\int_0^t \nos{X^{\epsilon,\delta}_{n}(s)-Z(s)} \d s}. \nonumber
\end{align}
{Since $X^{\epsilon,\delta}_n \in \mathbb{H}^1_0$ and $Z \in \mathbb{H}^1_0$ it holds that $\mathcal{J}_{\epsilon,\lambda}(X^{\epsilon,\delta}_n)=\bar{\mathcal{J}}_{\epsilon,\lambda}(X^{\epsilon,\delta}_n)$ and $\mathcal{J}_{\epsilon,\lambda}(Z)=\bar{\mathcal{J}}_{\epsilon,\lambda}(Z)$}. The lower-semicontinuity of {{$\bar{\mathcal{J}}_{\epsilon,\lambda}$ in $BV(\O)$}} with respect to convergence in $\mathbb{L}^1$, cf. \cite{Ambrosio_functions_of_BV},
and (\ref{delta_limit}), (\ref{lim_nnew}) {and the strong convergence $g^n \rightarrow g $ in $\L$}
imply that for $\delta \rightarrow 0$ and $n \rightarrow \infty$
the limiting process $X^{\epsilon} \in L^2(\Omega; C([0,T];\L)$ satisfies \eqref{reg.SVI}.
To conclude that $X^{\epsilon}$ is a SVI solution of (\ref{reg.TVF}) it remains to show that $X^{\epsilon} \in L^1(\Omega; L^1((0,T);BV(\O)))$.
Setting $G\equiv 0$ in \eqref{test} (which implies $Z\equiv 0$ by (\ref{test})) yields
\begin{align}\label{cont_ener}
\frac{1}{2}& \mathbb{E}{\nos{X^{\epsilon}(t)}}+\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\epsilon,\lambda}}(X^{\epsilon}(s)) \d s} \nonumber\\
&\leq \frac{1}{2} \mathbb{E}{\nos{x_0}}+\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\epsilon,\lambda}}(0) \d s}
+ \mathbb{E}{\int_0^t \nos{X^{\epsilon}(s)} \d s}\,.
\end{align}
On noting that (cf. Definition~\ref{Bounded Variation} or \cite[proof of Theorem 1.3]{Prohl_TVF_numerics})
\begin{align*}
{\bar{\mathcal{J}}_{\epsilon,\lambda}}(X^{\epsilon})
\geq& {\bar{\mathcal{J}}_{\lambda}}(X^{\epsilon}),
\end{align*}
and
${\bar{\mathcal{J}}_{\epsilon,\lambda}}(0)=\epsilon\be{\O} + \frac{\lambda}{2}\nos{g}$,
we deduce from (\ref{cont_ener}) that
\begin{align*}
\frac{1}{2}& \mathbb{E}{\nos{X^{\epsilon}(t)}}+\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\lambda}}(X^{\epsilon}(s)) \d s} \nonumber\\
&\leq \frac{1}{2} \mathbb{E}{\nos{x_0}}+\mathbb{E}{\int_0^t\epsilon\Big(\be{\O} + \frac{\lambda}{2}\nos{g}\Big) \d s}
+ \mathbb{E}{\int_0^t \nos{X^{\epsilon}(s)} \d s}\,.\nonumber
\end{align*}
Hence, by the Tonelli and Gronwall lemmas it follows that
\begin{align}\label{BV-eps-esti.}
\frac{1}{2} \mathbb{E}{\nos{X^{\epsilon}(t)}}&+\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\lambda}}(X^{\epsilon}(s)) \d s} \nonumber \\
&\leq C_T\exp(T)\left(\mathbb{E}{\nos{x_0}}+ \be{\O} + \lambda\nos{g} \right).
\end{align}
Hence $X^{\epsilon} \in L^2(\Omega; C([0,T];\L))\cap L^1(\Omega; L^1((0,T);BV(\O)))$ is a SVI solution of (\ref{reg.TVF}) for $\epsilon \in (0,1]$.
In the next step we show the uniqueness of the SVI solution.
Let $X^{\epsilon}_1, X^{\epsilon}_2$ be two SVI solutions to \eqref{reg.TVF}
for a fixed $\epsilon\in(0,1]$
with initial values
$x_0\equiv x^1_0, x^2_0$ and $g \equiv g^1,g^2$, respectively. Let $\{x^{2,n}_0\}_{n\in\mathbb{N}} \subset L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$ be a sequence,
s.t. $x^{2,n}_0 \rightarrow x^2_0$ in $L^2(\Omega,\mathcal{F}_0;\L)$ and {$\{g^{2,n}\}_{n\in\mathbb{N}} \subset \mathbb{H}^1_0$ be a sequence,
s.t. $g^{2,n}_0 \rightarrow g^2$ in $\L$
for $n \rightarrow \infty$ and let $\{X^{\epsilon,\delta}_{2,n}\}_{n\in\mathbb{N},\delta>0}$
be a sequence of variational solutions of \eqref{vis.TVF} (for fixed $\epsilon>0$)
with $x_0\equiv x^{2,n}_0$, $g\equiv g^{2,n}$}.
We note that the first part of the proof implies that $X^{\epsilon,\delta}_{2,n} \rightarrow X^{\epsilon}_2$
in $L^2(\Omega; C([0,T];\L)$ for $\delta\rightarrow 0$, $n\rightarrow \infty$.
We set $Z=X^{\epsilon,\delta}_{2,n}, G=A^{\epsilon,\delta}(X^{\epsilon,\delta}_{2,n})$ in (\ref{reg.SVI})
and observe that
\begin{align}\label{eps to 0 inequality}
\frac{1}{2} \mathbb{E}{\nos{X^{\epsilon}_1(t)-X^{\epsilon,\delta}_{2,n}(t)}}+&\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\epsilon,\lambda}}(X^{\epsilon}_1(s)) \d s} \nonumber
\\
&\leq \frac{1}{2} \mathbb{E}{\nos{x^1-x^{2,n}_0}}+\mathbb{E}{\int_0^t {\bar{\mathcal{J}}_{\epsilon,\lambda}}(X^{\epsilon,\delta}_{2,n}(s)) \d s} \nonumber
\\
&-{\mathbb{E}{\delta \int_0^t \ska{ X^{\epsilon}_1(s)- X^{\epsilon,\delta}_{2,n}(s), \Delta X^{\epsilon,\delta}_{2,n}(s))}} \d s}
\\
&-{\mathbb{E}{\int_0^t \ska{ X^{\epsilon}_1 (s)- X^{\epsilon,\delta}_{2,n}(s), \div\fe{X^{\epsilon,\delta}_{2,n}(s)} }\d s}} \nonumber
\\ \nonumber
&+ \mathbb{E}{\int_0^t \lambda\ska{X^{\epsilon}_1(s)- X^{\epsilon,\delta}_{2,n}(s),X^{\epsilon,\delta}_{2,n}(s)-{g^{2,n}}} \d s}
\\ \nonumber
&+ \mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-X^{\epsilon,\delta}_{2,n}(s)} \d s}
\\\nonumber
& := I+II+III+IV+V+VI.
\end{align}
{The term $III$ is estimated using Young's inequality as
\begin{align*}
III
&\leq C\mathbb{E}{\int_0^t \delta^{\frac{2}{3}}\nos{ X^{\epsilon}_1(s)-X^{\epsilon,\delta}_{2,n}(s)}+\delta^{\frac{4}{3}} \nos{\Delta X^{\epsilon,\delta}_{2,n}(s))} \d s}\,.
\end{align*}
{We have to show the following estimate
\begin{align*}
IV=\ska{X_1^{\epsilon}-X^{\epsilon,\delta}_{2,n},-\div \fe{X^{\epsilon,\delta}_{2,n}}} \leq \bar{\mathcal{J}}_{\epsilon,0}(X^{\epsilon}_1)-\bar{\mathcal{J}}_{\epsilon,0}(X^{\epsilon,\delta}_{2,n}).
\end{align*}
We consider an approximating sequence $x_k \in C^{\infty}(\O)\cap BV(\O)$, s.t., $x_k \rightarrow X_1^{\epsilon}$ strongly in $L^1(\O)$ and $\mathcal{J}_{\epsilon,\lambda}(x_k) \rightarrow \mathcal{J}_{\epsilon,\lambda}(X_1^{\epsilon})$ for $k \rightarrow \infty$, cf. \cite[Theorem 13.4.1]{book_attouch}.
Integration by parts then yields
\begin{align*}
&\ska{x_k-X^{\epsilon,\delta}_{2,n},-\div \fe{X^{\epsilon,\delta}_{2,n}}}
=\ska{\nabla(x_k-X^{\epsilon,\delta}_{2,n}), \fe{X^{\epsilon,\delta}_{2,n}}}
\\
&+\int_{\partial \O} \gamma_0(x_k) \fe{X^{\epsilon,\delta}_{2,n}}\nu \d \mathcal{H}^{n-1}-\int_{\partial \O} \gamma_0(X^{\epsilon,\delta}_{2,n}) \fe{X^{\epsilon,\delta}_{2,n}}\nu \d \mathcal{H}^{n-1}\,,
\end{align*}
where $\nu $ is the outer normal vector at $\mathcal{H}^{n-1}$ almost all $x \in \partial \O$. Since $X^{\epsilon,\delta}_{2,n}(\omega,t)\in\mathbb{H}^1_0$ for almost all $(\omega,t) \in \Omega \times [0,T]$, the second boundary integral vanishes.
We estimate the first boundary integral as
\begin{align*}
\int_{\partial \O} \gamma_0(x_k) \fe{X^{\epsilon,\delta}_{2,n}}\nu \d \mathcal{H}^{n-1}
\leq& \int_{\partial \O} \be{\gamma_0(x_k)} \be{\fe{X^{\epsilon,\delta}_{2,n}}\nu} \ d\mathcal{H}^{n-1}
\\
\leq&\int_{\partial \O} \be{\gamma_0(x_k)}\d \mathcal{H}^{n-1}.
\end{align*}
{On noting that $\no{X_1^{\epsilon}(\omega,t)} \leq C$ for a.a. $(\omega,t) \in \Omega \times (0,T)$, the convergence $x_k \rightarrow X_1^{\epsilon} $ for $k \rightarrow \infty$ in $L^1(\O)$ implies $x_k \rightharpoonup X_1^{\epsilon}$ in $\L$
a.e. in $\Omega \times (0,T)$.}
We further assert that the trace of each approximating function $x_k \in C^{\infty}(\O)\cap BV(\O)$, coincides with the trace of $X_1^{\epsilon}$ on the boundary of $\O$, see \cite[Remark 10.2.1]{book_attouch}.
Hence we obtain by taking the limit for $k\rightarrow \infty$ that
\begin{align*}
\ska{X^{\epsilon}_1-X^{\epsilon,\delta}_{2,n},-\div \fe{X^{\epsilon,\delta}_{2,n}}}&\leq \bar{\mathcal{J}}_{\epsilon,0}(X^{\epsilon}_1)-\bar{\mathcal{J}}_{\epsilon,0}(X^{\epsilon,\delta}_{2,n}).
\end{align*}
}
Next, we obtain
\begin{align*}
V
& = \lambda \mathbb{E}{\int_0^t \ska{X^{\epsilon}_1(s)- X^{\epsilon,\delta}_{2,n}(s),X^{\epsilon,\delta}_{2,n}(s)-{g^{2,n}}} \d s}
\\
&\leq \frac{\lambda}{2}\mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-{g^{2,n}}}-\nos{X^{\epsilon,\delta}_{2,n}(s)-{g^{2,n}}}\d s}.
\end{align*}
}
After substituting $III$-$V$ into \eqref{eps to 0 inequality} we arrive at
\begin{align}\label{Uniquness.eps to 0}
\frac{1}{2} \mathbb{E}{\nos{X^{\epsilon}_1(t)-X^{\epsilon,\delta}_{2,n}(t)}}&+\frac{\lambda}{2}\mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-g^1}\d s} \nonumber
\\ \nonumber
&\leq \frac{1}{2} \mathbb{E}{\nos{x^1_0-x^{2,n}_0}} +\frac{\lambda}{2}\mathbb{E}{\int_0^t \nos{X^{\epsilon,\delta}_{2,n}(s)-g^1}\d s}
\\\nonumber
&+ C\mathbb{E}{\int_0^t \delta^{\frac{2}{3}}\nos{ X^{\epsilon}_1(s)-X^{\epsilon,\delta}_{2,n}(s)}+\delta^{\frac{4}{3}} \nos{\Delta X^{\epsilon,\delta}_{2,n}(s))} \d s}
\\
& \frac{\lambda}{2}\mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-{g^{2,n}}}\d s} - \frac{\lambda}{2}\mathbb{E}{\int_0^t \nos{X^{\epsilon,\delta}_{2,n}(s)-{g^{2,n}}}\d s}
\\\nonumber
& +\mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-X^{\epsilon,\delta}_{2,n}(s)} \d s}
.
\end{align}
The convergences (\ref{delta_limit}), (\ref{lim_nnew}) imply the convergence $X^{\epsilon,\delta}_{2,n} \rightarrow X^{\epsilon}_2$ in $L^2(\Omega;C([0,T];\L))$
for $\delta \rightarrow 0$, $n \rightarrow \infty$.
We note that for $\delta \rightarrow 0$ the fourth term on the right-hand side of (\ref{Uniquness.eps to 0}) vanishes due to Lemma \ref{laplace_energy_estimate}.
Hence, by taking the limits for $\delta \rightarrow 0$, $n \rightarrow \infty$ in \eqref{Uniquness.eps to 0},
{using the strong convergence $g^{2,n} \rightarrow g^2 $ in $\L$ for $n \rightarrow \infty$} , the lower-semicontinuity of norms
and (\ref{delta_limit}), (\ref{lim_nnew}) we obtain
\begin{align*}
&\mathbb{E}{\nos{X^{\epsilon}_1(t)-X^{\epsilon}_2(t)}}\leq C \mathbb{E}{\nos{x^1_0-x^2_0}}
\\
&+\frac{\lambda}{2}\mathbb{E}{\int_0^t\nos{X^{\epsilon}_1(s)-g^2}+\nos{X^{\epsilon}_2(s)-g^1}-\nos{X^{\epsilon}_{1}(s)-g^1}-\nos{X^{\epsilon}_2(s)-g^2} \d s}
\\
&+\mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-{X^{\epsilon}_{2}(s)}} \d s}
\\
&\leq C\left(\mathbb{E}{\nos{x^1_0-x^2_0}}+\mathbb{E}{\int_0^t \nos{X^{\epsilon}_1(s)-X^{\epsilon}_2(s)} \d s} +\nos{g^1-g^2} \right )\, ,
\end{align*}
for all $t \in [0,T]$.
After applying the Tonelli and Gronwall lemmas we obtain \eqref{reg_stability_inequality}.
\end{proof}
Our second main theorem establishes existence and uniqueness of a SVI solution to \eqref{TVF}
in the sense of Definition~\ref{def_varsoleps}.
The solution is obtained as a limit of solutions of the regularized gradient flow \eqref{reg.TVF}
for $\epsilon \rightarrow 0$.
\begin{thms}\label{Thm.SVI}
Let $0 < T < \infty$ and {$x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$, $ g \in \L$} be fixed.
Let $\{X^{\epsilon}\}_{\epsilon>0}$ be the {SVI} solutions of \eqref{reg.TVF} for $\epsilon\in(0,1]$.
Then $X^{\epsilon}$ converges to the unique SVI variational solution $X$ of (\ref{TVF})
in $L^2(\Omega;C([0,T];\L))$ for $\epsilon\rightarrow 0$, i.e., there holds
\begin{align}\label{epsilon goes to 0}
\lim\limits_{\epsilon \rightarrow 0} \mathbb{E}{\sup\limits_{t \in [0,T]}\nos{X^{\epsilon}(t)-X(t)}}=0.
\end{align}
Furthermore, the following estimate holds
\begin{align}\label{stability_inequality}
&\mathbb{E}{\nos{X_1(t)-X_2(t)}}\leq C\left(\mathbb{E}{\nos{x^1_0-x^2_0}}+\nos{g^1-g^2} \right)\quad \mathrm{for\,\,all\,\,} t \in [0,T]\,,
\end{align}
where $X_1$ and $X_2$ are SVI solutions of (\ref{TVF}) with $x_0\equiv x^1_0$, $g\equiv g^1$ and $x_0\equiv x^2_0$, $g\equiv g^2$, respectively.
\end{thms}
\begin{proof}[\textbf{Proof of Theorem \ref{Thm.SVI}}]\hfill \\
We consider $\L$-approximating sequences {$\{x_0^n\}_{n\in\mathbb{N}} \subset L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$
and $\{g^n\}_{n\in\mathbb{N}} \subset \mathbb{H}^1_0$ of the initial condition $x_0\in L^2(\Omega,\mathcal{F}_0;\L)$ and $g \in \L$, respectively.}
For $n\in\mathbb{N}$, $\delta>0$ we denote by $X^{\epsilon_1,\delta}_{n},X^{\epsilon_2,\delta}_{n}$ the variational solutions of \eqref{vis.TVF}
with $\epsilon\equiv\epsilon_1$, $\epsilon\equiv\epsilon_2$, respectively.
By It\^o's formula the difference satisfies
\begin{align}\label{eps.difference}
\frac{1}{2}& \nos{X^{\epsilon_1,\delta}_{n}(t)-X^{\epsilon_2,\delta}_{n}(t)}
\nonumber\\
=& -\delta \int_0^t \nos{\nabla (X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s))}\d s
\nonumber \\
&-\int_0^t \ska{\frac{\nabla X^{\epsilon_1,\delta}_{n}(s)}{ \sqrt{ \vert \nabla X^{\epsilon_1,\delta}_{n}(s) \vert^2 +\epsilon_1^2}}-\frac{\nabla X^{\epsilon_2,\delta}_{n}(s)}{ \sqrt{ \vert \nabla X^{\epsilon_2,\delta}_{n}(s) \vert^2 +\epsilon_2^2}},\nabla (X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s))}\d s
\\
&-\lambda\int_0^t \nos{(X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s)}\d s +\int_0^t \nos{X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s)}\d W(s)\nonumber
\\
&+\int_0^t \nos{X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s)} \nonumber\d s.
\end{align}
We estimate the second term on the right-hand side of \eqref{eps.difference} using the convexity \eqref{eps.convexity.inequality}
\begin{align}\label{est_convex}
& \ska{\frac{\nabla X^{\epsilon_1,\delta}_{n}(s)}{ \sqrt{ \vert \nabla X^{\epsilon_1,\delta}_{n}(s) \vert^2 +\epsilon_1^2}}-\frac{\nabla X^{\epsilon_2,\delta}_{n}(s)}{ \sqrt{ \vert \nabla X^{\epsilon_2,\delta}_{n}(s) \vert^2 +\epsilon_2^2}},\nabla (X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s))}
\nonumber \\
=&\ska{\frac{\nabla X^{\epsilon_1,\delta}_{n}(s)}{ \sqrt{ \vert \nabla X^{\epsilon_1,\delta}_{n}(s) \vert^2 +\epsilon_1^2}},\nabla (X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s))}
\nonumber \\
&+\ska{\frac{\nabla X^{\epsilon_2,\delta}_{n}(s)}{ \sqrt{ \vert \nabla X^{\epsilon_2,\delta}_{n}(s) \vert^2 +\epsilon_2^2}},\nabla (X^{\epsilon_2,\delta}_{n}(s)-X^{\epsilon_1,\delta}_{n}(s))}
\\
\geq& \int\limits_{\mathcal{O}} \sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}-\sqrt{\bes{\nabla X^{\epsilon_2,\delta}_{n}}+\epsilon_1^2} \d x
\nonumber\\
\nonumber
&+\int\limits_{\mathcal{O}} \sqrt{\bes{\nabla X^{\epsilon_2,\delta}_{n}}+\epsilon_2^2}-\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2} \d x.
\end{align}
Next, we observe that
\begin{align*}
\int\limits_{\mathcal{O}}& \left(\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}-\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}\right) \d x
\\
&=\int\limits_{\mathcal{O}} \frac{\left(\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}-\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}\right)\left(\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}+\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}\right)}{\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}+\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}}\d x
\\
&= \int\limits_{\mathcal{O}} \frac{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2-\bes{\nabla X^{\epsilon_1,\delta}_{n}}-\epsilon_2^2}{\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}+\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}}\d x
\\
&= \int\limits_{\mathcal{O}} \frac{(\epsilon_1+\epsilon_2)(\epsilon_1-\epsilon_2)}{\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}+\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}}\d x
\\
&\leq \int\limits_{\mathcal{O}} \be{\epsilon_1 -\epsilon_2}\left( \frac{\epsilon_1}{\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}}+ \frac{\epsilon_2}{\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2}} \right)\d x
\leq C(\epsilon_1+\epsilon_2).
\end{align*}
Using the inequality above, we get
\begin{align*}
&\int\limits_{\mathcal{O}} \sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}-\sqrt{\bes{\nabla X^{\epsilon_2,\delta}_{n}}+\epsilon_2^2} \d x
+\int\limits_{\mathcal{O}} \sqrt{\bes{\nabla X^{\epsilon_2,\delta}_{n}}+\epsilon_2^2}-\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2} \d x\\
&\geq - \left|\int\limits_{\mathcal{O}} \sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_1^2}-\sqrt{\bes{\nabla X^{\epsilon_1,\delta}_{n}}+\epsilon_2^2} \d x \right|\\
&-\left|\int\limits_{\mathcal{O}} \sqrt{\bes{\nabla X^{\epsilon_2,\delta}_{n}}+\epsilon_1^2}-\sqrt{\bes{\nabla X^{\epsilon_2,\delta}_{n}}+\epsilon_2^2} \d x\right| \\
&\geq -C(\epsilon_1+\epsilon_2).
\end{align*}
Substituting (\ref{est_convex}) along with the last inequality into \eqref{eps.difference} yields
\begin{align}\label{eps.estimate}
\frac{1}{2} \nos{X^{\epsilon_1,\delta}_{n}(t)-X^{\epsilon_2,\delta}_{n}(t)} \leq& C(\epsilon_1+\epsilon_2) \nonumber\\
&+\int_0^t \nos{X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s)}\d W(s)\\
&+\int_0^t \nos{X^{\epsilon_1,\delta}_{n}(s)-X^{\epsilon_2,\delta}_{n}(s)}\nonumber\d s.
\end{align}
After using the Burkholder-Davis-Gundy inequality for $p=1$, the Tonelli and Gronwall lemmas we obtain that
\begin{align}\label{Cauchy eps}
\mathbb{E}{\sup\limits_{t \in [0,T]}\nos{X^{\epsilon_1,\delta}_{n}(t)-X^{\epsilon_2,\delta}_{n}(t)}}\leq C(\epsilon_1+\epsilon_2).
\end{align}
We take the limit for $\delta\rightarrow 0$ in \eqref{Cauchy eps}
for fixed $n$ and $\epsilon_1,\epsilon_2$,
and obtain using \eqref{delta_limit} by the lower-semicontinuity of norms that
\begin{align}\label{delta to 0}
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon_1}_n(t)-X^{\epsilon_2}_n(t)}}
\leq& \liminf \limits_{\delta \rightarrow 0 }
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon_1,\delta}_n(t)-X^{\epsilon_2,\delta}_n(t)}} \nonumber \\
\leq& C(\epsilon_1+\epsilon_2).
\end{align}
Hence, by \eqref{lim_nnew} and the lower-semicontinuity of norms, after taking the limit $n\rightarrow \infty$ in \eqref{delta to 0}
for fixed $\epsilon_1$, $\epsilon_2$ we get
\begin{align}\label{n to infty}
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon_1}(t)-X^{\epsilon_2}(t)}}
\leq& \liminf \limits_{n \rightarrow \infty }
\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon_1}_n(t)-X^{\epsilon_2}_n(t)}}
\\ \nonumber
\leq& C(\epsilon_1+\epsilon_2).
\end{align}
The above inequality implies that $\{X^{\epsilon}\}_{\epsilon>0}$ is a Cauchy Sequence in $\epsilon$.
Consequently there exists a unique $\{\mathcal{F}_t\}$-adapted process $X \in L^2(\Omega;C([0,T];\L))$ with $X(0)=x_0 $ such that
\begin{align}\label{epsilon gegen 0}
\lim\limits_{\epsilon \rightarrow 0}\mathbb{E}{\sup\limits_{t \in [0,T]} \nos{X^{\epsilon}(t)-X(t)}}
= 0.
\end{align}
This concludes the proof of \eqref{epsilon goes to 0}.
Next, we show that the limiting process $X$ is the SVI solution of (\ref{TVF}), i.e., we show that (\ref{SVIeps0}) holds.
We note that \eqref{BV-eps-esti.} implies that
\begin{align}\label{BVestimate}
\sup\limits_{\epsilon \in (0,1]} \mathbb{E}{\int_0^t \bar{\mathcal{J}}_{\lambda}(X^{\epsilon}(s)\d s } \leq C.
\end{align}
Hence using \eqref{epsilon gegen 0}, \eqref{BVestimate} we get by
Fatou's lemma and \cite[Proposition 11.3.2]{book_attouch} that
\begin{align*}
\liminf\limits_{\epsilon \rightarrow 0} \mathbb{E}{ \int_0^t \bar{\mathcal{J}}_{\lambda}(X^{\epsilon}(s)) \d s } \geq \mathbb{E} { \int_0^t \bar{\mathcal{J}}_{\lambda}(X(s)) \d s} .
\end{align*}
By Theorem~\ref{Thm_reg.SVI} we know that $X^{\epsilon}$ satisfies (\ref{reg.SVI}) for any $\epsilon \in (0,1]$.
By taking the limit for $\epsilon\rightarrow 0$ in (\ref{reg.SVI}),
using the above inequality and (\ref{epsilon gegen 0})
it follows that $X$ satisfies $\eqref{SVIeps0}$.
Finally, inequality $\eqref{stability_inequality}$ follows after taking the limit for $\epsilon \rightarrow 0$ in
(\ref{epsilon gegen 0}), by \eqref{reg_stability_inequality} and the lower semicontinuity of norms.
\end{proof}
\section{Numerical Approximation}\label{sec_num}
We construct a fully-discrete approximation of the STVF equation \eqref{TVF}
via an implicit time-discretization of the regularized STVF equation \eqref{reg.TVF}.
For $N \in \mathbb{N}$ we consider the time-step $\tau := T/N$,
set $t_i:=i\tau$ for $i=0,\ldots,N$ and denote the discrete Wiener increments as $\Delta_i W:= W(t_i)-W(t_{i-1})$.
We combine the discretization in time with a the standard $\mathbb{H}^1_0$-conforming finite element method,
see, e.g., \cite{BrennerS02}, \cite{Prohl_TVF_numerics}, \cite{bm16}.
Given a family of quasi-uniform triangulations $\big\{\mathcal{T}_h\big\}_{h>0}$ of $\O$ into open simplices with mesh size $h=\max_{K\in \mathcal{T}_h}\{\mathrm{diam}(K)\}$
we consider the associated space of piecewise linear, globally continuous functions
$\mathbb{V}_h = \{v_h \in C^0(\overline{\O});\, v_h|_K \in \mathcal{P}^1(K)\,\, \forall K\in \mathcal{T}_h\}\subset \mathbb{H}^1_0$
and set $L\equiv\text{dim}\mathbb{V}_h$ for the rest of the paper.
We set $X^h_0:=\mathcal{P}_h x_0$, $g^h:=\mathcal{P}_h g$, where $\mathcal{P}_h$ is the $\L$-projection onto $\mathbb{V}_h$.
The implicit fully-discrete approximation of (\ref{reg.TVF}) is defined as follows:
fix $N\in\mathbb{N}$, $h>0$ set $X^0_\varepsilon = x^h_0\in\mathbb{V}_h$ and determine $\Xi\in \mathbb{V}_h$, $i=1,\dots, N$ as the solution of
\begin{align}\label{num.reg.TVF}
\ska{\Xi,v_h}&=\ska{X_{\epsilon,h}^{i-1},v_h}-\tau \ska{\fe{\Xi},\nablav_h } \\
&-\tau\lambda\ska{\Xi -g^h,v_h}+\ska{X_{\epsilon,h}^{i-1},v_h}\Delta_i W &&\forall v_h \in \mathbb{V}_h \nonumber.
\end{align}
\begin{comment}
\begin{bems}
Throughout this section, to simplify the notation,
whenever we take the limit w.r.t. to the time-step $\tau$ we implicitly assume that
the limit is simultaneously taken with respect to the spatial discretization parameter $h$, i.e.,
we will denote by $\displaystyle \lim_{\tau\rightarrow 0}\equiv \displaystyle \lim_{\tau\rightarrow 0, h \rightarrow 0}$.
We note that there is no explicit coupling between the time-step $\tau$ and the the mesh size $h$.
\end{bems}
\end{comment}
To show convergence of the solution of the numerical scheme \eqref{num.reg.TVF}
we need to consider a discretization of the regularized problem (\ref{vis.TVF}).
Given $x_0\in L^2(\Omega,\mathcal{F}_0;\L)$, $g \in \L$
and $n \in \mathbb{N}$ we choose $x^n_0:=\mathcal{P}_n x_0\in\mathbb{V}_n$, $g^n:=\mathcal{P}_n g\in\mathbb{V}_n$ in (\ref{vis.TVF}).
Since $\mathbb{V}_n\subset \mathbb{H}^1_0$ the sequences $\{x^n_0\}_{n\in \mathbb{N}} \subset L^2(\Omega,\mathcal{F}_0;\mathbb{H}^1_0)$, $\{g_n\}_{n\in\mathbb{N}} \in \mathbb{H}^1_0$
constitute $\mathbb{H}^1_0$-approximating sequences of $x_0\in L^2(\Omega,\mathcal{F}_0;\L)$, $g \in \L$, respectively.
We set $x^{h,n}_0:=\mathcal{P}_h x_0^n$, $g^{h,n}:=\mathcal{P}_h g^n$, where $\mathcal{P}_h$ is the $\L$-projection onto $\mathbb{V}_h$.
The fully-discrete Galerkin approximation of (\ref{vis.TVF}) for fixed $n \in \mathbb{N} $ is then defined as follows:
fix $ N\in \mathbb{N}$, $h>0$ set $X_{\epsilon,\delta,n,h}^{0}=x^{h,n}_0$
and determine $\Yi\in \mathbb{V}_h$, $i=1,\dots, N$ as the solution of
\begin{align}\label{num.visc.TVF}
\ska{\Yi,v_h}&= \ska{\Ymin,v_h}-\tau \delta \ska{\nabla \Yi,\nabla v_h}-\tau \ska{\fe{\Yi},\nablav_h }
\nonumber \\
&-\tau\lambda\ska{\Yi -g^{h,n},v_h}+\ska{\Ymin,v_h}\Delta_i W &&\forall v_h \in \mathbb{V}_h .
\end{align}
{
The next lemma, cf. \cite[Lemma II.1.4]{book_temam} is used to show $\mathbb{P}$-a.s. existence of discrete solutions $\{\Xi\}_{i=1}^N$, $\{\Yi\}_{i=1}^N$ of numerical schemes \eqref{num.reg.TVF}, \eqref{num.visc.TVF}, respectively.
\begin{lems}\label{lemma_Existence_for_numerical_scheme}
Let $h: \mathbb{R}^L \rightarrow \mathbb{R}^L$ be continuous. If there is $R>0 $ such that $h(v)v\geq 0$ whenever $\no{v}_{\mathbb{R}^L}=R $ then there exist $\bar{v}$ satisfying $\no{\bar{v}}_{\mathbb{R}^L} \leq R$ and $h(\bar{v})=0$.
\end{lems}
In order to show $\{\mathcal{F}_{t_i}\}_{i=1}^{N}$-measurability of the random variables $\{\Xi\}_{i=1}^N$, $\{\Yi\}_{i=1}^N$
we make use of the following lemma, cf. \cite{gm_05,em_sis_18}.
\begin{lems}\label{lemma_measurability_for_numerical_scheme}
Let $(S,\Sigma)$ be a measure space. Let $f :S\times \mathbb{R}^L\rightarrow \mathbb{R}^L$ be a function that is $\Sigma$-measurable in its first argument for every $x \in \mathbb{R}^L$, that is continuous in its second argument for every $\alpha \in S$ and moreover such that for
every $\alpha \in S$ the equation $f(\alpha, x)=0$ has an unique solution $x=g(\alpha)$. Then $g : S \rightarrow \mathbb{R}^L$ is $\Sigma$-measurable.
\end{lems}
}
Below we show the existence, uniqueness and measurability of numerical solutions of (\ref{num.reg.TVF}), (\ref{num.visc.TVF}).
We state the result for the scheme (\ref{num.visc.TVF}) only,
since the proof also holds for $\delta=0$ (i.e. for (\ref{num.reg.TVF})) without any modifications.
\begin{lems}\label{Lemma_Existence_measurability_num.Scheme}
Let $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$, $g\in \L$
and let $L,n,N \in \mathbb{N}$ be fixed.
The for any $\delta\geq 0$, $\epsilon > 0$, $i=1,\ldots,N,$ there exist $\mathcal{F}_{t_i}$-measurable $\mathbb{P}$-a.s. unique
random variables $\Yi\in\mathbb{V}_h$ which solves (\ref{num.visc.TVF}).
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{Lemma_Existence_measurability_num.Scheme}}]
Assume that the $\mathbb{V}_h$-valued random variables $X_{\epsilon,\delta,n,h}^{0},\ldots,X_{\epsilon,\delta,n,h}^{i-1}$ satisfy \eqref{num.visc.TVF}
and that $X_{\epsilon,\delta,n,h}^{k}$ is $\mathcal{F}_{t_k}$-measurable for $k=1,\ldots,i-1$.
We show that there is a $\mathcal{F}_{t_i}$ measurable random variable $\Yi$, that satisfies \eqref{num.reg.TVF}.
Let $\{\phi_\ell\}_{\ell=1}^L$ be the basis of $\mathbb{V}_h$.
We identify every $v \in \mathbb{V}_h$ with a vector $\bar{v} \in \mathbb{R}^L$ with $v = \sum_{\ell=1}^L \bar{v}_\ell \phi _\ell$
and define a norm on $\mathbb{R}^L$ as $\no{\bar{v}}_{\mathbb{R}^L}:= \no{v}_{\mathbb{H}^1_0}$.
For an arbitrary $\omega \in \Omega$ we represent $X_\omega \in \mathbb{V}_h$ as a vector $\bar{X}_\omega \in \mathbb{R}^L$
and define a function $h: \Omega \times \mathbb{R}^L \rightarrow \mathbb{R}^L$ component-wise for $\ell=1,\ldots,L$ as
\begin{align*}
h(\omega,\bar{X}_\omega)_\ell:=(X_\omega-\Ymin(\omega),\phi_\ell)+\tau\delta (\nabla X_\omega, \nabla \phi_\ell)+\tau(\fe{X_\omega},\nabla \phi_\ell)\\
+\tau\lambda(X_\omega- g^{h,n},\phi_\ell) -(\Ymin(\omega),\phi_\ell)\Delta_i W(\omega).
\end{align*}
We show, that for each $\omega \in \Omega$ there exists an $\bar{X}_\omega$ such that $h(\omega,\bar{X}_\omega)=0$.
We note the following inequality
\begin{align*}
h(\omega,\bar{X}_\omega)\cdot \bar{X}_\omega=&(X_\omega-\Ymin(\omega),X)+\tau\delta\nos{\Delta X_\omega}+\tau(\fe{X_\omega},\nabla X_\omega)
\\
&+\tau\lambda(X_\omega- g^{h,n},X_\omega) -(\Ymin(\omega),X_\omega)\Delta_i W(\omega)
\\
\geq & \nos{X_\omega} -(\Ymin(\omega),X_\omega)+\tau(\fe{X_\omega},\nabla X_\omega)\\\
&-(\Ymin(\omega),X_\omega)\Delta_i W+\tau\lambda \no{X_\omega} -\tau\lambda ( g^n,X_\omega)\\
\geq & \no{ X_\omega}\left( \no{ X_\omega} -\no{ \Ymin(\omega)} -\no{\Ymin(\omega)}|\Delta_i W(\omega)|- \no{ g^{h,n}} \right)\,.
\end{align*}
On choosing $\no{X_\omega}=R_\omega$ large enough,
the existence of $X_{\epsilon,\delta,n}^{i}(\omega)\in \mathbb{V}_h$ for each $\omega\in\Omega$
then follows by Lemma \ref{lemma_Existence_for_numerical_scheme}, since $h(\omega,\cdot)$ is continuous
by the demicontinuity of the operator $A^{\epsilon,\delta}$, which follows from hemicontinuity and and monotonicty of $A^{\epsilon,\delta}$ for $\delta\geq 0$, $\epsilon >0$, see \cite[Remark 4.1.1]{Roeckner_book}.
The $\mathcal{F}_{t_i}$-measurabilty follows by Lemma~\ref{lemma_measurability_for_numerical_scheme} for unique $X_{\epsilon,\delta,n}^{i}$.
Hence, it remains to show that $X_{\epsilon,\delta,n,h}^{i}$ is $\mathbb{P}$-a.s. unique.
Assume there are two different solution $X_{1},X_{2}$, s.t. $h(\omega,\overline{X}_{1}(\omega))=0=h(\omega,\overline{X}_{2}(\omega))$ for $\omega\in\Omega$.
Then by the convexity (\ref{eps.convexity.inequality}) we observe that
\begin{align*}
0=&\big(h(\omega,\overline{X}_1(\omega))-h(\omega,\overline{X}_2(\omega))\big)\cdot\big(\overline{X}_1(\omega)-\overline{X}_2(\omega)\big)\\
=&(1+\tau \lambda)\nos{ X_1(\omega) - X_2(\omega)}+\tau \delta \nos{\nabla (X_1-X_2)(\omega)}\\
&+\tau\left(\fe{X_1(\omega)} -\fe{X_2(\omega)},\nabla X_1(\omega) - \nabla X_2(\omega)\right)\\
\geq& (1+\tau \lambda )\nos{X_1(\omega)-X_2(\omega)}+\tau \delta \nos{\nabla(X_1-X_2)(\omega)}\,.
\end{align*}
Hence ${X}_1\equiv {X}_2$ $\mathbb{P}$-a.s.
\end{proof}
We define the discrete Laplacian $\Delta_h : \mathbb{V}_h \rightarrow \mathbb{V}_h$ by
\begin{align}\label{disc_Lap}
-\ska{\Delta_h w_h,v_h}=\ska{\nabla w_h,\nabla v_h} ~~\forall w_h,v_h \in \mathbb{V}_h.
\end{align}
To obtain the required the stability properties of the numerical approximation \eqref{num.visc.TVF} we need the following lemma.
\begin{lems}\label{Lemma_Discret_Laplace}
Let $\Delta_h$ be the discrete Laplacian defined by \eqref{disc_Lap}. Then for any $v_h \in \mathbb{V}_h$, $\varepsilon,h>0$ the following inequality holds:
\begin{align}
-\ska{\fe{v_h},\nabla (\Delta_h v_h)} \geq 0\,.
\end{align}
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{Lemma_Discret_Laplace}}]
Let $\{\phi_\ell\}_{\ell=1}^L$ be the basis of $\mathbb{V}_h$ consisting of continuous piecewise linear Lagrange basis functions
associated with the nodes of the triangulation $\mathcal{T}_h$. Then any $v_h \in \mathbb{V}_h$ has the
representation $v_h=\sum_{\ell=1}^L (v_h)_{\ell} \phi_{\ell}$, where $(v_h)_{\ell} \in \mathbb{R}, \ell=1,\ldots,L $
and analogically $\Delta_hv_h=\sum_{\ell=1}^L (\Delta_h v_h)_\ell \phi_\ell$, with coefficients $(\Delta v_h)_\ell \in \mathbb{R}, \ell=1,\ldots,L$.
From (\ref{disc_Lap}) it follows that
\begin{align}\label{coeff}
\sum_{\ell=1}^L (\Delta v_h)_\ell \ska{\phi_\ell,\phi_k}
=-\sum_{\ell=1}^L (v_h)_{\ell} \ska{\nabla \phi_{\ell},\nabla \phi_k}=-\sum_{K\in \mathcal{T}_h}\sum_{\ell=1}^L (v_h)_{\ell} \ska{\nabla \phi_{\ell},\nabla \phi_k}_{K}\,,
\end{align}
where we denote $(v,w)_K := \int_{K}v(x)w(x)\mathrm{d}x$ for $K\in \mathcal{T}_h$.
We rewrite \eqref{coeff} with the mass matrix $M:=\{M\}_{i,k}:=(\phi_i,\phi_k)$ and the stiffness matrix $A:=\{A\}_{i,k}:=(\nabla \phi_i,\nabla \phi_k)$ and
$A_K := \{A_K\}_{i,k}:=(\nabla \phi_i,\nabla \phi_k)_{K}$ as
\begin{align}\label{coeff_Matrix}
\Delta_h \bar{v}_h=M^{-1}A \bar{v}_h=M^{-1}\sum_{K \in \mathcal{T}_h} A_K \bar{v}_h\, ,
\end{align}
where $\Delta_h\bar{v}_h\in \mathbb{R}^L$ is the vector $((\Delta_h v_h)_1,\ldots,(\Delta_h v_h)_L)^T$ and
$\bar{v}_h\in \mathbb{R}^L$ is the vector $((v_h)_1,\ldots,(v_h)_L)^T$.
Since $\mathbb{V}_h$ consists of functions, which are piecewise linear on the triangles $K \in \mathcal{T}_h$, $(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}$ is constant on every triangle $T$.
We note, that the matrices $M$ and $M^{-1}$ are positive definite.
We get using the Young's inequality
\begin{align*}
-&\ska{\fe{v_h},\nabla (\Delta_h v_h)}
=-\sum_{K\in \mathcal{T}_h}(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K\ska{\nabla v_h,\nabla (\Delta_h \nablav_h)}_K
\\
=&-\sum_{K\in \mathcal{T}_h}\sum_{k,\ell=1}^L(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K(v_h)_{\ell}(\Delta_h v_h)_k\ska{\nabla\phi_{\ell}, \nabla \phi_k }_K
\\
=&-\sum_{K\in \mathcal{T}_h}(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K \bar{v}_h^T A^T_K\Delta_h\bar{v}_h
\\
=&\sum_{K,K' \in \mathcal{T}_h}(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K \bar{v}_h^TA^T_K M^{-1} A_{K'}\bar{v}_h
\\
=&\frac{1}{2}\sum_{K,K'\in \mathcal{T}_h}(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K \bar{v}_h^TA^T_K M^{-1}A_{K'}\bar{v}_h\\
&+\frac{1}{2}\sum_{K,K'\in \mathcal{T}_h}(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_{K'} \bar{v}_h^TA^T_K M^{-1}A_{K'}\bar{v}_h
\\
=&\frac{1}{2}\sum_{K,K'\in \mathcal{T}_h} \bar{v}_h^T A^T_K M^{-1}A_{K'}\bar{v}_h\left((\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K+(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_{K'} \right)\\
\\
\geq&\frac{1}{2}\sum_{K,K'\in \mathcal{T}_h} \sqrt{(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_K}\bar{v}_h^T A^T_K M^{-1}A_{K'}\bar{v}_h\sqrt{(\bes{\nabla v_h}+\epsilon^2)^{-\frac{1}{2}}_{K'}} \\
\geq& 0\, ,
\end{align*}
since $M^{-1}$ is positive definite.
\end{proof}
In the next lemma we state the stability properties of the numerical solution of the scheme (\ref{num.visc.TVF})
which are discrete analogues of estimates in Lemma~\ref{lem_visc_exist}~and Lemma~\ref{laplace_energy_estimate}.
\begin{lems}\label{Lemma_Discrete a priori estimates}
Let $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$ and $g\in\L$ be given.
Then there exists a constant $C \equiv C(\mathbb{E}{\|x_0\|_{\L}}, \|g\|_{\L}) > 0$ such that for any $n \in \mathbb{N}$, $\tau,h>0$
the solution of scheme (\ref{num.visc.TVF}) satisfies
\begin{align}\label{discrete_energy_estimate_viscTVF}
\sup_{i=1,\ldots,N}\mathbb{E}{\nos{\Yi}}&+\frac{1}{4}\mathbb{E}{\sum_{k=1}^N\nos{X_{\epsilon,\delta,n,h}^{k}-X_{\epsilon,\delta,n,h}^{k-1}}} \nonumber \\+& \tau \delta \mathbb{E}{\sum_{k=1}^N\nos{\nabla X_{\epsilon,\delta,n,h}^{k}}}+\frac{\tau \lambda}{2}\mathbb{E}{\sum_{k=1}^N \nos{X_{\epsilon,\delta,n,h}^{k}}}\leq C\,,
\end{align}
and a constant $C_{n} \equiv C( \mathbb{E}[\|x_0^n\|_{\mathbb{H}^1_0}], \|g^n\|_{\mathbb{H}^1_0}) > 0$ such that for any $\tau,h>0$
\begin{align}\label{discrete_H1_estimate_viscTVF}
\sup_{i=1,\ldots,N}\mathbb{E}{\nos{\nabla \Yi}}+\frac{1}{4}\mathbb{E}{\sum_{k=1}^N\nos{\nabla(X_{\epsilon,\delta,n,h}^{k}-X_{\epsilon,\delta,n,h}^{k-1})}}+\tau \delta \mathbb{E}{\sum_{k=1}^N\nos{\Delta_h X_{\epsilon,\delta,n,h}^{k}}}\leq C_{n}.
\end{align}
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{Lemma_Discrete a priori estimates}}]
We set $v_h=\Yi$ in \eqref{num.visc.TVF}, use the identity
$2(a-b)a =a^2 - b^2 + (a-b)^2$
and get for $i=1,\ldots,N$
\begin{align}\label{num.visc.energie.estimate}
\frac{1}{2}\nos{\Yi} +\frac{1}{2}\nos{\Yi-\Ymin}+\tau \delta \nos{\nabla \Yi}+\tau \ska{\fe{\Yi},\nabla \Yi} \nonumber\\
=\frac{1}{2}\nos{\Ymin}-\tau \lambda\left(\nos{\Yi}-\ska{ g^{h,n},\Yi}\right) +\ska{\Ymin,\Yi}\Delta_i W.
\end{align}
We take expected value in (\ref{num.visc.energie.estimate})
and on noting the properties of
Wiener increments $\mathbb{E}{\Delta_i W}=0$, $\mathbb{E}{ \bes{\Delta_i W}}=\tau$ and the independence of $\Delta_i W$ and $\Ymin$
we estimate the stochastic term as
\begin{align*}
\mathbb{E}{\ska{\Ymin,\Yi}\Delta_i W}&=\mathbb{E}{\ska{\Ymin,\Yi-\Ymin}\Delta_i W}+\mathbb{E}{\ska{\Ymin,\Ymin}\Delta_i W}
\\
\leq& \mathbb{E}{\frac{1}{4}\nos{\Ymin-\Yi}+\nos{\Ymin}\bes{\Delta_i W}} + \mathbb{E}{\|\Ymin\|^2}\mathbb{E}{\Delta_i W}
\\
=&\frac{1}{4}\mathbb{E}{\nos{\Yi-\Ymin}}+\tau\mathbb{E}{\nos{\Ymin}}.
\end{align*}
We neglect the positive term
\begin{align*}
\ska{\fe{\Yi},\nabla \Yi} \geq 0\, ,
\end{align*}
and get from (\ref{num.visc.energie.estimate}) that
\begin{align*
\frac{1}{2}\mathbb{E}{\nos{\Yi}} +&\frac{1}{4}\mathbb{E}{\nos{\Yi-\Ymin}}+\tau \delta \mathbb{E}{\nos{\nabla \Yi}}+\frac{\tau\lambda}{2}\mathbb{E}{\nos{\Yi}}\\
\leq& \frac{1}{2}\mathbb{E}{\nos{\Ymin}}+\tau\mathbb{E}{\nos{\Ymin}}+\tau\lambda{\nos{g^{h,n}}}\,. \nonumber
\end{align*}
We sum up the above inequality for $k=1,\ldots,i$ and obtain
\begin{align}\label{num.visc.energie.estimate3}
\frac{1}{2}\mathbb{E}{\nos{\Yi}} +&\frac{1}{4}\mathbb{E}{\sum_{k=1}^i\nos{X_{\epsilon,\delta,n,h}^{k}-X_{\epsilon,\delta,n,h}^{k-1}}}+\tau \delta \mathbb{E}{\sum_{k=1}^i\nos{\nabla X_{\epsilon,\delta,n,h}^{k}}}+\frac{\tau \lambda}{2}\mathbb{E}{\sum_{k=1}^i \nos{ X_{\epsilon,\delta,n,h}^{k}}}
\nonumber \\
\leq& \frac{1}{2}\mathbb{E}{\nos{x^n_0}}+\tau\mathbb{E}{\sum_{k=1}^{i}\nos{X_{\epsilon,\delta,n,h}^{k-1}}} +T\lambda{\nos{ g^{h,n}}}\,.
\end{align}
By the discrete Gronwall lemma it follows from (\ref{num.visc.energie.estimate3}) that
\begin{align*}
\sup\limits_{i=1,\ldots,N}\mathbb{E}{\nos{ \Yi}} \leq \exp(2T)\Big(\mathbb{E}{\nos{x_0}}+2T\lambda{\nos{g}}\Big).
\end{align*}
We substitute the above estimate into the right-hand side of \eqref{num.visc.energie.estimate3}
to conclude (\ref{discrete_energy_estimate_viscTVF}).
To show the estimate
(\ref{discrete_H1_estimate_viscTVF}) we set $v_h = \Delta_h \Yi$ in (\ref{num.visc.TVF}) use integration by parts
and proceed analogically to the first part of the proof.
We note that by Lemma~\ref{Lemma_Discret_Laplace} it holds that
\begin{align}\label{discrete_resolvent_estimate}
\ska{\fe{\Yi},\nabla \Delta_h \Yi} \geq 0.
\end{align}
Hence we may neglect the positive term and get that
\begin{align*}
\frac{1}{2}&\mathbb{E}{\nos{\nabla \Yi}} +\frac{1}{4}\mathbb{E}{\sum_{k=1}^i\nos{\nabla(X_{\epsilon,\delta,n,h}^{k}-X_{\epsilon,\delta,n,h}^{k-1})}}+\tau \delta \mathbb{E}{\sum_{k=1}^i\nos{\Delta_h X_{\epsilon,\delta,n,h}^{k}}}\nonumber\\
&+\frac{\tau \lambda}{2}\mathbb{E}{\sum_{k=1}^i \nos{ \nabla X_{\epsilon,\delta,n,h}^{k}}}
\leq \frac{1}{2}\mathbb{E}{\nos{\nabla x^n_0}}+\tau\mathbb{E}{\sum_{k=1}^{i}\nos{\nabla X_{\epsilon,\delta,n,h}^{k-1}}} +T\lambda{\nos{\nabla g^n}}\,.
\end{align*}
and obtain (\ref{discrete_H1_estimate_viscTVF}) after an application of the discrete Gronwall lemma.
\end{proof}
We define piecewise constant time-interpolants of the numerical solution
$\{\Yi\}_{i=0}^{N}$ of (\ref{num.visc.TVF}) for $t\in [0,T]$ as
\begin{align}\label{eps_delta_interpol1}
\overline{X}_{\tau,h}^{\epsilon,\delta,n}(t):= \Yi \quad \mathrm{if}\quad t\in (t_{i-1},t_i]
\end{align}
and
\begin{align}\label{eps_delta_interpol2}
\overline{X}_{\tau_-,h}^{\epsilon,\delta,n}(t):= \Ymin\quad \mathrm{if}\quad t\in [t_{i-1},t_i)\,.
\end{align}
We note that (\ref{num.visc.TVF})
can be reformulated as
\begin{align}\label{Integralformulation}
&\ska{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(t),v_h}+\left\langle\int_0^{\theta_{+}(t)} A^{\epsilon,\delta}\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s) \d s,v_h\right\rangle \nonumber \\
&=\ska{X_{\epsilon,\delta,n}^0,v_h}+\ska{\int_0^{\theta_+(t)} \overline{X}_{\tau_-,h}^{\epsilon,\delta,n}(s) \d W(s),v_h} \qquad \mathrm{for}\,\, t\in [0,T],
\end{align}
where $\theta_+(0):=0$ and $\theta_+(t):=t_i$ if $t\in (t_{i-1},t_{i}]$.
Estimate \eqref{discrete_energy_estimate_viscTVF}
yields the bounds
\begin{align}\label{Constant interpolation estimate}
\sup\limits_{t\in [0,T]}\mathbb{E}{\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(t)}} &\leq C, &\sup\limits_{t\in [0,T]}\mathbb{E}{\nos{\overline{X}_{\tau_-,h}^{\epsilon,\delta,n}(t)}} \leq C,\\
~ \delta \mathbb{E}{\int_0^T \nos{\nabla\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)}\d s} &\leq C .\nonumber
\end{align}
Furthermore, \eqref{Constant interpolation estimate} and \eqref{a_bnd} imply
\begin{align}\label{Aed Abschaetzung}
\mathbb{E}{\int_0^T \nos{A^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)}_{\mathbb{H}^{-1}} \d s} \leq C.
\end{align}
The estimates in (\ref{Constant interpolation estimate}) imply for fixed $n \in \mathbb{N}$, $\epsilon,\,\delta>0$
the existence of a subsequence, still denoted by $\{\overline{X}_{\tau,h}^{\epsilon,\delta,n}\}_{\tau,h>0}$,
and a $Y \in L^2(\Omega\times (0,T);\L)\cap L^2(\Omega\times (0,T);\mathbb{H}^1_0)\cap L^{\infty}((0,T);L^2(\Omega;\L)$, s.t., for $\tau,h \rightarrow 0$
\begin{align}\label{limit_process}
\overline{X}_{\tau,h}^{\epsilon,\delta,n} &\rightharpoonup Y ~\text{in}~ L^2(\Omega\times (0,T);\L), \nonumber\\
\overline{X}_{\tau,h}^{\epsilon,\delta,n} &\rightharpoonup Y ~\text{in}~ L^2(\Omega\times (0,T);\mathbb{H}^1_0),\\
\overline{X}_{\tau,h}^{\epsilon,\delta,n} &\rightharpoonup^* Y ~\text{in}~ L^{\infty}((0,T);L^2(\Omega;\L)) \nonumber.
\end{align}
In addition, there exists a $\nu \in L^2(\Omega;\L)$ such that $\overline{X}_{\tau,h}^{\epsilon,\delta,n}(T) \rightharpoonup \nu$
in $L^2(\Omega;\L)$ as $\tau,h \rightarrow 0$ and the estimate (\ref{Aed Abschaetzung})
implies the existence of a $a^{\epsilon,\delta} \in L^2(\Omega\times (0,T);\mathbb{H}^{-1})$, s.t.,
\begin{align}\label{lim_a}
A^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n} &\rightharpoonup a^{\epsilon,\delta} ~\text{in}~ L^2(\Omega\times (0,T);\mathbb{H}^{-1})\quad \mathrm{for} \,\,\tau,h \rightarrow 0.
\end{align}
The estimates in (\ref{Constant interpolation estimate}) also implies for fixed $n \in \mathbb{N}$, $\epsilon,\, \delta > 0$
the existence of a subsequence, still denoted by $\{\overline{X}_{\tau_-,h}^{\epsilon,\delta,n}\}_{\tau>0}$,
and a $Y^- \in L^2(\Omega\times (0,T);\L)$, s.t.,
\begin{align}\label{limit_process_minus}
\overline{X}_{\tau_-,h}^{\epsilon,\delta,n} &\rightharpoonup Y^- ~\text{in}~ L^2(\Omega\times (0,T);\L)\quad \mathrm{for} \,\,\tau,h \rightarrow 0.
\end{align}
Finally, the inequality \eqref{num.visc.energie.estimate3} implies
\begin{align}\label{same_weak}
\lim\limits_{\tau \rightarrow 0}\mathbb{E}{\int_0^T \nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-\overline{X}_{\tau_-,h}^{\epsilon,\delta,n}(s)} \d s }=&\lim\limits_{\tau \rightarrow 0}\tau \mathbb{E}{\sum_{k=1}^N\nos{X_{\epsilon,\delta,n}^k-X_{\epsilon,\delta,n}^{k-1}}} \nonumber\\
\leq& \lim\limits_{\tau \rightarrow 0} C\tau =0\,.
\end{align}
which shows that the weak limits of $Y$ and $Y^-$ coincide.
The following result shows that the limit $Y\equiv X^{\epsilon,\delta}_{n}$, i.e.,
that the numerical solution of scheme (\ref{num.visc.TVF}) converges to the unique variational solution of (\ref{vis.TVF}) for $\tau,h \rightarrow 0$.
Owing to the properties (\ref{Monotonicity}), (\ref{a_bnd})
the convergence proof follows standard arguments for the convergence of numerical approximations of monotone equations, see for instance \cite{gm_05}, \cite{em_sis_18},
and is therefore omitted.
We note that the convergence of the whole sequence $\{\overline{X}_{\tau,h}^{\epsilon,\delta,n}\}_{\tau,h>0}$ follows by the uniqueness of the variational solution.
\begin{lems}\label{lemma_Limiten_Gleichung}
Let $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$ and $g\in\L$ be given, let $\epsilon, \delta, \lambda>0$, {$n \in \mathbb{N}$} be fixed.
Further, let $X^{\epsilon,\delta}_{n}$ be the unique variational solution of \eqref{vis.TVF}
for $ x^n_0=\mathcal{P}_nx_0$, $g^n = \mathcal{P}_n g$
and $\overline{X}_{\tau,h}^{\epsilon,\delta,n}$, $\overline{X}_{\tau_-,h}^{\epsilon,\delta,n}$ be the respective time-interpolant (\ref{eps_delta_interpol1}), (\ref{eps_delta_interpol2}) of the numerical solution $\{\Yi\}_{i=1}^N$ of \eqref{num.visc.TVF}.
Then $\overline{X}_{\tau,h}^{\epsilon,\delta,n}$, $\overline{X}_{\tau_-,h}^{\epsilon,\delta,n}$ converge to $X^{\epsilon,\delta}_{n}$ for $\tau,h \rightarrow 0$ in the sense that
the weak limits from (\ref{limit_process}), (\ref{lim_a}) satisfy $Y\equiv X^{\epsilon,\delta}_{n}$, $a^{\epsilon,\delta}\equiv A^{\epsilon,\delta} Y \equiv A^{\epsilon,\delta} X^{\epsilon,\delta}_{n}$ and $\nu=Y(T)\equiv X^{\epsilon,\delta}_{n}(T)$.
In addition it holds for almost all $(\omega,t) \in \Omega\times (0,T)$ that
\begin{align*
Y(t)=Y(0){-}\int_0^t A^{\epsilon,\delta} Y(s) \d s+\int_0^t Y(s)\d W(s),
\end{align*}
and there is an $\L$-valued continuous modification of $Y$ (denoted again as $Y$) such that for all $t \in [0,T]$
\begin{align}\label{Ito-Formule_fuer_Limiten}
\frac{1}{2}\nos{Y(t)}= & \frac{1}{2}\nos{Y(0)}{-}\int_0^t \langle A^{\epsilon,\delta} Y(s),Y(s) \rangle +\frac{1}{2}\nos{Y(s)} \d s
\\ \nonumber
& +\int_0^t (Y(s),Y(s))\d W(s).
\end{align}
\end{lems}
{The strong monotonicity property \eqref{Monotonicity} of the operator $A^{\epsilon,\delta}$ implies strong convergence of the numerical
approximation in $L^2(\Omega\times(0,T);\L)$.}
\begin{lems}\label{Lemma_Convergence_num.vis.Scheme}
Let $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$ and $g\in\L$ be given,
let $\epsilon, \delta, \lambda>0$, {$n \in \mathbb{N}$} be fixed. Further, let $X^{\epsilon,\delta}_{n}$ be the variational solution of \eqref{vis.TVF}
for $ x^n_0=\mathcal{P}_nx_0$, $g^n = \mathcal{P}_n g$
and $\overline{X}_{\tau,h}^{\epsilon,\delta,n}$ be the time-interpolant (\ref{eps_delta_interpol1}) of the numerical solution $\{\Yi\}_{i=1}^N$ of \eqref{num.visc.TVF}.
Then the following convergence holds true
\begin{align}
\lim\limits_{\tau,h \rightarrow 0}\nos{X^{\epsilon,\delta}_n-\overline{X}_{\tau,h}^{\epsilon,\delta,n}}_{L^2(\Omega \times (0,T);\L)}\rightarrow 0.
\end{align}
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{Lemma_Convergence_num.vis.Scheme}}]
The proof follows along the lines of \cite{gm_05}, \cite{em_sis_18}.
We sketch the main steps of the proof for the convenience of the reader.
We note that $\overline{X}_{\tau,h}^{\epsilon,\delta,n}$ satisfies (cf. proof of Lemma~\ref{Lemma_Discrete a priori estimates})
\begin{align}\label{psi estimate}
e^{-\kappa T} \mathbb{E}{\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(T)}} \leq& \mathbb{E}{\nos{x_0^n}} -\kappa\int_0^T e^{-\kappa s} \mathbb{E}{\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)}}\d s \nonumber\\
-& 2\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)\rangle \d s} \\
+&\mathbb{E}{\int_0^T e^{-\kappa s} \nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)} \d s}
+ \kappa \int_0^T e^{-\kappa s}\be{R_\tau(s)}\d s \nonumber,
\end{align}
where $\displaystyle R_{\tau}(t):= \mathbb{E}{\int_t^{\theta_+(t)}2\langleA^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)\rangle-\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)}\d s }$.
We reformulate the third term on the right-hand side in \eqref{psi estimate} as
\begin{align*}
&\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)\rangle \d s}\\
&= \mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-A^{\epsilon,\delta} X^{\epsilon,\delta}_n(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon,\delta}_n(s)\rangle \d s}\\
&+\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} X^{\epsilon,\delta}_n(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon,\delta}_n(s)\rangle+\langleA^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s),X^{\epsilon,\delta}_n(s)\rangle \d s}.
\end{align*}
We substitute the equality above into \eqref{psi estimate} and obtain for $\kappa \geq 1$ that
\begin{align*}
& e^{-\kappa T} \mathbb{E}{\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(T)}} +2\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-A^{\epsilon,\delta} X^{\epsilon,\delta}_n(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon}(s)\rangle \d s}
\\
& \leq \mathbb{E}{\nos{x_0^n}}
- 2\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} X^{\epsilon,\delta}_n(s),\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon,\delta}_n(s)\rangle+\langleA^{\epsilon,\delta} \overline{X}_{\tau,h}^{\epsilon,\delta,n}(s),X^{\epsilon,\delta}_n(s)\rangle \d s}
\\
&\qquad + \kappa\int_0^T e^{-\kappa s}\be{R_\tau(s)}\d s.
\end{align*}
We observe that $\displaystyle \int_0^T e^{-\kappa s}\be{R_\tau(s)}\d s \rightarrow 0$ for $\tau$.
Hence, by the lower-semicontinuity of norms
using the convergence properties from Lemma~\ref{lemma_Limiten_Gleichung}
and the monotonicity property \eqref{Monotonicity}
we get for $\tau,h \rightarrow 0$ that
\begin{align}\label{inequality 3}
& e^{-\kappa T} \mathbb{E}{\nos{X^{\epsilon,\delta}_n (T)}} +2\lambda\lim\limits_{\tau,h\rightarrow 0} \mathbb{E}{\int_0^T e^{-\kappa s} \nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon,\delta}_n(s)} \d s} \nonumber
\\
& \qquad + 2\delta\lim\limits_{\tau,h\rightarrow 0} \mathbb{E}{\int_0^T e^{-\kappa s} \nos{\nabla\big(\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon,\delta}_n(s)\big)} \d s}
\\ \nonumber
& \leq \mathbb{E}{\nos{x_0^n}} - 2\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta} X^{\epsilon,\delta}_n(s),X^{\epsilon,\delta}_n(s)\rangle \d s} .
\end{align}
It is not difficult to see that \eqref{Ito-Formule_fuer_Limiten} for $Y\equiv X^{\epsilon,\delta}_n$ implies
\begin{align} \label{Equalitiy}
& e^{-\kappa T} \mathbb{E}{\nos{X^{\epsilon,\delta}_n(T)}} = \mathbb{E}{\nos{x_0^n}}
- 2\mathbb{E}{\int_0^T e^{-\kappa s} \langle A^{\epsilon,\delta}(s)X^{\epsilon,\delta}_n,X^{\epsilon,\delta}_n(s)\rangle \d s }\\
& - \kappa\int_0^T e^{-\kappa s} \mathbb{E}{\nos{X^{\epsilon,\delta}_n(s)}}\d s +\mathbb{E}{\int_0^T e^{-\kappa s}\nos{X^{\epsilon,\delta}_n(s)} \d s} \nonumber.
\end{align}
We subtract the equality \eqref{Equalitiy} from \eqref{inequality 3} and obtain for $\kappa \geq 1$
\begin{align*}
\lambda \lim\limits_{\tau,h\rightarrow \infty} \mathbb{E}{\int_0^T e^{-\kappa s} \nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}(s)-X^{\epsilon,\delta}_n(s)} \d s} \leq 0.
\end{align*}
Hence, we conclude that $ \overline{X}_{\tau,h}^{\epsilon,\delta,n} \rightarrow X^{\epsilon,\delta}_{n} $ in $L^2(\Omega;L^2((0,T);\L)$.
\end{proof}
{\begin{bems}
It is obvious from the proof of Lemma~\ref{Lemma_Convergence_num.vis.Scheme}
that the strong convergence in $L^2(\Omega\times(0,T);\L)$
remains valid for $\lambda=0$ due to (\ref{Monotonicity}) by the Poincar\'e inequality.
\end{bems}}
Next lemma guarantees the convergence of the numerical solution of scheme \eqref{num.visc.TVF}
to the numerical solution of scheme (\ref{num.reg.TVF}) for $\delta \rightarrow 0$.
\begin{lems}\label{Lemma_Difference-num.Schemes}
{Let $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$ and $g\in\L$ be given.
Then for each $n \in \mathbb{N}$ there exists a constant $C\equiv C(T)>0$, $C_n\equiv C(\mathbb{E}[\|x_0^n\|_{\mathbb{H}^1_0}], \|g^n\|_{\mathbb{H}^1_0})>0$ }
such that for any $N\in\mathbb{N}$, $\delta>0$, $n\in \mathbb{N}$, $h,\epsilon \in (0,1]$ the following estimate holds for the difference of numerical solutions of (\ref{num.reg.TVF}) and (\ref{num.visc.TVF}):
\begin{align*}
\max\limits_{i=1,\ldots,N}\mathbb{E}{\nos{\Xi-\Yi}} \leq C(C_{n}\delta+ \mathbb{E}{\nos{x_0^h-x_0^{h,n}}}+ \lambda\nos{g^h-g^{h,n}}).
\end{align*}
\end{lems}
We note that the $n$-dependent constant $C_n$ in the estimate above is due to the a priori estimate
(\ref{discrete_H1_estimate_viscTVF}), for $\mathbb{H}^1_0$-regular data $x_0$, $g$ it holds that $C_n\equiv C(\mathbb{E}[\|x_0\|_{\mathbb{H}^1_0}], \|g\|_{\mathbb{H}^1_0})$
by the stability of the discrete $\L$-projection $\mathcal{P}_h:\mathbb{H}^1_0\rightarrow \mathbb{V}_h$ in $\mathbb{H}^1_0$.
\begin{proof}[\textbf{Proof of Lemma \ref{Lemma_Difference-num.Schemes}}]
We define $Z^{i}_{\epsilon,h}:=\Xi -\Yi$. From \eqref{num.reg.TVF} and \eqref{num.visc.TVF} we get
\begin{align*}
\ska{Z^{i}_{\epsilon,h},v_h}=& \ska{Z^{i-1}_{\epsilon,h}.v_h}-\tau \delta \ska{\Delta_h \Yi,v_h}\\&-\tau \ska{\fe{\Xi},v_h }-\tau \ska{\fe{\Yi},\nablav_h }\\&-\tau\lambda\ska{Z^{i}_{\epsilon,h},v_h}-{\tau\lambda\ska{g^h - g^{h,n},v_h}}
\\
&+\ska{Z^{i-1}_{\epsilon,h},v_h}\Delta_i W.
\end{align*}
We set $v_h=Z^{i}_{\epsilon,h}$ and obtain
\begin{align*}
\ska{Z^{i}_{\epsilon,h}-Z^{i-1}_{\epsilon,h},Z^{i}_{\epsilon,h}} =& -\tau \delta \ska{\Delta_h \Yi,Z^{i}_{\epsilon,h}}\\&-\tau \ska{\fe{\Xi}- \fe{\Yi},\nablaZ^{i}_{\epsilon,h} }\\&-\tau\lambda\nos{Z^{i}_{\epsilon,h}}-\tau\lambda\ska{g^h- g^{h,n},Z^{i}_{\epsilon,h}}
\\
&+\ska{Z^{i-1}_{\epsilon,h},Z^{i}_{\epsilon,h}}\Delta_i W.
\end{align*}
We note that
\begin{align*}
\ska{Z^{i}_{\epsilon,h}-Z^{i-1}_{\epsilon,h},Z^{i-1}_{\epsilon,h}}= \frac{1}{2}\nos{Z^{i}_{\epsilon,h}}-\frac{1}{2}\nos{Z^{i-1}_{\epsilon,h}}+\frac{1}{2}\nos{Z^{i}_{\epsilon,h} -Z^{i-1}_{\epsilon,h}}\,,
\end{align*}
and by the Cauchy-Schwarz and Young's inequalities
\begin{align*}
&\tau \delta \ska{\Delta_h \Yi,Z^{i}_{\epsilon,h}}\leq \frac{\tau \delta^2}{2\lambda } \nos{\Delta_h \Yi} +\frac{\tau \lambda}{2}\nos{Z^{i}_{\epsilon,h}},\\
&\tau \lambda \ska{g^h-g^{h,n} ,Z^{i}_{\epsilon,h}}\leq \frac{\tau \lambda}{2} \nos{g^h-g^{h,n}} +\frac{\tau \lambda}{2}\nos{Z^{i}_{\epsilon,h}}.
\end{align*}
From the convexity (\ref{eps.convexity.inequality}) it follows that
\begin{align*}
-\tau \ska{\fe{\Xi}-\fe{\Yi},\nabla(\Xi-\Yi)}\leq 0.
\end{align*}
Hence, we obtain that
\begin{align}\label{zest1}
&\frac{1}{2}\nos{Z^{i}_{\epsilon,h}}+\frac{1}{2}\nos{Z^{i}_{\epsilon,h} -Z^{i-1}_{\epsilon,h}}\\
&\leq \frac{1}{2}\nos{Z^{i-1}_{\epsilon,h}}+\frac{\tau \delta^2}{2\lambda } \nos{\Delta_h \Yi}+ \frac{\tau \lambda}{2} \nos{g^h-g^{h,n}} +\ska{Z^{i-1}_{\epsilon,h},Z^{i}_{\epsilon,h}}\Delta_i W\,. \nonumber
\end{align}
We estimate the last term on the right-hand side above as
\begin{align*}
\ska{Z^{i-1}_{\epsilon,h},Z^{i}_{\epsilon,h}}\Delta_i W=\ska{Z^{i-1}_{\epsilon,h},Z^{i}_{\epsilon,h}-Z^{i-1}_{\epsilon,h}}\Delta_i W+\nos{Z^{i-1}_{\epsilon,h}}\Delta_i W\\
\leq \frac{1}{2}\nos{Z^{i}_{\epsilon,h}-Z^{i-1}_{\epsilon,h}}+\frac{1}{2}\nos{Z^{i-1}_{\epsilon,h}}\bes{\Delta_i W} +\nos{Z^{i-1}_{\epsilon,h}}\Delta_i W,
\end{align*}
and substitute the above identity into (\ref{zest1})
\begin{align*}
\frac{1}{2}\nos{Z^{i}_{\epsilon,h}}+\frac{1}{2}\nos{Z^{i}_{\epsilon,h} -Z^{i-1}_{\epsilon,h}}\leq& \frac{1}{2}\nos{Z^{i-1}_{\epsilon,h}}+\frac{\tau \delta^2}{2\lambda } \nos{\Delta_h \Yi}+ \frac{\tau \lambda}{2} \nos{g^h-g^{h,n}} +\frac{1}{2}\nos{Z^{i}_{\epsilon,h}-Z^{i-1}_{\epsilon,h}}\\
&+\frac{1}{2}\nos{Z^{i-1}_{\epsilon,h}}\bes{\Delta_i W}+\nos{Z^{i-1}_{\epsilon,h}}\Delta_i W\,.
\end{align*}
Next, we sum up the above inequality up to $i\leq N$ and obtain
\begin{align*}
\frac{1}{2}\nos{Z^{i}_{\epsilon,h}} \leq& \frac{1}{2}\nos{Z^0_{\epsilon,h}}+\frac{\tau \delta^2}{2\lambda } \sum\limits_{k=1}^{i} \nos{\Delta_h X_{\epsilon,\delta,h}^k}
+\frac{1}{2}\sum\limits_{k=1}^{i}\nos{Z^{k-1}_{\epsilon,h}}\bes{\Delta_k W}+\sum\limits_{k=1}^{i}\nos{Z^{k-1}_{\epsilon,h}}\Delta_k W\\&+ \frac{T \lambda}{2} \nos{g^h- g^{h,n}}.
\end{align*}
After taking expectation in the above and using the independence properties of Wiener increments
and the estimate (\ref{discrete_H1_estimate_viscTVF}) we arrive at
\begin{align*}
\frac{1}{2}\mathbb{E}{\nos{Z^{i}_{\epsilon,h}}} \leq &\frac{1}{2}\nos{Z^0_{\epsilon,h}}+\frac{\tau \delta^2}{2\lambda } \mathbb{E}{\sum\limits_{k=1}^{i} \nos{\Delta_h X_{\epsilon,\delta,n,h}^k}}
+\frac{\tau}{2}\sum\limits_{k=1}^{i}\mathbb{E}{\nos{Z^{k-1}_{\epsilon,h}}}\\
\leq&C_n \delta+\frac{1}{2}\mathbb{E}{\nos{Z^0_{\epsilon,h}}} +\frac{\tau}{2}\sum\limits_{k=0}^{i-1}\mathbb{E}{\nos{Z^k_{\epsilon,h}}}+ \frac{T \lambda}{2} \nos{g^h-g^{h,n}}.
\end{align*}
with $C_n \equiv C(\|x_0^n\|_{\mathbb{H}^1_0}, \|g^n\|_{\mathbb{H}^1_0})$.
Finally, the Discrete Gronwall lemma yields for $i=1,\ldots,N$ that
\begin{align}\label{m,n gronwall}
\mathbb{E}{\nos{Z^{i}_{\epsilon,h}}} \leq \exp(T)(C_{n}\delta+ \frac{1}{2}\mathbb{E}{\nos{x_0^h- x_0^{h,n}}}+ \frac{T \lambda}{2}\nos{g^h-g^{h,n}}).
\end{align}
which concludes the proof .
\end{proof}
We define piecewise constant time-interpolant
of the discrete solution $\{\Xi \}_{i=0}^N$ of (\ref{num.reg.TVF}) for $t\in[0,T)$ as
\begin{align}\label{eps_interpol}
\overline{X}_{\tau,h}^{\epsilon}(t) = \Xi\quad \mathrm{if}\quad t \in (t_{i-1},t_i].
\end{align}
We are now ready to state the second main result of this paper
which is the convergence of the numerical approximation (\ref{num.reg.TVF}) to the unique SVI solution of the total variation flow (\ref{TVF})
(cf. Definition~\ref{def_varsoleps}).
\begin{thms}\label{Thm_Convergence_num.reg.Scheme}
Let $X$ be the SVI solution of \eqref{TVF} and let $\overline{X}_{\tau,h}^{\epsilon}$ be the time-interpolant (\ref{eps_interpol})
of the numerical solution of the scheme \eqref{num.reg.TVF}.
Then the following convergence holds true
\begin{align}\label{numconv}
\lim\limits_{\epsilon\rightarrow 0}\lim\limits_{\tau,h \rightarrow 0}\nos{X-\overline{X}_{\tau,h}^{\epsilon}}_{L^2(\Omega \times (0,T);\L)}\rightarrow 0.
\end{align}
\end{thms}
\begin{proof}[\textbf{Proof of Theorem \ref{Thm_Convergence_num.reg.Scheme}}]
For $x_0 \in L^2(\Omega,\mathcal{F}_0;\L)$ and $g\in\L$
we define the $\mathbb{H}^1_0$-approximating sequences $\{x_0^n\}_{n\in \mathbb{N}}\subset \mathbb{H}^1_0$, $x_0^n\rightarrow x_0\in L^2(\Omega,\mathcal{F}_0;\L)$,
$\{g^n\}_{n\in \mathbb{N}}\subset \mathbb{H}^1_0$, $n\in\mathbb{N}$, $g^n\rightarrow g\in\L$
via the $\L$-projection onto $\mathbb{V}_n\subset\mathbb{H}^1_0$.
We consider the solutions $X^{\epsilon},X^{\epsilon,\delta}_{n}$ of \eqref{reg.TVF}, \eqref{vis.TVF}, respectively,
and denote by $X^{\epsilon}_n$ the SVI solution of \eqref{reg.TVF} for $x_0\equiv x_0^n$, $g\equiv g^n$.
Furthermore, we recall that the interpolant $\overline{X}_{\tau,h}^{\epsilon,\delta,n}$ of the numerical solution of (\ref{num.visc.TVF}) was defined in (\ref{eps_delta_interpol1}).
We split the numerical error as
{\begin{align}\label{err_ineq}
\nonumber
\frac{1}{5}\nos{X-\overline{X}_{\tau,h}^{\epsilon}}_{L^2(\Omega \times (0,T);\L)} \leq& \nos{X-X^{\epsilon}}_{L^2(\Omega \times (0,T);\L)} +\nos{X^{\epsilon}-X^{\epsilon}_n}_{L^2(\Omega \times (0,T);\L)}
\\
&+ \nos{X^{\epsilon}_n-X^{\epsilon,\delta}_{n}}_{L^2(\Omega \times (0,T);\L)}+\nos{X^{\epsilon,\delta}_{n}-\overline{X}_{\tau,h}^{\epsilon,\delta,n}}_{L^2(\Omega \times (0,T);\L)}
\\
\nonumber
&+\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}-\overline{X}_{\tau,h}^{\epsilon}}_{L^2(\Omega \times (0,T);\L)}
\\
\nonumber
& =:I+II+III+IV+V.
\end{align}
}
By Theorem \ref{Thm_reg.SVI} it follows that
\begin{align*}
\lim\limits_{\epsilon \rightarrow 0}\, I = \lim\limits_{\epsilon \rightarrow 0} \nos{X-X^{\epsilon}}_{L^2(\Omega \times (0,T);\L)}=0.
\end{align*}
To estimate the second term we consider the solutions $X^{\epsilon}_n$ of (\ref{reg.TVF}) with $x_0\equiv x_0^n$ and $g\equiv g^n$.
From (\ref{reg_stability_inequality}) we deduce that
{
\begin{align*}
\lim\limits_{n \rightarrow \infty}\, II = & \lim_{n\rightarrow \infty}
\nos{X^{\epsilon}-X^{\epsilon}_n}_{L^2(\Omega \times (0,T);\L)}
= 0\, .
\end{align*}
}
We use (\ref{delta_limit}) to estimate the third term as
$$
\lim\limits_{\delta\rightarrow 0}III=\lim_{\delta\rightarrow 0}\nos{X^{\epsilon}_n-X^{\epsilon,\delta}_{n}}_{L^2(\Omega \times (0,T);\L)} = 0\,.
$$
The fourth term is estimated by Lemma~\ref{Lemma_Convergence_num.vis.Scheme}
\begin{align*}
\lim\limits_{\tau,h\rightarrow 0}IV=\lim\limits_{\tau,h\rightarrow 0}\nos{X^{\epsilon,\delta}_{n}-\overline{X}_{\tau,h}^{\epsilon,\delta,n}}_{L^2(\Omega \times (0,T);\L)} =0.
\end{align*}
For the last term we use Lemma \ref{Lemma_Difference-num.Schemes}
\begin{align*}
\lim\limits_{n \rightarrow \infty}\limsup_{\delta \rightarrow 0} V=\lim\limits_{n \rightarrow \infty}\limsup\limits_{\delta \rightarrow 0}\nos{\overline{X}_{\tau,h}^{\epsilon,\delta,n}-\overline{X}_{\tau,h}^{\epsilon}}_{L^2(\Omega \times (0,T);\L)}=0.
\end{align*}
Finally, we consecutively take $\tau,h \rightarrow 0$, $\delta\rightarrow 0$,
$n\rightarrow \infty$ and $\epsilon \rightarrow 0$ in (\ref{err_ineq})
and use the above convergence of $I-V$ to obtain (\ref{numconv}).
\end{proof}
\begin{bems}
We note that the convergence analysis simplifies in the case that the problem data have higher regularity. For $x_0,g\in\mathbb{H}^1_0$
it is possible to show that the problem \eqref{reg.TVF} admits a unique variational solution (which is also a SVI solution of \eqref{reg.TVF} by uniqueness)
by a slight modification of standard monotonicity arguments.
This is due to the fact that the operator (\ref{Operator}) retains all its properties for $\delta=0$
except for the coercivity. The coercivity is only required to guarantee $\mathbb{H}^1_0$-stability of the solution, nevertheless the stability
can also be obtained directly by the It\^o formula on the continuous level, cf. Lemma~\ref{laplace_energy_estimate}, or analogically to Lemma~\ref{Lemma_Discrete a priori estimates} on the discrete level, even for $\delta=0$.
Consequently, for $\mathbb{H}^1_0$-data the convergence of the numerical solution $\overline{X}_{\tau,h}^{\epsilon}$ can be shown as in Theorem~\ref{Thm_Convergence_num.reg.Scheme} without the additional $\delta$-regularization step.
\end{bems}
We conclude this section by showing unconditional stability of scheme \eqref{num.reg.TVF}, i.e.,
we show that the numerical solution satisfies a discrete energy law which is an analogue of the energy estimate (\ref{cont_ener}).
\begin{lems}\label{eps.num.energy.estimates}
Let $x_0, g\in\L$ and $T> 0$.
Then there exist a constant $C\equiv C(T)$ such that
the solutions of scheme \eqref{num.reg.TVF} satisfy for any $\epsilon,h\in (0,1]$, $N\in \mathbb{N}$
\begin{align}\label{disc_ener}
\sup_{i=1,\ldots,N}\frac{1}{2}\mathbb{E}{\nos{\Xi}}&+\tau \mathbb{E}{\sum\limits_{i=1}^N\mathcal{J}_{\epsilon}(\Xi) +\frac{\lambda}{2}\nos{\Xi-g^h}}\ \nonumber\\
&\leq C\left(\frac{1}{2}\nos{x_0}+ T\epsilon\be{\O}+\frac{T\lambda}{2}\nos{g})\right)\,.
\end{align}
\end{lems}
\begin{proof}[\textbf{Proof of Lemma \ref{eps.num.energy.estimates}}]
We set $v_h\equiv\Xi$ in \eqref{num.reg.TVF} and obtain
\begin{align}\label{1}
& \frac{1}{2}\nos{\Xi}+\frac{1}{2}\nos{\Xi-X_{\epsilon,h}^{i-1}}+\tau \ska {\fe{\Xi},\nabla \Xi}+\tau \lambda(\Xi-g^h,\Xi)
\nonumber \\
& \qquad =\frac{1}{2}\nos{X_{\epsilon,h}^{i-1}}+(X_{\epsilon,h}^{i-1},\Xi)\Delta_i W.
\end{align}
Using the the convexity of $\mathcal{J}_\epsilon$ along with the identity
$$
(X_{\epsilon,h}^{i-1},\Xi)\Delta_i W = (X_{\epsilon,h}^{i-1},\Xi-X_{\epsilon,h}^{i-1})\Delta_i W + \|X_{\epsilon,h}^{i-1}\|^2\Delta_i W\,,
$$
we get from (\ref{1}) that
\begin{align}\label{2}
\frac{1}{2}\nos{\Xi}&+\frac{1}{2}\nos{\Xi-X_{\epsilon,h}^{i-1}}+\tau \mathcal{J}_{\epsilon}(\Xi) +\frac{\tau\lambda}{2}\nos{\Xi-g^h}\nonumber\\
\leq&\tau\mathcal{J}_{\epsilon}(0) +\frac{\tau\lambda}{2}\nos{g^h}+\frac{1}{2}\nos{X_{\epsilon,h}^{i-1}}
+\frac{1}{2}\nos{\Xi-X_{\epsilon,h}^{i-1}} \\
&+\frac{1}{2}\nos{X_{\epsilon,h}^{i-1}}|\Delta_i W|+\nos{X_{\epsilon,h}^{i-1}}\Delta_i W.\nonumber
\end{align}
After taking the expectation and summing up over $i$ in \eqref{2},
and noting that $\mathcal{J}_{\epsilon}(0) =\epsilon\be{\O}$ we obtain
\begin{align*}
\frac{1}{2}\mathbb{E}{\nos{\Xi}}&+\tau \mathbb{E}{\sum\limits_{k=1}^i\mathcal{J}_{\epsilon}(X^{k}_{\epsilon,h}) +\frac{\lambda}{2}\nos{X^{k}_{\epsilon,h}-g^h}} \nonumber\\
\leq &\frac{1}{2}\nos{x_0}+T\Big(\epsilon\be{\O}+\frac{\lambda}{2}\nos{g}\Big)+\frac{\tau}{2}\mathbb{E}{\sum\limits_{k=0}^{i-1}\nos{X^{k}_{\epsilon,h}}}.
\end{align*}
Hence (\ref{disc_ener}) follows after an application of the discrete Gronwall lemma.
\end{proof}
}
\section{Numerical experiments}\label{sec_sim}
We perform numerical experiments using a generalization of the fully discrete finite element (\ref{num.reg.TVF})
on the unit square $\O= (0,1)^2$.
The scheme for $i=1, \dots, N$ then reads as
\begin{align}\label{fem_scheme}
\ska{X^i_{\varepsilon,h},v_h} =& \ska{X^{i-1}_{\varepsilon,h},v_h}-\tau \ska{\fe{X^i_{\varepsilon,h}},\nablav_h } \nonumber \\
& -\tau\lambda\ska{X^i_{\varepsilon,h} - g^h,v_h}+ \mu \ska{\sigma(X^{i-1}_{\varepsilon,h}) {\Delta_i W^h},v_h} &&\forall v_h \in \mathbb{V}_h\,,
\\ \nonumber
X^0_{\varepsilon,h} = & x_0^h\,,
\end{align}
where $g^h,\,x_0^h\in\mathbb{V}_h$ are suitable approximations of $g$, $x_0$ (e.g., the orthogonal projections onto $\mathbb{V}_h$), respectively,
and $\mu>0$ is a constant. The multiplicative space-time noise $\sigma(X^{i-1}_{\varepsilon,h})\Delta_i W^h$ is constructed as follows.
The term $W^h$ is taken to be a $\mathbb{V}_h$-valued space-time noise of the form
$$
\Delta_i {W}^h(x) = \sum_{\ell=1}^{L} \phi_\ell (x) \Delta_{i} {\beta}_\ell
\qquad \forall\, x \in \overline{\mathcal D}\,,
$$
where ${\beta}_\ell$, $\ell=1,\dots, L$ are independent scalar-valued Wiener processes and
$\{\phi_\ell\}_{\ell=1}^L$ is the standard 'nodal' finite element basis of $\mathbb{V}_h$.
In the simulations below we employ three practically relevant choices of $\sigma$:
a tracking-type noise $\sigma(X) \equiv \sigma_1(X) = |X - g^h|$,
a gradient type noise
$\sigma(X) \equiv \sigma_2(X) = |\nabla X |$ and the additive noise $\sigma(X)\equiv \sigma_3 = 1$;
in the first case the noise is small when the solution is close to the 'noisy image' $g^h$,
in the gradient noise case the noise is localized along the edges of the image.
We note that the fully discrete finite element scheme (\ref{fem_scheme}) corresponds to an approximation of the regularized equation (\ref{reg.TVF})
with a slightly more general space-time noise term of the form $\mu \sigma(X^{\epsilon}) \mathrm{d}W$.
In all experiments we set $T=0.05$, $\lambda=200$, $x_0 \equiv x_0^h\equiv 0$.
If not mentioned otherwise we use
the time step $\tau = 10^{-5}$, the mesh size $h = 2^{-5}$ and set $\epsilon=h=2^{-5}$, $\mu=1$.
We define $g \in \mathbb{V}_h$ as a piecewise linear interpolation
of the characteristic function of a circle with radius $0.25$ on the finite element mesh, see Figure~\ref{fig_data} (left),
and set $g^h = g + \xi_h \in \mathbb{V}_h$ with $\displaystyle \xi_h(x) = \nu \sum_{\ell=1}^{L} \phi_\ell (x) \xi_\ell$, $x\in\O$
where $\xi_\ell$, $\ell=1,\dots, L$ are realizations of independent $\mathcal{U}(-1,1)$-distributed random variables.
If not indicated otherwise we use $\nu=0.1$; the corresponding realization of $\xi_h$ is displayed in Figure~\ref{fig_data} (right).
\begin{figure}[!htp]
\center
\includegraphics[width=0.3\textwidth]{g}
\includegraphics[width=0.3\textwidth]{ng}
\includegraphics[width=0.3\textwidth]{g_noise}
\caption{The function $g$ (left), the noise $\xi_h$ (middle) and the noisy image (right).}
\label{fig_data}
\end{figure}
We choose $\epsilon=h=2^{-5}$, $\mu=1$, $\sigma\equiv \sigma_1$ as parameters for the 'baseline' experiment;
the individual parameters are then varied in order to demonstrate their influence on the evolution.
The time-evolution of the discrete energy functional $\mathcal{J}_{\epsilon,\lambda}(\Xi)$, $i=1,\dots,N$
for a typical realization of the space-time noise $W^h$ is displayed in Figure~\ref{fig_ener}; in the legend of the graph we state parameters
which differ from the parameters of the baseline experiment, e.g., the legend '$sigma_2,\, mu=0.125$'
corresponds to the parameters $\sigma\equiv \sigma_2$, $\mu=0.125$ and the remaining parameters are left unchanged, i.e., $\epsilon=h=2^{-5}$.
For all considered parameter setups, except for the case of noisier image $\nu=0.2$,
the evolution remained close to the discrete energy of the deterministic problem (i.e., (\ref{fem_scheme}) with $\mu=0$).
The energy decreases over time until the solution is close to the (discrete) minimum of $\mathcal{J}_{\epsilon,\lambda}$;
to highlight the differences we display a zoom at the graphs.
We observe that in the early stages (not displayed) the energy of stochastic evolutions with sufficiently small noise typically remained
below the energy of the deterministic problems and the situation reversed as the solution approached the stationary state.
\begin{figure}[!htp]
\center
\includegraphics[width=0.3\textwidth]{ener1}
\includegraphics[width=0.3\textwidth]{ener2}
\includegraphics[width=0.3\textwidth]{ener2_zoom}
\caption{Evolution of the discrete energy: $\sigma\equiv\sigma_1$, $h=2^{-5}$, $\epsilon=h,\frac{h}{2}$, $\mu=1,2$ (left);
$\sigma\equiv\sigma_1,\sigma_2,\sigma_3$, $\sigma\equiv\sigma_1$, $h=2^{-5}$, $\epsilon=2h$, $\sigma=\sigma_1$, $\epsilon=h=2^{-6}$ and $\nu=0.2$ (middle and right).
}
\label{fig_ener}
\end{figure}
In Figure~\ref{fig_finalu} we display the solution at the final time
computed with $\sigma \equiv \sigma_1$, $\epsilon=h$ for $h=2^{-5}, 2^{-6}$, respectively,
and $\sigma \equiv \sigma_2$, $\epsilon=h=2^{-5}$; graphically the results of the remaining simulations
did not significantly differ from the first case.
The displayed results may indicate that the noise $\sigma_2$ yields worse results than the noise $\sigma_1$ and $\sigma_2$;
however, for sufficiently small value of $\mu$ the results would remain close to the deterministic simulation as well.
We have magnified noise intensity $\mu$ to
highlight the differences to the other noise types (i.e., the noise is concentrated along the edges of the image).
We note that the gradient type noise $\sigma_2$ might be a preferred choice for practical computations, cf. \cite{swp14}.
\begin{figure}[!htp]
\center
\includegraphics[width=0.3\textwidth]{uh_final}
\includegraphics[width=0.3\textwidth]{uh_final_h_small}
\includegraphics[width=0.3\textwidth]{uh_final_gradnoise}
\caption{From let to right:
the solution for $\sigma \equiv \sigma_1$ with $\epsilon= h=2^{-5}$,
$\sigma \equiv \sigma_1$ with $\epsilon= h =2^{-6}$ and
$\sigma \equiv \sigma_2$ with $\epsilon=2^{-5}$.}
\label{fig_finalu}
\end{figure}
\section*{Acknowledgement}
This work was supported by the Deutsche Forschungsgemeinschaft through SFB 1283 “Taming uncertainty and profiting from randomness and low regularity in analysis, stochastics and their applications”.
The authors would like to thank the referee for careful reading of the manuscript and constructive comments, as well as to Lars Diening for stimulating discussions.
We would also like to thank Martin Ondrej\'at for pointing to us inaccuracies in the proof of uniqueness in Theorem~\ref{Thm_reg.SVI}.
\bibliographystyle{plain}
|
\section{Introduction}
\IEEEPARstart{Q}{uantifying} the correlations between disjoint subsystems of a quantum state is a fundamental problem in quantum information theory. Since correlations cannot be generated by local operations, measures of correlation must be non-increasing under local processing. For measures which are functions of the von Neumann entropy ($S(\rho) = -\text{Tr}~\rho\log\rho$), this is equivalent to being non-increasing under partial trace \footnote{This is because any processing can be written as an isometry followed by a partial trace, and the isometry will not affect entropies.}, i.e.
\begin{align}
E(\rho_{(A_1A_2)(B_1B_2)})&\geq E(\rho_{A_1B_1})\label{mon}
\end{align}
for a correlation measure $E$.
In this work, we will take (\ref{mon}) to be the defining property of a bipartite correlation measure. This property has been studied by \cite{alhejji2018monotonicity} for linear entropic quantities, but here we wish to identify bipartite correlation measures formed by minimizing a linear entropic quantity over all purifications of a state $\rho_{AB}$. More formally, we will study the space of quantities of the form
\begin{align}
E_\alpha(\rho_{AB}) = \inf_{\psi:\text{Tr}_{A'B'}\dyad{\psi}_{AA'BB'} = \rho_{AB}}f^\alpha(\dyad{\psi}_{AA'BB'})\nonumber
\end{align}
where $\alpha\in\mathbb{R}^{15}$ and
\begin{align}
f^\alpha(\dyad{\psi}_{AA'BB'}) = \sum_{\emptyset\neq\mathcal{J}\subseteq\{A,B,A',B'\}}\alpha_{\mathcal{J}}S_{\mathcal{J}}\nonumber
\end{align}
(each entry of $\alpha$ corresponds to a non-empty subset of $\{A,B,A',B'\}$), and identify instances which satisfy (\ref{mon}). Quantities of this form are of particular interest, since they often admit operational interpretations, usually in the form of bounds on performance in information theoretic tasks. Examples include the squashed entanglement \cite{doi:10.1063/1.1643788}, the entanglement assisted capacity \cite{bennett2002entanglement}, and the entanglement of purification \cite{doi:10.1063/1.1498001}.
In a pure state, the entropy of any subsystem is equal to the entropy of its complement, so we can remove redundancy from our search space by rewriting it as the set of quantities of the form
\begin{align}
E_\alpha(\rho_{AB}) = \inf_{\rho_{ABV}:\text{Tr}_V\rho_{ABV}=\rho_{AB}} f^\alpha(\rho_{ABV})\label{form2}
\end{align}
where $\alpha \in \mathbb{R}^7$ and
\begin{align}
f^\alpha(\rho_{ABV}) = \sum_{\emptyset\neq\mathcal{J}\subseteq\{A,B,V\}}\alpha_{\mathcal{J}}S_{\mathcal{J}}.
\end{align}
Note that the minimization is now over all extensions $\rho_{ABV}$, not only purifications $\dyad{\psi}_{AA'BB'}$.
By first examining the entanglement of purification, a well-known instance of (\ref{form2}) which satisfies (\ref{mon}), we are led to the construction of four convex polyhedral cones in $\mathbb{R}^7$. These four cones consist of $\alpha$ vectors which give rise to optimized bipartite correlation measures which all satisfy (\ref{mon}). We examine the extreme rays of these cones and find four nontrivial rays, two of which are new.
We study these two correlation measures and find several useful properties, including lower and upper bounds, additivity, and a relationship to the regularized entanglement of purification.
This paper is organized as follows. In Section II we present a proof of the monotonicity of the entanglement of purification, in order to illustrate our method for identifying monotones of the form (\ref{form2}). Guided by the proof in Section II, in Section III we define two different types of monotonicity and identify all monotones of each type. In Section IV, we examine the monotones found in Section III and find that many are trivial in the sense that they are equal to $I_{A:B}$ or $0$. After identifying some nontrivial monotones in Section IV, in Section V we go on to prove several important properties of these monotones.
Throughout this paper, for compactness of notation, we will denote all entropic quantities using subscripts. The entropy of subsystem $A$ will be denoted $S_A$, the entropy of $A$ conditioned on $B$ ($\equiv S_{AB} - S_B$) will be denoted $S_{A|B}$, the mutual information of $A$ and $B$ ($\equiv S_A + S_B - S_{AB}$) will be denoted $I_{A:B}$, and the mutual information of $A$ and $B$ conditioned on $V$ ($\equiv S_{AV} + S_{BV} - S_{ABV} - S_V$) will be denoted $I_{A:B|V}$.
\section{The entanglement of purification}
A well-known example of an optimized bipartite correlation measure is the entanglement of purification \cite{doi:10.1063/1.1498001}
\begin{align}
E_P(\rho_{AB}) = \inf_{\rho_{ABV}:\text{Tr}_V\rho_{ABV}=\rho_{AB}} S_{AV},\nonumber
\end{align}
i.e. $\alpha_{AV} = 1$ and $\alpha_\mathcal{J} = 0$ for all other $\mathcal{J}$. In this section we prove that (\ref{mon}) holds for $E_\alpha = E_P$, i.e. that $E_P$ is monotonically non-increasing under local processing of both subsystems of a bipartite state\footnote{This was first shown by \cite{doi:10.1063/1.1498001}, using a different method from the one used here.}. The proofs for monotocity under $A$-processing and $B$-processing are different, and point towards a method for identifying instances of (\ref{form2}) which satisfy (\ref{mon}).
First we show that $E_P(\rho_{AB})$ is monotone under $B$-processing. For each extension $\rho_{AB_1B_2V}$ of $\rho_{AB_1B_2}$, we will construct an extension of $\rho_{AB_1}$ whose value of $S_{AV}$ is no greater than that of $\rho_{AB_1B_2V}$. Given the extension $\rho_{AB_1B_2V}$ of $\rho_{AB_1B_2}$, consider the state
\begin{align}
\rho_{AB_1V}' = \text{Tr}_{B_2}\left[\rho_{AB_1B_2V}\right],\nonumber
\end{align}
which is an extension of the state $\rho_{AB_1}$. Now note that
\begin{align}
S_{AV}'\equiv S\left(\text{Tr}_{B_1}\left[\rho_{AB_1V}'\right]\right) = S\left(\text{Tr}_{B_1B_2}\left[\rho_{AB_1B_2V}\right]\right)\equiv S_{AV}.\nonumber
\end{align}
Therefore, every value of $S_{AV}$ achievable by an extension of the unprocessed state $\rho_{AB_1B_2}$ can also be achieved by an extension of the processed state $\rho_{AB_1}$. Thus, the minimum value of $S_{AV}$ for extensions of the processed state is no greater than the minimum value of $S_{AV}$ for extensions of the unprocessed state, which is exactly the statement that
\begin{align}
E_P(\rho_{AB_1B_2})\geq E_P(\rho_{AB_1}).\nonumber
\end{align}
This was a roundabout way of saying that $E_P(\rho_{AB})$ is monotone under $B$-processing because $S_{AV}$ itself (without the minimization) is monotone under $B$-processing.
Now we show that $E_P(\rho_{AB})$ is monotone under $A$-processing. Our method is the same, i.e., for each extension $\rho_{A_1A_2BV}$ of $\rho_{A_1A_2B}$, we construct an extension of $\rho_{A_1B}$ whose value of $S_{AV}$ is no greater than that of $\rho_{A_1A_2BV}$. Given the extension $\rho_{A_1A_2BV}$ of $\rho_{A_1A_2B}$, consider the state $\rho'_{\hat{A}B\hat{V}}=\rho_{A_1B(A_2V)}'$, where $\hat{A} = A_1$ and $\hat{V} = A_2V$. This is the same global state but written as an extension of the processed state $\rho_{A_1B}$. Now note that
\begin{align}
S_{\hat{A}\hat{V}}'\equiv S\left(\text{Tr}_B[\rho_{A_1B(A_2V)}']\right) = S\left(\text{Tr}_B\left[\rho_{A_1A_2BV}\right]\right)\equiv S_{AV},\nonumber
\end{align}
so we have shown that
\begin{align}
E_P(\rho_{A_1A_2B})\geq E_P(\rho_{A_1B}),\nonumber
\end{align}
which completes the proof of inequality (\ref{mon}).
\section{Monotones}
\subsection{Monotonicity types}
The main point to take away from the previous section is that for quantities of the form (\ref{form2}), there are two types of monotonicity we can identify. One way for a quantity $E_\alpha$ to be monotonic under processing of a subsystem $X\in \{A,B\}$ is for the associated $f^\alpha$ to be monotonic under processing of $X$. In this case, monotonicity of $E_\alpha$ is proved by starting with an extension of an unprocessed state and constructing from it an extension of an $X$-processed state by simply tracing out a subsystem $X_2$ of $X=X_1X_2$, as in the above proof of monotonicity of $E_P$ under $B$-processing. The monotonicity of $f^\alpha$ then implies the monotonicity of $E_\alpha$. This type of monotonicity (under, say, $A$-processing) is therefore characterized by the inequality
\begin{align}
f^\alpha(\rho_{A_1A_2BV})\geq f^\alpha(\rho_{A_1BV}).\label{ineq0}
\end{align}
But, as we saw for $E_P$, monotonicity of $f^\alpha$ is not necessary for monotonicity of $E_\alpha$. All that is necessary is for $f^\alpha$ to be monotonic under some operation which constructs an extension of a processed state from an extension of an unprocessed state. One such operation is a rearrangement of the subsystems making up the unprocessed state, as in the proof of monotonicity of the $E_P$ under $A$-processing. In this case, monotonicity of $E_\alpha$ under $A$-processing is implied by monotonicity of $f^\alpha$ under the operation $\rho_{A_1A_2BV}\to\rho_{A_1B(A_2V)}$, i.e. placing $A_2$ with $V$ in order to write the state as an extension of the processed state $\rho_{A_1B}$. This type of monotonicity, again under $A$-processing, is therefore characterized by the inequality
\begin{align}
f^\alpha(\rho_{A_1A_2BV})\geq f^\alpha(\rho_{A_1B(A_2V)}).\label{ineq1}
\end{align}
We will refer to monotonicity of the types characterized by (\ref{ineq0}) and (\ref{ineq1}) as 0-monotonicity and 1-monotonicity, respectively. A quantity $E_\alpha$ can now be monotonic under both $A$- and $B$-processing in four (not mutually exclusive) ways that we can identify. These quantities can be 00-, 01-, 10-, or 11-monotonic, where the first bit indicates whether $E_\alpha$ is 0- or 1-monotonic on $A$, and the second on $B$. As an example, we have shown $E_P$ to be 10-monotonic.
Note that there is still a redundancy in the $\alpha$ vectors, due to a purification symmetry. Given an extension $\rho_{ABV}$ of $\rho_{AB}$, we can form a canonical dual extension by purifying to $\rho_{ABVW}$, and tracing out $V$ to form $\rho_{ABW}$. Now, given $f^\alpha$, there exists $f^\beta$ for which $f^\alpha(\rho_{ABV}) = f^\beta(\rho_{ABW})$ (implying $E_\alpha = E_\beta$). Using the fact that entropies of complimentary subsystems are equal in a pure state, we see that $\beta_{AV} = \alpha_{BV}$, $\beta_{BV} = \alpha_{AV}$, $\beta_{ABV} = \alpha_V$, and $\beta_V = \alpha_{ABV}$. Also note that this symmetry takes 0-monotones to 1-monotones, and vice-verse. To see this, observe that under the purification symmetry, the operations defining 0- and 1-monotonicity ($\rho_{A_1A_2BV}\to\rho_{A_1BV}$ and $\rho_{A_1A_2BV}\to\rho_{A_1B(A_2V)}$, respectively) become
\begin{align}
&0:~~\rho_{A_1A_2BV}\to\rho_{A_1A_2BVW}\to\rho_{A_1BV(A_2W)}\to \rho_{A_1B(A_2W)}\nonumber
\\&1:~~\rho_{A_1A_2BV}\to\rho_{A_1A_2BVW}\to\rho_{A_1B(A_2V)W}\to\rho_{A_1BW}.\nonumber
\end{align}
This means we need only study the 00- and 10-monotones, since the 11- and 01-monotones are redundant via the purification symmetry.
\subsection{Monotonicity cones}
Expanding (\ref{ineq0}) and (\ref{ineq1}) in terms of the coefficients $\alpha_{\mathcal{J}}$ and moving all terms to one side, we see that quantities $E_\alpha$ which are 0- or 1-monotonic on $A$ are those for which $\alpha$ satisfies
\begin{multline}
\alpha_AS_{A_2|A_1} + \alpha_{AB}S_{A_2|A_1B}
\\ + \alpha_{AV}S_{A_2|A_1V} + \alpha_{ABV}S_{A_2|A_1BV}\geq 0\stepcounter{equation}\tag{\arabic{equation}-0A}\label{0A}
\end{multline}
or
\begin{multline}
\alpha_AS_{A_2|A_1} + \alpha_{AB}S_{A_2|A_1B}
\\- \alpha_{BV}S_{A_2|BV} - \alpha_VS_{A_2|V} \geq 0,\tag{\arabic{equation}-1A}\label{1A}
\end{multline}
respectively. Here $S_{A|B} = S_{AB} - S_B$ is the conditional entropy. Swapping the roles of $A$ and $B$ in inequalities(\ref{0A}) and (\ref{1A}) gives inequalities
\begin{multline}
\alpha_BS_{B_2|B_1} + \alpha_{AB}S_{B_2|B_1A}
\\+ \alpha_{BV}S_{B_2|B_1V} + \alpha_{ABV}S_{B_2|B_1AV}\geq 0\tag{\arabic{equation}-0B}\label{0B}
\end{multline}
and
\begin{multline}
\alpha_BS_{B_2|B_1} + \alpha_{AB}S_{B_2|B_1A}
\\- \alpha_{AV}S_{B_2|AV} - \alpha_VS_{B_2|V} \geq 0,\tag{\arabic{equation}-1B}\label{1B}
\end{multline}
satisfied by those $\alpha$ for which $E_\alpha$ is 0- or 1-monotonic on $B$, respectively. The set of all $\alpha\in\mathbb{R}^7$ for which (\ref{0A}) or (\ref{1A}) is implied by strong subadditivity (SSA) ($I_{A:B|C}\geq 0$) \cite{doi:10.1063/1.1666274, pippenger2003inequalities} and weak monotonicity (WM) ($S_{C|A} + S_{C|B}\geq 0$) \cite{doi:10.1063/1.1666274} of the von Neumann entropy, for any 4-partite state $\rho_{A_1A_2BV}$, form convex polyhedral cones in $\mathbb{R}^7$. Similarly, the set of all $\alpha\in\mathbb{R}^7$ for which (\ref{0B}) or (\ref{1B}) is implied by SSA and WM for any 4-partite state $\rho_{AB_1B_2V}$ also form convex polyhedral cones. Since the intersection of two convex polyhedral cones is a convex polyhedral cone, the set of all 00-, 10-, 01-, and 11-monotonic quantities (i.e., those $\alpha$ which satisfy, respectively, (\ref{0A}) and (\ref{0B}), (\ref{1A}) and (\ref{0B}), (\ref{0A}) and (\ref{1B}), (\ref{1A}) and (\ref{1B})) each form a convex polyhedral cone. Using SAGE's\footnote{SAGE is a Python-based open-source mathematics software available at www.sagemath.org} rational convex polyhedral cone module, together with the constraints on entropy vectors implied by SSA, one can determine that the 00- and 10-cones are generated by the extreme rays given by the rows of Table \ref{tab:table1}.
\begin{table}[h]
\begin{center}
\caption{Rows are the extreme rays of the 00- and 10-monotone cones in $\mathbb{R}^7$.}
\label{tab:table1}
\begin{tabular}{c|c|c|c|c|c|c|c}
\text{Cone} & $\alpha_A$ & $\alpha_B$ & $\alpha_V$ & $\alpha_{AB}$ & $\alpha_{AV}$ & $\alpha_{BV}$ & $\alpha_{ABV}$
\\\hline
00 & 1 & 1 & 0 & -1 & 0 & 0 & 0
\\~& 1 & 0 & 0 & 0 & -1 & 0 & 0
\\~& 0 & 0 & 0 & 0 & 1 & 1 & -1
\\~& 0 & 0 & 0 & 1 & 0 & 0 & -1
\\~& 0 & 1 & 0 & 0 & 0 & -1 & 0
\\~& 0 & 0 & 1 & 0 & 0 & 0 & 0
\\~& 0 & 0 & -1 & 0 & 0 & 0 & 0
\\\hline
10 & 1 & 1 & 0 & -1 & 0 & 0 & 0
\\~& 0 & 0 & -1 & 0 & 0 & 1 & -1
\\~& 1 & 0 & -1 & 0 & 0 & 0 & 0
\\~& 0 & 1 & 0 & 0 & 0 & 0 & -1
\\~& 1 & 1 & 0 & 0 & 0 & -1 & 0
\\~& 0 & 0 & -1 & 1 & 0 & 0 & -1
\\~& 0 & 0 & 0 & 0 & 1 & 0 & 0
\\~& 0 & 0 & 0 & 0 & -1 & 0 & 0
\end{tabular}
\end{center}
\end{table}
\subsection{Non-negativity in $V$}
For certain $\alpha$, $E_\alpha$ is $-\infty$. If, for some $\alpha$,
\begin{align}
\sum_{\substack{\mathcal{J}\subseteq\{A,B,V\}\\V\in\mathcal{J}}} \alpha_{\mathcal{J}}<0,\label{balance}
\end{align}
then we can achieve an arbitrarily large negative value of $f^\alpha$ for any state $\rho_{AB}$ by choosing an extension of the form $\rho_{ABV} = \rho_{AB}\otimes\mathbb{I}_k/k$, for sufficiently large $k$. So for $\alpha$ satisfying (\ref{balance}), $E_\alpha$ is $-\infty$. Therefore we are only interested in those $\alpha$ which satisfy
\begin{align}
\sum_{\substack{\mathcal{J}\subseteq\{A,B,V\}\\V\in\mathcal{J}}} \alpha_{\mathcal{J}}\geq 0,\nonumber
\end{align}
or equivalently, those $\alpha$ which satisfy
\begin{align}
\alpha\cdot (0,0,1,0,1,1,1)\geq 0.\label{balance2}
\end{align}
The set of all $\alpha$ satisfying (\ref{balance2}) form another convex cone $\mathcal{C}$ in $\mathbb{R}^7$, in fact they form the halfspace whose boundary is the plane through the origin with normal vector $(0,0,1,0,1,1,1)$. Now we can intersect each of the three cones shown in Table \ref{tab:table1} with the cone $\mathcal{C}$, in order to keep only those $\alpha$ satisfying (\ref{balance2}). The resulting cones (also obtained via SAGE's rational convex polyhedral cone module) are given by the extreme rays in Table \ref{tab:table2}.
\begin{table}[h]
\begin{center}
\caption{Rows are the extreme rays of the cones formed by intersecting the cones of 00- and 10-monotones with the cone $\mathcal{C}$ of vectors $\alpha\in\mathbb{R}^7$ which are non-negative in $V$ (i.e. satisfy (\ref{balance2})).}
\label{tab:table2}
\begin{tabular}{c|c|c|c|c|c|c|c||c}
\text{Cone} & $\alpha_A$ & $\alpha_B$ & $\alpha_V$ & $\alpha_{AB}$ & $\alpha_{AV}$ & $\alpha_{BV}$ & $\alpha_{ABV}$ & \text{label}
\\\hline
$\mathcal{C}\cap 00$ & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 1
\\~& 1 & 1 & 0 & -1 & 0 & 0 & 0 & 2
\\~& 0 & 0 & -1 & 0 & 1 & 1 & -1 & 3
\\~& 0 & 0 & 1 & 1 & 0 & 0 & -1 & 4
\\~& 0 & 1 & 1 & 0 & 0 & -1 & 0 & 5
\\~& 1 & 0 & 1 & 0 & -1 & 0 & 0 & 6
\\\hline
$\mathcal{C}\cap 10$ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 7
\\~& 1 & 1 & 0 & -1 & 0 & 0 & 0 & 8
\\~& 0 & 0 & -1 & 0 & 1 & 1 & -1 & 9
\\~& 1 & 1 & 0 & 0 & 1 & -1 & 0 & 10
\\~& 0 & 0 & -1 & 1 & 2 & 0 & -1 & 11
\\~& 0 & 1 & 0 & 0 & 1 & 0 & -1 & 12
\\~& 1 & 0 & -1 & 0 & 1 & 0 & 0 & 13
\end{tabular}
\end{center}
\end{table}
\section{A closer look at the monotone cones}
The extreme rays of a convex polyhedral cone generate the cone via conical combinations (real linear combinations with non-negative coefficients), so any conical combination of the extreme rays of one of the two cones above gives a monotonic $E_\alpha$\footnote{Since the infimum of a conic combination is not generally equal to the conic combination of infima, the extreme rays are somewhat less priveleged in the optimized setting. In other words there may be interesting quantities in these cones, other than the ones given by the extreme rays. In this paper we do not discuss these quantities, but study the extreme rays as a starting point.}. But in some cases these quantities can be trivial. We will see that many of the rays in the cones in Table \ref{tab:table2} give $E_\alpha=0$, or $E_\alpha\propto I_{A:B}$.
We first examine $\mathcal{C}\cap 00$. Extreme ray 1 gives $f^\alpha = S_V$, which is non-negative and achieves the value 0 for any $\rho_{AB}$ via the trivial extension. So for this $\alpha$, we have $E_\alpha=0$. Ray 2 gives $f^\alpha = I_{A:B}$, which gives $E_\alpha = I_{A:B}$ and the minimum is achieved by any extension. Rays 4, 5, and 6 give, respectively, $f^\alpha = I_{AB:V}$, $f^\alpha = I_{B:V}$, and $f^\alpha = I_{A:V}$, which are non-negative and achieve the value 0 for any $\rho_{AB}$ via the trivial extension. So for these three $\alpha$'s we also have $E_\alpha=0$. So the only extreme ray of $\mathcal{C}\cap 00$ which is not minimized by the trivial extension and does not have $E_\alpha=0$ or $E_\alpha=I_{A:B}$ is ray 3, which gives $E_\alpha(\rho_{AB}) = E_{sq}(\rho_{AB})$ (the squashed entanglement \cite{doi:10.1063/1.1643788}, $E_{sq}(\rho_{AB}) = \inf_{\rho_{ABV}}(I_{A:B|V})$). So any ray in $\mathcal{C}\cap 00$ which can be written as a conical combination of extreme rays that does not include the $E_{sq}$ ray will give $E_\alpha(\rho_{AB}) \propto I_{A:B}$ and is therefore trivial.
$\mathcal{C}\cap 10$ is where we will find an abundance of nontrivial quantities. There are only three trivial extreme rays, and they cannot be simultaneously minimized as in the two previous cones. Ray 8 is equal to rays 15 and 2, and again gives $E_\alpha = I_{A:B}$. Ray 12 gives $f^\alpha = I_{B:AV}$, which SSA implies is bounded below by $I_{A:B}$, and achieves the value $I_{A:B}$ via the trivial extension. Ray 13 gives $f^\alpha = S_A + S_{AV} - S_V$, which WM implies is bounded below by $I_{A:B}$, and achieves the value $I_{A:B}$ via any purification of $\rho_{AB}$.
The four remaining extreme rays of $\mathcal{C}\cap 10$ are non-trivial. Ray 9 is $E_{sq}$ which, interestingly, appears in all four monotonicity cones. Rays 7, 10, and 11 are (up to a scaling by 1/2, the reason for which will be clear in the next section)
\begin{align}
f^{P} & = S_{AV}\nonumber
\\f^{Q} &= \frac{1}{2}(S_A + S_B + S_{AV} - S_{BV})\nonumber
\\f^{R} &= \frac{1}{2}(S_{AB} + 2S_{AV} - S_{ABV} - S_V),\nonumber
\end{align}
respectively. $f^P$ gives $E_P$, which we expected to find. $E_Q$ and $E_R$ are new, and we will see that $E_Q$ and $E_R$ have several useful properties.
\section{Properties of $E_Q$ and $E_R$}
\subsection{Lower and upper bounds}
\begin{theorem}\footnote{This was proven for $E_P$ in \cite{doi:10.1063/1.1498001}}
$E_Q$ and $E_R$ satisfy
\begin{align}
\frac{1}{2}I_{A:B}\leq E(\rho_{AB}) &\leq \min\{S_A,S_B\}.\label{bounds}
\end{align}
\end{theorem}
\begin{proof}
We start with the upper bound. Both $f^{Q}$ and $f^{R}$ achieve a value of $S_A$ via the trivial extension, and a value of $S_B$ via any purification. Therefore $E_Q$ and $E_R$ satisfy the upper bound. To prove the lower bound in (\ref{bounds}) for $E_Q$, observe that
\begin{multline}
2f^{Q} - I_{A:B} = S_{AV} + S_{AB} - S_{BV}
\\\geq S_{AV} + S_{AB} - S_B - S_V = S_{A|V} + S_{A|B}\geq 0,\nonumber
\end{multline}
where the first inequality follows from subadditivity and the second from WM. Therefore $E_Q$ satisfies the lower bound. To prove the lower bound for $E_R$, observe that
\begin{multline}
2f^{R} - I_{A:B} = (S_{AB} + S_{AV} - S_{ABV} - S_A)
\\+ (S_{AB} + S_{AV} - S_V - S_B)
\\ = I_{B:V|A} + S_{A|B} + S_{A|V}\geq I_{B:V|A}\geq 0,\nonumber
\end{multline}
where the first inequality follows from WM and the second from SSA. Therefore $E_R$ satisfies the lower bound.
\end{proof}
\subsection{Additivity}
\begin{theorem}\label{Add}
$E_Q$ and $E_R$ are additive\footnote{$E_P$ is believed to be non-additive \cite{chen2012non}.}, i.e.
\begin{align}
E(\rho_{A_1B_1}\otimes\rho_{A_2B_2})= E(\rho_{A_1B_1}) + E(\rho_{A_2B_2}).\nonumber
\end{align}
\end{theorem}
\begin{proof} For brevity we will exclude the factor of 1/2. Fix two bipartite states $\rho_{A_1B_1}$ and $\rho_{A_2B_2}$, and form the bipartite state $\rho_{AB} = \rho_{A_1B_1}\otimes\rho_{A_2B_2}$ with $A=A_1A_2$ and $B=B_1B_2$. We start with $E_Q$.
First we show that
\begin{align}
E_Q(\rho_{A_1B_1}\otimes\rho_{A_2B_2})\geq E_Q(\rho_{A_1B_1}) + E_Q(\rho_{A_2B_2}).\nonumber
\end{align}
Let $\rho_{ABV}$ be an extension of $\rho_{AB}$. Now let $V_1\equiv A_2V$ and $V_2\equiv B_1V$ and consider the extensions $\rho_{A_1B_1V_1}$ and $\rho_{A_2B_2V_2}$ of $\rho_{A_1B_1}$ and $\rho_{A_2B_2}$. Since $\rho_{A} = \rho_{A_1}\otimes\rho_{A_2}$ and $\rho_{B} = \rho_{B_1}\otimes\rho_{B_2}$, we have
\begin{multline}
f^{Q}(\rho_{A_1B_1V_1}) + f^{Q}(\rho_{A_2B_2V_2})
\\= (S_{A_1} + S_{B_1} + S_{A_1V_1} - S_{B_1V_1})
\\+ (S_{A_2} + S_{B_2} + S_{A_2V_2} - S_{B_2V_2})
\\= S_A + S_B + S_{AV} - S_{B_1A_2V} + S_{B_1A_2V} - S_{BV}
\\= S_A + S_B + S_{AV} - S_{BV} = f^Q(\rho_{ABV}).\nonumber
\end{multline}
Now we show that
\begin{align}
E_Q(\rho_{A_1B_1}) + E_Q(\rho_{A_2B_2})\geq E_Q(\rho_{A_1B_1}\otimes\rho_{A_2B_2}).\nonumber
\end{align}
Let $\rho_{A_1B_1V_1}$ and $\rho_{A_2B_2V_2}$ be extensions of $\rho_{A_1B_1}$ and $\rho_{A_2B_2}$, and consider the extension of $\rho_{AB}$ given by
\begin{align}
\rho_{ABV} = \rho_{A_1B_1V_1}\otimes \rho_{A_2B_2V_2},
\end{align}
with $V \equiv V_1V_2$. Now using the fact that $\rho_{AV} = \rho_{A_1V_1}\otimes\rho_{A_2V_2}$ and $\rho_{BV} = \rho_{B_1V_1}\otimes\rho_{B_2V_2}$ we have
\begin{multline}
f^Q(\rho_{ABV}) = S_A + S_B + S_{AV} - S_{BV}
\\=S_{A_1} + S_{A_2} + S_{B_1} + S_{B_2} + (S_{A_1V_1} + S_{A_2V_2}) - (S_{B_1V_1} + S_{B_2V_2})
\\= (S_{A_1} + S_{B_1} + S_{A_1V_1} - S_{B_1V_1}) + (S_{A_2} + S_{B_2} + S_{A_2V_2} - S_{B_2V_2})
\\= f^{Q}(\rho_{A_1B_1V_1})+ f^Q(\rho_{A_2B_2V_2}).\nonumber
\end{multline}
Therefore $E_Q$ is additive.
Now we wish to prove the same thing for $E_R$, i.e.
\begin{align}
E_R(\rho_{A_1B_1}\otimes\rho_{A_2B_2})= E_R(\rho_{A_1B_1}) + E_R(\rho_{A_2B_2}).\nonumber
\end{align}
First we show that
\begin{align}
E_R(\rho_{A_1B_1}\otimes\rho_{A_2B_2})\geq E_R(\rho_{A_1B_1}) + E_R(\rho_{A_2B_2}).\nonumber
\end{align}
Let $\rho_{ABV}$ be an extension of $\rho_{AB}$. Now let $V_1 \equiv A_2V$ and $V_2 \equiv A_1V$ (see \cite{PhysRevLett.118.040501}) and consider the extensions $\rho_{A_1B_1V_1}$ and $\rho_{A_2B_2V_2}$ of $\rho_{A_1B_1}$ and $\rho_{A_2B_2}$. We wish to show that $f^R(\rho_{ABV})\geq f^R(\rho_{A_1B_1V_1}) + f^R(\rho_{A_2B_2V_2}),$ i.e.
\begin{multline}
S_{A_1A_2B_1B_2} + 2S_{A_1A_2V} - S_{A_1A_2B_1B_2V} - S_{V}\\\geq(S_{A_1B_1} + 2S_{A_1VA_2} - S_{A_1B_1VA_2} - S_{VA_2})
\\+(S_{A_2B_2} + 2S_{A_2VA_1} - S_{A_2B_2VA_1} - S_{VA_1}).\label{stuff}
\end{multline}
Note that inequality (\ref{stuff}) is equivalent to
\begin{align}
I_{A_1:A_2|V} + I_{B_1:B_2|A_1A_2V} \geq I_{A_1B_1:A_2B_2}.\label{stuuff}
\end{align}
Since $I_{A_1B_1:A_2B_2}=0$ by assumption, (\ref{stuuff}) is true by SSA.
Now we show that
\begin{align}
E_R(\rho_{A_1B_1}) + E_R(\rho_{A_2B_2})\geq E_R(\rho_{A_1B_1}\otimes\rho_{A_2B_2}).\nonumber
\end{align}
Let $\rho_{A_1B_1V_1}$ and $\rho_{A_2B_2V_2}$ be extensions of $\rho_{A_1B_1}$ and $\rho_{A_2B_2}$, and as we did with $E_Q$, consider the extension of $\rho_{AB}$ given by
\begin{align}
\rho_{ABV} = \rho_{A_1B_1V_1}\otimes \rho_{A_2B_2V_2},
\end{align}
with $V\equiv V_1V_2$.
Now using the fact that systems 1 and 2 are in a product state, we have
\begin{multline}
f^R(\rho_{ABV}) = S_{AB} + 2S_{AV} - S_{ABV} - S_V
\\=S_{A_1B_1} + S_{A_2B_2} + 2(S_{A_1V_1} + S_{A_2V_2})
\\- (S_{A_1B_1V_1} + S_{A_2B_2V_2}) - (S_{V_1} + S_{V_2})
\\= S_{A_1B_1} + 2S_{A_1V_1} - S_{A_1B_1V_1} - S_{V_1}
\\+ S_{A_2B_2} + 2S_{A_2V_2} - S_{A_2B_2V_2} - S_{V_2}
\\= f^{R}(\rho_{A_1B_1V_1}) + f^R(\rho_{A_2B_2V_2}).\nonumber
\end{multline}
Therefore $E_R$ is also additive.
\end{proof}
\subsection{Relationship to regularized $E_P$}
The regularized $E_P$, defined as
\begin{align}
E_P^\infty(\rho_{AB}) = \lim_{n\to\infty}\frac{1}{n}E_P(\rho_{AB}^{\otimes n}),
\end{align}
has an important operational interpretation. $E_P^\infty$ is the number of EPR pairs required to create $\rho_{AB}$ using only local operations and asymptotically vanishing communication \cite{doi:10.1063/1.1498001}. In general, $E_P^\infty$ is difficult to calculate. But there is a relationship between $E_P^\infty$ and the quantities $E_Q$ and $E_R$, which may provide a way to learn about $E_P^\infty$.
\begin{theorem}
\begin{align}
E(\rho_{AB})\leq E_P^\infty(\rho_{AB}),\nonumber
\end{align}
for $E=E_Q,E_R$.
\end{theorem}
\begin{proof}
First note that
\begin{align}
f^{Q} - f^{P} &= \frac{1}{2}(S_A + S_B - S_{AV} - S_{BV}) \nonumber
\\&= -\frac{1}{2}(S_{V|A} + S_{V|B})\nonumber
\\&\leq 0\nonumber
\end{align}
and
\begin{align}
f^{R} - f^{P} &= \frac{1}{2}(S_{AB} - S_V - S_{ABV})\nonumber
\\&= -\frac{1}{2}(S_V + S_{V|AB})\nonumber
\\&\leq 0,\nonumber
\end{align}
where both inequalities follow from WM. Therefore $E_P$ is lower bounded by both $E_Q$ and $E_R$. Additivity of $E_Q$ and $E_R$ (Thm. \ref{Add}) now allows us to write
\begin{align}
E(\rho_{AB}) = \frac{1}{n}E(\rho_{AB}^{\otimes n})\leq\frac{1}{n}E_P(\rho_{AB}^{\otimes n})~~\forall n,\nonumber
\end{align}
for $E = E_Q,E_R$. Taking the $n\to\infty$ limit gives the theorem.
\end{proof}
\section{Evaluation of $E_P$, $E_Q$ and $E_R$}
\subsection{Pure state}
For any pure state $\ket\psi_{AB}$, all extensions are of the form $\rho_{ABV} = \dyad{\psi}_{AB}\otimes\rho_V$, which makes calculation of $E_P$, $E_Q$, and $E_R$ trivial:
\begin{align}
E_P(\ket\psi_{AB}) &= \inf_{\rho_{ABV}}S_{AV} = \inf_{\rho_{ABV}}(S_A + S_V) = S_A\nonumber
\\E_Q(\ket\psi_{AB})&= \frac{1}{2}\inf_{\rho_{ABV}}(S_A + S_B + S_{AV} - S_{BV})\nonumber
\\&= S_A + \frac{1}{2}\inf_{\rho_{ABV}}(S_A + S_V - S_B - S_V)\nonumber
\\&=S_A\nonumber
\\E_R(\ket\psi_{AB})&=\frac{1}{2}\inf_{\rho_{ABV}}(S_{AB} + 2S_{AV} - S_{ABV} - S_V)\nonumber
\\&=\frac{1}{2}\inf_{\rho_{ABV}}(2(S_A + S_V) - (S_{AB} + S_V) - S_V)\nonumber
\\&=S_A.\nonumber
\end{align}
So for a pure state, $E_P = E_Q = E_R = S_A = S_B$.
\subsection{Classically correlated state}
We can also evaluate all three correlation measures for the classically correlated state
\begin{align}
\rho_{AB}^c = \sum_ip_i\dyad{ii}_{AB}.\nonumber
\end{align}
First note that an arbitrary extension of $\rho_{AB}^c$ takes the form
\begin{align}
\rho_{ABV}^c=\sum_{i,j}\sqrt{p_ip_j}\dyad{ii}{jj}\otimes\rho_V^{ij},\nonumber
\end{align}
with $\text{Tr}~\rho_V^{ij} = \delta_{ij}$.
From this we can see that $\rho_{AV}^c = \sum_ip_i\dyad i_A\otimes\rho_V^{ii}$, and $\rho_{BV}^c = \sum_ip_i\dyad i_B\otimes\rho_V^{ii}$.
\subsubsection{$E_P$} Since $S_{V|A} = \Sigma_ip_iS(\rho_V^{ii})\geq 0,$ we have that $f^P = S_A + S_{V|A}\geq S_A$, which is saturated by the trivial extension. Therefore $E_P(\rho_{AB}^c) = S_A$.
\subsubsection{$E_Q$} From the form of $\rho_{AV}^c$ and $\rho_{BV}^c$, we see that $S_{AV} = S_{BV}$ for any extension of $\rho_{AB}^c$, so $f^{Q} = \frac{1}{2}(S_A + S_B + S_{AV} - S_{BV}) = S_A$, so $E_Q(\rho_{AB}^c) = S_A = H(\{p_i\})$.
\subsubsection{$E_R$} By (\ref{bounds}), $E_R\geq \frac{1}{2}I_{A:B} = \frac{1}{2}H(\{p_i\})$ for the state $\rho_{AB}^c$. It is easy to check that this value is achieved by the extension
\begin{align}
\rho_{ABV}^c = \sum_ip_i\dyad{iii}_{ABV},\nonumber
\end{align}
so $E_R(\rho_{AB}^c) = \frac{1}{2}I_{A:B} = \frac{1}{2}H(\{p_i\})$.
\subsection{Symmetric or antisymmetric state}
For states with support entirely within the symmetric or antisymmetric subspace, we have
\subsubsection{$E_P$} In this case, \cite{christandl2005uncertainty} showed that $E_P = S_A$ and that $E_P$ is additive.
\subsubsection{$E_Q$} States $\rho_{ABV}$ with the reduced state on $\rho_{AB}$ supported within the symmetric or antisymmetric subspace are invariant under the swap operator $F_{AB} = \sum_{ij}\dyad{ij}{ji}_{AB}$, i.e.
\begin{align}
\rho_{BAV} = (F_{AB}\otimes\mathbb{I}_V)\rho_{ABV}(F_{AB}^\dagger\otimes\mathbb{I}_V) = \rho_{ABV}.\nonumber
\end{align}
To see this, note that the most general pure state $\ket\psi_{ABV}$ for which $\text{Tr}_V\dyad\psi_{ABV}$ is supported entirely in the symmetric or antisymmetric subspace of $\mathcal{H}_A\otimes\mathcal{H}_B$ is of the form
\begin{align}
\ket\psi_{ABV} = \sum_i\ket{\phi_i}_{AB}\ket{\xi_i}_C,\label{symm}
\end{align}
where all $\ket{\phi_i}_{AB}$ are symmetric or all $\ket{\phi_i}_{AB}$ are antisymmetric. In the former case we have
\begin{align}
F_{AB}\ket\psi_{ABV} = \sum_i\left(F_{AB}\ket{\phi_i}_{AB}\right)\ket{\xi_i}_C = \ket\psi_{ABV},\nonumber
\end{align}
while in the latter case we have
\begin{align}
F_{AB}\ket\psi_{ABV} = \sum_i\left(F_{AB}\ket{\phi_i}_{AB}\right)\ket{\xi_i}_C = -\ket\psi_{ABV}.\nonumber
\end{align}
In both cases, we see that
\begin{align}
F_{AB}\dyad\psi_{ABV}F_{AB}^\dagger = \dyad\psi_{ABV}.\nonumber
\end{align}
Since $\rho_{ABV}$ with symmetric or antisymmetric $\rho_{AB}$ is generally a mixture of states of the form (\ref{symm}), invariance under $F_{AB}$ follows. In particular, this means that for any extension $\rho_{ABV}$ of a symmetric or antisymmetric state $\rho_{AB}$, we have $S_{AV} = S_{BV}$. Therefore $f^{Q} = S_A$, so $E_Q(\rho_{AB}) = S_A$.
\subsubsection{$E_R$} Again, for any extension $\rho_{ABV}$, we have $S_{AV} = S_{BV}$, so
\begin{align}
f^{R} &= \frac{1}{2}(S_{AB} + (S_{AV} + S_{BV}) - S_V - S_{ABV})\nonumber
\\&= \frac{1}{2}(S_{AB} + I_{A:B|V}).\nonumber
\end{align}
Therefore $E_R(\rho_{AB}) = \frac{1}{2}S_{AB} + E_{sq}(\rho_{AB})$.
\section{Conclusion}
We have identified four quantities of the form (\ref{form2}) which have the properties of monotonicity and the lower and upper bounds in (\ref{bounds}). Two of them, the entanglement of purification and the squashed entanglement, have appeared in the literature \cite{doi:10.1063/1.1498001}\cite{doi:10.1063/1.1666274}\cite{doi:10.1063/1.1643788} and have been thoroughly studied. The other two, which we have called $E_Q$ and $E_R$, are new. We have shown that these two quantities are additive. We have also shown that they are lower bounds for $E_P^\infty$, which could potentially provide a calculational handle for $E_P^\infty$.
It should be noted that the search method used in this paper is not exhaustive. In particular, we are restricted to showing monotonicity by (\ref{ineq0}) and (\ref{ineq1}), while there may be other inequalities which imply monotonicity of a optimized correlation measure. A possible future research thrust is identifying such new correlation measures, or confirming their non-existence. Additionally, we have not yet found operational interpretations for the correlation measures $E_Q$ and $E_R$ identified in this paper. The examples presented in Section VI suggest that while $E_Q$ and $E_R$ capture both classical and quantum correlation, $E_R$ distinguishes between the two, whereas $E_Q$ seems not to tell them apart. Finally, we have considered only bipartite correlations in this work, but we expect that a search for optimized multipartite correlation measures may yield new and exciting formulas to study and interpret. Understanding the two new correlation measures identified here, as well as other possibly new optimized correlation measures (both bipartite and multipartite) will be a step forward towards the broader goal of understanding the structures of quantum correlations.
\section*{Acknowledgment}
We would like to thank Mohammad Alhejji, Michael Perlin, and the anonymous referees for useful conversations and insightful comments. This work was supported by NSF CAREER award CCF 1652560.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
|
\section{Introduction}\label{sec:intro}
The detection of binary black hole (BH) and binary neutron star (NS) mergers via gravitational waves (GWs)\citep[e.g.,][]{2019arXiv190201557T,2016PhRvL.116f1102A,2018arXiv181112940T,2017PhRvL.119p1101A} has opened a new window to explore the Universe. A lot of interest has been devoted to understanding the physical origins of these events, and in particular the formation scenarios of the binaries which merge. The most popular hypotheses
involve field binaries \citep[e.g.,][]{2003MNRAS.341..385P, 2013ApJ...779...72D, 2014ApJ...789..120B, 2016Natur.534..512B, 2016A&A...588A..50M, 2016MNRAS.458.2634M} and dynamical formation \citep[e.g.,][]{2000ApJ...528L..17P,2006ApJ...637..937O, 2009ApJ...692..917M, 2013MNRAS.429.2298M, 2014MNRAS.444...29L, 2016ApJ...824L...8R, 2018ApJ...866L...5R, 2016ApJ...816...65A, 2017ApJ...834...68C, 2018PhRvD..97j3014S, 2018MNRAS.481.5445S, 2018MNRAS.478.4030G, 2019MNRAS.486.5008A, 2018MNRAS.481.5123B, 2018MNRAS.473..909B, 2018PhRvL.121p1103F, 2018arXiv181110627F, 2019arXiv190100863D,Perna2019}.
Once binaries are formed, their subsequent evolution up to merger, and hence the locations of their mergers, are also of great importance. Compact object (CO) binaries located in the vicinity of a supermassive black hole (SMBH) are subject to Lidov-Kozai (LK) oscillations, which accelerate their merger times, thus enhancing their merger rates \citep[e.g.,][]{2017ApJ...841...77A, 2017ApJ...836L..26C, 2016ApJ...831..187A, 2012ApJ...757...27A, 2002ApJ...578..775B, 2018ApJ...856..140H, 2017ApJ...836...39S, 2016ApJ...828...77V, 2016ARA&A..54..441N, 2003ApJ...598..419W}. This channel of CO mergers has been shown to have rates which are competitive with other proposed channels, due to the combination of the large star density in the vicinity of the central SMBH, and the shorter merger times resulting from dynamical interactions.
The phenomenology associated with dynamical interactions becomes richer when two SMBHs are paired together, making up an SMBH binary (SMBHB). In this case, it has been shown that, in addition to Hypervelocity stars (or CO), also Hypervelocity binaries (HVBs) can be produced by the dynamical interaction between the CO binary and the SMBHB \citep[][]{2007ApJ...666L..89L,2009MNRAS.392L..31S,2018MNRAS.475.4595W,2019MNRAS.482.3206W,Coughlin2018}. Therefore, detection of HVBs can serve as an unique diagnostic for the presence of a secondary SMBH in galactic nuclei.
Besides ejection, another outcome of strong dynamics around SMBHs is a stellar tidal disruption event \citep[TDE, e.g.,][]{1988Natur.333..523R}. TDEs can also be used to probe SMBHBs in many different ways \citep[e.g.,][]{2017MNRAS.469.2042B,2017MNRAS.471L.115C,2018MNRAS.477.4009D}.
TD rates have been found to be particularly enhanced by unequal SMBHBs, with mass ratios $<0.1$ \citep{2011ApJ...729...13C,2018arXiv181101960T}. Moreover, since main sequence stars are generally disrupted only by SMBHs with masses below $10^8M_\odot$ (more massive black holes just swallow the whole star without tearing it apart), a TDE flare in a galaxy expected to host a billion solar mass SMBH would be indicative of the presence of a secondary, much lighter, SMBH companion, as proposed by \citet{2018MNRAS.479.3181F}. Interestingly, few peculiar transient events have been interpreted as candidate TDEs from low mass ratio SMBHB binaries \citep{2009ApJ...706L.133L,2018MNRAS.474.3857C}. Conversely, the HVB mechanism quantified by \citet{2018MNRAS.475.4595W} operates most efficiently at SMBHB mass ratios close to unity. Hence, the combination of these two diagnostics cover a significant fraction of the total parameter space available to SMBHBs. The identification of traditional, main sequence star HVBs is limited to the Milky Way and the local group. On the other hand, due to their compactness, the parameter space available to HVBs is much higher for degenerate stars than for main-sequence stars, since binaries containing remnants, especially neutron stars (NSs) and black holes (BHs), can exist at smaller orbital separations. It is therefore interesting to model the generation of CO-HVB ejection from SMBHBs and their potential observable features, which is one of the main focuses of this work.
The observable phenomenology associated with a merger event in the electromagnetic (EM) domain depends on the type of objects making up the CO binary.
Mergers of two NSs have been confirmed to be associated with short gamma ray bursts (GRBs) \citep{Abbott2017}, and it is expected that this should also be the case for mergers of NS-BH binaries for small mass ratios (see e.g. the review in \citealt{Bartos2013}). BH-BH mergers are most likely to be EM-quiet, albeit ideas for accompanying radiation have been proposed \citep{Perna2016,Loeb2016,Zhang2016,Woosley2016,DeMink2017,Bartos2017,Liebling2016,Murase2016,Janiuk2017,Fraschetti2018}.
The location of the merger site has long been considered as an important diagnostic of the progenitor type. For CO mergers, the predictions are of a wide distribution of distances from the galaxy hosts, primarily reflecting the natal kicks, the merger times, and the mass of the host galaxy (e.g. \citealt{Belczynski2006,Oshau2017,Perna2018}). For all three types of CO binaries (i.e. NS-NS, NS-BH and BH-BH), a small fraction of events at $\sim$~Mpc scale is expected only in small galaxies, with masses $\lesssim 10^{10}M_\odot$
(e.g. \citealt{Perna2002}), and hence yielding a small contribution to the total merger rates.
For larger, Milky-Way type of galaxies, the gravitational potential of the galaxy prevents very large distances to be reached before mergers. Therefore, the production of HVBs presents the special possibility that some fraction of short GRBs may occur in isolated regions of the Universe between galaxies, or in the intergalactic medium (IGM), and not be directly associated with any particular host galaxy. These isolated short GRBs could thus indicate the presence of an SMBH-IMBH/SMBH binary hosted in the nearest galaxy or galaxy group/cluster.
Identifying these off-host GRB events as a function of redshift opens up the possibility of using them to constrain the SMBHB fraction in galaxies and how it changes as a function of cosmic time. The sophisticated $N$-body simulations performed by \citet{2019MNRAS.482.3206W} revealed how the post-ejection properties of the HVBs depend on the properties of the ejecting SMBHB. Thus, combining observations of the electromagnetic properties of off-host GRBs with gravitational wave detection -- specifically the information extracted from the waveform data -- could further constrain the properties of the SMBHBs as well as the environment in their immediate vicinity. Moreover, if Type Ia supernovae result from mergers of two WDs \citep[i.e.,][]{1984ApJS...54..335I}, the detection of 'orphan' (i.e. in the IGM, away from their host galaxies) Type Ia SNe, could provide another powerful probe of SMBH binarity. We will in fact see that hypervelocity WD-WD binaries are also expected to be produced by dynamical interactions with SMBHBs.
In this paper, we combine the cosmic SMBHB population extracted from a large scale cosmological simulation \citep[The Millennium-II,][]{2009MNRAS.398.1150B} with high-resolution, N-body simulations of the interaction of CO binaries with those SMBHBs, to predict the cosmological distribution of CO binaries, and their properties (including radial separation from their host galaxies) at the time of merger.
Additionally, we compute the cosmological merger rate of these events, and their contribution to the GW signal, and discuss our results in the context of present and future GW observations and accompanying electromagnetic counterparts.
The paper is organized as follows. In \sect{sec:SMBHBformation}, we discuss the cosmic SMBHB formation rate and cusp erosion process that provides the ideal environment to produce various types of CO merger events. In \sect{sec:relfraction} we perform the scattering experiments in the reference frame of the SMBHB to obtain the relative fraction of different types of merger events and their dependences on initial parameters, and discuss the properties of the outcomes. In \sect{sec:cosmic}, we combine the outcomes of the scattering experiments and SMBHB formation rate to estimate the cosmic rates of different types of merger events triggered by the SMBHBs. Our conclusions are summarized in \sect{sec:summary}.
\section{SMBHB formation and cusp erosion}\label{sec:SMBHBformation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/SMBHB_formation}
\caption{ The observed SMBHB formation rate derived from the Millennium II simulation as a function of the primary SMBH mass $M_1$, the binary mass ratio $q$, the redshift $z$ and the binary semi-major axis $a_\mathrm{bh}$. These distributions are used in our simulations to sample the parameters of the SMBHBs.}
\label{fig:GammaSMBHB}
\end{figure*}
\subsection{SMBHB formation rate}
{ The cosmic population of SMBHBs is derived using the data from Millennium-II \citep[e.g.,][]{2009MNRAS.398.1150B}, which is a large scale N-body simulation with sufficient resolution to distinguish dark matter halos down to $\sim10^8M_\odot$. On top of this dark matter backbone, semianalytic galaxy evolution models have been included to reconstruct the cosmic history of galaxies down to $\approx10^6M_\odot$, hosting central black holes with masses down to $\sim 10^4M_\odot$. In particular, data are taken from the semianalytic model developed in \citet[][]{2011MNRAS.413..101G}, following the procedure described in \citet[][]{2018arXiv181101960T}. For consistency, all the post processing is performed using the same cosmology as in Millennium-II: $h=H_0/(100\,\mathrm{km\,s^{-1}\,Mpc^{-1}})=0.73$, $\Omega_\Lambda=0.75$ and $\Omega_M=0.25$. Although these values of the cosmological parameters are outdated, a $\sim$ few percent level change in $h$, $\Omega_\Lambda$ or $\Omega_M$ does not have a significant impact on our results.}
We select all mergers involving two galaxies both hosting an SMBH, for a total of $N_\bullet=169435$ events across the history of the Millennium-II simulation. We then construct a 3D grid by logarithmically binning $\log_{10}(M_1)$ and the mass ratio $q$ within the ranges $M_1\in[10^5,10^{10}]M_\odot$ and $q\in[10^{-4},1]$, and by evenly binning the redshift $z$ from 0 to 5 according to the redshift of each Millennium-II snapshot. The redshift of each merger is identified with the redshift at which the progenitor galaxy is identified in the simulation. In doing this, we are implicitly assuming that the timescale for dynamical friction (DF) to bring the two SMBHs together in a relatively close binary following the galaxy merger is shorter than the typical time elapsed between two snapshots of the simulation, which is $\approx 300$ Myr. Although the DF timescales might be longer \citep{1987gady.book.....B}, we ignore this complication in this work.
Dividing the SMBHB number in each bin by the co-moving volume $V_c$ of the Millennium-II simulation, we get an approximation of the differential SMBHB formation rate,
\begin{equation*}
\frac{\mathrm{d}^4 N_\bullet}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} V_c}\,.
\label{eq:SMBHBdens}
\end{equation*}
This quantity is simply the differential number density of SMBHBs per unit redshift, logarithmic mass and mass ratio. If we now multiply it by the standard comoving volume shell ${\mathrm{d} V_c}/{\mathrm{d} z}$ and convert redshift into time by multiplying ${\mathrm{d} z}/{\mathrm{d} t_r}\times{\mathrm{d} t_r}/{\mathrm{d} t}$ with $t_r=t/(1+z)$, we finally obtain the differential SMBHB formation rate observed on Earth,
\begin{equation}
\gamma_\mathrm{\bullet}(z,\log q,\log M_1) = \frac{\mathrm{d}^4 N_\bullet}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} t}\,.
\end{equation}
\fig{fig:GammaSMBHB} shows the distribution of the SMBHB formation rate $\gamma_\mathrm{\bullet}(z,\log q,\log M_1)$ in a 2-D plane by integrating over each single variable $z$, $\log M_1$ and $\log q$ separately.
The SMBHB formation rate as observed on Earth is dominated by systems with $M<10^6$M$_\odot$, with mass ratios broadly distributed around $0.1$, reaching out to $z\approx 4$.
\subsection{The SMBHB-bound binary interaction rate}
\label{sec:intrate}
If we take a snapshot of the sky, the differential number of SMBHBs that are actively interacting with the surrounding population of bound stars is given by the differential SMBHB formation rate $\gamma_\mathrm{\bullet}(z,\log q,\log M_1)$ times the duty cycle $D_\mathrm{erd}(z,\log q,\log M_1)$ that the SMBHB takes to erode the cusp of stars bound to $M_1$. Note that $D_\mathrm{erd}$ is generally a function of $z,\log q,\log M_1$. During this duty cycle, the SMBHB interacts with a mass $M_\mathrm{erd}(z,\log q,\log M_1)$, so that the rate at which the SMBHB interacts with bound stars is simply $\dot{M}_{\mathrm{erd}}=M_\mathrm{erd}/D_\mathrm{erd}$ (where we omitted the parametric dependencies). The differential interaction rate is therefore given by the differential number of active SMBHBs times the rate at which those binaries interact with ambient stars:
\begin{eqnarray}
\dot{\mathcal{M}}_\mathrm{act}(z,\log q,\log M_1)&=&\frac{\mathrm{d}^4 M_\mathrm{act}}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} t}\\
&=&\gamma_\mathrm{\bullet} D_{\mathrm{erd}} \dot{M}_{\mathrm{erd}}\\
&=& \frac{\mathrm{d}^4 N_\bullet}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} t}{M}_{\mathrm{erd}}\,.
\label{eq:Mactive}
\end{eqnarray}
Therefore, to know the differential interaction rate, we simply need to estimate the total mass interacting with the binary, ${M}_{\mathrm{erd}}$. Numerical simulations of unequal mass SMBHBs embedded in stellar cusps have shown that the secondary hole efficiently inspirals down to a separation at which the mass in stars enclosed within its orbit around $M_1$ is of the order of $\approx 2M_2$, and it does this without significantly affecting the stellar distribution \citep[e.g.,][]{2007ApJ...656..879M}. At that point, further shrinking of the orbit proceeds via efficient ejection of bound stars, thus eroding the cusp \citep[e.g.,][]{2008ApJ...686..432S}. Pending numerical factors of order unity, the number of interacting stars can therefore be approximated as $M_\mathrm{erd}\simM_2=qM_1$.
The specific { mass interaction rate} that is observed on Earth up to redshift $z$ can then be computed by integrating $\gamma_\mathrm{erd}(z,\log q,\log M_1)$ from 0 to $z$,
\begin{equation}
\Lambda(z,\log q,\log M_1) = \int_0^z\dot{\mathcal{M}}_\mathrm{act}(z^\prime,\log q,\log M_1) \mathrm{d} z^\prime\,.
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/most_interact}
\caption{Contour plots of the mass in CO binaries that is actively interacting with a SMBHB as a function of SMBHB primary mass and mass ratio. The four panels represent this quantity integrated up to different redshifts $z=0.5$, $z=1$, $z=1.5$ and $z=2$.}
\label{fig:most_interact}
\end{figure*}
\fig{fig:most_interact} shows $\Lambda(z,\log q,\log M_1)$ at different redshifts: $z=0.5$, $z=1.0$, $z=1.5$ and $z=2.0$. This figure indicates that for relatively low redshifts, most of the CO binaries that are actively interacting with SMBHBs correspond to high mass ratio (0.1-1) SMBHBs with primary masses $M_1$ between $10^6\,M_\odot$ and $10^9\,M_\odot$. { Integrating the contours in the figure yields that about $10^6\,M_\odot$ of stars and COs actively interacting with SMBHBs out to $z=2$ is potentially observable from Earth every year.}
In order to set up the scattering experiments in \sect{sec:relfraction}, we need an estimate of the SMBHB separation at which stars are efficiently ejected from the cusp. This corresponds to $a_0$ so that $M_*(r<a_0)=2M_2$. For simplicity, following \citet[][]{2010ApJ...719..851S}, we assume the stellar density profile to follow an isothermal sphere. With this assumption, and by using the $M-\sigma$ relation from \citet[][]{2000ApJ...539L...9F}, we get
\begin{equation}\label{eq:a0}
a_0\sim 0.8\bigg(\frac{q}{1+q}\bigg)\bigg(\frac{M_1}{M_6}\bigg)^{1/2}~\mathrm{pc}\,.
\end{equation}
\section{Scattering experiments in the reference frame of the SMBHB}\label{sec:relfraction}
Not all actively interacting CO binaries will produce outcomes of interest to us, in particular hypervelocity binaries that escape their host galaxies and merge in the intergalactic medium. Therefore, scattering experiments between different types of CO binaries with SMBHBs with different $M_1$ and $q$ values are needed to obtain good statistics in the relevant regions of parameter space.
\subsection{Numerical setup}
We perform extensive scattering experiments using our high-precision code {SpaceHub} (Yihan 2019 in prep.; see also Yihan et al. 2019 for more details), which employs the {\tt ARCHAIN} algorithm \cite[e.g.,][]{1993CeMDA..57..439M} to accurately trace the motion of tight binaries with arbitrarily large mass ratios and eccentricities. The original chain structure in the algorithm, combined with the Kahan summation introduced by us to the regularized code, significantly reduces the round-off errors during close encounters.
\begin{figure}
\center
\includegraphics[width=0.8\columnwidth]{figs/config}
\caption{Schematic illustration of the stellar binary orbiting around the super-massive black hole binary (not to scale).}
\label{fig:config}
\end{figure}
In order to cover the relevant SMBHB masses and mass ratios, as indicated by \fig{fig:GammaSMBHB} and \fig{fig:most_interact}, the $q$ and $M_1$ values of the SMBHBs are logarithmically generated in the range $[10^{-4},1]$ and $[10^5,10^{10}]\,M_\odot$, respectively. The semi-major axis $a_{\bh}$ of the SMBHB is set to $a_0/3$, where $a_0$ is the active ejection radius and the eccentricity of the SMBHB $e_{\bh}$ is uniformly distributed in the range $[0,1]$. The choice of $a_0/3$ is motivated by the fact that N-body simulations and scattering experiments have shown that the whole cusp erosion shrinks the semi-major axis of the SMBHB by roughly a factor of 10 \citep[e.g.,][]{2008ApJ...686..432S,2011MNRAS.415L..35S} starting from $a_0$ given by \eqn{eq:a0}. Therefore, our scattering experiments are performed at a specific semi-major axis of the SMBHB $a_{\bh} = a_0/3$, which is the geometric average of $a_0$ and $a_0/10$. \fig{fig:config} shows the configuration of our scattering experiments and corresponding orbital parameters.
We explore five types of stellar mass binaries: WD-WD, BH-BH, NS-NS, BH-NS and EBBH (eccentric binary BHs). {250k scattering experiments are performed for BH-BH, NS-NS and BH-NS binaries while 500k scattering experiments are performed for WD-WD binaries and EBBHs.}
\subsubsection{WD-WD binary population}
For the WD-WD binaries, we take the masses $m_1$ and $m_2$ of each WD from a normal distribution with $\langle M_{WD}\rangle=0.619 M_\odot$ and $\sigma_m = 0.108 M_\odot$ \citep[e.g.,][]{Tremblay2016}. The separation $a$ of the WD-WD binary is taken from the distribution \citep[e.g.][]{Maoz2017}
\begin{equation}
N(x)\propto x^3 \ln ( 1 + x^{-4})\,,
\end{equation}
where
\begin{equation}
x=\frac{a}{(Kt_0)^{1/4}}
\end{equation}
with
\begin{equation}
K = \frac{256}{5}\frac{G^3}{c^5}m_1m_2(m_1+m_2)\,,
\end{equation}
and $t_0$ is the age of the host galaxy. This is evidently a time-dependent model in which high redshift WD-WD binaries would be more compact. However, for simplicity in our simulations, we take $t_0 = 13.6~{\rm Gyr}$ throughout. The eccentricity $e$ of the WD-WD binaries is drawn from a uniform distribution within $[0, 1]$ \citep{Geller2019}.
The radius of the WD is set to be $0.01R_\odot({m_\odot}/{m_\mathrm{wd}})^{1/3}$.
\subsubsection{BH-BH, NS-NS and BH-NS binary populations}
For the BH-BH, NS-NS and BH-NS binaries, we take the distributions for $m_{\star 1},\, m_{\star 2},\,a_{\star}$ and $e_{\star}$ from the BSE code \href{https://www.syntheticuniverse.org/}{\tt StarTrack} \citep{Belczynski2008} with metallicity $Z=0.1Z_\odot$.
The BH radii are set equal to their Schwarzschild radii, while the radii of the NSs are set to be 10 $\mathrm{km}$.
\subsubsection{Eccentric BH-BH binary population}
The properties of the CO binary populations in galactic nuclei are essentially unknown. \citet{Gond2018} suggested that eccentric BH-BH binaries can form via dynamical capture. Therefore, we perform an additional set of simulations with eccentric binary BHs (labelled EBBHs). Following \citet{Gond2018}, the masses of the BHs are drawn from a Salpeter initial mass function (IMF), $N(m) \propto m^{-2.35}$ \citep[e.g.,][]{1955ApJ...121..161S}. The separation is drawn from a logarithmically uniform distribution where $a_{\rm min}=0.01\,\mathrm{au}$ and $a_{\rm max}=1000\,\mathrm{au}$. The circularity $c=1- e^2$ also obeys a logarithmically uniform distribution, where $c_{\rm min} = 10^{-6}$ and $c_{\rm max} = 1$.\\\\
\begin{figure*}
\includegraphics[width=\textwidth]{figs/m1m2_pop}
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/ae_pop}
\caption{Distribution of the compact object binary population in the $m_{\star 1}$-$m_{\star 2}$ (top series of panels) and $a_{\star}$-$e_{\star}$ (bottom series of panels) plane. In each panel, the main plot shows the 2D probability density distribution (PDF), whereas the side plots show the 1D distribution marginalized over the other quantity. \textit{Top Left} panel: BH-BH; \textit{Top middle} panel: NS-NS; \textit{Top right} panel: BH-NS; \textit{Bottom left} panel: WD-WD; \textit{Bottom right} panel: EBBH. }
\label{fig:COb-pop}
\end{figure*}
\fig{fig:COb-pop} shows the parameters of the five types of CO binaries described above. For the BH-BH case, almost all binaries have circular orbits. Conversely, NS-NS binaries tend to be eccentric, and NS-BH binaries tend to be uniformly distributed in eccentricity. The semi-major axes of the BH-BH, NS-NS and BH-NS binaries are mostly clustered in the $10^{-2}$AU-1AU range, thus forming a population of relatively tight binaries.
The orbits of the centre of mass (CoM) of the CO binaries are created such that $a_{\cm}$ is isothermally distributed (i.e., $p(a)\propto a^{-2}$) around the primary SMBHB in the range of $[R_t, 10a_{\bh}]$, where
\begin{equation}
R_t = 3.7\bigg( \frac{M_1}{m_{\star 1}+m_{\star 2}}\bigg)^{1/3}\frac{ 1+e_{\star}}{1-e_{\cm}}a_{\star}
\end{equation}
is the minimum semimajor axis of the stellar binary CoM around $M_1$ ensuring that the stellar binary will not be tidally disrupted at its apocentre when its CoM orbit approaches pericentre around $M_1$.
All angles and inclinations are sampled according to a spherically symmetric distribution. All anomalies are sampled uniformly in time. For each simulation, at least 50 cycles of the CoM orbit of the CO binary are first performed in isolation to ensure that the CO binaries are stable before interacting directly with the SMBHB.
\subsection{Stability Map}
\label{sec:stability}
Since stars reside deep in the potential well of $M_1$, relaxation processes due to the extended stellar distribution are negligible and the typical star-SMBHB interaction time is much shorter than the relaxation time $T_{\rm relax}$ of the surrounding stellar system. The cusp erosion can therefore be treated as a set of individual star-SMBHB interactions that can be tackled by means of 3-body scattering experiments. If a fraction of ambient stars is in fact made up by binaries of compact objects (WDs, NSs or BHs) then, during the 4-body interaction, several processes may lead to CO binary mergers.
To understand the different physics behind the merger events, we follow the stability map method in our previous work \citep[][]{2019MNRAS.482.3206W}. Briefly, the four body system can be divided into two triples: an inner triple formed by $m_{\star 1}$-$m_{\star 2}$-$M_1$, and an outer triple formed by $M_1$-$M_2$-$(m_{\star 1},m_{\star 2})$. In each triple, the system can be either stable or unstable. In stable systems, secular effects like LK oscillations may accelerate the merger process. In unstable systems, there is a significant chance of a close interaction occurring between two or more of the components of the triple. The stability of each triple is expressed in terms of the stability factor $S$. For the inner triple we have
\begin{equation}
S_{\mathrm{in}} = \log\bigg(\frac{a_{\cm}(1-e_{\cm})}{a_{\star}(1+e_{\star})} / Y_\mathrm{in,crit}\bigg)
\end{equation}
where
\begin{equation}\label{eq:crit1}
Y_\mathrm{in,crit} = \frac{3.7}{\beta_\mathrm{out}}-\frac{2.2}{1+\beta_\mathrm{out}} + \frac{1.4}{\beta_\mathrm{in}}\frac{\beta_\mathrm{out}-1}{\beta_\mathrm{out}+1}
\end{equation}
\citep[e.g.,][]{1996ASPC...90..433K} with $\beta_\mathrm{in}=\mathrm{max}\bigg(\frac{m_{\star 1}}{m_{\star 2}}, \frac{m_{\star 2}}{m_{\star 1}}\bigg)^{1/3}$, and $\beta_\mathrm{out}=\bigg(\frac{m_{\star 1}+m_{\star 2}}{M_1}\bigg)^{1/3}$.
For the outer triple system we have
\begin{equation}
S_\mathrm{out} = \log\bigg(\frac{a_{\bh}(1-e_{\bh})}{a_{\cm}(1+e_{\cm})} / Y_\mathrm{out,crit} \bigg)
\end{equation}
where,
\begin{equation}\label{eq:crit2}
Y_\mathrm{out,crit} = \frac{3.7}{\zeta_\mathrm{out}}-\frac{2.2}{1+\zeta_\mathrm{out}} + \frac{1.4}{\zeta_\mathrm{in}}\frac{\zeta_\mathrm{out}-1}{\zeta_\mathrm{out}+1}
\end{equation}
with $\zeta_\mathrm{in}=\bigg(\frac{M_1}{m_{\star 1}+m_{\star 2}}\bigg)^{1/3}$, and $\zeta_\mathrm{out}=\bigg(\frac{m_{\star 1}+m_{\star 2}+M_1}{M_2}\bigg)^{1/3}$.
If $S_\mathrm{in,out}>0$, then the corresponding triple system is stable, otherwise it is unstable and tends to be disrupted.
We can thus divide the parameter space of the four-body system into four parts, where $(S_\mathrm{in}>0,S_\mathrm{out}>0)$, $(S_\mathrm{in}<0,S_\mathrm{out}>0)$, $(S_\mathrm{in}>0,S_\mathrm{out}<0)$ and $(S_\mathrm{in}<0,S_\mathrm{out}<0)$. For stable triple systems where $S_\mathrm{in,out}>0$, hierarchical LK oscillations can occur. In such systems, the long time scale secular perturbations induced by the outer orbit can drive the eccentricity of the inner orbit to extreme values, while simultaneously decreasing the inclination between the inner and outer orbital planes and conserving angular momentum.
\subsection{Classification of mergers in SMBHBs}
We now classify the different outcomes of the SMBHB-CO interactions with the aid of the stability map introduced in the previous section and visualized in Figure \ref{fig:stability}. In general, when $S_\mathrm{in}<0$, the tidal force exerted by the central SMBH is large enough to break apart the CO binary on a short timescale. {Those systems have been excluded from the initial conditions as we discussed above. Therefore, the legend 'decoupled' in \fig{fig:stability} indicates binary disruption from interaction with the secondary massive BH (i.e., disruption directly by the secondary massive BH or by the primary massive BH but as a result of a perturbation by the secondary massive BH)}. The right part of the stability map, where $S_\mathrm{in}>0$, is where all the interesting action occurs. Here, the interaction can lead to different flavors of ejected and merging systems, as we now explore in detail.
\subsubsection{Ejected mergers}
In the stability map where $S_\mathrm{out}<0$, the outer triple is unstable, indicating that the influence of $M_2$ severely perturbs the orbit of the centre of mass of the CO binary. In fact, in this case the CO binary can be significantly affected by close encounters with $M_2$, causing its ejection from the SMBHB sphere of influence with high velocity. Due to energy and angular momentum exchange with $M_2$, ejected CO binaries can get tightened enough to merge within a Hubble time. { Depending on the ejection velocity and on the depth of the potential well of the host galaxy, these CO binaries can merge in the outskirts of the dark matter halo of their host galaxy, or even in the IGM.}
This kind of merger event has two prerequisites: the CO binaries have to be ejected by the SMBHB (to produce HVBs), and the ejected CO binaries need to be tight enough to merge in a Hubble time due to GW radiation. To produce the HVBs, their ejection by the SMBHB must be efficient. This requires that the semi-major axis of the SMBHB is at least as small as $a_0$, and that the number density of CO binaries around $M_2$ is sufficiently large. At the same time, $S_\mathrm{out}$ needs to be smaller than zero to ensure that the CO binaries have a close encounter with $M_2$, while $S_\mathrm{in}$ needs to be larger than zero to ensure that the CO binaries will not be disrupted by $M_1$. To get compact HVBs that will merge within a Hubble time, the CO binaries need either to be sufficiently hard before the ejection or sufficient transfer of energy and angular momentum must occur from the binary's orbit to its centre of mass orbit around the SMBHB. Detailed descriptions of these processes can be found in \citet{2019MNRAS.482.3206W}. The distribution of ejected merger events in the stability map is shown in \fig{fig:stability} for all the simulated CO binary species.
\subsubsection{Quick mergers (i.e., collisions)}
In the inner triple system, LK cycles can directly create collisions between the components of the compact object binary if the eccentricity becomes sufficiently high. The typical (quadrupole) timescale for such an excitation cycle is \citep[e.g.,][]{1962P&SS....9..719L,1962AJ.....67..591K},
\begin{equation}\label{eq:LK}
\begin{split}
\tau_\mathrm{LK,in} &\sim \frac{1}{n_\star}\bigg( \frac{m_{\star 1}+m_{\star 2}}{M_1}\bigg) \bigg( \frac{a_{\cm}}{a_{\star}}\bigg)^3(1-e_{\cm}^2)^{3/2}\\
&=1.3\mathrm{Myr}\bigg( \frac{M_1}{M_6}\bigg)^{-1}\bigg( \frac{m_{\star 1}+m_{\star 2}}{M_\odot}\bigg)^{1/2}\\
&\times \bigg( \frac{a_{\cm}}{0.1\mathrm{pc}}\bigg)^{3}\bigg( \frac{a_{\star}}{\mathrm{au}}\bigg)^{-3/2}(1-e_{\cm}^2)^{3/2}\,,
\end{split}
\end{equation}
where $n_\star=\sqrt{G(m_{\star 1}+m_{\star 2})/a_{\star}^3}$ is the mean motion of the compact object binary. The strength of the next higher order LK effect (i.e., the octupole effect) is described by \citep[][]{2013MNRAS.431.2155N}
\begin{equation}\label{eq:oct_LK}
\epsilon_\mathrm{oct} = \frac{|m_{\star 1}-m_{\star 2}|}{m_{\star 1}+m_{\star 2}}\frac{a_{\star}}{a_{\cm}}\frac{e_{\cm}}{1-e_{\cm}^2}\,.
\end{equation}
Since the stability criterion in the inner triple requires the ratio ${a_{\star}}/{a_{\cm}}$ to be extremely small, the CO binary will not be disrupted by the SMBHB and the octupole effect is usually negligible.
To maintain the LK oscillations in the inner triple, the perturbations from the SMBH need to be stable. This requires that the centre of mass orbit of the CO binary will not change significantly during the inner LK cycles. Thus, the outer triple system also needs to be stable such that the perturbation from $M_2$ will not break the centre of mass orbit of the CO binary. In such a system, for which $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$, we find a relatively short corresponding value for $\tau_{LK,in}$, If and hence we frequently detect collisions in the simulations. Merger events from this channel occur on a short time scale in the simulations, especially for CO binaries with component stars that have the largest radii (i.e., WD-WD binaries). Thus, throughout this paper, we call this kind of merger 'quick mergers'.
\begin{figure*}
\includegraphics[width=\textwidth]{figs/stability}
\caption{Distribution of the possible outcomes of our simulations in terms of the stability map. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH. The marginalized distributions are normalized to 1 for visualization, thus they do not reflect the true event numbers of each event set. Also note that the ejected-merger is a subset of HVBs.}
\label{fig:stability}
\end{figure*}
\fig{fig:stability} shows the quick mergers in the stability map ($S_\mathrm{in}$,\,$S_\mathrm{out}$). We can clearly see that for different CO binaries (i.e., WD-WD, BH-BH and NS-NS), almost all quick mergers occur in the upper right quadrant where $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$. As expected, for the WD-WD binaries, the distribution for $S_\mathrm{in}$ is more dispersed. This is because the WD-WD binaries need only relatively low eccentricities to merge due to their larger radii, which requires only weaker LK oscillations that correspond to larger values for $S_{in}$ (larger ${a_{\cm}}/{a_{\star}}$ in \eqn{eq:LK}). A signature of quick merger events is that they have non-negligible eccentricities in the LIGO band. This is because the CO binaries directly collide in our simulations before circularization can occur due to GW radiation. Note, however, that LK oscillations around a single SMBH can also lead to eccentric LIGO events \citep[e.g.][]{2018ApJ...860....5G}; therefore, eccentric mergers should not be considered signposts of SMBHBs.
\subsubsection{Slow mergers}
Some CO binaries neither experience ejections by $M_2$ nor are subject to strong LK cycles. {This occurs for a subset of systems located within the stable region in our stability map} where $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$. Alternatively, they could remain bound to the SMBHB after a close encounter with $M_2$ in the region where $S_\mathrm{in}>0$ and $S_\mathrm{out}<0$. CO binaries in the region where $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$ which experience negligible LK effects are in the ideal environment to allow for binary stellar evolution (BSE) to occur. Thus, some tight CO binaries could form from BSE via common envelope evolution (for example), and then merge due to GW radiation within {the cosmological time elapsed since their formation, i.e. the lookback time $\tau_{\rm lb}$}. CO binaries in the region where $S_\mathrm{in}>0$ and $S_\mathrm{out}<0$ which remain bound to the SMBHB after a close encounter with the secondary $M_2$ tend to become harder post-encounter. Therefore, a fraction of CO binaries in this region of phase space will be pushed into the merger band of LIGO by close encounters with $M_2$.
Despite the specific formation channel (BSE or encounter) for the tight CO binaries observed in the galactic centre, our simulations suggest a high formation rate if an IMBH is indeed present. Unlike quick mergers that are directly detected in the simulations, the 'slow' merger events will undergo significant circularization before merger due to GW radiation. Therefore, the eccentricity of the merger events for this channel will be completely negligible once in the LIGO band. Slow merger events are also shown in the stability map of \fig{fig:stability} (denoted with "S-merger").
\subsection{Outcomes of Scattering experiments}
\subsubsection{Relative fraction of HVBs}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/TV-distribution}
\caption{Contour plots of the 2D PDF $\int f_{\mathrm{hvb}}(m_\mathrm{chirp},v_\mathrm{ej},\tau) \mathrm{d} m_\mathrm{chirp}$ of HVBs, i.e. the normalized differential distribution of the HVBs as a function of $v_\mathrm{ej}$ and $\tau$. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:TV-dist}
\end{figure*}
From our scattering experiments, {for each value of} $M_1$ and $q$, the distribution of HVBs can be expressed as
\begin{equation*}
\frac{\mathrm{d}^4 N_\mathrm{hvb}}{\mathrm{d}a_{\star}\mathrm{d}e_{\star}\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej}}\,,
\end{equation*}
where $a_{\star}$, $e_{\star}$, $m_\mathrm{chirp}$ and $v_\mathrm{ej}$ are the semi-major axis, eccentricity, chirp mass and ejection velocity of the HVB, respectively. All these quantities are directly obtained from the scattering experiments.
After the ejection, due to GW radiation, some HVBs will merge within an Hubble time. The merger time scale due to GW radiation is given by \citep[][]{1964PhRv..136.1224P}
\begin{equation}
\tau_\mathrm{GW}(m_\mathrm{chirp},a_{\star},e_{\star}) = \frac{3}{85}\frac{c^5a_{\star}^4(1-e_{\star}^2)^{7/2}}{G^3m_{\star 1}m_{\star 2}(m_{\star 1}+m_{\star 2})}\,.
\end{equation}
Since not all HVBs will merge within {$\tau_{\rm lb}$}, we need to select only those HVBs whose merger time $\tau_\mathrm{GW}$ is shorter than {$\tau_{\rm lb}$}. Previous works \citep[][]{2019MNRAS.482.3206W} have shown that from the close interaction with the SMBHB, many of the ejected binaries tend to become harder after the ejection due to energy and angular momentum exchange between the IMBH and the compact binary. Thus, we expect a significant fraction of the ejected HVBs to rapidly escape the SMBHBs and eventually merge (within $\tau_{\rm lb}$) in the dark matter halo of their host galaxy or even in the IGM. To compute the distribution of merger times due to GW emission, $\tau_\mathrm{GW}$, we have to convert the distribution of HVBs to the $(m_\mathrm{chirp},v_\mathrm{ej},\tau)$-space by integrating over the parameter $\tau$,
\begin{equation}
\begin{split}
&f_{\mathrm{hvb}}(m_\mathrm{chirp},v_\mathrm{ej},\tau)=\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej} \mathrm{d} \tau}\\
&=\frac{\mathrm{d}}{\mathrm{d}\tau}\oiint_{\tau_\mathrm{GW}(m_\mathrm{chirp},a_{\star},e_{\star})<\tau}\frac{\mathrm{d}^4 N_\mathrm{hvb}}{\mathrm{d}a_{\star}\mathrm{d}e_{\star}\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej}}\mathrm{d}a_{\star}\mathrm{d}e_{\star}\,.\\
\end{split}
\end{equation}
Figure \ref{fig:TV-dist} shows $f_{\mathrm{hvb}}(m_\mathrm{chirp},v_\mathrm{ej},\tau)$,
which is the 2D-PDF of the hypervelocity binaries as a function of the ejection velocity $v_\mathrm{ej}$ and the GW merger time $\tau_{\rm GW}$. Clearly, for each type of CO binary, we get some HVBs for which the final velocity is high enough relative to the local escape speed that these objects are expected to escape from the sphere of influence of their host SMBHBs. After applying a specific galactic potential, one can compute whether or not a given binary will merge within the dark matter (DM) halo of their host galaxy or the IGM. { Detailed calculations regarding the cosmic rate of this type of merger will be discussed in \sect{sec:cosmic}.}
\subsubsection{Relative fraction of quick merger events}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/q-merger_mq}
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/q-merger_ae_cm}
\caption{The 2D PDF of the relative fraction of quick mergers in the $M_1$-$q$ (top series of panels) and $a_{\cm}/a_{\bh}$-$a_{\star}/a_{\cm}$ (bottom series of panels) planes. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:quick_rel}
\end{figure*}
Similarly, we can examine the properties of quick merger events. Those are shown in \fig{fig:quick_rel}. The upper panels show, for the various types of CO binaries, the distribution
\begin{equation*}
\frac{1}{N_\mathrm{quick}}\frac{\mathrm{d}^2 N_\mathrm{quick}}{\mathrm{d}\log M_1\mathrm{d}\log q}\,,
\end{equation*}
{which indicates that the relative rate of quick mergers is independent of the mass ratio of the SMBHB and has a large probability for SMBHBs with large $M_1$.}
{Indeed, the quick merger rate depends on the time scale of the LK oscillations from the primary massive BH which is more sensitive to the location of the CO binaries in the SMBHB potential well ($a_{\cm}/a_{\bh}$), as well as the mass of $M_1$.} The bottom panels show the distribution
\begin{equation*}
\frac{1}{N_\mathrm{quick}}\frac{\mathrm{d}^2 N_\mathrm{quick}}{\mathrm{d}e_{\cm}\mathrm{d}\log a_{\cm}}\,.
\end{equation*}
When BH-BH, NS-NS and BH-NS binaries are involved, quick mergers tend to occur when $a_{\cm}/a_{\bh}\lesssim 10^{-1}$. To produce quick mergers, in fact, the LK-effect in the inner triple must be very efficient. The time scale for the leading quadrupole and octupole LK-effects are given by, respectively, \eqn{eq:LK} and \eqn{eq:oct_LK}. In this region, the strong LK oscillations efficiently drive the CO binary orbital eccentricity to extremely high values and cause a direct collision. The bottom two panels show that quick mergers in WD-WD binaries and EBBHs can also frequently occur when $a_{\rm cm}/a_0>1$. This is because WDs are much larger in size than either NSs or BHs, and binaries containing them can merge at lower eccentricities. EBBHs themselves are much more eccentric, and thus need less help from LK oscillations. Therefore, only a moderate LK effect is required, which can occur farther from the central SMBH.
From our previous stability analysis, we know that quick mergers correspond mostly to the region $S_\mathrm{in}~>~0$ and $S_\mathrm{out}~>~0$, for which both the inner and outer triple systems form stable hierarchical systems. In \fig{fig:stability}, we see that, as $S_\mathrm{in}$ approaches zero from above (indicating that the perturber comes closer to the compact object binary) the quick merger rate increases significantly. This is explained by the stronger inner LK effect due to a closer perturber. As $S_\mathrm{in}$ crosses zero, the perturbation from $M_1$ becomes so strong that the inner triple becomes unstable, leading to binary disruption. With that said, this is not seen in the stability map directly due to the 'hard cut' imposed in our initial condition at $S_\mathrm{in}=0$. Note that, since the WD radii are much larger than those of NSs and BHs and the EBBHs are very eccentric from the beginning, quick mergers of binary WD-WD and EBBHs require weaker LK oscillations, which is reflected by a wider dispersion in $S_\mathrm{in}$ in the bottom panels of Figure \ref{fig:stability}.
Note that, in our simulations, the dissipative 2.5PN term is not included {due to the fact that the computation becomes prohibitively expensive for very tight CO binaries}. This narrows the eccentricity window in our collision detections. Eccentricities lower than what required for collisions to occur (but still extremely high) can generate quick mergers due to the high efficiency of GW radiation at very high $e_{\star}$. These GW-induced mergers could have non-negligible eccentricities in the LIGO band, as was found in \citet[][]{2018PhRvD..97j3014S, 2012ApJ...757...27A}. Therefore, the quick merger rate obtained from our simulations corresponds to a lower limit.
\subsubsection{Relative fraction of slow merger events}
As for {ejected} mergers, we obtain the relative fraction of slow merger events as described in the following section.
From our scattering experiments, we obtain a number of unresolved interactions. After selecting those for which $\tau_\mathrm{GW} < \tau_\mathrm{hubble}$, we divide the results into two parts, namely $S_\mathrm{out}>0$ and $S_\mathrm{out}<0$. These two subsets correspond to the two different cases described above:
binaries that could be formed from BSE and undergo long term LK oscillations from $M_1$, or {tight binaries that remain bound to the SMBHB after close encounters with $M_2$}. The two subsets can clearly be seen from the lower panels of \fig{fig:incenter_rel}, which shows the 2D-PDF of the slow merger rate in the ($M_1$, $q$) and ($e_{\cm}$, $a_{\cm}$) planes, namely
\begin{equation*}
\frac{1}{N_\mathrm{slow}}\frac{\mathrm{d}^2 N_\mathrm{slow}}{\mathrm{d}\log M_1\mathrm{d}\log q}\,
\end{equation*}
and
\begin{equation*}
\frac{1}{N_\mathrm{slow}}\frac{\mathrm{d}^2 N_\mathrm{slow}}{\mathrm{d}e_{\cm}\mathrm{d}\log a_{\cm}}\,.
\end{equation*}
The upper sets of panels in \fig{fig:incenter_rel} show that slow mergers from LK oscillations rarely occur for SMBHBs with mass ratios larger than 0.1.
The two subsets of events with $S_\mathrm{out}>0$ and $S_\mathrm{out}<0$ can clearly be recognized in the bi-modal distributions of the lower set of panels. The mode peaking at $a_{\cm} \sim [0.001, 0.1] a_{\bh}$ corresponds to stable systems with $S_\mathrm{out}>0$, whereas the mode extending at $a_{\cm} > a_{\bh}$ corresponds to unstable systems with $S_\mathrm{out}<0$. Note that since EBBHs merge much faster than the other four types of CO binaries in the region $a_{\cm} \sim [0.001, 0.1] a_{\bh}$, the distribution of slow mergers of EBBHs is dominated by the mode at $a_{\cm} > a_{\bh}$.
\begin{figure*}
\includegraphics[width=\textwidth]{figs/s-merger_mq}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/s-merger_ae_cm}\\
\caption{Same as \fig{fig:quick_rel} but for the relative fraction of 'slow mergers'. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:incenter_rel}
\end{figure*}
\section{Cosmic rate of merger events from SMBHBs}\label{sec:cosmic}
{We now combine the outcomes of our scattering experiments with the SMBHB population extracted from the Millennium-II simulation, to estimate the cosmic merger rates of different types of mergers produced by SMBHB-CO binary interactions. Before proceeding, however, we shall stress an important caveat. Our simulations are meant to be representative of the dynamical interactions occurring as the SMBHB erodes the cusp of bound stars surrounding it as it first becomes bound. The cusp erosion process shrinks the binaries by a factor of $\approx 10$ starting from $a_0$, which is given by \eqn{eq:a0} if we assume that the cusp distribution follows an isothermal sphere. This might not be sufficient to efficiently drive the SMBHB to coalescence, and further shrinking, via the interaction of unbound stars scattered in the SMBHB loss cone, might be required. In general, shrinking by only a few e-folds is needed \citep[e.g.,][]{2010ApJ...719..851S} and each e-fold requires the interaction with a mass in stars of the order of $M_2$ \citep[see, e.g.,][]{2005LRR.....8....8M}. Since the erosion of the bound cusp ejects a mass in stars of order $\approx M_2$, the rates estimated in the following should be considered as conservative, and might be higher by up to a factor of a few.}
{We caution, however, that interaction with unbound binaries is inherently different than for bound binaries. Most stars, especially for SMBHB mass ratios close to unity, are either ejected in a single fly-by, or captured in relatively short-lived metastable orbits. LK oscillations are therefore irrelevant in this case, and the rate of quick and slow mergers should be strongly suppressed. On the other hand, CO binaries can still be ejected as HVBs, but the ejection velocities tend to be less extreme than those achieved in bound scattering. Unbound SMBHB-CO binary scattering experiments will eventually be needed to quantify the contribution of this late phase of the SMBHB evolution to the cosmic rate of ejected CO binary mergers. This is deferred to future work.}
\subsection{Ejected merger rate}
From the scattering experiments, we have obtained the HVB distribution ${\mathrm{d}^3 N_\mathrm{hvb}}/({\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej} \mathrm{d} \tau})$. Note that this distribution varies with $M_1$ and $q$, thus it is a function of $m_\mathrm{chirp}$, $\tau$ and $v_{\rm eject}$ but also of $\log M_1$ and $\log q$. After the ejection, a given HVB will escape from the galactic potential with initial velocity $v_\mathrm{ej}$ and merge in the dark matter halo or in the IGM at a distance $R$ away from the host galactic centre. This distance depends on the flight time $\tau$, and can be computed explicitly by assuming a specific galactic potential. In this work, we adopt the characteristic galactic potential of elliptical galaxies \citep[e.g.,][]{2008ApJ...680..312K,2014ApJ...793..122K},
\begin{equation}
\begin{split}
\phi(r)&=\phi_\mathrm{bh}(r)+\phi_b(r) + \phi_h(r)\\
&= -\frac{G(M_1+M_2)}{r} -\frac{GM_b}{r+r_b} - \frac{GM_h}{r}\ln\left(1+\frac{r}{r_h}\right)\,,
\end{split}
\label{eq:pot}
\end{equation}
where $M_b$ is the mass of the bulge, $M_h$ is the mass of the dark matter halo, $r_b$ is the scale radius of the bulge and $r_h$ is the scale radius of the dark matter halo. The choice of the spherical potential in equation \eqref{eq:pot} is dictated by computational convenience, as it allows us to perform 1D calculations to compute the ejected CO binary trajectory. The addition of a disk component would produce a deflection of the ejected trajectories, without significantly affecting the overall distribution of distances travelled. We scale the variables $M_b$, $M_h$, $r_b$ and $r_h$ with the total mass of the SMBHB $M_1(1+q)$ by means of the relations \cite[e.g.,][]{2013ARA&A..51..511K, 2015MNRAS.452.4013K}
\begin{equation}
\begin{aligned}
&\frac{M_1(q+1)}{10^9M_\odot}&=&\,0.49\left(\frac{M_b}{10^{11}M_\odot}\right)^{1.16}\,,\\
&\frac{M_h}{10^{13}M_\odot}&=&\,0.30\left(\frac{M_b}{10^{11}M_\odot}\right)\,,\\
&\frac{M_b}{10^6 M_\odot}\bigg(\frac{\mathrm{kpc}}{r_b}\bigg)^3&=&\, 4000\,,\\
&\frac{M_h}{10^6 M_\odot}\bigg(\frac{ \mathrm{kpc}}{r_h}\bigg)^3&=&\,125\,.
\end{aligned}
\end{equation}
The ejection properties are recorded at a distance of $r\,\sim\,50a_{\bh}$, given the condition that the total energy of the compact binary must be positive. Thus, we can calculate the traveled distance $R=D(M_1(1+q),\tau,v_\mathrm{ej})$
\begin{equation}
\int_0^D\frac{dr}{\sqrt{2E_0-2\phi(r)- 2l_0^2/r^2}} =\int_0^{\tau_\mathrm{GW}} dt\,,
\end{equation}
where $E_0$ and $l_0$ are, respectively, the total energy and total angular momentum per unit mass at the time of ejection. The distribution of HVBs as a function of mass, traveling time $\tau$ and distance from the galaxy centre at a merger distance $R$ can then be calculated as
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} \tau\mathrm{d} R}(\log q,\log M_1,m_\mathrm{chirp},\tau,R)\\
&=\frac{\mathrm{d}}{\mathrm{d} R}\oint_{D(M_1,q,\tau,v_\mathrm{ej}) < R}\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej}\mathrm{d} \tau}\mathrm{d} v_\mathrm{ej}\\
\end{split}\,.
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/eRV_dist}\\
\caption{{The 2D-PDF of the ejected merger events as a function of the merger distance $R$ from their host galaxy, as well as the corresponding CoM velocity when they merge}.\textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH. }
\label{fig:RV}
\end{figure*}
\fig{fig:RV} shows the 2D-PDF of the distance $R$ and the corresponding velocity at that distance of the simulated HVBs. Note that not all ejected mergers can be observed within the age of the universe.
The results of our scattering experiments yield the time at which the SMBHB ejects the CO binaries, which corresponds to the redshift $z$. Merger events occur a time $\tau$ after the ejection. Therefore, there is a significant displacement of the merger redshift with respect to $z$. Only those events with $\tau\lesssim \tau_{\rm lb}$ can be observed; let us define the merger redshift to be $z^\prime=g(z,\tau)$, where
\begin{equation}
\begin{split}
\tau &= \int_{z^\prime}^z \frac{\mathrm{d} t}{\mathrm{d} \bar{z}}\mathrm{d} \bar{z}\\
&= \frac{1}{H_0}\int_{z^\prime}^z\frac{\mathrm{d} \bar{z}}{(1+\bar{z})\sqrt{\Omega_M(1+\bar{z})^3+\Omega_k(1+\bar{z})^2+\Omega_\Lambda}}\,.\\
\end{split}
\end{equation}
We solve this equation numerically to obtain the value of $z^\prime$ for each given $z$ and $\tau$. With this relationship, we can calculate the redshift distribution of the ejected mergers. Only mergers with corresponding redshift $z^\prime > 0$ can be observed today.
To compute the cosmic ejected merger rate in the rest frame of the Earth, we need to convolve the fraction of {HVBs} obtained from the scattering experiments with the cosmic distribution of compact binaries actively interacting with SMBHBs per unit time. As discussed in Section \ref{sec:intrate}, we compute the distribution of the actively interacting mass per unit time via \eqn{eq:Mactive};
thus it is easy to obtain the cosmic distribution for the actively interacting binary number per unit time,
\begin{eqnarray}
\dot{n}_\mathrm{act}(z,\log q,\log M_1)&=&\frac{\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\\
&=&\frac{\mathrm{d}^4 (n_\mathrm{bi} M_\mathrm{act})}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\\
&=&n_{\mathrm{bi}}\dot{\mathcal{M}}_\mathrm{act}(z,\log q,\log M_1)\,,
\end{eqnarray}
where $n_\mathrm{bi}$ is the number of compact binaries (BHB, WDB, NSB) per unit mass in the galactic centre, which can be estimated via consideration of stellar evolution with a given IMF, and the binary fraction typical of galactic centres.
From our scattering experiments, {for each} value of $M_1$ and $q$, we obtain the HVB distribution ${\mathrm{d}^3 N_\mathrm{hvb}}/({\mathrm{d} m_\mathrm{chirp}\mathrm{d} \tau\mathrm{d} R})$ as a function of the parameters $m_\mathrm{chirp}$, $\tau$ and $R$. Note that this distribution varies with $M_1$ and $q$, thus it is not only a function of $m_\mathrm{chirp}$, $\tau$ and $R$ but also a function of $\log M_1$ and $\log q$. Therefore, the cosmic rate of HVBs can be calculated as
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^4N_\mathrm{eject}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t}(z^\prime, m_\mathrm{chirp}, R)\\
&=\theta(z^\prime)\frac{\mathrm{d}}{\mathrm{d} z^\prime}\int_{g(z,\tau)<z^\prime}\mathrm{d}\tau\mathrm{d} z\int\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} \tau\mathrm{d} R}\\
&\times\frac{\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\frac{1}{N_{\mathrm{exp}}(\log q,\log M_1)}\mathrm{d} \log M_1 \mathrm{d}\log q\,,\\
\end{split}
\end{equation}
where $N_{\mathrm{exp}}$ is the total number of scattering experiments performed for each $M_1$ {and $\theta(z^\prime)$ selects out the merger events that can be observed to date ($z^\prime\ge 0$).}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/emerger_RM}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/emerger_RZ}
\caption{Cosmic distribution of ejected merger events and their dependences on the parameters $R$, $m_\mathrm{chirp}$ and $z^\prime$ normalized by the undetermined parameter $n_\mathrm{bi}$, where $n_\mathrm{bi}$ is the CO binary number per unit solar mass in the galactic centre. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:cosmic_rate}
\end{figure*}
The upper panels of \fig{fig:cosmic_rate} show the distribution of the rate of cosmic '{ejected}' mergers {that can be observed up to date, i.e. ${\mathrm{d}^3N_\mathrm{eject}}/({\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t})$}. For all types of CO binaries, we find merger events occurring both in dark matter halos and in the interstellar and intergalactic media ($\unit{kpc-Mpc}$). There is a peak in $R$ around $\unit{Mpc}$ for WD-WD, BH-BH, NS-NS and BH-NS binaries, indicating that those systems can merge deep in the IGM. Conversely, EBBHs preferentially merge at $R<10^{-2}\unit{Mpc}$, due to their much shorter GW radiation merger times. The bottom subplots of \fig{fig:cosmic_rate} show the distribution of the cosmic {ejected} merger rate ${\mathrm{d}^3N_\mathrm{eject}}/({\mathrm{d} z^\prime\,\mathrm{d} R\,\mathrm{d} t})$. We see that most of the {ejected} mergers should be detected in the range $z^\prime \sim [0.5,3]$.
By integrating ${\mathrm{d}^4N_\mathrm{eject}}/({\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t})$ we get
\begin{equation}
\Gamma_\mathrm{eject} = \frac{\mathrm{d} N_\mathrm{eject}}{\mathrm{d} t} = \int \frac{\mathrm{d}^4N_\mathrm{eject}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t}\mathrm{d} z^\prime \mathrm{d} R\mathrm{d} m_\mathrm{chirp}
\end{equation}
which is the {ejected} merger event rate up to redshift $5$. The numerical values of the merger rate can be found in \tab{tab:R}, where $n_\mathrm{wd-wd}$, $n_\mathrm{ns-ns}$, $n_\mathrm{bh-bh}$, $n_\mathrm{bh-ns}$ and $n_\mathrm{ebbh}$ are the binary number per unit solar mass in the galactic centre for WD-WD, NS-NS, BH-BH, BH-NS and EBBH, respectively. {In the table, a subset of WD-WD mergers for which the total mass $>1.4 M_\odot$ has been listed separately. We simply label these as Type Ia SNe.}
To assess the detectability of emitted GWs, we must consider the signal-to-noise ratio (SNR) of the GW detectors that can detect the events out to $z^\prime=f(m_\mathrm{chirp})$. With this, the detectable {ejected} merger event rate as a function of $R$ and $z^\prime$ can be derived from
\begin{equation}
\frac{\mathrm{d}^3N_\mathrm{eject}}{\mathrm{d} R\mathrm{d} z^\prime\mathrm{d} t}(R,z^\prime)=\int_{f(m_\mathrm{chirp})<z^\prime}\frac{\mathrm{d}^4N_\mathrm{eject}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t}\mathrm{d} m_\mathrm{chirp}\,.
\end{equation}
\subsection{Quick merger rate}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/qmerger_RM}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/qmerger_RZ}
\caption{Cosmic distribution of quick merger events and their dependences on the parameters $R$, $m_\mathrm{chirp}$ and $z^\prime$ normalized by the undetermined parameter $n_\mathrm{bi}$, where $n_\mathrm{bi}$ is the CO binary number per unit solar mass in the galactic centre. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:cosmic_quick_rate}
\end{figure*}
Following the same procedure as for {ejected} merger events, we can calculate the cosmic rate of quick merger events as follows,
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^3N_\mathrm{quick}}{\mathrm{d} z^\prime,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}(z^\prime, m_\mathrm{chirp})\\
&=\int N_{\mathrm{quick}}\frac{\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\frac{\delta(z-z^\prime)}{N_{\mathrm{exp}}(\log q,\log M_1)}\mathrm{d} \log M_1 \mathrm{d}\log q\,,
\end{split}
\end{equation}
where $N_{\mathrm{quick}}$ is verified to be independent of $M_1$ and $q$ in \sect{sec:relfraction}.
\fig{fig:cosmic_quick_rate} shows the distribution of ${\mathrm{d}^3N_\mathrm{quick}}/({\mathrm{d} z^\prime,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t})$; the corresponding rates can be obtained by
\begin{equation}
\frac{\mathrm{d} N_\mathrm{quick}}{\mathrm{d} t} = \int \frac{\mathrm{d}^3N_\mathrm{quick}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}\mathrm{d} z^\prime \mathrm{d} m_\mathrm{chirp}
\end{equation}
The numerical values of the merger rate can be found in \tab{tab:R}.
\subsection{Slow merger rate}
As for the quick mergers, we estimate the cosmic rate of {slow} mergers as
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^3N_\mathrm{slow}}{\mathrm{d} z^\prime,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}(z^\prime, m_\mathrm{chirp})\\
&=\int \frac{\delta(z-z^\prime)\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\frac{N_{\mathrm{slow}}(\log M_1,\log q)}{N_{\mathrm{exp}}(\log q,\log M_1)}\mathrm{d} \log M_1 \mathrm{d}\log q\,.
\end{split}
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/smerger_RM}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/smerger_RZ}
\caption{Cosmic distribution of slow merger events and their dependences on the parameters $R$, $m_\mathrm{chirp}$ and $z^\prime$ normalized by the undetermined parameter $n_\mathrm{bi}$, where $n_\mathrm{bi}$ is the CO binary number per unit solar mass in the galactic centre. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:cosmic_in_rate}
\end{figure*}
\fig{fig:cosmic_in_rate} shows the distribution ${\mathrm{d}^3N_\mathrm{slow}}/({\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t})$, which is almost identical to the quick merger distribution. However, the non-negligible residual eccentricities for quick mergers determine the region in phase-space where the merger events tend to occur, as indicated by \fig{fig:incenter_rel} and \fig{fig:quick_rel}. The corresponding rates are computed via the relation
\begin{equation}
\frac{\mathrm{d} N_\mathrm{slow}}{\mathrm{d} t} = \int \frac{\mathrm{d}^3N_\mathrm{slow}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}\mathrm{d} z^\prime \mathrm{d} m_\mathrm{chirp}
\end{equation}
The numerical values of the merger rate can be found in \tab{tab:R}.
\subsection{Merger distance from the galactic centre}
Due to differences in the surrounding environments, merger events occurring at different distances from the galactic centre are likely to show different signatures in their electromagnetic (EM) counterparts, as discussed in the following subsection. {Quick mergers} induced by very strong LK oscillations are naturally produced close to the SMBHB. However, the post-encounter, {slow mergers} created by the secondary SMBH and {ejected mergers} may occur in the outskirts or even outside the host galaxy.
\begin{figure*}
\includegraphics[width=\textwidth]{figs/R_dist_all}
\caption{The distribution of the merger distance from the galactic centre $R$ for different types of mergers. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:r_dist}
\end{figure*}
\fig{fig:r_dist} shows the merger distances of quick mergers, {slow} mergers and {ejected} mergers. Independent of the CO species, quick mergers are found within $100~\mathrm{pc}$. Conversely, all {ejected} mergers will be found at $>10^2\mathrm{pc}$. As a representative subdivision, we consider $R<10^2\mathrm{pc}$, $10^2\mathrm{pc}<R<10^5\mathrm{pc}$ and $R>10^5\mathrm{pc}$, which correspond to, respectively, the galactic centre, the galactic field and the IGM. The merger rates of different CO binaries in the three regions of parameter space are given in \tab{tab:R}. The parameter $n_{bi}$ quantifies the number of CO binaries per unit mass, and can be determined by combining a specific stellar IMF, with stellar evolution models and empirical estimates for the binary fraction in galactic nuclei.
\begin{table}
\begin{tabular} { | l | l | l | l |}
\hline
& $R < 10^2\mathrm{pc}$ & $R \in [10^2,10^5]\mathrm{pc}$ & $R > 10^5\mathrm{pc}$ \\
\hline
\textit{Ejected}:& & &\\
BH-BH & $1.25 \times 10^{1} n_\mathrm{bh-bh}$ & $7.40 \times 10^{2} n_\mathrm{bh-bh}$& $8.96 \times 10^{2} n_\mathrm{bh-bh}$\\
NS-NS & $7.65 \times 10^{1} n_\mathrm{ns-ns}$ & $3.87 \times 10^{2} n_\mathrm{ns-ns}$& $6.13 \times 10^{2} n_\mathrm{ns-ns}$\\
BH-NS & $3.72 \times 10^{1} n_\mathrm{bh-ns}$ & $4.56 \times 10^{2} n_\mathrm{bh-ns}$& $7.20 \times 10^{2} n_\mathrm{bh-ns}$\\
WD-WD & $3.87 \times 10^{0} n_\mathrm{wd-wd}$ & $1.41 \times 10^{2} n_\mathrm{wd-wd}$& $2.09 \times 10^{2} n_\mathrm{wd-wd}$\\
Ia SNe & $3.87 \times 10^{0} n_\mathrm{wd-wd}$ & $1.43 \times 10^{1} n_\mathrm{wd-wd}$& $1.83 \times 10^{1} n_\mathrm{wd-wd}$\\
EBBH & $2.47 \times 10^{3} n_\mathrm{bh-bh}$ & $4.02 \times 10^{3} n_\mathrm{bh-bh}$& $1.38 \times 10^{3} n_\mathrm{bh-bh}$\\
\hline
\textit{Quick}:& & &\\
BH-BH & $2.83 \times 10^{4} n_\mathrm{bh-bh}$ & $\times$ & $\times$\\
NS-NS & $7.33 \times 10^{4} n_\mathrm{ns-ns}$ & $\times$ & $\times$\\
BH-NS & $2.30 \times 10^{4} n_\mathrm{bh-ns}$ & $\times$ & $\times$\\
WD-WD & $6.45 \times 10^{4} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
Ia SNe & $8.72 \times 10^{3} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
EBBH & $1.09 \times 10^{5} n_\mathrm{bh-bh}$ & $\times$ & $\times$\\
\hline
\textit{Slow}:& & &\\
BH-BH & $6.63 \times 10^{4} n_\mathrm{bh-bh}$ & $\times$ & $\times$\\
NS-NS & $2.77 \times 10^{4} n_\mathrm{ns-ns}$ & $\times$ & $\times$\\
BH-NS & $3.82 \times 10^{4} n_\mathrm{bh-ns}$ & $\times$ & $\times$\\
WD-WD & $8.42 \times 10^{3} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
Ia SNe & $1.30 \times 10^{3} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
EBBH & $3.08 \times 10^{4} n_\mathrm{bh-bh}$ & $7.73 \times 10^{1} n_\mathrm{bh-bh}$ & $\times$\\
\hline
\textit{Total}:& & &\\
BH-BH & $9.47 \times 10^{4} n_\mathrm{bh-bh}$ & $7.39 \times 10^{2} n_\mathrm{bh-bh}$& $8.96 \times 10^{2} n_\mathrm{bh-bh}$\\
NS-NS & $1.01 \times 10^{5} n_\mathrm{ns-ns}$ & $3.86 \times 10^{2} n_\mathrm{ns-ns}$& $6.12 \times 10^{2} n_\mathrm{ns-ns}$\\
BH-NS & $6.12 \times 10^{4} n_\mathrm{bh-ns}$ & $4.56 \times 10^{2} n_\mathrm{bh-ns}$& $7.20 \times 10^{2} n_\mathrm{bh-ns}$\\
WD-WD & $7.29 \times 10^{4} n_\mathrm{wd-wd}$ & $1.41 \times 10^{2} n_\mathrm{wd-wd}$& $2.09 \times 10^{2} n_\mathrm{wd-wd}$\\
Ia SNe & $1.00 \times 10^{4} n_\mathrm{wd-wd}$ & $1.43 \times 10^{1} n_\mathrm{wd-wd}$& $1.83 \times 10^{1} n_\mathrm{wd-wd}$\\
EBBH & $1.42 \times 10^{5} n_\mathrm{bh-bh}$ & $4.10 \times 10^{3} n_\mathrm{bh-bh}$& $1.38 \times 10^{3} n_\mathrm{bh-bh}$\\
\hline
\hline
\end{tabular}
\caption{{Model-independent rate of merger events classified by $R$, where $n_\mathrm{wd-wd}$, $n_\mathrm{ns-ns}$, $n_\mathrm{bh-bh}$ and $n_\mathrm{bh-ns}$ are the binary number per unit solar mass in the galactic centre for WD-WD, NS-NS, BH-BH and BH-NS binaries, respectively. The units for all entries are in $M_\odot\,\mathrm{yr}^{-1}$}.}
\label{tab:R}
\end{table}
The first step in the determination of $n_{bi}$ is to estimate the fraction of single WDs, NSs, and BHs in the galactic nuclei, which requires the evolution of the stellar population from a specific IMF. Since we use the outcome of the {\tt StarTrack} code for our scattering experiments, we follow the same procedure as \citet[][]{2012ApJ...749...91F}. It is generally assumed that the remnant mass is directly determined by the core mass $M_{\rm core}$. Here, we map the IMF to the core mass by using \citep[e.g.,][]{2015MNRAS.451.4086S},
\begin{equation}
M_{\rm core} = -2 + (B + 2)(g(Z, M_\star; K_1, \delta_1) + g(Z, M_\star; K_2, \delta_2))
\label{eq:mcore}
\end{equation}
where $Z$ is the metallicity, $M_\star$ is the initial stellar mass and
\begin{equation}
\begin{aligned}
&g(Z, m; x, y) = \frac{0.5}{1+ 10^{(x(Z)-m)y(Z)}}\\
&B=40.98 + 3.415\times 10^4Z - 8.064\times 10^6 Z^2\\
&K_1 = 35.17 + 1.548\times 10^4Z - 3.759\times 10^6 Z^2\\
&k_2 = 20.36 + 1.162\times 10^5Z - 2.276\times 10^7 Z^2\\
&\delta_1 = 2.5\times 10^{-2} -4.346Z +1.34\times 10^3 Z^2\\
&\delta_2 = 1.75\times 10^{-2} + 11.39Z - 2.902\times 10^3 Z^2
\end{aligned}
\end{equation}
for $Z\in[0.1, 0.4] Z_\odot$. The relationship between the remnant mass and the core mass is
\begin{equation}
M_{\rm rem} =
\left\{
\begin{aligned}
&0.619 M_\odot, \quad M_{\rm core} \leq M_{\rm NS,core}^{\rm lower}&\\
&\mathrm{max}\{h(M_{\rm core}, Z), 1.27M_\odot\}, \quad M_{\rm NS,core}^{\rm lower} < M_{\rm core} \leq 5\, M_\odot&\\
&h(M_{\rm core}, Z),\quad 5\, M_\odot < M_{\rm core} \leq 10\, M_\odot&\\
&\mathrm{max}\{h(M_{\rm core}, Z), f(M_{\rm core}, Z)\}, \quad 10\, M_\odot < M_{\rm core}&
\end{aligned}
\right.
\end{equation}
where
\begin{eqnarray}
h(M_{\rm core}, Z) &=& A_1(Z) + \frac{A_2(Z)- A_1(Z)}{1 + 10 ^{ (L(Z) - M_{\rm core})\eta(Z) }}\\
f(M_{\rm core}, Z) &=& m(Z)M_{\rm core} + q(Z)\\
m(Z) &=& -6.476\times 10^{2}Z + 1.911\\
q(Z) &=& 2.300\times 10^{3}Z + 11.67\\
A_1(Z) &=& 1.340 - \frac{29.46}{ 1 + (\frac{Z}{1.110\times 10^{-3}})^{2.361}}\\
A_2(Z) &=& 80.22 - 74.73\frac{Z^{0.965}}{2.700\times 10^{-3} + Z^{0.965} }\\
L(Z) &=& 5.683 + \frac{3.533}{ 1 + (\frac{Z}{7.43\times 10^{-3}})^{1.993}}\\
\eta(Z) &=&1.066 - \frac{1.121}{ 1 + (\frac{Z}{2.558\times 10^{-2}})^{0.609}}
\end{eqnarray}
$M_{\rm NS,core}^{\rm lower}$ is the lower mass limit for the NS core, found by substituting $M_\star$ with
\begin{equation}
M_{\rm NS}^{\rm lower} = 9M_\odot + 0.9\log_{10}(Z/Z_\odot)
\end{equation}
in equation \eqref{eq:mcore} \citep[e,g,.][]{2012ApJ...749...91F}.
If, we take the Salpeter IMF \citep[e.g.,][]{1955ApJ...121..161S}
\begin{equation}
\xi(m)dm \propto \bigg(\frac{m}{M_\odot}\bigg)^{-2.35}dm
\end{equation}
with $m_{\rm min} = 0.4\,M_\odot$, $m_{\rm max} = 150\,M_\odot$ and $Z=0.1Z_\odot$ as typically used in {the initial binary populations from {\tt StarTrack}}, the stellar evolution process yields $1.19$ WDs per unit solar mass and $2.04\times 10^{-2}$ COs per unit solar mass.
The {\tt StarTrack} code gives the BH-BH, NS-NS and BH-NS number fractions among all CO binaries to be $50\%$, $39.7\%$ and $10.3\%$, respectively. Therefore, if we take the binary number fraction in galactic nuclei to be 0.1
as a representative value, we find $n_\mathrm{wd-wd}=5.9\times 10^{-2} \,M_\odot^{-1}$, $n_\mathrm{bh-bh}=1.0\times 10^{-3}\,M_\odot^{-1}$, $n_\mathrm{ns-ns}=8.1\times 10^{-4}\,M_\odot^{-1}$ and $n_\mathrm{bh-ns}=2.1\times 10^{-4}\,M_\odot^{-1}$. The corresponding cosmic merger rates are reported in \tab{tab:rates}.
\begin{table}
\begin{tabular} { | l | l | l | l |}
\hline
& $R < 10^2\mathrm{pc}$ & $R \in [10^2,10^5]\mathrm{pc}$ & $R > 10^5\mathrm{pc}$ \\
\hline
\textit{Ejected}:& & &\\
BH-BH & $1.25 \times 10^{-2} $ & $7.34 \times 10^{-1} $& $8.96 \times 10^{-1} $\\
NS-NS & $6.20 \times 10^{-2} $ & $3.13 \times 10^{-1} $& $4.96 \times 10^{-1} $\\
BH-NS & $7.82 \times 10^{-3} $ & $9.57 \times 10^{-2} $& $1.51 \times 10^{-1} $\\
WD-WD & $2.29 \times 10^{-1} $ & $8.34 \times 10^{0} $& $1.23 \times 10^{1} $\\
Ia SNe & $2.29 \times 10^{-1} $ & $8.42 \times 10^{-1} $& $1.08 \times 10^{0} $\\
EBBH & $2.47 \times 10^{0} $ & $4.02 \times 10^{0} $& $1.38 \times 10^{0} $\\
\hline
\textit{Quick}:& & &\\
BH-BH & $2.83 \times 10^{1} $ & $\times$ & $\times$\\
NS-NS & $5.94 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $4.82 \times 10^{0} $ & $\times$ & $\times$\\
WD-WD & $3.80 \times 10^{3} $ & $\times$ & $\times$\\
Ia SNe & $5.14 \times 10^{2} $ & $\times$ & $\times$\\
EBBH & $1.09 \times 10^{2} $ & $\times$ & $\times$\\
\hline
\textit{Slow}:& & &\\
BH-BH & $6.63 \times 10^{1} $ & $\times$ & $\times$\\
NS-NS & $2.25 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $8.02 \times 10^{0} $ & $\times$ & $\times$\\
WD-WD & $4.97 \times 10^{2} $ & $\times$ & $\times$\\
Ia SNe & $7.69 \times 10^{1} $ & $\times$ & $\times$\\
EBBH & $3.08 \times 10^{1} $ & $7.73 \times 10^{-2} $ & $\times$\\
\hline
\textit{Total}:& & &\\
BH-BH & $9.47 \times 10^{1} $ & $7.39 \times 10^{-1} $& $8.96 \times 10^{-1} $\\
NS-NS & $8.19 \times 10^{1} $ & $3.13 \times 10^{-1} $& $4.96 \times 10^{-1} $\\
BH-NS & $1.29 \times 10^{1} $ & $9.57 \times 10^{-2} $& $1.51 \times 10^{-1} $\\
WD-WD & $4.30 \times 10^{3} $ & $8.34 \times 10^{0} $& $1.23 \times 10^{1} $\\
Ia SNe & $5.91 \times 10^{2} $ & $8.42 \times 10^{-1} $& $1.08 \times 10^{0} $\\
EBBH & $1.42 \times 10^{2} $ & $4.10 \times 10^{0} $& $1.38 \times 10^{0} $\\
\hline
\hline
\end{tabular}
\caption{Cosmic rate of merger events classified by $R$ with a Salpeter IMF. The units of all entries are in $\mathrm{yr}^{-1}$}
\label{tab:rates}
\end{table}
The same process can also be performed for a log-flat IMF
\begin{equation}
\xi(m)dm \propto \bigg(\frac{m}{M_\odot}\bigg)^{-1}dm
\end{equation}
which yields $n_\mathrm{wd-wd}=1.7\times 10^{-3}\,M_\odot^{-1}$, $n_\mathrm{bh-bh}=1.9\times 10^{-3}\,M_\odot^{-1}$, $n_\mathrm{ns-ns}=1.3\times 10^{-3}\,M_\odot^{-1}$ and $n_\mathrm{bh-ns}=3.5\times 10^{-4}\,M_\odot^{-1}$. The corresponding cosmic merger rates are listed in \tab{tab:log_flat_rates}.
\begin{table}
\begin{tabular} { | l | l | l | l |}
\hline
& $R < 10^2\mathrm{pc}$ & $R \in [10^2,10^5]\mathrm{pc}$ & $R > 10^5\mathrm{pc}$ \\
\hline
\textit{Ejected}:& & &\\
BH-BH & $2.38 \times 10^{-2} $ & $1.41 \times 10^{0} $& $1.70 \times 10^{0} $\\
NS-NS & $9.95 \times 10^{-2} $ & $5.03 \times 10^{-1} $& $7.97 \times 10^{-1} $\\
BH-NS & $1.30 \times 10^{-2} $ & $1.59 \times 10^{-1} $& $2.52 \times 10^{-1} $\\
WD-WD & $6.59 \times 10^{-3} $ & $2.40 \times 10^{-1} $& $3.11 \times 10^{-1} $\\
Ia SNe & $6.59 \times 10^{-3} $ & $2.43 \times 10^{-2} $& $3.11 \times 10^{-2} $\\
EBBH & $4.70 \times 10^{0} $ & $7.65 \times 10^{0} $& $2.63 \times 10^{0} $\\
\hline
\textit{Quick}:& & &\\
BH-BH & $5.39 \times 10^{1} $ & $\times$ & $\times$\\
NS-NS & $9.53 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $8.03 \times 10^{0} $ & $\times$ & $\times$\\
WD-WD & $1.09 \times 10^{2} $ & $\times$ & $\times$\\
Ia SNe & $1.48 \times 10^{1} $ & $\times$ & $\times$\\
EBBH & $2.07 \times 10^{2} $ & $\times$ & $\times$\\
\hline
\textit{Slow}:& & &\\
BH-BH & $1.26 \times 10^{2} $ & $\times$ & $\times$\\
NS-NS & $3.60 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $1.34 \times 10^{1} $ & $\times$ & $\times$\\
WD-WD & $1.43 \times 10^{1} $ & $\times$ & $\times$\\
Ia SNe & $2.22 \times 10^{0} $ & $\times$ & $\times$\\
EBBH & $5.85 \times 10^{1} $ & $1.47 \times 10^{-1} $ & $\times$\\
\hline
\textit{Total}:& & &\\
BH-BH & $1.80 \times 10^{2} $ & $1.41 \times 10^{0} $& $1.70 \times 10^{0} $\\
NS-NS & $1.31 \times 10^{2} $ & $5.03 \times 10^{-1} $& $7.97 \times 10^{-1} $\\
BH-NS & $2.14 \times 10^{1} $ & $1.59 \times 10^{-1} $& $2.52 \times 10^{-1} $\\
WD-WD & $1.24 \times 10^{2} $ & $2.40 \times 10^{-1} $& $3.56 \times 10^{-1} $\\
Ia SNe & $1.70 \times 10^{1} $ & $2.43 \times 10^{-2} $& $3.11 \times 10^{-2} $\\
EBBH & $2.70 \times 10^{2} $ & $7.79 \times 10^{0} $& $2.63 \times 10^{0} $\\
\hline
\hline
\end{tabular}
\caption{Cosmic rate of merger events classified by $R$ with a log-flat IMF. The units of all entries are $\mathrm{yr}^{-1}$}
\label{tab:log_flat_rates}
\end{table}
\subsection{Electromagnetic counterparts for different types of merger events}
The possibility of drawing astrophysical inferences from the merger
events discussed here largely relies on our ability to identify their merger locations with respect to their galaxy hosts. Studies of host galaxies have a long history in the context of both long and short GRBs, and have been used to probe the nature of their progenitors. The events discussed here span a very large range of distances; however, while {slow} merger events can be produced in a variety of astrophysical conditions (such as dynamical interactions with a single SMBH), the production of HVBs can only occur in the presence of a SMBHB, and hence are especially interesting, as discussed in this paper.
WD-WD mergers are not observable by LIGO/Virgo in gravitational waves, and hence their EM signatures provide the only available empirical information. Mergers of two WDs in a binary have been proposed as candidates of Type Ia SNe \citep[the so-called double-degenerate scenario,][]{1984ApJS...54..335I}.
These transients are typically found to trace the blue light of galaxies \citep[e.g.][]{2015yCat..74480732A}, hence generally occurring within their hosts. The detection of a Type Ia SNe in the IGM would signal an origin from an HVB, whether a WD-WD or a WD and a companion main sequence star \citep[as in the single degenerate scenario,][]{1973ApJ...186.1007W}.
For CO binaries made up of combinations of NSs and BHs, detection by GWs -- when within the LIGO horizon -- constitutes the initial event trigger. However, localization relies on EM counterparts. Among the three possible types of CO binaries, BH-BH mergers are the least likely to be EM bright. For merger events occurring within an AGN disk, accretion by the local gas may provide some source of accretion at the time of merger \citep{2018arXiv180702859S}. For binaries in the intergalactic medium, on the other hand, special conditions are required, such as, e.g., a fossil disk around one of the two BHs of the binary \citep{Perna2019}, or very strong charges \citep{Zhang2016}. However, even if such conditions were realized and a relativistic jet was launched, the likelihood of observing it both in $\gamma$-rays as well as at longer wavelengths would be limited by the fact that relatively bright events could only be detected if the jet
points towards the observer. This is because, lacking ejecta material in the close environment of the merger for the jet to interact with, side emission is highly suppressed. Therefore, identification and localization of a BH-BH merger from an HVB in the intergalactic medium may not be feasible.
On the other hand, NS-NS mergers have been confirmed to produce
a relativistic jet and electromagnetic radiation covering a very
broad spectrum, from $\gamma$-rays to the radio band \citep{Abbott2017}.
Even when observed at a sizable angle with respect to the jet axis,
the interaction of the jet with the tidally disrupted material of the NS produces a bright cocoon, which makes these events more easily detectable within the LIGO horizon even for more unfavorable viewing configurations. In terms of EM radiation, NS-NS mergers from HVBs would be characterized by a much weaker afterglow (i.e. X-rays through radio emission) relative to the prompt $\gamma$-ray emission. This is because, unlike the $\gamma$-ray radiation, which only depends on the properties of the jet and the ejecta material, the afterglow radiation also depends on the density
$n$ of the medium in which the jet propagates. Lower densities result in dimmer afterglows \citep{Sari1998}. Hence a binary NS merger which is very bright in $\gamma$-rays but very dim at longer wavelengths could signal a merger site in the intergalactic medium, even if not accurately localized due to the lack of a bright afterglow.
Last, the situation for an HVB made up of an NS and a BH is more
complex. Only binaries with mass ratios $q$ not exceeding $\sim 3-5$ (with the precise value dependent on the equation of state of the NS) are expected to lead to the formation of an accretion disk during the merger, and hence possibly drive a relativistic jet, similar to the NS-NS merger case. On the other hand, large mass ratios will result in the NS being swallowed by the BH without being tidally disrupted. Sources of radiation in this latter case, if at all, would be more similar to those discussed earlier for binary BH mergers.
\section{Conclusions}\label{sec:summary}
In this work we have studied interactions between CO binaries and massive black hole
binaries that may lead to merger events detectable by GW detectors.
We have performed high-precision N-body simulations in the local reference frame of an SMBHB and combined the scattering outcomes with results from the cosmology simulation Millennium-II to estimate the cosmic rate of different types of CO merger events from dynamical interactions with SMBHBs. Our main conclusions are summarized in the following.
Our simulations show that the interactions between the CO binaries and the SMBHB can produce different types of merger events, which we have dubbed 'quick mergers', '{slow} mergers' and '{ejected} mergers'. Different types of merger events occur at different distances from the galactic centre. Quick mergers with non-negligible eccentricities in the LIGO band can be produced/detected in the innermost region of the SMBHB, in proximity of the primary SMBH. {Slow} mergers, with fully circularized orbits in the LIGO band, can be produced from BSE in the stable region of SMBHBs, or by post-encounters with the secondary SMBH, and can be detected at moderate distances from the galactic centre. The {ejected} mergers that are produced from the hypervelocity binaries could potentially be detected in the galactic halo or in the IGM. Different types of merger events can be precisely classified by the stability map of the SMBHB introduced in Section \ref{sec:stability} and might be distinguished observationally by measuring the distribution of merger distances to the galactic centre from EM counterparts, if available.
Close encounters between CO binaries and SMBHBs can create hypervelocity binaries that could escape the host galaxies and eventually merge in the intergalactic medium. Such a high velocity cannot be produced by interactions in globular clusters nor multiple stellar systems; thus an off-centre merger event might indicate the existence of a nearby SMBHB, especially a high mass ratio SMBHB formed from a prior major galaxy-galaxy merger, which would kick out any merged CO binaries occurring in its orbital plane.
Assuming a binary fraction $n_\mathrm{bi}$ of $10\%$ and a Salpeter IMF, we estimate the merger rate of WD-WD binaries to be $4.32\times 10^{3}\mathrm{yr}^{-1}$, of which $12.3\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $8.34\mathrm{yr}^{-1}$ occur in the DM halo, and $4.30\times 10^{3}\mathrm{yr}^{-1}$ occur in the galactic nucleus. The merger rate of BH-BH binaries is $96.3\mathrm{yr}^{-1}$, of which $0.896\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.740\mathrm{yr}^{-1}$ occur in the DM halo, and $94.7\mathrm{yr}^{-1}$ occur in the galactic nucleus. The merger rate of NS-NS binaries is $82.7\mathrm{yr}^{-1}$, of which $0.496\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.313\mathrm{yr}^{-1}$ occur in the DM halo, and $81.9\mathrm{yr}^{-1}$ occur in the galactic nucleus. The merger rate of EBBHs is $148\mathrm{yr}^{-1}$, of which $1.08\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.842\mathrm{yr}^{-1}$ occur in the DM halo, and $142\mathrm{yr}^{-1}$ occur in the galactic nucleus. Finally, the merger rate of BH-NS binaries is $13.1\mathrm{yr}^{-1}$ of which $0.151\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.0957\mathrm{yr}^{-1}$ occur in DM halos and $12.9\mathrm{yr}^{-1}$ occur in galactic nuclei.
Finally, we caution that our scattering experiments can only describe the cusp erosion phase, in which the SMBHB ejects stars in the bound cusp surrounding it as it first becomes bound. Further ejection of unbound CO binaries in the later stages of the SMBHB evolution might contribute to the cosmic rate of fast, slow and ejected CO binary mergers. We defer the study of this later phase to future work.
\section*{Acknowledgements}
RP and NL acknowledge support by NSF award AST-1616157.
The Center for Computational Astro-physics at the Flatiron Institute is supported by the Simons Foundation.
\section{Introduction}\label{sec:intro}
The detection of binary black hole (BH) and binary neutron star (NS) mergers via gravitational waves (GWs)\citep[e.g.,][]{2019arXiv190201557T,2016PhRvL.116f1102A,2018arXiv181112940T,2017PhRvL.119p1101A} has opened a new window to explore the Universe. A lot of interest has been devoted to understanding the physical origins of these events, and in particular the formation scenarios of the binaries which merge. The most popular hypotheses
involve field binaries \citep[e.g.,][]{2003MNRAS.341..385P, 2013ApJ...779...72D, 2014ApJ...789..120B, 2016Natur.534..512B, 2016A&A...588A..50M, 2016MNRAS.458.2634M} and dynamical formation \citep[e.g.,][]{2000ApJ...528L..17P,2006ApJ...637..937O, 2009ApJ...692..917M, 2013MNRAS.429.2298M, 2014MNRAS.444...29L, 2016ApJ...824L...8R, 2018ApJ...866L...5R, 2016ApJ...816...65A, 2017ApJ...834...68C, 2018PhRvD..97j3014S, 2018MNRAS.481.5445S, 2018MNRAS.478.4030G, 2019MNRAS.486.5008A, 2018MNRAS.481.5123B, 2018MNRAS.473..909B, 2018PhRvL.121p1103F, 2018arXiv181110627F, 2019arXiv190100863D,Perna2019}.
Once binaries are formed, their subsequent evolution up to merger, and hence the locations of their mergers, are also of great importance. Compact object (CO) binaries located in the vicinity of a supermassive black hole (SMBH) are subject to Lidov-Kozai (LK) oscillations, which accelerate their merger times, thus enhancing their merger rates \citep[e.g.,][]{2017ApJ...841...77A, 2017ApJ...836L..26C, 2016ApJ...831..187A, 2012ApJ...757...27A, 2002ApJ...578..775B, 2018ApJ...856..140H, 2017ApJ...836...39S, 2016ApJ...828...77V, 2016ARA&A..54..441N, 2003ApJ...598..419W}. This channel of CO mergers has been shown to have rates which are competitive with other proposed channels, due to the combination of the large star density in the vicinity of the central SMBH, and the shorter merger times resulting from dynamical interactions.
The phenomenology associated with dynamical interactions becomes richer when two SMBHs are paired together, making up an SMBH binary (SMBHB). In this case, it has been shown that, in addition to Hypervelocity stars (or CO), also Hypervelocity binaries (HVBs) can be produced by the dynamical interaction between the CO binary and the SMBHB \citep[][]{2007ApJ...666L..89L,2009MNRAS.392L..31S,2018MNRAS.475.4595W,2019MNRAS.482.3206W,Coughlin2018}. Therefore, detection of HVBs can serve as an unique diagnostic for the presence of a secondary SMBH in galactic nuclei.
Besides ejection, another outcome of strong dynamics around SMBHs is a stellar tidal disruption event \citep[TDE, e.g.,][]{1988Natur.333..523R}. TDEs can also be used to probe SMBHBs in many different ways \citep[e.g.,][]{2017MNRAS.469.2042B,2017MNRAS.471L.115C,2018MNRAS.477.4009D}.
TD rates have been found to be particularly enhanced by unequal SMBHBs, with mass ratios $<0.1$ \citep{2011ApJ...729...13C,2018arXiv181101960T}. Moreover, since main sequence stars are generally disrupted only by SMBHs with masses below $10^8M_\odot$ (more massive black holes just swallow the whole star without tearing it apart), a TDE flare in a galaxy expected to host a billion solar mass SMBH would be indicative of the presence of a secondary, much lighter, SMBH companion, as proposed by \citet{2018MNRAS.479.3181F}. Interestingly, few peculiar transient events have been interpreted as candidate TDEs from low mass ratio SMBHB binaries \citep{2009ApJ...706L.133L,2018MNRAS.474.3857C}. Conversely, the HVB mechanism quantified by \citet{2018MNRAS.475.4595W} operates most efficiently at SMBHB mass ratios close to unity. Hence, the combination of these two diagnostics cover a significant fraction of the total parameter space available to SMBHBs. The identification of traditional, main sequence star HVBs is limited to the Milky Way and the local group. On the other hand, due to their compactness, the parameter space available to HVBs is much higher for degenerate stars than for main-sequence stars, since binaries containing remnants, especially neutron stars (NSs) and black holes (BHs), can exist at smaller orbital separations. It is therefore interesting to model the generation of CO-HVB ejection from SMBHBs and their potential observable features, which is one of the main focuses of this work.
The observable phenomenology associated with a merger event in the electromagnetic (EM) domain depends on the type of objects making up the CO binary.
Mergers of two NSs have been confirmed to be associated with short gamma ray bursts (GRBs) \citep{Abbott2017}, and it is expected that this should also be the case for mergers of NS-BH binaries for small mass ratios (see e.g. the review in \citealt{Bartos2013}). BH-BH mergers are most likely to be EM-quiet, albeit ideas for accompanying radiation have been proposed \citep{Perna2016,Loeb2016,Zhang2016,Woosley2016,DeMink2017,Bartos2017,Liebling2016,Murase2016,Janiuk2017,Fraschetti2018}.
The location of the merger site has long been considered as an important diagnostic of the progenitor type. For CO mergers, the predictions are of a wide distribution of distances from the galaxy hosts, primarily reflecting the natal kicks, the merger times, and the mass of the host galaxy (e.g. \citealt{Belczynski2006,Oshau2017,Perna2018}). For all three types of CO binaries (i.e. NS-NS, NS-BH and BH-BH), a small fraction of events at $\sim$~Mpc scale is expected only in small galaxies, with masses $\lesssim 10^{10}M_\odot$
(e.g. \citealt{Perna2002}), and hence yielding a small contribution to the total merger rates.
For larger, Milky-Way type of galaxies, the gravitational potential of the galaxy prevents very large distances to be reached before mergers. Therefore, the production of HVBs presents the special possibility that some fraction of short GRBs may occur in isolated regions of the Universe between galaxies, or in the intergalactic medium (IGM), and not be directly associated with any particular host galaxy. These isolated short GRBs could thus indicate the presence of an SMBH-IMBH/SMBH binary hosted in the nearest galaxy or galaxy group/cluster.
Identifying these off-host GRB events as a function of redshift opens up the possibility of using them to constrain the SMBHB fraction in galaxies and how it changes as a function of cosmic time. The sophisticated $N$-body simulations performed by \citet{2019MNRAS.482.3206W} revealed how the post-ejection properties of the HVBs depend on the properties of the ejecting SMBHB. Thus, combining observations of the electromagnetic properties of off-host GRBs with gravitational wave detection -- specifically the information extracted from the waveform data -- could further constrain the properties of the SMBHBs as well as the environment in their immediate vicinity. Moreover, if Type Ia supernovae result from mergers of two WDs \citep[i.e.,][]{1984ApJS...54..335I}, the detection of 'orphan' (i.e. in the IGM, away from their host galaxies) Type Ia SNe, could provide another powerful probe of SMBH binarity. We will in fact see that hypervelocity WD-WD binaries are also expected to be produced by dynamical interactions with SMBHBs.
In this paper, we combine the cosmic SMBHB population extracted from a large scale cosmological simulation \citep[The Millennium-II,][]{2009MNRAS.398.1150B} with high-resolution, N-body simulations of the interaction of CO binaries with those SMBHBs, to predict the cosmological distribution of CO binaries, and their properties (including radial separation from their host galaxies) at the time of merger.
Additionally, we compute the cosmological merger rate of these events, and their contribution to the GW signal, and discuss our results in the context of present and future GW observations and accompanying electromagnetic counterparts.
The paper is organized as follows. In \sect{sec:SMBHBformation}, we discuss the cosmic SMBHB formation rate and cusp erosion process that provides the ideal environment to produce various types of CO merger events. In \sect{sec:relfraction} we perform the scattering experiments in the reference frame of the SMBHB to obtain the relative fraction of different types of merger events and their dependences on initial parameters, and discuss the properties of the outcomes. In \sect{sec:cosmic}, we combine the outcomes of the scattering experiments and SMBHB formation rate to estimate the cosmic rates of different types of merger events triggered by the SMBHBs. Our conclusions are summarized in \sect{sec:summary}.
\section{SMBHB formation and cusp erosion}\label{sec:SMBHBformation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/SMBHB_formation}
\caption{ The observed SMBHB formation rate derived from the Millennium II simulation as a function of the primary SMBH mass $M_1$, the binary mass ratio $q$, the redshift $z$ and the binary semi-major axis $a_\mathrm{bh}$. These distributions are used in our simulations to sample the parameters of the SMBHBs.}
\label{fig:GammaSMBHB}
\end{figure*}
\subsection{SMBHB formation rate}
{ The cosmic population of SMBHBs is derived using the data from Millennium-II \citep[e.g.,][]{2009MNRAS.398.1150B}, which is a large scale N-body simulation with sufficient resolution to distinguish dark matter halos down to $\sim10^8M_\odot$. On top of this dark matter backbone, semianalytic galaxy evolution models have been included to reconstruct the cosmic history of galaxies down to $\approx10^6M_\odot$, hosting central black holes with masses down to $\sim 10^4M_\odot$. In particular, data are taken from the semianalytic model developed in \citet[][]{2011MNRAS.413..101G}, following the procedure described in \citet[][]{2018arXiv181101960T}. For consistency, all the post processing is performed using the same cosmology as in Millennium-II: $h=H_0/(100\,\mathrm{km\,s^{-1}\,Mpc^{-1}})=0.73$, $\Omega_\Lambda=0.75$ and $\Omega_M=0.25$. Although these values of the cosmological parameters are outdated, a $\sim$ few percent level change in $h$, $\Omega_\Lambda$ or $\Omega_M$ does not have a significant impact on our results.}
We select all mergers involving two galaxies both hosting an SMBH, for a total of $N_\bullet=169435$ events across the history of the Millennium-II simulation. We then construct a 3D grid by logarithmically binning $\log_{10}(M_1)$ and the mass ratio $q$ within the ranges $M_1\in[10^5,10^{10}]M_\odot$ and $q\in[10^{-4},1]$, and by evenly binning the redshift $z$ from 0 to 5 according to the redshift of each Millennium-II snapshot. The redshift of each merger is identified with the redshift at which the progenitor galaxy is identified in the simulation. In doing this, we are implicitly assuming that the timescale for dynamical friction (DF) to bring the two SMBHs together in a relatively close binary following the galaxy merger is shorter than the typical time elapsed between two snapshots of the simulation, which is $\approx 300$ Myr. Although the DF timescales might be longer \citep{1987gady.book.....B}, we ignore this complication in this work.
Dividing the SMBHB number in each bin by the co-moving volume $V_c$ of the Millennium-II simulation, we get an approximation of the differential SMBHB formation rate,
\begin{equation*}
\frac{\mathrm{d}^4 N_\bullet}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} V_c}\,.
\label{eq:SMBHBdens}
\end{equation*}
This quantity is simply the differential number density of SMBHBs per unit redshift, logarithmic mass and mass ratio. If we now multiply it by the standard comoving volume shell ${\mathrm{d} V_c}/{\mathrm{d} z}$ and convert redshift into time by multiplying ${\mathrm{d} z}/{\mathrm{d} t_r}\times{\mathrm{d} t_r}/{\mathrm{d} t}$ with $t_r=t/(1+z)$, we finally obtain the differential SMBHB formation rate observed on Earth,
\begin{equation}
\gamma_\mathrm{\bullet}(z,\log q,\log M_1) = \frac{\mathrm{d}^4 N_\bullet}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} t}\,.
\end{equation}
\fig{fig:GammaSMBHB} shows the distribution of the SMBHB formation rate $\gamma_\mathrm{\bullet}(z,\log q,\log M_1)$ in a 2-D plane by integrating over each single variable $z$, $\log M_1$ and $\log q$ separately.
The SMBHB formation rate as observed on Earth is dominated by systems with $M<10^6$M$_\odot$, with mass ratios broadly distributed around $0.1$, reaching out to $z\approx 4$.
\subsection{The SMBHB-bound binary interaction rate}
\label{sec:intrate}
If we take a snapshot of the sky, the differential number of SMBHBs that are actively interacting with the surrounding population of bound stars is given by the differential SMBHB formation rate $\gamma_\mathrm{\bullet}(z,\log q,\log M_1)$ times the duty cycle $D_\mathrm{erd}(z,\log q,\log M_1)$ that the SMBHB takes to erode the cusp of stars bound to $M_1$. Note that $D_\mathrm{erd}$ is generally a function of $z,\log q,\log M_1$. During this duty cycle, the SMBHB interacts with a mass $M_\mathrm{erd}(z,\log q,\log M_1)$, so that the rate at which the SMBHB interacts with bound stars is simply $\dot{M}_{\mathrm{erd}}=M_\mathrm{erd}/D_\mathrm{erd}$ (where we omitted the parametric dependencies). The differential interaction rate is therefore given by the differential number of active SMBHBs times the rate at which those binaries interact with ambient stars:
\begin{eqnarray}
\dot{\mathcal{M}}_\mathrm{act}(z,\log q,\log M_1)&=&\frac{\mathrm{d}^4 M_\mathrm{act}}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} t}\\
&=&\gamma_\mathrm{\bullet} D_{\mathrm{erd}} \dot{M}_{\mathrm{erd}}\\
&=& \frac{\mathrm{d}^4 N_\bullet}{\mathrm{d} z\,\mathrm{d}\log q\,\mathrm{d}\log M_1\,\mathrm{d} t}{M}_{\mathrm{erd}}\,.
\label{eq:Mactive}
\end{eqnarray}
Therefore, to know the differential interaction rate, we simply need to estimate the total mass interacting with the binary, ${M}_{\mathrm{erd}}$. Numerical simulations of unequal mass SMBHBs embedded in stellar cusps have shown that the secondary hole efficiently inspirals down to a separation at which the mass in stars enclosed within its orbit around $M_1$ is of the order of $\approx 2M_2$, and it does this without significantly affecting the stellar distribution \citep[e.g.,][]{2007ApJ...656..879M}. At that point, further shrinking of the orbit proceeds via efficient ejection of bound stars, thus eroding the cusp \citep[e.g.,][]{2008ApJ...686..432S}. Pending numerical factors of order unity, the number of interacting stars can therefore be approximated as $M_\mathrm{erd}\simM_2=qM_1$.
The specific { mass interaction rate} that is observed on Earth up to redshift $z$ can then be computed by integrating $\gamma_\mathrm{erd}(z,\log q,\log M_1)$ from 0 to $z$,
\begin{equation}
\Lambda(z,\log q,\log M_1) = \int_0^z\dot{\mathcal{M}}_\mathrm{act}(z^\prime,\log q,\log M_1) \mathrm{d} z^\prime\,.
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/most_interact}
\caption{Contour plots of the mass in CO binaries that is actively interacting with a SMBHB as a function of SMBHB primary mass and mass ratio. The four panels represent this quantity integrated up to different redshifts $z=0.5$, $z=1$, $z=1.5$ and $z=2$.}
\label{fig:most_interact}
\end{figure*}
\fig{fig:most_interact} shows $\Lambda(z,\log q,\log M_1)$ at different redshifts: $z=0.5$, $z=1.0$, $z=1.5$ and $z=2.0$. This figure indicates that for relatively low redshifts, most of the CO binaries that are actively interacting with SMBHBs correspond to high mass ratio (0.1-1) SMBHBs with primary masses $M_1$ between $10^6\,M_\odot$ and $10^9\,M_\odot$. { Integrating the contours in the figure yields that about $10^6\,M_\odot$ of stars and COs actively interacting with SMBHBs out to $z=2$ is potentially observable from Earth every year.}
In order to set up the scattering experiments in \sect{sec:relfraction}, we need an estimate of the SMBHB separation at which stars are efficiently ejected from the cusp. This corresponds to $a_0$ so that $M_*(r<a_0)=2M_2$. For simplicity, following \citet[][]{2010ApJ...719..851S}, we assume the stellar density profile to follow an isothermal sphere. With this assumption, and by using the $M-\sigma$ relation from \citet[][]{2000ApJ...539L...9F}, we get
\begin{equation}\label{eq:a0}
a_0\sim 0.8\bigg(\frac{q}{1+q}\bigg)\bigg(\frac{M_1}{M_6}\bigg)^{1/2}~\mathrm{pc}\,.
\end{equation}
\section{Scattering experiments in the reference frame of the SMBHB}\label{sec:relfraction}
Not all actively interacting CO binaries will produce outcomes of interest to us, in particular hypervelocity binaries that escape their host galaxies and merge in the intergalactic medium. Therefore, scattering experiments between different types of CO binaries with SMBHBs with different $M_1$ and $q$ values are needed to obtain good statistics in the relevant regions of parameter space.
\subsection{Numerical setup}
We perform extensive scattering experiments using our high-precision code {SpaceHub} (Yihan 2019 in prep.; see also Yihan et al. 2019 for more details), which employs the {\tt ARCHAIN} algorithm \cite[e.g.,][]{1993CeMDA..57..439M} to accurately trace the motion of tight binaries with arbitrarily large mass ratios and eccentricities. The original chain structure in the algorithm, combined with the Kahan summation introduced by us to the regularized code, significantly reduces the round-off errors during close encounters.
\begin{figure}
\center
\includegraphics[width=0.8\columnwidth]{figs/config}
\caption{Schematic illustration of the stellar binary orbiting around the super-massive black hole binary (not to scale).}
\label{fig:config}
\end{figure}
In order to cover the relevant SMBHB masses and mass ratios, as indicated by \fig{fig:GammaSMBHB} and \fig{fig:most_interact}, the $q$ and $M_1$ values of the SMBHBs are logarithmically generated in the range $[10^{-4},1]$ and $[10^5,10^{10}]\,M_\odot$, respectively. The semi-major axis $a_{\bh}$ of the SMBHB is set to $a_0/3$, where $a_0$ is the active ejection radius and the eccentricity of the SMBHB $e_{\bh}$ is uniformly distributed in the range $[0,1]$. The choice of $a_0/3$ is motivated by the fact that N-body simulations and scattering experiments have shown that the whole cusp erosion shrinks the semi-major axis of the SMBHB by roughly a factor of 10 \citep[e.g.,][]{2008ApJ...686..432S,2011MNRAS.415L..35S} starting from $a_0$ given by \eqn{eq:a0}. Therefore, our scattering experiments are performed at a specific semi-major axis of the SMBHB $a_{\bh} = a_0/3$, which is the geometric average of $a_0$ and $a_0/10$. \fig{fig:config} shows the configuration of our scattering experiments and corresponding orbital parameters.
We explore five types of stellar mass binaries: WD-WD, BH-BH, NS-NS, BH-NS and EBBH (eccentric binary BHs). {250k scattering experiments are performed for BH-BH, NS-NS and BH-NS binaries while 500k scattering experiments are performed for WD-WD binaries and EBBHs.}
\subsubsection{WD-WD binary population}
For the WD-WD binaries, we take the masses $m_1$ and $m_2$ of each WD from a normal distribution with $\langle M_{WD}\rangle=0.619 M_\odot$ and $\sigma_m = 0.108 M_\odot$ \citep[e.g.,][]{Tremblay2016}. The separation $a$ of the WD-WD binary is taken from the distribution \citep[e.g.][]{Maoz2017}
\begin{equation}
N(x)\propto x^3 \ln ( 1 + x^{-4})\,,
\end{equation}
where
\begin{equation}
x=\frac{a}{(Kt_0)^{1/4}}
\end{equation}
with
\begin{equation}
K = \frac{256}{5}\frac{G^3}{c^5}m_1m_2(m_1+m_2)\,,
\end{equation}
and $t_0$ is the age of the host galaxy. This is evidently a time-dependent model in which high redshift WD-WD binaries would be more compact. However, for simplicity in our simulations, we take $t_0 = 13.6~{\rm Gyr}$ throughout. The eccentricity $e$ of the WD-WD binaries is drawn from a uniform distribution within $[0, 1]$ \citep{Geller2019}.
The radius of the WD is set to be $0.01R_\odot({m_\odot}/{m_\mathrm{wd}})^{1/3}$.
\subsubsection{BH-BH, NS-NS and BH-NS binary populations}
For the BH-BH, NS-NS and BH-NS binaries, we take the distributions for $m_{\star 1},\, m_{\star 2},\,a_{\star}$ and $e_{\star}$ from the BSE code \href{https://www.syntheticuniverse.org/}{\tt StarTrack} \citep{Belczynski2008} with metallicity $Z=0.1Z_\odot$.
The BH radii are set equal to their Schwarzschild radii, while the radii of the NSs are set to be 10 $\mathrm{km}$.
\subsubsection{Eccentric BH-BH binary population}
The properties of the CO binary populations in galactic nuclei are essentially unknown. \citet{Gond2018} suggested that eccentric BH-BH binaries can form via dynamical capture. Therefore, we perform an additional set of simulations with eccentric binary BHs (labelled EBBHs). Following \citet{Gond2018}, the masses of the BHs are drawn from a Salpeter initial mass function (IMF), $N(m) \propto m^{-2.35}$ \citep[e.g.,][]{1955ApJ...121..161S}. The separation is drawn from a logarithmically uniform distribution where $a_{\rm min}=0.01\,\mathrm{au}$ and $a_{\rm max}=1000\,\mathrm{au}$. The circularity $c=1- e^2$ also obeys a logarithmically uniform distribution, where $c_{\rm min} = 10^{-6}$ and $c_{\rm max} = 1$.\\\\
\begin{figure*}
\includegraphics[width=\textwidth]{figs/m1m2_pop}
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/ae_pop}
\caption{Distribution of the compact object binary population in the $m_{\star 1}$-$m_{\star 2}$ (top series of panels) and $a_{\star}$-$e_{\star}$ (bottom series of panels) plane. In each panel, the main plot shows the 2D probability density distribution (PDF), whereas the side plots show the 1D distribution marginalized over the other quantity. \textit{Top Left} panel: BH-BH; \textit{Top middle} panel: NS-NS; \textit{Top right} panel: BH-NS; \textit{Bottom left} panel: WD-WD; \textit{Bottom right} panel: EBBH. }
\label{fig:COb-pop}
\end{figure*}
\fig{fig:COb-pop} shows the parameters of the five types of CO binaries described above. For the BH-BH case, almost all binaries have circular orbits. Conversely, NS-NS binaries tend to be eccentric, and NS-BH binaries tend to be uniformly distributed in eccentricity. The semi-major axes of the BH-BH, NS-NS and BH-NS binaries are mostly clustered in the $10^{-2}$AU-1AU range, thus forming a population of relatively tight binaries.
The orbits of the centre of mass (CoM) of the CO binaries are created such that $a_{\cm}$ is isothermally distributed (i.e., $p(a)\propto a^{-2}$) around the primary SMBHB in the range of $[R_t, 10a_{\bh}]$, where
\begin{equation}
R_t = 3.7\bigg( \frac{M_1}{m_{\star 1}+m_{\star 2}}\bigg)^{1/3}\frac{ 1+e_{\star}}{1-e_{\cm}}a_{\star}
\end{equation}
is the minimum semimajor axis of the stellar binary CoM around $M_1$ ensuring that the stellar binary will not be tidally disrupted at its apocentre when its CoM orbit approaches pericentre around $M_1$.
All angles and inclinations are sampled according to a spherically symmetric distribution. All anomalies are sampled uniformly in time. For each simulation, at least 50 cycles of the CoM orbit of the CO binary are first performed in isolation to ensure that the CO binaries are stable before interacting directly with the SMBHB.
\subsection{Stability Map}
\label{sec:stability}
Since stars reside deep in the potential well of $M_1$, relaxation processes due to the extended stellar distribution are negligible and the typical star-SMBHB interaction time is much shorter than the relaxation time $T_{\rm relax}$ of the surrounding stellar system. The cusp erosion can therefore be treated as a set of individual star-SMBHB interactions that can be tackled by means of 3-body scattering experiments. If a fraction of ambient stars is in fact made up by binaries of compact objects (WDs, NSs or BHs) then, during the 4-body interaction, several processes may lead to CO binary mergers.
To understand the different physics behind the merger events, we follow the stability map method in our previous work \citep[][]{2019MNRAS.482.3206W}. Briefly, the four body system can be divided into two triples: an inner triple formed by $m_{\star 1}$-$m_{\star 2}$-$M_1$, and an outer triple formed by $M_1$-$M_2$-$(m_{\star 1},m_{\star 2})$. In each triple, the system can be either stable or unstable. In stable systems, secular effects like LK oscillations may accelerate the merger process. In unstable systems, there is a significant chance of a close interaction occurring between two or more of the components of the triple. The stability of each triple is expressed in terms of the stability factor $S$. For the inner triple we have
\begin{equation}
S_{\mathrm{in}} = \log\bigg(\frac{a_{\cm}(1-e_{\cm})}{a_{\star}(1+e_{\star})} / Y_\mathrm{in,crit}\bigg)
\end{equation}
where
\begin{equation}\label{eq:crit1}
Y_\mathrm{in,crit} = \frac{3.7}{\beta_\mathrm{out}}-\frac{2.2}{1+\beta_\mathrm{out}} + \frac{1.4}{\beta_\mathrm{in}}\frac{\beta_\mathrm{out}-1}{\beta_\mathrm{out}+1}
\end{equation}
\citep[e.g.,][]{1996ASPC...90..433K} with $\beta_\mathrm{in}=\mathrm{max}\bigg(\frac{m_{\star 1}}{m_{\star 2}}, \frac{m_{\star 2}}{m_{\star 1}}\bigg)^{1/3}$, and $\beta_\mathrm{out}=\bigg(\frac{m_{\star 1}+m_{\star 2}}{M_1}\bigg)^{1/3}$.
For the outer triple system we have
\begin{equation}
S_\mathrm{out} = \log\bigg(\frac{a_{\bh}(1-e_{\bh})}{a_{\cm}(1+e_{\cm})} / Y_\mathrm{out,crit} \bigg)
\end{equation}
where,
\begin{equation}\label{eq:crit2}
Y_\mathrm{out,crit} = \frac{3.7}{\zeta_\mathrm{out}}-\frac{2.2}{1+\zeta_\mathrm{out}} + \frac{1.4}{\zeta_\mathrm{in}}\frac{\zeta_\mathrm{out}-1}{\zeta_\mathrm{out}+1}
\end{equation}
with $\zeta_\mathrm{in}=\bigg(\frac{M_1}{m_{\star 1}+m_{\star 2}}\bigg)^{1/3}$, and $\zeta_\mathrm{out}=\bigg(\frac{m_{\star 1}+m_{\star 2}+M_1}{M_2}\bigg)^{1/3}$.
If $S_\mathrm{in,out}>0$, then the corresponding triple system is stable, otherwise it is unstable and tends to be disrupted.
We can thus divide the parameter space of the four-body system into four parts, where $(S_\mathrm{in}>0,S_\mathrm{out}>0)$, $(S_\mathrm{in}<0,S_\mathrm{out}>0)$, $(S_\mathrm{in}>0,S_\mathrm{out}<0)$ and $(S_\mathrm{in}<0,S_\mathrm{out}<0)$. For stable triple systems where $S_\mathrm{in,out}>0$, hierarchical LK oscillations can occur. In such systems, the long time scale secular perturbations induced by the outer orbit can drive the eccentricity of the inner orbit to extreme values, while simultaneously decreasing the inclination between the inner and outer orbital planes and conserving angular momentum.
\subsection{Classification of mergers in SMBHBs}
We now classify the different outcomes of the SMBHB-CO interactions with the aid of the stability map introduced in the previous section and visualized in Figure \ref{fig:stability}. In general, when $S_\mathrm{in}<0$, the tidal force exerted by the central SMBH is large enough to break apart the CO binary on a short timescale. {Those systems have been excluded from the initial conditions as we discussed above. Therefore, the legend 'decoupled' in \fig{fig:stability} indicates binary disruption from interaction with the secondary massive BH (i.e., disruption directly by the secondary massive BH or by the primary massive BH but as a result of a perturbation by the secondary massive BH)}. The right part of the stability map, where $S_\mathrm{in}>0$, is where all the interesting action occurs. Here, the interaction can lead to different flavors of ejected and merging systems, as we now explore in detail.
\subsubsection{Ejected mergers}
In the stability map where $S_\mathrm{out}<0$, the outer triple is unstable, indicating that the influence of $M_2$ severely perturbs the orbit of the centre of mass of the CO binary. In fact, in this case the CO binary can be significantly affected by close encounters with $M_2$, causing its ejection from the SMBHB sphere of influence with high velocity. Due to energy and angular momentum exchange with $M_2$, ejected CO binaries can get tightened enough to merge within a Hubble time. { Depending on the ejection velocity and on the depth of the potential well of the host galaxy, these CO binaries can merge in the outskirts of the dark matter halo of their host galaxy, or even in the IGM.}
This kind of merger event has two prerequisites: the CO binaries have to be ejected by the SMBHB (to produce HVBs), and the ejected CO binaries need to be tight enough to merge in a Hubble time due to GW radiation. To produce the HVBs, their ejection by the SMBHB must be efficient. This requires that the semi-major axis of the SMBHB is at least as small as $a_0$, and that the number density of CO binaries around $M_2$ is sufficiently large. At the same time, $S_\mathrm{out}$ needs to be smaller than zero to ensure that the CO binaries have a close encounter with $M_2$, while $S_\mathrm{in}$ needs to be larger than zero to ensure that the CO binaries will not be disrupted by $M_1$. To get compact HVBs that will merge within a Hubble time, the CO binaries need either to be sufficiently hard before the ejection or sufficient transfer of energy and angular momentum must occur from the binary's orbit to its centre of mass orbit around the SMBHB. Detailed descriptions of these processes can be found in \citet{2019MNRAS.482.3206W}. The distribution of ejected merger events in the stability map is shown in \fig{fig:stability} for all the simulated CO binary species.
\subsubsection{Quick mergers (i.e., collisions)}
In the inner triple system, LK cycles can directly create collisions between the components of the compact object binary if the eccentricity becomes sufficiently high. The typical (quadrupole) timescale for such an excitation cycle is \citep[e.g.,][]{1962P&SS....9..719L,1962AJ.....67..591K},
\begin{equation}\label{eq:LK}
\begin{split}
\tau_\mathrm{LK,in} &\sim \frac{1}{n_\star}\bigg( \frac{m_{\star 1}+m_{\star 2}}{M_1}\bigg) \bigg( \frac{a_{\cm}}{a_{\star}}\bigg)^3(1-e_{\cm}^2)^{3/2}\\
&=1.3\mathrm{Myr}\bigg( \frac{M_1}{M_6}\bigg)^{-1}\bigg( \frac{m_{\star 1}+m_{\star 2}}{M_\odot}\bigg)^{1/2}\\
&\times \bigg( \frac{a_{\cm}}{0.1\mathrm{pc}}\bigg)^{3}\bigg( \frac{a_{\star}}{\mathrm{au}}\bigg)^{-3/2}(1-e_{\cm}^2)^{3/2}\,,
\end{split}
\end{equation}
where $n_\star=\sqrt{G(m_{\star 1}+m_{\star 2})/a_{\star}^3}$ is the mean motion of the compact object binary. The strength of the next higher order LK effect (i.e., the octupole effect) is described by \citep[][]{2013MNRAS.431.2155N}
\begin{equation}\label{eq:oct_LK}
\epsilon_\mathrm{oct} = \frac{|m_{\star 1}-m_{\star 2}|}{m_{\star 1}+m_{\star 2}}\frac{a_{\star}}{a_{\cm}}\frac{e_{\cm}}{1-e_{\cm}^2}\,.
\end{equation}
Since the stability criterion in the inner triple requires the ratio ${a_{\star}}/{a_{\cm}}$ to be extremely small, the CO binary will not be disrupted by the SMBHB and the octupole effect is usually negligible.
To maintain the LK oscillations in the inner triple, the perturbations from the SMBH need to be stable. This requires that the centre of mass orbit of the CO binary will not change significantly during the inner LK cycles. Thus, the outer triple system also needs to be stable such that the perturbation from $M_2$ will not break the centre of mass orbit of the CO binary. In such a system, for which $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$, we find a relatively short corresponding value for $\tau_{LK,in}$, If and hence we frequently detect collisions in the simulations. Merger events from this channel occur on a short time scale in the simulations, especially for CO binaries with component stars that have the largest radii (i.e., WD-WD binaries). Thus, throughout this paper, we call this kind of merger 'quick mergers'.
\begin{figure*}
\includegraphics[width=\textwidth]{figs/stability}
\caption{Distribution of the possible outcomes of our simulations in terms of the stability map. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH. The marginalized distributions are normalized to 1 for visualization, thus they do not reflect the true event numbers of each event set. Also note that the ejected-merger is a subset of HVBs.}
\label{fig:stability}
\end{figure*}
\fig{fig:stability} shows the quick mergers in the stability map ($S_\mathrm{in}$,\,$S_\mathrm{out}$). We can clearly see that for different CO binaries (i.e., WD-WD, BH-BH and NS-NS), almost all quick mergers occur in the upper right quadrant where $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$. As expected, for the WD-WD binaries, the distribution for $S_\mathrm{in}$ is more dispersed. This is because the WD-WD binaries need only relatively low eccentricities to merge due to their larger radii, which requires only weaker LK oscillations that correspond to larger values for $S_{in}$ (larger ${a_{\cm}}/{a_{\star}}$ in \eqn{eq:LK}). A signature of quick merger events is that they have non-negligible eccentricities in the LIGO band. This is because the CO binaries directly collide in our simulations before circularization can occur due to GW radiation. Note, however, that LK oscillations around a single SMBH can also lead to eccentric LIGO events \citep[e.g.][]{2018ApJ...860....5G}; therefore, eccentric mergers should not be considered signposts of SMBHBs.
\subsubsection{Slow mergers}
Some CO binaries neither experience ejections by $M_2$ nor are subject to strong LK cycles. {This occurs for a subset of systems located within the stable region in our stability map} where $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$. Alternatively, they could remain bound to the SMBHB after a close encounter with $M_2$ in the region where $S_\mathrm{in}>0$ and $S_\mathrm{out}<0$. CO binaries in the region where $S_\mathrm{in}>0$ and $S_\mathrm{out}>0$ which experience negligible LK effects are in the ideal environment to allow for binary stellar evolution (BSE) to occur. Thus, some tight CO binaries could form from BSE via common envelope evolution (for example), and then merge due to GW radiation within {the cosmological time elapsed since their formation, i.e. the lookback time $\tau_{\rm lb}$}. CO binaries in the region where $S_\mathrm{in}>0$ and $S_\mathrm{out}<0$ which remain bound to the SMBHB after a close encounter with the secondary $M_2$ tend to become harder post-encounter. Therefore, a fraction of CO binaries in this region of phase space will be pushed into the merger band of LIGO by close encounters with $M_2$.
Despite the specific formation channel (BSE or encounter) for the tight CO binaries observed in the galactic centre, our simulations suggest a high formation rate if an IMBH is indeed present. Unlike quick mergers that are directly detected in the simulations, the 'slow' merger events will undergo significant circularization before merger due to GW radiation. Therefore, the eccentricity of the merger events for this channel will be completely negligible once in the LIGO band. Slow merger events are also shown in the stability map of \fig{fig:stability} (denoted with "S-merger").
\subsection{Outcomes of Scattering experiments}
\subsubsection{Relative fraction of HVBs}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/TV-distribution}
\caption{Contour plots of the 2D PDF $\int f_{\mathrm{hvb}}(m_\mathrm{chirp},v_\mathrm{ej},\tau) \mathrm{d} m_\mathrm{chirp}$ of HVBs, i.e. the normalized differential distribution of the HVBs as a function of $v_\mathrm{ej}$ and $\tau$. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:TV-dist}
\end{figure*}
From our scattering experiments, {for each value of} $M_1$ and $q$, the distribution of HVBs can be expressed as
\begin{equation*}
\frac{\mathrm{d}^4 N_\mathrm{hvb}}{\mathrm{d}a_{\star}\mathrm{d}e_{\star}\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej}}\,,
\end{equation*}
where $a_{\star}$, $e_{\star}$, $m_\mathrm{chirp}$ and $v_\mathrm{ej}$ are the semi-major axis, eccentricity, chirp mass and ejection velocity of the HVB, respectively. All these quantities are directly obtained from the scattering experiments.
After the ejection, due to GW radiation, some HVBs will merge within an Hubble time. The merger time scale due to GW radiation is given by \citep[][]{1964PhRv..136.1224P}
\begin{equation}
\tau_\mathrm{GW}(m_\mathrm{chirp},a_{\star},e_{\star}) = \frac{3}{85}\frac{c^5a_{\star}^4(1-e_{\star}^2)^{7/2}}{G^3m_{\star 1}m_{\star 2}(m_{\star 1}+m_{\star 2})}\,.
\end{equation}
Since not all HVBs will merge within {$\tau_{\rm lb}$}, we need to select only those HVBs whose merger time $\tau_\mathrm{GW}$ is shorter than {$\tau_{\rm lb}$}. Previous works \citep[][]{2019MNRAS.482.3206W} have shown that from the close interaction with the SMBHB, many of the ejected binaries tend to become harder after the ejection due to energy and angular momentum exchange between the IMBH and the compact binary. Thus, we expect a significant fraction of the ejected HVBs to rapidly escape the SMBHBs and eventually merge (within $\tau_{\rm lb}$) in the dark matter halo of their host galaxy or even in the IGM. To compute the distribution of merger times due to GW emission, $\tau_\mathrm{GW}$, we have to convert the distribution of HVBs to the $(m_\mathrm{chirp},v_\mathrm{ej},\tau)$-space by integrating over the parameter $\tau$,
\begin{equation}
\begin{split}
&f_{\mathrm{hvb}}(m_\mathrm{chirp},v_\mathrm{ej},\tau)=\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej} \mathrm{d} \tau}\\
&=\frac{\mathrm{d}}{\mathrm{d}\tau}\oiint_{\tau_\mathrm{GW}(m_\mathrm{chirp},a_{\star},e_{\star})<\tau}\frac{\mathrm{d}^4 N_\mathrm{hvb}}{\mathrm{d}a_{\star}\mathrm{d}e_{\star}\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej}}\mathrm{d}a_{\star}\mathrm{d}e_{\star}\,.\\
\end{split}
\end{equation}
Figure \ref{fig:TV-dist} shows $f_{\mathrm{hvb}}(m_\mathrm{chirp},v_\mathrm{ej},\tau)$,
which is the 2D-PDF of the hypervelocity binaries as a function of the ejection velocity $v_\mathrm{ej}$ and the GW merger time $\tau_{\rm GW}$. Clearly, for each type of CO binary, we get some HVBs for which the final velocity is high enough relative to the local escape speed that these objects are expected to escape from the sphere of influence of their host SMBHBs. After applying a specific galactic potential, one can compute whether or not a given binary will merge within the dark matter (DM) halo of their host galaxy or the IGM. { Detailed calculations regarding the cosmic rate of this type of merger will be discussed in \sect{sec:cosmic}.}
\subsubsection{Relative fraction of quick merger events}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/q-merger_mq}
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/q-merger_ae_cm}
\caption{The 2D PDF of the relative fraction of quick mergers in the $M_1$-$q$ (top series of panels) and $a_{\cm}/a_{\bh}$-$a_{\star}/a_{\cm}$ (bottom series of panels) planes. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:quick_rel}
\end{figure*}
Similarly, we can examine the properties of quick merger events. Those are shown in \fig{fig:quick_rel}. The upper panels show, for the various types of CO binaries, the distribution
\begin{equation*}
\frac{1}{N_\mathrm{quick}}\frac{\mathrm{d}^2 N_\mathrm{quick}}{\mathrm{d}\log M_1\mathrm{d}\log q}\,,
\end{equation*}
{which indicates that the relative rate of quick mergers is independent of the mass ratio of the SMBHB and has a large probability for SMBHBs with large $M_1$.}
{Indeed, the quick merger rate depends on the time scale of the LK oscillations from the primary massive BH which is more sensitive to the location of the CO binaries in the SMBHB potential well ($a_{\cm}/a_{\bh}$), as well as the mass of $M_1$.} The bottom panels show the distribution
\begin{equation*}
\frac{1}{N_\mathrm{quick}}\frac{\mathrm{d}^2 N_\mathrm{quick}}{\mathrm{d}e_{\cm}\mathrm{d}\log a_{\cm}}\,.
\end{equation*}
When BH-BH, NS-NS and BH-NS binaries are involved, quick mergers tend to occur when $a_{\cm}/a_{\bh}\lesssim 10^{-1}$. To produce quick mergers, in fact, the LK-effect in the inner triple must be very efficient. The time scale for the leading quadrupole and octupole LK-effects are given by, respectively, \eqn{eq:LK} and \eqn{eq:oct_LK}. In this region, the strong LK oscillations efficiently drive the CO binary orbital eccentricity to extremely high values and cause a direct collision. The bottom two panels show that quick mergers in WD-WD binaries and EBBHs can also frequently occur when $a_{\rm cm}/a_0>1$. This is because WDs are much larger in size than either NSs or BHs, and binaries containing them can merge at lower eccentricities. EBBHs themselves are much more eccentric, and thus need less help from LK oscillations. Therefore, only a moderate LK effect is required, which can occur farther from the central SMBH.
From our previous stability analysis, we know that quick mergers correspond mostly to the region $S_\mathrm{in}~>~0$ and $S_\mathrm{out}~>~0$, for which both the inner and outer triple systems form stable hierarchical systems. In \fig{fig:stability}, we see that, as $S_\mathrm{in}$ approaches zero from above (indicating that the perturber comes closer to the compact object binary) the quick merger rate increases significantly. This is explained by the stronger inner LK effect due to a closer perturber. As $S_\mathrm{in}$ crosses zero, the perturbation from $M_1$ becomes so strong that the inner triple becomes unstable, leading to binary disruption. With that said, this is not seen in the stability map directly due to the 'hard cut' imposed in our initial condition at $S_\mathrm{in}=0$. Note that, since the WD radii are much larger than those of NSs and BHs and the EBBHs are very eccentric from the beginning, quick mergers of binary WD-WD and EBBHs require weaker LK oscillations, which is reflected by a wider dispersion in $S_\mathrm{in}$ in the bottom panels of Figure \ref{fig:stability}.
Note that, in our simulations, the dissipative 2.5PN term is not included {due to the fact that the computation becomes prohibitively expensive for very tight CO binaries}. This narrows the eccentricity window in our collision detections. Eccentricities lower than what required for collisions to occur (but still extremely high) can generate quick mergers due to the high efficiency of GW radiation at very high $e_{\star}$. These GW-induced mergers could have non-negligible eccentricities in the LIGO band, as was found in \citet[][]{2018PhRvD..97j3014S, 2012ApJ...757...27A}. Therefore, the quick merger rate obtained from our simulations corresponds to a lower limit.
\subsubsection{Relative fraction of slow merger events}
As for {ejected} mergers, we obtain the relative fraction of slow merger events as described in the following section.
From our scattering experiments, we obtain a number of unresolved interactions. After selecting those for which $\tau_\mathrm{GW} < \tau_\mathrm{hubble}$, we divide the results into two parts, namely $S_\mathrm{out}>0$ and $S_\mathrm{out}<0$. These two subsets correspond to the two different cases described above:
binaries that could be formed from BSE and undergo long term LK oscillations from $M_1$, or {tight binaries that remain bound to the SMBHB after close encounters with $M_2$}. The two subsets can clearly be seen from the lower panels of \fig{fig:incenter_rel}, which shows the 2D-PDF of the slow merger rate in the ($M_1$, $q$) and ($e_{\cm}$, $a_{\cm}$) planes, namely
\begin{equation*}
\frac{1}{N_\mathrm{slow}}\frac{\mathrm{d}^2 N_\mathrm{slow}}{\mathrm{d}\log M_1\mathrm{d}\log q}\,
\end{equation*}
and
\begin{equation*}
\frac{1}{N_\mathrm{slow}}\frac{\mathrm{d}^2 N_\mathrm{slow}}{\mathrm{d}e_{\cm}\mathrm{d}\log a_{\cm}}\,.
\end{equation*}
The upper sets of panels in \fig{fig:incenter_rel} show that slow mergers from LK oscillations rarely occur for SMBHBs with mass ratios larger than 0.1.
The two subsets of events with $S_\mathrm{out}>0$ and $S_\mathrm{out}<0$ can clearly be recognized in the bi-modal distributions of the lower set of panels. The mode peaking at $a_{\cm} \sim [0.001, 0.1] a_{\bh}$ corresponds to stable systems with $S_\mathrm{out}>0$, whereas the mode extending at $a_{\cm} > a_{\bh}$ corresponds to unstable systems with $S_\mathrm{out}<0$. Note that since EBBHs merge much faster than the other four types of CO binaries in the region $a_{\cm} \sim [0.001, 0.1] a_{\bh}$, the distribution of slow mergers of EBBHs is dominated by the mode at $a_{\cm} > a_{\bh}$.
\begin{figure*}
\includegraphics[width=\textwidth]{figs/s-merger_mq}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/s-merger_ae_cm}\\
\caption{Same as \fig{fig:quick_rel} but for the relative fraction of 'slow mergers'. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:incenter_rel}
\end{figure*}
\section{Cosmic rate of merger events from SMBHBs}\label{sec:cosmic}
{We now combine the outcomes of our scattering experiments with the SMBHB population extracted from the Millennium-II simulation, to estimate the cosmic merger rates of different types of mergers produced by SMBHB-CO binary interactions. Before proceeding, however, we shall stress an important caveat. Our simulations are meant to be representative of the dynamical interactions occurring as the SMBHB erodes the cusp of bound stars surrounding it as it first becomes bound. The cusp erosion process shrinks the binaries by a factor of $\approx 10$ starting from $a_0$, which is given by \eqn{eq:a0} if we assume that the cusp distribution follows an isothermal sphere. This might not be sufficient to efficiently drive the SMBHB to coalescence, and further shrinking, via the interaction of unbound stars scattered in the SMBHB loss cone, might be required. In general, shrinking by only a few e-folds is needed \citep[e.g.,][]{2010ApJ...719..851S} and each e-fold requires the interaction with a mass in stars of the order of $M_2$ \citep[see, e.g.,][]{2005LRR.....8....8M}. Since the erosion of the bound cusp ejects a mass in stars of order $\approx M_2$, the rates estimated in the following should be considered as conservative, and might be higher by up to a factor of a few.}
{We caution, however, that interaction with unbound binaries is inherently different than for bound binaries. Most stars, especially for SMBHB mass ratios close to unity, are either ejected in a single fly-by, or captured in relatively short-lived metastable orbits. LK oscillations are therefore irrelevant in this case, and the rate of quick and slow mergers should be strongly suppressed. On the other hand, CO binaries can still be ejected as HVBs, but the ejection velocities tend to be less extreme than those achieved in bound scattering. Unbound SMBHB-CO binary scattering experiments will eventually be needed to quantify the contribution of this late phase of the SMBHB evolution to the cosmic rate of ejected CO binary mergers. This is deferred to future work.}
\subsection{Ejected merger rate}
From the scattering experiments, we have obtained the HVB distribution ${\mathrm{d}^3 N_\mathrm{hvb}}/({\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej} \mathrm{d} \tau})$. Note that this distribution varies with $M_1$ and $q$, thus it is a function of $m_\mathrm{chirp}$, $\tau$ and $v_{\rm eject}$ but also of $\log M_1$ and $\log q$. After the ejection, a given HVB will escape from the galactic potential with initial velocity $v_\mathrm{ej}$ and merge in the dark matter halo or in the IGM at a distance $R$ away from the host galactic centre. This distance depends on the flight time $\tau$, and can be computed explicitly by assuming a specific galactic potential. In this work, we adopt the characteristic galactic potential of elliptical galaxies \citep[e.g.,][]{2008ApJ...680..312K,2014ApJ...793..122K},
\begin{equation}
\begin{split}
\phi(r)&=\phi_\mathrm{bh}(r)+\phi_b(r) + \phi_h(r)\\
&= -\frac{G(M_1+M_2)}{r} -\frac{GM_b}{r+r_b} - \frac{GM_h}{r}\ln\left(1+\frac{r}{r_h}\right)\,,
\end{split}
\label{eq:pot}
\end{equation}
where $M_b$ is the mass of the bulge, $M_h$ is the mass of the dark matter halo, $r_b$ is the scale radius of the bulge and $r_h$ is the scale radius of the dark matter halo. The choice of the spherical potential in equation \eqref{eq:pot} is dictated by computational convenience, as it allows us to perform 1D calculations to compute the ejected CO binary trajectory. The addition of a disk component would produce a deflection of the ejected trajectories, without significantly affecting the overall distribution of distances travelled. We scale the variables $M_b$, $M_h$, $r_b$ and $r_h$ with the total mass of the SMBHB $M_1(1+q)$ by means of the relations \cite[e.g.,][]{2013ARA&A..51..511K, 2015MNRAS.452.4013K}
\begin{equation}
\begin{aligned}
&\frac{M_1(q+1)}{10^9M_\odot}&=&\,0.49\left(\frac{M_b}{10^{11}M_\odot}\right)^{1.16}\,,\\
&\frac{M_h}{10^{13}M_\odot}&=&\,0.30\left(\frac{M_b}{10^{11}M_\odot}\right)\,,\\
&\frac{M_b}{10^6 M_\odot}\bigg(\frac{\mathrm{kpc}}{r_b}\bigg)^3&=&\, 4000\,,\\
&\frac{M_h}{10^6 M_\odot}\bigg(\frac{ \mathrm{kpc}}{r_h}\bigg)^3&=&\,125\,.
\end{aligned}
\end{equation}
The ejection properties are recorded at a distance of $r\,\sim\,50a_{\bh}$, given the condition that the total energy of the compact binary must be positive. Thus, we can calculate the traveled distance $R=D(M_1(1+q),\tau,v_\mathrm{ej})$
\begin{equation}
\int_0^D\frac{dr}{\sqrt{2E_0-2\phi(r)- 2l_0^2/r^2}} =\int_0^{\tau_\mathrm{GW}} dt\,,
\end{equation}
where $E_0$ and $l_0$ are, respectively, the total energy and total angular momentum per unit mass at the time of ejection. The distribution of HVBs as a function of mass, traveling time $\tau$ and distance from the galaxy centre at a merger distance $R$ can then be calculated as
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} \tau\mathrm{d} R}(\log q,\log M_1,m_\mathrm{chirp},\tau,R)\\
&=\frac{\mathrm{d}}{\mathrm{d} R}\oint_{D(M_1,q,\tau,v_\mathrm{ej}) < R}\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} v_\mathrm{ej}\mathrm{d} \tau}\mathrm{d} v_\mathrm{ej}\\
\end{split}\,.
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/eRV_dist}\\
\caption{{The 2D-PDF of the ejected merger events as a function of the merger distance $R$ from their host galaxy, as well as the corresponding CoM velocity when they merge}.\textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH. }
\label{fig:RV}
\end{figure*}
\fig{fig:RV} shows the 2D-PDF of the distance $R$ and the corresponding velocity at that distance of the simulated HVBs. Note that not all ejected mergers can be observed within the age of the universe.
The results of our scattering experiments yield the time at which the SMBHB ejects the CO binaries, which corresponds to the redshift $z$. Merger events occur a time $\tau$ after the ejection. Therefore, there is a significant displacement of the merger redshift with respect to $z$. Only those events with $\tau\lesssim \tau_{\rm lb}$ can be observed; let us define the merger redshift to be $z^\prime=g(z,\tau)$, where
\begin{equation}
\begin{split}
\tau &= \int_{z^\prime}^z \frac{\mathrm{d} t}{\mathrm{d} \bar{z}}\mathrm{d} \bar{z}\\
&= \frac{1}{H_0}\int_{z^\prime}^z\frac{\mathrm{d} \bar{z}}{(1+\bar{z})\sqrt{\Omega_M(1+\bar{z})^3+\Omega_k(1+\bar{z})^2+\Omega_\Lambda}}\,.\\
\end{split}
\end{equation}
We solve this equation numerically to obtain the value of $z^\prime$ for each given $z$ and $\tau$. With this relationship, we can calculate the redshift distribution of the ejected mergers. Only mergers with corresponding redshift $z^\prime > 0$ can be observed today.
To compute the cosmic ejected merger rate in the rest frame of the Earth, we need to convolve the fraction of {HVBs} obtained from the scattering experiments with the cosmic distribution of compact binaries actively interacting with SMBHBs per unit time. As discussed in Section \ref{sec:intrate}, we compute the distribution of the actively interacting mass per unit time via \eqn{eq:Mactive};
thus it is easy to obtain the cosmic distribution for the actively interacting binary number per unit time,
\begin{eqnarray}
\dot{n}_\mathrm{act}(z,\log q,\log M_1)&=&\frac{\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\\
&=&\frac{\mathrm{d}^4 (n_\mathrm{bi} M_\mathrm{act})}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\\
&=&n_{\mathrm{bi}}\dot{\mathcal{M}}_\mathrm{act}(z,\log q,\log M_1)\,,
\end{eqnarray}
where $n_\mathrm{bi}$ is the number of compact binaries (BHB, WDB, NSB) per unit mass in the galactic centre, which can be estimated via consideration of stellar evolution with a given IMF, and the binary fraction typical of galactic centres.
From our scattering experiments, {for each} value of $M_1$ and $q$, we obtain the HVB distribution ${\mathrm{d}^3 N_\mathrm{hvb}}/({\mathrm{d} m_\mathrm{chirp}\mathrm{d} \tau\mathrm{d} R})$ as a function of the parameters $m_\mathrm{chirp}$, $\tau$ and $R$. Note that this distribution varies with $M_1$ and $q$, thus it is not only a function of $m_\mathrm{chirp}$, $\tau$ and $R$ but also a function of $\log M_1$ and $\log q$. Therefore, the cosmic rate of HVBs can be calculated as
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^4N_\mathrm{eject}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t}(z^\prime, m_\mathrm{chirp}, R)\\
&=\theta(z^\prime)\frac{\mathrm{d}}{\mathrm{d} z^\prime}\int_{g(z,\tau)<z^\prime}\mathrm{d}\tau\mathrm{d} z\int\frac{\mathrm{d}^3 N_\mathrm{hvb}}{\mathrm{d} m_\mathrm{chirp}\mathrm{d} \tau\mathrm{d} R}\\
&\times\frac{\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\frac{1}{N_{\mathrm{exp}}(\log q,\log M_1)}\mathrm{d} \log M_1 \mathrm{d}\log q\,,\\
\end{split}
\end{equation}
where $N_{\mathrm{exp}}$ is the total number of scattering experiments performed for each $M_1$ {and $\theta(z^\prime)$ selects out the merger events that can be observed to date ($z^\prime\ge 0$).}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/emerger_RM}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/emerger_RZ}
\caption{Cosmic distribution of ejected merger events and their dependences on the parameters $R$, $m_\mathrm{chirp}$ and $z^\prime$ normalized by the undetermined parameter $n_\mathrm{bi}$, where $n_\mathrm{bi}$ is the CO binary number per unit solar mass in the galactic centre. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:cosmic_rate}
\end{figure*}
The upper panels of \fig{fig:cosmic_rate} show the distribution of the rate of cosmic '{ejected}' mergers {that can be observed up to date, i.e. ${\mathrm{d}^3N_\mathrm{eject}}/({\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t})$}. For all types of CO binaries, we find merger events occurring both in dark matter halos and in the interstellar and intergalactic media ($\unit{kpc-Mpc}$). There is a peak in $R$ around $\unit{Mpc}$ for WD-WD, BH-BH, NS-NS and BH-NS binaries, indicating that those systems can merge deep in the IGM. Conversely, EBBHs preferentially merge at $R<10^{-2}\unit{Mpc}$, due to their much shorter GW radiation merger times. The bottom subplots of \fig{fig:cosmic_rate} show the distribution of the cosmic {ejected} merger rate ${\mathrm{d}^3N_\mathrm{eject}}/({\mathrm{d} z^\prime\,\mathrm{d} R\,\mathrm{d} t})$. We see that most of the {ejected} mergers should be detected in the range $z^\prime \sim [0.5,3]$.
By integrating ${\mathrm{d}^4N_\mathrm{eject}}/({\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t})$ we get
\begin{equation}
\Gamma_\mathrm{eject} = \frac{\mathrm{d} N_\mathrm{eject}}{\mathrm{d} t} = \int \frac{\mathrm{d}^4N_\mathrm{eject}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t}\mathrm{d} z^\prime \mathrm{d} R\mathrm{d} m_\mathrm{chirp}
\end{equation}
which is the {ejected} merger event rate up to redshift $5$. The numerical values of the merger rate can be found in \tab{tab:R}, where $n_\mathrm{wd-wd}$, $n_\mathrm{ns-ns}$, $n_\mathrm{bh-bh}$, $n_\mathrm{bh-ns}$ and $n_\mathrm{ebbh}$ are the binary number per unit solar mass in the galactic centre for WD-WD, NS-NS, BH-BH, BH-NS and EBBH, respectively. {In the table, a subset of WD-WD mergers for which the total mass $>1.4 M_\odot$ has been listed separately. We simply label these as Type Ia SNe.}
To assess the detectability of emitted GWs, we must consider the signal-to-noise ratio (SNR) of the GW detectors that can detect the events out to $z^\prime=f(m_\mathrm{chirp})$. With this, the detectable {ejected} merger event rate as a function of $R$ and $z^\prime$ can be derived from
\begin{equation}
\frac{\mathrm{d}^3N_\mathrm{eject}}{\mathrm{d} R\mathrm{d} z^\prime\mathrm{d} t}(R,z^\prime)=\int_{f(m_\mathrm{chirp})<z^\prime}\frac{\mathrm{d}^4N_\mathrm{eject}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} R\,\mathrm{d} t}\mathrm{d} m_\mathrm{chirp}\,.
\end{equation}
\subsection{Quick merger rate}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/qmerger_RM}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/qmerger_RZ}
\caption{Cosmic distribution of quick merger events and their dependences on the parameters $R$, $m_\mathrm{chirp}$ and $z^\prime$ normalized by the undetermined parameter $n_\mathrm{bi}$, where $n_\mathrm{bi}$ is the CO binary number per unit solar mass in the galactic centre. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:cosmic_quick_rate}
\end{figure*}
Following the same procedure as for {ejected} merger events, we can calculate the cosmic rate of quick merger events as follows,
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^3N_\mathrm{quick}}{\mathrm{d} z^\prime,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}(z^\prime, m_\mathrm{chirp})\\
&=\int N_{\mathrm{quick}}\frac{\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\frac{\delta(z-z^\prime)}{N_{\mathrm{exp}}(\log q,\log M_1)}\mathrm{d} \log M_1 \mathrm{d}\log q\,,
\end{split}
\end{equation}
where $N_{\mathrm{quick}}$ is verified to be independent of $M_1$ and $q$ in \sect{sec:relfraction}.
\fig{fig:cosmic_quick_rate} shows the distribution of ${\mathrm{d}^3N_\mathrm{quick}}/({\mathrm{d} z^\prime,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t})$; the corresponding rates can be obtained by
\begin{equation}
\frac{\mathrm{d} N_\mathrm{quick}}{\mathrm{d} t} = \int \frac{\mathrm{d}^3N_\mathrm{quick}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}\mathrm{d} z^\prime \mathrm{d} m_\mathrm{chirp}
\end{equation}
The numerical values of the merger rate can be found in \tab{tab:R}.
\subsection{Slow merger rate}
As for the quick mergers, we estimate the cosmic rate of {slow} mergers as
\begin{equation}
\begin{split}
&\frac{\mathrm{d}^3N_\mathrm{slow}}{\mathrm{d} z^\prime,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}(z^\prime, m_\mathrm{chirp})\\
&=\int \frac{\delta(z-z^\prime)\mathrm{d}^4 N_\mathrm{act}}{\mathrm{d} z\mathrm{d}\log q\mathrm{d}\log M_1\mathrm{d} t}\frac{N_{\mathrm{slow}}(\log M_1,\log q)}{N_{\mathrm{exp}}(\log q,\log M_1)}\mathrm{d} \log M_1 \mathrm{d}\log q\,.
\end{split}
\end{equation}
\begin{figure*}
\includegraphics[width=\textwidth]{figs/smerger_RM}\\
\rule{\textwidth}{0.02cm}
\includegraphics[width=\textwidth]{figs/smerger_RZ}
\caption{Cosmic distribution of slow merger events and their dependences on the parameters $R$, $m_\mathrm{chirp}$ and $z^\prime$ normalized by the undetermined parameter $n_\mathrm{bi}$, where $n_\mathrm{bi}$ is the CO binary number per unit solar mass in the galactic centre. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:cosmic_in_rate}
\end{figure*}
\fig{fig:cosmic_in_rate} shows the distribution ${\mathrm{d}^3N_\mathrm{slow}}/({\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t})$, which is almost identical to the quick merger distribution. However, the non-negligible residual eccentricities for quick mergers determine the region in phase-space where the merger events tend to occur, as indicated by \fig{fig:incenter_rel} and \fig{fig:quick_rel}. The corresponding rates are computed via the relation
\begin{equation}
\frac{\mathrm{d} N_\mathrm{slow}}{\mathrm{d} t} = \int \frac{\mathrm{d}^3N_\mathrm{slow}}{\mathrm{d} z^\prime\,\mathrm{d} m_\mathrm{chirp}\,\mathrm{d} t}\mathrm{d} z^\prime \mathrm{d} m_\mathrm{chirp}
\end{equation}
The numerical values of the merger rate can be found in \tab{tab:R}.
\subsection{Merger distance from the galactic centre}
Due to differences in the surrounding environments, merger events occurring at different distances from the galactic centre are likely to show different signatures in their electromagnetic (EM) counterparts, as discussed in the following subsection. {Quick mergers} induced by very strong LK oscillations are naturally produced close to the SMBHB. However, the post-encounter, {slow mergers} created by the secondary SMBH and {ejected mergers} may occur in the outskirts or even outside the host galaxy.
\begin{figure*}
\includegraphics[width=\textwidth]{figs/R_dist_all}
\caption{The distribution of the merger distance from the galactic centre $R$ for different types of mergers. \textit{Top Left} panel: BH-BH, \textit{Top middle} panel: NS-NS, \textit{Top right} panel: BH-NS, \textit{Bottom left} panel: WD-WD. \textit{Bottom right} panel: EBBH.}
\label{fig:r_dist}
\end{figure*}
\fig{fig:r_dist} shows the merger distances of quick mergers, {slow} mergers and {ejected} mergers. Independent of the CO species, quick mergers are found within $100~\mathrm{pc}$. Conversely, all {ejected} mergers will be found at $>10^2\mathrm{pc}$. As a representative subdivision, we consider $R<10^2\mathrm{pc}$, $10^2\mathrm{pc}<R<10^5\mathrm{pc}$ and $R>10^5\mathrm{pc}$, which correspond to, respectively, the galactic centre, the galactic field and the IGM. The merger rates of different CO binaries in the three regions of parameter space are given in \tab{tab:R}. The parameter $n_{bi}$ quantifies the number of CO binaries per unit mass, and can be determined by combining a specific stellar IMF, with stellar evolution models and empirical estimates for the binary fraction in galactic nuclei.
\begin{table}
\begin{tabular} { | l | l | l | l |}
\hline
& $R < 10^2\mathrm{pc}$ & $R \in [10^2,10^5]\mathrm{pc}$ & $R > 10^5\mathrm{pc}$ \\
\hline
\textit{Ejected}:& & &\\
BH-BH & $1.25 \times 10^{1} n_\mathrm{bh-bh}$ & $7.40 \times 10^{2} n_\mathrm{bh-bh}$& $8.96 \times 10^{2} n_\mathrm{bh-bh}$\\
NS-NS & $7.65 \times 10^{1} n_\mathrm{ns-ns}$ & $3.87 \times 10^{2} n_\mathrm{ns-ns}$& $6.13 \times 10^{2} n_\mathrm{ns-ns}$\\
BH-NS & $3.72 \times 10^{1} n_\mathrm{bh-ns}$ & $4.56 \times 10^{2} n_\mathrm{bh-ns}$& $7.20 \times 10^{2} n_\mathrm{bh-ns}$\\
WD-WD & $3.87 \times 10^{0} n_\mathrm{wd-wd}$ & $1.41 \times 10^{2} n_\mathrm{wd-wd}$& $2.09 \times 10^{2} n_\mathrm{wd-wd}$\\
Ia SNe & $3.87 \times 10^{0} n_\mathrm{wd-wd}$ & $1.43 \times 10^{1} n_\mathrm{wd-wd}$& $1.83 \times 10^{1} n_\mathrm{wd-wd}$\\
EBBH & $2.47 \times 10^{3} n_\mathrm{bh-bh}$ & $4.02 \times 10^{3} n_\mathrm{bh-bh}$& $1.38 \times 10^{3} n_\mathrm{bh-bh}$\\
\hline
\textit{Quick}:& & &\\
BH-BH & $2.83 \times 10^{4} n_\mathrm{bh-bh}$ & $\times$ & $\times$\\
NS-NS & $7.33 \times 10^{4} n_\mathrm{ns-ns}$ & $\times$ & $\times$\\
BH-NS & $2.30 \times 10^{4} n_\mathrm{bh-ns}$ & $\times$ & $\times$\\
WD-WD & $6.45 \times 10^{4} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
Ia SNe & $8.72 \times 10^{3} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
EBBH & $1.09 \times 10^{5} n_\mathrm{bh-bh}$ & $\times$ & $\times$\\
\hline
\textit{Slow}:& & &\\
BH-BH & $6.63 \times 10^{4} n_\mathrm{bh-bh}$ & $\times$ & $\times$\\
NS-NS & $2.77 \times 10^{4} n_\mathrm{ns-ns}$ & $\times$ & $\times$\\
BH-NS & $3.82 \times 10^{4} n_\mathrm{bh-ns}$ & $\times$ & $\times$\\
WD-WD & $8.42 \times 10^{3} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
Ia SNe & $1.30 \times 10^{3} n_\mathrm{wd-wd}$ & $\times$ & $\times$\\
EBBH & $3.08 \times 10^{4} n_\mathrm{bh-bh}$ & $7.73 \times 10^{1} n_\mathrm{bh-bh}$ & $\times$\\
\hline
\textit{Total}:& & &\\
BH-BH & $9.47 \times 10^{4} n_\mathrm{bh-bh}$ & $7.39 \times 10^{2} n_\mathrm{bh-bh}$& $8.96 \times 10^{2} n_\mathrm{bh-bh}$\\
NS-NS & $1.01 \times 10^{5} n_\mathrm{ns-ns}$ & $3.86 \times 10^{2} n_\mathrm{ns-ns}$& $6.12 \times 10^{2} n_\mathrm{ns-ns}$\\
BH-NS & $6.12 \times 10^{4} n_\mathrm{bh-ns}$ & $4.56 \times 10^{2} n_\mathrm{bh-ns}$& $7.20 \times 10^{2} n_\mathrm{bh-ns}$\\
WD-WD & $7.29 \times 10^{4} n_\mathrm{wd-wd}$ & $1.41 \times 10^{2} n_\mathrm{wd-wd}$& $2.09 \times 10^{2} n_\mathrm{wd-wd}$\\
Ia SNe & $1.00 \times 10^{4} n_\mathrm{wd-wd}$ & $1.43 \times 10^{1} n_\mathrm{wd-wd}$& $1.83 \times 10^{1} n_\mathrm{wd-wd}$\\
EBBH & $1.42 \times 10^{5} n_\mathrm{bh-bh}$ & $4.10 \times 10^{3} n_\mathrm{bh-bh}$& $1.38 \times 10^{3} n_\mathrm{bh-bh}$\\
\hline
\hline
\end{tabular}
\caption{{Model-independent rate of merger events classified by $R$, where $n_\mathrm{wd-wd}$, $n_\mathrm{ns-ns}$, $n_\mathrm{bh-bh}$ and $n_\mathrm{bh-ns}$ are the binary number per unit solar mass in the galactic centre for WD-WD, NS-NS, BH-BH and BH-NS binaries, respectively. The units for all entries are in $M_\odot\,\mathrm{yr}^{-1}$}.}
\label{tab:R}
\end{table}
The first step in the determination of $n_{bi}$ is to estimate the fraction of single WDs, NSs, and BHs in the galactic nuclei, which requires the evolution of the stellar population from a specific IMF. Since we use the outcome of the {\tt StarTrack} code for our scattering experiments, we follow the same procedure as \citet[][]{2012ApJ...749...91F}. It is generally assumed that the remnant mass is directly determined by the core mass $M_{\rm core}$. Here, we map the IMF to the core mass by using \citep[e.g.,][]{2015MNRAS.451.4086S},
\begin{equation}
M_{\rm core} = -2 + (B + 2)(g(Z, M_\star; K_1, \delta_1) + g(Z, M_\star; K_2, \delta_2))
\label{eq:mcore}
\end{equation}
where $Z$ is the metallicity, $M_\star$ is the initial stellar mass and
\begin{equation}
\begin{aligned}
&g(Z, m; x, y) = \frac{0.5}{1+ 10^{(x(Z)-m)y(Z)}}\\
&B=40.98 + 3.415\times 10^4Z - 8.064\times 10^6 Z^2\\
&K_1 = 35.17 + 1.548\times 10^4Z - 3.759\times 10^6 Z^2\\
&k_2 = 20.36 + 1.162\times 10^5Z - 2.276\times 10^7 Z^2\\
&\delta_1 = 2.5\times 10^{-2} -4.346Z +1.34\times 10^3 Z^2\\
&\delta_2 = 1.75\times 10^{-2} + 11.39Z - 2.902\times 10^3 Z^2
\end{aligned}
\end{equation}
for $Z\in[0.1, 0.4] Z_\odot$. The relationship between the remnant mass and the core mass is
\begin{equation}
M_{\rm rem} =
\left\{
\begin{aligned}
&0.619 M_\odot, \quad M_{\rm core} \leq M_{\rm NS,core}^{\rm lower}&\\
&\mathrm{max}\{h(M_{\rm core}, Z), 1.27M_\odot\}, \quad M_{\rm NS,core}^{\rm lower} < M_{\rm core} \leq 5\, M_\odot&\\
&h(M_{\rm core}, Z),\quad 5\, M_\odot < M_{\rm core} \leq 10\, M_\odot&\\
&\mathrm{max}\{h(M_{\rm core}, Z), f(M_{\rm core}, Z)\}, \quad 10\, M_\odot < M_{\rm core}&
\end{aligned}
\right.
\end{equation}
where
\begin{eqnarray}
h(M_{\rm core}, Z) &=& A_1(Z) + \frac{A_2(Z)- A_1(Z)}{1 + 10 ^{ (L(Z) - M_{\rm core})\eta(Z) }}\\
f(M_{\rm core}, Z) &=& m(Z)M_{\rm core} + q(Z)\\
m(Z) &=& -6.476\times 10^{2}Z + 1.911\\
q(Z) &=& 2.300\times 10^{3}Z + 11.67\\
A_1(Z) &=& 1.340 - \frac{29.46}{ 1 + (\frac{Z}{1.110\times 10^{-3}})^{2.361}}\\
A_2(Z) &=& 80.22 - 74.73\frac{Z^{0.965}}{2.700\times 10^{-3} + Z^{0.965} }\\
L(Z) &=& 5.683 + \frac{3.533}{ 1 + (\frac{Z}{7.43\times 10^{-3}})^{1.993}}\\
\eta(Z) &=&1.066 - \frac{1.121}{ 1 + (\frac{Z}{2.558\times 10^{-2}})^{0.609}}
\end{eqnarray}
$M_{\rm NS,core}^{\rm lower}$ is the lower mass limit for the NS core, found by substituting $M_\star$ with
\begin{equation}
M_{\rm NS}^{\rm lower} = 9M_\odot + 0.9\log_{10}(Z/Z_\odot)
\end{equation}
in equation \eqref{eq:mcore} \citep[e,g,.][]{2012ApJ...749...91F}.
If, we take the Salpeter IMF \citep[e.g.,][]{1955ApJ...121..161S}
\begin{equation}
\xi(m)dm \propto \bigg(\frac{m}{M_\odot}\bigg)^{-2.35}dm
\end{equation}
with $m_{\rm min} = 0.4\,M_\odot$, $m_{\rm max} = 150\,M_\odot$ and $Z=0.1Z_\odot$ as typically used in {the initial binary populations from {\tt StarTrack}}, the stellar evolution process yields $1.19$ WDs per unit solar mass and $2.04\times 10^{-2}$ COs per unit solar mass.
The {\tt StarTrack} code gives the BH-BH, NS-NS and BH-NS number fractions among all CO binaries to be $50\%$, $39.7\%$ and $10.3\%$, respectively. Therefore, if we take the binary number fraction in galactic nuclei to be 0.1
as a representative value, we find $n_\mathrm{wd-wd}=5.9\times 10^{-2} \,M_\odot^{-1}$, $n_\mathrm{bh-bh}=1.0\times 10^{-3}\,M_\odot^{-1}$, $n_\mathrm{ns-ns}=8.1\times 10^{-4}\,M_\odot^{-1}$ and $n_\mathrm{bh-ns}=2.1\times 10^{-4}\,M_\odot^{-1}$. The corresponding cosmic merger rates are reported in \tab{tab:rates}.
\begin{table}
\begin{tabular} { | l | l | l | l |}
\hline
& $R < 10^2\mathrm{pc}$ & $R \in [10^2,10^5]\mathrm{pc}$ & $R > 10^5\mathrm{pc}$ \\
\hline
\textit{Ejected}:& & &\\
BH-BH & $1.25 \times 10^{-2} $ & $7.34 \times 10^{-1} $& $8.96 \times 10^{-1} $\\
NS-NS & $6.20 \times 10^{-2} $ & $3.13 \times 10^{-1} $& $4.96 \times 10^{-1} $\\
BH-NS & $7.82 \times 10^{-3} $ & $9.57 \times 10^{-2} $& $1.51 \times 10^{-1} $\\
WD-WD & $2.29 \times 10^{-1} $ & $8.34 \times 10^{0} $& $1.23 \times 10^{1} $\\
Ia SNe & $2.29 \times 10^{-1} $ & $8.42 \times 10^{-1} $& $1.08 \times 10^{0} $\\
EBBH & $2.47 \times 10^{0} $ & $4.02 \times 10^{0} $& $1.38 \times 10^{0} $\\
\hline
\textit{Quick}:& & &\\
BH-BH & $2.83 \times 10^{1} $ & $\times$ & $\times$\\
NS-NS & $5.94 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $4.82 \times 10^{0} $ & $\times$ & $\times$\\
WD-WD & $3.80 \times 10^{3} $ & $\times$ & $\times$\\
Ia SNe & $5.14 \times 10^{2} $ & $\times$ & $\times$\\
EBBH & $1.09 \times 10^{2} $ & $\times$ & $\times$\\
\hline
\textit{Slow}:& & &\\
BH-BH & $6.63 \times 10^{1} $ & $\times$ & $\times$\\
NS-NS & $2.25 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $8.02 \times 10^{0} $ & $\times$ & $\times$\\
WD-WD & $4.97 \times 10^{2} $ & $\times$ & $\times$\\
Ia SNe & $7.69 \times 10^{1} $ & $\times$ & $\times$\\
EBBH & $3.08 \times 10^{1} $ & $7.73 \times 10^{-2} $ & $\times$\\
\hline
\textit{Total}:& & &\\
BH-BH & $9.47 \times 10^{1} $ & $7.39 \times 10^{-1} $& $8.96 \times 10^{-1} $\\
NS-NS & $8.19 \times 10^{1} $ & $3.13 \times 10^{-1} $& $4.96 \times 10^{-1} $\\
BH-NS & $1.29 \times 10^{1} $ & $9.57 \times 10^{-2} $& $1.51 \times 10^{-1} $\\
WD-WD & $4.30 \times 10^{3} $ & $8.34 \times 10^{0} $& $1.23 \times 10^{1} $\\
Ia SNe & $5.91 \times 10^{2} $ & $8.42 \times 10^{-1} $& $1.08 \times 10^{0} $\\
EBBH & $1.42 \times 10^{2} $ & $4.10 \times 10^{0} $& $1.38 \times 10^{0} $\\
\hline
\hline
\end{tabular}
\caption{Cosmic rate of merger events classified by $R$ with a Salpeter IMF. The units of all entries are in $\mathrm{yr}^{-1}$}
\label{tab:rates}
\end{table}
The same process can also be performed for a log-flat IMF
\begin{equation}
\xi(m)dm \propto \bigg(\frac{m}{M_\odot}\bigg)^{-1}dm
\end{equation}
which yields $n_\mathrm{wd-wd}=1.7\times 10^{-3}\,M_\odot^{-1}$, $n_\mathrm{bh-bh}=1.9\times 10^{-3}\,M_\odot^{-1}$, $n_\mathrm{ns-ns}=1.3\times 10^{-3}\,M_\odot^{-1}$ and $n_\mathrm{bh-ns}=3.5\times 10^{-4}\,M_\odot^{-1}$. The corresponding cosmic merger rates are listed in \tab{tab:log_flat_rates}.
\begin{table}
\begin{tabular} { | l | l | l | l |}
\hline
& $R < 10^2\mathrm{pc}$ & $R \in [10^2,10^5]\mathrm{pc}$ & $R > 10^5\mathrm{pc}$ \\
\hline
\textit{Ejected}:& & &\\
BH-BH & $2.38 \times 10^{-2} $ & $1.41 \times 10^{0} $& $1.70 \times 10^{0} $\\
NS-NS & $9.95 \times 10^{-2} $ & $5.03 \times 10^{-1} $& $7.97 \times 10^{-1} $\\
BH-NS & $1.30 \times 10^{-2} $ & $1.59 \times 10^{-1} $& $2.52 \times 10^{-1} $\\
WD-WD & $6.59 \times 10^{-3} $ & $2.40 \times 10^{-1} $& $3.11 \times 10^{-1} $\\
Ia SNe & $6.59 \times 10^{-3} $ & $2.43 \times 10^{-2} $& $3.11 \times 10^{-2} $\\
EBBH & $4.70 \times 10^{0} $ & $7.65 \times 10^{0} $& $2.63 \times 10^{0} $\\
\hline
\textit{Quick}:& & &\\
BH-BH & $5.39 \times 10^{1} $ & $\times$ & $\times$\\
NS-NS & $9.53 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $8.03 \times 10^{0} $ & $\times$ & $\times$\\
WD-WD & $1.09 \times 10^{2} $ & $\times$ & $\times$\\
Ia SNe & $1.48 \times 10^{1} $ & $\times$ & $\times$\\
EBBH & $2.07 \times 10^{2} $ & $\times$ & $\times$\\
\hline
\textit{Slow}:& & &\\
BH-BH & $1.26 \times 10^{2} $ & $\times$ & $\times$\\
NS-NS & $3.60 \times 10^{1} $ & $\times$ & $\times$\\
BH-NS & $1.34 \times 10^{1} $ & $\times$ & $\times$\\
WD-WD & $1.43 \times 10^{1} $ & $\times$ & $\times$\\
Ia SNe & $2.22 \times 10^{0} $ & $\times$ & $\times$\\
EBBH & $5.85 \times 10^{1} $ & $1.47 \times 10^{-1} $ & $\times$\\
\hline
\textit{Total}:& & &\\
BH-BH & $1.80 \times 10^{2} $ & $1.41 \times 10^{0} $& $1.70 \times 10^{0} $\\
NS-NS & $1.31 \times 10^{2} $ & $5.03 \times 10^{-1} $& $7.97 \times 10^{-1} $\\
BH-NS & $2.14 \times 10^{1} $ & $1.59 \times 10^{-1} $& $2.52 \times 10^{-1} $\\
WD-WD & $1.24 \times 10^{2} $ & $2.40 \times 10^{-1} $& $3.56 \times 10^{-1} $\\
Ia SNe & $1.70 \times 10^{1} $ & $2.43 \times 10^{-2} $& $3.11 \times 10^{-2} $\\
EBBH & $2.70 \times 10^{2} $ & $7.79 \times 10^{0} $& $2.63 \times 10^{0} $\\
\hline
\hline
\end{tabular}
\caption{Cosmic rate of merger events classified by $R$ with a log-flat IMF. The units of all entries are $\mathrm{yr}^{-1}$}
\label{tab:log_flat_rates}
\end{table}
\subsection{Electromagnetic counterparts for different types of merger events}
The possibility of drawing astrophysical inferences from the merger
events discussed here largely relies on our ability to identify their merger locations with respect to their galaxy hosts. Studies of host galaxies have a long history in the context of both long and short GRBs, and have been used to probe the nature of their progenitors. The events discussed here span a very large range of distances; however, while {slow} merger events can be produced in a variety of astrophysical conditions (such as dynamical interactions with a single SMBH), the production of HVBs can only occur in the presence of a SMBHB, and hence are especially interesting, as discussed in this paper.
WD-WD mergers are not observable by LIGO/Virgo in gravitational waves, and hence their EM signatures provide the only available empirical information. Mergers of two WDs in a binary have been proposed as candidates of Type Ia SNe \citep[the so-called double-degenerate scenario,][]{1984ApJS...54..335I}.
These transients are typically found to trace the blue light of galaxies \citep[e.g.][]{2015yCat..74480732A}, hence generally occurring within their hosts. The detection of a Type Ia SNe in the IGM would signal an origin from an HVB, whether a WD-WD or a WD and a companion main sequence star \citep[as in the single degenerate scenario,][]{1973ApJ...186.1007W}.
For CO binaries made up of combinations of NSs and BHs, detection by GWs -- when within the LIGO horizon -- constitutes the initial event trigger. However, localization relies on EM counterparts. Among the three possible types of CO binaries, BH-BH mergers are the least likely to be EM bright. For merger events occurring within an AGN disk, accretion by the local gas may provide some source of accretion at the time of merger \citep{2018arXiv180702859S}. For binaries in the intergalactic medium, on the other hand, special conditions are required, such as, e.g., a fossil disk around one of the two BHs of the binary \citep{Perna2019}, or very strong charges \citep{Zhang2016}. However, even if such conditions were realized and a relativistic jet was launched, the likelihood of observing it both in $\gamma$-rays as well as at longer wavelengths would be limited by the fact that relatively bright events could only be detected if the jet
points towards the observer. This is because, lacking ejecta material in the close environment of the merger for the jet to interact with, side emission is highly suppressed. Therefore, identification and localization of a BH-BH merger from an HVB in the intergalactic medium may not be feasible.
On the other hand, NS-NS mergers have been confirmed to produce
a relativistic jet and electromagnetic radiation covering a very
broad spectrum, from $\gamma$-rays to the radio band \citep{Abbott2017}.
Even when observed at a sizable angle with respect to the jet axis,
the interaction of the jet with the tidally disrupted material of the NS produces a bright cocoon, which makes these events more easily detectable within the LIGO horizon even for more unfavorable viewing configurations. In terms of EM radiation, NS-NS mergers from HVBs would be characterized by a much weaker afterglow (i.e. X-rays through radio emission) relative to the prompt $\gamma$-ray emission. This is because, unlike the $\gamma$-ray radiation, which only depends on the properties of the jet and the ejecta material, the afterglow radiation also depends on the density
$n$ of the medium in which the jet propagates. Lower densities result in dimmer afterglows \citep{Sari1998}. Hence a binary NS merger which is very bright in $\gamma$-rays but very dim at longer wavelengths could signal a merger site in the intergalactic medium, even if not accurately localized due to the lack of a bright afterglow.
Last, the situation for an HVB made up of an NS and a BH is more
complex. Only binaries with mass ratios $q$ not exceeding $\sim 3-5$ (with the precise value dependent on the equation of state of the NS) are expected to lead to the formation of an accretion disk during the merger, and hence possibly drive a relativistic jet, similar to the NS-NS merger case. On the other hand, large mass ratios will result in the NS being swallowed by the BH without being tidally disrupted. Sources of radiation in this latter case, if at all, would be more similar to those discussed earlier for binary BH mergers.
\section{Conclusions}\label{sec:summary}
In this work we have studied interactions between CO binaries and massive black hole
binaries that may lead to merger events detectable by GW detectors.
We have performed high-precision N-body simulations in the local reference frame of an SMBHB and combined the scattering outcomes with results from the cosmology simulation Millennium-II to estimate the cosmic rate of different types of CO merger events from dynamical interactions with SMBHBs. Our main conclusions are summarized in the following.
Our simulations show that the interactions between the CO binaries and the SMBHB can produce different types of merger events, which we have dubbed 'quick mergers', '{slow} mergers' and '{ejected} mergers'. Different types of merger events occur at different distances from the galactic centre. Quick mergers with non-negligible eccentricities in the LIGO band can be produced/detected in the innermost region of the SMBHB, in proximity of the primary SMBH. {Slow} mergers, with fully circularized orbits in the LIGO band, can be produced from BSE in the stable region of SMBHBs, or by post-encounters with the secondary SMBH, and can be detected at moderate distances from the galactic centre. The {ejected} mergers that are produced from the hypervelocity binaries could potentially be detected in the galactic halo or in the IGM. Different types of merger events can be precisely classified by the stability map of the SMBHB introduced in Section \ref{sec:stability} and might be distinguished observationally by measuring the distribution of merger distances to the galactic centre from EM counterparts, if available.
Close encounters between CO binaries and SMBHBs can create hypervelocity binaries that could escape the host galaxies and eventually merge in the intergalactic medium. Such a high velocity cannot be produced by interactions in globular clusters nor multiple stellar systems; thus an off-centre merger event might indicate the existence of a nearby SMBHB, especially a high mass ratio SMBHB formed from a prior major galaxy-galaxy merger, which would kick out any merged CO binaries occurring in its orbital plane.
Assuming a binary fraction $n_\mathrm{bi}$ of $10\%$ and a Salpeter IMF, we estimate the merger rate of WD-WD binaries to be $4.32\times 10^{3}\mathrm{yr}^{-1}$, of which $12.3\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $8.34\mathrm{yr}^{-1}$ occur in the DM halo, and $4.30\times 10^{3}\mathrm{yr}^{-1}$ occur in the galactic nucleus. The merger rate of BH-BH binaries is $96.3\mathrm{yr}^{-1}$, of which $0.896\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.740\mathrm{yr}^{-1}$ occur in the DM halo, and $94.7\mathrm{yr}^{-1}$ occur in the galactic nucleus. The merger rate of NS-NS binaries is $82.7\mathrm{yr}^{-1}$, of which $0.496\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.313\mathrm{yr}^{-1}$ occur in the DM halo, and $81.9\mathrm{yr}^{-1}$ occur in the galactic nucleus. The merger rate of EBBHs is $148\mathrm{yr}^{-1}$, of which $1.08\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.842\mathrm{yr}^{-1}$ occur in the DM halo, and $142\mathrm{yr}^{-1}$ occur in the galactic nucleus. Finally, the merger rate of BH-NS binaries is $13.1\mathrm{yr}^{-1}$ of which $0.151\mathrm{yr}^{-1}$ occur in the inter-galactic medium, $0.0957\mathrm{yr}^{-1}$ occur in DM halos and $12.9\mathrm{yr}^{-1}$ occur in galactic nuclei.
Finally, we caution that our scattering experiments can only describe the cusp erosion phase, in which the SMBHB ejects stars in the bound cusp surrounding it as it first becomes bound. Further ejection of unbound CO binaries in the later stages of the SMBHB evolution might contribute to the cosmic rate of fast, slow and ejected CO binary mergers. We defer the study of this later phase to future work.
\section*{Acknowledgements}
RP and NL acknowledge support by NSF award AST-1616157.
The Center for Computational Astro-physics at the Flatiron Institute is supported by the Simons Foundation.
|
\section*{Abstract}
To date, insect scale robots capable of controlled flight have used flapping wings for generating lift, but this requires a complex and failure-prone mechanism.
A simpler alternative is electrohydrodynamic (EHD) thrust, which requires no moving mechanical parts. In EHD, corona discharge generates a flow of ions in an electric field between two electrodes; the high-velocity ions transfer their kinetic energy to neutral air molecules through collisions, accelerating the gas and creating thrust. We introduce a fabrication process for EHD thruster based on 355~nm laser micromachining and our approach allows for greater flexibility in materials selection. Our four-thruster device measures 1.8 $\times$ 2.5~cm and is composed of steel emitters and a lightweight carbon fiber mesh. The current and thrust characteristics of each individual thruster of the quad thruster is determined and agrees with Townsend relation. The mass of the quad thruster is 37~mg and the measured thrust is greater than its weight (362.6~$\upmu \text{N}$). The robot is able to lift off at a voltage of 4.6 kV with a thrust to weight ratio of 1.38.
\section*{Author summary}
Agile small aerial robots can be useful in exploring areas where humans cannot reach. Applications of these robots include gas leak detection in pipes, search and rescue in case of natural disasters, etc. There have been recent developments in flapping wing robots at an insect scale, but these are mechanically complex and difficult to fabricate. Recently, researchers have reported progress in developing aerial vehicles using electrohydrodynamic (EHD) thrust. This phenomenon, also known as ionic wind, produces ions that collide with neutral air molecules to create thrust. As there are no mechanically moving parts, this propulsion scheme eliminates fatigue failure, simplifies assembly, and makes the robots more robust to crashes. Here we report a new means to fabricate these devices more rapidly using a laser-based process, demonstrating a four-thruster robot able to lift its own weight.
\section*{Introduction}
Insect-scale robotics has been an area of interest for its possible uses in agriculture, search and rescue, and biomedicine, among other areas. The small size and reduced manufacturing cost of insect robots have facilitated microrobotic research. To date, the primary emphasis in insect-scale robotics has been on drawing inspiration from biology, because biology has found solutions whose existence proves they work. One example is a flapping-wing robot fly\cite{ma2013controlled}. Robots of this type have subsequently incorporated onboard sensors for flight stabilization \cite{oceli}, and lifted off for the first time without needing a wire tether reaching to the ground using a laser power source \cite{james2018liftoff}, and subsequently, incandescent light~\cite{Jafferis2019}. Other developments include using explosives to break the surface tension of water \cite{chen2017biologically} and RoboFly, which is capable of performing multi-modal locomotion including walking in addition to flying \cite{chukewad2018}. While flapping wings are well suited to insect-sized aerial vehicles, they impose a significant cost in terms of mechanical complexity \cite{WoodPico2012}. In this paper, we focus on an alternative means of generating thrust that is not seen in biology: electrohydrodynamic (EHD) thrust. EHD thrust requires sustained high voltage, which may be why it is not observed in biology. From an engineering perspective, EHD has the appealing characteristic that it requires no moving mechanical parts, simplifying fabrication. A recent advance indicating the promise of EHD thrust was fully EHD-powered 2.5~kg airplane~\cite{MIT2018}
The work in~\cite{drew2018} reported the demonstration of a very small EHD-based aircraft, the ``Ionocraft" measuring only 3~cm across and consisting of four-thrusters. It was able to take off using an external power source. The strength of that work was how the low outflow velocity from low-voltage EHD thrusters was a good match to the low mass of that device. Additionally, EHD is potentially simpler than flapping-wing flight because even a four-thruster device consists of only a single moving part. However, the device in~\cite{drew2018} was fabricated using expensive semiconductor-based cleanroom fabrication facilities. Significant engineering development is required before a small EHD-powered robot can perform aerial locomotion without wires and fully autonomously, which will be necessary for such robots to have a useful application. More rapid and less expensive methods to build the robots with EHD propulsion could facilitate faster design iteration times, which is highly desired when technology is still in its infancy.
\begin{figure}[!htbp]
\centering
\includegraphics[width=13cm]{Fig1.jpg}
\caption{An assembled Quad-thruster robot next to a U.S.penny
The $1.8\times 2.5$~cm quad-thruster having a mass of 37~mg is shown. The robot components consist of a carbon fiber collector grid, four blue tempered steel emitters, and eight fiber optic glass tubes. All components are hand-assembled using external jigs.}
\label{fig1}
\end{figure}
In this paper, we utilize laser machining fabrication to build centimeter-sized aerial robots with EHD thrusters. The process allows for use of a greater variety of electrode materials and eliminates the need for a cleanroom facility. For example, it allows fabricating a complete four-thruster device in a matter of minutes. The quad-thruster robot presented in this paper, as shown in Fig.~\ref{fig1} has generated thrust exceeding its weight. We report the fabrication methodology and the experimental measurements of the corona current, thrust and energy transfer efficiency for each individual thruster of the quad-thruster robot. Finally, we present the takeoff of our quad-thruster robot.
\section*{Electrohydrodynamics (EHD)}
Electrohydrodynamics (EHD) is an interdisciplinary field describing the interaction of fluids with an electric field. Insights into complex multi-physics interactions are essential for understanding EHD flows: (1) ion generation; (2) the ion motion in the electric field; (3) the interaction between the motion of ions and the neutral molecules; and (4) the inertial and viscous forces in the complex flow.
\subsection*{Corona discharge driven flow}
Corona discharge generates a flow of ions in a strong electric field between two electrodes; the high-velocity ions transfer their kinetic energy to the neutral air molecules by collisions that accelerate the gas in the direction of ion drift. This electrohydrodynamic (EHD) flow propulsion phenomenon, also referred to in the literature as ionic wind, is used in many practical applications, such as convective cooling \cite{go2008enhancement,jewell2008cfd}, electrostatic precipitators (ESP) \cite{vaddi2019particle}, plasma-assisted combustion \cite{ju2015plasma}, airflow control \cite{moreau2007airflow, roth2003aerodynamic}, and as a turbulent boundary layer actuators \cite{choi2011turbulent}. The corona induced EHD flow converts electric energy into kinetic energy directly and requires no moving parts. The voltage-current relation during the corona discharge characterizes the ion motion between the electrodes. This phenomenon has been studied since the early 20th century. The classic relationship was derived by Townsend \cite{townsend1915electricity} in 1914 and validated for a coaxial corona configuration. Some recent studies modify Townsend’s quadratic relationship to better describe the relationship for different electrode configurations \cite{yanallah2012semi,zhang2015characteristics,martins2012simulation,guan2018analytical}. A generalized analytical model for voltage to current and voltage to velocity relationship for EHD driven flow has been recently described \cite{guan2018analytical}; the analytical model has a good agreement with the experimental data in the accelerating flow regions (EHD dominated flow). Previous studies have reported that maximum velocity for point-to-ring electrode configuration was recorded at ~9 m/s \cite{guan2018experimental} and have assessed the use of ionic winds in propulsion applications \cite{moreau2013electrohydrodynamic}.
Stuetzer \cite{stuetzer1959ion} presented the first experimental and theoretical analysis of pressure drag produced by the ions, where he determined the pressure generation over a wide range of carrier media. Previous work performed by Masuyama \cite{masuyama2013performance} determined the achievable thrust to power ratios of EHD propulsion on the orders of 5-10 N.kW$^{-1}$. Thrust to power ratio was found to be dependent on electrode distance and the potential difference between the electrodes. Similar results were observed for an ionocraft with a wireless power supply onboard and transmitted power up to 100 W to ionocraft at the voltages up to 12 kV \cite{khomich2018atmosphere}. The EHD propulsion can be utilized for UAV propulsion; the experimentally measured maximum thrust density of 15 N.m$^{-3}$ was reported recently \cite{gilmore2015electrohydrodynamic}. Drew et al. showed that higher thrust density can be achieved for insect-scale robots \cite{drew2017first,drew2018} and EHD flow can be used for flight control.
\subsection*{Electrohydrodynamic Force}
A one dimensional model for an EHD thruster yields an expression in terms of the current, distance between the anode and cathode. Space charge effect is ignored here, however, it can be important at high electric field strengths. The current is determined by integrating charge density
\begin{equation}
I = \int {J.dA} = \int {\rho_e E \mu dA}
\label{current_eq}
\end{equation}
where $\rho_e$\ is the charge density, $\mu$ is the ion mobility in the air, $E$ is the electric field. Ion mobility value of $\mu = 2\times10^{-4}$~m$^2$/V-s is used here. For energy transfer efficiency analysis, consider that thrust is equal to the Coulomb force acting on the volume of fluid between the anode and cathode
\begin{equation}
F = \int {\rho_e E dV}= \frac{Id}{\mu}
\label{force_eq}
\end{equation}
where $F$ is the thrust, $I$ is the ion current, $d$ is the distance between electrodes. The corona power can be written as in Eq. \ref{power_eqn}, and efficiency which is defined as $F/P$ is given by equation Eq. \ref{forceoverP_eq}.
\begin{equation}
P = IV
\label{power_eqn}
\end{equation}
\begin{equation}
\frac{F}{P}=\frac{d}{\mu V}=\frac{1}{E \mu}
\label{forceoverP_eq}
\end{equation}
where \textit{E} is the electric field strength, and \textit{V} is the applied voltage. Drew et al. report the minimum efficiency in their design should be 2~mN/W. The analysis sheds insight into the importance of the electrode distance and applied voltage. Eq. \ref{forceoverP_eq} shows that for the larger electrode spacing higher efficiency values can be reached as observed by Guan et al. \cite{guan2018experimental}. Related to electrode configuration, it is important to revisit Townsend's relations \cite{townsend1914xi}
\begin{equation}
I=CV(V-V_{crit})
\label{current_eq_townsend}
\end{equation}
where $V_{crit}$ is the onset voltage and \textit{V} is the voltage applied. \textit{C} is a constant related to the geometry of the electrodes \cite{masuyama2013performance}. Thrust can be determined using Townsend's relation
\begin{equation}
F=\frac{CV(V-V_{crit})d}{\mu}
\label{force_eq_townsend}
\end{equation}
In practical thruster design to achieve maximum thrust, the constant $C$ needs to be maximized and $V_{crit}$ needs to be minimized. Among the considerations related to the thruster design are the effects of non-linear ionization region, secondary flow structures, cathode blockage ratio, the transition from glow to streaming corona discharge and eventually to sparkover. The full optimization of the EHD thruster is beyond the scope of this paper.
\section*{Fabrication}
Our fabrication process emphasizes speed and simplicity by minimizing the number of components and fabrication time.
Previous work has used a silicon-on-insulator process for the fabrication of EHD thruster \cite{drew2017first}. The emitter and collector electrodes were made from silicon patterned with a photolithographic mask. After a deep reactive ion etch (DRIE) process to ablate through the wafer, the electrodes are once more etched with hydrofluoric acid (HF). In that work, insulating standoffs to separate the emitter from the collector were made from fused silica capillary tubing with an outer diameter of 400~$\upmu \text{m}$. Connections between the tubing and electrodes were made with UV-curable epoxy. Power connections are made with the application of silver epoxy. An external jig was used to align the assembly. Our robot has 13 components (8 capillary tubes, 4 emitters and 1 collector) in comparison to the most recently reported design for the ``Ionocraft" that has 41 components (including sensor components) \cite{drew2018}.
Here, we use laser micro-machining instead of lithographically-patterned silicon. Our laser is a diode-pumped solid-state (DPSS) frequency tripled Nd:Yag laser with 355~nm wavelength (PhotoMachining, Inc., Massachusetts). The DPSS laser output power is 2~W, its beam diameter is 20~$\upmu \text{m}$, and position repeatability is about 3~$\upmu \text{m}$. With this system, we are able to machine both the emitters and collectors in about ten minutes. Our proposed methodology involves machining using the following steps:
\begin{enumerate}
\item Blue tempered steel shim and carbon fiber sheet are laser machined with features for emitters and grid respectively, using the DPSS laser. Their corresponding CAD drawings and actual machined parts are shown in Fig. \ref{fig2} (a), (b) and Fig. \ref{fig2} (e), (g), respectively
\item Jig-1 (holding jig) and jig-2 (spacing jig) (Fig.~\ref{fig2}~(c) and Fig.~\ref{fig2}~(f), respectively) are fabricated out of a sheet of acrylic using a standard CO$_2$ laser cutter.
\item Poles made out of glass fiber optic tubing of an inner diameter of 250~$\upmu \text{m}$ and an outer diameter of 350~$\upmu \text{m}$ are used for maintaining a uniform gap between the electrodes. One of the poles is shown in Fig.~\ref{fig2}~(d).
\end{enumerate}
\begin{figure}[!htbp]
\centering
\includegraphics[width=13cm]{Fig2.jpg}
\caption{\textbf{Individual components of the quad-thruster robot.}
Top view of (a) an emitter drawing, (b) a quad-thruster collector grid drawing. (c) A holding jig used for placing eight glass fiber-optic poles shown in (d) and also for keeping the grid in a plane perpendicular to the poles. (d) One of the eight poles required for the assembly. (e) One of the four emitters involved in the robot. (f) A spacing jig used for keeping emitters at a uniform distance from the grid. (g) A quad-thruster collector grid.}
\label{fig2}
\end{figure}
\section*{Design Analysis}
In this section, we discuss various parameters involved in the design and assembly of the robot. We walk through design considerations to optimize the thrust generated while selecting values for these parameters.
\subsection*{Emitter}
The emitter (corona electrode) material must be rigid, conductive, and with high curvature features (points). The design analysis of the emitter electrode included material, curvature, number of emitter tips, and orientation.
The material initially used was a 50~$\upmu \text{m}$ stainless steel but was changed to a 100~$\upmu \text{m}$ blue tempered stainless steel as it was a stiffer and more durable material. We explored different tip angles, starting with 30$^\circ$, before reducing to 10$^\circ$ and then 5$^\circ$. A smaller radius of curvature of an emitter tip creates a stronger electric field gradient and a high ion concentration. Due to the limitations on the laser beam diameter used to fabricate the emitter and local heating due to the beam, we found that 5$^\circ$ was the sharpest tip that the machine could fabricate.
The number of emitter tips corresponded to the number of electric field localization for corona discharge to occur. We explored different numbers of tips. In each case, the thruster showed similar thrust values, so we settled on eight tips. The last factor evaluated was the emitter tip orientation. The emitter design in the baseline iteration had electrode tips that were designed to be in the plane parallel to the collector grid. This case involved ions having an initial velocity component in the horizontal direction. It was concluded that the horizontal velocity component and hence the kinetic energy loss can be avoided by pointing the tips directly towards the grid.
\subsection*{Collector}
The collector is the heaviest of all the components in the robot. It should have a low blockage ratio to allow the thrust-causing air molecules to flow through it while remaining stiff. We explored different grid spacing and material thicknesses.
We started with a collector grid with 150~$\upmu \text{m}$ spacing between grid marks made out of readily available 50~$\upmu \text{m}$ stainless steel. Due to issues with weight and bending of the stainless steel grid from the strain of other components, we switched to unidirectional carbon fiber reinforced composite. The carbon fiber sheet was made by laying up the 69~GSM (69~g/m$^2$) carbon fiber (TenCate M49J) in 0-90-0 directions. After curing, this lay-up measures about 180~$\upmu \text{m}$ thick. The mass of the single thruster carbon fiber grid with this configuration was 5.9~mg, compared to the previous 8.3~mg stainless steel grid.
We further optimized spacing and weight by reducing the spacing from 150~$\upmu \text{m}$ to 100~$\upmu \text{m}$. We were unable to achieve a functional collector using thinner, 90~$\upmu \text{m}$ carbon fiber and reduced grid spacing of 50~$\upmu \text{m}$ because of excessive breakage during fabrication (when laser cutter has a spot radius of 20~$\upmu \text{m}$ which ends up removing material unevenly on both sides of a grid line). There are 15$\times$9 square openings in the collector grid each with a side length of 100~$\upmu \text{m}$. This works out to a total flow area of 21.6~mm$^2$ corresponding to a blockage of 32.75$\%$.
\subsection*{Quad-thruster design}
After individual thrusters were designed, the next design steps involved putting four of the single thrusters together to make a quad-thruster. In the quad thruster, each single thruster was separated by 7~mm and the inter-electrode spacing is 3.5~mm. Inter-electrode spacing is chosen iteratively based on the observed thrust to weight ratio for the current robot design. Effect of inter-thruster spacing (7~mm) and orientation is neither explored nor optimized in this current work. We also created a single thruster, as shown in Fig.~\ref{fig3}, for testing and performance characterization.
\begin{figure}[!h]
\centerline{\includegraphics[width=8cm]{Fig3.jpg}}
\caption{
Dimetric view of collector grid, emitter electrode, and connector poles for the current iteration of a single thruster.}
\label{fig3}
\end{figure}
\section*{Assembly} \label{assembly}
In this section, we discuss the assembly process for the quad-thruster. The assembly takes about 15 minutes to complete after the components are fabricated. The steps involved in the assembly process are summarized below.
\begin{enumerate}
\item The set of four poles are placed through holes on jig-1 as shown in Fig.~\ref{fig4}~(a).
\item The grid is carefully placed on the holding jig through the poles, it is then glued down with these poles from top to avoid accidentally gluing poles of the grid with the jig. It then looks as shown in Fig.~\ref{fig4}~(b).
\item Four spacing jigs (having a height of 3.5~mm) are now placed on each single collector grid as shown in Fig.~\ref{fig4}~(c).
\item Four emitters are now slid into the poles on top of single collector grids as shown in Fig.~\ref{fig4}~(d). It is made sure that all of the tips are in contact with their spacing jigs. These emitters are then glued down (Cyanoacrylate) with the poles.
\item Once the glue is dry, all four spacing jigs are slid out, and the whole assembly is then taken out of the holding jig. The assembly looks like the CAD shown in Fig.~\ref{fig4}~(e). Fig.~\ref{fig4}~(f) shows a picture of an actual assembly sitting on a holding jig.
\item The whole system is powered through external tethers. The quad-thruster has 2 external wires; a 58-gauge copper wire is attached to one of the ends of the emitters such that the connection is closer to the center of the collector grid, and another 58-gauge copper wire is attached to the center of the collector grid. Each wire is placed over electrode and silver paste is added using a probe tip. After an electrical connection is established, a bit of glue is added to further reinforce the joint.
\end{enumerate}
\begin{figure}[!h]
\centering
\includegraphics[width=13cm]{Fig4.jpg}
\caption{ \textbf{Assembly steps of the quad-thruster robot.}
(a) Fiber-optic glass poles placed into the holes of the holding jig. (b) Grid is then placed on holding jig through the poles. Note: grid thickness exaggerated for 3D view. (c) Four spacing jigs placed on each single collector grid. (d) Four emitters are slid into the poles on top of spacing jigs. (e) Dimetric view of a quad-thruster after the jigs are removed. (f) Picture of a quad-thruster fully assembled in the external jig-1 that is used for assembly.}
\label{fig4}
\end{figure}
\section*{Experimental Method and Results}
Electrical characterization of the individual thruster in the quad thruster was performed first. A high voltage positive DC power supply (Bertan 205B-20R) was used to create the potential difference between the emitter and collector. The current associated with the discharge was determined from the power supply's built-in ammeter. Results are shown in Fig.~\ref{fig5} for each of the four thrusters that comprise a single unit. The current and voltage trends are similar to previously reported quadratic trends \cite{guan2018analytical, townsend1915electricity, townsend1914xi}. Using equation~(\ref{current_eq_townsend}) the current-voltage relationship is plotted, showing that our data agree with the model. A peak force of 1.2 mN for the quad thruster is predicted from Eq.~(\ref{force_eq_townsend}), which corresponds to a thrust-to-weight ratio of 3.5.
\begin{figure}[!htbp]
\centerline{\includegraphics[width=10cm]{Fig5.jpg}}
\caption{\textbf{Corona current vs the applied voltage for each individual thruster.} The measurements are fitted with a Townsend current model and accurately capture the theoretical trend. Calculated corona onset voltage is 3600~V with a standard deviation of 100~V}
\label{fig5}
\end{figure}
The thrust force calculated from the model is validated using experimental thrust measurement as shown in Fig.~\ref{fig6}; A similar setup has been used in previous studies~\cite{moreau2013electrohydrodynamic}. The thruster was held directly above the balance with 0.1 mg resolution (Mettler Toledo) such that the collector grid was aligned parallel to the scale surface using a ceramic tweezer. This arrangement, with the thruster fixed, reduces the confounding effect of electrostatic forces acting on the aircraft through the tether wire. The distance between the collector grid and the weight scale is 21 mm. The scale reading was set to zero and each individual thruster was energized to measure the thrust. A piece of Teflon was placed between the balance plateau and the collector to electrically isolate the balance and to avoid any leakage current. The measured thrust is the force exerted by the accelerated ionic wind on the precision scale. It can be seen that the thrust increases with the voltage applied across the electrodes as shown in Fig.~\ref{fig7}. The thrust trends follow the previously reported quadratic relationship with the applied voltage in~\cite{moreau2013electrohydrodynamic} and~\cite{masuyama2013performance}. The maximum thrust generated occurs immediately before the sparkover is initiated, at which point thrust drops to zero and destroying the mesh. The peak force generated by each thruster was around 260~$\upmu \text{N}$ with a standard deviation of 20~$\upmu \text{N}$. The model predictions were higher than the experimental results, but the model captured the trend. The difference between the model and experiments can be attributed to blockage and drag losses as shown in~\cite{moreau2013electrohydrodynamic}. The model described in Eq.~(\ref{force_eq_townsend}) helped us to validate the thrust measurements and can form the basis for future constrained parametric optimization studies of the quad-thruster robot.
\begin{figure}[!htbp]
\centerline{\includegraphics[width=8cm]{Fig6.jpg}}
\caption{\textbf{Schematic of thrust measurement apparatus.}
Thrust generated by the EHD thruster was measured by measuring the force produced by the ionic wind on the precision scale. Tethers are not shown for simplicity and the robot is not resting on the scale.}
\label{fig6}
\end{figure}
\begin{figure}[!htbp]
\centering
\includegraphics[width=10cm]{Fig7.jpg}
\caption{\textbf{Thrust variation with applied voltage is depicted for each thruster of the quad-thruster robot.} The input voltages range from 3 kV to 5.2 kV. The Townsend current model shown in fig.\ref{fig6} is used to compute theoretical upper limit for the measured Coulomb force and the data captures this trend.}
\label{fig7}
\end{figure}
Thrust density and efficiency are important parameters in understanding the working and performance of EHD thruster compared to other designs. Thrust density is defined as the amount of thrust generated per unit area, whereas efficiency is defined as thrust per unit power. Thrust density for EHD thrusters is calculated from the effective area where EHD flow exists, i.e., the mesh area. Fig.~\ref{fig8} shows how the thrust density varies with the corona input power of EHD thruster. The electrical power was calculated from equation \ref{power_eqn}. A maximum thrust of 0.295~mN corresponds to 13.67~N/\text{m$^2$} thrust density achieved at an input electrical power (aka corona power) of 90.4~mW. Therefore, the thrust density per unit power for the EHD thruster is 151.17~N/\text{m$^2$}W. The efficiency is about 3.265~mN/W. This data can be compared to a piezo-actuated flapping wing such as the RoboFly \cite{chukewad2018}, which has a measured efficiency of 12.2~mN/W. For a thrust of 0.736~mN, input power of 60~mW, and 308~\text{mm$^2$} effective swept area of the wing, the thrust density is 2.39~N/\text{m$^2$}. Therefore, the thrust density per unit power is 39.8~N/{m$^2$W}. Therefore, while the efficiency of the EHD thruster is lower than a flapping-wing robot of comparable size (the efficiency of the flapping wing robot is 3.74 times higher), the thrust density per unit power consumed is 3.8 times higher for the EHD thruster. This is important because the thrust density correlates to the mass of the thruster, and therefore this metric represents a scale-independent (and propulsion-type-independent) measure of efficiency.
\begin{figure}[!htbp]
\centering
\includegraphics[width=10cm]{Fig8.jpg}
\caption{\textbf{Efficiency in terms of thrust density versus the corona power for each thruster in the quad-thruster robot.} The data points shown displayed are mean values of thrust density and corona power.}
\label{fig8}
\end{figure}
For free flight experiments, the robot was placed on a wooden table. The four emitters are interconnected with 51 gauge wires as mentioned above in the assembly section and the quad thruster was actuated using two 58 gauge copper wires in front of a high-speed camera (Sony RX100). One 58-gauge connection is attached from the top to one of the inner ends of an emitter and the other connection is attached to the center of the collector grid. The power tethers were held using ceramic tweezers and strain relieved. The inner legs were removed and attached to the outer legs of the robot to increase the height and diminish the electrostatic interaction with the takeoff plane. With a voltage of 4.6~kV, lift off of quad thruster was achieved. Fig.\ref{fig9} an image sequence from the flight for the first 0.32 seconds before the wires touched each other, which ended the free flight. This shows conclusively that the robot is able to lift its own weight and we believe that a vertical liftoff can be achieved in future experiments by trimming the robot as illustrated in \cite{DakshDevice}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=13cm]{Fig9.jpg}
\caption{\textbf{Frames captured at a frame rate of 240 fps from the quad-thruster in flight.} Robot is resting on the table with the collector connection dangling down and the emitter connection straight up.}
\label{fig9}
\end{figure}
Finally, we conclude this section by comparing the performance of the single-emitter thrusters of our four-thruster device to the single-emitter, four-thruster device presented in~\cite{drew2017first} (which did not directly measure thrust measurements). Hence, to compare the performance of the two robots, the thrust generated (estimated from the takeoff video presented in~\cite{drew2017first}) and thrust-to-weight ratio are chosen as the criteria. The single emitter version of the Ionocraft takes off at 2400~V with a corona current measuring close to 20~$\upmu$A. This works out to a corona power of 0.048~W, and from the observed peak acceleration of the Ionocraft, they conclude that the corresponding thrust is approximately 200 $\upmu$N which amounts to a thrust to weight ratio of 2.04. At the same input power, our quad-thruster device generates 675~$\upmu$N which gives a thrust-to-weight ratio of 1.86. Table.\ref{table1} includes a comprehensive summary of manufacturing methodology and performance comparison between the Ionocraft and our laser-fabricated quad-thruster robot.
\begin{table}[!htbp]
\begin{adjustwidth}{-2.25in}{0in}
\RaggedLeft
\captionsetup{justification=RaggedLeft}
\caption{
{\bf Comparison of our work with earlier work by Drew \cite{drew2017first}.}}
\begin{tabular}{|l+l|l|l|l|l|l|l|}
\hline
\multicolumn{1}{|l|}{\bf Comparison Criterion} & \multicolumn{1}{|l|}{\bf Drew \cite{drew2017first}} & \multicolumn{1}{|l|}{\bf Our work}\\ \thickhline
Electrode Material & Silicon & Carbon fiber (grid) \\
\ & &Stainless steel (emitter) \\ \hline
Total weight ($\upmu$N) & 98~$\upmu$\text{N} (10~\text{mg}) & 362.6~$\upmu$ \text{N} (37~\text{mg}) \\ \hline
Thrust at 0.048~W ($\upmu$\text{N}) & 200~$\upmu$\text{N} & 675~$\upmu$\text{N} \\ \hline
Thrust to Weight ratio at 0.048~W & 2.04 & 1.86 \\ \hline
Assembly time in minutes & 30 & 15 \\ \hline
Clean-room facility & Required & Not required \\ \hline
\end{tabular}
\label{table1}
\end{adjustwidth}
\end{table}
\section*{Conclusion}
A quad thruster was fabricated using UV laser micro-machining and the performance of the quad thruster is characterized. Measured current and thrust values are in good agreement with the EHD theory. The variation across all the four thrusters in the quad is minimal and the lift off of quad-thruster robot is demonstrated. The quad-thruster robot was able to lift its own weight, as indicated by thrust measurements and free-flight connected to a wire tether. The thrust-to-weight ratio of our robot at takeoff voltage of 4.6~kV is 1.38 with a power consumption of 0.037~W. The peak thrust-to-weight ratio of our robot at the maximum actuation voltage of 5.2kV (with typical operation range of 3.6kV to 5.2kV) is 3.03 with a power consumption of 0.104~W, which is slightly below the ratio of 4.5 reported previously~\cite{drew_hilton_head}. We used fewer emitter rows than~\cite{drew_hilton_head}, and we expect that adding rows of tips, will substantially increase lift with little added weight. We also believe there is ample opportunity to reduce the mass of our device through the use of thinner and lighter material. We plan to conduct an exploration of different designs in finite element simulation to explore the configuration space in greater detail. All of this work will contribute significantly in boosting the thrust to weight ratio of our quad-thruster robot which is fundamental for steps towards autonomy.
The fabrication time, from raw materials to complete assembly takes less than 25 minutes and this approach allows for greater flexibility in the selection of materials. The process is viable and a faster alternative to a silicon-on-insulator fabrication process of an EHD thruster at the laboratory prototyping stage. This compares favorably with the process reported in literature~\cite{drew2017first}, which takes 2-3 days at best. The fabrication time includes next-day shipping time for the masks and other queue time. In addition to short fabrication time, we remark that laser micro-fabrication allows for a much more diverse material set. While silicon has a high strength-to-weight ratio, there are other materials that can provide better performance for certain applications, such as the even higher strength-to-weight ratio of unidirectional carbon fiber composites. Furthermore, if there are other materials that may improve the lifetime of the sharp emitter tips, it is almost certainly possible to machine it using a DPSS laser, and the fabrication of these electrodes does not require a cleanroom facility.
The path to an autonomy mandates advancement on a number of fronts. Various components are needed as payload. A small on-board camera (as demonstrated in \cite{balasubramanian2018insect} for a flapping wing insect-scale robot) along with other sensing units such as IMUs (has mass of 37~mg in \cite{drew2018}) can be added, which will be instrumental in controlled flight and sensing around surroundings. Future work also includes work on on-board power supply and associated power-electronics. Until recently, micro-robots have been powered through external connections. The Autonomous Insect Robotics Lab at the University of Washington has developed light weight circuit that requires no battery and provides wireless power to a robot \cite{james2018liftoff}. This circuit was capable of developing 200~V to drive piezo actuators in a 100~mg package. We expect that a similar approach could extend to the kV potential differences needed for EHD thrusters.
\section*{Acknowledgment}
The authors also wish to thank Johannes James and Vikram Iyer for insightful discussion during experimental set up and initial design, and TenCate, Inc. for donating the composite materials used in this study.
|
\section{Introduction} \label{sec:introduction}
Word Sense Disambiguation (WSD) is a core task of Natural Language Processing (NLP) which consists in assigning the correct sense to a word in a given context, and has many potential applications \cite{Navigli2009WordSD}. Despite breakthroughs in distributed semantic representations (i.e. word embeddings), resolving lexical ambiguity has remained a long-standing challenge in the field. Systems using non-distributional features, such as It Makes Sense (IMS, \citealp{Zhong2010ItMS}), remain surprisingly competitive against neural sequence models trained end-to-end. A baseline that simply chooses the most frequent sense (MFS) has also proven to be notoriously difficult to surpass.
Several factors have contributed to this limited progress over the last decade, including lack of standardized evaluation, and restricted amounts of sense annotated corpora. Addressing the evaluation issue, \citet{raganato2017word} has introduced a unified evaluation framework that has already been adopted by the latest works in WSD. Also, even though SemCor \cite{Miller1994UsingAS} still remains the largest manually annotated corpus, supervised methods have successfully used label propagation \cite{Yuan2016SemisupervisedWS}, semantic networks \cite{Vial2018ImprovingTC} and glosses \cite{Luo2018IncorporatingGI} in combination with annotations to advance the state-of-the-art. Meanwhile, task-specific sequence modelling architectures based on BiLSTMs or Seq2Seq \cite{D17-1120} haven't yet proven as advantageous for WSD.
Until recently, the best semantic representations at our disposal, such as word2vec \cite{Mikolov2013DistributedRO} and fastText \cite{Bojanowski2017EnrichingWV}, were bound to word types (i.e. distinct tokens), converging information from different senses into the same representations (e.g. `play song' and `play tennis' share the same representation of `play'). These word embeddings were learned from unsupervised Neural Language Modelling (NLM) trained on fixed-length contexts. However, by recasting the same word types across different sense-inducing contexts, these representations became insensitive to the different senses of polysemous words. \citet{CamachoCollados2018FromWT} refer to this issue as the meaning conflation deficiency and explore it more thoroughly in their work.
Recent improvements to NLM have allowed for learning representations that are context-specific and detached from word types. While word embedding methods reduced NLMs to fixed representations after pretraining, this new generation of contextual embeddings employs the pretrained NLM to infer different representations induced by arbitrarily long contexts. Contextual embeddings have already had a major impact on the field, driving progress on numerous downstream tasks. This success has also motivated a number of iterations on embedding models in a short timespan, from context2vec \cite{Melamud2016context2vecLG}, to GPT \cite{radford2018improving}, ELMo \cite{peters2018deep}, and BERT \cite{bert_naacl}.
Being context-sensitive by design, contextual embeddings are particularly well-suited for WSD. In fact, \citet{Melamud2016context2vecLG} and \citet{peters2018deep} produced contextual embeddings from the SemCor dataset and showed competitive results on \citet{raganato2017word}'s WSD evaluation framework, with a surprisingly simple approach based on Nearest Neighbors ($k$-NN). These results were promising, but those works only produced sense embeddings for the small fraction of WordNet \cite{Fellbaum2000WordNetA} senses covered by SemCor, resorting to the MFS approach for a large number of instances. Lack of high coverage annotations is one of the most pressing issues for supervised WSD approaches \cite{Le2018ADD}.
Our experiments show that the simple $k$-NN w/MFS approach using BERT embeddings suffices to surpass the performance of all previous systems. Most importantly, in this work we introduce a method for generating sense embeddings with full-coverage of WordNet, which further improves results (additional 1.9\% F1) while forgoing MFS fallbacks. To better evaluate the fitness of our sense embeddings, we also analyse their performance without access to lemma or part-of-speech features typically used to restrict candidate senses. Representing sense embeddings in the same space as any contextual embeddings generated from the same pretrained NLM eases introspections of those NLMs, and enables token-level intrinsic evaluations based on $k$-NN WSD performance. We summarize our contributions\footnote{Code and data: \href{https://github.com/danlou/lmms}{github.com/danlou/lmms}} below:
\begin{itemize}
\item A method for creating sense embeddings for all senses in WordNet, allowing for WSD based on $k$-NN without MFS fallbacks.
\item Major improvement over the state-of-the-art on cross-domain WSD tasks, while exploring the strengths and weaknesses of our method.
\item Applications of our sense embeddings for concept-level analyses of NLMs.
\end{itemize}
\section{Language Modelling Representations} \label{sec:lmreps}
Distributional semantic representations learned from Unsupervised Neural Language Modelling (NLM) are currently used for most NLP tasks. In this section we cover aspects of word and contextual embeddings, learned from from NLMs, that are particularly relevant for our work.
\subsection{Static Word Embeddings} \label{sec:static}
Word embeddings are distributional semantic representations usually learned from NLM under one of two possible objectives: predict context words given a target word (Skip-Gram), or the inverse (CBOW) (word2vec, \citealp{Mikolov2013DistributedRO}). In both cases, context corresponds to a fixed-length window sliding over tokenized text, with the target word at the center. These modelling objectives are enough to produce dense vector-based representations of words that are widely used as powerful initializations on neural modelling architectures for NLP. As we explained in the introduction, word embeddings are limited by meaning conflation around word types, and reduce NLM to fixed representations that are insensitive to contexts. However, with fastText \cite{Bojanowski2017EnrichingWV} we're not restricted to a finite set of representations and can compositionally derive representations for word types unseen during training.
\subsection{Contextual Embeddings} \label{sec:contextual}
The key differentiation of contextual embeddings is that they are context-sensitive, allowing the same word types to be represented differently according to the contexts in which they occurr. In order to be able to produce new representations induced by different contexts, contextual embeddings employ the pretrained NLM for inferences. Also, the NLM objective for contextual embeddings is usually directional, predicting the previous and/or next tokens in arbitrarily long contexts (usually sentences). ELMo \cite{peters2018deep} was the first implementation of contextual embeddings to gain wide adoption, but it was shortly after followed by BERT \cite{bert_naacl} which achieved new state-of-art results on 11 NLP tasks. Interestingly, BERT's impressive results were obtained from task-specific fine-tuning of pretrained NLMs, instead of using them as features in more complex models, emphasizing the quality of these representations.
\section{Word Sense Disambiguation (WSD)} \label{sec:wsd}
\noindent There are several lines of research exploring different approaches for WSD \cite{Navigli2009WordSD}. Supervised methods have traditionally performed best, though this distinction is becoming increasingly blurred as works in supervised WSD start exploiting resources used by knowledge-based approaches (e.g. \citealp{luo2018leveraging,Vial2018ImprovingTC}). We relate our work to the best-performing WSD methods, regardless of approach, as well as methods that may not perform as well but involve producing sense embeddings. In this section we introduce the components and related works that are most relevant for our approach.
\subsection{Sense Inventory, Attributes and Relations} \label{sec:wordnet}
The most popular sense inventory is WordNet, a semantic network of general domain concepts linked by a few relations, such as synonymy and hypernymy. WordNet is organized at different abstraction levels, which we describe below. Following the notation used in related works, we represent the main structure of WordNet, called synset, with $lemma_{POS}^{\#}$, where $lemma$ corresponds to the canonical form of a word, $_{POS}$ corresponds to the sense's part-of-speech (\underline{n}oun, \underline{v}erb, \underline{a}djective or adve\underline{r}b), and $^{\#}$ further specifies this entry.
\begin{itemize}
\item Synsets: groups of synonymous words that correspond to the same sense, e.g. $dog_n^{1}$.
\item Lemmas: canonical forms of words, may belong to multiple synsets, e.g. \textit{dog} is a lemma for $dog_n^{1}$ and $chase_v^{1}$, among others.
\item Senses: lemmas specifed by sense (i.e. sensekeys), e.g. \textit{{dog\small{\%1:05:00::}}}, and \textit{{domestic\_dog\small{\%1:05:00::}}} are senses of $dog_n^{1}$.
\end{itemize}
Each synset has a number of attributes, of which the most relevant for this work are:
\begin{itemize}
\item Glosses: dictionary definitions, e.g. $dog_n^{1}$ has the definition `a member of the genus Ca...'.
\item Hypernyms: `type of' relations between synsets, e.g. $dog_n^{1}$ is a hypernym of $pug_n^{1}$.
\item Lexnames: syntactical and logical groupings, e.g. the lexname for $dog_n^{1}$ is \textit{noun.animal}.
\end{itemize}
In this work we're using WordNet 3.0, which contains 117,659 synsets, 206,949 unique senses, 147,306 lemmas, and 45 lexnames.
\subsection{WSD State-of-the-Art} \label{sec:sota}
While non-distributional methods, such as \citet{Zhong2010ItMS}'s IMS, still perform competitively, there are have been several noteworthy advancements in the last decade using distributional representations from NLMs. \citet{iacobacci2016embeddings} improved on IMS's performance by introducing word embeddings as additional features.
\citet{Yuan2016SemisupervisedWS} achieved significantly improved results by leveraging massive corpora to train a NLM based on an LSTM architecture. This work is contemporaneous with \citet{Melamud2016context2vecLG}, and also uses a very similar approach for generating sense embeddings and relying on $k$-NN w/MFS for predictions. Although most performance gains stemmed from their powerful NLM, they also introduced a label propagation method that further improved results in some cases. Curiously, the objective \citet{Yuan2016SemisupervisedWS} used for NLM (predicting held-out words) is very evocative of the cloze-style Masked Language Model introduced by \citet{bert_naacl}. \citet{Le2018ADD} replicated this work and offers additional insights.
\citet{D17-1120} trained neural sequencing models for end-to-end WSD. This work reframes WSD as a translation task where sequences of words are translated into sequences of senses. The best result was obtained with a BiLSTM trained with auxilliary losses specific to parts-of-speech and lexnames. Despite the sophisticated modelling architecture, it still performed on par with \citet{iacobacci2016embeddings}.
The works of \citet{Melamud2016context2vecLG} and \citet{peters2018deep} using contextual embeddings for WSD showed the potential of these representations, but still performed comparably to IMS.
Addressing the issue of scarce annotations, recent works have proposed methods for using resources from knowledge-based approaches. \citet{luo2018leveraging} and \citet{Luo2018IncorporatingGI} combine information from glosses present in WordNet, with NLMs based on BiLSTMs, through memory networks and co-attention mechanisms, respectively. \citet{Vial2018ImprovingTC} follows \citet{D17-1120}'s BiLSTM method, but leverages the semantic network to strategically reduce the set of senses required for disambiguating words.
All of these works rely on MFS fallback. Additionally, to our knowledge, all also perform disambiguation only against the set of admissible senses given the word's lemma and part-of-speech.
\subsection{Other methods with Sense Embeddings} \label{sec:senseembeddings}
Some works may no longer be competitive with the state-of-the-art, but nevertheless remain relevant for the development of sense embeddings. We recommend the recent survey of \citet{CamachoCollados2018FromWT} for a thorough overview of this topic, and highlight a few of the most relevant methods. \citet{Chen2014AUM} initializes sense embeddings using glosses and adapts the Skip-Gram objective of word2vec to learn and improve sense embeddings jointly with word embeddings. \citet{Rothe2015AutoExtendEW}'s AutoExtend method uses pretrained word2vec embeddings to compose sense embeddings from sets of synonymous words. \citet{CamachoCollados2016NasariIE} creates the NASARI sense embeddings using structural knowledge from large multilingual semantic networks.
These methods represent sense embeddings in the same space as the pretrained word embeddings, however, being based on fixed embedding spaces, they are much more limited in their ability to generate contextual representations to match against. Furthermore, none of these methods (or those in \S \ref{sec:sota}) achieve full-coverage of the +200K senses in WordNet.
\section{Method} \label{sec:method}
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{img/lmms_nn.pdf}
\caption{Illustration of our $k$-NN approach for WSD, which relies on full-coverage sense embeddings represented in the same space as contextualized embeddings. For simplification, we label senses as synsets. Grey nodes belong to different lemmas (see \S \ref{sec:usm}).}
\label{fig:knn}
\end{figure}
Our WSD approach is strictly based on $k$-NN (see Figure \ref{fig:knn}), unlike any of the works referred previously. We avoid relying on MFS for lemmas that do not occur in annotated corpora by generating sense embeddings with full-coverage of WordNet. Our method starts by generating sense embeddings from annotations, as done by other works, and then introduces several enhancements towards full-coverage, better performance and increased robustness. In this section, we cover each of these techniques.
\subsection{Embeddings from Annotations} \label{sec:annotations}
Our set of full-coverage sense embeddings is bootstrapped from sense-annotated corpora. Sentences containing sense-annotated tokens (or spans) are processed by a NLM in order to obtain contextual embeddings for those tokens. After collecting all sense-labeled contextual embeddings, each sense embedding is determined by averaging its corresponding contextual embeddings. Formally, given $n$ contextual embeddings $\vec{c}$ for some sense $s$:
$$\vec{v}_{s} = \frac{1}{n}\sum_{i=1}^{n}\vec{c}_i , dim(\vec{v}_{s}) = 1024$$
In this work we use pretrained ELMo and BERT models to generate contextual embeddings. These models can be identified and replicated with the following details:
\begin{itemize}
\item ELMo: 1024 (2x512) embedding dimensions, 93.6M parameters. Embeddings from top layer (2).
\item BERT: 1024 embedding dimensions, 340M parameters, cased. Embeddings from sum of top 4 layers ([-1,-4])\footnote{This was the configuration that performed best out of the ones on Table 7 of \citet{devlin2018bert}.}.
\end{itemize}
BERT uses WordPiece tokenization that doesn't always map to token-level annotations (e.g. `multiplication' becomes `multi', `\#\#plication'). We use the average of subtoken embeddings as the token-level embedding. Unless specified otherwise, our LMMS method uses BERT.
\subsection{Extending Annotation Coverage} \label{sec:extend}
As many have emphasized before \cite{Navigli2009WordSD,CamachoCollados2018FromWT,Le2018ADD}, the lack of sense annotations is a major limitation of supervised approaches for WSD. We address this issue by taking advantage of the semantic relations in WordNet to extend the annotated signal to other senses. Semantic networks are often explored by knowledge-based approaches, and some recent works in supervised approaches as well \cite{luo2018leveraging,Vial2018ImprovingTC}. The guiding principle behind these approaches is that sense-level representations can be imputed (or improved) from other representations that are known to correspond to generalizations due to the network's taxonomical structure. \citet{Vial2018ImprovingTC} leverages relations in WordNet to reduce the sense inventory to a minimal set of entries, making the task easier to model while maintaining the ability to distinguish senses. We take the inverse path of leveraging relations to produce representations for additional senses.
On \S \ref{sec:wordnet} we covered synsets, hypernyms and lexnames, which correspond to increasingly abstract generalizations. Missing sense embeddings are imputed from the aggregation of sense embeddings at each of these abstraction levels. In order to get embeddings that are representative of higher-level abstractions, we simply average the embeddings of all lower-level constituents. Thus, a synset embedding corresponds to the average of all of its sense embeddings, a hypernym embedding corresponds to the average of all of its synset embeddings, and a lexname embedding corresponds to the average of a larger set of synset embeddings. All lower abstraction representations are created before next-level abstractions to ensure that higher abstractions make use of lower generalizations. More formally, given all missing senses in WordNet $\hat{s} \in {W}$, their synset-specific sense embeddings $S_{\hat{s}}$, hypernym-specific synset embeddings $H_{\hat{s}}$, and lexname-specific synset embeddings $L_{\hat{s}}$, the procedure has the following stages:
$$
\begin{matrix}
(1) & if |S_{\hat{s}}| > 0 , & \vec{v}_{\hat{s}} = \frac{1}{|S_{\hat{s}}|}\sum\vec{v}_{s} , \forall \vec{v}_{s} \in S_{\hat{s}} \\\\
(2) & if |H_{\hat{s}}| > 0 , & \vec{v}_{\hat{s}} = \frac{1}{|H_{\hat{s}}|}\sum\vec{v}_{syn} , \forall \vec{v}_{syn} \in H_{\hat{s}} \\\\
(3) & if |L_{\hat{s}}| > 0 , & \vec{v}_{\hat{s}} = \frac{1}{|L_{\hat{s}}|}\sum\vec{v}_{syn} , \forall \vec{v}_{syn} \in L_{\hat{s}}
\end{matrix}
$$
In Table \ref{tab:extension} we show how much coverage extends while improving both recall and precision.
\begin{table}[h]
\centering
\resizebox{0.48\textwidth}{!}{%
\begin{tabular}{@{}lccc@{}}
\toprule
\textbf{} & \textbf{} & \multicolumn{2}{c}{\small{\textbf{F1 / P / R (without MFS)}}} \\ \cmidrule(l){3-4}
\multicolumn{1}{c}{\textbf{Source}} & \textbf{Coverage} & \small{\textbf{BERT}} & \small{\textbf{ELMo}} \\ \midrule \midrule
\multicolumn{1}{l}{SemCor} & 16.11\% & 68.9 / 72.4 / 65.7 & 63.0 / 66.2 / 60.1 \\ \midrule
\multicolumn{1}{l}{+ synset} & 26.97\% & 70.0 / 72.6 / 70.0 & 63.9 / 66.3 / 61.7 \\ \midrule
\multicolumn{1}{l}{+ hypernym} & 74.70\% & 73.0 / 73.6 / 72.4 & 67.2 / 67.7 / 66.6 \\ \midrule
\multicolumn{1}{l}{+ lexname} & 100\% & 73.8 / 73.8 / 73.8 & 68.1 / 68.1 / 68.1 \\
\bottomrule
\end{tabular}
}
\caption{Coverage of WordNet when extending to increasingly abstract representations along with performance on the ALL test set of \citet{raganato2017word}.}
\label{tab:extension}
\end{table}
\subsection{Improving Senses using the Dictionary} \label{sec:dictionary}
There's a long tradition of using glosses for WSD, perhaps starting with the popular work of \citet{Lesk1986AutomaticSD}, which has since been adapted to use distributional representations \cite{Basile2014AnEL}. As a sequence of words, the information contained in glosses can be easily represented in semantic spaces through approaches used for generating sentence embeddings. There are many methods for generating sentence embeddings, but it's been shown that a simple weighted average of word embeddings performs well \cite{Arora2017ASB}.
Our contextual embeddings are produced from NLMs using attention mechanisms, assigning more importance to some tokens over others, so they already come `pre-weighted' and we embed glosses simply as the average of all of their contextual embeddings (without preprocessing). We've also found that introducing synset lemmas alongside the words in the gloss helps induce better contextualized embeddings (specially when glosses are short). Finally, we make our dictionary embeddings ($\vec{v}_d$) sense-specific, rather than synset-specific, by repeating the lemma that's specific to the sense, alongside the synset's lemmas and gloss words. The result is a sense-level embedding, determined without annotations, that is represented in the same space as the sense embeddings we described in the previous section, and can be trivially combined through concatenation or average for improved performance (see Table \ref{tab:ablation}).
Our empirical results show improved performance by concatenation, which we attribute to preserving complementary information from glosses. Both averaging and concatenating representations (previously $L_2$ normalized) also serves to smooth possible biases that may have been learned from the SemCor annotations. Note that while concatenation effectively doubles the size of our embeddings, this doesn't equal doubling the expressiveness of the distributional space, since they're two representations from the same NLM. This property also allows us to make predictions for contextual embeddings (from the same NLM) by simply repeating those embeddings twice, aligning contextual features against sense and dictionary features when computing cosine similarity. Thus, our sense embeddings become:
$$
\vec{v}_s = \begin{bmatrix}
||\vec{v}_s||_2\\
||\vec{v}_d||_2
\end{bmatrix}, dim(\vec{v}_s) = 2048
$$
\begin{table*}[htb]
\centering
\resizebox{0.85\textwidth}{!}{%
\begin{tabular}{@{}c|c|c|c|c|c|c|c@{}}
\toprule
\textbf{Configurations} & \textbf{LMMS$_{1024}$} & \hspace{1.5cm} & \hspace{1.5cm} & \textbf{LMMS$_{2048}$} & \hspace{1.5cm} & \hspace{1.5cm} & \textbf{LMMS$_{2348}$} \\ \hline \hline
\textbf{Embeddings} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} \\
Contextual \small{(d=1024)} & \xmark & & \xmark & \xmark & \xmark & & \xmark \\
Dictionary \small{(d=1024)} & & \xmark & \xmark & \xmark & & \xmark & \xmark \\
Static \small{(d=300)} & & & & & \xmark & \xmark & \xmark \\
\textbf{Operation} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} \\
Average & & & \xmark & & & & \\
Concatenation & & & & \xmark & \xmark & \xmark & \xmark \\ \hline
\textbf{Perf. (F1 on ALL)} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} & \textbf{} \\
Lemma \& POS & 73.8 & 58.7 & 75.0 & \textbf{75.4} & 73.9 & 58.7 & \textbf{75.4} \\
Token \small{(Uninformed)} & 42.7 & 6.1 & 36.5 & 35.1 & 64.4 & 45.0 & \textbf{66.0} \\
\bottomrule
\end{tabular}
}
\caption{Overview of the different performance of various setups regarding choice of embeddings and combination strategy. All results are for the 1-NN approach on the ALL test set of \citet{raganato2017word}. We also show results that ignore the lemma and part-of-speech features of the test sets to show that the inclusion of static embeddings makes the method significantly more robust to real-world scenarios where such gold features may not be available.}
\label{tab:ablation}
\end{table*}
\subsection{Morphological Robustness} \label{sec:morphology}
WSD is expected to be performed only against the set of candidate senses that are specific to a target word's lemma. However, as we'll explain in \S \ref{sec:usm}, there are cases where it's undesirable to restrict the WSD process.
We leverage word embeddings specialized for morphological representations to make our sense embeddings more resilient to the absence of lemma features, achieving increased robustness. This addresses a problem arising from the susceptibility of contextual embeddings to become entirely detached from the morphology of their corresponding tokens, due to interactions with other tokens in the sentence.
We choose fastText \cite{Bojanowski2017EnrichingWV} embeddings (pretrained on CommonCrawl), which are biased towards morphology, and avoid Out-of-Vocabulary issues as explained in \S \ref{sec:static}. We use fastText to generate static word embeddings for the lemmas ($\vec{v}_l$) corresponding to all senses, and concatenate these word embeddings to our previous embeddings. When making predictions, we also compute fastText embeddings for tokens, allowing for the same alignment explained in the previous section. This technique effectively makes sense embeddings of morphologically related lemmas more similar. Empirical results (see Table \ref{tab:ablation}) show that introducing these static embeddings is crucial for achieving satisfactory performance when not filtering candidate senses.
\noindent Our final, most robust, sense embeddings are thus:
$$
\vec{v}_s = \begin{bmatrix}
||\vec{v}_s||_2 \\
||\vec{v}_d||_2 \\
||\vec{v}_l||_2
\end{bmatrix}, dim(\vec{v}_s) = 2348
$$
\section{Experiments} \label{sec:experiments}
Our experiments centered on evaluating our solution on \citet{raganato2017word}'s set of cross-domain WSD tasks. In this section we compare our results to the current state-of-the-art, and provide results for our solution when disambiguating against the full set of possible senses in WordNet, revealing shortcomings to be improved.
\begin{table*}[htb]
\centering
\resizebox{0.85\textwidth}{!}{%
\begin{tabular}{@{}ccccccc@{}}
\toprule
\multirow{2}{*}{\textbf{Model}} & \textbf{Senseval2} & \textbf{Senseval3} & \textbf{SemEval2007} & \textbf{SemEval2013} & \textbf{SemEval2015} & \textbf{ALL} \\
& \small{(n=2,282)} & \small{(n=1,850)} & \small{(n=455)} & \small{(n=1,644)} & \small{(n=1,022)} & \small{(n=7,253)} \\ \midrule \midrule
MFS$^\dagger$ (Most Frequent Sense) & {65.6} & {66.0} & {54.5} & {63.8} & {67.1} & {64.8} \\
IMS$^\dagger$ \citeyearpar{Zhong2010ItMS} & {70.9} & {69.3} & {61.3} & {65.3} & {69.5} & {68.4} \\
IMS + embeddings$^\dagger$ \citeyearpar{iacobacci2016embeddings} & {72.2} & {70.4} & {62.6} & {65.9} & {71.5} & {69.6} \\
context2vec $k$-NN$^\dagger$ \citeyearpar{Melamud2016context2vecLG} & {71.8} & {69.1} & {61.3} & {65.6} & {71.9} & {69.0} \\
word2vec $k$-NN \citeyearpar{Yuan2016SemisupervisedWS} & {67.8} & {62.1} & {58.5} & {66.1} & {66.7} & {-} \\
LSTM-LP (Label Prop.) \citeyearpar{Yuan2016SemisupervisedWS} & {\underline{73.8}} & {\underline{71.8}} & {\underline{63.5}} & {69.5} & {72.6} & {-} \\
Seq2Seq (Task Modelling) \citeyearpar{D17-1120} & {70.1} & {68.5} & {63.1*} & {66.5} & {69.2} & {68.6*} \\
BiLSTM (Task Modelling) \citeyearpar{D17-1120} & {72.0} & {69.1} & {64.8*} & {66.9} & {71.5} & {69.9*} \\
ELMo $k$-NN \citeyearpar{peters2018deep} & {71.5} & {67.5} & {57.1} & {65.3} & {69.9} & {67.9} \\
HCAN (Hier. Co-Attention) \citeyearpar{luo2018leveraging} & {72.8} & {70.3} & {-*} & {68.5} & {\underline{72.8}} & {-*} \\
BiLSTM w/Vocab. Reduction \citeyearpar{Vial2018ImprovingTC} & {72.6} & {70.4} & {61.5} & {\underline{70.8}} & {71.3} & {70.8} \\ \midrule
BERT $k$-NN & {\textbf{76.3}} & {73.2} & {66.2} & {71.7} & {74.1} & {73.5} \\
LMMS$_{2348}$ (ELMo) & {68.1} & {64.7} & {53.8} & {66.9} & {69.0} & {66.2} \\
LMMS$_{2348}$ (BERT) & {\textbf{76.3}} & {\textbf{75.6}} & {\textbf{68.1}} & {\textbf{75.1}} & {\textbf{77.0}} & {\textbf{75.4}} \\
\bottomrule
\end{tabular}%
}
\caption{Comparison with other works on the test sets of \citet{raganato2017word}. All works used sense annotations from SemCor as supervision, although often different pretrained embeddings. $^\dagger$ - reproduced from \citet{raganato2017word}; * - used as a development set; bold - new state-of-the-art (SOTA); underlined - previous SOTA.}
\label{tab:eval}
\end{table*}
\subsection{All-Words Disambiguation} \label{sec:eval}
In Table \ref{tab:eval} we show our results for all tasks of \citet{raganato2017word}'s evaluation framework. We used the framework's scoring scripts to avoid any discrepancies in the scoring methodology. Note that the $k$-NN referred in Table \ref{tab:eval} always refers to the closest neighbor, and relies on MFS fallbacks.
The first noteworthy result we obtained was that simply replicating \citet{peters2018deep}'s method for WSD using BERT instead of ELMo, we were able to significantly, and consistently, surpass the performance of all previous works. When using our method (LMMS), performance still improves significantly over the previous impressive results (+1.9 F1 on ALL, +3.4 F1 on SemEval 2013). Interestingly, we found that our method using ELMo embeddings didn't outperform ELMo $k$-NN with MFS fallback, suggesting that it's necessary to achieve a minimum competence level of embeddings from sense annotations (and glosses) before the inferred sense embeddings become more useful than MFS.
In Figure \ref{fig:knn_add} we show results when considering additional neighbors as valid predictions, together with a random baseline considering that some target words may have less senses than the number of accepted neighbors (always correct).
\begin{figure}[htb]
\centering
\includegraphics[width=0.35\textwidth]{img/plot_knn.pdf}
\caption{Performance gains with LMMS$_{2348}$ when accepting additional neighbors as valid predictions.}
\label{fig:knn_add}
\end{figure}
\subsection{Part-of-Speech Mismatches} \label{sec:pos}
The solution we introduced in \S \ref{sec:morphology} addressed missing lemmas, but we didn't propose a solution that addressed missing POS information. Indeed, the confusion matrix in Table \ref{tab:confusion} shows that a large number of target words corresponding to verbs are wrongly assigned senses that correspond to adjectives or nouns. We believe this result can help motivate the design of new NLM tasks that are more capable of distinguishing between verbs and non-verbs.
\begin{table}[htb]
\centering
\resizebox{0.40\textwidth}{!}{%
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{@{}c||cccc@{}}
\toprule
\textbf{WN-POS} & \textbf{NOUN} & \textbf{VERB} & \textbf{ADJ} & \textbf{ADV} \\ \hline \hline
\textbf{NOUN} & 96.95\% & 1.86\% & 0.86\% & 0.33\% \\ \hline
\textbf{VERB} & {\ul 9.08\%} & 70.82\% & {\ul 19.98\%} & 0.12\% \\ \hline
\textbf{ADJ} & {\ul 4.50\%} & 0\% & 92.27\% & 2.93\% \\ \hline
\textbf{ADV} & 2.02\% & 0.29\% & 2.60\% & 95.09\% \\
\bottomrule
\end{tabular}
\renewcommand{\arraystretch}{1}
}
\caption{POS Confusion Matrix for Uninformed Sense Matching on the ALL testset using LMMS$_{2348}$.}
\label{tab:confusion}
\end{table}
\subsection{Uninformed Sense Matching} \label{sec:usm}
WSD tasks are usually accompanied by auxilliary parts-of-speech (POSs) and lemma features for restricting the number of possible senses to those that are specific to a given lemma and POS. Even if those features aren't provided (e.g. real-world applications), it's sensible to use lemmatizers or POS taggers to extract them for use in WSD. However, as is the case with using MFS fallbacks, this filtering step obscures the true impact of NLM representations on $k$-NN solutions.
Consequently, we introduce a variation on WSD, called Uninformed Sense Matching (USM), where disambiguation is always performed against the full set of sense embeddings (i.e. +200K vs. a maximum of 59). This change makes the task much harder (results on Table \ref{tab:ablation}), but offers some insights into NLMs, which we cover briefly in \S \ref{sec:knowledge}.
\subsection{Use of World Knowledge} \label{sec:knowledge}
It's well known that WSD relies on various types of knowledge, including commonsense and selectional preferences \cite{Lenat1986CYCUC,Resnik1997SelectionalPA}, for example. Using our sense embeddings for Uninformed Sense Matching allows us to glimpse into how NLMs may be interpreting contextual information with regards to the knowledge represented in WordNet. In Table \ref{tab:knowledge} we show a few examples of senses matched at the token-level, suggesting that entities were topically understood and this information was useful to disambiguate verbs. These results would be less conclusive without full-coverage of WordNet.
\begin{table*}[htb]
\centering
\resizebox{0.90\textwidth}{!}{%
\begin{tabular}{@{}p{3cm}p{3cm}p{3cm}p{3cm}p{3cm}p{3cm}@{}}
\textbf{Marlon$^\star$} & \textbf{Brando$^\star$} & {\ul \textbf{played}} & \textbf{Corleone$^\star$} & \textbf{in} & \textbf{Godfather$^\star$} \\
\small{$person_{n}^{1}$} & \small{$person_{n}^{1}$} & \small{$act_{v}^{3}$} & \small{$syndicate_{n}^{1}$} & \small{$movie_{n}^{1}$} & \small{$location_{n}^{1}$} \\
\small{$womanizer_{n}^{1}$} & \small{$group_{n}^{1}$} & \small{$make_{v}^{42}$} & \small{$mafia_{n}^{1}$} & \small{$telefilm_{n}^{1}$} & \small{$here_{n}^{1}$} \\
\small{$bustle_{n}^{1}$} & \small{$location_{n}^{1}$} & \small{$emote_{v}^{1}$} & \small{$person_{n}^{1}$} & \small{$final$\_$cut_{n}^{1}$} & \small{$there_{n}^{1}$} \\ \midrule
\pbox{1.2\textwidth}{$\mathbf{act_{v}^{3}}$: play a role or part; $\mathbf{make_{v}^{42}}$: represent fictiously, as in a play, or pretend to be or act like; $\mathbf{emote_{v}^{1}}$: give expression or emotion to, in a stage or movie role.} \\
& & & & & \\
\textbf{Serena$^\star$} & \textbf{Williams} & {\ul \textbf{played}} & \textbf{Kerber$^\star$} & \textbf{in} & \textbf{Wimbledon$^\star$} \\
\small{$person_{n}^{1}$} & \small{$professional$\_$tennis_{n}^{1}$} & \small{$play_{v}^{1}$} & \small{$person_{n}^{1}$} & \small{$win_{v}^{1}$} & \small{$tournament_{n}^{1}$} \\
\small{$therefore_{r}^{1}$} & \small{$tennis_{n}^{1}$} & \small{$line$\_$up_{v}^{6}$} & \small{$group_{n}^{1}$} & \small{$romp_{v}^{3}$} & \small{$world$\_$cup_{n}^{1}$} \\
\small{$reef_{n}^{1}$} & \small{$singles_{n}^{1}$} & \small{$curl_{v}^{5}$} & \small{$take$\_$orders_{v}^{2}$} & \small{$carry_{v}^{38}$} & \small{$elimination$\_$tournament_{n}^{1}$} \\ \midrule
\pbox{1.2\textwidth}{$\mathbf{play_{v}^{1}}$: participate in games or sport; $\mathbf{line}$\_$\mathbf{up_{v}^{6}}$: take one's position before a kick-off; $\mathbf{curl_{v}^{5}}$: play the Scottish game of curling.} \\
& & & & & \\
\textbf{David} & \textbf{Bowie$^\star$} & {\ul \textbf{played}} & \textbf{Warszawa$^\star$} & \textbf{in} & \textbf{Tokyo} \\
\small{$person_{n}^{1}$} & \small{$person_{n}^{1}$} & \small{$play_{v}^{14}$} & \small{$poland_{n}^{1}$} & \small{$originate$\_$in_{n}^{1}$} & \small{$tokyo_{n}^{1}$} \\
\small{$amati_{n}^{2}$} & \small{$folk$\_$song_{n}^{1}$} & \small{$play_{v}^{6}$} & \small{$location_{n}^{1}$} & \small{$in_{r}^{1}$} & \small{$japan_{n}^{1}$} \\
\small{$guarnerius_{n}^{3}$} & \small{$fado_{n}^{1}$} & \small{$riff_{v}^{2}$} & \small{$here_{n}^{1}$} & \small{$take$\_$the$\_$field_{v}^{2}$} & \small{$japanese_{a}^{1}$} \\ \midrule
\pbox{1.2\textwidth}{$\mathbf{play_{v}^{14}}$: perform on a certain location; $\mathbf{play_{v}^{6}}$: replay (as a melody); $\mathbf{riff_{v}^{2}}$: play riffs.} \\
\end{tabular}%
}
\caption{Examples controlled for syntactical changes to show how the correct sense for `played' can be induced accordingly with the mentioned entities, suggesting that disambiguation is supported by world knowledge learned during LM pretraining. Words with $^\star$ never occurred in SemCor. Senses shown correspond to the top 3 matches in LMMS$_{1024}$ for each token's contextual embedding (uninformed). For clarification, below each set of matches are the WordNet definitions for the top disambiguated senses of `played'.}
\label{tab:knowledge}
\end{table*}
\section{Other Applications} \label{sec:otherapps}
Analyses of conventional word embeddings have revealed gender or stereotype biases \cite{Bolukbasi2016ManIT,Caliskan2017SemanticsDA} that may have unintended consequences in downstream applications. With contextual embeddings we don't have sets of concept-level representations for performing similar analyses. Word representations can naturally be derived from averaging their contextual embeddings occurring in corpora, but then we're back to the meaning conflation issue described earlier. We believe that our sense embeddings can be used as representations for more easily making such analyses of NLMs. In Figure \ref{fig:gender} we provide an example that showcases meaningful differences in gender bias, including for lemmas shared by different senses ($doctor$: PhD vs. medic, and $counselor$: therapist vs. summer camp supervisor). The bias score for a given synset $s$ was calculated as following:
$$ bias(s) = sim(\vec{v}_{man_{n}^{1}}, \vec{v}_{s}) - sim(\vec{v}_{woman_{n}^{1}}, \vec{v}_{s})$$
\begin{figure}[htb]
\centering
\includegraphics[width=0.49\textwidth]{img/bias.pdf}
\caption{Examples of gender bias found in the sense vectors. Positive values quantify bias towards $man_{n}^{1}$, while negative values quantify bias towards $woman_{n}^{1}$.}
\label{fig:gender}
\end{figure}
Besides concept-level analyses, these sense embeddings can also be useful in applications that don't rely on a particular inventory of senses. In \citet{LoureiroWiC}, we show how similarities between matched sense embeddings and contextual embeddings are used for training a classifier that determines whether a word that occurs in two different sentences shares the same meaning.
\section{Future Work} \label{sec:future}
In future work we plan to use multilingual resources (i.e. embeddings and glosses) for improving our sense embeddings and evaluating on multilingual WSD. We're also considering exploring a semi-supervised approach where our best embeddings would be employed to automatically annotate corpora, and repeat the process described on this paper until convergence, iteratively fine-tuning sense embeddings.
We expect our sense embeddings to be particularly useful in downstream tasks that may benefit from relational knowledge made accessible through linking words (or spans) to commonsense-level concepts in WordNet, such as Natural Language Inference.
\section{Conclusion} \label{sec:conclusion}
This paper introduces a method for generating sense embeddings that allows a clear improvement of the current state-of-the-art on cross-domain WSD tasks. We leverage contextual embeddings, semantic networks and glosses to achieve full-coverage of all WordNet senses. Consequently, we're able to perform WSD with a simple $1$-NN, without recourse to MFS fallbacks or task-specific modelling. Furthermore, we introduce a variant on WSD for matching contextual embeddings to all WordNet senses, offering a better understanding of the strengths and weaknesses of representations from NLM. Finally, we explore applications of our sense embeddings beyond WSD, such as gender bias analyses.
\section{Acknowledgements} \label{sec:acknowledgements}
This work is financed by National Funds through the Portuguese funding agency, FCT - Funda\c{c}\~{a}o para a Ci\^{e}ncia e a Tecnologia within project: UID/EEA/50014/2019.
|
\section{Introduction} \label{Introduction}
Accurate information of states, parameters and disturbances is essential for effective real-time operation of any system. Many of its relevant variables are often not measurable or too expensive to measure on line. A cost effective approach is to employ estimation techniques to obtain the required information from measurements of other variables and a mathematical model of the system.
Linear and nonlinear estimation have been an active researcher field during the past several decades \citep{patwardhan2012nonlinear}. Linear estimation methods use a simpler representation of the system and can provide acceptable performance only around an operating point and the steady state operational conditions. However, as nonlinearities in the system dynamics become dominants, the performance of linear approaches deteriorates and the estimation algorithms will not necessarily converge to an accurate solutions. Although optimal state estimation solutions for linear systems exists, nonlinear estimation algorithms suffer from generating near-optimal solutions. Consequently, research of nonlinear estimation and filtering problems remains a challenging research area.
Countless studies have been conducted in the literature to address and analyse nonlinear estimation problems. These methods can be broadly categorized into \citep{patwardhan2012nonlinear}: i) linearization methods \citep{kushner1977probability}; ii) approximation methods \citep{benevs1981exact}; iii) Bayesian recursive methods \citep{doucet2001introduction}; iv) moment methods \citep{crisan1998convergence}; and v) higher dimensional nonlinear filter methods \citep{arasaratnam2009cubature}. However, there exist some approaches that approximate the nonlinear behaviour of systems with a linear
parameter--varying (LPV) models \citep{shamma1991guaranteed,shamma1993gain}.
LPV systems are linear systems with matrices depending on time-varying parameters that can evolve over wide operating ranges \citep{apkarian1995self}. These parameters, called \textit{scheduling variables}, depend on exogenous signals that can be measured. When the bounds of these signals are known, the LPV model can be reformulated into a convex linear combination of linear time-invariant \citep{leith2000survey}. If the scheduling variables are functions of endogenous signals such as states, inputs or outputs of the system instead of exogenous signals, \emph{LPV} system describes a large class of nonlinear systems \citep{toth2011state}. The most common technique to obtain an LPV system is the polytopic approach, where the system depends affinely on a time-varying parameter vector that evolves within a polytopic set. In practical situations they could be inaccessible by the fact that scheduling variables are functions of the system states \citep{theilliol2011design}.
In the polytopic LPV observer design, trust full knowledge of the scheduling variables is of paramount importance, because this information is needed to design the observer. Many researchers have proposed solutions to this problem in the polytopic framework. LPV observers with unmeasured scheduling parameters can be designed using proportional observer \citep{ichalal2016auxiliary}, proportional-integral observer \citep{aouaouda2013multi}, generalized dynamic observer \citep{gao2016new,osorio2016new} and adaptive observer \citep{bezzaoucha2013nonlinear,bezzaoucha2018new} framework, respectively.
The main contribution of this paper is the design and analysis of a robust estimator for nonlinear systems under bounded disturbances combining quasi-\emph{LPV} models and dual estimation using a receding horizon framework. The proposed algorithm simultaneously estimates the mixing parameters and the states using a dual estimation approach within a multiple iteration scheme that improve the performance of the estimation at each sample. The conditions to guarantee the robust stability and a convergence to the true system and states for the case of vanishing disturbances are derived. To achieve these results is crucial that the prior weighting in the cost function and the length of the estimation horizon are properly chosen. The assumption on the prior weighting can be verified a prior design.
The paper is organized as follows: Section 2 introduces the notation, definitions and properties that will be used through the paper. Section 3 presents the main results and shows its connections with previous results. The stability and convergence to the true state in the dual iteration is discussed in the initial part of section. Then, the robust regional stability and convergence to the true state and parameters of the estimator are analysed. In section 4 two simple examples are discussed to illustrate the concepts and to show the difference with the estate of the art. Finally, Section 5 presents conclusions.
\section{Preliminaries and setup}
\subsection{Notation}
Let $\mathbb{Z}_{\left[a,b \right]}$ denotes the set of integers in the interval $\left[ a,b \right] \subseteq \mathbb{R} \textnormal{, and } \mathbb{Z}_{\geq a}$ denotes the set of integers greater or equal to $a$. Boldface symbols denote sequences of finite or infinite length, i.e., $\boldsymbol{w} \coloneqq \{ w_{k_1}, \ldots, w_{k_2} \} \textnormal{ for some } k_1, k_2 \in \mathbb{Z}_{\geq 0} \textnormal{ and } k_1 < k_2$, respectively. We denote $x_{j\vert k}$ as the element of the sequence $\boldsymbol{x}$ given at time $k \in \mathbb{Z}_{\geq 0} \textnormal{ and } j \in \left[k_1, k_2 \right]$.
By $\norm{x}$ we denote the Euclidean norm of a vector $x \in \mathbb{R}^n$. Let $\supofseq{x} \coloneqq \sup_{k\in \mathbb{Z}_{\geq 0}} \norm{x_k} $ denote the supreme norm of the sequence $\boldsymbol{x} \textnormal{ and } \supofseq{x}_{\left[a, b \right]} \coloneqq \sup_{k\in \mathbb{Z}_{\left[a, b \right]}} \norm{x_k} $ .
A function $\gamma : \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class $\mathcal{K}$ if $\gamma$ is continuous, strictly increasing and $\gamma \left(0\right) = 0$ . If $\gamma$ is also unbounded, it is of class $\mathcal{K}_\infty$. A function $\zeta : \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class $\mathcal{L}$ if $\zeta \left( k \right)$ is non increasing and $\lim_{k \rightarrow \infty} \zeta\left(k \right) = 0$. A function $\beta : \mathbb{R}_{\geq 0}\times\mathbb{Z}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class $\mathcal{KL}$ if $\beta \left(\cdot,k \right)$ is of class $\mathcal{K}$ for each fixed $k \in \mathbb{Z}_{\geq 0}$, and $\beta \left(r,\cdot \right)$ of class $\mathcal{L}$ for each fixed $r \in \mathbb{R}_{\geq 0}$.
The following inequalities hold for all $\beta \in \mathcal{KL}, \; \gamma \in \mathcal{K} \textnormal{ and } a_j \in \mathbb R_{\geq 0} \textnormal{ with } j \in \mathbb{Z}_{\left[1,n \right]}$
\begin{equation} \label{property_1}
\begin{split}
\gamma\left(\sum_{j=1}^{n}a_i \right) \leq \sum_{j=1}^{n} \gamma\left (n\, a_i\right),&\quad
\beta\left(\sum_{j=1}^{n}a_i, k \right) \leq \sum_{j=1}^{n}\beta\left(n\,
a_i, k\right).
\end{split}
\end{equation}
The preceding inequalities hold since $\max\{a_j \}$ is included in the sequence $\{ a_1, a_2, \ldots, a_n \}$ and $\mathcal{K}$ functions are non-negative strictly increasing functions.
\textbf{Bounded sequences:} A sequence $\boldsymbol{w}$ is bounded if $\supofseq{w}$ is finite. The set of bounded sequences $\boldsymbol{w}$ is denoted as $\mathcal{W}\left(w_{\max}\right) \coloneqq \{w : \boldsymbol{w} \leq w_{\max} \}$ for some $w_{\max} \in \mathbb{R}_{\geq 0}$.
\textbf{Convergent sequences:}
A bounded infinite sequence $\boldsymbol{w}$ is convergent if $\norm{w_k} \rightarrow 0$ as $k \rightarrow \infty$. Let us denote the set of convergent sequences
\begin{equation*}
\mathcal{C}_w \coloneqq \{ \boldsymbol{w} \in \mathcal{W}\left(w_{\max}\right)
\vert \boldsymbol{w}\text{ is convergent} \}.
\end{equation*}
Analogously, the sequence \textbf{$\boldsymbol{v}$} and $\mathcal{C}_v$ can be defined in similar way.
\subsection{Problem statement}
Let us consider a nonlinear discrete-time system with the following behaviour
\begin{equation} \label{eq_nonlinsys}
\begin{split}
x_{k+1} =& f\left(x_k, w_k, \right) \qquad x_0 = \mathtt{x}_0, \: \forall k \in \mathbb{Z}_{\geq 0},\\
y_{k} =& h\left(x_k\right) + v_k
\end{split}
\end{equation}
where $x_{k} \in \mathcal{X} \subset \mathbb{R}^{n_x}$ is the system state, $w_{k} \in \mathcal{W} \subset \mathbb{R}^{n_x}$ is the additive process disturbance, $y_k \in \mathcal{Y} \subset \mathbb{R}^{n_p}$ is the system measurements and $v_{k} \in \mathcal{V} \subset \mathbb{R}^{n_p}$ is the measurement noise.
The sets $\mathcal{X}, \mathcal{W}, \mathcal{Y}$ and $\mathcal{V}$ are known compact and convex with the null vector $\textbf{0}$ in their interior. In the following we assume that $f:\mathbb{R}^{n_x}\times\mathbb{R}^{n_m}\rightarrow \mathbb{R}^{n_x}$ is at least $C^1$ and locally Lipschitz on $x_k$ and $h:\mathbb{R}^{n_x}\rightarrow \mathbb{R}^{n_p}$ is continuous.
Finally, the solution of system \eqref{eq_nonlinsys} at time $k$ is denoted by $x(k,x_0, \boldsymbol{w}, \boldsymbol{d})$, with initial condition $\mathtt{x}_0$ and disturbance sequence $\boldsymbol{w}$. Furthermore, the initial conditions $x_0$ and $\alpha_0$ are unknown, but priors knowledge $\bar{x}_0$ and $\bar{\alpha}_0$ are assumed to be available and their errors are assumed to be bounded, i.e., $\bar{x}_0 \in \mathcal{X}_0 \coloneqq \left\{ \bar{x}_0 : |x_0 - \bar{x}_0| \leq e_{x\,max} \right\}$ such that $\mathcal{X}_0 \subseteq \mathcal{X}$ and $\bar{\alpha}_0 \in \mathcal{A}_0 \coloneqq \left\{ \bar{\alpha}_0 : |\alpha_0 - \bar{\alpha}_0| \leq e_{\alpha\,max} \right\}$ such that $\mathcal{A}_0 \subseteq \mathcal{A}$, respectively.
The solution of the estimation problem aims to find at time $k$ an estimate $\hat{x}_{k|k}$ of the current state $x_k$ using a moving horizon estimator (\emph{MHE}). At each sampling time $k$ the only information available are the previous $N$ measurements $\boldsymbol{y} \coloneqq \left\{ y_{k-N} , \dotsc , y_{k} \right\}$ and a matrix $G(x_k,w_k) \in \Omega(\mathcal{A})$, where $\mathcal{A}$ denotes a polytopic set of matrices
such that
\begin{equation}
A_k = \sum_{i=1}^{q} \alpha_{i,k} A_i, \quad \sum_{i=1}^{q} \alpha_{i,k} C_i
\end{equation}
with $\mathcal{A}$ the unit simplex
\begin{equation}
\mathcal{A} \coloneqq \left\{ \displaystyle\sum_{i=1}^{q} \alpha_{i,k} = 1, \alpha_{i,k} \geq 0 \right\}
\end{equation}
Then, any property ensured for the uncertain \emph{LPV} model
\begin{equation}
\begin{array}{rl}
x_{k+1} =& \displaystyle \sum_{i=1}^{q} \alpha_{i,k} A_i x_{k} + w_{k} + d_{k} , \\
y_{k} =& \displaystyle \sum_{i=1}^{q} \alpha_{i,k} C_i x_{k} + v_{k},
\end{array}
\end{equation}
holds true also for the nonlinear system \eqref{eq_nonlinsys} \citep{angelis2003system}.
Therefore, in this work we propose a moving horizon estimation algorithm to simultaneously estimate the state of the system $\hat{x}_{k \vert k}$ and the mixing parameter of the \emph{LPV} model $\hat{\alpha}_{k \vert k}$. The optimization problem to be solved at each sampling time is the following
\begin{equation} \label{nonconvex problem}
\begin{array}{c}
\underset{\boldsymbol{\hat{x}_{k-N|k}}, \boldsymbol{\hat{w}}, \boldsymbol{\hat{d}}, \boldsymbol{\hat{\alpha}}_k, \boldsymbol{\hat{w}_{\alpha}} }{\operatorname{min}}
\begin{array}{rl}
\Psi_{x,\alpha} \coloneqq & \Gamma_{k-N} \left(\hat{x}_{k-N\vert k} \right) + \sum\limits_{j=k-N}^{k} \ell \left( \hat{w}_{j\vert k},\hat{v}_{j\vert k},\hat{w}_{{\alpha}_{j\vert k}},\hat{d}_{j\vert k}\right) \\
& + \Lambda_k( \hat{\alpha}_{k-N\vert k})
\end{array} \\
\text{s.t.}\left\{
\begin{array}{l}
\begin{array}{rll}
{\hat{\alpha}_{j+1\vert k}} =& {\hat{\alpha}_{j\vert k}} + {\hat{w}_{{\alpha}_{j\vert k}}} \hspace{4cm} j \in \mathbb{Z}_{\left[k-N, k-1\right]}, & \\
{\hat{x}_{j+1|k}} =& \sum_{i=1}^{q} {\hat{\alpha}_{i,k\vert k}} A_i{\hat{x}_{j\vert k}} + {\hat{w}_{j\vert k}}+{\hat{d}_{j\vert k}} , \\
y_{j} =& \sum_{i=1}^{q} {\hat{\alpha}_{i,k \vert k}} C_i {\hat{x}_{j\vert k}} + \hat{v}_{j\vert k} \hspace{2cm} j \in \mathbb{Z}_{\left[k-N, k\right]}, \\
\sum_{i=1}^{q} {\hat{\alpha}_{i,k\vert k}} =& 1,\\
{\hat{\alpha}_{i,k\vert k}} \geq& 0 \hspace{6cm} i \in \mathbb{Z}_{\left[1, q\right]}, \\
\end{array} \\
\hat{x}_{j\vert k} \in \mathcal{X}, \ \hat{w}_{j\vert k} \in \mathcal{W},\hat{w}_{\alpha} \in \mathcal{W}_{\alpha}, \ \hat{v}_{j\vert k} \in \mathcal{V}, \ \hat{d}_{j\vert k} \in \mathcal{D}.
\end{array} \right.
\end{array}
\end{equation}
where $\hat{x}_{j|k}$ is the optimal estimated, $\hat{w}_{j|k}$ is the optimal process noise estimate and $\hat{\alpha}_{j|k}$ is the optimal mixing parameter and $\hat{w}_{\alpha_{j|k}}$ the noise associated to it at sample $k-j \quad j = 0, 1, \dotsc, N$ based on measurements $y_{k-j}$ available at time $k$. The process noise $\boldsymbol{\hat{w}} \coloneqq \left\{ \hat{w}_{k-N-1|k}, \dotsc, \hat{w}_{k-1|k} \right\}$, the mixing parameters ${\alpha} \coloneqq \left[ \alpha_{1,k|k}, \dotsc, \alpha_{q,k|k} \right]^T$, $\boldsymbol{\hat{w}_{\alpha}} \coloneqq \left\{ \hat{w}_{\alpha_{1,k|k}}, \dotsc, \hat{w}_{\alpha_{q,k|k}} \right\}$ and $\hat{x}_{k-N|k}$ are the optimization variables. The stage cost $\ell \left(\hat{w}_{j\vert k},\hat{v}_{j\vert k},\hat{w}_{{\alpha}_{j\vert k}},\hat{d}_{j\vert k}\right)$ penalizes the estimated process noise sequence $\boldsymbol{\hat{w}}_{j|k}$ and the estimation residuals $\boldsymbol{\hat{v}}_{j|k} = y_j - h\left(\hat{x}_{j|k}\right)$, while $\Gamma_{k-N} \left(\hat{x}_{k-N|k}\right)$ and $\Lambda\left(\hat{\alpha}_{k-N|k}\right)$ are the prior weights that penalizes the prior estimates $\hat{x}_{k-N|k}$ and $\hat{\alpha}_{k-N|k}$
The robust stability of estimator \eqref{nonconvex problem} can be achieved by combining a suitable choice of the stage cost $\ell\left( \hat{w}_{j\vert k}, \hat{v}_{j\vert k}, \hat{d}_{j\vert k} \right)$ and the time--varying prior weights
\begin{equation}
\begin{array}{rl}
\Gamma_{k-N \vert k}\left( \hat{x}_{k-N\vert k} \right) =& \lvert \, \hat{x}_{k-N\vert k} - \bar{x}_{k-N} \, \rvert_{P^{-1}_{x,k-N\vert k}}, \\
\Lambda_{k-N \vert k} \left(\hat{w}_{\alpha,k-N\vert k} \right) =& \lvert \, \hat{\alpha}_{k-N\vert k} - \bar{\alpha}_{k-N} \, \rvert_{P^{-1}_{\alpha,k-N\vert k}},
\end{array}
\end{equation}
whose parameters $\left( P^{-1}_{x,k-N\vert k},\bar{x}_{k-N}, P^{-1}_{\alpha,k-N\vert k},\bar{\alpha}_{k-N} \right)$
are recursively updated using the information available at time $k$. In this approach, the prior weight matrix $P_{*,k-N\vert k}$ are given by \citep{sanchez2017adaptive}
\begin{equation} \label{eq:updatePk}
\begin{split}
\epsilon_{k-N} & = y_{k-N}-\hat{y}_{k-N \vert k}, \\
N_{*,k} & = \left[ 1 + \hat{*}_{k-N\vert k-1}^T\, P_{*,k-N-1} \hat{*}_{*,k-N\vert k-1} \right]\frac{\sigma}
{\norm{\epsilon_{k-N}}_2^2} \\
\theta_{*,k} & = 1 - \frac{1}{N_{*,k}},\\
W_{*,k} & = \left[ I - \frac{P_{*,k-N-1} \hat{*}_{k-
N\vert k-1} \hat{*}_{k-N\vert k-1}^T}{1 + \hat{*}_{k-N\vert k-1}^T P_{*,k-N-1} \hat{*}_{k-N\vert k-1}}
\right] P_{*,k-N-1}, \\
P_{*,k-N} & = \left\{ \begin{array}{ccc}
\frac{1}{\theta_{*,k}} W_{*,k} & & \text{if } \frac{1}{\theta_{*,k}}
\Tr{W_{*,k}} \leq c, \\
W_{*,k} & & \text{otherwise},
\end{array} \right.
\end{split}
\end{equation}
where $* \coloneqq \left[x,\alpha \right]$, $\sigma, \, \sigma_w, \, c, \, \lambda \in R_{>0}, \, c>\lambda, \, P_{0}=\lambda I_{n\times n}$ and $\sigma \gg \sigma_w$, where $\sigma_w$ denotes the process noise variance. The prior knowledges of the window $\bar{x}_{k-N}$ and $\bar{\alpha}_{k-N}$ are updated using a smoothed estimate
\begin{equation} \label{eq:updateXA}
\begin{array}{rl}
\bar{x}_{k-N} =& \hat{x}_{k-N\vert k-1}, \\
\bar{\alpha}_{k-N} =& \hat{\alpha}_{k-N\vert k-1}.
\end{array}
\end{equation}
\subsection{Dual estimation formulation}
The joint estimator simultaneously estimates states and mixing parameters. For systems with many parameters, augmenting the state vector can cause a significant increase to the state dimension. This may be problematic as the dimension of the state vector grows, the errors accumulate and the convexity of the optimization problem is lost. To overcome this problem, a dual estimation setup is introduced: the estimation problem \eqref{nonconvex problem} solves separately the state estimation problem (assuming that mixing parameters $\alpha$ remains constant) and the model identification problem (assuming that estimated states $\hat{x}_k$ remains constant) at each sampling time. The problems to be solved iteratively are
\begin{equation} \label{mhe_x problem}
\begin{array}{c}
\underset{\boldsymbol{\hat{x}_{k-N|k}}, \boldsymbol{\hat{w}_x} }{\operatorname{min}} \Psi_x \coloneqq \Gamma_{k-N} \left(\hat{x}_{k-N\vert k} \right) + \sum\limits_{j=k-N}^{k} \ell \left( \hat{w}_{j\vert k}, \hat{v}_{j\vert k}\right) \\%[0.5cm]
\text{s.t.} \left\{
\begin{array}{l}
\begin{array}{rll}
{\hat{x}_{k-N\vert k}} =& \bar{x}_{k-N} + {\hat{w}_{k-N\vert k}} & \\
{\hat{x}_{j+1|k}} =& \sum_{i=1}^{q} \alpha_{i,k} A_i {\hat{x}_{j\vert k}} + {\hat{w}_{x \; j\vert k}} & \quad j \in \mathbb{Z}_{\left[k-N, k-1\right]}, \\
y_{j} =& \sum_{i=1}^{q} \alpha_{i,k} C_i {\hat{x}_{j\vert k}} + \hat{v}_{j\vert k} & \quad j \in \mathbb{Z}_{\left[k-N, k\right]}, \\
\end{array} \\
\hat{x}_{j|k} \in \mathcal{X}, \ \hat{w}_{j|k} \in \mathcal{W}, \ \hat{v}_{j|k} \in \mathcal{V}.
\end{array} \right.
\end{array}
\end{equation}
where the decision variables are $\hat{x}_{k-N\vert k}$ and $\hat{w}_{k-j\vert k} \quad j\coloneqq 1,2, \dots , N$
and
\begin{equation} \label{mhe_alpha problem}
\begin{array}{c}
\underset{\boldsymbol{\alpha},\boldsymbol{\hat{w}_\alpha}}{\operatorname{min}} \Psi_{\alpha} \coloneqq \Lambda_{k-N} \left(\hat{\alpha}_{k-N\vert k} \right) + \sum\limits_{j=k-N}^{k} \ell \left( \hat{d}_{j\vert k}, \hat{v}_{j\vert k},\hat{w}_{{\alpha}_{j\vert k}}\right) \\%[0.5cm]
\text{s.t.} \left\{
\begin{array}{l}
\begin{array}{rll}
{\alpha_{k-N\vert k}} =& \bar{\alpha}_{k-N} + {\hat{w}_{{\alpha}_{k-N\vert k}}} & \\
{\alpha_{j+1\vert k}} =& {\alpha_{j\vert k}} + {\hat{w}_{{\alpha}_{j\vert k}}} & \; j \in \mathbb{Z}_{\left[k-N, k-1\right]} \\
x_{j+1|k} =& \sum_{i=1}^{q} {\alpha_{i,j\vert k}} A^i x_{j\vert k} +
{\hat{d}_{j\vert k}}, \\% \; j \in \mathbb{Z}_{\left[k-N, k-1\right]}.\\
y_{j} =& \sum_{i=1}^{q} {\alpha_{i,j\vert k}} C^i x_{j\vert k} + \hat{v}_{j\vert k} &\; j \in \mathbb{Z}_{\left[k-N, k\right]},\\
\sum_{i=1}^{q}{\alpha_{i,j\vert k}} =& 1 \\
{\alpha_{i,j\vert k}} \geq& 0 & j \in \mathbb{Z}_{\left[1, q\right]}
\end{array} \\
\hat{d}_{j|k} \in \mathcal{D},\hat{w}_{\alpha} \in \mathcal{W}_{\alpha}, \ \hat{v}_{j|k} \in \mathcal{V}.
\end{array} \right.
\end{array}
\end{equation}
where the decision variables are $\hat{w}_{\alpha,k-j\vert k}$ $\quad j\coloneqq 0,1, \dots , N$ and $\quad \hat{\alpha}_{k-N\vert k}$. Problems \eqref{mhe_x problem} and \eqref{mhe_alpha problem} are solved iteratively several times for the same sampling-time. The main novelty of the proposed algorithm is that an improvement in the state estimation and model identification can be guaranteed for a certain number of iterations when some assumptions are fulfilled. Moreover, the number of iterations can be computed offline.
The sequence $P_{k \vert k } \quad k \geq 0$ is positive definite, it is decreasing in norm and it is bounded. The proof of these properties follows similar steps as in \cite{sanchez2017adaptive}.
\begin{assumption} \label{prior weighting assumption}
The prior weighting $\Gamma_{k-N}\left(\hat{x}_{k-N\vert k}\right)$ is a continuous function $ \mathbb{R}^n \rightarrow \mathbb{R}$ lower bounded by $\underline{\gamma}_p\left(\cdot\right) \in \mathcal{K}_\infty{}$ and upper bounded by $\bar{\gamma}_p\left(\cdot\right) \in \mathcal{K}_{\infty}$ such that
\begin{equation} \label{prior weighting bounds}
\underline{\gamma}_p\left(\norm{\hat{x}_{k-N\vert k} - \bar{x}_{k- N}} \right) \leq \Gamma_{k-N}\left(\norm{\hat{x}_{k-N\vert k} - \bar{x}_{k- N}} \right)
\leq \bar{\gamma}_p \left(\norm{\hat{x}_{k-N\vert k} - \bar{x}_{k-N}} \right)
\end{equation}
for all $\hat{x} \in \mathcal{X}$ and
\begin{equation} \label{eq:prior weighting equivalence}
\underline{\gamma}_p\left(r \right) \geq \underline{c}_p \, r^{a},
\quad \bar{\gamma}_p\left(r \right) \leq \bar{c}_p \, r^{a}
\end{equation}
where $0 \leq \underline{c}_p \leq \bar{c}_p$ and $a \in R_{\geq 1}$. Moreover, if the arrival cost is updated using equation \eqref{eq:updatePk}, the bounds $\underline{\gamma}_p$ and $\bar{\gamma}_p$ are bounded by
\begin{equation}
\underline{\gamma}_p\left(r \right) \geq \norm{P^{-1}_0} \, r^{a},
\quad \overline{\gamma}_p\left(r \right) \leq \norm{P^{-1}_{\infty}} \, r^{a}
\end{equation}
\end{assumption}
\begin{definition} \label{ioss_definition}
The system \eqref{eq_nonlinsys} is \textit{incrementally input/output-to-state stable} if there exist some functions $\beta \in \mathcal{KL}$ and $\gamma_1, \gamma_2 \in \mathcal{K}$ such that for every two initial states $z_1$, $z_2 \in \mathbb{R}^n$, and any two disturbances sequences $\boldsymbol{w_1}, \boldsymbol{w_2}$ the following holds for all $k \in \mathbb{Z}_{\geq 0}$:
\begin{equation} \label{i_IOSS_properterty}
\begin{array}{rl}
\norm{x\left( k,z_1,\boldsymbol{w_1} \right) - x\left( k,z_2,\boldsymbol{w_2} \right)} \leq& \max \left \{ \beta\left( \norm{z_1 - z_2},k \right),\gamma_1\left( \supofseq{w_1 - w_2}_{\left[0,k-1 \right]} \right),\right. \\
&\left. \gamma_2\left( \supofseq{y_1 - y_2}_{\left[0,k-1 \right]} \right) \right \} \\
\leq& \beta\left( \norm{z_1 - z_2},k \right) + \gamma_1\left( \supofseq{w_1 - w_2}_{\left[0,k-1 \right]} \right) + \\
& \gamma_2\left( \supofseq{y_1 - y_2}_{\left[0,k-1 \right]} \right)
\end{array}
\end{equation}
\end{definition}
for all $k$ \citep{sontag1997output}.
\begin{assumption} \label{assumption beta function ineq}
The function $\beta(r,s) \in \mathcal{KL}$ and satisfies the following inequality:
\begin{equation} \label{beta_function_ineq}
\beta(r,s) \leq c_{\beta}r^p s^{-q}
\end{equation}
for some $c_{\beta} \in \mathbb{R}_{\geq 0}$, $p \in \mathbb{R}_{\geq 0}$ and $q \in \mathbb{R}_{\geq 0}$ and $q\geq p$.
\end{assumption}
\begin{assumption} \label{stage cost assumption}
The stage cost $\ell\left(\cdot \right) : \mathbb{R}^n \times \mathbb{R}^m \rightarrow \mathbb{R}$ is a continuous function bounded by $\underline{\gamma}_w, \underline{\gamma}_v, \bar{\gamma}_w, \bar{\gamma}_v$ $\in \mathcal{K}_\infty{}$ such that the following inequalities are satisfied $\forall w \in \mathcal{W} \textnormal{ and } v \in \mathcal{V}$
\begin{equation} \label{stage cost inequalities}
\underline{\gamma}_w\left(\hat{w} \right) + \underline{\gamma}_v\left(\hat{v} \right) \leq \ell\left(\hat{w}, \hat{v} \right) \leq \bar{\gamma}_w\left(\hat{w} \right) + \bar{\gamma}_v\left(\hat{v} \right)
\end{equation}
\end{assumption}
\noindent Functions $\gamma_1$ and $\gamma_2$ from Definition \ref{ioss_definition} are related with the bounds of stage cost $\bar{\gamma}_w, \underline{\gamma}_w, \bar{\gamma}_v$ and $\underline{\gamma}_v$ through the following inequalities
\begin{equation} \label{gamma1 and gamma2 inequalities}
\gamma_1\left( 6\underline{\gamma}_w^{-1} \left( r \right)\right) \leq c_1
r^{b_1}, \;
\gamma_2\left( 6\underline{\gamma}_v^{-1} \left( r \right)\right) \leq c_2
r^{b_2}
\end{equation}
for $c_{1}, c_{2}, b_{1}, b_{2} > 0.$
\begin{assumption} \label{assum:equivalencia funcion gamma_alpha}
The prior weighting $\Lambda_{k-N}\left(\hat{\alpha}_{k-N\vert k}\right)$ is a continuous function $ \mathbb{R}^q \rightarrow \mathbb{R}$ lower bounded by $\underline{\gamma}_{\Lambda}\left(\cdot\right) \in \mathcal{K}_\infty{}$ and upper bounded by $\overline{\gamma}_{\Lambda}\left(\cdot\right) \in \mathcal{K}_{\infty}$ such that
\begin{equation} \label{eq:equivalencia funcion gamma_alpha}
\begin{array}{rcl}
\underline{\gamma}_{\Lambda}\left(\norm{\hat{\alpha}_{k-N}-\bar{\alpha}_{k-N}}\right) &\leq \Lambda_{k-N}\left(\norm{\hat{\alpha}_{k-N}-\bar{\alpha}_{k-N}}\right) \leq& \overline{\gamma}_{\Lambda}\left(\norm{\hat{\alpha}_{k-N}-\bar{\alpha}_{k-N}}\right)
\end{array}
\end{equation}
where
\begin{equation}
\begin{array}{rcl}
\underline{\gamma}_{\Lambda}\left(\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}\right)&\geq& \underline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^a \\
\overline{\gamma}_{\Lambda}\left(\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}\right) &\leq& \overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^a
\end{array}
\end{equation}
for some $\underline{c}_{\Lambda}\in\mathbb{R}_{\geq0}$, $\overline{c}_{\Lambda}\in\mathbb{R}_{\geq0}$, $a\in\mathbb{R}_{>0}$, $\overline{c}_{\Lambda}>\underline{c}_{\Lambda}$.
\end{assumption}
In this work, we claim that the proposed estimator holds the property of being robust
asymptotic stable, which is defined as follows.
\begin{definition}
Consider the system described by Equation \eqref{mhe_alpha problem} subject to disturbances $\boldsymbol{w} \in \mathcal{W}\left(w_{\max} \right)$ and $\boldsymbol{v} \in \mathcal{V}\left(v_{\max} \right)$ for $w_{\max} \in \mathbb{R}_{\geq 0} $, $v_{max} \in \mathbb{R}_{\geq 0}$ with prior estimate $\bar{x}_0 \in \mathcal{X}\left(e_{\max} \right)$ for $e_{\max} \in \mathbb{R}_{\geq 0}$. The moving horizon state estimator given by Equation \eqref{mhe_alpha problem} is robustly
asymptotically stable (RAS) if there exists functions $\Phi \in \mathcal{KL} $ and $\pi_w$, $\pi_v \in \mathcal{K}$ such that for all $x_0 \in \mathcal{X}$, all $\bar{x}_0 \in \mathcal{X}_0\left(e_{\max} \right)$, the following is satisfied for all $k \in \mathbb{Z}_{\geq 0}$
\begin{equation} \label{RGAS def property}
\norm{x_k - \hat{x}_k} \leq \Phi\left(\norm{x_0 - \bar{x}_0},k \right) + \pi_w\left(\supofseq{w}_{[0,k-1]} \right) + \pi_v\left(\supofseq{v}_{[0,k-1]} \right)
\end{equation}
\end{definition}
We want to show that if system \eqref{mhe_x problem} is i-IOSS and Assumptions \eqref{prior weighting assumption}, \eqref{assumption beta function ineq} and \eqref{stage cost assumption} are fulfilled, then the proposed estimator with adaptive arrival cost is RGAS. Furthermore, if the process disturbance and measurement noise sequences are convergent (i.e., $\boldsymbol{w}, \boldsymbol{v} \in \mathcal{C}$), the estimation converges to the true state.
\section{Theoretical properties}
Now we are ready to derive the main results: \textit{i}) the stability of the iterated dual estimation and \textit{ii} the robust
asymptotic stability (\emph{RAS}) of the proposed estimator with an estimation horizon $\mathcal{N}$ for nonlinear detectable systems under bounded disturbances
\subsection{Stability of the dual estimation iterations}
\label{sec:stability of the fixed point iteration}
As stated formerly, problems \eqref{mhe_x problem}--\eqref{mhe_alpha problem} are solved sequentially within a dual estimation iteration for each sampling time.
In the following, we will state the conditions required to achieve effectively a decreasing behaviour of the costs inside the dual estimation iteration.
We have now all the necessary ingredients to enunciate the first theorem,
\begin{theorem} \label{minimum value of l}
The sequences of costs $\{ \Psi_x^1,\Psi_x^2,\ldots,\Psi_x^l \}$ and $\{\Psi_{\alpha}^1,\Psi_{\alpha}^2, \ldots,\Psi_{\alpha}^l \}$ generated by the dual estimation iteration are decreasing if the number of iterations $l$ satisfies
\begin{equation}
l \geq \displaystyle\log_2\left( \frac{\displaystyle\epsilon \Psi_x^1 - \Gamma_{k-N}\left( \hat{x}^l_{k-N\vert k} \right)}{\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)} +1 \right) +1,
\end{equation}
where $\epsilon \in \mathbb{R}_{\geq 0}$ and
\begin{equation*}
\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)\coloneqq \underset{i \in \mathbb{Z}_{\left[ 1,l \right]}}{\operatorname{min}} \left\{ \Gamma_{k-N}\left( \hat{x}^i_{k-N\vert k} \right) \right\}.
\end{equation*}
\end{theorem}
\textbf{Proof.}
Let us consider the sequence of costs $\Psi_x^i(\cdot)$ and $\Psi_{\alpha}^i(\cdot)$ $\forall i \in \mathbb{Z}_{>1}$ generated in the \textit{i}--th iteration of the optimization problems \eqref{mhe_x problem} and \eqref{mhe_alpha problem}. Due to the optimality of the solutions the following inequalities are satisfied
\begin{equation} \label{decreasing costs}
\begin{array}{rl}
\Psi_x^1\left( \cdot \right) &\geq \Psi_x^2\left( \cdot \right) \geq, \ldots, \geq \Psi_x^l\left( \cdot \right), \\ \Psi_{\alpha}^1\left( \cdot \right) &\geq \Psi_{\alpha}^2\left( \cdot \right) \geq, \ldots, \geq \Psi_{\alpha}^l\left( \cdot \right).
\end{array}
\end{equation}
Since any iteration takes into account both $\Psi^i_x\left( \cdot \right)$ and $\Psi^i_{\alpha}\left( \cdot \right)$, and due to the sequences are non increasing, we only need to prove the decreasing behaviour of only one of these sequences, let's say $\Psi^i_x\left( \cdot \right)$.
Defining the normalized cost
\begin{equation} \label{gl}
g(k,i) \coloneqq \frac{\Psi_x^i\left( \cdot \right)}{\Psi_x^1\left( \cdot \right)} \quad \forall k,i \geq 1
\end{equation}
the necessary and sufficient conditions to guarantee its decrement along the dual iteration can be obtained using the Gronwall inequality (see \cite{ames1997inequalities, holte2009discrete}). It states that given any three non-negative sequences $y_n, f_n$ and $g_n$ that satisfy
\begin{equation} \label{gronwall lemma condition}
y_n \leq f_n + \displaystyle \sum_{k=0}^{n}g_k \; y_k \qquad \forall n > 0,
\end{equation}
they also verify
\begin{equation} \label{gronwall lemma inequality}
y_n \leq f_n + \displaystyle \sum_{k=0}^{n-1}f_k g_k \displaystyle \prod_{j=k+1}^{n-1}\left( 1+g_j \right) \qquad \forall n > 0.
\end{equation}
Taking the sequences of costs $y_i = \Psi^i_x\left( \cdot \right)$, of arrival-costs $f_i = \Gamma_{k-N}\left( \cdot \right)$ and the normalized costs $g_i = g\left( k,i \right)$ for $n=l$
, which verify \eqref{gronwall lemma condition},
Gronwall inequality \eqref{gronwall lemma inequality} can be written as follows
\begin{equation}
\Psi_x^l \leq \Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right) + \displaystyle\sum_{i=1}^{l-1} \Gamma_{k-N}\left( \hat{x}_{k-N\vert k}^i \right) g(k,i) \prod_{j = i+1}^{l-1} \left( 1 + g(k,j) \right).
\end{equation}
Dividing by $\Psi_x^1$ we obtain
\begin{equation*} \label{Gronwall ineq.}
g\left( k,l \right) \leq \frac{\Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right) + \displaystyle\sum_{i=1}^{l-1} \Gamma_{k-N}\left( \hat{x}_{k-N\vert k}^i \right) g(k,i) \prod_{j = i+1}^{l-1} \left( 1 + g(k,j) \right)}{\displaystyle\Psi^1_x} = \epsilon,
\end{equation*}
which leads to
\begin{equation} \label{ineq intermedia 1}
\displaystyle\sum_{i=1}^{l-1} \Gamma_{k-N}\left( \hat{x}_{k-N\vert k}^i \right) g(k,i) \prod_{j = i+1}^{l-1} \left( 1 + g(k,j) \right) = \epsilon \Psi_x^1 - \displaystyle\Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right).
\end{equation}
Defining
\begin{equation*}
\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)\coloneqq \underset{i \in \mathbb{Z}_{\left[ 1,l \right]}}{\operatorname{min}} \left\{ \Gamma_{k-N}\left( \hat{x}^i_{k-N\vert k} \right) \right\}
\end{equation*}
and recalling that $g\left( k,i\right)$ is a non-increasing sequence, $g\left( k,l \right)\leq g\left( k,i \right) \; \forall l \geq i$, equation \eqref{ineq intermedia 1} can be rewritten as follows
\begin{equation} \label{ineq intermedia 2}
\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)g\left( k,l \right)\displaystyle\sum_{i=1}^{l-1}\prod_{j=i+1}^{l-1}\left( 1+g\left( k,l \right) \right) < \epsilon \Psi_x^1 - \displaystyle\Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right).
\end{equation}
Since
\begin{equation} \label{geometric serie}
\displaystyle\sum_{i=1}^{l-1}\prod_{j=i+1}^{l-1}\left( 1+g\left( k,l \right) \right) = \displaystyle \sum_{i=0}^{l-2}\left( 1+g\left( k,l \right) \right)^i = \frac{\displaystyle\left( 1+g\left( k,l \right) \right)^{l-1} -1 }{\displaystyle g\left( k,l \right)},
\end{equation}
inequality \eqref{ineq intermedia 2} can be rewritten as follows
\begin{equation} \label{ineq intermedia 3}
\begin{array}{rcl}
\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)g\left( k,l \right)\left(\frac{\left( 1+g\left( k,l \right) \right)^{l-1} -1 }{ g\left( k,l \right)}\right) &\stackrel{\eqref{geometric serie}}{<}& \epsilon \Psi_x^1 - \Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right) \\
\end{array}
\end{equation}
and finally $g\left(k,l\right)$ is bounded by
\begin{equation} \label{g(k,l)}
g\left( k,l \right) < \left( \displaystyle\frac{\epsilon \Psi_x^1 - \Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right)}{\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)} + 1 \right)^{\displaystyle\frac{1}{l-1}}-1 < 1.
\end{equation}
Selecting a $l$ large enough, we can guarantee the decrement of sequence $g\left( k,i \right)$ and cost function $\Psi_x^i$ within the dual estimation iteration. Solving for inequality \eqref{g(k,l)}, an upper bound for the number of iterations is given by
\begin{equation} \label{value of l}
l \geq \ceil*{\log_2\left( \displaystyle\frac{\epsilon \Psi_x^1 - \Gamma_{k-N}\left(\hat{x}^l_{k-N\vert k} \right)}{\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)} + 1 \right)} + 1.
\end{equation}
A conservative estimate of $l$ can be computed taking into account the worst case scenario
\begin{equation} \label{value of l conservative 1}
l \geq \displaystyle \ceil*{\log_2\left( \mathcal{E} N\left( \bar{\gamma}_w\left( \supofseq{w} \right) + \bar{\gamma}_v\left( \supofseq{v} \right) \right) + 1 \right)} + 1,
\end{equation}
where $\mathcal{E} \coloneqq \epsilon/\Gamma_{k-N}\left( \hat{x}^m_{k-N\vert k} \right)$.
Inequalities \eqref{value of l} and \eqref{value of l conservative 1} allow to compute the required value of $l$ to guarantee the costs decreasing within the dual estimation iteration. \qed
\begin{remark}
Note that for the noiseless case, only one iteration is needed after the transient due to the uncertainty in the initial condition has vanished.
\end{remark}
\subsection{Robust stability}
In the previous subsection it was shown that the sequence of cost decreases within the dual estimation iteration. At each sampling time, the model used by the estimator is replaced with the newly available until satisfied the stopping criteria. In the following paragraphs we will prove robust stability for the estimator under bounded disturbances and model uncertainty assuming that the system is \textit{i}-IOSS. Moreover, if the length $N$ of the horizon of the estimator is larger than a certain value $\mathcal{N}$ that can be computed offline, the number of iterations $l$ is chosen according to equations \eqref{value of l} and \eqref{value of l conservative 1}, the effects of uncertainty in the initial condition vanish, as well as the disturbances due to model uncertainty. Besides, in the absence of process and measurement noises, states and model converges to the true ones.
\begin{theorem} \label{Theorem_1}
Consider an i-IOSS system \eqref{eq_nonlinsys} with disturbances $\boldsymbol{w} \in \mathcal{W}\left(w_{\max} \right)$, $\boldsymbol{v} \in \mathcal{V}\left(v_{\max} \right)$. Assume that the arrival cost weight matrix of the MHE problem $\Gamma_{k-N}$ is updated using the adaptive algorithm \eqref{eq:updatePk}. Moreover, Assumptions \ref{prior weighting assumption}, \ref{assumption beta function ineq} and \ref{stage cost assumption} are fulfilled and initial condition $x_0$ and $\alpha_0$ are unknown, but prior estimates $\bar{x}_0 \in \mathcal{X}_0$ and $\bar{\alpha}_0 \in \mathcal{A}_0$ are available. Then, the MHE estimator resulting from problems \eqref{mhe_x problem}-- \eqref{mhe_alpha problem} is $RAS$.
\end{theorem}
\textbf{Proof.}
In order to proof stability for the estimator, we start comparing the costs of the first iteration
and the resulting estimated state $\hat{x}_{k-N|k}$ for sampling time $k$
\begin{equation*}
\begin{array}{rl}
\Psi\left( \hat{x}_{k-N\vert k}, \hat{\alpha}, \left\{ \hat{\boldsymbol{w}}_{j\vert k} \right\}, \left\{ \hat{\boldsymbol{d}}_{j\vert k} \right\} \right)
&= g\left(k,l\right)\Psi^1\left( \hat{x}_{k-N\vert k}^1, \hat{\alpha}^1, \left\{ \hat{\boldsymbol{w}}_{j\vert k}^1 \right\}, \left\{ \hat{\boldsymbol{d}}_{j\vert k}^1 \right\} \right) \\
&\leq g\left(k,l\right)\Psi\left( x_{k-N}, \alpha, \left\{\boldsymbol{w}\right\}, \left\{\boldsymbol{0}\right\} \right)
\end{array}
\end{equation*}
Note that if $\hat{\alpha}=\alpha$, the estimated model match with the system, therefore there is no model uncertainty, i.e. $\left\{\boldsymbol{d}\right\} = \left\{\boldsymbol{0}\right\}$. Replacing the sequence of estimated process noises $\left\{\hat{\boldsymbol{w}}\right\}$ by the true sequence $\left\{\boldsymbol{w}\right\}$, the only feasible solution is the true sequence of states $\left\{\boldsymbol{x}\right\}$, and due optimality the inequality is verified.
By mean of Assumptions \ref{prior weighting assumption} - \ref{assum:equivalencia funcion gamma_alpha}, the cost $\Psi\left( \cdot \right)$ is bounded by
\begin{equation} \label{eq:detectability}
\begin{array}{rl}
\underline{\gamma}_p\left(\norm{\hat{x}_{k-N\vert k}-\bar{x}_{k-N}}\right) + \underline{\gamma}_{\Lambda}\left(\norm{\hat{\alpha}_{k-N\vert k}-\bar{\alpha}_{k-N}}\right) +
N\underline{\gamma}_w\left( \norm{\hat{w}_{j\vert k}}\right) + \quad\\
\underline{\gamma}_{\alpha}\left(\norm{\hat{w}_{\alpha, j\vert k}}\right)+
N\underline{\gamma}_v\left(\norm{\hat{v}_{j\vert k}}\right) + N\underline{\gamma}_d\left(\norm{\hat{d}_{j\vert k}}\right) \leq& \\
\left(\overline{\gamma}_p\left(\norm{x_{k-N}-\bar{x}_{k-N}}\right) + \overline{\gamma}_{\Lambda}\left(\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}\right) + N\overline{\gamma}_w\left( \|\boldsymbol{w}\|_{\left[k-N,k\right]}\right) +\right. \quad\\
\left.N\overline{\gamma}_{\alpha}\left( \|\boldsymbol{w}_{\alpha}\|_{\left[k-N,k\right]}\right) +
N\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)\right)g\left(k,l\right) \leq& \\
\overline{c}_p\norm{x_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)+\overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}g\left(k,l\right)+N\left(\overline{\gamma}_w\left(\|\boldsymbol{w}\|\right) +\right. \;\\
\left.
\overline{\gamma}_v\left(\|\boldsymbol{v}\|\right)+\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|\right)
\right)g\left(k,l\right).\;
\end{array}
\end{equation}
Solving for $\norm{\hat{x}_{k-N\vert k}-\bar{x}_{k-N}}$ and using relations \eqref{property_1}, we can write
\begin{equation} \label{eq:bound of arrival cost 1-iteration}
\begin{array}{rl}
\norm{\hat{x}_{k-N\vert k}-\bar{x}_{k-N}} \leq& \underline{\gamma}_p^{-1}\left(\overline{c}_p\norm{x_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)\right) + \\
& \underline{\gamma}_p^{-1}\left(5\overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}g\left(k,l\right)\right) + \\
& \underline{\gamma}_p^{-1}\left(5N\overline{\gamma}_{w}\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)g\left(k,l\right)\right) + \\
& \underline{\gamma}_p^{-1}\left(5N\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)g\left(k,l\right)\right) +\\
& \underline{\gamma}_p^{-1}\left(5N\overline{\gamma}_{v}\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)g\left(k,l\right)\right).
\end{array}
\end{equation}
Using again Assumptions \ref{prior weighting assumption} - \ref{assum:equivalencia funcion gamma_alpha}, one can write
\begin{equation} \label{eq:bound of arrival cost 1-iteration with upper bound of functions}
\begin{array}{rl}
\norm{\hat{x}_{k-N\vert k}-\bar{x}_{k-N}} \leq& \displaystyle \left(\frac{5\overline{c}_p\norm{x_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{\underline{}{c}_p}\right)^{1/a} + \\
& \displaystyle \left(\frac{5\overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^a g\left(k,l\right)}{\underline{c}_p}\right)^{1/a} + \\
& \displaystyle \left(\frac{5N\overline{\gamma}_{w}\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)g\left(k,l\right)}{\underline{c}_p}\right)^{1/a} + \\
& \displaystyle \left(\frac{5N\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)g\left(k,l\right)}{\underline{c}_p}\right)^{1/a} + \\
& \displaystyle \left(\frac{5N\overline{\gamma}_{v}\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)g\left(k,l\right)}{\underline{c}_p}\right)^{1/a}
\end{array}
\end{equation}
From now on, we will drop the superindex $l$.
Using Definition 1, the estimation error at time $k$, given the error at initial conditions $(k=0)$, is bounded by
\begin{equation*}
\norm{x_k - \hat{x}_{k\vert k}}
\leq \beta\left( \norm{x_0 - \hat{x}_{0\vert k}},k \right)+\gamma_1\left( \supofseq{w - \hat{w}}_{\left[0,k-1 \right]} \right) +
\gamma_2\left(\supofseq{v - \hat{v}}_{\left[0,k-1 \right]} \right),
\end{equation*}
and, assuming that $k=N$, we have
\begin{equation} \label{eq:estimation error l-th it.}
\begin{array}{rl}
\norm{x_k - \hat{x}_{k\vert k}} \leq& \beta\left( \norm{x_{k-N} - \hat{x}_{k-N\vert k}},N \right)+\gamma_1\left( \supofseq{w - \hat{w}}_{\left[k-N,k-1 \right]} \right) + \\
& \gamma_2\left( \supofseq{v - \hat{v}}_{\left[k-N,k-1 \right]} \right).
\end{array}
\end{equation}
To found a bound for the estimation error we need to bounds for the terms of the right hand of \eqref{eq:estimation error l-th it.}. Let us start with the first term using inequalities \eqref{property_1} such that the effect of estimation error at the beginning of the estimation window is bounded by
\begin{equation} \label{eq:partial bound of beta function}
\begin{array}{rl}
\beta\left( \norm{x_{k-N} - \hat{x}_{k-N\vert k}},N \right)
&= \beta\left(\norm{x_{k-N}-\bar{x}_{k-N}+\bar{x}_{k-N} - \hat{x}_{k-N\vert k}},N\right)\\
&\leq \beta\left(\norm{x_{k-N}-\bar{x}_{k-N}}+\norm{\bar{x}_{k-N} - \hat{x}_{k-N\vert k}},N\right)\\
&\leq \beta\left(\norm{x_{k-N}-\bar{x}_{k-N}}+\norm{\hat{x}_{k-N\vert k}-\bar{x}_{k-N}},N\right) \\
& \leq \beta\left(2\norm{x_{k-N}-\bar{x}_{k-N}},N\right)+\beta\left(2\norm{\hat{x}_{k-N\vert k}-\bar{x}_{k-N}},N\right).
\end{array}
\end{equation}
Now, the first term of \eqref{eq:partial bound of beta function} can be rewritten using Assumption \ref{assumption beta function ineq}, and the second term with the use of \eqref{eq:bound of arrival cost 1-iteration with upper bound of functions}
\begin{equation}
\begin{array}{rl}
\beta\left(\norm{x_{k-N}-\hat{x}_{k-N\vert k}},N\right)
\leq & \displaystyle \frac{c_{\beta}2^p \norm{x_{k-N}-\bar{x}_{k-N}}^p}{N^q} + \\
& \beta\left(\displaystyle \frac{2\,5^{1/a}\overline{c}_p^{1/a}\norm{x_{k-N}-\bar{x}_{k-N}}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}} \right.+\\
& \left.\displaystyle \frac{2\,5^{1/a}\overline{c}_{\Lambda}^{1/a}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}} \right.+\\
& \left.\displaystyle \frac{2\,5^{1/a}N^{1/a}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}} \right.+\\
& \left.\displaystyle \frac{2\,5^{1/a}N^{1/a}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}} \right.+\\
& \left.\displaystyle \frac{2\,5^{1/a}N^{1/a}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}}, N \right)\\
\end{array}
\end{equation}
Using inequalities \eqref{property_1}
\begin{equation}
\begin{array}{rl}
\beta\left(\norm{x_{k-N}-\hat{x}_{k-N\vert k}},N\right)
\leq & \displaystyle \frac{c_{\beta}2^p \norm{x_{k-N}-\bar{x}_{k-N}}^p}{N^q} + \\
& \beta\left(\displaystyle \frac{10\,5^{1/a}\overline{c}_p^{1/a}\norm{x_{k-N}-\bar{x}_{k-N}}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}}, N\right)+\\
& \beta\left(\displaystyle \frac{10\,5^{1/a}\overline{c}_{\Lambda}^{1/a}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}}, N\right)+\\
& \beta\left(\displaystyle \frac{10\,5^{1/a}N^{1/a}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}}, N\right)+\\
& \beta\left(\displaystyle \frac{10\,5^{1/a}N^{1/a}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}}, N\right)+\\
& \beta\left(\displaystyle \frac{10\,5^{1/a}N^{1/a}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{\underline{c}_p^{1/a}}, N\right)+\\
\end{array}
\end{equation}
Now, by mean of Assumption \ref{assumption beta function ineq}
\begin{equation}
\begin{array}{rl}
\beta\left(\norm{x_{k-N}-\hat{x}_{k-N\vert k}},N\right)
\leq & \displaystyle \frac{c_{\beta}2^p \norm{x_{k-N}-\bar{x}_{k-N}}^p}{N^q} + \\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}\overline{c}_p^{p/a}\norm{x_{k-N}-\bar{x}_{k-N}}^p g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}\overline{c}_{\Lambda}^{p/a}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^p g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{1/a}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{p/a}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k-1\right]}\right)^{p/a}g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{p/a}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k-1\right]}\right)^{p/a}g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}}
\end{array}
\end{equation}
Rearranging terms
\begin{equation}
\label{eq:bound of first term of i-IOSS}
\begin{array}{rl}
\beta\left(\norm{x_{k-N}-\hat{x}_{k-N\vert k}},N\right)
\leq & \displaystyle \frac{\norm{x_{k-N}-\bar{x}_{k-N}}^p}{N^q}\left(
\displaystyle c_{\beta}2^p + \frac{c_{\beta}10^p\,5^{p/a}\overline{c}_p^{p/a}g\left(k,l\right)^{p/a}}{\underline{c}_p^{p/a}}\right) +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}\overline{c}_{\Lambda}^{p/a}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^p g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{1/a}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{p/a}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k-1\right]}\right)^{p/a}g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{p/a}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k-1\right]}\right)^{p/a}g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}}
\end{array}
\end{equation}
Once we have found an upper bound for the first term of \eqref{eq:estimation error l-th it.}, we will follow a similar procedure to find a bound for the second and third terms. Using \eqref{eq:detectability} for the $l-th$ iteration, we can write
\begin{equation}
\begin{array}{rl}
\norm{\hat{w}_{j\vert k}} \stackrel{\eqref{eq:detectability}\eqref{property_1}}{\leq}& \underline{\gamma}_w^{-1}\left( \frac{5\overline{c}_p \norm{x_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{N}\right) \, + \underline{\gamma}_w^{-1}\left( \frac{5\overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{N}\right) + \\
& \underline{\gamma}_w^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)}{N}\right)+\underline{\gamma}_w^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)}{N} \, \right) + \\
& \underline{\gamma}_{w}^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)}{N}\right)
\end{array}
\end{equation}
Introducing this bound in the second term of Equation \eqref{eq:estimation error l-th it.}:
\begin{align}
\nonumber
\gamma_1\left(\|\boldsymbol{w}-\hat{\boldsymbol{w}}\|_{\left[k-N,k\right]}\right) \leq& \gamma_1\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]} + \|\hat{\boldsymbol{w}}\|_{\left[k-N,k\right]}\right) \\
\nonumber
\leq & \gamma_1\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]} + \underline{\gamma}_w^{-1}\left( \frac{5\overline{c}_p \norm{x_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{N}\right)+ \right.\\ \nonumber
&\left. \underline{\gamma}_w^{-1}\left( \frac{5\overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{N}\right) + \right.\\
&\left. \underline{\gamma}_w^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)}{N}\right) + \right.\\ \nonumber
&\left. \underline{\gamma}_w^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)}{N}\right) + \right.\\ \nonumber
&\left. \underline{\gamma}_{w}^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)}{N}\right)
\right)
\end{align}
Recalling Inequalities \eqref{property_1} we obtain the bound
\begin{equation} \label{eq:bound of gamma_1}
\begin{array}{rl}
\gamma_1\left( \|\boldsymbol{w}_j - \hat{\boldsymbol{w}}_{j\vert k}\|_{\left[k-N,k\right]} \right) \leq& \gamma_1\left(6\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right) + \gamma_1\left(6 \underline{\gamma}_w^{-1}\left( \frac{5\overline{c}_p \norm{x_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{N}\right)\right) +\\
& \gamma_1\left(6 \underline{\gamma}_w^{-1}\left( \frac{5\overline{c}_{\Lambda}\norm{\alpha_{k-N}-\bar{x}_{k-N}}^a g\left(k,l\right)}{N}\right) \right)+\\
& \gamma_1\left(6 \underline{\gamma}_w^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)}{N}\right) \right) + \\
& \gamma_1\left(6 \underline{\gamma}_w^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)}{N}\right) \right) + \\
& \gamma_1\left(6 \underline{\gamma}_{w}^{-1}\left(\frac{5g\left(k,l\right)\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)}{N}\right) \right).
\end{array}
\end{equation}
With the use of Assumption \ref{stage cost assumption}, the bound can be finally written as
\begin{equation} \label{eq:final bound of gamma_1}
\begin{array}{rl}
\gamma_1\left( \|\boldsymbol{w}_j - \hat{\boldsymbol{w}}_{j\vert k}\|_{\left[k-N,k\right]} \right) \leq& \gamma_1\left(6\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)+ \displaystyle \frac{c_1 5^{b_1}\overline{c}_p^{b_1}\norm{x_{k-N}-\bar{x}_{k-N}}^{ab_1}g\left(k,l\right)^{b_1}}{N^{b_1}}+\\
& \displaystyle \frac{c_1 5^{b_1}\overline{c}_{\Lambda}^{b_1}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^{ab_1}g\left(k,l\right)^{b_1}}{N^{b_1}}+ \\
& \displaystyle c_1 5^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_1}+\\
& \displaystyle c_1 5^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_1}+\\
& \displaystyle c_1 5^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)^{b_1}
\end{array}
\end{equation}
With a similar procedure, a bound for the third term of inequality \eqref{eq:estimation error l-th it.} is found
\begin{equation} \label{eq:final bound of gamma_2}
\begin{array}{rl}
\gamma_2\left( \|\boldsymbol{v}_j - \hat{\boldsymbol{v}}_{j\vert k}\|_{\left[k-N,k\right]} \right) \leq& \gamma_2\left(6\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)+ \displaystyle \frac{c_2 5^{b_2}\overline{c}_p^{b_2}\norm{x_{k-N}-\bar{x}_{k-N}}^{ab_2}g\left(k,l\right)^{b_2}}{N^{b_2}}+\\
& \displaystyle \frac{c_2 5^{b_2}\overline{c}_{\Lambda}^{b_2}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^{ab_2}g\left(k,l\right)^{b_2}}{N^{b_2}}+ \\
& \displaystyle c_2 5^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_2}+\\
& \displaystyle c_2 5^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_2}+\\
& \displaystyle c_2 5^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)^{b_2}
\end{array}
\end{equation}
The estimation error given in Equation \eqref{eq:estimation error l-th iteration} can be bounded as
\begin{equation} \label{eq:estimation error l-th iteration}
\begin{array}{rl}
\norm{x_k-\hat{x}_{k\vert k}} \leq & \displaystyle \frac{\norm{x_{k-N}-\bar{x}_{k-N}}^p}{N^q}\left(
\displaystyle c_{\beta}2^p + \frac{c_{\beta}10^p\,5^{p/a}\overline{c}_p^{p/a}g\left(k,l\right)^{p/a}}{\underline{c}_p^{p/a}}\right) +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}\overline{c}_{\Lambda}^{p/a}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^p g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{1/a}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k-1\right]}\right)^{1/a}g\left(k,l\right)^{1/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{p/a}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k-1\right]}\right)^{p/a}g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}} +\\
& \displaystyle \frac{c_{\beta}10^p\,5^{p/a}N^{p/a}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k-1\right]}\right)^{p/a}g\left(k,l\right)^{p/a}}{N^q\,\underline{c}_p^{p/a}}+\\
& \gamma_1\left(6\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)+ \displaystyle \frac{c_1 5^{b_1}\overline{c}_p^{b_1}\norm{x_{k-N}-\bar{x}_{k-N}}^{ab_1}g\left(k,l\right)^{b_1}}{N^{b_1}}+\\
& \displaystyle \frac{c_1 5^{b_1}\overline{c}_{\Lambda}^{b_1}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^{ab_1}g\left(k,l\right)^{b_1}}{N^{b_1}}+ \\
& \displaystyle c_1 5^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_1}+\\
& \displaystyle c_1 5^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_1}+\\
& \displaystyle c_1 5^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)^{b_1} + \\
& \gamma_2\left(6\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)+ \displaystyle \frac{c_2 5^{b_2}\overline{c}_p^{b_2}\norm{x_{k-N}-\bar{x}_{k-N}}^{ab_2}g\left(k,l\right)^{b_2}}{N^{b_2}}+\\
& \displaystyle \frac{c_2 5^{b_2}\overline{c}_{\Lambda}^{b_2}\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^{ab_2}g\left(k,l\right)^{b_2}}{N^{b_2}}+ \\
& \displaystyle c_2 5^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_2}+\\
& \displaystyle c_2 5^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_2}+\\
& \displaystyle c_2 5^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_{\alpha}\left(\|\boldsymbol{w_{\alpha}}\|_{\left[k-N,k\right]}\right)^{b_2}
\end{array}
\end{equation}
Since the vector $\alpha$ (and its estimated $\hat{\alpha}$) satisfies $\sum_{i=1}^{q}\alpha_i=1$ and $\alpha_i\geq 0$, the maximal value of $\norm{\alpha-\hat{\alpha}}$ is upper bounded by $\sqrt{2}$, i.e., $\max\{\norm{\alpha-\hat{\alpha}}\} = \max\{\norm{w_{\alpha}}\}\leq \sqrt{2}$.
Defining the constants as follows
\begin{equation*}
q\geq p/a, \; \zeta \coloneqq \max \left \{p, a\,b_1, a\,b_2\right \} \; \eta \coloneqq \min\left\{ q, b_1, b_2\right\},
\end{equation*}
inequality \eqref{eq:estimation error l-th iteration} can be rewritten as follows
\begin{equation} \label{eq:estimation error l-th iteration conservative bound}
\begin{array}{rl}
\norm{x_k - \hat{x}_{k\vert k}} \leq& \displaystyle\frac{\norm{x_{k-N}-\bar{x}_{k-N}}^{\zeta}}{N^{\eta}} \left( \left(1 + \frac{5^{p+q}\overline{c}_p^{q} g\left(k,l\right)^q}{\underline{c}_p^{p/a}}\right)c_{\beta}2^p + c_1 5^{b_1}\overline{c}_1^{b_1}g\left(k,l\right)^{b_1} + \right.\\
& \left. c_2 5^{b_2}\overline{c}_p^{b_2}g\left(k,l\right)^{b_2} \right) + \displaystyle\frac{\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}^{\zeta}}{N^{\eta}}\left( \frac{c_{\beta}10^p5^q\overline{c}_{\Lambda}^qg\left(k,l\right)^q}{\overline{c}_p^{p/a}} + \right.\\
& \left. c_1 5^{b_1}\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^{b_1} + c_25^{b_2}\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^{b_2} \right) + g\left(k,l\right)^{\eta}\left( \frac{c_{\beta}10^p5^q\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^q}{\underline{c}_p^{p/a}} + \right. \\
& \left. c_15^{b_1}\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^{b_1} + c_25^{b_2}\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^{b_2} \right) + \frac{c_{\beta}10^p5^qg\left(k,l\right)^q\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^q}{\underline{c}_p^{p/a}} + \\
& \gamma_1\left(6\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right) + c_15^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_1}+\\
& c_25^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_2} + \frac{c_{\beta}10^p5^q g\left(k,l\right)^q\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^q}{\underline{c}_p^{p/a}} + \\
& \gamma_2\left(6\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right) + c_15^{b_1}g\left(k,l\right)^{b_1}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_1} + \\
& c_25^{b_2}g\left(k,l\right)^{b_2}\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_2}.
\end{array}
\end{equation}
Noting again that $\norm{\alpha_{k-N}-\bar{\alpha}_{k-N}}\leq \sqrt{2}$ and defining the functions and constants as follows
\begin{align}
k_{1} \coloneqq & c_{\beta}2^p \\
k_{2}\coloneqq & \left(5^{p+q}\left(\frac{\overline{c}_p}{\underline{c}_p}\right)^qc_{\beta}2^p+ c_15^{b_1}\overline{c}_p^{b_1}+c_25^{b_2}\overline{c}_p^{b_2}\right), \\
K \coloneqq & \frac{c_{\beta}10^p5^q}{\underline{c}_p^{p/a}}+c_15^{b_1}+c_25^{b_2} \\
\psi_{w1} \coloneqq & \gamma_1\left(6\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right), \\
\psi_{w2} \coloneqq & \frac{c_{\beta}10^p5^q\overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^q}{\underline{c}_p^{p/a}} + \left(c_15^{b_1} + c_25^{b_2}\right) \overline{\gamma}_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]}\right)^{b_1} \\
\psi_{v1} \coloneqq & \gamma_2\left(6\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right), \\
\psi_{w2} \coloneqq & \frac{c_{\beta}10^p5^q\overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^q}{\underline{c}_p^{p/a}} + \left(c_15^{b_1} + c_25^{b_2}\right) \overline{\gamma}_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]}\right)^{b_1} \\
\end{align}
the bound of the estimation error can be rewritten as follows
\begin{equation} \label{eq:bound of estimation error pre final}
\begin{array}{rl}
\norm{x_k -\hat{x}_{k\vert k}} \leq & \displaystyle\frac{\norm{x_{k-N}-\bar{x}_{k-N}}^{\zeta}}{N^{\eta}}\left(k_1 + g\left(k,l\right)^{\eta}k_2\right) + \psi_{w1}+g\left(k,l\right)\psi_{w2}+ \\
& \psi_{v1}+g\left(k,l\right)\psi_{v2}+g\left(k,l\right)\left(\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^{\zeta} + \displaystyle\frac{\overline{c}_{\Lambda}2^{\zeta/2}}{N^{\eta}}\right)K
\end{array}
\end{equation}
Defining others constants and functions
\begin{align}
\bar{k}_{\beta}\left(l\right) \coloneqq & k_1 + g\left(k,l\right)k_2 \\
\Phi_w\left(w,l\right) \coloneqq & \psi_{w1} + g\left(k,l\right)^{\eta}\psi_{w2} \\
\Phi_v\left(v,l\right) \coloneqq & \psi_{v1} + g\left(k,l\right)^{\eta}\psi_{v2} \\
\Phi_{\alpha}\left(l,N\right) \coloneqq & g\left(k,l\right)\left(\overline{\gamma}_{\alpha}\left(\sqrt{2}\right)^{\zeta} + \frac{\overline{c}_{\Lambda}^{\zeta}\,2^{\zeta/2}}{N^{\eta}}\right) K \\
\bar{\beta}\left(r,s\right) \coloneqq & \frac{\bar{k}_{\beta}r^p}{s^q}
\end{align}
with $p=\zeta$, $q=\eta$, $\bar{\beta}\left(r,s\right)\in \mathcal{KL}$, one can write the estimation error as follows
\begin{equation}
\label{eq:RAS}
\begin{array}{rl}
\norm{x_k -\hat{x}_{k\vert k}} \leq & \bar{\beta}\left(\norm{x_{k-N}-\bar{x}_{k-N}}, N\right) + \Phi_w\left(\|\boldsymbol{w}\|_{\left[k-N,k\right]},l\right) + \Phi_v\left(\|\boldsymbol{v}\|_{\left[k-N,k\right]},l\right) +\\
& \Phi_{\alpha}\left(l, N\right)
\end{array}
\end{equation}
The reader can verify that the same result is obtained for $k\in \mathbb{Z}_{\left[1,N-1\right]}$. To guarantee the validity of previous results to the entire time horizon the definition of $\beta\left( r, s\right)$ must be extended to $s = 0$ . Because of $\bar{\beta}\left(r, s \right) \in \mathcal{KL}$, $\bar{\beta}\left(r, 0 \right) \in \mathcal{KL}$ and $\bar{\beta}\left(r, 0 \right) \geq \bar{\beta}\left(r, k \right)$, $\forall k \in \mathbb{Z}_{\geq 1}$, it is sufficient to define $\bar{\beta}\left(r, 0 \right) \coloneqq k_{\beta} \; \bar{\beta}\left(r, 1 \right)$ for some $k_{\beta} \in \mathbb{R}_{> 1}$ to extend the definition of these function for all $k \in \mathbb{Z}_{\left[0,N\right]}$.
Let us select some $\epsilon \in \mathbb{R}_{>0}$ and
\begin{equation*}
\begin{array}{rl}
r_{\max}\coloneqq& \left\{ \displaystyle \bar{\beta}\left(e_{\max},0\right)+\Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right),\right.\\
& \left.\left(1+\epsilon\right)\left(\Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right) \right\}
\end{array}
\end{equation*}
Let us define $\mathcal{N}$ as
\begin{equation}
\begin{array}{rl}
\displaystyle\left(\frac{2\left(1+\epsilon\right)\left(k_1+k_2g\left(k,l\right)^{\eta}\right)e_{\max}^{\zeta}}{r_{\max}}\right)^{1/\eta} \leq & \mathcal{N}
\end{array}
\end{equation}
\begin{remark}
Note that due to the model uncertainty, the horizon must be enlarged.
\end{remark}
Adopting an estimator with a window length greater or equal to $\mathcal{N}$, one will have
\begin{equation} \label{beta equiv for N0}
\bar{\beta}\left(r, \; N \right) \leq \frac{r}{2},
\end{equation}
the effects of the initial conditions will vanish. As $k \rightarrow \infty$, the estimation error will entry to the bounded set $\mathcal{X}\left(w,v \right) \subset \mathcal{X}$ defined by the noises of the system and the uncertainty
\begin{equation}
\mathcal{X} \left(w,v,l \right) \coloneqq \{ \norm{x_{k} -
\hat{x}_{k\vert k}} \leq \left(1 + \epsilon \right)\left(\Phi_w \left(\supofseq{w},1 \right) + \right.
\left. \Phi_v\left(\supofseq{v},1 \right) +\Phi_{\alpha}\left(1,\mathcal{N}\right)\right) \}.
\end{equation}
This set define the minimum size region of error space $\mathcal{X}$ that the error can achieve by removing the effect of errors in initial conditions ($e_{max}$). Equation \eqref{beta equiv for N0} establish a trade off between speed of convergence and window length, which is related with the size of $\mathcal{X}\left(w,v,l \right)$.
For any MHE with adaptive arrival cost and window length $N \geq \mathcal{N}$ two situations can be considered
\begin{itemize}
\item The estimator has removed the effects of $x_0$ on $\hat{x}_{k\vert k}$ such that
$\norm{x_{k}-\hat{x}_{k\vert k}} \in \mathcal{X}\left(w,v,l \right)$, and
\item The estimator has not removed the effects of $x_0$ on $\hat{x}_{k \vert k}$ such that $\norm{x_{k}-\hat{x}_{k\vert k}}
\notin \mathcal{X}\left(w,v,l \right)$,
\end{itemize}
Let us assume that the estimation error is
\begin{equation*}
\norm{x_k-\hat{x}_{k\vert k}}\leq 2\left(1+\epsilon\right)\left(\Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right),
\end{equation*}
the estimation error will be given by
\begin{align}
\norm{x_{k+N}-\hat{x}_{k+N\vert k+N}} \leq&\, \bar{\beta}\left(\norm{x_k-\hat{x}_{k\vert k+N}}, \mathcal{N}\right)
+ \Phi_w\left(w_{\max},1\right)+\nonumber\\
& \Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right),\nonumber \\
\leq &\, \frac{\norm{x_k-\hat{x}_{k\vert k}}}{2\left(1+\epsilon\right)} +\Phi_w\left(w_{\max},1\right)+ \Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\nonumber\\
\leq &\, 2\left(\Phi_w\left(w_{\max},1\right)+ \Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right)\nonumber \\
\leq &\, 2\left(1+\epsilon\right)\left(\Phi_w\left(w_{\max},1\right)+ \Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right)
\end{align}
Therefore, when
\begin{equation*}
\norm{x_k-\hat{x}_{k\vert k}}\leq 2\left(1+\epsilon\right)\left(\Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\overline{K}\left(1\right)\right),
\end{equation*}
the error will no become larger. Assuming now
\begin{equation*}
r_{\max}\geq\norm{x_k -\hat{x}_{k\vert k}}>2\left( 1 + \epsilon \right) \left(\Phi_w\left(w_{\max}\right)+\Phi_v\left(v_{\max}\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right)
\end{equation*}
the estimation error is given by
\begin{equation}
\begin{array}{rl}
\norm{x_{k+N}-\hat{x}_{k+N\vert k+N}} \leq& \bar{\beta}\left(\norm{x_k-\hat{x}_{k\vert k+N}}, \mathcal{N}\right)+\Phi_w\left(w_{\max},1\right)+\nonumber\\
& \Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\nonumber\\
\leq& \frac{\norm{x_k-\hat{x}_{k\vert k+N}}}{2\left(1+\epsilon\right)} + \Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\\
\leq& \frac{\norm{x_k-\hat{x}_{k\vert k+N}}}{2\left(1+\epsilon\right)}+ \frac{\norm{x_k-\hat{x}_{k\vert k+N}}}{2\left(1+\epsilon\right)}\nonumber\\
\leq & \norm{x_k-\hat{x}_{k\vert k+N}}\left(\frac{1}{1+\epsilon}\right)
\end{array}
\end{equation}
with $\xi\coloneqq\left(\frac{1}{1+\epsilon}\right)<1$, since $\epsilon>0$
In the latter case, the estimator error behaves contractively. By mean of some definitions, i.e., $i\coloneqq \lfloor \frac{k}{\mathcal{N}} \rfloor$, $j\coloneqq t \mod \mathcal{N}$, time $k$ can be expressed as $k=i\mathcal{N}+j$. For $N\geq\mathcal{N}$, the $\mathcal{KL}$ functions $\bar{\beta}\left(r,s\right)$ is decreasing every $\mathcal{N}$ samples. Writing the estimation error with this notation for time $k$
\begin{equation}
\begin{array}{rl}
\norm{x_k-\hat{x}_{k\vert k}} \leq& \max\left\{ \xi^i\norm{x_j-\hat{x}_{j\vert k}}, 2\left(1+\epsilon\right) \left(\Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right) \right\} \\
\leq& \bar{\beta}\left(\norm{x_0-\bar{x}_0},j\right)\xi^i + 2\left(1+\mu\right) \left(\Phi_w\left(w_{\max},1\right)+\Phi_v\left(v_{\max},1\right)+\Phi_{\alpha}\left(1,\mathcal{N}\right)\right) \\
\end{array}
\end{equation}
Defining
\begin{equation}
\begin{array}{rl}
\overline{\Phi}\left(\norm{x_0-\bar{x}_0}\right) \coloneqq& \xi^i\bar{\beta}\left(\norm{x_0-\bar{x}_0},j\right), \\
\overline{\Phi}_w\left(w_{\max}\right) \coloneqq& 2\left(1+\mu\right)\Phi_w\left(w_{\max},1\right), \\
\overline{\Phi}_v\left(v_{\max}\right) \coloneqq& 2\left(1+\mu\right)\Phi_v\left(v_{\max},1\right), \\
\overline{\Phi}_{\alpha}\left(l_{\min},\mathcal{N}\right) \coloneqq& 2\left(1+\mu\right)\Phi_{\alpha}\left(1,\mathcal{N}\right).
\end{array}
\end{equation}
Finally we can write
\begin{equation}
\norm{x_k-\hat{x}_{k\vert k}} \leq \overline{\Phi}\left(\norm{x_0-\bar{x}_0},k\right) + \overline{\Phi}_w\left(w_{\max}\right) + \overline{\Phi}_v\left(v_{\max}\right) + \overline{\Phi}_{\alpha}\left(l_{\min},\mathcal{N}\right)
\end{equation}
Taking $w_{\max}$ from $\|\boldsymbol{w}\|_{\left[0,k\right]}$ instead $\|\boldsymbol{w}\|_{\left[k-N,k\right]}$ and $v_{\max}$ from $\|\boldsymbol{v}\|_{\left[0,k\right]}$ instead $\|\boldsymbol{v}\|_{\left[k-N,k\right]}$, and noting that Equation \eqref{eq:RAS} still being valid, the robust regional practical stability is proved.
On the other hand, the convergence of the estimator to the true state in the case of decaying disturbances can be established.
Assuming $\lim_{k\rightarrow\infty}w_k=0$ and $\lim_{k\rightarrow\infty}v_k=0$, given Equation \eqref{eq:RAS}, one can choose some $k\geq K_1$ for which $\max\left\{ w_{\max},v_{\max} \right\} \leq \min \left\{ \Phi_w^{-1}\left(\frac{\varepsilon}{4}\right), \Phi_v^{-1}\left(\frac{\varepsilon}{4}\right) \right\}$. At the same time, one can choose some large enough value of $l$ such that $\overline{\Phi}_{\alpha}\left(l,\mathcal{N}\right) \leq \frac{\varepsilon}{4}$. Note that according to Equation \eqref{g(k,l)}, the value of $l$ will be getting smaller as $\|\boldsymbol{w}\|\rightarrow 0$ and $\|\boldsymbol{v}\|\rightarrow 0$.
Recalling that $\overline{\Phi}\left(\cdot\right) \in \mathcal{KL}$, there will exist some $k\geq K_2$ such that $\overline{\Phi}\left(\norm{x_{k-K_2}-\bar{x}_{k-K_2}},k\right) \leq \frac{\varepsilon}{4}$. Under these conditions, there exists some time $k\geq \max\left\{K_1,K_2 \right\}+\mathcal{N}$ such that
\begin{equation}
\begin{array}{rl}
\norm{x_k-\hat{x}_{k\vert k}} \leq & \overline{\Phi}\left(\norm{x_{k-\mathcal{N}}-\bar{x}_{k-\mathcal{N}}},k\right) + \overline{\Phi}_w\left(w_{\max}\right) + \overline{\Phi}_v\left(v_{\max}\right) + \overline{\Phi}_{\alpha}\left(l,\mathcal{N}\right) \\
\norm{x_k-\hat{x}_{k\vert k}} \leq & \displaystyle\frac{\varepsilon}{4}+\frac{\varepsilon}{4}+\frac{\varepsilon}{4}+\frac{\varepsilon}{4} \\
\norm{x_k-\hat{x}_{k\vert k}} \leq & \varepsilon
\end{array}
\end{equation}
Since one can choose any value of $\displaystyle\varepsilon$, $\lim_{k\rightarrow \infty} \norm{x_k-\hat{x}_{k\vert k}} = 0$ can be guaranteed when $\lim_{k\rightarrow\infty}w_k=0$ and $\lim_{k\rightarrow\infty}v_k=0$ as claimed.
\qed
\begin{remark}
As expected, the model uncertainty deteriorates state estimation. However, for a large enough value of $l$, this effect can be mitigated. Moreover, when $\lim_{l\rightarrow\infty}\Phi_{\alpha}\left(l,\mathcal{N}\right) = 0$
\end{remark}
\section{Simulation and results}
The following examples will be used to illustrate the results presented in the previous sections and to evaluate and compare the performance of the proposed estimator with others from the state of the art. The process disturbance is white Gaussian noise acting as an additive exogenous input to the system.
\subsection{Unknown linear system}
Let us considers the linear system
\begin{align*}
x_{k+1} &= A_p x_k + w_k \\
y_k &= C_p x_k + v_k
\end{align*}
whose matrices are unknown an they have the following structure
\begin{align} \label{linear sys}
A_p &= \left[ \begin{array}{cc}
0 & a_1 \\
1 & a_2
\end{array} \right],
C_p = \left[ \begin{array}{cc}
c_1 & c_2
\end{array} \right]
\end{align}
The system is affected with additive process and measurement noise $w$ and $v$ drawn from normal distributions with zero mean and covariance $Q_w = S_w^2 I_2$ and $R_v = S_v^2$, respectively. The polytope is defined using three \emph{LTI} models
\begin{equation} \label{linear_pol}
\begin{array}{c}
A_1 = \left[ \begin{array}{cc}
0 & 0.72 \\
1 & 0.28
\end{array} \right],
A_2 = \left[ \begin{array}{cc}
0 & -0.59 \\
1 & 1.57
\end{array} \right],
A_3 = \left[ \begin{array}{cc}
0 & -0.35 \\
1 & 1.26
\end{array} \right], \\
C_1 = \left[ \begin{array}{cc}
-1.46 & -1.29 \\
\end{array} \right],
C_2 = \left[ \begin{array}{cc}
-4.84 & -2.90 \\
\end{array} \right],
C_3 = \left[ \begin{array}{cc}
-0.09 & -0.03
\end{array} \right],
\end{array}
\end{equation}
such that the system belongs to it.
The matrices of the system $\left(A_{S}, C_{S} \right)$ and its model $\left(A_{M}, C_{M} \right)$ were generated as a convex combination of the polytope with mixing parameters $\mathbf{\alpha_{S}}$ ($\alpha_{S,1}=0.22, \alpha_{S,2}=0.76, \alpha_{S,3}=0.02$) and $\mathbf{\alpha_{M}}$ ($\alpha_{M,1}=0.41, \alpha_{M,2}=0.22, \alpha_{M,3}=0.37$), respectively.
The stage cost of the receding horizon estimators is chosen as $\ell(w,v) = w^T Q^{-1} w + v^T R^{-1} v$ with $R^{-1} = 5$ and $Q^{-1}=diag\left(0.1, 0.1\right)$.
The proposed moving horizon estimator ($MHE_{A}$) the prior weighting matrix is given by $\Gamma_{k-N} \left(\chi\right) = (\chi - \hat{x}(0|k))^T P^{-1}_{k|k} (\chi - \hat{x}(0|k))$, where $P^{-1}_0 = 0.1 \boldsymbol{I}_2$ and $P_{k|k}$ is updated using equations \eqref{eq:updatePk} and \eqref{eq:updateXA} with $\sigma = 1e^{-4}$ and $c=5$.
The robust moving horizon estimator ($MHE_{R}$) implements the algorithm proposed by \cite{muller2017nonlinear} with the nominal model ($A_M,C_M$), the prior weight given by $\Gamma \left(\chi\right) = L(\chi - \hat{x}(0|k))^T (\chi - \hat{x}(0|k))$ and parameters $\delta = 1$, $\delta_1 = \kappa^N$ ($\kappa=0.89$) and $\delta_2 = 1/N$ (see equation (3) of \cite{muller2017nonlinear}).
The full information (\emph{FIE}, see \cite{ji2016robust}) is configured with the true model and the same parameters used by the $MHE_{R}$ with $\delta = 1$, $\delta_1 = \kappa^k$ and $\delta_2 = 1/k$ with the system matrices ($A_S,C_S$)
The robust Kalman filter ($KF_R$) was designed following the design procedure proposed by \cite{zhu2002design} using the nominal model ($A_M,C_M$) and computing the bounds from the models of the polytope. The Kalman filter ($KF$) was designed using the matrices of the system ($A_S,C_S$)
\begin{table}[th]
\centering
\caption{Averaged MSE for $S_w = 0.1, S_v = 0.05, N=8$.}
\begin{tabular}{|l|c|c|c|c|c|} \hline
& \small$KF_{R}$ & \small $KF$ &\small $FIE$ &\small $MHE_{R}$ &\small $MHE_{A}$ \\ \hline
$x_0$ &\small0.78467 &\small0.030851 &\small0.0054 &\small0.2662 &\small0.0212 \\ \hline
$x_1$ &\small1.9946 &\small0.069122 &\small0.0039 &\small0.4675 &\small0.0389 \\ \hline
\end{tabular} \label{tab:linear system 1}
\end{table}
Table \ref{tab:linear system 1} shows the mean square estimation error ($MSE$) of each estimator averaged over 100 trials for $S_w = 0.1$, $S_v = 0.05$ and $N=8$ for all receding horizon estimators ($FIE$, $MHE_R$ and $MHE_A$). It can be seen that the proposed estimator average mean square estimation error is smaller than the Kalman filters ($KF$ and $KF_R$) and $MHE_R$. Only the $FIE$ provides better performance than the proposed algorithm. The performance difference between the estimators that employ the nominal model ($KF_R$ and $MHE_R$) is due to the adaptation capabilities of $MHE_A$ that allows to reduce the uncertainty of the estimator model. The main performance difference between $MHE_A$ and $FIE$ estimators is due to the model employed by estimator and the amount of information employed to estimate $\hat{x}_{k|k}$. While the $FIE$ estimator use of the exact model and all the system output available until $k$, the $MHE_A$ identifies the model in the initial samples and only use the last $N$ system outputs.
\begin{figure}[tb]
\centering
\includegraphics[width=0.90\textwidth]{Figures/Nvsl.png}
\caption{Estimation error for different values of $l$ and $N$.}
\label{fig:decreasing cost with l 3D}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=0.80\textwidth]{Figures/evolution_alpha_Sw0-01-05-1.png}
\caption{Estimated mixing parameters $\hat{\alpha}$ for $N=8$, $l=$, and noises variance (\myinvertedtriangle{white} $S_w=0.0$, \mysquare{white} $S_w=0.1$, \mytriangle{white} $S_w=0.5$ and \mycircle{white} $S_w=1.0$)}
\label{fig:estimation of alpha}
\end{figure}
Figure \ref{fig:decreasing cost with l 3D} shows the behaviour of the estimation error as a function of $l$ and $N$. This figure shows that the main factor in the reduction of the estimation error is the number of iterations $l$ used to update the estimates. It can be also seen that there a significant improvement in the initial iterations ($l < 10$), then after iterations there is no significant improvement in the estimation error. It is worth nothing that the error is decreasing with the iterations as it was shown in Section 4.
Figure \ref{fig:estimation of alpha} shows the time evolution of the estimated vector of mixing parameters $\hat{\alpha}$ for different values of process noise variance. The true values are representing as continuous line. When $S_w$ is smaller than the value of states ($S_w \leq 0.5$), the mixing parameters $\hat{\alpha}$ converge quickly to the true value or remain closer to it.
\subsection{Example 2: Nonlinear time-varying system}
As a second example, we consider a second order time--varying nonlinear system whose dynamic is given by
behavior
\begin{equation}
\begin{array}{rl}
x_{k+1} &= \left[ \begin{array}{c}
x_2\left(k\right) \\
p_1\left(k\right)x_1\left(k\right) + \sin(p_2\,x_2\left(k\right))
\end{array} \right] + w_k \\
y_k &= C x_k + v_k
\end{array}
\end{equation}
with the parameters $p_1(k)$ and $p_2(k)$ given by
\begin{equation}
\begin{array}{rlr}
p_1\left(k+1\right) &= 0.01 \, p_1 \left(k\right)\sin\left(\frac{5\pi k}{N}\right) & \hspace{0.5cm} \forall k\in \mathbb{Z}_{\left[1,3N/4\right]}, \\
p_1\left(k+1\right) &= p_1\left(k\right) & \hspace{0.5cm} \forall k \in \mathbb{Z}_{> 3N/4}, \\
p_2 &= 0.05
\end{array}
\end{equation}
The polytope was designed to guarantee that the nonlinear system always remains inside it. The polytope is defined using three \emph{LTI} models
\begin{equation}
\begin{array}{c}
A_1 = \left[ \begin{array}{cc}
0 & 1.30 \\
1 & -1.52
\end{array} \right],
A_2 = \left[ \begin{array}{cc}
0 & -2.44 \\
1 & 0.66
\end{array} \right],
A_3 = \left[ \begin{array}{cc}
0 & 1.31 \\
1 & 2.81
\end{array} \right],
C = \left[ \begin{array}{cc}
0.5 & 0.5 \\
\end{array} \right].
\end{array}
\end{equation}
The stage cost for all receding horizon estimators is chosen as $\ell(w,v) = w^T Q^{-1} w + v^T R^{-1} v$ with $R^{-1} = 5e^2$ and $Q^{-1}=diag\left(1e^3, 5e^3\right)$.
The proposed moving horizon estimator ($MHE_{A}$) the prior weighting matrix is given by $\Gamma_{k-N} \left(\chi\right) = (\chi - \hat{x}(0|k))^T P^{-1}_{k|k} (\chi - \hat{x}(0|k))$, where $P^{-1}_0 = 0.1 \boldsymbol{I}_2$ and $P_{k|k}$ is updated using equations \eqref{eq:updatePk} and \eqref{eq:updateXA} with $\sigma = 1$ and $c=1$.
The robust moving horizon estimator ($MHE_{R}$) implements the algorithm proposed by \cite{muller2017nonlinear} with the nominal model, the prior weight given by $\Gamma \left(\chi\right) = L(\chi - \hat{x}(0|k))^T (\chi - \hat{x}(0|k))$ and parameters $\delta = 1$, $\delta_1 = \kappa^N$ ($\kappa=0.89$) and $\delta_2 = 1/N$ (see equation (3) of \cite{muller2017nonlinear}).
The full information (\emph{FIE}, see \cite{ji2016robust}) is configured with the linearized model updated at each sampling time and the same parameters used by the $MHE_{ROB}$ with $\delta = 1$, $\delta_1 = \kappa^k$ and $\delta_2 = 1/k$.
The robust Kalman filter was designed following the design procedure proposed by \cite{zhu2002design} using the nominal model and computing the bounds from the models of the polytope. The guess for the initial condition is $\bar{x}_0 = \left[0,0 \right]^T$, whereas $x_0 = \left[ 0.5, 0.3\right]^T$.
\begin{table}[!h]
\centering
\caption{Averaged MSE for $Sw = 0.1$, $Sv = 0.05$ and $N=8$.}
\begin{tabular}{|l|c|c|c|c|} \hline
& \small$EKF_{R}$ & \small $FIE$ &\small $MHE_{R}$ &\small $MHE_{A}$ \\ \hline
$x_0$ &\small0.40879 &\small0.32324 &\small0.31541 &\small0.018498 \\ \hline
$x_1$ &\small0.4297 &\small0.30082 &\small1.0106 &\small0.064984 \\ \hline
\end{tabular} \label{tab:average_estimation_error_for_nls}
\end{table}
Table \ref{tab:average_estimation_error_for_nls} shows the mean square estimation error ($MSE$) of each estimator averaged over 100 trials for $S_w = 0.1$, $S_v = 0.05$ and $N=8$ for all receding horizon estimators ($FIE$, $MHE_{R}$ and $MHE_{A}$). It can be seen that the average mean square estimation error of $MHE_{A}$ is smaller than the other estimators ($EKF_{R}$, $MHE_{R}$ and $FIE$).
The performance difference between the estimators that employ the nominal model ($EKF_{R}$ and $MHE_{R}$) is mainly due to the adaptation capabilities of $MHE_{A}$. The main performance difference between $MHE_A$ and $FIE$ estimators is due to the FIE attempts to reconstruct the state trajectory of a nonlinear time-varying system with a \emph{LTI} system.
\begin{figure}[b]
\centering
\includegraphics[width=0.90\textwidth]{Figures/NvslNL.png}
\caption{Performance of the proposed algorithm for different values of $l$ and $N$.}
\label{fig:error NLS Nvsl}
\end{figure}
Figure \ref{fig:error NLS Nvsl} shows the behaviour of the estimation error as a function of $l$ and $N$. This figure shows that the main factor in the reduction of the estimation error is the number of iterations $l$ used to update the estimates. It can be also seen that there a significant improvement in the initial iterations ($l < 15$), then after iterations there is no significant improvement in the estimation error.
It can also see how the estimation error increases for higher values of $N$. This behavior is due to the estimator use only one model along the entire estimation horizon, whereas the nonlinear system is changing its parameters every sample.
The $EKF$ aim to improve the estimation error in comparison with the $MHE_{A}$. This improvement is due to the $EKF$ update the model at every sampling time with the true model. Besides, the $MHE_{A}$ use the same model along the entire horizon estimation to estimate the optimal state trajectory of the nonlinear system.
\begin{figure}[tb]
\centering
\includegraphics[width=0.90\textwidth]{Figures/evolution_parameters__Sw01-025-05.png}
\caption{Estimation of the parameters of the nonlinear time-varying system ( $N=8$, and noises \mysquare{white}=0.1, \mytriangle{white}=0.25 and \mycircle{white}=0.5).}
\label{fig:parameters_estimated_NLS}
\end{figure}
Figure \ref{fig:parameters_estimated_NLS} shows the time evolution of the estimated parameters $\hat{p}_1$ and $\hat{p}_2$ for different values of process noise variance. The true values are representing as continuous line. When $S_w$ is smaller than the value of states ($S_w \leq 0.25$), the parameters $\hat{p}$ converge quickly to the true value or remain closer to it. However,
\section{Conclusions}
In this paper we introduce an adaptive polytopic observer for nonlinear systems under bounded disturbances based on moving horizon estimator and dual estimation techniques and proved their stability properties. In a first stage we proved the stability of the dual estimation iteration. Then, in a second stage we established robust
asymptotic stability for the adaptive moving horizon estimator. It was also shown that the estimation error converges to zero in case that disturbances converge to zero.
An advantage of this updating mechanism is that the required conditions on prior weighting are such that it can be chosen off-line. Furthermore, it introduces a feedback mechanism between the arrival cost weight and the estimation errors that automatically controls the amount of information used to compute it, which allows to shorten the estimation horizon.
\begin{ack}
The authors wish to thank the Consejo Nacional
de Investigaciones Cientificas y Tecnicas (CONICET)
from Argentina, for their support.
\end{ack}
\bibliographystyle{agsm
\clearpage
|
\section{Introduction}
Among the many odd properties of neutron stars (NS), one of the weirdest is certainly the existence of stable hyperon and muon layers. Much like the neutrons, these usually unstable particles are stabilized by the degenerate Fermi-gas phase, which Pauli-blocks the would-be daughter particles of a decay.
A typical old NS contains roughly $10^{55}$ muons, compared to about $10^{57}$ neutrons~\cite{Bahcall:1965zza,Bahcall:1965zzb,Potekhin:2013qqa,Goriely:2013xba,Goriely:2010bm}. This lets NS interact with and potentially capture dark matter (DM) that couples mainly to muons, unique among astrophysical objects~\cite{Garani:2018kkd}.
Muonphilic DM might seem far-fetched, but considering our lengthy and so far unsuccessful quest for DM that interacts with first-generation particles in (in)direct detection experiments and at colliders it behooves us to consider alternatives to standard WIMPs~\cite{Bertone:2018xtm,Kopp:2009et,Kile:2014jea}. Furthermore, several anomalies that hint at physics beyond the Standard Model (SM) reside in the muon sector, e.g.~$(g-2)_\mu$~\cite{Bennett:2006fi,Keshavarzi:2018mgv} and $B\to K^{(*)}\mu^+\mu^-$~\cite{Aaij:2014ora,Aaij:2017vbb,Aaij:2019wad}, and typically require dominantly muonphilic interactions. A popular example here is the gauge boson $Z'$ of an anomaly-free $U(1)_{L_\mu-L_\tau}$~\cite{He:1990pn,Foot:1990mn,He:1991qd,Heeck:2011wj}, which can resolve either the $(g-2)_\mu$ anomaly~\cite{Gninenko:2001hx,Baek:2001kca,Carone:2013uh,Altmannshofer:2014pba} or the $B$-meson discrepancies~\cite{Altmannshofer:2014cfa,Crivellin:2015mga,Crivellin:2015lwa,Altmannshofer:2016jzy} depending on the $Z'$ mass $m_{Z'}$ and its gauge coupling $g'$. It is but a small step to connect such a new particle to DM, e.g., by using the $U(1)_{L_\mu-L_\tau}$ charge as a DM stabilization mechanism and the $Z'$ interactions to obtain the correct relic abundance. This results in dominantly \emph{muonphilic} DM, which can only interact non-gravitationally with NS.
Assuming a large enough muon--DM cross section to capture a significant number of DM particles, the infalling DM unavoidably transfers heat to the NS, see
Refs.~\cite{Baryakhtar:2017dbj,Raj:2017wrv,Bell:2018pkk,Camargo:2019wou,Bell:2019pyc}.
This can increase the temperature of old NS from $\mathcal{O}(\unit[100]{K})$ to $\mathcal{O}(\unit[2000]{K})$, leading to an infrared blackbody spectrum that is in principle within range of future telescopes such as the James Webb Space Telescope, the Thirty Meter Telescope, or the European
Extremely Large Telescope~\cite{Baryakhtar:2017dbj}.
Further heating can occur through DM annihilations inside the NS.
In this paper we will discuss this capture of muonphilic DM in NS and the resulting heating, allowing for both DM and the mediator to be light to go beyond existing calculations. As a well-motivated example for muonphilic DM we consider an $U(1)_{L_\mu-L_\tau}$ model that can also ameliorate existing anomalies in the muon sector.
\section{Dark matter in neutron stars}
\label{sec:neutron_stars}
The capture of DM particles by the NS depends on the macroscopic NS properties such as its mass $M_\star$ and radius $R_\star$, and also on the properties of its Fermi-degenerate medium, especially chemical potential and lepton fractions. We consider a NS profile based on realistic two- and three-nucleon forces obtained in Refs.~\cite{Potekhin:2013qqa,Goriely:2013xba}, where phenomenological fits are performed for equations of state (EoS) of varying stiffness, neglecting the possibility of exotic quark phases. We present our results for the low-mass configuration of model BSK20 which corresponds to the NS values $M_\star = 1.52\,{\rm M_\odot}$, $R_\star=\unit[11.6]{km}$, the core muon fraction $Y_\mu=2\times 10^{-2}$, and muon chemical potential in the core $\mu_\mu =\unit[65]{MeV}$.
The above parameters for the NS are consistent with the observed properties of old NS. For example, number fractions of leptons in the core of a NS are important in the theory of NS cooling as they determine whether the direct Urca processes of neutrino emission operate efficiently or not~\cite{Klahn:2006ir}. Existence of beta equilibrium via the reactions $n \to p + e^- + \bar\nu_e$, $n \to p + \mu^- + \bar\nu_\mu$, and
$e^- \to \mu^- + \nu_e + \bar\nu_\mu$ together with charge neutrality ensures $\mu_n = \mu_p +\mu_e = \mu_p +\mu_\mu $ for the chemical potentials and $n_p =n_e +n_\mu$ for the number densities~\cite{Cohen1970}. Observations indicate that the direct Urca processes operate only in a relatively small number of NS, which could be explained by the so called {\it minimal cooling paradigm}, i.e.~not involving the direct Urca process~\cite{Page:2004fy,Pearson:2018tkr}. Moreover, the exceptions to the above paradigm can be explained by internal heating mechanisms~\cite{Potekhin:2015qsa} (none of which require DM inside the NS). All of the above indicate that an acceptable EoS should not allow the direct Urca process to occur in NS with masses below $\unit[1.5]{M_\odot}$~\cite{Klahn:2006ir}.
More recently, advanced versions of phenomenological EoS called BSK24 and BSK26~\cite{Pearson:2018tkr} were constructed by adding extra energy density functionals and by considering new atomic mass evaluation data. Since these models are fitted to the same EoS of moderate stiffness of Akmal--Pandharipande--Ravenhall~\cite{Akmal:1998cf} as that of BSK20, the qualitative predictions are very similar~\cite{Pearson:2018tkr}. Also note that these EoS are still allowed by the latest constraints from LIGO's observation of a binary NS merger~\cite{Abbott:2018exr,Most:2018hfd}. We refer to App.~A of Ref.~\cite{Garani:2018kkd} for more details regarding radial profiles of BSK20 and possible uncertainties on the DM capture rate.
\paragraph{Dark matter capture:}
The \emph{maximal} capture rate, i.e.~the rate at which all DM particles $\chi$ with mass $m_\chi$ that intercept the NS are captured, is given by the geometric rate~\cite{Garani:2018kkd}
\begin{align}
C_\star \simeq \frac{5.6 \times 10^{25}}{\unit{s}} \, \frac{\rho_\chi}{\unit{GeV/cm^3}} \, \frac{\unit{GeV}}{m_\chi} \, \frac{R_\star}{\unit[11.6]{km} }\,\frac{ M_\star}{\rm 1.52\,M_\odot} \,,
\label{eq:capturegeom-sim}
\end{align}
with DM energy density $\rho_\chi$ around the NS.
For DM--muon contact interactions this rate can be achieved for cross sections above $\pi R_\star^2/N_\mu \simeq \unit[5 \times 10^{-43}]{cm^2}$.
The geometric rate $C_\star $ is shown in Fig.~\ref{fig:cap-muons-med} as a black dashed line fixing $\rho_\chi=\unit[0.4]{GeV/cm^3}$.
\begin{figure}
\includegraphics[width=0.49\textwidth]{cap-muons-med}
\caption{DM capture rates on muons in NS for several $Z'$ masses (colored lines) as well as the maximal geometric capture rate $C_\star$ (black dashed line). The solid colored lines correspond to the exact capture rate~\cite{Garani:2018kkd}, the dotted lines show the approximation from Eq.~\eqref{eq:cap-approx}.
}
\label{fig:cap-muons-med}
\end{figure}
We will consider interactions between DM particles $\chi$ and muons that are mediated by a potentially light gauge boson $Z'$ with vector coupling $g' q_\mu$ to muons and coupling $g' q_\chi$ to $\chi$.
The differential elastic scattering cross section in the non-relativistic limit is then
\begin{equation}
\frac{{\rm d} \sigma }{{\rm d} E_R}(\chi\mu \to\chi\mu) = \frac{ (g^\prime)^4 q_\chi^2 q_\mu^2 }{2 \pi}\frac{m_\mu}{w^2(2 m_\mu E_R + m_{Z^\prime}^2)^2}\, ,
\end{equation}
valid both for scalar and fermion DM.
$w^2 = u_\chi^2 +v_\text{esc}^2$ is the squared speed of DM inside the NS, $E_R = m_\text{red}^2 w^2 (1- \cos \theta_\text{CM})/m_\mu$ is the recoil energy in the muon rest frame, and $m_\text{red}$ the reduced mass of the DM--muon system.
The rate of accretion of DM particles from the halo into a NS by scattering off muons with density $n_\mu$ then reads~\cite{Feng:2016ijc}, in the zero-temperature limit,
\begin{eqnarray}
C &\simeq& \int_0^{R_\star} {\rm d} r \, 4 \pi r^2 n_\mu(r) \int_0^\infty {\rm d} u_\chi \, \left(\frac{\rho_\chi}{m_\chi}\right) \, f_{v_\star}(u_\chi) u_\chi \, \nonumber \\
&& \times w(r)^2 \, \zeta(r) \int_{E_R^\text{min}}^{E_R^\text{max}} {\rm d} E_R\, \frac{{\rm d} \sigma}{{\rm d} E_R} \, .
\label{eq:cap-approx}
\end{eqnarray}
To approximate Pauli blocking we introduce the factor $\zeta(r) \equiv {\rm Min}[1,\delta p(r)/p_\text{F}(r)]$ following Refs.~\cite{McDermott:2011jp,Bell:2013xk,Garani:2018kkd}.
Here $\delta p(r) \simeq \sqrt{2} m_\text{red} v_\text{esc}(r)$ is the typical momentum transfer in the muon rest frame and $p_\text{F}(r)=\sqrt{2 m_\mu \mu_\mu(r)}$ the Fermi momentum.
$f_{v_\star}$ is the DM velocity distribution in the NS frame and $u_\chi$ the DM speed in the halo~\cite{Garani:2018kkd}. To capture a DM particle from the halo, the recoil energy should be between
\begin{align}
E_R^\text{min} = m_\chi u_\chi^2/2 && \text{ and } && E_R^\text{max} = 2 m_\text{red}^2 w^2/m_\mu \,.
\end{align}
The $Z'$-mass dependence of the capture rate is given by
\begin{align}
\int_{E_R^\text{min}}^{E_R^\text{max}} &{\rm d} E_R\, \frac{{\rm d} \sigma}{{\rm d} E_R} = \frac{ (g^\prime)^4 q_\chi^2 q_\mu^2}{4 \pi w^2} \, \Theta(4 w^2 m_\text{red}^2 - m_\mu m_\chi u_\chi^2) \nonumber \\
& \times\frac{(4 w^2 m_\text{red}^2 - m_\mu m_\chi u_\chi^2)}{(m^2_{Z^\prime} + m_\mu m_\chi u_\chi^2) (m^2_{Z^\prime} + 4 m_\text{red}^2 w^2)} \,.
\end{align}
For $m_\chi \gg m_\mu > m_{Z^\prime}$ the capture cross section and rate thus become independent of the mediator mass.
For a close-by NS we take the local DM density $\rho_\chi=\unit[0.4]{GeV/cm^3}$ and present the DM capture rate on muons in Fig.~\ref{fig:cap-muons-med}. The colored dotted lines correspond to the capture rate obtained using Eq.~\eqref{eq:cap-approx}, whereas the solid curves are more accurate rates which are numerically obtained by using the methods of Ref.~\cite{Garani:2018kkd}, where possible non-trivial kinematics due to the degenerate nature of NS matter are taken into account.
For $m_{Z^\prime} >m_\chi > m_\mu$ the rate $C$ scales as $m_\chi^{-1}$; for $m_\chi \lesssim p_\text{F}/( \sqrt{2} v_\text{esc}) < m_{Z^\prime}$, Pauli blocking of the final state muon becomes efficient and the capture rate saturates and scales as $m_\chi^2$. For $m_{Z^\prime} < m_\mu$, Pauli blocking instead becomes important for $m_\chi \lesssim m_{Z^\prime}$. The full calculation of the capture rate differs from Eq.~(\ref{eq:cap-approx}) only when Pauli blocking is relevant and suppresses $C$ by a factor up to $ 35$. Thermal effects in the capture rate are negligible for the range of DM masses of interest here.
The main theoretical uncertainties in our calculation are the NS muon content and chemical potential, which should be at most off by a factor of two~\cite{Garani:2018kkd} (see also Ref.~\cite{Bell:2019pyc}).
\paragraph{Neutron star heating:}
As shown in Refs.~\cite{Baryakhtar:2017dbj,Raj:2017wrv,Bell:2018pkk,Camargo:2019wou,Bell:2019pyc}, the NS temperature increase due to DM capture is
\begin{align}
T_\text{\rm kin} \simeq \unit[1700]{K}\left(\frac{C}{C_\star} \right)^{1/4}\left(\frac{\rho_\text{DM}}{\unit[0.4]{GeV/cm^3}} \right)^{1/4} .
\end{align}
As can be appreciated from Fig.~\ref{fig:cap-muons-med}, it is not difficult to saturate $C\sim C_\star$ for much of the interesting parameter space.
The SM--DM interaction underlying $C$ is of course impossible to reconstruct from $T_\text{\rm kin}$, but the observation of NS with $T < T_\text{\rm kin} $ in DM-rich environments would still allow us to set limits on DM--$\mu$ interactions. Importantly, truly muonphilic DM can \emph{only} be captured in NS, much like inelastic DM~\cite{Baryakhtar:2017dbj,Bell:2018pkk}, whereas most other DM models also allow for capture in other objects, including Earth.
Additional NS heating comes from annihilation of symmetric DM into SM particles. NS are expected to cool via neutrino emission through the modified Urca process for the first million years and through photon emission afterwards~\cite{Kouvaris:2010vv,Kouvaris:2007ay}. Once the rate of DM accretion equilibrates with DM annihilation the emissivity due to DM annihilation does not depend on temperature. This implies that once the temperature of the NS is sufficiently small, the power from DM annihilations that heat up the NS equals that of the photon emission, resulting in a constant temperature. For an old NS ten parsecs away, the maximal heating due to annihilations gives $T_\text{ann} \simeq \unit[2480]{K}\,[\rho_\text{DM}/(\unit[0.4]{GeV/cm^3})]^{0.45}$~\cite{Kouvaris:2007ay,Baryakhtar:2017dbj}.
Finally, comments on the impact of the NS profile on DM constraints from NS heating are in order. As discussed above the NS temperature increase due to capture and annihilation of DM are proportional to the capture rate, so uncertainties in the capture rate directly translate to the uncertainties on $T_{\rm kin}$ and $T_{\rm ann}$. In Ref.~\cite{Garani:2018kkd} it is found that the capture rate on muons differ by a factor $\simeq 5$ between the low-mass configuration of BSK20 ($\mu_\mu =\unit[65]{MeV}$, $N_\mu = 10^{55}$) and the high-mass configuration of BSK21 ($\mu_\mu =\unit[160]{MeV}$, $N_\mu = 3.8\times 10^{55}$). Note that $\mu_n$ and $N_\mu$ are not independent parameters. Consequently, the constraints obtained on the coupling will at most vary by a factor three. It is also important to note that uncertainties due to the NS profile choice could be as large as the uncertainty on the local DM density.
For \emph{asymmetric} DM~\cite{Zurek:2013wia}, which cannot self-annihilate by construction, it seems possible to accumulate enough DM inside the NS to form a black hole that destroys its host~\cite{Goldman:1989nd}. The observation of very old NS in DM-rich regions would then allow us to put strong constraints on, e.g., the muon--DM cross section~\cite{Garani:2018kkd}. \emph{Fermionic} DM is typically difficult to constrain in this way due to the additional pressure from the Pauli exclusion principle. \emph{Bosonic} DM can in principle collapse much easier; however, as pointed out in Refs.~\cite{Kouvaris:2011fi,Bell:2013xk}, DM self-interactions play an important role and can hinder black hole formation even for bosonic DM, as discussed below.
\section{Dark matter charged under \texorpdfstring{$L_\mu-L_\tau$}{Lmu-Ltau} }
\label{sec:model}
\begin{figure}[tb]
\includegraphics[width=0.48\textwidth]{Lmu-Ltau_limits.pdf}
\caption{
Limits on the $U(1)_{L_\mu-L_\tau}$ gauge boson mass $m_{Z'}$ and coupling $g'$, assuming vanishing kinetic mixing.
The dark (light) green region can resolve the $(g-2)_\mu$ anomaly~\cite{Keshavarzi:2018mgv} at $2\sigma$ ($1\sigma$), the cyan region ameliorates the Hubble $H_0$ tension~\cite{Escudero:2019gzq}, and the yellow region can resolve the $b\to s\mu^+\mu^-$ anomaly~\cite{Altmannshofer:2016jzy} while satisfying $B_s$--$\bar{B}_s$ mixing constraints~\cite{DiLuzio:2017fdq}.
The other shaded regions are excluded by $N_\text{eff}$~\cite{Kamada:2015era,Kamada:2018zxi}, BaBar~\cite{TheBABAR:2016rlg}, CMS~\cite{Sirunyan:2018nnz}, LEP~\cite{Altmannshofer:2014cfa}, and neutrino trident production in CCFR~\cite{Mishra:1991bv,Altmannshofer:2014pba}.
The dashed lines show the expected sensitivities of Belle-II~\cite{Jho:2019cxq}, DUNE~\cite{Altmannshofer:2019zhy,Ballett:2019xoj} and NA62 (in $K\to \mu +\text{inv}$)~\cite{Krnjaic:2019rsv}. Not shown are the sensitivities of M$^3$~\cite{Kahn:2018cqs} and NA64$\mu$~\cite{Gninenko:2014pea,Chen:2018vkr}.
The two black stars denote the benchmark values used in the main text.
}
\label{fig:Lmu-Ltau_limits}
\end{figure}
As a well-motivated example for muonphilic DM we consider an extension of the SM by the anomaly-free gauge group $U(1)_{L_\mu-L_\tau}$~\cite{He:1990pn,Foot:1990mn,He:1991qd,Heeck:2011wj}. Its gauge boson $Z'$ does not couple to first-generation particles and is thus only weakly constrained, as shown in Fig.~\ref{fig:Lmu-Ltau_limits}. It is arguably the simplest explanation of the $(g-2)_\mu$ anomaly if $g'\sim 5\times 10^{-4}$ and $M_{Z'} \sim 10$--$\unit[100]{MeV}$~\cite{Altmannshofer:2014pba}, fully testable with currently running experiments. For masses around $\unit[10]{MeV}$ this $Z'$ could furthermore ameliorate the observed tension in the Hubble parameter $H_0$ by contributing slightly to $N_\text{eff}$~\cite{Escudero:2019gzq}, and could also affect the high-energy neutrino flux measured in IceCube~\cite{Araki:2014ona,Araki:2015mya}.
In a different region of parameter space, a heavy $L_\mu-L_\tau$ gauge boson could resolve the persistent anomalies in $b\to s\mu^+\mu^-$ transitions, as long as additional $Z_\mu'$ couplings to $\overline{b}\gamma^\mu P_L s$ are generated, e.g.~via vector-like fermions~\cite{Altmannshofer:2014cfa,Crivellin:2015mga,Crivellin:2015lwa,Altmannshofer:2016jzy}. This requires $m_{b}\ll m_{Z'}$ in order to obtain the desired contact-operator coefficient $C_9^\mu \simeq -0.95$~\cite{Aebischer:2019mlg} (see also~\cite{Alguero:2019ptt,Alok:2019ufo,Ciuchini:2019usw,Datta:2019zca,Kowalska:2019ley}) and $m_{Z'}/g' \lesssim \unit[1.1]{TeV}$ to evade the strong constraints from $B_s$--$\bar{B}_s$ mixing (at $2\sigma$)~\cite{DiLuzio:2017fdq}. In the following we will focus on two benchmark values for $L_\mu-L_\tau$ that satisfy the existing constraints from Fig.~\ref{fig:Lmu-Ltau_limits} and could play a role either for $(g-2)_\mu$ and $H_0$ ($m_{Z'}=\unit[10]{MeV}$, $g'=5\times 10^{-4}$) or $b\to s\mu^+\mu^-$ ($m_{Z'}=\unit[100]{GeV}$, $g'=0.1$).
\begin{figure*}[t]
\includegraphics[height=0.38\textwidth]{NSexclusion10MeV.pdf}
\includegraphics[height=0.38\textwidth]{NSexclusion100GeV.pdf}
\caption{
Complex scalar $L_\mu-L_\tau$ DM $\chi$ for two benchmark values of $m_{Z'}$ and $g'$, one relevant for $(g-2)_\mu$ (left) and one for $b\to s\mu^+\mu^-$ (right). The black line denotes the WIMP scenario where the thermal symmetric abundance $\Omega_\text{sym}$ matches the observed DM abundance $\Omega_\text{obs}$; above is the region of asymmetric DM.
In the orange parameter space we expect kinetic heating of old nearby NS to $\unit[1700]{K}$ by DM capture on muons; in the red region additional annihilation increases the NS temperature to $\unit[2500]{K}$.
The blue line shows the DM self-interaction transfer cross section $\sigma_T (\chi\chi\to\chi\chi)/m_\chi = \unit[1]{cm^2/g}$ that can resolve small-scale structure issues~\cite{Kaplinghat:2015aga}, esp.~if suppressed on cluster scales as here for $ m_\chi\gtrsim\unit{GeV}$ (green ellipse).}
\label{fig:abundance}
\end{figure*}
DM charged under $U(1)_{L_\mu-L_\tau}$ has been discussed for some time, typically as a WIMP with thermal abundance~\cite{Cirelli:2008pk,Baek:2008nz}. An appropriately chosen $L_\mu-L_\tau$ charge $q_\chi$ can render a new particle $\chi$ stable with \emph{muonphilic} $Z'$ connection to the SM -- the DM interactions with tauons being practically irrelevant.
For most values of $q_\chi$ the full Lagrangian actually has an additional \emph{global} symmetry $U(1)_\chi$ that corresponds to conserved DM number. In analogy to the observed baryon number asymmetry one could assume that the cosmological history also led to a DM asymmetry~\cite{Zurek:2013wia} which results in a density $\Omega_\text{asym}$ proportional to the asymmetry.
Unless the $Z'$ couplings are tiny, DM $\chi$ will also be produced thermally, leading to a \emph{symmetric} DM component~$\Omega_\text{sym}$. We will consider both options here but remain agnostic about the potential DM asymmetry.
As a concrete example we will discuss \emph{scalar} DM, the fermionic case is similar.
For the two benchmark points specified earlier we show the scalar-DM parameter space in Fig.~\ref{fig:abundance}, obtained using {\sc MicrOMEGAs}~\cite{Belanger:2006is}, illustrating both the WIMP case, $\Omega_\text{sym}=\Omega_\text{obs}\simeq 0.26$, and the asymmetric DM region that arises for larger couplings.
The annihilation channels $\chi\bar\chi\to \mu\bar\mu$, $\tau\bar\tau$, $\nu\bar\nu$ are $p$-wave suppressed for scalar DM, whereas $\chi\bar\chi\to Z'Z'$ is $s$-wave and thus dominates the annihilation for $m_\chi>m_{Z'}$.
The orange regions indicate that the DM capture rate is maximized, $C\simeq C_*$, so a nearby NS would be kinetically heated to $\unit[1700]{K}$. $C\simeq C_*$ holds away from the resonance region $m_\chi \sim m_{Z'}/2$.
Notice that our asymmetric scalar DM unavoidably has repulsive self-interactions mediated by the $Z'$. As a result, even asymmetric scalar DM will not lead to black hole formation inside the NS~\cite{Kouvaris:2011fi,Bell:2013xk}, unless the DM masses are far above the values considered here.
In the light-mediator case these repulsive self-interactions can also have an impact on structure formation, which has been argued to prefer DM--DM transfer cross sections $\sigma_T/m_\chi \sim\unit[1]{cm^2/g}$ on dwarf-galaxy scales (DM velocities $v_\text{dwarf}\sim\unit[10]{km/s}$)~\cite{Tulin:2017ara}. We show this contour in Fig.~\ref{fig:abundance} using the formulae of Refs.~\cite{Tulin:2012wi,Tulin:2013teo,Cyr-Racine:2015ihg} for $\sigma_T (\chi\chi\to\chi\chi)$. Cross sections of similar size do not seem to be favored on cluster scales, although the issue is currently far from settled. Suppressing $\sigma_T/m_\chi$ on cluster scales ($v_\text{cluster}>\unit[2000]{km/s}$) requires a velocity-dependent cross section~\cite{Kaplinghat:2015aga}, which arises in our model for $m_{Z'}/m_\chi \lesssim v_\text{cluster}/c$, as can be seen in Fig.~\ref{fig:abundance} (left) [see also Ref.~\cite{Kamada:2018zxi}].
The DM mass region $1$--$\unit[10]{GeV}$ in the light $Z'$ case is thus potentially preferred due to small-scale structure formation, while the region above the dashed blue line in Fig.~\ref{fig:abundance} is disfavored by observations such as the bullet cluster~\cite{Robertson:2016xjh}.
For \emph{symmetric} DM there is the possibility of additional NS heating through DM annihilations. Even the $p$-wave channels will reach equilibrium with capture~\cite{Kouvaris:2007ay,Kouvaris:2010vv} and lead to nearby NS temperatures up to $\unit[2480]{K}$~\cite{Baryakhtar:2017dbj}; only low-energy neutrinos ($E_\nu \lesssim \unit[0.1]{GeV}$) are able to escape the NS.
DM annihilations also give rise to indirect-detection signatures;
for the light $Z'$ case of Fig.~\ref{fig:abundance} (left) this is irrelevant due to $p$-wave suppression and difficult-to-observe $s$-wave neutrinos~\cite{Bringmann:2016din}.
For heavier $Z'$, the $s$-wave $\chi\bar\chi\to Z'Z'\to 4\ell$ could lead to indirect-detection signatures, but is currently only relevant for $m_{Z'}< m_\chi \lesssim \unit[100]{GeV}$~\cite{Leane:2018kjk}, i.e.~not for Fig.~\ref{fig:abundance} (right).
So far we have ignored additional interactions that arise from kinetic mixing or the Higgs portal, which will induce DM couplings to non-muonic matter and are highly constrained by direct-detection experiments~\cite{Altmannshofer:2016jzy,Araki:2017wyg,Arcadi:2018tly,Bauer:2018egk}.
For example, the one-loop scattering of DM via $Z'$ on the protons inside a nucleus $N$ takes the form~\cite{Kopp:2009et,Altmannshofer:2016jzy}
\begin{align}
\hspace{-1ex}\sigma_{\chi N} =\frac{Z^2}{A^2}\frac{m_{\text{red},\chi N}^2 }{\pi m_{Z'}^4}\left(g' q_\chi\right)^2\left[e \epsilon+ \frac{ \alpha g'}{3 \pi } \log \left(\frac{m_\tau^2}{m_\mu^2}\right)\right]^2 ,
\label{eq:DM-proton}
\end{align}
keeping both the tree-level contribution from a Lagrangian term $\tfrac{\epsilon }{2}\, Z'_{\mu\nu} A^{\mu\nu}$~\cite{Galison:1983pa,Holdom:1985ag} and the finite one-loop contribution from muon and tauon loops~\cite{Holdom:1985ag,Araki:2017wyg}.
For heavy DM and $\epsilon=0$, XENON1T~\cite{Aprile:2018dbl} seemingly excludes the entire parameter space of Fig.~\ref{fig:abundance} (right) using Eq.~\eqref{eq:DM-proton}, while CRESST-III~\cite{Abdelhameed:2019hmk} excludes the entire $m_\chi \gtrsim \unit[1]{GeV}$ parameter space of Fig.~\ref{fig:abundance} (left). However, there is no reason to ignore the tree-level kinetic mixing angle $\epsilon$ and it is reasonable to expect other new particles in the model to contribute further to $\gamma$--$Z'$ mixing. As a result, the full kinetic mixing should be treated as a free parameter, constrained by direct-detection experiments. The observation of old cold NS can on the other hand set a constraint that is independent of the kinetic mixing angle and thus perfectly complementary to Earth's direct-detection experiments.
This complementarity holds even more true for other muonphilic DM models. As a simple example, we can consider $U(1)_{L_\mu-L_\tau}$ with \emph{Majorana} DM $\chi$, which has a $Z'$ coupling $Z'_\rho\overline{\chi}\gamma^\rho\gamma_5\chi$ that leads to velocity-suppressed scattering cross sections. For the scattering on muons inside a NS this suppression is very mild because the infalling DM picks up a relativistic velocity, leading to a capture rate that is very similar to Fig.~\ref{fig:cap-muons-med}, at least for $m_\chi > m_\mu$. \emph{Direct detection} cross sections on Earth on the other hand will be heavily suppressed by $u_\chi^2 \sim 10^{-6}$ compared to the complex-scalar DM case of Eq.~\eqref{eq:DM-proton}, lowering the necessity to fine-tune the kinetic-mixing angle~$\epsilon$.
A similar disconnect between NS-capture and direct-detection cross sections can be achieved by replacing the $Z'$ mediator by, e.g., a pseudoscalar~\cite{Bell:2019pyc}.
\section{Conclusion}
\label{sec:conclusion}
The muon $g-2$ experiment at Fermilab (E989) is expected to confirm or dispute the longstanding magnetic-moment anomaly by the end of 2019~\cite{Keshavarzi:2019bjn}, while Belle~II will probe the $L_\mu-L_\tau$ solution to said anomaly. Together with LHCb, Belle~II is also expected to scrutinize the $B\to K^{(*)}\mu^+\mu^-$ anomalies that hint at lepton non-universality.
With these tantalizing and soon to be reevaluated hints for new physics in the muon sector it behooves us to consider muonphilic DM, which only has highly suppressed couplings to first-generation particles.
Using the curious fact that NS contain a large population of stable muons we can expect limits on DM--muon interactions from the DM capture on NS.
Indeed, the infalling DM will heat old NS up to $\mathcal{O}(\unit[2000]{K})$ due to elastic scattering and potentially annihilation, which is in reach of future infrared telescopes.
For light mediator masses, relevant for $(g-2)_\mu$, muonphilic DM can also have self-interactions that resolve structure-formation issues.
Overall we have shown that muonphilic DM is not as elusive as naively expected, both in terms of motivation and in terms of signatures.
\section*{Acknowledgements}
We thank Mauro Valli for discussions and comments on the manuscript.
RG is supported by the ULB-ARC grant ``Probing DM with Neutrinos'' and the Excellence of Science grant (EOS) convention~30820817.
JH is supported, in part, by the National Science Foundation under Grant No.~PHY-1620638 and by a Feodor Lynen Research Fellowship of the Alexander von Humboldt Foundation.
\bibliographystyle{utcaps_mod}
|
\section{Introduction}
As the capabilities and pervasiveness of machine learning (ML) and artificial intelligence (AI)
increasingly affect society, there is
increasing concern about the \emph{safety} of such systems, i.e.\ the potential of accidental harm from implementation errors and unintended consequences in ML algorithms. As a result, there has been increasing interest in the nascent field
of \emph{AI safety} \cite{amodei2016concrete,everitt2018agi,leike2018scalable,yudkowsky2004coherent,christiano2018supervising,irving2018ai}, which seeks to understand and solve the technical challenges
in developing and deploying AI that does what its designer intended it to do. The purpose of this chapter is to explore how the study of AI safety intersects with that of evolutionary computation (EC), to both highlight an exciting and important set of safety problems within EC, and to suggest that evolution and EC have important insights that could benefit the general study of AI safety.
To frame the problem of AI safety, we adopt the framework of \citet{amodei2016concrete}, which defines AI safety as concerned with accidents in ML systems, and defines five problems within three broad categories of issues: (1) specifying the wrong objective function, (2) making safe and efficient use of a true but expensive objective (e.g.\ human feedback), and (3) how to improve or adapt
safely while interacting with the real world. A running example in that paper, which we adopt here, describes a robot with the task of cleaning an office using common tools; we modify the example to assume that the controller for this robot has been evolved, i.e.\ with an EC technique like neuroevolution \cite{lehman2013neuroevolution,yao1999evolving} or genetic programming (GP; \cite{banzhaf1998genetic,koza1992genetic}) in the setting of evolutionary robotics (ER; \cite{nolfi2000evolutionary,lewis1992genetic}). While this running example is posed in the reinforcement learning (RL)
setting of ER, similar issues can arise whenever an EC-trained artifact
interacts with the real world; for example, a credit-scoring system trained with GP symbolic regression (e.g.\ as in \citet{ong2005building}) when deployed might enact unintended consequences on the real-world borrowers its decisions affect, e.g.\ by basing decisions on ethically- and legally-problematic borrower traits (e.g.\ race).
One motivation for this chapter is to draw attention within EC to a selection of interesting and important concrete research problems (as introduced by \citet{amodei2016concrete}), in hopes of encouraging progress towards one of EC's aspirations: to provide mature and reliably safe solutions for real-world AI problems. If EC systems are increasingly trained, refined, and applied in the
real-world, it becomes necessary to deal with real-world complications that are often side-stepped in
closed-world research benchmarks; grappling with these issues is thus necessary for EC to transition into a reliable approach for safely solving real-world problems.
For example, if evolution is occurring in an environment
alongside humans (e.g.\ evolving a robot controller that interacts with people in an office setting) much care is needed
to design an appropriate fitness function that at least does not cause harm in its early incarnations; in contrast,
fitness functions in more traditional closed-world ER simulations often undergo many iterations of free-form debugging, with no real danger or cost (beyond wasted time and computation), where initial attempts often create highly-unexpected outcomes \cite{lehman2018surprising}. To enable reliable real-world deployment of EC, it
may be useful to come up with new automated design procedures, to import tools from AI safety in statistical ML, or to perform new and directed EC research on solving technical safety problems.
A complementary motivation is to highlight AI safety problems for
which EC techniques might be particularly well-suited to make significant contributions.
For example, the subfields of
quality diversity (QD; \cite{pugh2016quality,lehman2011evolving}) and open-ended evolution \cite{taylor2016open,standish2003open}
might provide a natural mechanism to create a diverse set of
test-scenarios to illuminate rare but important potential failures modes of ML systems (that might
otherwise go unidentified). For example, the fooling images work of \citet{nguyen2015deep} shows how
EC can automatically identify diverse visual patterns that a deep neural network
will confidently misidentify). Overall, while most current AI safety work is
conducted with traditional statistical ML (e.g.\ gradient-based deep learning approaches),
EC might bring new ideas, perspectives, and techniques to bear on such problems.
A final motivation is to consider if and how natural evolution solved problems
similar to those tackled by AI safety researchers. For example, evolution has designed various means of
collaboration among social animals and between mutalistic species, that in effect minimize negative side-effects to other agents (an important topic in AI safety).
Additionally, evolution has uncovered ways to explore more safely both across an evolutionary timescale (i.e.\ through the evolution of
evolvability \cite{kirschner1998evolvability,wagner1996perspective}, whereby evolution favors improved variation) and an individual organism's lifetime (i.e.\ through the complementary instincts of curiosity and fear \cite{buss2015evolutionary}). The hope is that biological inspiration might point the way towards potential solutions to these kinds of safety problems in EC or in ML at large.
The conclusion is that AI safety is likely to be a growing field of interest in
coming years that offers a range of interesting technical challenges,
and that EC may both have important insights to offer and benefits to gain from research in that community.
\section{Background}
The next sections describe the field of AI safety, and how EC is applied
in the real world, which helps to understand safety concerns from an EC perspective.
\subsection{AI Safety}
The field of \emph{AI safety} \cite{amodei2016concrete,everitt2018agi} seeks to pose and solve technical challenges involved in developing AI that in practice does what its designer intends it to do. The hope is to help foresee and avoid harmful accidents that might result from good-intentioned AI gone astray, for example, through misspecified fitness functions or differences between the training and testing environments.
While the name ``AI safety'' naturally evokes ideas of direct physical safety (e.g.\ how to make sure there are sufficient guard-rails that prevent a robotic arm from accidentally hitting a human), the problems studied in AI safety also encompass more abstract and broad concerns. Such concerns include immediate and short-term ones, like how a mobile robot driven by RL can continually improve its policy by exploring, without taking any catastrophic actions (such as those that cause harm to itself, to the environment, or humans); they also include more speculative concerns about the future (e.g.\ how to make sure
an AI that surpassed human intelligence would still be controllable and aligned with our interests).
One central challenge in AI safety, relevant both to short and long-term concerns, is known as the \emph{value alignment} problem: How to align what a computational agent values with what we value. This problem might appear at first simple, because as designers of agents we have complete control over their incentives. However, such alignment remains an unsolved technical
challenge. Currently we do not know how in practice to algorithmically specify (or learn from data) the complexity of what humans care about, e.g.\ our
moral intuitions, common-sense knowledge, and cultural norms, all of which can potentially come to bear upon what we intend for a computational agent to do.
In other words, EC as of yet lacks a procedure to specify a correct and complete fitness function that encompasses all the background context that could be important
for a system that interacts appropriately with humans and society.
More concretely, even for an AI system that interacts with the real world in very limited ways, it is
still often a challenge to design a fitness function that truly measures or incentivizes correct behavior \cite{lehman2018surprising}.
Indeed, the typical paradigm in AI remains to specify a fixed and
relatively simple objective function (e.g.\ a fitness function in EC) that is then optimized through search; however, as practitioners in EC are well-aware, an intuitive fitness function can often be optimized in unexpected ways \cite{lehman2018surprising}. While there exist candidate approaches to value alignment \cite{irving2018ai,leike2018scalable,yudkowsky2004coherent}, the problem at core currently remains unsolved.
Interestingly, even if incentives are aligned, i.e.\ the learning system is provided
with the correct objective function, how to successfully (and
safely) optimize that objective function is still a difficult and unsolved problem
in its own right. For example, an RL agent that is given the correct objective
to optimize can still make mistakes \emph{while it is being optimized} (e.g.\ it can make harmful mistakes
while exploring how to improve its policy); or, the objective might be challenging to optimize
(e.g.\ it might instantiate a fitness landscape with many local optima), and the
locally-optimal policies found by search in practice might not be value-aligned.
One useful framework for categorizing technical challenges in AI safety
comes from \citet{amodei2016concrete}, which divides safety problems into five categories: avoiding negative side effects,
reward hacking, scalable oversight, safe exploration, and robustness to distributional shift (see table \ref{tab:tcais} for short descriptions of each).
We adopt this framework in this paper for relating AI safety problems to EC and evolution, and later in this paper describe each of these
problems in detail and how they emerge in EC.
\begin{table}[]
\begin{tabular}{|p{4.2cm}|p{7.3cm}|}
\hline
Avoiding Negative Side Effects & Negative side effects result from a fitness function that correctly specifies how to \emph{narrowly} achieve a goal, but does not penalize possible harms to the environment or other agents. \\ \hline
Reward Hacking & Reward hacking is when a fitness function fails to well-specify how to achieve a goal; evolution can therefore maximize fitness in an unexpected and undesirable way. \\ \hline
Scalable Oversight & Scalable oversight requires effectively and efficiently balancing the use of cheap proxy fitness functions (e.g.\ a simple heuristic) with expensive but more accurate fitness evaluations (e.g.\ human assessment). \\ \hline
Safe Exploration & Safe exploration studies how evolution can learn effective behavior while minimizing catastrophic actions taking during learning. \\ \hline
Robustness to Distributional Shift & Robustness to distributional shift requires real-world applications of evolution to safely deal with situations not seen during training. \\ \hline
\end{tabular}
\caption{The table describes five categories of technical challenges in AI safety, as identified by \citet{amodei2016concrete}.}
\label{tab:tcais}
\end{table}
One general consideration for AI safety is that it is most relevant when considering applying AI algorithms to real-world
situations, where human well-being, broadly speaking (e.g.\ including not only physical safety, but also social harm from biased high-stakes decisions \cite{whittaker2018ai} or offense from insensitive classifications \cite{usatoday}), might be at stake. Thus the next section reviews common paradigms for applying EC to the real world.
\subsection{EC and the Real World}
There are many different motivations for studying EC. While one entirely legitimate
such motivation is to understand the creative potential of
algorithms inspired by biological evolution for its own sake, researchers in EC often explicitly
aim towards real-world applications of their ideas, or at least paint a viable path towards
how their ideas might be translated into beneficial real-world impact. Below we describe how such translation
often happens in both supervised and reinforcement learning problems (see figure \ref{fig:ecrw} for a high-level summary).
\begin{figure}
\centering
\includegraphics[height=2.4in]{EC2RW}
\caption{Common paradigms for EC to impact the real world. In (a) supervised learning, a model is evolved by fitting training data, and then is deployed into a larger decision pipeline that involves real data; the decisions resulting from the pipeline (that are influenced by the trained model) translate into real world impact. In (b) evolutionary robotics or RL, an agent is often first trained in simulation, and then transferred across the reality gap, where it can potentially be further evolved. Alternatively, in embodied evolution, an agent is trained from the onset in the real world.}
\label{fig:ecrw}
\end{figure}
\subsubsection{Supervised Learning}
When EC is applied to supervised learning, i.e.\ training where the task is
to predict or classify over a labeled training set, it is important to recall that
supervised training performance
is rarely an end in itself. While improved accuracy on a benchmark is often critical
for publishing a paper about symbolic regression \cite{koza1994genetic} or neural classification \cite{rocha2007evolution}, such accuracy is only practically important insofar as it feeds into the downstream task the model is \emph{applied to}. For example, a classification model of credit-worthiness
might be applied to decide whether a loan should be granted or not. While improved accuracy will likely
contribute to such use cases, it will not take into account the nuances of the domain, like the differing impact of different kinds of mistakes \cite{usatoday}.
Thus, while applications of supervised learning might
at first not seem relevant to AI safety, the objective that a supervised-learning EC model is
trained towards (e.g.\ classification accuracy) nearly always serves only as a \emph{proxy} for the ``true'' downstream objective in the real world
(e.g.\ efficient loan allocations that abide by legal and moral norms). Notably, limitations
of such proxies are well-known; for example, the fairness, accountability, and transparency
community within ML has highlighted how maximizing training accuracy can result in models that
base decisions on societally-unacceptable criteria \cite{zafar2017fairness,whittaker2018ai}. This pervasive type of gap between the proxy and the true objective can
be seen as a manifestation of the general value alignment problem, and techniques for minimizing such a gap
highlight how AI safety research can be relevant to EC-based supervised learning.
\subsubsection{Reinforcement Learning}
When EC is instead applied to robotics or RL, evolution most often first occurs within
simulated environments. The idea is that policies trained in simulation can subsequently be transferred to
reality \cite{jakobi1995noise,koos2013transferability}, and potentially further evolved in the real world. The reasons for training
in simulation include that real-world evaluations can be slow, tedious, and expensive, often
risking damage to hardware (like a robot) and to the broader environment (like humans coexisting with the robot). Simulation enables more convenient large-scale
experimentation (given sufficient computation), although both how to design accurate simulations for complicated domains and how to successfully transfer policies from simulation
to the real world remain challenging areas of research \cite{pollack2000evolutionary,koos2013transferability,jakobi1995noise}.
Safety concerns in this paradigm can emerge from simulations that do not reveal safety-critical edge-cases later encountered when models are deployed in the real-world, or from changing circumstances in reality (i.e.\ distributional shift) that are not captured for in simulation.
Another paradigm in EC is embodied evolution \cite{watson2002embodied}, wherein evolution is conducted in the real world, to circumvent the challenges of
building accurate simulators and crossing the reality gap. In this setting, to the extent that evolved policies interact with humans or can damage their
robotic body or their environment, there may be the need for potentially expensive supervision (an AI safety issue discussed in more detail later). In
general, because there is never the protective buffer of simulation between a policy and the real-world, safety considerations in embodied
evolution may be more challenging than in other settings.
The conclusion is that as EC strives and achieves greater real-world impact, there will likely be a corresponding increased risk (albeit still potentially minor in many domains) of unintentional harm,
independent of the specific paradigm by which EC models are trained and deployed.
\section{EC and Concrete AI Safety Problems}
This section explores more concretely how ideas from EC intersect with those from AI safety. We adopt the
framework of \citet{amodei2016concrete}, which identifies five classes of concrete problems that can cause AI accidents: avoiding negative side effects,
reward hacking, scalable oversight, safe exploration, and robustness to distributional shift (described from a high level in table \ref{tab:tcais}).
For each of these five problems
we introduce the problem, describe how it can arise in EC, how it relates to various research areas in EC, and suggest directions for potential solutions to such problems.
Note that our main aim here is to frame AI safety for EC researchers and practitioners, and as a result, we will not comprehensively survey the broader study of AI safety within ML; for more comprehensive surveys, see \citet{amodei2016concrete} or \citet{everitt2018agi}.
\subsection{Avoiding Negative Side Effects}
The problem of negative side effects is that a well-specified fitness function must not only reward achieving a desired goal narrowly, but should also penalize possible negative consequences on the broader environment. That is, a fitness function is often under-specified in practice, even if the conditions of achieving the desired goal are correctly described. The reason is that there are many ways to short-sightedly accomplish a goal that humans would nonetheless find unacceptable. For example, borrowing from \citet{amodei2016concrete}, a robot might knock over an expensive vase en route to its destination; even if the robot arrives successfully at its destination (its goal), the damage to the vase is an unacceptable negative byproduct of the robot pursuing its goal.
If the fitness function does not penalize for breaking the vase, the resulting negative side effects could be viewed as a failure of the researcher to express the correct fitness function. However, while one could
attempt to anticipate and hard code into the fitness function every negative contingency, such exhaustive anticipation is often unrealistic, and at best tedious. Ideally, there would be a way to
automatically (or with minimal supervision) augment a goal-directed fitness function to penalize such undesired impacts.
The challenge in designing such an automated method relates to the value alignment
problem in AI safety, in that there is much background context (e.g.\ about what objects in the environment are fragile or important) that a human brings to their understanding of what an acceptable solution is; such context is difficult to effectively and exhaustively translate into a fitness function (although some projects do aim to distill such background knowledge \citep{lenat1995cyc}).
Interestingly, most EC and ER environments are constructed such that there is little potential for
negative side effects; the reason is that richer environments are more challenging to model, more computationally demanding to simulate, and
such complications are often orthogonal to the
research questions under study. In practice, simulated environments in ER are nearly always
closed-world and spartan, containing only elements directly relevant to the task at hand.
For example, a common variety of ER task involves simulated wheeled robots navigating through
an enclosed environment containing only walls and artifacts directly related to the task (e.g.\
a light switch that can be triggered, or tokens that can be collected). Negative side-effects
are often impossible by definition: The robot can not damage itself or anything of importance in
its environment.
In ER and EC experiments that involve the real world, or interacting with humans, there is
more potential for negative side-effects, although experimenters nearly always apriori
minimize that possibility by design. For example, when transferring policies evolved in simulation
to the real world, the real world environment is often engineered to mimic the spartan simulated one,
and often such transfers are one-off experiments (i.e.\ the robot will not then be operating
in an ongoing way) under intensive supervision.
However, despite the minimization by design of negative side effects,
the conclusion is that as (or if) EC and ER progresses, we likely will want or need
evolved agents to be deployed in
complex open-world or human-coinhabited environments; in such situations, the problem of
negative side effects can no longer be avoided. Thus,
when aiming toward the real world, simulated environments may need to
be augmented to include the potential for negative side-effects (and for learning
to avoid them), or automated techniques for mitigating side-effects from
real-world deployment may need to be developed.
So far, the problem of negative side-effects appears to be an
under-studied aspect of how to scale EC, one that may provide exciting future research
directions. One possible paradigm for minimizing negative-side effects is to train
EC agents through interactive evolutionary computation (IEC; \cite{takagi2001interactive}), i.e.\ to
involve humans directly in the breeding process. Due to the problem of user fatigue in IEC \cite{takagi2001interactive}, i.e.\ that
the task of breeding can become monotonous and exhausting,
it is difficult to scale IEC, which necessitates learning surrogate models \cite{jin2011surrogate} or applying distributed IEC \cite{secretan2008picbreeder}, i.e.\ systems that involve many humans breeding in potentially uncoordinated ways.
Overall, the interaction of IEC with embodied evolution or
ER in general (as in \citet{woolley2014novel}) could benefit from greater study
from a safety perspective. Current research directions in ML that address negative side effects
include penalizing for changes to the environment \cite{armstrong2017low}, or
algorithms that \emph{satisfice} instead of optimize unboundingly \cite{taylor2016quantilizers} (motivated by the idea
that side-effects may often result from extreme optimization). Both such approaches could potentially
be adapted for EC.
\subsection{Reward Hacking}
The problem of reward hacking, like that of negative side-effects, is caused by an
incompletely- or incorrectly-specified fitness function. While negative side-effects
are collateral damage incurred while successfully achieving the desired objective, reward hacking is
when optimization uncovers unexpected ways to maximize the fitness function \emph{without}
achieving the desired objective. For example, if the true objective
of a cleaning robot is to clean the office, but its fitness function rewards
for each individual mess the robot cleans, the robot may discover that it maximizes fitness by
creating new messes that it can subsequently clean \cite{amodei2016concrete}.
The phenomenon of reward hacking is familiar to most EC practitioners; nearly all of us
have encountered situations where an intuitive fitness function is maximized by
counter-intuitive (and undesirable) behavior. Indeed, that so many illuminating (and funny) anecdotes of reward
hacking existed in the EC community was one main inspiration behind the crowd-sourced documentation effort of \citet{lehman2018surprising}, which describes many reward-hacking examples. A representative example is found in
Karl Sims' seminal virtual creatures work \cite{sims1994evolving}. In early attempts to evolve locomotion gaits by rewarding forward motion, the result was not locomotion, but morphological evolution towards tall rigid bodies that could exploit their potential energy by falling or somersaulting forward.
Beyond EC, the challenge of constructing incentives for agents (like fitness functions) that cannot be
undermined is well known in other fields. For example, in economics, Goodhart's law \cite{goodhart1984problems} states
that ``when a measure becomes a target, it ceases to be a good measure.'' Similar understanding goes by the name of the principle-agent problem
in economics and political science \cite{ross1973economic}, and similar challenges exist in designing contracts in law \cite{hadfield2018incomplete}. Further, there
are many historical examples of perverse incentives, where an incentive to solve one problem instead exacerbates it; for example, a French colonial program in Hanoi paid citizens for turning in rat tails, in hopes of exterminating rats, but it instead led to \emph{farming} rats \cite{vann2003rats}.
This consilience of evidence suggests that designing incentives is generally difficult, and that
humans are habitually overconfident about their ability to skillfully do so, often failing to
anticipate subtle loopholes instantiated by intuitive reward structures.
In this way, reward hacking in EC and ML is one manifestation of a broader problem.
In practice, reward hacking in EC is often solved through iteration. First, an intuitive fitness
function leads to surprising and undesirable outcomes, that are understandable only in hindsight. The
experimenter then attempts to modify the fitness function to patch the problem, which
potentially may lead to a different kind of exploit that must also be patched.
Interestingly, because these failed incentives can be viewed as failures of the
experimenter, and happen within the loop of scientific experimentation that precedes
a polished experimental setup,
they are often not reported scientifically \cite{lehman2018surprising};
as a result, the prevalence and
importance of reward hacking in EC may be under-appreciated and understudied.
While frustrating, when evolution occurs in simulation
such reward hacking may not cause harm much beyond wasted experimenter
effort and time. However, the ability for EC practitioners to quickly and
safely explore new tasks, especially in settings such as embodied evolution or
reality-gap crossing, is undercut by the expertise and trial-and-error needed
to construct reliable fitness functions.
As in negative side-effects, IEC is one avenue for helping to overcome
reward hacking, by involving human judgment to assess quality during
evolution
rather than by crafting fixed heuristics. Beyond directed human breeding, humans may also supply other (potentially richer) forms of guidance to further constrain or replace traditional
fitness functions, like demonstrations of acceptable behavior or heuristic advice, as in \citet{karpov2011human}.
Such EC research directions can be seen as connected to similar potential solutions
in traditional ML, such as imitation learning \cite{argall2009survey},
wherein an agent learns how to imitate expert demonstrations of behavior; cooperative inverse reinforcement learning \cite{hadfield2018incomplete}, where a reinforcement learning
agent cooperates with a human to discover and optimize the human's preferences; or reward modeling \cite{leike2018scalable}, wherein a machine learning model is trained to
predict human preferences (similar to surrogate models used in IEC \cite{jin2011surrogate}.
Exploring if and how such ML methods could apply to EC (e.g.\ evolutionary imitation learning,
or applying deep learning models to learn models of human preferences to drive evolution) may
be a productive area of future research.
\subsection{Scalable Oversight}
The problem of scalable oversight is that in EC and learning systems in general, there
is often expensive-to-gather information that accurately reflects how acceptable a
solution is, but such guidance is too expensive to be applied as the primary
driver of search. For example, a very accurate measure of fitness for
a cleaning robot might require expensive manual testing of how dirty
a carpet is before and after the robot is deployed within a room. Other
proxy measures may be more cheaply available, such as a human giving a quick glance to a room, or
by the robot measuring how much dirt it is picking up. However, such proxies
might exacerbate problems such as negative side effects or reward hacking \cite{amodei2016concrete}. For example, a robot maximizing dirt picked up might knock over a plant to gain access to more dirt, or a robot maximizing human approval after a quick glance might hide messes under a rug.
The issue is how to efficiently and effectively apply combinations of cheap
proxy signals with occasional expensive feedback, to produce a practical (and well-behaved)
learning system.
One way the issue of scalable oversight emerges in EC is through
the practical construction of real-world fitness functions (e.g.\ fitness functions for
fine-tuning policies in reality that were first learned in simulation, or fitness functions applied in embodied evolution). In other
words, when applying evolution in a real-world situation, what sensors are available on a robot, what a human can easily evaluate, or how the environment can be augmented with automated sensors to evaluate aspects of behavior (e.g.\ motion capture equipment or ceiling-mounted cameras) will affect what
fitness functions are possible to automate, and the overall cost-effectiveness of executing different experiments.
However, scalable oversight, like other AI safety issues, is often eliminated by
design from simulated EC domains. Experiments in which cheap proxy fitness evaluations are not possible
or in which they fail (due to reward hacking or negative side-effects) are unlikely to be pursued or published. However, if progress could be made on
enabling more scalable oversight, it might extend the range of what kinds of embodied evolution
or real-world fine-tuning could be performed. In this way, scalable oversight is an interesting avenue of research not
only for safety reasons, but because it may help expand the complexity of domains for which real-world EC can
be applied.
The area of EC research most similar to scalable oversight is that of surrogate-assisted EC \cite{jin2011surrogate}, wherein expensive-to-calculate fitness functions are approximated
with a learned model; of particular interest (for their potential efficiency)
are surrogate models that intelligently
choose which points in the search space to subject to expensive ground-truth fitness queries.
For example, \citet{gaier2018data} applies Bayesian optimization to enable a data-efficient quality diversity \cite{pugh2016quality} algorithm. Another related area of EC study are methods that estimate which
genomes are likely to successfully cross the reality gap \cite{koos2013transferability}; the
reason is that there is an analogy between simulations (and their relation to reality) and
proxy fitness measures (and their relation to ground-truth fitness).
From a ML perspective, \citet{amodei2016concrete} propose that a semi-supervised formulation of reinforcement learning may be a productive paradigm for tackling scalable oversight. The idea is
that an agent only receives reward information on a small subset of its experience (as opposed to
the traditional RL setting where reward is observed for each action taken in the environment). In
particular, the agent must learn \emph{when} to request expensive reward information, and is
incentivized to learn cheap proxy measures that correlate with the expensive reward. Because EC
uses fitness functions that operate over an individual's entire evaluation, rather
than the per-timestep rewards of traditional RL, it may not be easy to translate
such a paradigm to EC (although it could be an interesting direction for research).
One potential way of framing semi-supervised RL for
evolutionary RL is to learn a semi-supervised reward predictor (with ML) that could
assign fitness to individuals by observing their sensory-motor stream.
\subsection{Safe Exploration}
The problem of safe exploration is how evolution (or individuals capable of life-time learning) can explore new solutions without ever (or only very rarely) taking catastrophic actions, i.e.\ ones that harm valuable aspects of the environment, including humans or expensive equipment such as robots.
Note that safe exploration remains a problem even if objectives are correctly specified: Even if a fitness
function correctly identifies all unacceptable negative side-effects, and a \emph{properly-trained} agent would thus avoid such effects, \emph{during learning} an agent might still undertake catastrophic actions. For example,
the cleaning robot may suffer a fitness penalty for breaking a vase, but it still needs to experience that penalty during training to learn to avoid breaking it. A related problem is that given a robotic controller that behaves safely, there is no guarantee that an arbitrary mutation of it will also
be safe. The danger of exploration is a deep philosophical problem, in that the very act of exploration seems inherently to be about stepping into the unknown. However, humans can often
successfully explore new possibilities and emerge relatively unscathed (sometimes using mental models to predict whether a new strategy would be catastrophic before trying it, somewhat similarly to model-based RL \citep{sutton1990integrated}), suggesting that practical solutions may be possible.
There are two main ways that real-world
accidents from safe exploration can emerge in EC. First, take the case of learning a plastic policy (e.g.\ a policy that learns from experience \emph{during its lifetime} \cite{soltoggio2008evolutionary,soltoggio2018born}). For example, a robot might be trained to explore any environment it is embedded within, in search of a particular goal. In effect, such an agent must
learn \emph{how to explore}, and if the deployment plan involves the real world (through embodied evolution, or crossing the
reality gap), then there are risks from unsafe exploration. For example, in a new environment, a learned exploratory strategy might
lead the robot to damage itself. Second, there is the case where a learned (non-plastic) policy is either trained
in the real world (embodied evolution), or is fine-tuned in the real world after being trained in simulation. In this
case, exploring the space of policies (through mutations of existing policies) may result in unsafe policies. For example,
in some robotics domains solutions are known to be fragile, i.e.\ that most mutations result in
degenerate (possibly damaging) behavior \cite{lehman2011improving,lehman2018more}. For concreteness, a robot trained to walk successfully
in simulation may lose some performance when transferred across the reality gap, and there is no guarantee that
perturbations of the transferred policy (explored in hopes they will improve the walking policy) will not cause the robot to fall and harm itself.
Overall, it may be impossible to solve the issue of safe exploration without involving
some form of human oversight. The reason is that learning what is unsafe
seemingly requires either: (1) an accurate model of the world that includes robust identification of
catastrophes, (2) labelled data of all possible causes of unsafe scenarios in a domain, or (3) active experience
in the domain with feedback from an overseer that prevents unsafe actions from being taken.
All three require either extensive domain knowledge, e.g.\ (1) or (2), or direct human intervention (3).
In this way, the problem
of safe exploration may be intrinsically tied (like some of the other problems) to that of scalable oversight: Given
that potentially expensive human feedback is needed, how can it be gathered and exploited in an efficient way to
enable reliable real-world exploration?
Interestingly,
like other problems mentioned here, often the issue of safe exploration in EC currently arises \emph{outside} the formal
scientific process: Domains are constructed that intrinsically minimize risk (e.g.\ through spartan closed-world design), and guard-rails to minimize damage to real-world robots and their environment are
engineered on a robot-by-robot or domain-by-domain basis by experimenters; failure modes (e.g. robot damage) encountered in such experiments are unlikely to be deemed of enough scientific import to be published. Thus, one contribution to studying safe
exploration in EC would be to introduce a variant of common ER benchmarks that simulate the idea of safe
embodied evolution; for example, a maze navigation task could include deep holes that would endanger a robot,
or fragile and valuable aspects of the environment.
Another possible avenue of research for contributing to safe evolutionary exploration is to
improve the robustness and evolvability of genomes. For example, some EC methods find parts of the search space that are more robust to mutation \cite{lehman2018more}, or adapt variation operators to increase robustness or evolvability \cite{lehman2011improving,wierstra2008natural}, or attempt to enforce small changes to an evolved policy \cite{lehman2018safe}. While not initially motivated by safe exploration, it may be possible to adapt such techniques towards that end. The idea is that with well-tuned variation, parent policies that are safe may be more likely to produce safe children policies, under the assumption that larger policy changes are more likely to be degenerate.
EC could also attempt to solve existing safe exploration
benchmarks from the RL community, e.g.\ the safe exploration grid-world of \citet{leike2017ai} or
domains explored by \citet{moldovan2012safe}.
Potential safe exploration techniques could also be imported or adapted from studies of safe exploration in RL \cite{garcia2015comprehensive}. Promising such techniques include the
approach of \citet{saunders2018trial}, wherein human oversight is used to train a supervised learning model
that blocks unsafe actions, or \citet{lipton2016combating}, wherein catastrophic actions are
explicitly stored and rehearsed to endow a RL agent with an intrinsic sense
of fear. Similar models could be trained to block unsafe actions for ER or
in embodied evolution.
\subsection{Robustness to Distributional Drift}
The problem of robustness to distributional shift is that when AI systems are deployed, they may encounter situations that deviate from
the ones it was trained upon. In such situations, a naively trained agent may demonstrate
arbitrarily inappropriate behavior, because extrapolating to novel circumstances is challenging.
Accidents can thus result in this paradigm if an agent's
policy results in ill-suited actions when encountering new situations.
In some EC communities, such as ER, experiments may not always explore how well a learned behavior
generalizes to situations other than the exact ones experienced in training; i.e.\
in the language of statistical ML, the training set doubles as the testing set.
As a result, there may be little understanding of how a policy would generalize, and
how pathological a robot's behavior would be if it encountered a novel situation.
Note that interestingly, the issue of poor generalization
is a topic of recent interest in deep RL
as well \cite{zhang2018study,justesen2018procedural,cobbe2018quantifying}.
While this paradigm may not be intrinsically problematic, i.e.\ if the research
question does not involve generalization or real-world deployment,
graceful degradation of out-of-training-distribution performance becomes critical as policies are
deployed in the real-world (especially open-world scenarios where it is well-understood that all possible situations cannot be anticipated, and that circumstances will likely shift over time).
Several EC communities study partial solutions to
this problem. For example, one subfield of EC studies
dynamic fitness landscapes \cite{branke2003designing,richter2009detecting},
wherein evolution continues as circumstances change, which could continually align the policy to the current
distribution of scenarios. Further, such fluid adaptation may favor (or be enabled by mechanisms that encourage) more \emph{evolvable} representations, i.e.\ representations offering diverse and adaptive variation, another important and related field of EC study \cite{wagner1996perspective,kashtan2007varying}. Complementarily, others in EC study meta-learning \cite{soltoggio2008evolutionary}, i.e.\ evolutionary approaches to learning \emph{how to learn}, which may enable a policy to quickly learn online from its own mistakes.
While these research communities provide
important insights for tackling distributional shift, new benchmark tasks may
be needed to ground out the risks from real-world distributional shift and to determine which (or which
combinations) of these techniques would help ameliorate such risks in practice.
For example, an ER domain could be introduced in which environments are produced
through procedural content generation (PCG; \cite{shaker2016procedural}), but where the
distribution of PCG parameters changes over evolutionary time; different approaches
could be compared by how many catastrophic failures are encountered across
evolutionary time.
Solutions could also take inspiration from the study of distributional shift
within ML. For example, the insight in Inverse Reward
Design \cite{hadfield2017inverse} is that the
fitness function
encountered during training should only be trusted insofar as
it reflects situations that occur during training (i.e.\ the human
designer of the fitness function designed it explicitly to solve
such training situations). An agent
should thus have uncertainty over what such a fitness function implies for
for situations that never appear in training environments. It may be possible to export such an
insight to an evolutionary context, perhaps by querying a human for guidance or forcing
a known safe policy to take over when anomalous circumstances
are encountered (e.g.\ as studied by the fields of novelty/anomoly detection \cite{markou2003novelty,chandola2009anomaly} or
uncertainty-aware RL \cite{kahn2017uncertainty,eysenbach2017leave}).
\section{Discussion}
One interesting question is if EC has unique contributions to make to the
general study of AI safety. A potential benefit of evolution relative to
traditional ML is its divergent creative potential -- evolution seems
well-suited to discovering a great diversity of well-adapted artifacts. Subfields of EC that
study artificial life \cite{langton1997artificial}, open-endedness \cite{standish2003open}, and quality diversity \cite{pugh2016quality} focus on
this facet of evolution, which may be of use for helping in particular
with the problem of robustness to distributional shift. That is, evolution
could be driven to discover a wide range of new training situations to
discover latent flaws in learned policies or models, to augment a limited
training set that might not cover the diversity of situations that could later
be encountered. For example, the
work of \citet{nguyen2015deep} applies a QD algorithm to find, in a
single evolutionary run, a set of diverse images that reliably fool
a deep neural network vision model; following work has shown that
these kinds of adversarial images can provide safety hazards for
real-world use cases of such vision models \cite{eykholt2017robust,kurakin2016adversarial}.
Similar QD approaches might also be used to evolve scenarios to stress-test robotic policies.
Work in this spirit includes \citet{goldsby2010automatically}, wherein
novelty search and GP are used to probe latent behavior of a robotic navigation
system and an automobile door locking control system. Similarly, the environments evolved by
open-ended systems like POET \cite{wang2019paired} could be adapted as a testing suite
for fixed policies.
A related question is to consider what lessons biological
evolution has for AI safety. Many problems faced by AI safety have been
solved, at least in some abstract sense, by biology. For example, the
problem of negative side effects in AI safety is related to the evolution of cooperation and sociality in biology, in
that cooperation often entails considering other agents and their goals in addition
to one's own goal (whether through behavioral convention, as in bees, or deliberative thought, as in humans). From this perspective,
the negative side effects of a robot pursuing its own limited agenda
result from not understanding or taking into account the broader preferences of outside agents
(e.g.\ that a vase is a valuable artifact and should not be broken while cleaning a room). Humans have evolved
moral instincts, the ability to empathize with others, and verbal and written language, all of which enables us to
understand the gestalt of a task another human might ask us to perform, thereby helping us avoid reward hacking and
negative side effects. Similarly, the robustness of our
genetic architecture to random mutations and the natural instincts of
curiosity and fear are nature's hard-won solution to the problem of safe exploration on a genetic and individual level, respectively.
In the same way that evolution (and EC)
have a privileged position in the study and understanding of human-level AI (because evolution is the only algorithm to so-far produce human-level intelligence), evolution and EC may also have a privileged position in understanding the AI safety challenges that
biology has in some sense solved.
An important question for future study
is if methods in EC manifest different kinds of AI safety concerns than those
considered within traditional ML, e.g.\ due to their lack of formal gradient-following or because some EAs produce AI as the result of a divergent creative process (as opposed to optimizing an explicit
objective function as common in most ML).
Because this question is yet unanswered, it is unclear whether the long-term safety agendas currently popular in ML \cite{leike2018scalable,christiano2018supervising,irving2018ai}
are applicable to AI produced by paradigms such as evolutionary artificial life or open-ended evolution, which in their grandest aspirations (just as in traditional
ML or AI) include producing agents with human-level intelligence \cite{stanley2017open}. If current safety agendas do not apply to ambitious forms of EC, then formulating new agendas that targeting them may be a valuable pursuit.
A final discussion topic is to draw together some of the recurring themes from considering each AI safety problem separately, in hopes of
highlighting promising research questions and paradigms. One theme is the
potential need for modifications of EC benchmarks to include safety considerations or the adoption of existing AI safety benchmarks within EC.
Benchmarks, for better or worse, help draw researcher attention,
and can render seemingly nebulous problems more concrete. Because existing EC domains and benchmarks minimize safety concerns by design (because
researchers most often are pursuing research questions orthogonal to safety), new benchmarks may help to catalyze safety research, especially
if they are variants of domains familiar to EC researchers.
For example, EC techniques could be applied to the AI safety grid-worlds of \citet{leike2017ai}. Alternatively, existing ER domains (such as maze
navigation or ball-gathering) could be augmented with catastrophic actions (for investigating safe exploration), or could include held-out
test environments that could test for robustness to distributional shift.
Another overarching theme is the potential for some form of IEC to help in the
solution to nearly all of the reviewed problems; this is not surprising, because many AI safety problems emerge precisely because
human insight is relegated to constructing a fixed setup (i.e.\ in EC the genetic encoding and the fitness function), and IEC is a
framework for allowing human choice to intervene during evolution. Safety considerations may drive more efficient ways to perform
IEC (through improved surrogate models), as well as the construction of new forms of IEC. For example, IEC most often helps steer
what individuals reproduce, but IEC solutions to problems such as safe exploration may require humans to interact more directly
with policies \emph{as they execute}, i.e.\ to intervene to prevent unsafe actions. One source of inspiration may be systems
such as the neuroevolution-based game NERO \cite{stanley2005real}, in which a human experimenter can interact in real time to dynamically change the
environment, parameters of the fitness function, and even embody a virtual agent to probe learned agent behaviors.
\section{Conclusion}
AI safety is an important research topic for enabling EC to reach one of its aspirations, which is to maximize its beneficial real-world impact. At first glance,
such research might seem uninteresting, because it can evoke sentiments of domain-specific engineering, rather than the pursuit of grand scientific questions; however,
AI safety enfolds interesting and philosophically deep unsolved technical challenges, including how to avoid catastrophe while learning about the world, and how to create
fitness functions that incentivize agents that abide by the spirit rather than the letter of the law. As ML and AI grow in import, we can expect funding and interest in AI safety to similarly grow,
and the hope of this paper is to advocate for EC researchers to both contribute and take note of advances in this developing field.
\begin{comment}
One broad aim of evolutionary computation (EC; \cite{TODO}) is to design algorithms with
creative and engineering potential similar to that of biological evolution \cite{TODO}. While outside
perception of EC often assumes the field is the study of biologically-inspired black-box optimization (BBO; i.e.\ gradient-free algorithms for optimizing a fixed objective function), EC also includes other rich paradigms, like open-ended evolution \cite{TODO}, digital evolution \cite{TODO}, and EC-based artificial life \cite{TODO}, each of which seeks understanding-through-engineering of how the volition-free
process of evolution created profoundly complex artifacts, e.g.\ human beings. In this way,
similarly to other fields of artificial intelligence (AI), if the grandest aims of EC were met, one potential product is human-level AI (HLAI).
Arguably, EC has a privileged position in the study of human-level artificial intelligence (HLAI)
because EC aims towards algorithmic understanding of the process by which human-level intelligence in fact came into
existence (i.e.\ biological evolution). That is, unlike other approaches to HLAI, EC takes inspiration from a natural search process that in fact produced HLAI, and EC's potential for success is not predicated on our ability to directly reverse engineer the keystones of intelligence. This argument rests on the assumption that the keystones of intelligence are more complicated and more difficult to discover or understand than those of evolution, which we discuss it in more detail in section \ref{TODO}. Even granting this claim, the privileged position does not imply that EC is the singularly most promising approach to generate HLAI, only that it is supported by a form of empirical evidence that other approaches lack.
While it is unclear if EC will ever meet the ambitious aim of HLAI,
it is clear that the creation of HLAI itself would have profound impacts on science, society, and humanity's
self-understanding. Indeed, as a result of significant progress in the field of statistical machine learning (ML; \cite{TODO}), such as superhuman performance in games such as Atari and Go \cite{TODO}, there is increased concern over societal issues such as whether HLAI is possible or imminent,
and what benefits or harms AI or HLAI will entail for humanity \cite{TODO}.
Such big picture concerns, as well as important short-term concerns about unintended consequences of ML \cite{TODO}, have
driven interest in the nascent field of AI safety \cite{TODO}, which researches how to ensure that artificial agents act as intended, in human-beneficial ways.
This paper also explores the ways in which EC and ideas from EC could benefit AI safety, and particular drawbacks or benefits of EC from an AI safety perspective.
\section{Background}
\subsection{EC beyond Black Box Optimization}
\subsection{AI Safety}
\section{AI Safety for EC}
This paper argues that just as EC has a similarly privileged position in the generation of HLAI, it too is privileged within
the study of AI safety \cite{TODO} (and in particular, to the value alignment problem). There are two main arguments: (1) to the extent that AI safety is concerned with HLAI, and EC is privileged with respect to generating HLAI, then EC's prospects for AI safety deserve specific attention; and (2) if natural evolution is what produced agents capable of understanding and respecting human values (to the extent we believe that humans understand or respect their own values), then from that empirical grounding we might then believe EC is a particularly promising approach to generating agents informed by, capable of learning, or respecting, human values. The first claim at some level is relatively incontroversial: It seems sensible at least to admit uncertainty of if and how HLAI will come about, and to expend some effort to understand safety considerations specific to different ways HLAI could come about (especially to the school of AI most similar to the process that itself generated HLAI). The second claim, however, is less obvious, and will be explored in more detail later in this paper.
This section explores AI safety in the specific context of EC, largely focusing on EC as applied beyond
BBO, as
\section{AI Safety through EC}
\section{Conclusion}
\end{comment}
\bibliographystyle{unsrtnat}
|
\section{Introduction}
\label{section:Intro}
The study and explanation of quantum decoherence is necessary from both a philosophical and a practical point of view. On one hand, because the world we see in our everyday life follows classical rules rather than quantum ones and the emergence of one from the other is a fundamental issue concerning our conception of physical reality. And on the other hand, because plenty of modern research, like the recent advances on quantum computation, are completely based in the quantum framework and holding a map that describes in detail the frontiers of the quantum realm is not optional.\\
Most extended treatments of quantum decoherence are based on the presence of an external agent, the `bath', that absorbs information from the system \cite{Zurek2003}. This is, quantum entanglement with external degrees of freedom limits the quantum behaviour of the system under study, forcing it to behave in a way that resembles the classical rules. This conception is correct but, if na\"ively understood, it may look as if classical behaviour of a system would be impossible to reach without the external assistance. Such interpretation, rather than answering the questions about the origin of decoherence, passes it further to the bath without solving it \cite{Kastner2014}.\\
This paradox is easily solved if we abstract the definition of the bath. This role can be played by any degrees of freedom, either external to the system or internal, that are not accessible to the observer. In this sense we can distinguish between the more traditional approach - the Environmentally-Induced Decoherence (EID) - caused by external agents, and the less apparent one, caused even when the system is isolated by degrees of freedom within it that cannot be traced - the Self-Induced Decoherence (SID) \cite{Castagnino2005}. The second definition can be very useful in the context of many-body theory because if our model of the system is, as usual, limited to its one- or two-body description while the particle number is much larger, then even when the system is isolated we can expect decoherence due to the coupling of our coarse-grained picture of the system to correlations of high order that we cannot measure.\\
Therefore, the statement claiming that an isolated quantum system cannot present decoherence would not be fully correct. Reflecting on this affirmation we can stablish an analogy to another very similar one extracted from the context of Classical Statistical Mechanics: the contradiction between the reversibility of microscopic dynamics (and its quasi-cyclical evolution, according to Poincar\'es Theorem) against the inevitable irreversibility established by the Second Law of Thermodynamics. Obviously, this contradiction is only superficial but reconciling microscopical reversibility/cyclicality with thermodynamic irreversibility requires an effort reconsidering the differences between our macroscopic picture of the system, ruled by the Second Law, and the microscopic model, that is time-reversible. Given the resemblance between both questions we will use the same conceptual tools used in the context of pure-state Quantum Statistical Mechanics \cite{Gogolin2016}, closely related to ergodicity and typicality \cite{Goldstein2010}.\\
This work is structured as follows. First, in Section \ref{section:SID}, we will present the concept of Self-Induced Decoherence, first explaining that it refers to the coarse-grained description of a many body system in \ref{subsection:CGRepresentations}, then we will present quantum purity as our quantitative guideline to determine if a system has undergone a decoherence process in \ref{subsection:PurityAsQuantifier} and we will add a short remark on temporal typicality as a core idea in our theory, inherited from Statistical Mechanics, in \ref{subsection:TimeAveraging}. Next, in Section \ref{section:Analysis}, we will not start by analysing the causes of decoherence - but the causes of coherence. In other words, we will shortly reflect on why non-interacting systems keep quantum purity constant. From there, in \ref{subsection:EDandES}, we will compare the roles of the two characteristics of interacting systems that deprives them from that feature of ideal systems: the Eigenstate Deformations and the Energy Shifts. Finally, we will further explore the role of the Energy Shifts in \ref{subsection:PurityEvolution} to estimate how much does the amplitude of purity fluctuations decrease as we increase the size of the system. We will end summarizing our conclusions in \ref{Conclusions}.\\
\section{Self-Induced Decoherence in the many-body context }\label{section:SID}
Literature related to quantum decoherence \cite{Zurek2003} departs almost invariably from the following assumption - that the system considered in some sense in contact with external agents. This can be either because of some interaction similar to friction \cite{Caldeira1983}, or because of the absorption and emission of electromagnetic radiation \cite{Pfau1994} or even because we are literally studying an open system that can exchange particles with another subsystem with known properties. Whatever the nature of such external agents is they are the ultimate responsible of the loss of quantum coherence that the system may undergo. While this point of view is legitimate and explains the vast majority of real situations, where the total isolation of the system is not possible, we may keep the feeling, also legitimate, that the question about what is the origin of decoherence has not been answered, but displaced instead: we have not explained how quantum systems stop behaving coherently, we have only discharged any responsibility upon `external agents', eluding to provide an actual answer \cite{Kastner2014}.\\
This dilemma has of course a solution. Nevertheless, we cannot properly understand it without turning back to the same foundational principles of Statistical Mechanics that allows to explain thermal equilibration in isolated systems. We now make a short review of these concepts.\\
\subsection{Coarse-grained description of quantum systems}\label{subsection:CGRepresentations}
We depart from this assumption, that a quantum system of $N$ particles can be described in terms of an N-body wave function $|\Psi_N\rangle$ and the N-body density matrix will be
\begin{equation}\label{eq:NPureState}
\hat \rho_N = |\Psi_N\rangle \langle \Psi_N|
\end{equation}
While the treatment could be generalized to non-pure N-body density matrices we will only address here systems that are pure in its full quantum representation, i.e. at the N-body level of description. One of the reasons why this is not necessary is because any model and measurements concerning a quantum many-body systems will be restricted anyway due to the limited amount of information about it that we can actually access and manage. Describing a N-particle wave function is basically impossible as soon as we hit a not so high number of particles due to the fast growth of the dimensionality of the Hilbert space required to represent it when N becomes large.\\
The conventional strategy is to work with reduced representations of the system, i.e. coarse-graining our model of it. Such representations are only as complicated as the one- or two-body correlations at most. For example, we can formally define the following correlation functions
\begin{eqnarray}
G_1(\vec r| \vec r^\prime) &= & \langle \hat \psi^\dagger(\vec r)\, \hat \psi(\vec r^\prime) \rangle\\
. \, . \, .& & \nonumber\label{eq:Traditional1BodyCorrelations} \\
G_n(\vec r_1, ..., \vec r_n| \vec r_1^\prime, ..., \vec r_n^\prime) & =& \left\langle \prod_{j=1}^n\hat \psi^\dagger(\vec r_j)\,\prod_{k=1}^n \hat \psi(\vec r_k^\prime) \right\rangle\label{eq:TraditionalNBodyCorrelations}
\end{eqnarray}
If we intend to work in terms of values {\it per particle} it is convenient to define an artifact that provides the same information than these correlation functions but `scaled'. We can then use the {\it n-body Reduced Density Matrices}, that can be defined recursively in the context of quantum systems of identical particles as follows
\begin{equation}\label{eq:DefinitionPartialTrace}
\hat \rho_{n-1} = \mbox{Ptr}_n (\hat \rho_n) = \frac{1}{n}\sum_{j=1}^M \hat a_j\, \hat \rho_n\, \hat a_j^\dagger
\end{equation}
These Reduced Density Matrices correspond to a (fictitious) system of $n$ particles that presents the same expectation values `per particle' as the actual N-body system that we are really studying, so that
\begin{equation}\label{eq:QuantumIntensiveVariables}
G_n(\vec r_1, ..., \vec r_n| \vec r_1^\prime, ..., \vec r_n^\prime) = {{N}\choose{n} }\langle \vec r_1, ..., \vec r_n |\hat \rho_n| \vec r_1^\prime, ..., \vec r_n^\prime \rangle
\end{equation}
As long as $n\ll N$ using the reduced representation is considerably simpler. In the present work we focus on $n=1$. In general, since the access to information corresponding to correlations of very many particles is strongly limited, it is not common to consider $G_n$ for $n>1$.
\subsection{Purity as a quantifier of quantum coherence}\label{subsection:PurityAsQuantifier}
The next aspect we must consider is when can we claim that the system has lost quantum coherence. Under the previously exposed assertion, i.e. that all the information of the system available is contained within its coarse-grained representation, the most natural criterion is to accept that a quantum many-body system will be coherent if its coarse-grained representation is. But once we accept this a less obvious aspect requires to be determined as well, and this is how are we going to quantify the degree of coherence of the system.\\
In an experimental context, for example, quantum decoherence manifests though the loss of imaging contrast in the interference patterns between two or more drops of a coherent gas sample, like a Bose-Einstein Condensate \cite{Ketterle1997,Ott2004}. But since our analysis is theoretical we need a more fundamental quantity, more general and simpler to calculate. Given the relation between decoherence and entanglement it makes sense to consider some form of entropy, which is used in the context of quantum information theory as an acceptable quantifier of the degree of entanglement between the system and either external agents or unobserved degrees of freedom. While there are several available forms and definitons of entropy we will make use here of the simplest possible from the algebraic point of view: the Renyi entropy of index two, $S_{2}$, also called linear entropy
\\
\begin{eqnarray}\label{eq:RenyiEntropy}
S_{\nu} &=& \frac{1}{1-\nu} \, \log{\mbox{Tr}\left( \hat \rho^\nu \right)}
\end{eqnarray}
where we have made use of the natural representation of the on-body reduced density matrix\\
\begin{equation}\label{eq:NaturalRepresentation}
\hat \rho_1 = \sum_{\alpha}p_\alpha \, |\alpha\rangle \langle \alpha|
\end{equation}
The greatest advantage of this particular choice is that it is specially simple to calculate because it is directly related to the so called quantum purity $P$ \\
\begin{eqnarray}\label{eq:DefPurity}
P &=& e^{-S_2} = \mbox{Tr}\left( \hat \rho_1^2\right)= \sum_{p,q=1}^M |\rho_1(p|q)|^2
\end{eqnarray}
The latter one requires no explicit diagonalization of the reduced density matrix to calculate it. We will use it in this work as a rough quantifier of the degree of decoherence achieved by our test system after undergoing a relaxation process.
\subsection{Time-averaging and its role in relaxation}\label{subsection:TimeAveraging}
Our concept of coherence loss will be based in this idea, that quantum purity will show the tendency to evolve towards low values. But against this criterion the following objection could be argued - that it is not guaranteed that for a given system any quantity that we measure (except for a few constants of motion) will evolve towards a constant or merely stable value.\\
Even more, in the case of quantum systems with a discrete energy spectrum any amount that we observe is always a sum of periodic terms, which frequencies are given by the Hamiltonian eigenvalues. Therefore, any measured quantity will return to its initial value or a very similar one after a long enough time lapse. Although it is not formally a quantum observable, quantum purity would be no exception.\\
This question is not new. A very similar one was posed in the context of Classical Statistical Mechanics, where Poincar\'e's Theorem ensures the return to conditions similar to the initial ones after a long enough time lapse. In that case one of the basic aspects to formulate the theory in a consistent way was not only to coarse-grained the description of the system to a macroscopic representation of it - but also to abandon the instantaneous values of the variables considered $A(t)$ in favor of its time-averaged counterparts $\bar A$
\begin{eqnarray}\label{eq:DefTimeAverage}
\bar A(T) = \frac{1}{T}\int_0^t dt^\prime\, A(t^\prime)
\end{eqnarray}
and that for practical purposes we could consider that the system evolved fast enough to take the infinite time limit $\bar A = \lim_{T\rightarrow \infty} \bar A(T)$.\\
The argument is that the time scales considered when we analyse the system from a macroscopic point of view are significantly larger than the time scales relevant to describe its microscopic evolution. Then, even admitting that the system is under constant evolution, it would not make much sense to examinate carefully the instantaneous values and it would be more useful to take the averaged value $\bar A$.\\
Once we admit that $\bar A$ should be considered the physically relevant quantity, one of the foundations of most of the formalism of Statistical Mechanics is to assume that the system considered will hold the Ergodic Hypothesis. This is, that given enough time, we can consider that it will evolve through all of the available region of phase space uniformly. Despite being a core aspect of the standard formalism, the Ergodic Hypothesis is far from being a settled question \footnote{Let this textbook quotation be a sample: {\it "The Ergodic Theorem has so far been an interesting mathematical exercise irrelevant to physics"}. From K. Huang, {\it Statistical Mechanics} (2nd Edition), John Wiley \& Sons Inc. (1987)}. In the quantum context such controversy only gets worse: for example, the first version of the Quantum Ergodic Theorem from J. von Neumann \cite{vonNeumann} was for a long time misinterpreted \cite{Goldstein2010} and the thermalization of isolated quantum systems has been the subject of very recent discussions \cite{Rigol2008, Cramer2008,Rigol2012,Gogolin2016}. Although these are fascinating questions, we refrain from arguing here whether the values after relaxation do coincide with those supplied by the Ergodic Hypothesis or not, or if they correspond to the values of a state of thermal equilibrium. We will be satisfied as long as the value of quantum purity after a relaxation process can be considered stable.\\
This is precisely the aspect that we should most worry about: even accepting that the quantities evaluated, purity $\bar P$ in our case, do `average away', we keep the uncertainty of having spontaneous revivals. Does this mean that a system that departs from a highly coherent state and undergoes a relaxation process that makes it lose its coherence could return to its initially coherent state given enough time? If such revivals do happen, does it still make sense to talk about decoherence even then? The answers are, respectively, it depends and yes.\\
Just as for a classical system, the question is not simply about fulfilling the Ergodic Hypothesis at the limit $T \rightarrow \infty$. We should wonder instead if the time required for the system to fulfil the hypothesis to a reasonable degree of precision is not too long; and how frequently and in what magnitude do the instantaneous values deviate from the time averaged one. In the latter sense, it is more meaningful to pay attention to the amplitude of the deviations
\begin{eqnarray}\label{eq:DefAverageFluctuations}
\bar \sigma_{A}^2(T) = \frac{1}{T}\int_0^t dt^\prime\, \left(A(t^\prime) - \bar A(T) \right)^2
\end{eqnarray}
The most convincing result in Statistical Machanics is precisely that for many systems the amplitude of such deviations is of order $\sigma_x \approx O(V^{-1/2})$ for intensive variable (i.e. for values {\it per particle}) $x=\frac{X}{V}$ when the system is very large.\\
In our case the parameter of interest is quantum purity. Being a non-linear function of the reduced density matrix, purity is not strictly a quantum observable and we must expect different behaviour in comparison with actual observables but we can be guided by exactly the same general criterion. When a quantum many-body system departs from an initial state with a coherent coarse-grained representation, we expect the corresponding quantum purity to be high and we will say that the system loses coherence if the time-averaged value $\bar P$ decays after some time to a value comparatively smaller. And we expect the value of purity fluctuations $\bar \sigma_P$ to be very small in comparison with the amount of purity loss. We can guess too that the amplitude of the fluctuations will drastically decay as we consider larger and larger systems.\\
Given the arguments exposed so far, it should be out of discussion already that Self-Induced Decoherence should be seriously considered as an alternative mechanism to Environmentally-Induced Decoherence. Nevertheless, we should still discuss in further depth the characteristics of a system that allow the phenomen, and consequently, which of its attributes require more attention if we are to predict at best this behaviour.
\subsection{The Bose-Hubbard Hamiltonian as a test system}\label{subsection:IntroBH}
In the following sections we will use the Bose-Hubbard Hamiltonian to illustrate the principles that we are explaining. We made this particular choice because it is simple to find numerical exact solutions for a few particles and because it has been studied exhaustively analytically, numerically and experimentally (as a model for one-dimensional lattices in magneto-optical traps) \cite{Bloch2005}.\\
\begin{eqnarray}\label{eq:DefBoseHubbardHamiltonian}
\hat H_{BH} &=& \sum_{m=1}^M \left( \hat a^\dagger_{m} \hat a_{m+1} + \hat a^\dagger_{m+1} \hat a_{m}+ V \, \hat n_m (\hat n_m-1)\right)\nonumber\\
& &
\end{eqnarray}
If this form is unfamiliar to the reader it is only because we used $\hbar = J = 1$ and $V=U/J$.\\
For our simulations we will take periodic boundary conditions and a number of lattice sites $M$ and a number of particles $N$ quite limited, $N,M\leq 6$. The figures will focus on the case $N=M=5$ and $V=1/4$ to avoid unnecessary redundancy. Although the systems so modelled are quite small we will find out that the principles here exposed are already fulfilled in these cases.\\
We will also choose a concrete initial state for our numerical simulations, one that has all of the particles in one lattice site at position $x$
\begin{eqnarray}\label{eq:InitialStateSingleModeBH}
|\Psi_N^0\rangle &=& \frac{(\hat a^\dagger_{x=1})^N}{\sqrt{N!}}|0\rangle = |n_1=N,n_{j\neq 1}=0\rangle
\end{eqnarray}
By exact diagonalization of this Hamiltonian we can show the N-body evolution of this particularly simple initial state at practically any time.
\section{Analytical description of Self-Induced Decoherence}
\label{section:Analysis}
To understand the causes of Self-Induced Decoherence from an analytical point of view our strategy is to formulate the opposite question. Given the vast amount of possible N-body Hamiltonians that could model a physical system, why would any Hamiltonian in particular conserve the quantum purity of the coarse-grained representation of the system? If we knew nothing about how the many-body Hamiltonian has been formulated, if we did not know that it has been written using the second quantization formalism out of one- or two-particle terms, and required only the Hamiltonian to be a hermitian operator in the N-particle space, the we would consider the conservation of coarse-grained quantities to be a very formidable and counter-intuitive feature that only a few Hamiltonians may fulfil. The key is first of all to understand that a Hamiltonian that does so should be the exception, not the rule. And once we determine the characteristics that allow such Hamiltonian to behave in this anomalous way, we may remove them one by one to evaluate the role of each.\\
The obvious shortcut offered by this strategy resides in the fact that we do indeed know a set of Hamiltonians that do conserve the quantum purity of the reduced representation of the system. They are of course the non-interacting Hamiltonians. In its second quantization representation they have the form
\begin{eqnarray}\label{eq:IdealHamiltonian2ndQuantization}
\hat H_0 &=& \sum_{k=1}^M \hbar \, \omega^0_k \, \hat a_k^\dagger \hat a_k
\end{eqnarray}
where $k=(1,...,M)$ are the $M$ energy eigenstates of the system when there is only one particle. Since we are not interested in open systems we will not work in second quantization any further and we will instead restrict our analysis to the Fock layer of $N$ particles, so that our Hamiltonian (restricted to that Fock layer) can be written as
\begin{eqnarray}\label{eq:IdealHamiltonianNBody}
\hat H_0^{(N)} &=& \sum_{\vec n \,/\, N,M} \hbar\, \omega_{\vec n}^0| \omega_{\vec n}^0\rangle \langle \omega_{\vec n}^0 |\\
&=& \sum_{\vec n \,/\, N,M} \hbar\, \vec n \cdot \vec \omega^0\,| \vec n\rangle \langle \vec n |
\end{eqnarray}
where the ideal eigenvalues are expressed as
\begin{eqnarray}\label{eq:IdealHamiltonianEigenvalues}
\omega_{\vec n}^0 &=& \vec n \cdot \vec \omega^0 = \sum_{p=1}^M n_p \omega_p^0
\end{eqnarray}
and the ideal eigenstates are
\begin{eqnarray}\label{eq:IdealHamiltonianEigenstates}
|\omega_{\vec k}^0\rangle &=& | \vec k \rangle
\end{eqnarray}
We have used the following non-conventional notation for the sums to shorten the expressions
$$ \sum_{\vec n\, /\, N,M} \rightarrow \sum_{n_1+n_2+...+n_M = N}$$
this is, $\vec n / N, M$ denotes all distributions of $N$ particles among $M$ single-particle levels, i.e. all vectors $\vec n = (n_1,n_2,...,n_M)$ with $\sum_{m=1}^M n_m = N$. We express in a similar way the vector of non-interacting eigenvalues $\vec \omega^0 = (\omega^0_1,...,\omega^0_M)$.\\
If we know the Hamiltonian in its diagonal form it will be simple also to determine the state of the system in any future instant by using the N-body time evolution operator
\begin{eqnarray}\label{eq:IdealTimeEvolutionOperator}
\hat U_0(t) &=& \sum_{\vec n \,/\, N,M} e^{-i\, \omega^0_{\vec n}}\,| \omega_{\vec n}^0\rangle \langle \omega_{\vec n}^0|\\
& =& \sum_{\vec n \,/\, N,M} e^{-i\, \vec n \cdot \vec \omega^0}\,| \vec n\rangle \langle \vec n |
\end{eqnarray}
Taking any initial N-body state and the above described operator we can iterate the partial trace operation (\ref{eq:DefinitionPartialTrace}) or use the conventional one-body correlations (\ref{eq:Traditional1BodyCorrelations}) until we obtain the one-body representation at that instant and we can see that its matrix elements in the one-body energy basis are
\begin{eqnarray}
\rho_1(p|q) &=& \frac{1}{N}\langle \Psi_N(t)|\hat a_p^\dagger \hat a_q | \Psi_N(t) \rangle\nonumber\\
&=& \sum_{\vec n \, /\, N-1,M} \frac{\sqrt{(n_p+1)(n_q+1)}}{N}\nonumber\\
& &\times \langle \vec n q| \,e^{-i \frac{t}{\hbar}\, \hat H_0 } |\Psi_N^0 \rangle \langle \Psi_N^0 | e^{i \frac{t}{\hbar}\, \hat H_0 } \,| \vec n p\rangle\label{eq:IdealEvolution1RDMnone}\\
&=& \sum_{\vec n \, /\, N-1} e^{-i t (\omega^0_{\vec n p} - \omega^0_{\vec n q})}\, \frac{\sqrt{(n_p+1)(n_q+1)}}{N}\nonumber\\
& & \langle \vec n p | \Psi_N^0\rangle \langle \Psi_N^0 | \vec n q \rangle\label{eq:IdealEvolution1RDMstates}\\
&=& e^{-i t (\omega^0_{p} - \omega^0_{q})}\, \sum_{\vec n \, /\, N-1} \frac{\sqrt{(n_p+1)(n_q+1)}}{N}\nonumber\\
& & \langle \vec n p | \Psi_N^0\rangle \langle \Psi_N^0 | \vec n q \rangle\label{eq:IdealEvolution1RDMenergies}\\
&=& e^{-i t (\omega_p^0 - \omega_q^0)}\,\rho^0(p|q)\label{eq:IdealEvolution1RDM}
\end{eqnarray}
where we used the simplified notation $\sqrt{n_p+1}|\vec n p \rangle = \hat a_p^\dagger | \vec n \rangle $.\\
Logically, we have obtained the same result that one could expect from the ideal Hamiltonian - that every matrix element in the one-body energy representation will only change by a phase factor $e^{-i t (\omega_p^0 - \omega_q^0)}$. The interesting step now is to pay attention to how this simplification was achieved. It has been so because of two happy coincidences that may not have to take place in an arbitrary N-body hermitian operator. One of them happens as we pass from (\ref{eq:IdealEvolution1RDMnone}) to (\ref{eq:IdealEvolution1RDMstates}), and it happens because $\langle \vec n p| \omega^0_{\vec m}\rangle = \delta_{\vec m, \vec n p}$. The second one happens as we pass from (\ref{eq:IdealEvolution1RDMstates}) to (\ref{eq:IdealEvolution1RDMenergies}), because $\omega^0_{\vec n p} - \omega^0_{\vec n q} = \omega^0_p - \omega^0_q$. These are the two features of the non-interacting Hamiltonian that allow it to conserve the purity of its coarse-grained representation. Now we will proceed to evaluate their individual roles.
\subsection{Roles of Energy Shifts and Eigenstate Deformation}\label{subsection:EDandES}
Among all possible N-body Hamiltonians that one may take the non-interacting ones are the exception, not the rule. Consequently, we must now study the effect of adding interactions. To make this analysis easier we will assume that this interaction is weak enough to consider that we are not `far' from the ideal case. This means that, departing from a known non-interacting Hamiltonian $\hat H_0^{(N)}$ as described in (\ref{eq:IdealHamiltonianNBody}), its interacting counterpart $\hat H^{(N)}$ can be written in the following form
\begin{eqnarray}\label{eq:defEDandES}
\hat H^{(N)} &=& \sum_{\vec k\, /\, N} \hbar \omega_{\vec k} \, |\omega_{\vec k}\rangle \langle \omega_{\vec k} |\\
\omega_{\vec k} &=& \vec k \cdot \vec \omega^0 + \Delta_{\vec k} \\
|\omega_{\vec k} \rangle &=& |\vec k \rangle + |\phi_{\vec k}\rangle
\end{eqnarray}
This is, as we apply the interaction, we cause each N-body energy level $\omega_{\vec n}$ to be displaced by a quantity that we will call {\it Energy Shift} $\Delta_{\vec n}$, and that the natural N-body basis of the Hamiltonian suffers a unitary transformation that displaces each eigenvector according to $|\omega_{\vec n}\rangle = |\omega_{\vec n}^0\rangle + |\phi_{\vec n}\rangle$. We will refer to $|\phi_{\vec n}\rangle$ or its components $\phi_{\vec m,\vec n}=\langle \vec m | \phi_{\vec n}\rangle$ as the {\it Eigenstate Deformation}. As we have seen above when we calculated (\ref{eq:IdealEvolution1RDM}) each one of these two attributes correspond to a qualitatively different deviation from the non-interacting behaviour: the Energy Shifts will stop us from taking the simplification used between (\ref{eq:IdealEvolution1RDMstates}) and (\ref{eq:IdealEvolution1RDMenergies}), while the Eigenstate Deformations forbid the step taken from (\ref{eq:IdealEvolution1RDMnone}) to (\ref{eq:IdealEvolution1RDMstates}).\\
\begin{table}[]
\begin{tabular}{|l|c|c|}
\hline
& $\Delta_{\vec n}=0$ & $\Delta_{\vec n}\neq 0$\\ \hline
$|\phi_{\vec n} \rangle=0$ & $\hat H_0$ & $\hat H_D$ \\ \hline
$|\phi_{\vec n} \rangle\neq 0$ & $\hat H_S$ & $\hat H$ \\ \hline
\end{tabular}\caption{A schema of the strategy followed to explore the role of Energy Shifts $\Delta_{\vec n}$ and Eigenstate Deformations $|\phi_{\vec n}\rangle$. We build two different Hamiltonians $\hat H_S$ and $\hat H_D$ combining properties from the interacting Hamiltonian $\hat H = \hat H_0 + \hat V$ and its non-interacting counterpart $\hat H_0$.}\label{tab:EDandES}
\end{table}
According to everything we have presented so far, it makes sense to ask what are the roles that the Energy Shifts and the Eigenstate Hamiltonians play individually. To answer this we will follow the strategy described in Table \ref{tab:EDandES}. We design artificial N-body Hamiltonians that show each one of these deviations from the ideal case in an isolated way. They will be the {\it Shifted Hamiltonian} $\hat H_S$ and the {\it Deformed Hamiltonian} $\hat H_D$\\
\begin{eqnarray}\label{def:HDandHS}
\hat H_S &=& \sum_{\vec k\, /\, N} \hbar \omega_{\vec k} \, |\vec k\rangle \langle \vec k |\\
\hat H_D &=& \sum_{\vec k\, /\, N} \hbar \vec k \cdot \vec \omega^0 \, |\omega_{\vec k}\rangle \langle \omega_{\vec k} |
\end{eqnarray}
The compared effect of time evolution under $\hat H_D$ or $\hat H_S$ can be seen in Fig.~\ref{fig:CompareEXESEDoccupationRelax} y Fig.~\ref{fig:CompareEXESEDoccupationSteady}, where we display the time evolution of the occupation number of the initially populated lattice site according to the four different Hamiltonians. In Fig.~\ref{fig:CompareEXESEDoccupationRelax} we see that for short times both $\hat H_D$ and $\hat H_S$ yield similar qualitatively similar results. This is because the both capture the essential component of the evolution, which is nothing but `free flight', i.e. the dynamics provided by the non-interacting Hamiltonian. Still, once a reasonably long time has passed (see Fig.~\ref{fig:CompareEXESEDoccupationSteady}), the real system behaves in a way that we could call closer to ``equilibrium'' in the sense explained in previously in Section \ref{subsection:TimeAveraging}: a low and stable average value and small fluctuations around it. This long term dynamics is reasonably well described by $\hat H_S$ but not so by $\hat H_D$, that looks more similar to the evolution given by the ideal Hamiltonian.\\
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{FIG_nj_relax_compare_exSSED_N5M5V2.eps}\caption{Compared evolution of the occupation of the site initially populated.$n_j(t=0)=1$ when we evolve under the Energy Shift (red) and using the Eigenstate Deformation (blue) Hamiltonians. At early stages of the relaxation process both approaches seem similar}\label{fig:CompareEXESEDoccupationRelax}
\includegraphics[width=0.5\textwidth]{FIG_nj_steady_compare_exSSED_N5M5V2.eps}
\caption{Same as Fig.~\ref{fig:CompareEXESEDoccupationRelax} but after a reasonably long time. After a long enough evolution time only the Energy Shift Hamiltonian yields qualitatively realistic results.}\label{fig:CompareEXESEDoccupationSteady}
\end{figure}
\end{center}
Nevertheless, evaluating the time-evolved matrix elements we can assert that the Shifted Hamiltonian has one important weakness
\begin{eqnarray}\label{eq:Evolution1RDMwithHS}
\rho_1^{S}(p|q) &=& e^{-i t (\omega_p^0 - \omega_q^0)}\, \sum_{\vec n \, /\, N-1,M} e^{-it(\Delta_{\vec n p}-\Delta_{\vec n q})}\nonumber\\
& & \frac{\sqrt{(n_p+1)(n_q+1)}}{N} \langle \vec n p | \Psi_N^0\rangle \langle \Psi_N^0 | \vec n q \rangle\nonumber\\
& &
\end{eqnarray}
The occupation numbers in the one-body energy representation do not evolve under the Shifted Hamiltonian because $[\hat H_S,\hat H_0]=0$. This does not happen for the Deformed Hamiltonian, where
\begin{eqnarray}\label{eq:MomentumOccupationNumbersED}
\rho_1^{D}(p|q) &=& \sum_{\vec m, \vec m^\prime / N} e^{-it(\vec m-\vec m^\prime)\cdot \vec \omega^0}\, R(\vec m, \vec m^\prime; p,q)\\
R(\vec m, \vec m^\prime; p,q) &=& \sum_{\vec n \, /\, N-1,M} \frac{\sqrt{(n_p+1)(n_q+1)}}{N} \nonumber\\
& & \times\, \langle \omega_{\vec m}|\Psi_N^0\rangle \langle \Psi_N^0| \omega_{\vec m^\prime}\rangle \,\langle \vec n p |\omega_{\vec m}\rangle \, \langle \omega_{\vec m^\prime}| \vec n q \rangle\nonumber\\
& &
\end{eqnarray}
because in this case $\langle \omega_{\vec m}| \vec n p \rangle \neq \delta_{\vec m,\vec n p}$.\\
If our goal was to obtain the values of physical observables from the reduced representation once the stationary state has been achieved, then the description offered by $\hat H_D$ would be more interesting because it keeps more detailed information about the components of the initial state that will remain stationary. In turn, if we are interested about the description of the mid- and long-term dynamics then $\hat H_S$ yields qualitatively more realistic results. We can glimpse the reason when we compare what makes (\ref{eq:Evolution1RDMwithHS}) different from (\ref{eq:MomentumOccupationNumbersED}).\\
In the first one (\ref{eq:Evolution1RDMwithHS}) the deviations from the ideal dynamics given by $\hat H_0$ come from phase factors that depend only on the differences between Energy Shifts $\Delta_{\vec n p}-\Delta_{\vec n q}$, and since these are relatively small in comparison to other phase changes they will produce changes that are much slower than the non-interacting evolution. But because such phase factors do affect each and every one of the components $\langle \vec n p | \Psi_N^0\rangle$, this implies that in the long run the differences with respect to ideal evolution will not be limited in amplitude.\\
On the contrary, the differences from ideal evolution and the evolution described by the Deformed Hamiltonian (\ref{eq:MomentumOccupationNumbersED}) will have different components because now $\langle \omega_{\vec m}| \vec n p \rangle = \delta_{\vec m,\vec n p} + \phi_{\vec m,\vec n p}$, with $\phi_{\vec m,\vec n p}\neq 0$. The component with the highest amplitude in (\ref{eq:MomentumOccupationNumbersED}) will follow from the terms where $\delta_{\vec m,\vec n p} \delta_{\vec m^\prime,\vec n q}$. But this component is precisely the equivalent to non-interacting evolution. Any deviation from ideal evolution will be given by the rest of the terms, all of them proportional in amplitude $\phi_{\vec m,\vec n p}$ o $\phi_{\vec m^\prime,\vec n q}$. This means that the deviations from ideality provided by $\hat H_D$ will be limited by the amplitude of the Eigenstate Deformation, that we could quantify through
\begin{eqnarray}\label{eq:TotalEigenstateDeformation}
D &=&\mbox{Max} \left[|\phi_{\vec n,\vec m} |^2\right]_{\vec n, \vec m \,/\, N,M}
\end{eqnarray}
\begin{table}[]
\begin{tabular}{|c|c|c|c|}
\hline
N \,/\, M & 4 & 5 & 6 \\ \hline
4 & >1.0e-45 & 9.0e-06 & 3.0e-05 \\ \hline
5 & 8.9e-07 & 7.0e-05 & 5.6e-05 \\ \hline
6 & 1.3e-05 & 1.5e-04 & 1.1e-04\\ \hline
\end{tabular}\caption{Estimation of the Eigenstate Deformation $D$ for the system sizes simulated, for $V=1/4$. We see that they remain at low values in all cases.}\label{tab:EigenstateDeformation}
\end{table}
In our numerical simulations this parameter is not larger than $O(10^{-4})$, as we can see in Table \ref{tab:EigenstateDeformation}. We can consider that this quantity will be always small as long as we an consider that the interactions are weak. Furthermore, considering cases beyond this scenario would not be consistent with our convention of using the same labelling $|\omega_{\vec n}\rangle$ for the eigenstates of both the ideal and the interacting Hamiltonian because our initial claim, i.e. that we can do this because the deformed eigenstates are very close to the non-deformed ones, will not be true any longer.\\
We are going to use a similar argumentation to analyse the evolution of quantum purity of the reduced representation. But since the non-interacting Hamiltonian conserves purity perfectly the deviations from ideal behaviour will be even more apparent.
\subsection{Evolution of quantum purity}\label{subsection:PurityEvolution}
While the bahaviour of all the four Hamiltonians is similar in the previous context, i.e. refered to the dynamics of quantum observables in the reduced representation (in the formal sense of observable), this is in part because the core of the dynamics was still what we informally called `free flight', this is, dynamics in the absence of interactions as given by $\hat H_0$. This changes drastically once we focus on the evolution of quantum purity, which is not an observable in the formal sense, and remains constant under such `free flight'.\\
We can see that this is true in Fig.~\ref{fig:PurityRelaxationEXES}, where the evolution of the quantum purity under the different Hamiltonians is represented. The system departs from a state that is pure $P(t=0)=1$ in its coarse-grained representation. The evolution under $\hat H_0$ keeps purity constant but the differences between $\hat H_D$ (blue) and $\hat H_S$ (red) are now much more evident, being still $\hat H_S$ the one that describes better the dynamics of the actual Hamiltonian $\hat H$ (black).
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{FIG_Purity_Relax_exact_SS_N5M5V2.eps}
\caption{Compared evolution of the purity of the one-body reduced density matrix when we evolve under the Energy Shift (red) and using the Eigenstate Deformation (blue) approximations. Even for systems of relatively small size ($N=M=5$ is represented, interaction strength $V=1/4$) the purity decays to values close to the minimum. Only the Energy Shift Hamiltonian predicts this decay.}\label{fig:PurityRelaxationEXES}
\end{figure}
\end{center}
The differences between them have an origin similar to that discussed in the previous section. $\hat H_D$ fluctuates fast and with limited amplitude, while $\hat HS$ still has phase factors with much smaller frequencies - therefore slower - but its fluctuations are unbounded in amplitude. Since quantum purity is - among all clues of coherence - the simplest to calculate analytically, we can use the previous result (\ref{eq:Evolution1RDMwithHS}) to explore this behaviour. Then we can write for the Shifted Hamiltonian
\begin{eqnarray}\label{eq:PurityEvolutionHS}
P(t)&=&\sum_{p,q} | \rho_1(p|q) |^2 \nonumber \\
&=& \sum_{p, q} \sum_{\vec n, \vec m\,/\, N-1,M} \, e^{-it(\Delta_{\vec n p}-\Delta_{\vec n q}-\Delta_{\vec m p}+\Delta_{\vec m q})}
\nonumber\\
& & \times \,\frac{\sqrt{(n_p+1)(n_q+1)(m_p+1)(m_q+1)}}{N^2} \nonumber\\
& & \times \,\langle \vec n p|\Psi_N^0\rangle \langle \Psi_N^0 | \vec n q \rangle\,\langle \Psi_N^0 | \vec m p\rangle \langle \vec m q|\Psi_N^0 \rangle\nonumber\\
& &
\end{eqnarray}
Consistently with the statement made, that our conception of Self-Induced Decoherence is based on the same principle of time-averaging and limited fluctuations borrowed from Statistical Mechanics, it makes sense to split the terms from the expression above in its stationary and fluctuating components
\begin{eqnarray}
P(t) &=& P_{AV} + \delta P(t)
\end{eqnarray}
where the stationary value $P_{AV}$ is the sum of all terms where the phase factors cancel (either because $p=q$ or else because $\vec n =\vec m$)
\begin{eqnarray}\label{eq:PurityBaseValueHS}
P_{AV} &=& \sum_{p,q}\sum_{\vec n \, /\, N-1}\frac{(n_p+1)(n_q+1)}{N^2}\,|\langle \vec n p | \Psi_N^0 \rangle |^2 |\langle \Psi_N^0 | \vec n q \rangle|^2\nonumber \\
& & + \sum_p \sum_{\vec n \neq \vec m} \frac{(n_p+1)(m_p+1)}{N^2}\left| \langle \vec n p|\Psi_N^0\rangle \right| ^2\left| \langle \vec m p|\Psi_N^0 \rangle \right| ^2\nonumber \\
& &
\end{eqnarray}
The fluctuations $\delta P(t)$ correspond to the terms where phase factors do not cancel ($p\neq q$ and $\vec n \neq \vec m$)
\begin{eqnarray}\label{eq:PurityFluctuationsHS}
\delta P(t) &=& \sum_{p \neq q} \sum_{\vec n \neq \vec m\,/\, N-1,M} \, e^{-it(\Delta_{\vec n p}-\Delta_{\vec n q}-\Delta_{\vec m p}+\Delta_{\vec m q})}\, \nonumber\\
& &\times\, \frac{\sqrt{(n_p+1)(n_q+1)(m_p+1)(m_q+1)}}{N^2}\nonumber\\
& & \times \, \langle \vec n p|\Psi_N^0\rangle \langle \Psi_N^0 | \vec n q \rangle\,\langle \Psi_N^0 | \vec m p\rangle \langle \vec m q|\Psi_N^0 \rangle \nonumber\\
& &
\end{eqnarray}
Our main interest concerning the fluctuations it to determine if their amplitude is going to remain bounded most of the time. If the contributing terms add constructively very often we would sporadically observe sudden peaks of high quantum purity, something that would not fit our concept of decoherence.\\
Our plan implies to consider that terms with phase factors of different frequencies can be treated as independent random variables. This will require those frequencies to respect certain conditions: they should of course be different from each other; they should not be an integer multiple of each other (i.e. they are incommesurable with respect to each other); and the differences between them should not be so small that we would not be able to distinguish them on a reasonable observational time scale (see the Appendix \ref{appendix:SumOscillating}). In practice this hypothesis may not always be fulfilled altogether. But this is not the case of the concrete model that we are using as an example, as we can see represented in Fig.~\ref{fig:DifferenceOfEnergyShifts}. From all terms with phase factors dependent on the differences between Energy Shifts $|\Delta_{\vec m} - \Delta_{\vec m^\prime}|$ that could contribute to the fluctuations $\delta P(t)$, all those that are not of the form $(\alpha,\beta)=(\vec n p, \vec n q)$ will have phase factors that cancel and these terms contribute to the average value instead, just as we expected - but among those that do have this form (represented in the figure) only a few are small enough to be considered non-fluctuating terms. If we take the simulation time represented in Fig.~\ref{fig:PurityRelaxationEXES} as our reference time scale then the threshold to discard the lowest values is $1/T\approx 0.033$, leaving 503 (72$\%$) of them above it. \\
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{FIG_elespectro_N5M5V2.eps}
\caption{Energy spectrum of the Bose-Hubbard Hamiltonian for $N=M=5$ and $V=1/4$. We see that the interaction displaces the energy levels (blue) from the non-interacting ones (red). These differences are the Energy Shifts $\Delta$}\label{fig:EnergyShifts}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{FIG_Difference_Shifts_N5M5V2.eps}
\caption{Differences between the energy shifts $|\Delta_{\vec n p}-\Delta_{\vec n q}|$ involved in each one of the phase factors in the sum (\ref{eq:Evolution1RDMwithHS}) that yields the time evolution of the reduced density matrix elements $\rho^S(p|q)$ under the Energy Shift Hamiltonian $\hat H_S$. The size of the Hilbert space for $N=M=5$ is $L_{5,5}= 126$, meaning that up to $7.875$ pairs of shift differences $(\vec m,\vec m^\prime)$ could be involved. But only the shift differences of the form $(\vec m,\vec m^\prime)=(\vec n p,\vec n q)$, $p \neq q$ are relevant (only $700$ possibilities).}\label{fig:DifferenceOfEnergyShifts}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{FIG_Diff_Of_Diff_N5M5V2.eps}
\caption{Differences between the difference between energy shifts $|\Delta_{\vec n p}-\Delta_{\vec n q} - \Delta_{\vec m p} + \Delta_{\vec m q}|$ involved in each one of the phase factors in the terms contributing to purity fluctuations (\ref{eq:PurityFluctuationsHS}). The amount of shift differences previously calculated amounted to $700$, meaning that up to $244.650$ terms with different frequencies could be involved. But again, only the differences of the form $(\vec n p,\vec n q,\vec m p,\vec m q)$, $\vec n\neq\vec m$, $p \neq q$ are relevant (only $24.150$ possibilities). We have represented in the picture only a sample of them (one out of $50$).}\label{fig:Diff_Of_Diff}
\end{figure}
\end{center}
Notice that the differences represented in Fig.~\ref{fig:DifferenceOfEnergyShifts} $|\Delta_{\vec n p}-\Delta_{\vec n q}|$ will be crucial when distinguishing stationary and fluctuating terms - in the evolution of $\rho^S(p|q)$, just as we expressed in (\ref{eq:Evolution1RDMwithHS}). For the evolution of quantum purity it becomes more complicated because the phase factors depend instead on the differences between the differences between the Energy Shifts $|\Delta_{\vec n p}-\Delta_{\vec n q} - \Delta_{\vec m p} + \Delta_{\vec m q}|$. They are represented in Fig.\ref{fig:Diff_Of_Diff}, obtaining similar results. The principles explained so far are still the same and the main hypothesis is that the resulting frequencies should not cancel or be redundant too often. If we apply the same criterion explained above (time scale from Fig.~\ref{fig:PurityRelaxationEXES} as reference) then 19.389 (80$\%$) of the terms would contribute to the fluctuations.\\
As long as this `incommesurability conjecture' is acceptable we can suppose that each oscillating term is a random variable independent from the others and we can calculate the variance of the sum of all of them as
\begin{eqnarray}\label{eq:Sigma2PurityHS}
\bar \sigma^2_P&=& \sum_{p \neq q} \sum_{\vec n \neq \vec m} \frac{(n_p+1)(n_q+1)(m_p+1)(m_q+1)}{N^4} \nonumber\\
& & \times \,\left| \langle \vec n p|\Psi_N^0\rangle \langle \Psi_N^0 | \vec n q \rangle\,\langle \Psi_N^0 | \vec m p\rangle \langle \vec m q|\Psi_N^0 \rangle \right| ^2\nonumber \\
& &
\end{eqnarray}
The two results above (\ref{eq:PurityBaseValueHS}) and (\ref{eq:Sigma2PurityHS}) can be easily checked in our test system numerically - a Bose-Hubbard Hamiltonian departing from an initial state (\ref{eq:InitialStateSingleModeBH}) with all particles localized in one lattice site at position $x$. This state can be rewritten as
\begin{eqnarray}\label{eq:InitialStateSingleMode}
|\Psi_N^0\rangle &=& \sum_{\vec n / N, M} \sqrt{\frac{N!}{n_1!...n_M!}}\langle 1|x \rangle^{n_1}...\langle M|x \rangle^{n_M}\, |\vec n\rangle\nonumber\\
& &
\end{eqnarray}
Notice that $x$ is an eigenstate of `position' (lattice site) representation, while the indexes in the sum $\vec n = (n_1,n_2,...,n_k,...,n_M)$, with $k=1,2,...,M$ are in the `quasimomentum' (one-body energy) eigenstates.\\
This initial state in particular is very simple because the components of a localized state have all the same amplitude
\begin{eqnarray}\label{eq:ScalarProductPositionMomentumBH}
|\langle x | k \rangle|^2 &=& 1/M
\end{eqnarray}
and this simplifies enormously the calculation. For the average value (\ref{eq:PurityBaseValueHS}) we obtain
\begin{eqnarray}\label{eq:PurityBaseValueSSPISAnsatz}
P_{AV} &=& \frac{1}{M}+M(M-1)\frac{a_{N-1}^{(M)}}{M^{2N}}
\end{eqnarray}
and for the variance of fluctuations(\ref{eq:Sigma2PurityHS}) it is
\begin{eqnarray}\label{eq:Sigma2PuritySSPISAnsatz}
\bar \sigma^2_P &=& \frac{M(M-1)}{2}\left[ \left(\frac{a_{N-1}^{(M)}}{M^{2N}} \right)^2 - \frac{b_{N-1}^{(M)}}{M^{4N}} \right]
\end{eqnarray}
where both results are expressed in terms of sums of powers of the multinomial series
\begin{eqnarray}\label{eq:MultinomialPowerSum}
a_n^{(m)} &=& \sum_{\vec k / n, m} {{n}\choose{\vec k} }^2\\
b_n^{(m)} &=& \sum_{\vec k / n, m} {{n}\choose{\vec k} }^4\\
{{n}\choose{\vec k} }&=&\left(\frac{n!}{k_1!k_2!...k_m!} \right)
\end{eqnarray}
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{FIG_An_Bn_coefficients.eps}
\caption{Behaviour of the coefficients $A_n=n^{-2n}\,a_n^{(n)}$ (red) and $B_n=n^{-4n}\,b_n^{(n)}$ (blue), on which the variance of purity fluctuations $\bar \sigma_P^2$ depend when we evolve under the Shifted Hamiltonian $\hat H_S$.}\label{fig:AnBnCoefficients}
\end{figure}
\end{center}
This result allows us to have an idea of how fast the amplitude of purity fluctuations decay as we consider systems of larger sizes. For example, in Fig.~\ref{fig:AnBnCoefficients} we can see that the terms on which the variance depends $A_n=n^{-2n}\,a_n^{(n)}$ y $B_n=n^{-2n}\,b_n^{(n)}$ decay at least as $\log{A_n} \propto n-1$ (same for $B_n$). The amplitude of purity fluctuations would be notoriously limited, as we expected.
\section{Conclusions}
\label{Conclusions}
In this work we have recalled and supported with numerical evidence the concept of Self-Induced Decoherence for the coarse-grained description of quantum systems. Although this concept is not new it has been mostly underestimated in the previous literature about quantum decoherence. While it is very hard to consider real systems as perfectly isolated, the effect is present and should not be disregarded.\\
The concept of Self-Induced Decoherence exposed here requires us to remember the basic principles of Pure-State Quantum Statistical Mechanics. This is, that our access to the system is limited to a simplified representation of it, based on a few-body (usually one body) Reduced Density Matrix. That when we refer to a `stationary state' of a many-body system after undergoing a relaxation process, it does not mean that the system actually remains stationary, but that the accessible quantities remain {\it at almost any time} around well defined and stable average values. And that the deviations from such average values are only fluctuations of an amplitude that, on average, is extremely small and proportionally less relevant as we consider systems of larger sizes.\\
Keeping these concepts in mind and accepting for the sake of simplicity that coherent quantum systems will present a higher quantum purity than incoherent ones we could verify indeed that the coarse-grained representation of a many-body system of isolated interacting particles loses quantum coherence. We added a qualitative analysis of the causes that make interacting Hamiltonians to lose quantum coherence in the way previously described. It is based on splitting the two obvious characteristics that distinguish an interacting Hamiltonian from its non-interacting counterpart: the Eigenstate Deformation and the Energy Shifts. Our analysis suggests that the first one (small in amplitude but fast) will play a major role in fluctuations, while the second one (large in amplitude but slow) is the responsible for the long time stationary behaviour of quantum purity.\\
About the aforementioned fluctuations, we could make an analytical estimation of its amplitude based on the Energy Shift Approximation and taking an initial state - the fully localized state - that is uniformly distributed among the non-interacting energy spectrum. We could see that this amplitude decreases extremely fast as the size of the system grows. But the actual amplitude of fluctuations will not be that small for two clear reasons. The first one, more obvious, is that the Eigenstate Deformations will add a significant contribution to the purity fluctuations. And the second one, more subtle, is that some of the oscillation frequencies involved in the fluctuating terms of the quantum purity may be either redundant - and therefore result in stationary contributions - or else so similar that they will not be distinguishable in a reasonable observational time scale. The fact that redundancies - in fact, the lack of them - in the N-body energy spectrum plays an important role in relaxation of isolated quantum systems has been already pointed out both explicitly \cite{Reimann2012,Short2012} and implicitly in relation to the Eigenstate Thermalization Hypothesis \cite{Rigol2012,Rigol2016}. But we could observe that for quantum purity this principle becomes slightly more complex because it appeals to the (non-)redundancy of differences of differences of energy shifts. \\
We have put under test all of these concepts in a system simple enough to admit exact solutions - the few-body Bose-Hubard model. With this we have provided numerical evidence supporting that the principles exposed are fulfilled in this case, even for very limited system sizes. For this system and this particular initial conditions we cannot talk properly about relaxation to a stationary state because we can see that observables (local occupation numbers, in particular) do not stop to oscillate even after a long time. But even then we can see that the quantum purity, in turn, does decay to a very stable lower value, a fact that we could interpret as a tendency to incoherent behaviour of the coarse-grained accessible representation.\\
Concerning possible experimental implementations, we can consider ourselves lucky because the Bose-Hubbard model has had its laboratory counterpart for a very long time already, in the form of trapped bosons in a one-dimensional magneto-optical lattice. This is another reason why Self-Induced Decoherence in the Bose-Hubbard model has obvious practical implications. As an example, the interference patterns of the bosons released from a lattice would lose contrast depending on how long they have been stored before being released. Their lifetime as quantum coherent systems would be limited regardless of how well isolated from outer sources of decoherence we keep the sample \cite{Ott2004}.
\begin{appendix}
\section{Sum of oscillating quantities with different frequencies}\label{appendix:SumOscillating}
Consider an oscillatory variable $y$, oscillating with frequency $\Omega$
\begin{eqnarray}
y(t) &=& \sin{(\Omega t + \alpha)}
\end{eqnarray}
This non-random variable can be considered random is the instant $t$ when we measure its value is taken randomly from a sampling interval $T$. If this observation interval is exactly one single oscillation period then the measured values are distributed according to
\begin{eqnarray}
p(y) &=& \frac{1}{\pi}\frac{1}{\sqrt{1-y^2}}
\end{eqnarray}
having for average value and variance
\begin{eqnarray}
\bar y &=& 0\\
\bar \sigma_y^2 &=& \frac{1}{2}
\end{eqnarray}
If instead of a single oscillation period our samplig time $T$ covers an exact multiple of full oscillation periods the results will distribute in the exact same way. But even if $T$ does not cover an exact integer multiple of full periods the results will not differ much in comparison with those obtained for an exact integer multiple as long as $T$ covers many full oscillation periods.\\
If our measured variable $y$ were the sum of many oscillating terms
\begin{eqnarray}\label{eq:SumOfOscillatingVariables}
y(t) &=& \sum_{j} y_j\, \sin{(\Omega_j t + \alpha_j)}
\end{eqnarray}
the terms contributing to $y$ could not be treated as independent random variables if, for example, their frequencies $\Omega_j$ were equal or too close to be distinguished within the sampling time considered $|\Omega_i-\Omega_j|\ll \frac{2\pi}{T}$. \\
But as long as we can consider all the frequencies invoveld as different within our sampling time interval, then we could treat each term as an independent variable from the rest and we can calculate the average value and variance of (\ref{eq:SumOfOscillatingVariables}) as
\begin{eqnarray}
\bar y &=& 0\\
\bar \sigma_y &=& \frac{1}{2}\sum_{j}y_j^2
\end{eqnarray}
\end{appendix}
\bibliographystyle{apsrev}
|
\section*{Acknowledgements}
This work was supported in part by the U.S. Department of Energy (DOE) Grant No. DE-SC0009937 A.K. also acknowledges support by the World Premier International Research Center Initiative (WPI), MEXT, Japan. V.T. would like to thank Aspen Center for Physics for hospitality, where part of the work was done.
|
\section{Introduction}\label{sec:intro}
Massive antenna array technology can help in realizing large beamforming and multiplexing gains~\cite{massive-MIMO}, as desired for the goal of sustainable ubiquitous Internet-of-Things (IoT) deployment~\cite{SP-Mag-Massive-IoT}. However, due to the usage of low-cost hardware components, the performance of these sustainable IoT systems is more prone to suffer from the radio frequency (RF) imperfections~\cite{schenk2008RF} like the in-phase-and-quadrature-phase-imbalance (IQI)~\cite{IQI-2019-Mag}. {Thus, generalized green signal processing techniques are being investigated to combat the adverse effect of hardware impairments~\cite{SPL2-R1-1,SPL2-R1-2,SPL2-R1-3} and the problem of carrier frequency offset (CFO) recovery in frequency-selective IQI~\cite{SPL2-R2-0}.} However, as these impairments adversely influence both channel estimation (CE) and precoding processes at transmitter (TX), new jointly-optimal estimator and beamformer designs are required.
\subsection{State-of-the-Art}\label{sec:RW}
In recent times, there have been increasing interests~\cite{IQI-2019-Mag,Virtual-IQI,Eriksson-TC17-IQI,Schober-TWC17-IQI,Access17-IQI,IQI-AF-Stat-MaxD,LMMSE-IQI-Close} on investigating the performance degradation in energy beamforming (EB) gains of the massive multiple-input-single-output (MISO) systems suffering from IQI. Specifically, as each single-antenna receiver (RX) in multiuser (MU) systems gets wrongly viewed as having a virtual port due to underlying IQI~\cite{Virtual-IQI}, it leads to an inaccurate CE at the multiantenna TX.
Noting it, sum rate limits in downlink (DL) MU MISO systems under IQI and CE errors were derived in~\cite{Eriksson-TC17-IQI}. In contrast~\cite{Access17-IQI,Schober-TWC17-IQI} were targeted towards the joint CE and IQI compensation in uplink (UL) MISO systems. More recently, performance analysis of dual-hop statistical channel state information (CSI) assisted cooperative communications was conducted via simulations in~\cite{IQI-AF-Stat-MaxD} to incorporate the effect of IQI. However, these works \cite{Virtual-IQI,Eriksson-TC17-IQI,Schober-TWC17-IQI,Access17-IQI,IQI-AF-Stat-MaxD} only presented linear-minimum-mean-square-error (LMMSE)~\cite{LMMSE-IQI-Close} based CE, that requires strong prior CSI.
On another front, there are also some works on least-squares (LS) based CE under IQI~\cite{LS-IQI-ICASSP10,ICASSP18,LSE-Pilot-OFDMA}.
A special structured pilot was used in \cite{LS-IQI-ICASSP10} to obtain LS estimator (LSE) for both actual and IQI-based virtual signal terms. However, these complex pilots are not suited for limited feedback settings involving low-power IoT RX. Therefore, LS and LMMSE estimates using conventional methods were presented in~\cite{ICASSP18} to quantify EB gains during MISO wireless power transfer under joint-TX-RX IQI and CE errors over Rician fading. Lately, an LSE using additional pilots to exploit the interference among symmetric subcarriers for mitigating effect of IQI was designed in \cite{LSE-Pilot-OFDMA}.
\subsection{Motivation and Scope}\label{sec:motiv}
All existing works \cite{Virtual-IQI,Eriksson-TC17-IQI,Schober-TWC17-IQI,Access17-IQI,IQI-AF-Stat-MaxD,LMMSE-IQI-Close,LS-IQI-ICASSP10,ICASSP18,LSE-Pilot-OFDMA}, investigating the impact of IQI on CE, considered the underlying additional virtual signal term as interference, and \textit{simply ignored} the information content in it. Likewise, the current precoder designs for multiantenna TX serving single-antenna RX are based on \textit{suboptimal} maximum ratio transmission (MRT) scheme, ignoring the impairment that signal undergoes due to IQI. \textit{To the best of our knowledge, the optimal CE and TX precoder designs respectively minimizing the underlying LS error and maximizing signal power at RX under IQI and CE errors have not been investigated yet.}
Unlike existing works, the proposed \textit{globally-optimal} LSE does not require any prior CSI. {The adopted \textit{novel and generic} complex-to-real-domain transformation based methodology to obtain the LSE and precoder in \textit{closed-form} can be extended for investigating designs in MU and multiantenna RX settings.} Lastly, the proposed precoder design holds for \textit{any CE scheme.}
\subsection{Contribution of This Letter}\label{sec:contrib}
Our contribution is three-fold. (1) \textit{Global-minimizer of LS error} during CE under TX-RX-IQI is derived in closed-form. (2) \textit{Novel precoder design} is proposed to globally-maximize the nonconvex received signal power over IQI-impaired MISO channels. {Extension of this design to multiuser settings is also discussed.} (3) To validate the nontrivial analysis for different system parameters, extensive simulations are conducted, which also \textit{quantify the achievable EB gains} over benchmarks. After outlining system model in Section~\ref{sec:model}, these three contributions are discoursed in Sections~\ref{sec:OCE}, \ref{sec:OTB}, and \ref{sec:results}, respectively.
\section{System Description}\label{sec:model}
In this section we present the system model details, followed by the adopted transmission protocol and IQI signal model.
\subsection{Wireless Channel Model and Transmission Protocol}
We consider DL MISO system comprising of an $N$ antenna source $\mathcal{S}$ and a single-antenna IoT user $\mathcal{U}$. Assuming flat quasi-static Rayleigh block fading~\cite[Ch 2.2]{simon2005digital}, the $\mathcal{U}$-to-$\mathcal{S}$ channel is represented by $\mathbf{h}\sim\mathbb{C} \mathbb{N}\left(\mathbf{0}_{N\times1},\beta\,\mathbf{I}_N\right)$, where $\beta$ incorporates the effect of both distance-dependent path loss and shadowing.
Transmission protocol involves estimation of $\mathbf{h}$ from the received IQI-impaired signal at $\mathcal{S}$. Exploiting channel reciprocity in the adopted time-division duplex mode~\cite{CSI-WET-Rician}, we can divide each coherence block of $\tau$ seconds (s) into two phases, namely CE and information transfer (IT). During CE phase of duration $ \tau_c\le \tau$, $\mathcal{U}$ transmits a pilot signal $\mathrm{s}$ with mean power $p_c$ and the resulting received baseband signal at $\mathcal{S}$ without any IQI is
\begin{equation}\label{eq:rxS}
\mathbf{y}=\mathbf{h}\,\mathrm{s} +\mathbf{n},
\end{equation}
where $\mathbf{n}\in\mathbb{C}^{N\times1}$ is received additive white Gaussian noise (AWGN) vector with zero mean entries having variance $\sigma_{1}^{2}$.
\subsection{Adopted Transmission Protocol}\label{sec:prot}
Our protocol involves estimation of $\mathbf{h}$ from the received IQI-impaired signal at $\mathcal{S}$. Here, exploiting channel reciprocity in the adopted time-division duplex mode~\cite{CSI-WET-Rician}, we can divide each coherence block into two phases, namely CE and information transfer (IT). During CE phase of duration $ \tau_c\le \tau$, $\mathcal{U}$ transmits a pilot signal $\mathrm{s}$ with mean power $p_c$ and the resulting received baseband signal $\mathbf{y}\in\mathbb{C}^{N\times1}$ at $\mathcal{S}$ without any IQI is
\begin{equation}\label{eq:rxS}
\mathbf{y}=\mathbf{h}\,\mathrm{s} +\mathbf{n},
\end{equation}
where $\mathbf{n}\in\mathbb{C}^{N\times1}$ is received additive white Gaussian noise (AWGN) vector with zero mean entries having variance $\sigma_{1}^{2}$.
\subsection{Signal Model for Characterizing IQ Impairments}
We assume that received baseband signal $\mathbf{y}$ in \eqref{eq:rxS} undergoes the joint-TX-RX-IQI. Therefore, the baseband signal $\mathrm{s}$ at $\mathcal{U}$ gets practically altered to $\mathrm{s}_{\mathrm{T}}$, defined below, due to TX-IQI~\cite{schenk2008RF}
\begin{equation}\label{eq:TXIQI}
\mathrm{s}_{\mathrm{T}}=\mathrm{T}_{\mathcal{U}1}\,\mathrm{s}+\mathrm{T}_{\mathcal{U}2}\,\mathrm{s}^*.
\end{equation}
Here, $\mathrm{T}_{\mathcal{U}1}\triangleq\frac{1+g_{{\mathrm T}_{\mathcal{U}}}\mathrm{e}^{j\phi_{{\mathrm T}_{\mathcal{U}}}}}{2}$ and $\mathrm{T}_{\mathcal{U}2}\triangleq\frac{1-g_{{\mathrm T}_{\mathcal{U}}}\mathrm{e}^{j\phi_{{\mathrm T}_{\mathcal{U}}}}}{2}$, with $g_{{\mathrm T}_{\mathcal{U}}}$ and $\phi_{{\mathrm T}_{\mathcal{U}}}$ respectively denoting TX amplitude and phase mismatch at IoT user $\mathcal{U}$.
Similarly, the baseband signal $\mathbf{y}$ received at $\mathcal{S}$ gets practically impaired due to RX-IQI as~\cite{schenk2008RF}
\begin{equation}\label{eq:RXIQI}
\mathbf{y}_{\mathrm{R}}=\mathbf{R}_{\mathcal{S}1}\,\mathbf{y}+\mathbf{R}_{\mathcal{S}2}\,\mathbf{y}^*,
\end{equation}
where $i$th diagonal entry of diagonal matrices $\mathbf{R}_{\mathcal{S}1}$ and $\mathbf{R}_{\mathcal{S}2}$ are $[\mathbf{R}_{\mathcal{S}1}]_i\triangleq\frac{1+g_{{\mathrm R}_{\mathcal{S}i}}\mathrm{e}^{-j\phi_{{\mathrm R}_{\mathcal{S}i}}}}{2}$ and $[\mathbf{R}_{\mathcal{S}2}]_i\triangleq\frac{1-g_{{\mathrm R}_{\mathcal{S}i}}\mathrm{e}^{j\phi_{{\mathrm R}_{\mathcal{S}i}}}}{2}$. Here $g_{{\mathrm R}_{\mathcal{S}i}}$ and $\phi_{{\mathrm R}_{\mathcal{S}i}}$ respectively denote the RX amplitude and phase mismatch at the $i$th antenna of $\mathcal{S}$.
Finally, combining \eqref{eq:TXIQI} and \eqref{eq:RXIQI} in \eqref{eq:rxS}, the baseband signal $\mathbf{y}_{\mathrm{J}}\in\mathbb{C}^{N\times1}$ as received at $\mathcal{S}$ during CE phase under joint-TX-RX-IQI is given by
\begin{equation}\label{eq:JXIQI}
\mathbf{y}_{\mathrm{J}}=\mathbf{h}_{\mathrm A}\,\mathrm{s}+\mathbf{h}_{\mathrm B}\,\mathrm{s}^* +\mathbf{n}_{\mathrm{J}},
\end{equation}
where $\mathbf{h}_{\mathrm A}\triangleq\mathbf{R}_{\mathcal{S}1}\,\mathbf{h}\,\mathrm{T}_{\mathcal{U}1}+\mathbf{R}_{\mathcal{S}2}\,\mathbf{h}^*\,\mathrm{T}_{\mathcal{U}2}^*$, $\mathbf{h}_{\mathrm B}\triangleq\mathbf{R}_{\mathcal{S}1}\,\mathbf{h}\,\mathrm{T}_{\mathcal{U}2}+\mathbf{R}_{\mathcal{S}2}\,\mathbf{h}^*\,\mathrm{T}_{\mathcal{U}1}^*$, and $\mathbf{n}_{\mathrm{J}}\triangleq\mathbf{R}_{\mathcal{S}1}\,\mathbf{n} +\mathbf{R}_{\mathcal{S}2}\,\mathbf{n}^*$. We recall that for addressing the demands of low-rate IoT settings using narrow band signals~\cite{Eriksson-TC17-IQI,Access17-IQI}, we have adopted this frequency-independent-IQI model~\cite{schenk2008RF}. Furthermore, as the IQI parameters change very slowly as compared to the channel estimates, we assume their perfect knowledge availability at $\mathcal{S}$~\cite{IQI-2019-Mag,Virtual-IQI,Eriksson-TC17-IQI,Schober-TWC17-IQI,Access17-IQI,schenk2008RF}. Using this practically-motivated assumption, we optimally exploit the information available in the IQI-based virtual signal term $\mathbf{h}_{\mathrm B}\,\mathrm{s}^*$ for designing the LSE and precoder at $\mathcal{S}$. Moreover, using this IQI-knowledge, our proposed solution methodology can also be applied to the frequency-dependent-IQI scenarios.
\section{Optimal Channel Estimation}\label{sec:OCE}
\subsection{Existing LSE $\widehat{\mathbf{h}}_{\mathrm{A}}$ for IQI-Impaired Channels}\label{sec:existing-LSE}
Current works~\cite{Virtual-IQI,Eriksson-TC17-IQI,Schober-TWC17-IQI,Access17-IQI,IQI-AF-Stat-MaxD,LMMSE-IQI-Close,LS-IQI-ICASSP10,ICASSP18,LSE-Pilot-OFDMA} considered $\mathbf{h}_{\mathrm B} \mathrm{s}^* \!+\mathbf{n}_{\mathrm{J}}$ as the effective noise signal under IQI, and thus, applied conventional pseudo-inverse method~\cite{kay1993fundamentals} on $\mathbf{y}_{\mathrm{J}}$ in \eqref{eq:JXIQI} with $\norm{\mathrm{s}}^2=p_c\tau_c$, to obtain LSE $\widehat{\mathbf{h}}_{\mathrm{A}}$ for the effective channel $\mathbf{h}_{\mathrm A}$ under IQI, defined below
\begin{equation}\label{eq:LS}
\widehat{\mathbf{h}}_{\mathrm{A}}=\mathbf{y}_{\mathrm{J}}\;\mathrm{s}^*\left(\mathrm{s}\,\mathrm{s}^*\right)^{-1}
=\mathbf{h}_{\mathrm A} +\widetilde{\mathbf{h}}_{\mathrm{A}},
\end{equation}
where $\widetilde{\mathbf{h}}_{\mathrm{A}}\triangleq\left(\mathbf{h}_{\mathrm B}\,\mathrm{s}^* +\mathbf{n}_{\mathrm{J}}\right)\mathrm{s}^*\left({p_c\,\tau_c}\right)^{-1}$ is underlying CE error.
\subsection{Proposed LS Approach and Challenges}
As mentioned in Section~\ref{sec:motiv}, we consider both the terms in $\mathbf{y}_{\mathrm{J}}$, i.e., actual $\mathbf{h}_{\mathrm A}\,\mathrm{s}$ and IQI-based virtual $\mathbf{h}_{\mathrm B}\,\mathrm{s}^*$, containing information on $\mathbf{h}$. Therefore, the proposed optimal LSE $\widehat{\mathbf{h}}$ for $\mathbf{h}$ can be obtained by solving the following LS problem in $ \mathbf{h}$,
\begin{align}\nonumber
\mathcal{O}_1:\, \underset{ \mathbf{h}}{\text{argmin}}\quad\mathcal{E}\triangleq
\norm{\mathbf{y}_{\mathrm{J}}-\mathbf{h}_{\mathrm A}\,\mathrm{s}-\mathbf{h}_{\mathrm B}\,\mathrm{s}^*}^2.
\end{align}
Although $\mathcal{O}_{\mathrm 1}$ is nonconvex due to the presence of $\mathbf{h}^*$ terms in $\mathbf{h}_{\mathrm A}$ and $\mathbf{h}_{\mathrm B}$, we can characterize all the possible candidates for the optimal solution of $\mathcal{O}_{\mathrm 1}$ by setting derivative of objective $\mathcal{E}$ to zero and then solve in $\mathbf{h}$. Below, we first simplify $\mathcal{E}$ as
\begin{align}\label{eq:tr-obj}
\mathcal{E}\!=&\,\mathbf{y}_{\mathrm{J}}^{\rm H}\mathbf{y}_{\mathrm{J}}- \mathbf{y}_{\mathrm{J}}^{\rm H}\mathbf{A}\mathbf{h}\!-\!\mathbf{y}_{\mathrm{J}}^{\rm H} \mathbf{B}\mathbf{h}^*\!+ \mathbf{h}^{\rm H}\mathbf{A}^{\rm H}\mathbf{A}\mathbf{h}- \mathbf{h}^{\rm H}\mathbf{A}^{\rm H}\mathbf{y}_{\mathrm{J}}+ \nonumber\\
&\; \mathbf{h}^{\rm H}\mathbf{A}^{\rm H}\mathbf{B}\mathbf{h}^*\!-\!\mathbf{h}^{\rm T}\mathbf{B}^{\rm H}\mathbf{y}_{\mathrm{J}}\!+\!\mathbf{h}^{\rm T}\mathbf{B}^{\rm H}\mathbf{A}\mathbf{h}\!+\! \mathbf{h}^{\rm T}\mathbf{B}^{\rm H}\mathbf{B}\mathbf{h}^*,
\end{align}
where $\mathbf{A}\triangleq\mathbf{R}_{\mathcal{S}1}\,\mathrm{T}_{\mathcal{U}1}\,\mathrm{s}+\mathbf{R}_{\mathcal{S}1}\,\mathrm{T}_{\mathcal{U}2}\,\mathrm{s}^*$ and $\mathbf{B}\triangleq\mathbf{R}_{\mathcal{S}2}\,\mathrm{T}_{\mathcal{U}2}^*\,\mathrm{s}+\mathbf{R}_{\mathcal{S}2}\,\mathrm{T}_{\mathcal{U}1}^*\,\mathrm{s}^*$ are diagonal matrices. Using complex-valued differentiation rules~\cite{complex-matrixbook} to the find derivative of \eqref{eq:tr-obj} with respect to $\mathbf{h}$ and setting resultant to zero, gives the following system
\begin{equation}\label{eq:der-tr-obj}
\mathbf{h}^{\rm H}\,\mathbf{Z}_{\rm A} +\mathbf{h}^{\rm T}\,\mathbf{Z}_{\rm B}=\mathbf{y}_{\mathrm{AB}}^{\rm T},
\end{equation}
where $\mathbf{y}_{\mathrm{AB}}\!\!\triangleq\!\mathbf{A}^{\rm T}\mathbf{y}_{\mathrm{J}}^*+\mathbf{B}^{\rm H}\mathbf{y}_{\mathrm{J}}\!\in\!\mathbb{C}^{N\times 1}\!$ and $\mathbf{Z}_{\rm B}\!\triangleq\!\mathbf{B}^{\rm H}\mathbf{A}\!+\!\mathbf{A}^{\rm T}\mathbf{B}^*\!\in\!\mathbb{C}^{N\times N}$. Here, $\mathbf{Z}_{\rm A}\triangleq\mathbf{A}^{\rm H}\mathbf{A}+ \mathbf{B}^{\rm T}\mathbf{B}^*\!\in\mathbb{R}^{N\times N}$ is a diagonal matrix with $\left[\mathbf{Z}_{\rm A}\right]_i=\left|\left[\mathbf{A}\right]_i\right|^2+\left|\left[\mathbf{B}\right]_i\right|^2,\forall i\in\mathcal{N}=\{1,2,\ldots,N\}.$
Though, we have been able to reduce the LS problem $\mathcal{O}_1$ of obtaining optimal LSE $\widehat{\mathbf{h}}$ for IQI-impaired channels to the nonlinear system of equations \eqref{eq:der-tr-obj} in the complex variable $\mathbf{h},$ solving the latter numerically is computationally-expensive and time-consuming, especially for $N\!\gg\!1$. Therefore, next we propose an \textit{equivalent} complex-to real transformation for \textit{efficiently} obtaining unique globally-optimal solution $\widehat{\mathbf{h}}$ of $\mathcal{O}_1$.
\subsection{Closed-Form Expression for Globally-Optimal LSE $\widehat{\mathbf{h}}$}\label{sec:CF-LSE}
Before deriving $\widehat{\mathbf{h}}$, let us define some key notations below.
\begin{definition}
We can define the real composite representations for any complex vector $\mathbf{u}\in\mathbb{C}^{n\times1}$ by $\underline{\mathbf{u}}\in\mathbb{R}^{2n\times 1}$ and for any complex matrix $\mathbf{U}\!\in\!\mathbb{C}^{n_1\times n_2}$ by $\underline{\mathbf{U}}\!\in\!\mathbb{R}^{2n_1\times2n_2}$ as below
\begin{eqnarray}
\underline{\mathbf{u}}\triangleq\left[\!\!\begin{array}{cc}
\mathrm{Re}\{\mathbf{u}\} \\
\mathrm{Im}\{\mathbf{u}\} \end{array}\!\!\right],\qquad\quad
\underline{\mathbf{U}}\triangleq\left[\!\!\begin{array}{ccc}
\mathrm{Re}\{\mathbf{U}\} &\;\, -\mathrm{Im}\{\mathbf{U}\} \\
\mathrm{Im}\{\mathbf{U}\}&\;\quad
\mathrm{Re}\{\mathbf{U}\} \end{array}\!\!\right].
\end{eqnarray}
\end{definition}
Using above definition, \eqref{eq:der-tr-obj} can be rewritten in real-domain as
\begin{equation}\label{eq:OCE-0}
\left[\!\!\begin{array}{ccc}
\mathbf{Z}_{\rm A} &\;\, \mathbf{0}_{N\times N} \\
\mathbf{0}_{N\times N} &\;\, -\mathbf{Z}_{\rm A} \end{array}\!\!\right]\,\underline{\mathbf{h}} +\underline{\mathbf{Z}_{\rm B}}\,\underline{\mathbf{h}} =\underline{\mathbf{y}_{\mathrm{AB}}}.
\end{equation}
Recalling $\mathbf{Z}_{\rm A}$ is real while solving \eqref{eq:OCE-0}, the real and imaginary terms of the proposed LSE $\widehat{\mathbf{h}}$ can be analytically expressed as
\begin{eqnarray}\label{eq:OCE}
\left[\!\!\!\begin{array}{cc}
\mathrm{Re}\{\widehat{\mathbf{h}}\}\\ \mathrm{Im}\{\widehat{\mathbf{h}}\}\end{array} \!\!\!\right]\!\triangleq\!\left[\!\!\!\begin{array}{ccc}
\mathbf{Z}_{\rm A}+\mathrm{Re}\{\mathbf{Z}_{\rm B}\} & -\mathrm{Im}\{\mathbf{Z}_{\rm B}\} \\
\mathrm{Im}\{\mathbf{Z}_{\rm B}\}
& -\mathbf{Z}_{\rm A}+\mathrm{Re}\{\mathbf{Z}_{\rm B}\} \end{array}\!\!\!\right]^{-1}\! \underline{\mathbf{y}_{\mathrm{AB}}}\,.\!
\end{eqnarray}
\section{Optimal Transmit Beamforming Design}\label{sec:OTB}
After optimizing LSE using CE phase, now we optimize the efficiency of IT (phase $2$) over IQI-impaired DL channel. Metric to be maximized here by optimally designing precoder $\mathbf{x}\in\mathbb{C}^{N\times 1}$ at $\mathcal{S}$ is the \textit{signal power at $\mathcal{U}$} during IT phase.
\subsection{Conventional Precoder Design}
With $\mathrm{s_{\mathcal{U}}}$ being unit-energy data symbol, the signal received at $\mathcal{U}$ due to IT, under perfect CSI and no IQI assumption, is
\begin{equation}\label{eq:DL-ET}
\mathrm{y_{\mathcal{U}}}= \mathbf{h}^{\mathrm T} \,\mathbf{x}\,\mathrm{s_{\mathcal{U}}}+\mathrm{n_{\mathcal{U}}},
\end{equation}
where precoder $\mathbf{x}$ satisfies $\left\lVert\mathbf{x}\right\rVert^2\!\!\le\! p_i,$ with $p_i$ being the transmit power of $\mathcal{S}$ and $\mathrm{n_{\mathcal{U}}}\sim\mathbb{C} \mathbb{N}\left(0,\sigma_2^{2}\right)$ is the received AWGN at $\mathcal{U}$. Like in case of CE, the existing works~\cite{Virtual-IQI,Eriksson-TC17-IQI,Schober-TWC17-IQI,Access17-IQI,IQI-AF-Stat-MaxD,LMMSE-IQI-Close,LS-IQI-ICASSP10,ICASSP18,LSE-Pilot-OFDMA} ignored the virtual term and designed the precoder as in conventional systems to perform MRT at $\mathcal{S}$ in the DL. Therefore, using the conventional LSE $\widehat{\mathbf{h}}_{\rm A}$ as defined in Section~\ref{sec:existing-LSE}, the benchmark precoder following MRT is given by $\mathbf{x}_{\rm A}\triangleq\frac{\sqrt{p_i}\;\widehat{\mathbf{h}}_{\rm A}^*}{\norm{\widehat{\mathbf{h}}_{\rm A}^*}}.$
\subsection{Maximizing Received Signal Strength under IQI}
Under joint-TX-RX-IQI, $\mathrm{y_{\mathcal{U}}}$ gets practically impaired to
\begin{align}\label{eq:Rx-signal}
\mathrm{y}_{\mathcal{U}\mathrm{J}}
=\mathrm{R}_{\mathcal{U}1}\,\mathrm{y}_{\mathcal{U}{\rm T}}+\mathrm{R}_{\mathcal{S}2}\,\mathrm{y}_{\mathcal{U}{\rm T}}^*
=\mathbf{a}\,\mathbf{x}+\mathbf{b}\,\mathbf{x}^*+\mathrm{n}_{\mathcal{U}{\rm J}},
\end{align}
where complex vectors $\mathbf{a},\mathbf{b},$ and $\mathrm{n}_{\mathcal{U}{\rm J}}$ are defined below
\begin{subequations}
\begin{gather}
\mathbf{a}\triangleq\left(\mathrm{R}_{\mathcal{U}1}\,\mathbf{h}^{\mathrm T}\,\mathbf{T}_{\mathcal{S}1}+\mathrm{R}_{\mathcal{U}2}\,\mathbf{h}^{\mathrm H}\,\mathbf{T}_{\mathcal{S}2}^*\right)\mathrm{s_{\mathcal{U}}}\in\mathbb{C}^{1\times N},\label{eq:a}\\
\mathbf{b}\triangleq\left(\mathrm{R}_{\mathcal{U}1}\,\mathbf{h}^{\mathrm T}\,\mathbf{T}_{\mathcal{S}2}+\mathrm{R}_{\mathcal{U}2}\,\mathbf{h}^{\mathrm H}\,\mathbf{T}_{\mathcal{S}1}^*\right)\mathrm{s}_{\mathcal{U}}^*\in\mathbb{C}^{1\times N},\label{eq:b}\\
\mathrm{n}_{\mathcal{U}{\rm J}}\triangleq\mathrm{R}_{\mathcal{U}1}\,\mathrm{n}_{\mathcal{U}}+\mathrm{R}_{\mathcal{S}2}\,\mathrm{n}_{\mathcal{U}}^*\sim\mathbb{C} \mathbb{N}\left(0,\sigma_{\rm J}^2\right).
\end{gather}
\end{subequations}
Here with $g_{{\mathrm R}_{\mathcal{U}}}$ and $\phi_{{\mathrm R}_{\mathcal{U}}}$ respectively denoting RX amplitude and phase mismatch at $\mathcal{U}$, $\mathrm{R}_{\mathcal{U}1}\triangleq\frac{1+g_{{\mathrm R}_{\mathcal{U}}}\mathrm{e}^{-j\phi_{{\mathrm R}_{\mathcal{U}}}}}{2}$ and $\mathrm{R}_{\mathcal{U}2}\triangleq\frac{1-g_{{\mathrm R}_{\mathcal{U}}}\mathrm{e}^{j\phi_{{\mathrm R}_{\mathcal{U}}}}}{2}$.
Therefore, $\sigma_{\rm J}^2\triangleq\left(\left|\mathrm{R}_{\mathcal{U}1}\right|^2+\left|\mathrm{R}_{\mathcal{U}2}\right|^2\right)\sigma_2^2$. $\mathrm{y}_{\mathcal{U}{\rm T}}\triangleq\mathbf{h}^{\mathrm T}\left(\mathbf{T}_{\mathcal{S}1}\,\mathbf{x}\,\mathrm{s_{\mathcal{U}}}+\mathbf{T}_{\mathcal{S}2}\,\mathbf{x}^*\mathrm{s_{\mathcal{U}}}^*\right)+\mathrm{n}_{\mathcal{U}}$ is TX-IQI impaired signal, where $\mathbf{T}_{\mathcal{S}1}$ and $\mathbf{T}_{\mathcal{S}2}$ represent diagonal matrices with $g_{{\mathrm T}_{\mathcal{S}i}}$ and $\phi_{{\mathrm T}_{\mathcal{S}i}}$ in their $i$th diagonal entries $[\mathbf{T}_{\mathcal{S}1}]_i\triangleq \frac{1+g_{{\mathrm T}_{\mathcal{S}i}}\mathrm{e}^{j\phi_{{\mathrm T}_{\mathcal{S}i}}}}{2}$ and $[\mathbf{T}_{\mathcal{S}2}]_i\triangleq \frac{1-g_{{\mathrm T}_{\mathcal{S}i}}\mathrm{e}^{j\phi_{{\mathrm T}_{\mathcal{S}i}}}}{2}$ respectively denoting TX amplitude and phase mismatch at $i$th antenna of $\mathcal{S}$ during the IT phase.
Noting that the received signal has two useful terms $\mathbf{a}\,\mathbf{x}$ and $\mathbf{b}\,\mathbf{x}^*$ in \eqref{eq:Rx-signal}, the proposed precoder optimization problem for maximizing the signal power at $\mathcal{U}$ is formulated as below
\begin{equation*}
\mathcal{O}_2\!:\, \underset{\mathbf{x}}{\text{argmax}}\;\left\lVert\mathbf{a}\,\mathbf{x}+\mathbf{b}\,\mathbf{x}^*\right\rVert^2,\quad\text{subject to}\;\mathrm{(C1)}:\left\lVert\mathbf{x}\right\rVert^2\le p_i.
\end{equation*}
The challenges here include non-convexity of $\mathcal{O}_2$ and need for fast-converging or closed-form globally-optimal design to obtain the desired solution $\mathbf{x}_{\rm o}$ in a computationally-efficient manner. Furthermore, this signal power as objective is actually closely-related to other \textit{key metrics} like ergodic capacity and detection error probability~\cite{Opt-R3} because former's higher value also implies better ergodic capacity or lower error probability.
\begin{figure}[!t]
\centering\includegraphics[width=3.48in]{MSE}
\caption{ Validating LSE $\widehat{\mathbf{h}}$ against benchmark for different SNR and IQI values.}
\label{fig:MSE}
\end{figure}
\begin{figure*}[!t]
\centering\includegraphics[width=6.7in]{All_in_One}\vspace{-2mm}
\caption{Comparing relative performance of proposed optimal LSE, precoder, and jointly-optimal designs against benchmark for different $\overline{\gamma}_{\rm E},N,\tau_c,\Delta$ values.}\vspace{-1mm}
\label{fig:All}
\end{figure*}
\subsection{Novel Globally-Optimal Precoder}\label{sec:optP}
Though $\mathcal{O}_2$ is nonconvex, its globally-optimal solution can be characterized via Karush-Kuhn-Tucker (KKT) point~\cite{Baz}. To obtain latter, below we define Lagrangian function for $\mathcal{O}_2$
\begin{align}
\mathcal{L}\!\triangleq&\left\lVert\mathbf{a}\,\mathbf{x}+\mathbf{b}\,\mathbf{x}^*\right\rVert^2-\ell\,\big(\!\left\lVert\mathbf{x}\right\rVert^2-p_i\big)= \mathbf{x}^{\rm H}\mathbf{a}^{\rm H}\mathbf{a}\,\mathbf{x}+ \mathbf{x}^{\rm H}\mathbf{a}^{\rm H}\mathbf{b}\,\mathbf{x}^*\nonumber\\
&\qquad\; +\mathbf{x}^{\rm T}\,\mathbf{b}^{\rm H}\mathbf{a}\,\mathbf{x} + \mathbf{x}^{\rm T}\mathbf{b}^{\rm H}\,\mathbf{b}\,\mathbf{x}^*-\ell\,\big(\mathbf{x}^{\rm H}\mathbf{x}-p_i\big),
\end{align}
where $\ell\ge0$ is the Lagrange multiplier corresponding to $\mathrm{(C1)}$.
\begin{eqnarray}\label{eq:partial}
\frac{\partial\mathcal{L}}{\partial\mathbf{x}}=\mathbf{x}^{\rm H}\mathbf{a}^{\rm H}\mathbf{a} +\mathbf{x}^{\rm T}\left(\mathbf{b}^{\rm H}\mathbf{a}+\mathbf{a}^{\rm T}\mathbf{b}^*\right)+ \mathbf{x}^{\rm T}\mathbf{b}^{\rm H}\,\mathbf{b}-\ell\, \mathbf{x}^{\rm H}.
\end{eqnarray}
Setting $\frac{\partial\mathcal{L}}{\partial\mathbf{x}}$ in \eqref{eq:partial} to $\mathbf{0}_{1\times N}$, yields the KKT condition below
\begin{align}\label{eq:KKT1}
\mathbf{x}^{\rm T}\,\mathbf{Z}_{\mathbf a}^*+\mathbf{x}^{\rm H}\,\mathbf{Z}_{\mathbf b}^*=\ell\, \mathbf{x}^{\rm T},
\end{align}
where $\mathbf{Z_a}\triangleq\mathbf{a}^{\rm H}\mathbf{a}+\mathbf{b}^{\rm T}\mathbf{b}^*$ and $\mathbf{Z_b}\triangleq\mathbf{b}^{\rm H}\mathbf{a}+\mathbf{a}^{\rm T}\mathbf{b}^*$. Using the composite real definition from Section~\ref{sec:CF-LSE} in \eqref{eq:KKT1}, we obtain
\begin{eqnarray}\label{eq:Eig}
\left(\underline{\mathbf{Z}_{\mathbf{a}}}\right)^*\underline{\mathbf{x}} +\left(\underline{\mathbf{Z}_{\mathbf{b}}}\right)^{\rm H}\left[\!\!\begin{array}{cc}
\;\;\,\mathrm{Re}\{\mathbf{x}\} \\
-\mathrm{Im}\{\mathbf{x}\} \end{array}\!\!\right]= \ell\,\underline{\mathbf{x}},\hspace{2mm}\text{or}\hspace{2mm}\mathbf{Z}_{\mathbf{ab}}\,\underline{\mathbf{x}} = \ell\,\underline{\mathbf{x}},
\end{eqnarray}
{where the real square matrix $\mathbf{Z}_{\mathbf{ab}}\in\mathbb{R}^{2N\times2N}$ is defined as}
\begin{eqnarray}\label{eq:Zab}
\mathbf{Z}_{\mathbf{ab}}\triangleq\left[\!\!\begin{array}{ccc}
\hspace{3mm}\mathrm{Re}\{{\mathbf{Z}_{\mathbf{a}}}\}+\mathrm{Re}\{{\mathbf{Z}_{\mathbf{b}}}\}&\; \mathrm{Im}\{{\mathbf{Z}_{\mathbf{a}}}\}+\mathrm{Im}\{{\mathbf{Z}_{\mathbf{b}}}\}\\
-\mathrm{Im}\{{\mathbf{Z}_{\mathbf{a}}}\}+\mathrm{Im}\{{\mathbf{Z}_{\mathbf{b}}}\}&\;
\mathrm{Re}\{{\mathbf{Z}_{\mathbf{a}}}\}-\mathrm{Re}\{{\mathbf{Z}_{\mathbf{b}}}\} \end{array}\!\!\right].
\end{eqnarray}
As \eqref{eq:Eig} possesses an eigenvalue problem form, the solution to \eqref{eq:Eig} in $\underline{\mathbf{x}}$ is given by the principal eigenvector $\mathrm{v}_{\max}\left\lbrace\mathbf{Z}_{\mathbf{ab}}\right\rbrace$ corresponding to the maximum eigenvalue $\lambda_{\max}\left\lbrace\mathbf{Z}_{\mathbf{ab}}\right\rbrace$ of $\mathbf{Z}_{\mathbf{ab}}$. Therefore, the globally-maximum signal power is attained at the proposed precoder $\mathbf{x}_{\rm o}\triangleq\mathrm{Re}\{\mathbf{x}_{\rm o}\}+j\,\mathrm{Im}\{\mathbf{x}_{\rm o}\}$, whose real and imaginary parts, obtained via eigen-decomposition are
\begin{eqnarray}\label{eq:optx}
\underline{\mathbf{x}_{\rm o}}=\left[\begin{array}{cc}
\mathrm{Re}\{\mathbf{x}_{\rm o}\} \\
\mathrm{Im}\{\mathbf{x}_{\rm o}\} \end{array}\right]\triangleq\sqrt{p_i}\;\frac{\mathrm{v}_{\max}\left\lbrace\mathbf{Z}_{\mathbf{ab}}\right\rbrace}{\left\lVert\mathrm{v}_{\max}\left\lbrace\mathbf{Z}_{\mathbf{ab}}\right\rbrace\right\rVert}\in\mathbb{R}^{2N\times1}.
\end{eqnarray}
\subsection{Extending Precoder Design to Multiuser Settings}\label{sec:extn}
For maximizing the sum received power among $K$ single-antenna users, the precoder optimization problem is given by
\begin{equation*}
\mathcal{O}_3\!:\, \underset{\mathbf{x}}{\text{argmax}}\;\left\lVert\boldsymbol{\mathcal{A}}\,\mathbf{x}+\boldsymbol{\mathcal{B}}\,\mathbf{x}^*\right\rVert^2,\qquad\text{subject to}\quad\mathrm{(C1)},
\end{equation*}
where the $K\times N$ matrices $\boldsymbol{\mathcal{A}}$ and $\boldsymbol{\mathcal{B}}$ are respectively obtained from $\mathbf{a}$ and $\mathbf{b}$ in \eqref{eq:a} and \eqref{eq:b}, but with $\mathrm{s_{\mathcal{U}}}$ replaced by unit-energy vector $\mathbf{s_{\mathcal{U}}}$, $\mathbf{h}$ replaced with $N\times K$ matrix $\mathbf{H}$ whose $i$th column corresponds to channel gain for $\mathcal{S}$ to $i$th user link, and the $K\times K$ diagonal matrices $\mathbf{R}_{\mathcal{U}1}$ and $\mathbf{R}_{\mathcal{U}2}$, respectively replacing $\mathrm{R}_{\mathcal{U}1}$ and $\mathrm{R}_{\mathcal{U}2}$. Here, $i$th diagonal entries of $\mathbf{R}_{\mathcal{U}1}$ and $\mathbf{R}_{\mathcal{U}2}$ incorporate the RX amplitude and phase mismatch at $i$th user. So, following Section~\ref{sec:optP}, the optimal precoder for $\mathcal{O}_3$ is given by \eqref{eq:optx}, but with $\boldsymbol{\mathcal{A}}$ and $\boldsymbol{\mathcal{B}}$ respectively replacing $\mathbf{a}$ and $\mathbf{b}$ in $\mathbf{Z}_{\mathbf{ab}}$ definition. The accuracy of this TX design in multiuser setting can be verified from the fact that for no IQI, $\mathbf{Z}_{\mathbf{ab}}$ reduces to $\underline{\mathbf{H}\,\mathbf{H}^{\rm H}}$ with result matching \cite[Theorem 1]{EBF-TWC14}.\color{black}
\section{Performance Evaluation and Conclusion}\label{sec:results}
Here we numerically validate the proposed CE analysis and precoder optimization while setting simulation parameters as $N=10$, $\tau=10$ms, $\tau_c=0.01\tau$, $p_i=30$dBm, $p_c=-30$dBm, $\sigma_{1}^{2}=\sigma_{2}^{2}=10^{-17}$ Joule, and $\beta=\frac{\varpi}{d^{\varrho}}$, where $\varpi=\left(\frac{3\times 10^8}{4\pi f}\right)^2$ is average channel attenuation at unit reference distance with $f=915$MHz as TX frequency, $d=100$m as $\mathcal{S}$-to-$\mathcal{U}$ distance, and $\varrho=2.5$ as path loss exponent. For the average simulation results, we have used $10^5$ independent channel realizations.
\subsection{Validation of Proposed LSE under Practical IQI Modelling}\label{sec:sim-valid}
We start with verifying the quality of proposed LSE $\widehat{\mathbf{h}}$ (cf. \eqref{eq:OCE}) in Fig.~\ref{fig:MSE} against the benchmark $\widehat{\mathbf{h}}_{\rm A}$ as defined in \eqref{eq:LS}. For IQI incorporation, we adopt the following practical model~\cite{TCOM-HEB}
\begin{eqnarray}\label{eq:g-p}
g\triangleq1-\Delta_{g}\left(1+\Psi_{g}\right),\qquad\quad\phi\triangleq \Delta_{\phi}\left(1+\Psi_{\phi}\right),
\end{eqnarray}
where $g$ and $\phi$ respectively can incorporate any amplitude and phase mismatch, with the constants $\Delta_{g}$ and $\Delta_{\phi}$ representing the errors due to fixed sources. Whereas, $\Psi_{g}$ and $\Psi_{\phi}$, respectively denoting errors due to random sources, are assumed to follow the uniform distribution~\cite{IQI-AF-Stat-MaxD,LMMSE-IQI-Close} over the interval $\left[-\frac{1}{2}\Phi_{g},\frac{1}{2}\Phi_{g}\right]$ and $\left[-\frac{1}{2}\Phi_{\phi},\frac{1}{2}\Phi_{\phi}\right]$, respectively. {Since the practical ranges for the constants $\left(\Delta_g,\,\Delta_{\phi},\,\Phi_g,\,\Phi_{\phi}\right)$ corresponding to the means and variances of amplitude and phase errors (in radians) are similar~\cite{TWC-IQI,SPL-HEB}, we set $\Delta_{g}=\Delta_{\phi}=\Phi_{g}=\Phi_{\phi}=\Delta=0.4$, for each of the $8$ IQI parameters.}
Results plotted in Fig.~\ref{fig:MSE} show the trend in mean square error (MSE)~\cite{massive-MIMO} between the actual channel $\mathbf{h}$ and its LSE (proposed $\widehat{\mathbf{h}}$ and benchmark $\widehat{\mathbf{h}}_{\rm A}$) against increasing average received signal-to-noise-ratio (SNR) $\overline{\gamma}_{\rm E}=\frac{\beta\,p_c\tau_c}{\sigma_1^2}$ at $\mathcal{S}$ during CE phase. The quality of both proposed and existing LSE improve with increasing $\overline{\gamma}_{\rm E}$ because the underlying CE errors reduce for both considered values of IQI degradation parameter $\Delta$. However, for the benchmark LSE, the error floor region starts at $\overline{\gamma}_{\rm E}=20$dB and $\overline{\gamma}_{\rm E}=30$dB for $\Delta=0.4$ and $\Delta=0.1$, respectively. Whereas, MSE for the proposed globally-optimal LSE for the IQI-impaired channel keeps on decreasing at the same rate \textit{without} having any error floor. This corroborates the \textit{significantly-higher} practical utility of our proposed LSE $\widehat{\mathbf{h}}$ for the IQI-influenced MISO communications, in terms of our CE design providing about $-3$dB and $-11$dB improvement in MSE over benchmark for $\Delta=0.1$ and $\Delta=0.4$, respectively.
\subsection{Comparison of Proposed Designs Against Benchmark}\label{sec:comp}
{Here we compare the mean signal power performance of the three proposed schemes: (i) \textit{jointly-optimal} LSE $\widehat{\mathbf{h}}$ and precoder $\mathbf{x}_{\rm o}$, (ii) \textit{optimal precoder} $\mathbf{x}_{\rm o}$ with conventional LSE $\widehat{\mathbf{h}}_{\rm A}$, (iii) \textit{optimal LSE} $\widehat{\mathbf{h}}$ with MRT-based precoder $\mathbf{x}_{\rm A}$, against the \textit{benchmark} having LSE $\widehat{\mathbf{h}}_{\rm A}$ and precoder $\mathbf{x}_{\rm A}$.} Starting with comparison for different $\overline{\gamma}_{\rm E}$ in Fig.~\ref{fig:All}(a), we notice that jointly-optimal performs the \textit{best}, followed by optimal precoder and proposed LSE. The gaps between the optimal and benchmark designs increase with $\overline{\gamma}_{\rm E}$ due to lower errors at higher SNRs.
Next in Fig.~\ref{fig:All}(b), we plot the comparison for different array sizes $N$ at $\mathcal{S}$. Here, with $N$ increased from $4$ to $20$, mean signal power at $\mathcal{U}$ gets enhanced by $7$dB for each of the four schemes. However, their relative gap remains \textit{invariant} of $N$.
Now, shifting focus to CE time $\tau_c$, we shed insights on how to optimally set it. From Fig.~\ref{fig:All}(c), we notice that the relative trend among four schemes is similar, but more importantly, the optimal $\tau_c$ for each scheme is \textit{practically the same} ($\approx 10^{-3}\tau$).
Next we investigate the impact of increased mismatch $\Delta$ in the amplitude and phase terms modelling the IQI. In particular, by plotting the variation of $\Delta$ from $0$ to $0.5$~\cite{IQI-AF-Stat-MaxD,LMMSE-IQI-Close,TCOM-HEB} in Fig.~\ref{fig:All}(d), we observe that degradation in the mean signal power performance gets enhanced with increased IQI (i.e., $\Delta$) for each scheme. However, this performance degradation for jointly-optimal, optimal precoder, optimal LSE, and benchmark schemes when parameter $\Delta$ increases from $0$ (no IQI) to $0.5$ is $-1.6$dB,$-1.9$dB,$-2.7$dB, and $-2.8$dB, respectively.
\begin{figure}[!t]
\centering\includegraphics[width=3.4in]{comp}\vspace{-2mm}
\caption{Average performance gains of our proposed designs over benchmark.}
\label{fig:comp}
\end{figure}
Lastly, in Fig.~\ref{fig:comp}, we have plotted the average performance gains as achieved by the proposed LSE $\widehat{\mathbf{h}}$, precoder $\mathbf{x}_{\rm o}$, and the jointly-optimal design over benchmark for different values of critical parameters $\overline{\gamma}_{\rm E},N,\tau_c,$ and $\Delta$. {We observe that jointly-optimal design provides an \textit{overall improvement of $24$\%}.} Here, optimal precoder, providing about $18$\% enhancement alone in mean signal power at $\mathcal{U}$, proved to be a \textit{better semi-adaptive scheme} than optimizing LSE, which yields $6\%$ improvement.
\subsection{Concluding Remarks}\label{sec:conclusion}
This letter exploiting the additional channel gain information in the signal received during IQI-impaired MISO communication, came up with a \textit{novel LSE} that is shown to reduce the overall MSE in CE by $-8$dB, while totally \textit{removing the error floor}. {To maximize the practical EB gains in both single and multiple user set-ups, we derive new \textit{globally-optimal precoder} in the form of \textit{principal eigenvector} of the matrix composed of IQI parameters and LSE.} Numerical results have shown that the proposed jointly-optimal LSE and precoder design can provide an overall improvement of $24\%$ over the benchmark. This corroborates the fact that our proposed design is the \textit{way-forward} to maximize practical utility of low-cost hardware in multiantenna transmission supported sustainable IoT systems.
|
\section{Introduction}
Supernova remnants (SNRs) are key objects to study the chemical enrichment of the interstellar medium (ISM) and stellar evolution, while SNRs produced by core-collapse supernovae are tightly linked to star-formation sites in galaxies. Supernova explosions represent the tragic end of the life of many types of stars, while they interact with the ISM by injecting enriched material and energy into their environment, often being at the origin of superwinds expanding far away from the galaxies \citep{doi:10.1146/annurev.aa.15.090177.001135}. Depending on the shock velocity and the density of the surrounding ISM, the high-energy shock waves resulting from SN explosions are proposed to either trigger or repress star formation (e.g. \citep{2009MNRAS.399.2183N, 2006MNRAS.373..811M, 2013NatSR....E1738B}.
The strength of their radiation and the energy dissipated by the shocks are such that a vast bubble of ionized gas is generated and the expansion shell is particularly bright. For instance, they are a non-negligeable contributor to the observed H$\alpha$ emission line in galaxies, which can be easily mixed with the emission from H\,{\sevensize II}\ regions often used in more distant objects to measure the star formation rate (SFR). Nevertheless, the exact contribution from SNRs to the H$\alpha$ flux is not well known and could vary from one galaxy to another,
e.g. \cite{2015MNRAS.446..943V} found a flux contribution of 5 $\pm$ 5 \%.
The morphology of SNRs depends on different factors like their evolutionary phase, the input energy released by the explosion, the mass loss history of the progenitor, and the density of the surrounding ISM. The SNR evolution can be summarized by a four-stage model \citep{1972ARA&A..10..129W}. During the first stage, called the free expansion phase, the supernova ejecta sweeps up ISM material as it expands freely until its mass becomes equal to the mass of the swept up material. Theoretically, this phase can last between 90 and 300 years and is characterized by constant temperature and a constant expansion velocity of the shell. The second evolutionary stage, called the adiabatic phase or Sedov-Taylor phase, is described by the Sedov-Taylor self-similar solution \citep{1959sdmm.book.....S, 1950RSPSA.201..159T}. In this phase, the quantity of the material swept up from the circumstellar and interstellar medium becomes so important that the expanding shell starts to slow down and cool. In general, the estimated duration of this phase is between 100 and 10\,000 years but it can also be missed completely in some cases. However, if the expansion is happening in a hot interstellar medium, it will last longer and the SNR diameter may reach 180\,pc \citep{1977ApJ...218..148M}. During the third stage, called the radiative phase or the snowplow phase, the mass of the swept up material is dramatically increased, which forces the velocity of the shock front to decrease down to $\sim$200\,km\,s$^{-1}$. The temperature behind the shock front drops to $\sim$10$^{5}$\,K and the energy lost due to recombination becomes significant, creating a cooling region behind the shock front and producing shock-heated collisionally ionized species (such as [S\,{\sevensize II}] and [O\,{\sevensize III}]) along with hydrogen recombination lines. In this phase, the SNR diameter expands from about 14\,pc to 50\,pc in a typical ISM with a density of 1.0\,cm$^{-3}$ \citep{1990ASIC..305....1C}. During this phase, which can last up to 20\,000 years, the SNR becomes visible in the optical band. The last evolutionary stage is a dissipative process during which the velocity of the shock reaches the sound speed of the ambient ISM and the SNR is dispersed. The size of the SNR during this phase is from 50\,pc to over 100\,pc.
In the optical, the remnant diameter can vary from a few parsecs up to more than a 100\,pc \citep{2009SerAJ.179...55C}. With larger diameters, superbubbles are another type of emission region with shocks related emission lines, which may involve massive stars and/or SN explosions.
Superbubbles have typical velocity shock below 100\,km\,s$^{-1}$ \citep{2015yCat..74460943V}.
Because of their proximity, SNRs observed in the Milky Way allow a detailed study of their physical properties and interactions with the surrounding ISM. According to the latest version of the Galactic SNR catalogs \citep{2017yCat.7278....0G,2012AdSpR..49.1313F}\footnote{http://www.physics.umanitoba.ca/snr/SNRcat}, our galaxy is hosting at least 294~remnants. But these catalogs remain limited because of the high absorption in the Galactic plane and the large uncertainty associated with the distance measurement. On the other hand, these constraints are often less important in the case of extragalactic SNRs in nearly face-on galaxies.
Extragalactic SNRs are very important to understand the physics of the SNRs, as individual objects and as a component of the ISM and galaxies. First, a distance measurement (and SNR flux and size) is less of a burden as it may be given by other indicators for the galaxy. Second, the advancement of astronomical instruments allows us to observe simultaneously complete samples of SNRs which help to take into account observational selection effects. Historically, the first extragalactic SNRs were identified in the Magellanic Clouds using radio and optical observations \citep{1963Natur.199..681M, 1973ApJ...180..725M}. Today, 25~nearby galaxies are known to host almost 1200~extragalactic SNRs \citep{2015MNRAS.446..943V}. It is interesting to highlight that with the exception of SNRs in the SMC and LMC, almost all the extragalactic SNRs were first identified in the optical. The number of observed SNRs in a galaxy, from the optical, varies from a small number to a few hundred. 25~SNRs have been identified in the SMC \citep{2012A&A...545A.128H}, 26 in NGC\,6946 \citep{1997ApJS..112...49M}, 53 in the LMC \citep{2016A&A...585A.162M}, 93 in M101 \citep{2012AJ....143...85F}, 217 in M33 \citep{2010ApJS..187..495L, 2018ApJ...855..140L}, and 296 in M83 \citep{2014ApJ...788...55B}. \cite{2013MNRAS.429..189L} detected more than 400 SNRs in six nearby galaxies (NGC\,2403, NGC\,3077, NGC\,4214, NGC\,4395, NGC\,4449, and NGC\,5204) based on optical photometric and spectroscopic observations.
The identification and confirmation of extragalactic SNRs are mainly done using data from the radio \citep{1997ApJS..109..417L, 2001ApJ...560..719L, 2001ApJ...551..702H}, visible \citep{2010A&A...517A..91S, 2009A&A...493.1061S, 2004ApJS..155..101B, 1998ApJS..117...89G, 1997ApJS..113..333M, 1997ApJS..112...49M}, and X-Ray \citep{2005AJ....130..539G, 2001ApJ...561..189P} wavelength ranges. Only a few extragalactic SNRs have been identified in the infrared using the [Fe\,{\sevensize II}]\,1.64\,$\micron$ emission line \citep{1997ApJ...476..105G}.
The radiation of SNRs in different wavelength ranges is under the influence of biases as they cover different aspects of the ISM environment and SNR age and evolution \citep{2000ApJ...544..780P, 2010ApJ...725..842L, 2012Ap&SS.337..573S, 2017hsn..book.2005L}.
In the radio, only SNR candidates associated with H\,{\sevensize II}\ regions are identified, which means that radio SNR samples are biased towards star-forming regions.
In the X-Ray, candidates are selected if they display a soft spectrum and if they are associated with an H\,{\sevensize II}\ region, which means that X-Ray samples have the same bias than the radio samples and are also biased against SNR candidates with hard spectra and no optical counterparts. In the optical, samples are under the influence of biases privileging the identification of SNRs located in low density environments. A multi-wavelength (X-Ray, optical, and radio) study of SNRs in NGC\,300 by \cite{2000ApJ...544..780P} revealed 16 new SNRs, 2 in the radio and 14 in the X-Ray, in addition to the 28 SNRs previously identified in the optical. The lack of new optical detection is explained by the fact that optical SNRs can only be detected when they represent relatively low confusion with other H$\alpha$ emission sources. The optical SNRs found here are generally located well away from star-forming regions. Consequently, SNR samples identified optically are often not complete.
Another technique, based on the search for Large-Velocity-Width Sources (LVWS), is used to identify SNRs with optical spectroscopy \citep{1986ApJ...311...85C, 1990ApJ...365..164C}. While H\,{\sevensize II}\ regions show low-velocity dispersion
\citep[$\sigma_{v} \le~30$~km\,s$^{-1}$;][]{1977ApJ...213...15M, 1983ApJ...274..141G},
broad emission line widths observed in the LVWS can be caused by stellar winds or supernova. Using this technique, \citet{1986ApJ...311...85C} discovered four LVWS inside the Giant H\,{\sevensize II}\ regions NGC\,5471 A,B,C and NGC\,5461 in the nearby galaxy M101.
In the visible, a criterion often used for the identification of extragalactic SNRs is based on a value of the emission lines ratio [S\,{\sevensize II}]$\lambda\lambda$6716,6731/H$\alpha$\,$\geq$\,0.4 \citep{1973ApJ...182..697M}.
This criterion was used, for example, to produce the catalogue of \cite{2015yCat..74460943V}.
It allows differentiation between photoionized nebulae, like H\,{\sevensize II}\ regions, and SNRs shock-heated nebulae \citep{2004ApJS..155..101B}. Physically, in a typical H\,{\sevensize II}\ region, where the photoionization of the gas by hot stars is dominating, the second ionization state of the Sulfur, S++, is favored over S+, and the expected theoretical ratio [S\,{\sevensize II}]/H$\alpha$ is between 0.1 and 0.3 \citep{2009A&A...493.1061S}. In an SNR, shock waves are propagating into the ISM after the supernova explosion. The material cools behind these shock waves and it increases the quantity of S+. Consequently, the ratio [S\,{\sevensize II}]/H$\alpha$ becomes higher than 0.3. Regions of diffuse ionized gas (DIG), as seen in many galaxies \citep{1999ApJ...523..223H, 2009ApJ...704..842B} also presents a relatively strong [S\,{\sevensize II}]/H$\alpha$ ratio (as well as a strong [N\,{\sevensize II}]/H$\alpha$ ratio; \citep{1994ApJ...431..156W}; \citep{2005ApJ...632..277E}. These are regions with temperatures similar to H\,{\sevensize II}\ regions ($\sim$10$^4$\,K), lower densities (0.1\,cm$^{-3}$), and slightly lower ionization states. The origin of the DIG is not well understood yet and seems to be complex; it is often suspected of being caused by ionizing photons escaping H\,{\sevensize II}\ regions and traveling large distances, but it can also be related to a generation of post-AGB stars, fast shocks, a weak AGN, and dust diffusion \citep[e.g. ][]{2011MNRAS.415.2182F, 2012A&A...544A..57H, 2012ApJ...758..109S, 2014MNRAS.444.3961D, 2017ApJ...834...40H}
Physical and chemical properties of SNRs can also be determined using emission line diagnostics in the visible; for instance, one can get the electron density with [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731, the shock velocity with [O\,{\sevensize III}]$\lambda$5007/H$\beta$, and the chemical abundances with [N\,{\sevensize II}]$\lambda\lambda$6548,6583/H$\alpha$ \citep[e.g.][]{2017arXiv170107840L}. Some emission lines ratios have been found to display various correlations which may be useful to understand the SNR evolution and environment. For example, a study of the SNRs in M33 by \citet{1998ApJS..117...89G} shows a galactocentric gradient for [N\,{\sevensize II}]$\lambda\lambda$6548,6583/H$\alpha$, no relation between [N\,{\sevensize II}]$\lambda\lambda$6548,6583/H$\alpha$ and [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731, a weak correlation between the SNRs diameter and
[N\,{\sevensize II}]$\lambda\lambda$6548,6583/H$\alpha$ or
[S\,{\sevensize II}]$\lambda\lambda$6716,6731/H$\alpha$, a strong correlation between [N\,{\sevensize II}]$\lambda\lambda$6548,6583/H$\alpha$ and
[S\,{\sevensize II}]$\lambda\lambda$6716,6731/H$\alpha$, and also between [O\,{\sevensize I}]$\lambda$6300/H$\alpha$ and [S\,{\sevensize II}]$\lambda\lambda$6716,6731/H$\alpha$.
More correlations have been seen in the case of SNRs in M31; \citet{1981ApJ...247..879B, 1982ApJ...254...50B} found a clear galactocentric gradient for the ratio [N\,{\sevensize II}]$\lambda\lambda$6548,6583/H$\alpha$, a little evidence of a gradient for [S\,{\sevensize II}]$\lambda\lambda$6716,6731/H$\alpha$, [O\,{\sevensize III}]$\lambda\lambda$4959,5007/H$\beta$, and
[O\,{\sevensize III}]$\lambda\lambda$4959,5007/[O\,{\sevensize II}]$\lambda$3727.
NGC\,3344 is an isolated galaxy \citep{1973AISAO...8....3K}, located at 6.1\,Mpc \citep{2013AJ....146...86T}, and classified as a (R)SAB(r)bc by \cite{1991rc3..book.....D}.
Seen nearly face-on, NGC\,3344 presents a bright inner ring with multiple spiral arms, and numerous H\,{\sevensize II}\ regions. A weak bar can be seen inside the ring.
Its optical diameter is about 8$^{\prime}$, while
its H\,{\sevensize I}\ diameter is 18$^{\prime}$
\citep{1983AJ.....88..272H}. Recently, a Type Ibc supernova (SN 2012fh) was discovered at RA = 10:43:34.05 and DEC = +24:53:29.0 \citep{2012CBET.3263....1N}.
As part of a SITELLE study of the ionized gas emission in NGC\,3344, this first paper focuses on the identification and confirmation of the galaxy's SNR population. The paper is organized as follows: Section~2 describes the observation of NGC\,3344 using SITELLE while section~3 presents the data reduction process and the technique used to measure the emission lines. In Section~4, we describe the method applied to identify the emission line regions, the method developed to subtract the galaxy stellar populations contribution along with the DIG component, the equations used to calculate the internal extinction, and the selection criteria for the SNR candidates.
Section~5 presents our method, based on the emission lines ratios along with Sabbadin plots \citep{1977A&A....60..147S} and Baldwin, Phillips \& Terlevich (BPT) diagrams \citep{1981PASP...93....5B}, used to classify the candidates into three categories, including the Confirmed SNRs. The physical proprieties of the Confirmed SNR are discussed in Section~6 considering, among others, shock models. Conclusions from this work are then presented in Section~7. The detailed study of the H\,{\sevensize II}\ regions of NGC\,3344, along with a discussion of the galaxy SFR and chemical evolution, using the same SITELLE data, will follow in a upcoming paper (Moumen et al. in prep.).
\begin{table*}
\centering
\caption{Observing parameters and sky conditions for NGC\,3344}
\label{tab:param_obs_table}
\begin{tabular}{lccc}
\hline
Filter & SN1 & SN2 & SN3 \\
& 365-385 nm & 480-520 nm & 651-685 nm \\
\hline\hline
Date & 2016 March 07 & 2016 March 06 & 2016 March 05 \\
Seeing & 0.8$^{\prime\prime}$ & 0.8$^{\prime\prime}$ & 0.8$^{\prime\prime}$ \\
Sky quality & Clear & Clear & Clear \\
Folding order & 8 & 6 & 8 \\
Step size [nm] & 1647 & 1680 & 2943 \\
Number of steps & 70 & 134 & 258 \\
Spectral resolution (R) & 400 & 600 & 1500 \\
Exposure time per step [s] & 138.0 & 79.5 & 34.0 \\
Total observing time [h] & 2.8 & 3.1 & 2.7 \\
\hline
\end{tabular}
\end{table*}
\section{OBSERVATION}
NGC\,3344 was observed in March 2016 (RUNID: 16AH41; PI: R.P.~Martin) with the imaging Fourier transform spectrometer (iFTS) SITELLE \citep{2010SPIE.7735E..0BD, 2019MNRAS.485.3930D} installed on the 3.6-m Canada-France-Hawaii Telescope (CFHT).
With a large 11$^{\prime}$ $\times$ 11$^{\prime}$ field of view (FoV) and a
seeing-limited spatial resolution sampled at 0.32$^{\prime\prime}$, SITELLE allows us to obtain more than 4 million spectra in the wavelength range from 350 to 900~nm. In the case of NGC\,3344, the seeing was typically 0.8$^{\prime\prime}$. We selected the three filters SN1 (365-385\,nm), SN2 (480-520\,nm), and SN3 (651-685\,nm) in order to get the strong emission lines useful to study the ionized gas properties ([O\,{\sevensize II}]$\lambda$3727, [O\,{\sevensize III}]$\lambda\lambda$4959,5007, H$\beta$, [N\,{\sevensize II}]$\lambda\lambda$6548,6584, H$\alpha$, and [S\,{\sevensize II}]$\lambda\lambda$6716,6731).
Observation with each filter provides a wavelength calibrated data cube.
The spectral resolution and observing time for each filter, along with other observing parameters are listed in Table~\ref{tab:param_obs_table}. A higher spectral resolution was used for the SN3 filter to allow us to properly separate the emission lines of [N\,{\sevensize II}]$\lambda\lambda$6548,6583 and H$\alpha$, and to study the ionized gas kinematics.
\section{DATA REDUCTION AND LINES FITTING}
\label{sec:datareduction}
The data reduction was performed with ORBS \citep[v. 4.0-DR1-beta;][]{2015ASPC..495..327M, 2016MNRAS.463.4223M}. ORBS is an automatic data reduction software designed for iFTS data obtained with SITELLE freely available on the web\footnote{https://sourceforge.net/projects/orb-orbs/}. ORBS follows in a few steps to transform the interferogram cubes (from the two cameras) into a spectral datacube for each filter \citep[see][]{2019MNRAS.485.3930D}. First, the standard CCD image calibrations are applied, including the bias subtraction, the flat-field correction, and the cosmic rays removal. Second, due to a slight optical misalignment between the cameras, an alignment of the interferogram cubes is performed before combining the two cubes and correcting for atmospheric variations (airmass and clouds). Finally, a discrete Fourier transform is applied, along with the phase correction, on each pixel of the interferogram cube to create the spectral datacube. The wavelength calibration is carried out by ORBS using a high-resolution HeNe data cube collected during the observing run. Images of the standard star GD71 were available for the flux calibration in the three filters.
Given that the spatial resolution is seeing limited, the spaxels have been spatially binned 3$\times$3~pixels to increase the signal-to-noise ratio (S/N) of the fainter regions. The Galactic extinction for NGC\,3344 is negligible ($A_{V} = 0.091$; NED\footnote{NASA/IPAC Extragalactic Database: http://ned.ipac.caltech.edu}) so we ignore it here and in the following analysis.
The sky background was subtracted using the median spectrum built from more than 2$\times$10$^{4}$ pixels in the FoV, located away from the galaxy. As an example, Figure~\ref{fig:skyLineSN3} shows the sky spectrum obtained for the SN3 filter. Many OH lines are seen in this wavelength range. These lines are actually used to refine the wavelength calibration in the SN3 filter \citep[i.e. to take into account the flexure in the instrument during the observation, while the HeNe datacube used for the first order calibration is obtained with the telescope at the zenith; for more details, see][]{2019MNRAS.485.3930D}.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/skyline-sn3.png}
\caption{An example of the sky background spectrum in the case of the SN3 filter. The strong OH lines seen in this wavelength range are identified.}
\label{fig:skyLineSN3}
\end{figure}
The three datacubes have been spatially aligned relative to each other using a homemade Python program that enables a sub-pixel accuracy. Using the datacubes deep image (i.e. the image obtained after summing all the interferograms collected in one filter, prior to the application of the Fourier transform), this program calculates the shifts along the spatial axis $x$ and $y$ using the function \texttt{chi2\_shift} in the Astropy Python package \texttt{image\_registration} and it applies a correction for the misalignment using the function \texttt{shiftnd} from the same package. The SN3 deep image was used as the reference since it is the deepest one. We measured a shift [$dx$, $dy$] of [$-$0.41, 0.94] for SN1 and [0.48, $-$0.54] for SN2. The image of NGC\,3344 obtained from the combination of two of the filter deep images is presented in Figure~\ref{fig:figure_deep}.
\begin{figure*}
\includegraphics[width=\textwidth]{figures/NGC3344-2.png}
\caption{The SITELLE image of NGC\,3344 obtained from the combination of the SN1 (blue) and SN3 (red) deep images. The image is centred on the galaxy (at RA\,=\,10h43m31.15s and DEC\,=\,+24$^\circ$55$^{\prime}$20.0$^{\prime\prime}$) and fills the FoV of SITELLE (11$^{\prime}$ $\times$ 11$^{\prime}$). Bright lines are artefacts caused by saturated stars. North is up and East is left.}
\label{fig:figure_deep}
\end{figure*}
SITELLE has an instrumental profile best reproduced by a cardinal sine (sinc) function. ORCS \citep[v.~1.0.1;][]{2015ASPC..495..327M, 2016MNRAS.463.4223M}, a data extraction software developed specifically for SITELLE used to fit simultaneously all the lines in each datacube. ORCS returned maps, for each emission line, of the amplitude (intensity), FWHM, continuum height, flux, and velocity (based on one or multiple lines centroid). More maps are also returned for the parameters uncertainty. For example, Figure~\ref{fig:snr_spectra} presents the spectrum of one pixel within two SNR candidates studied in this paper, along with their fit obtained with ORCS.
\begin{figure*}
\includegraphics[width=0.8\textwidth]{figures/spectra-fit.png}
\caption{
Examples of SITELLE spectra for one pixel within the SNR candidates SNR-C\,1050 (top) and SNR-C\,159 (bottom). The segments for the three filters are shown including the emission lines, from right to left: SN1 (R = 400), SN2 (R = 600), and SN3 (R = 1500). In black, the observed spectrum and in red, the fit obtained with ORCS. The strong emission lines are identified. The S/N$_{{\rm H}\alpha}$\,=\,19 for SNR-C\,1050 and 80 for SNR-C\,159.}
\label{fig:snr_spectra}
\end{figure*}
In the Appendix~\ref{Flux_maps_A}, Figure~\ref{fig:figure_fluxmaps} shows maps of the flux for all the lines used in this work. It is interesting to notice how the [O\,{\sevensize III}]$\lambda\lambda$4959,5007 emission lines are absent in the galaxy inner ring, while this structure is well marked by the other emission lines, including [O\,{\sevensize II}]$\lambda$3727. Figure~\ref{fig:exVel_map} presents the ionized gas velocity map gathered considering only spaxels with S/N$_{{\rm H}\alpha} \geq 3$. With a small inclination ($i = 18.7^\circ$; HyperLeda\footnote{http://leda.univ-lyon1.fr/}), NGC\,3344 displays a small velocity gradient, of the order of 200~km~s$^{-1}$, between its receding and approaching side. The systemic velocity measured at the galaxy centre (579$\pm$18~km~s$^{-1}$) is in agreement with the value of NED (580$\pm$1~km~s$^{-1}$).
\begin{figure}
\includegraphics[width=\columnwidth]{figures/velocity_map_rac-dec.png}
\caption{NGC\,3344 velocity map. This map was obtained considering
only the spaxels with S/N$_{{\rm H}\alpha}$\,$\geq$\,3.
}
\label{fig:exVel_map}
\end{figure}
\section{SNR CANDIDATES}
\subsection{Automatic Detection of Emission Regions}
\label{sec:auto}
In order to identify the SNR candidates in a non-subjective way, we used the automated identification technique for ionized gas regions described by \citet{2018MNRAS.477.4152R}. This technique was initially created to study the star-forming regions in the nearby, nearly face-on spiral galaxy NGC\,628 where more than 4200 H\,{\sevensize II}\ region candidates have thus been identified. We used this technique, although with some adaptations for our purpose to find SNRs, following the same main three steps:
\noindent (i) The identification of the emission peaks is done using a combination of the [S\,{\sevensize II}]$\lambda$6716 and [S\,{\sevensize II}]$\lambda$6731 continuum-subtracted images. \citet{2018MNRAS.477.4152R} considered the three emission lines H$\alpha$, H$\beta$, and [O\,{\sevensize III}]$\lambda\lambda$4959,5007 for the identification of star-forming regions in NGC\,628. In the case of SNRs, compared to H\,{\sevensize II}\ regions, the [S\,{\sevensize II}] emission lines are rather strong, relative to H$\alpha$, and may be related to a different emission process (i.e. shocks vs. photoionization), and are therefore considered here as better suited to define the SNR's boundary. As an example, Figure~\ref{fig:exsnr159} presents the H$\alpha$ and [S\,{\sevensize II}] maps near one of our best SNR candidates, SNR-C\,159; the difference is striking with a relatively stronger [S\,{\sevensize II}] flux for the SNR compared to the surrounding H\,{\sevensize II}\ regions. Therefore, using the combined [S\,{\sevensize II}] continuum-subtracted image, an emission peak is confirmed if its intensity is greater than the intensity of the 5 surrounding pixels (this number allows us to reduce the amount of false detection due to the noise, without missing fainter peaks in crowded regions) and if the total intensity of a box centred on the emission peak is higher than the adopted detection threshold. The size of the box is 3$\times$3~pixels and the threshold adopted, which is slightly different from a CCD quadrant to another, vary from 2.5$\times$10$^{-19}$ in faint regions to 8$\times$10$^{-18}$~erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$ in the brightest regions. We also add a spectral constraint which consists in considering only pixels with S/N\,$>$\,3 for both of the [S\,{\sevensize II}] lines.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/HaSII-SNR159.png}
\caption{H$\alpha$ (top) and [S\,{\sevensize II}]$\lambda$6716+[S\,{\sevensize II}]$\lambda$6731 (bottom) flux maps near the SNR candidate SNR-C\,159. While the scale of the images was selected so the brightness of the H\,{\sevensize II}\ regions in the H$\alpha$ and the [S\,{\sevensize II}] maps is comparable, SNR-C\,159 appears clearly in the [S\,{\sevensize II}] map which is known to be a good tracer of shock-heated gas.
}
\label{fig:exsnr159}
\end{figure}
\noindent (ii) The determination of the zone of influence (an area influenced by the ionizing photons and shocks from an emission peak) around each emission peak is done by studying the distance between each pixel and the surrounding peaks. A maximum radius of 30 pixels was adopted (i.e. at a distance of 6.1~Mpc for NGC\,3344, this radius corresponds to 285~pc), much larger than the maximum size known for SNRs \citep{2009SerAJ.179...55C} but allowing to detect more extended objects like superbubbles and DIG regions. In general a pixel is associated to its closest peak. If more than one peak is located at the same distance, the pixel is associated to the brightest peak. Finally, the position of the most distant pixels defines the boundaries of the zone of influence.
\noindent (iii) The outer limit of an emission region is defined as the distance where the slope of the flux profile decreases by less than 2\% within the zone of influence. The flux profile is calculated by summing the flux of pixels located in circular annuli 25~pc thick centred on the peak. Figure~\ref{fig:exzoom-outerlim} presents the [S\,{\sevensize II}] map near the SNR candidate SNR-C\,159, where the domain, i.e. the outer limit, for a sample of emission regions is shown.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/contour1.png}
\caption{
The [S\,{\sevensize II}] map near the SNR candidate SNR-C\,159, where the outer limit for various emission regions are shown.}
\label{fig:exzoom-outerlim}
\end{figure}
For each emission region thus defined, the code then fits a radial profile to the flux. The theoretical flux profile adopted is a pseudo-Voight profile (i.e. a combination of a Gaussian and Lorentzian function; see Eq.~3 of \citet{2018MNRAS.477.4152R}), allowing to accurately fit a wide variety of morphology for the emission regions (i.e. from highly peaked to very diffused). The code then returns for each emission region a list of parameters for the best profile fit: the peak intensity ($A$), the FWHM of the profile, the correlation coefficient ($R$) between the profile fit and the flux points, and the background level. Figure~\ref{fig:exProfil_snr} shows examples of the pseudo-Voight profiles obtained for two emission regions: the extended and more diffused region SNR-C\,348 (most probably a DIG region) and the more compact region SNR-C\,159 (one of the best SNR candidates).
The number of emission regions found, using the technique of Rousseau-Nepton et al. (2018) but with the combined [S\,{\sevensize II}] map,
is rather large; 2192 emission regions are found. At this point, it is clear that most of these regions are not SNRs, but H\,{\sevensize II}\, regions (almost 3000 emission regions have been found using the same technique but with the H$\alpha$+H$\beta$+[O\,{\sevensize III}] map; Moumen et al. in prep.) and also DIG regions.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/snr348-profile.png}
\includegraphics[width=\columnwidth]{figures/snr159-profile.png}
\caption{Pseudo-Voight profile of the diffused region SNR-C\,348 (most probably a DIG region) and the compact region SNR-C\,159 (one of the best SNR candidates), as obtained with the technique of Rousseau-Nepton et al. (2018) adapted to SNR candidates.
}
\label{fig:exProfil_snr}
\end{figure}
\subsection{Removal of the Stellar Populations and Diffuse Ionized Gas Background}
\label{sec:bg}
In order to isolate and study the emission lines associated with SNRs, it is important to take into account the light on the line of sight from other sources in the galaxy. NGC\,3344 is rather massive \citep{2013AJ....145..101K}, it is therefore expected that its disk and bulge stellar populations will be responsible for absorption lines that will be superimposed to the emission lines (mainly the Balmer lines). H\,{\sevensize II}\ regions are numerous in NGC\,3344 -- large complexes can be seen in its ring for example -- and may be the place where numerous SN explosions occur. Therefore, H\,{\sevensize II}\ regions are as well expected to pollute the spectrum of many SNRs while: 1) the ionized gas produces additional emission lines, and 2) the young stellar population adds another absorption component underneath the emission lines. In this massive star-forming galaxy, one can also anticipate DIG emission lines to be mixed with the SNR signatures.
We have considered two different methods to create the galaxy background spectrum to be subtracted from each emission region. For the first method, a Local Galaxy Background (LGB) was obtained for each region considering a circular aperture centred on each region. The LGB internal radius corresponds to the outer limit of the region, while its external radius is given when its spectrum reaches S/N\,$\simeq$\,25 in the continuum (a value larger than the best signal-to-noise ratio obtained for the integrated region spectrum). Prior to summing individual spectra in order to create the LGB spectra, a mask was applied to the datacubes to avoid all the other emission regions identified (up to their outer limit) and also all the H\,{\sevensize II}\ region candidates found independently using the technique presented in Section~\ref{sec:auto} but, in this case, with the H$\alpha$+H$\beta$+[O\,{\sevensize III}] image.
For the second method, we considered a Global Galaxy Background (GGB) obtained by creating median spectra in rings centred on the galaxy. The spaxels position and galactocentric radius have been first calculated considering the galaxy inclination $i = 18.7^\circ$ (Hyperleda) and position angle $PA = 150^\circ$ (NED). The width of the different rings is variable as it aims for S/N\,$\simeq$\,25 in the continuum of the GGB spectra (except at large galactocentric radius, where the noise becomes too important as discussed below). Prior to the summation of the GGB spectrum, all the SNR and H\,{\sevensize II}\ region candidates have been masked and all spaxels have been shifted to the galaxy velocity rest frame using the velocity map shown in Figure~\ref{fig:exVel_map}.
We find that both methods used to calculate the galaxy background for the different emission regions have pros and cons. While the LGB takes more easily into account the local distribution of the emission regions, it is not ideal where a high density of emission regions forces a much larger external background radius. For example, in the inner ring, the LGB can select spaxels which are quite far, up to 350\,pc, from the region itself. On the other hand, the GGB may represents better the overall behaviour of the old stellar populations and the DIG at a given galactocentric radius, but it does not take into account the local distribution of the surrounding emission regions. Figure~\ref{fig:bg3} shows, as an example, the flux of the emission lines H$\alpha$, [N\,{\sevensize II}], and [S\,{\sevensize II}], along with the SN3 continuum level, measured in the galaxy background spectra obtained with both methods. First, the two background methods display lines which are only in emission, i.e. pure absorption lines of H$\alpha$ or H$\beta$ from the galaxy stellar populations are never seen in the background spectra. The DIG emission is important all over the galaxy disk. The background H$\alpha$ emission line has a maximum flux corresponding to $\sim$12.56\% of the average H$\alpha$ flux for all the emission regions. Second, both methods indicate stronger emission lines in the inner ring (at a galactocentric distance of $\sim$1\,kpc), with a decreasing intensity toward the outer disk. The LGB displays on average stronger lines and more variations in the spiral arms. In this paper, where we aim at identifying SNRs in an automatic and systematic way, we will favor the GGB method. For some specific purposes, i.e. to discuss the extinction and the effect of the environment on the identification of the SNRs, a comparison using the LGB method will also be done.
Figure~\ref{fig:bg_sp} presents a sequence of normalized GGB SN3 spectra plotted, for comparison, as a function of their galactocentric distance. As already indicated by Figure~\ref{fig:bg3}, these GGB spectra display a significant change in the emission lines strength with their position; the H$\alpha$ emission is quite strong in the galaxy inner ring (at $\sim$1\,kpc). It is clear that the noise becomes very important at a very large radius (even though we tried to reach S/N\,$\simeq$\,25 by increasing the width of the ring). This is mainly explained by the fact that the continuum flux from the disk stellar populations drops quickly. Because of the high level of noise seen at larger radius, and because the spectra are quite similar between 2.5 and 4\,kpc, we simply used the GGB spectrum at 3.5\,kpc for all the region located further away. In the following, unless stated otherwise, for each of the emission region we have subtracted its corresponding GGB spectrum, i.e. for each of the spaxel in an emission region, a GGB spectrum is assigned, scaled to the continuum level of the spaxel, and subtracted. The scaling factor is in general very close to one and assures us that the continuum level of the emission region is never negative. Figure~\ref{fig:spectra_bg} shows the background corrected spectrum of two SNR candidates, one away in the disk and one inside the inner ring, along with the scaled GGB spectrum used.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/BGa-BGc-2.png}
\caption{Comparison of the SN3 emission lines' flux in the galaxy background spectra created with the local LGB (in blue) and global GGB method (in red). There is one data point for all the emission regions, corresponding to the position (i.e. the galactocentric distance) of the emission region peak. All the background spectra have been scaled to the continuum level of the regions prior to line measurements. The galaxy inner ring is located at a galactocentric distance of $\sim$1\,kpc.
}
\label{fig:bg3}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{figures/normal-25-new-2.png}
\caption{SN3 global galaxy background (GGB) spectra for different galactocentric distances. For each spectrum, the central radius of the background annulus is indicated on the plot, along with the annulus half width. NGC\,3344 shows an important DIG contribution that dominates these spectra. At a galactocentric radius of $\geq$\,4\,kpc, the noise level becomes important.
}
\label{fig:bg_sp}
\end{figure}
\begin{figure*}
\includegraphics[width=0.8\textwidth]{figures/1050sp.png}
\includegraphics[width=0.8\textwidth]{figures/159sp.png}
\caption{Spectra of two SNR candidates after the subtraction of the GGB background. SNR-C\,1050 is located inside the galaxy inner ring while SNR-C\,159 has a lower background level at a distance of 3.25\,kpc from the galaxy centre. The weighted spectrum is the background spectrum assigned to the region, before it is scaled to the continuum level of the region.
}
\label{fig:spectra_bg}
\end{figure*}
\subsection{SNR Candidate Selection}
\label{sec:iden-snr}
In order to select the SNR candidates among the emission regions identified in the previous subsections, we used four criteria. These criteria often involve the size (i.e. diameter) of the region. This size parameter is represented by $\Sigma$ through all the paper and is simply set equal to the FWHM of the pseudo-Voight profile obtained for each region. This definition may underestimate the real size of a region, but it allows a secure integrated flux measurement for the region emission lines while avoiding an uncertainty due to the immediate background contamination. The selection criteria are:
\noindent 1) The integrated spectrum of a region must have a line ratio ([S\,{\sevensize II}]$\lambda$6716+[S\,{\sevensize II}]$\lambda$6731)/H$\alpha$ $\geq$\,0.4. This is supported by the work of \citet{1963Natur.199..681M} and \citet{1973ApJ...180..725M}. A total of 1058 emission regions satisfy this criterion alone.
\noindent 2) The spectrum of each region must have S/N\,$\geq$\,5 for the three emission lines H$\alpha$, [S\,{\sevensize II}]$\lambda$6716, and [S\,{\sevensize II}]$\lambda$6731. The number of emission regions that meet this criterion alone is 778.
\noindent 3) The size $\Sigma$ of a region should not exceed 120\,pc. This value is slightly larger than the expected diameter of 100\,pc for Galactic SNRs \citep{2009SerAJ.179...55C}, to take into account mainly the galaxy distance uncertainty. In total, 1030 emission regions satisfy this criterion alone.
\noindent 4) The pseudo-Voight profile of a region must present a correlation coefficient $R \geq 0.5$. In general, emission regions with a lower correlation coefficient are seen in the images as extended DIG regions or blended emission regions of all kinds. A total of 1637 emission regions meet this requirement alone.
Combining these four criteria, we obtain a list of 129 SNR candidates. Figure~\ref{fig:r2c_histo} shows that our selected candidates display a wide range of [S\,{\sevensize II}]/H$\alpha$ and H$\alpha$ flux.
The median H$\alpha$ flux measured for the selected SNR candidates is 3.90$\times$10$^{-16}$\,erg\,s$^{-1}$\,cm$^{-2}$\,\AA$^{-1}$. On average, the size $\Sigma$ of the regions is near 90\,pc, and the profile correlation coefficient $R$ is $\sim$\,80\%. We find no SNR candidate with $\Sigma$\,$<$\,40\,pc, but this limit corresponds to the seeing at the time of the observations.
In Appendix~\ref{cat_snr_cand}, Figure~\ref{fig:figure_snrs2Ha} presents the H$\alpha$+[S\,{\sevensize II}] image with a close-up of the 129 SNR candidates. The candidates appear as green sources due to their relatively stronger [S\,{\sevensize II}] emission. Table~\ref{tab:example_table_snr} gives the SNR candidates properties: their identification number, coordinates, integrated H$\alpha$ flux and ([S\,{\sevensize II}]$\lambda$6716+[S\,{\sevensize II}]$\lambda$6731)/H$\alpha$ flux ratio, region size $\Sigma$, pseudo-Voight profile correlation coefficient $R$, and galactocentric distance. All the emission lines integrated flux for the 129 candidates are listed in Table~\ref{tab:table_emission_lines_flux}.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/Histo-SNR.png}
\caption{Histograms of the selected SNR candidates as a function of: (a)~the [S\,{\sevensize II}]/H$\alpha$ flux ratio, (b)~the H$\alpha$ Flux, (c)~the size $\Sigma$, and (d)~the profile correlation coefficient $R$.
All these candidates satisfy the selection requirements given in Section~\ref{sec:iden-snr}.
}
\label{fig:r2c_histo}
\end{figure}
\subsection{Intrinsic Extinction}
\label{sec:ext}
In general, studies of SNRs make use of emission line ratios that are built from lines near each other in the spectrum, and therefore consider that an extinction correction is not required (e.g. \citealt{2018ApJ...855..140L}). As done in previous optical studies of SNRs, we will consider similar line ratios, but we will also use the [O\,{\sevensize II}]$\lambda$3727/H$\beta$ ratio, since it is available with our SITELLE data. In this case, a reliable extinction correction is necessary. The study of the galaxy intrinsic extinction is done here considering the Balmer decrement. Our sample of emission line regions detected with our technique (\S~\ref{sec:auto}) have different physical gas conditions, but since we are concentrating on the study of the SNRs (and most of the other type of emission regions will be put aside during our selection and confirmation process, as presented below), we adopt a theoretical flux ratio $(F_{{\rm H}\alpha}/F_{{\rm H}\beta})_{\rm theo}$\,=\,3.0, as prescribed for the physical properties of SNRs \citep{2006agna.book.....O}. A similar approach was also taken by \cite{1985ApJ...289..582B} to study the SNRs in M33. The color excess in the SNR candidates obtained in our study is then given by the equation:
\begin{equation}
\label{eq_ebv}
E(B-V) = \frac {2.5} {1.07} \log \bigg(
\frac { ( F_{{\rm H}\alpha} / F_{{\rm H}\beta} )_{\rm obs} } {3.00} \bigg),
\end{equation}
where $E(\beta - \alpha)/E(B-V)$\,=\,1.07 (as extracted from the analysis of \citealt{1989ApJ...345..245C}), and where $({F_{{\rm H}\alpha}/F_{{\rm H}\beta}})_{\rm obs}$ is the observed flux ratio of the two main Balmer lines.
As shown in the Appendix~\ref{fig:figure_fluxmaps}, the H$\alpha$ map is deep enough to study the whole galaxy pixel per pixel. However, the H$\beta$ map presents far less pixels (only 12\% of the whole sample of the H$\alpha$ pixels) with S/N\,$\geq$\,3. In this case, instead of calculating the extinction for each pixel, we are using the integrated flux of each emission region (i.e. by combining the spectrum of the pixels located within an aperture centred on the emission peak with a diameter equal to $\Sigma$, i.e. the FWHM of the region pseudo-Voight profile). When considering the integrated flux, 80\% of the regions have an H$\beta$ detection with S/N\,$\geq$\,3.
Figure~\ref{fig:hbhamap} shows the flux ratio of the H$\alpha$ line over H$\beta$ observed (before and after the subtraction of the galaxy background -- GGB and LGB) for each region as a function of their galactocentric distance (GCD). Prior to the background subtraction, the average ratio is larger than the selected theoretical value of 3 in the galaxy inner 2-3\,kpc, while it drops below this theoretical value at a larger distance. Different technical effects have been investigated to explain values below the theoretical ratio: datacube misalignment, aperture size vs atmospheric refraction, and flux calibration; but no clear correlation has appeared. On the other hand, the region with the lowest ratio, SNR-C2161 with a ratio H$\alpha$/H$\beta = 1.37\pm0.37$, is actually rather extended and a physical explanation involving an important contamination from the DIG is proposed to explain values below the theoretical ratio. After the background subtraction, the H$\alpha$/H$\beta$ ratio is on average near 3 (considering the uncertainties; see Fig.~\ref{fig:hbhamap}) through the whole disk, but again an important scatter is seen. As shown in the figure, the subtraction of the local background (LGB) produce more scatter in the H$\alpha$/H$\beta$ ratio than the subtraction of the global background (GGB). As suspected, the LGB is more easily influenced by the non uniformity of the emission from surrounding nebulae.
Since the background subtracted H$\alpha$/H$\beta$ ratio is on average very close to the theoretical value, one might conclude that the galaxy internal extinction is rather low (values of $E(B-V)$ are also shown in Fig.~\ref{fig:hbhamap}). But this is not so clear considering the importance of the dispersion and the complex nature of the galaxy background. As discussed in Section~\ref{sec:bg}, the background contribution is a combination of the underlying stellar populations absorption lines and DIG emission lines. As it can be seen with the examples of Figure~\ref{fig:spectra_bg}, these two effects seems to cancel out well in the H$\beta$ line compared to H$\alpha$, resulting in a relatively weaker H$\alpha$ emission line and a ratio H$\alpha$/H$\beta$ closer to the theoretical value. On the other hand, the background is more complex than presented here, as its components are spatially variable, relative to each other and relative to the SNRs, as the DIG present distinct physical properties (e.g. electron temperatures and densities) compared to SNRs, and as these components are mixed with non-homogeneous layers of dust. We therefore suspect that the scatter seen in the background subtracted H$\alpha$/H$\beta$ ratio is related to the complexity of the background. With the data in hand, we cannot separate these multiple components and their extinction (as mentioned previously, we never get to observe simple Balmer absorption lines in the galaxy disk, which would help us for example to isolate the effect of the stellar populations). It is therefore difficult to estimate here the level at which the background subtraction affects the extinction measurement
Considering that the background corrected H$\alpha$/H$\beta$ ratio is on average near 3, no extinction correction was actually applied to the data. In the remaining of the paper, only the background (GGB) correction was made.
\begin{figure*}
\includegraphics[width=\textwidth]{figures/hahbebc.png}
\caption{The H$\alpha$/H$\beta$ ratio for all the SNR candidates. Left: before the background subtraction. Centre: after the background subtraction using the LGB (blue) and the GGB (red) measurements.
Right: the galactocentric variation of the color excess obtained using Equation~\ref{eq_ebv} and considering only the SNR candidates with $S/N > 5$ for all the emission lines H$\beta$, H$\alpha$, [SII]$\lambda$6716, and [SII]$\lambda$6731.
}
\label{fig:hbhamap}
\end{figure*}
\section{SNR Optical Confirmation}
The emission lines available in the SITELLE data are here considered
in Sabbadin plots (\S~\ref{sec:sabaddin}) and BPT diagrams (\S~\ref{sec:bpt}),
in order to identify the ionization mechanism at play in the regions and confirm the best SNRs in our sample of 129~candidates. Figure~\ref{fig:decision_figure} presents our tree of decisions used to classify the SNR candidates into three categories: Confirmed SNRs, Probable SNRs, and Less likely SNRs. The following subsections describe in detail our analysis and classification scheme for the SNR candidates.
\begin{figure*}
\includegraphics[width=0.8\textwidth]{figures/New_Decision.png}
\caption{
Tree of decision used to confirm the SNRs in NGC\,3344.
Sabbadin plots and BPT diagrams, based on emission line ratios, offer constrains for the ionization mechanism.
By the end of this analysis, the 129 SNR candidates fall into
one of the three categories: Confirmed (42), Probable (45), or Less likely SNRs (42).
}
\label{fig:decision_figure}
\end{figure*}
\subsection{Sabbadin Plots}
\label{sec:sabaddin}
\citet{1977A&A....60..147S} performed a spectroscopic and photometric investigations on the nearby Planetary Nebula (PN) Sharpless 176 and compared its emission line ratios with those of SNRs, H\,{\sevensize II}\ regions, and other PNe. In their plots, where they compare emission line ratios of H$\alpha$/([N\,{\sevensize II}]$\lambda$6548+[N\,{\sevensize II}]$\lambda$6583), H$\alpha$/([S\,{\sevensize II}]$\lambda$6716+[S\,{\sevensize II}]$\lambda$6731), and [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731, SNRs, H\,{\sevensize II}\ regions, and PNe are distinct phenomena as they occupy different zones.
While the electron density is expected to be higher in SNRs during the radiative phase (compared to H\,{\sevensize II}\ regions), a smaller ratio [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731 and larger ratios of a line involving electron collisions (like [N\,{\sevensize II}] and [S\,{\sevensize II}]) over the H$\alpha$ resonance line is expected. Following the work of Sabbadin et al., more emission regions from our galaxy and from nearby galaxies have been considered so that it confirmed the behavior originally found. For example, a detailed version of the Sabbadin plots considering many SNRs identified in the optical was produced by \citet{2013MNRAS.429..189L}.
The first rule in confirming the SNR nature of our 129 candidates is to find them in the SNR zone of at least one of the three Sabbadin plots.
The choice for only one plot is justified by the fact that some physical factors (like the contamination by an H\,{\sevensize II}\ region and/or by the DIG) may move the SNR away from the shock region in some of the Sabbadin plots.
Figure~\ref{fig:sabaddin1} shows the Sabbadin plots, where our 129 candidates are represented by filled circles and where the domain for Galactic SNRs, H\,{\sevensize II}\ regions, and PNe are delimited by shaded zones and lines. The color of the circles gives the final classification as discussed in Section~\ref{sec:fc}. Circles with a `$+$' symbol indicate SNR candidates with S/N\,$<$\,3 for the [N\,{\sevensize II}] lines (the criterion used to select the SNR candidates already implied S/N\,$\geq$\,5 for the H$\alpha$ and [S\,{\sevensize II}] lines). The 42 circles with a `$+$' are in black as they are automatically considered Less likely SNRs. The remaining candidates (87/129), fall within (considering the line ratio uncertainties) the SNR zone in at least one of the three Sabbadin plots.
In the Sabbadin plots, three candidates, SNR-C\,233, 332, and 1135, display a large value of the [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731 ratio, but with a large uncertainty due to weaker lines. Although older SNRs may have, in theory, a ratio larger that 1.43, we believe that these candidates, because of their weak lines, are better explained by very low density DIG regions. The candidate with a very high density (low [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731 ratio) and the lowest H$\alpha$/[N\,{\sevensize II}] ratio is SNR-C\,1050. This candidate is the only one located inside the galaxy inner ring. Within the inner ring, the DIG emission is important, with strong nitrogen emission, but we are confident that along with the galaxy background subtraction, the DIG contribution was removed properly. No star-forming regions are surrounding SNR-C\,1050 which may indicate that it was produced more recently by a type Ia supernovae.
\begin{figure*}
\includegraphics[width=\textwidth]{figures/sabaddin-1.png}
\caption{Sabbadin plots with the 129 SNR candidates found in NGC\,3344.
The final classification of the candidates is shown using different colors: red for the Confirmed SNRs, green for the Probable SNRs, and black for the Less likely SNRs. Shaded zones and lines are used to separate the position of the SNRs, H\,{\sevensize II}\ regions, and planetary nebulae according to \citet{1977A&A....60..147S}.
Circles with a `$+$' symbol indicate SNR candidates with S/N\,$<$\,3 for the [N\,{\sevensize II}]$\lambda\lambda$6548,6583 lines, while circles with a `$\times$' symbol indicate candidates with S/N\,$<$\,3 for the H$\beta$ or [O\,{\sevensize III}] lines.
}
\label{fig:sabaddin1}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{figures/sabaddin-2.png}
\contcaption{}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{figures/sabaddin-3.png}
\contcaption{}
\end{figure*}
\subsection{BPT Diagrams}
\label{sec:bpt}
The BPT diagrams, named after Baldwin, Phillips, and Terlevich \citep{1981PASP...93....5B}, represent another tool to identify the principal excitation mechanism in the nebular gas. Baldwin et al. showed empirically that several combinations of optical emission lines can be used to separate objects into four categories according to the principal excitation mechanism. These categories are: normal H\,{\sevensize II} regions, planetary nebulae, objects photoionized by a power-law continuum, and objects excited by shock-wave heating. Using these emission lines, Baldwin et al. created three diagrams: BPT-NII with [N\,{\sevensize II}]$\lambda$6583/H$\alpha$ vs [O\,{\sevensize III}]$\lambda$5007/H$\beta$; BPT-SII with [S\,{\sevensize II}]$\lambda\lambda$6716,6731/H$\alpha$ vs [O\,{\sevensize III}]$\lambda$5007/H$\beta$; and BPT-OI with [O\,{\sevensize I}]$\lambda$6300/H$\alpha$ vs [O\,{\sevensize III}]$\lambda$5007/H$\beta$. All the emission lines used in the BPT diagrams are available with SITELLE except the [O\,{\sevensize I}]$\lambda$6300 line.
The second rule in confirming the nature of the SNR candidates is for them to fall outside the H\,{\sevensize II}\ region zone in at least one of the BPT diagrams that we can draw with the available data.
Again, if a background residue is present, it may move the
SNR away from the shock region and, therefore, the constraint is put for only one BPT diagram.
Figure~\ref{fig:BPT1} shows the BPT-NII and BPT-SII diagrams,
where the filled circles (with different colors for the final classification as discussed in \S~\ref{sec:fc}) represent our 129 SNR candidates. In these diagrams, the domains for the H\,{\sevensize II}\ regions, shocks, LINERS, and Seyfert are indicated based on the work of \citet{2003MNRAS.346.1055K} and \citet{2001ApJ...556..121K}. As in Section~\ref{sec:sabaddin}, circles with a `$+$' symbol indicate candidates with S/N\,$<$\,3 for the [N\,{\sevensize II}] lines. Circles with a `$\times$' symbol are used in the case of candidates with S/N\,$<$\,3 for the [O\,{\sevensize III}]$\lambda$5007 or H$\beta$ lines (and the selection criterion for the SNR candidates already implied S/N\,$\geq$\,5 for the H$\alpha$ and [S\,{\sevensize II}] lines). Along with the 42 regions with a `$+$' symbol, there are 29 regions with a `$\times$' symbol.
The remaining 42 candidates (with S/N\,$\geq$\,3 for H$\beta$, [O\,{\sevensize III}], and [N\,{\sevensize II}]), fall (considering the line ratio uncertainties) inside the zone for shock regions in, at least, one of the two BPT diagrams. No foreground AGN is known in the field of view which could correspond to any of our SNR candidates.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/BPT_NII}
\includegraphics[width=\columnwidth]{figures/BPT_SII.png}
\caption{BPT-NII and BPT-SII diagrams of the 129 SNR candidates found in NGC\,3344. Colors and symbols are the same as presented in Figure~\ref{fig:sabaddin1}. The domain for the H\,{\sevensize II}\ regions (dotted line), the composite/transition regions, and pure shock regions labelled LINERS, Seyfert, and AGN (continuous line) are indicated based on the work of \citet{2003MNRAS.346.1055K} and \citet{2001ApJ...556..121K}, respectively.
}
\label{fig:BPT1}
\end{figure}
\subsection{Final Classification}
\label{sec:fc}
If both the Sabbadin plots and BPT diagrams related classification rules are satisfied, the candidate falls in the category of Confirmed SNRs, and it is represented by a filled circle in red in Figures~\ref{fig:sabaddin1} and \ref{fig:BPT1}. If the candidate is within an SNR domain in at least one of the three Sabbadin plots and is not satisfying the rule for the BPT diagrams, it is considered as a Probable SNR and it is shown in green in the figures. If the candidate is outside the SNR domain in all five plots (with the required signal-to-noise ratio for the lines), it is a Less likely SNR and it is represented in black in the figures. From the start, all 129 candidates have a ratio [S\,{\sevensize II}]/H$\alpha$\,$\ge$\,0.4, but only 42 end as Confirmed SNRs, 45 as Probable SNRs, and 42 are called Less likely SNRs.
Because of their poor signal-to-noise ratio, many candidates (with `$+$' or `$\times$' symbols in Fig.~\ref{fig:sabaddin1} and \ref{fig:BPT1}) have been put in the category of Less likely SNRs and Probable SNRs, despite the fact that some of them are located in the SNR zones in Sabbadin plots and BPT diagrams. A deeper observation could help confirm their category as, for example, an old SNR may indeed present a very weak [O\,{\sevensize III}]$\lambda$5007 line.
On the other hand, Less likely SNRs that fall in the domain for H\,{\sevensize II}\ regions in the BPT diagrams (Fig.~\ref{fig:BPT1}) and in the Sabbadin plots (Fig.~\ref{fig:sabaddin1}) raise the question of some remaining contamination by the immediate environment: are we looking at SNR superposed with an H\,{\sevensize II}\ region, are we confusing an H\,{\sevensize II}\ region bathing into the DIG with an SNR, or are we seeing clumps in the DIG?
For the three SNR categories, we actually find no significant difference between the average distance ($\sim$\,$7.5 \pm 2.5$~pixels) between the SNR and the closest emission peak (in H$\alpha$ and [S\,{\sevensize II}]) or between the average number ($\sim$\,$8 \pm 3$~pixels) of emission peaks surrounding an SNR within a radius of 15 pixels.
For the line ratios presented in Figures~\ref{fig:sabaddin1} and \ref{fig:BPT1}, we have subtracted a global galaxy background (GGB) which may not be ideal to consider the immediate environment of the regions, but, as we have shown in Figure~\ref{fig:bg3}, it displayed less dispersion compared to the local galaxy background (LGB).
\section{SNR Properties}
\subsection{Shock Models}
Shock models have been developed by various teams and are often used to study the physical properties of SNRs. The models from \cite{2008ApJS..178...20A} and \cite{1984ApJ...276..653D} have been of particular interest for us. \cite{2008ApJS..178...20A} produced a library of fully radiative shock models calculated with the
photoionization and shocks code MAPPINGS III. This code produces grids
of models taking into account different emission line ratios with shock velocities in the range [100-1000]~km~s$^{-1}$ and the magnetic parameter \textit{B/n$^{1/2}$} between 10$^{-4}$ to 10~$\mu$G cm$^{3/2}$ for five different atomic abundance sets and for a preshock density of 1.0~cm$^{-3}$. \cite{1984ApJ...276..653D} derived an extensive grid of models for fixed physical conditions appropriate to a typical evolved SNR.
In Figure~\ref{fig:ModelsAllen}, we present grids from \citet{2008ApJS..178...20A} for the emission line ratios [O\,{\sevensize III}]$\lambda$5007/H$\beta$ versus [N\,{\sevensize II}]$\lambda$6583/H$\alpha$ for the shock-only and shock+precursor models considering different abundances.
The first plot (shock-only) represents the physical extreme case of the absence of ionized gas ahead of the shock front, while the second plot (shock+precursor) represents the case of having an extensive, radiation bounded, precursor region ahead of the shock. For these grids, the vertical lines represent different values of the magnetic field, while the horizontal lines represent the shock velocity. This figure indicates a strong dependence of the emission lines ratio [N\,{\sevensize II}]$\lambda$6583/H$\alpha$ with the metallicity, while the [O\,{\sevensize III}]$\lambda$5007/H$\beta$ is more sensitive to the shock velocity. Our sample of Confirmed SNRs have been superposed on these plots.
In both models, while different values of the magnetic field may be considered, a very low metallicity (i.e. SMC type) is excluded and a low value of the shock velocity is favored. More particularly in the case of the shock+precursor model, a shock velocity below 250~km\,s$^{-1}$ is prescribed for the Dopita2005, solar, and 2$\times$solar metallicity. With the shock+precusor model, many Confirmed SNRs are best reproduced with Dopita2005, solar, the 2$\times$solar metallicity grid which is consistent with the value found in literature (12 + LOG(O/H) = 8.72) for H\,{\sevensize II}\ regions \citep{2014AJ....147..131P}.
The model of \citet{2008ApJS..178...20A} also predicts a weak [O\,{\sevensize III}]$\lambda$5007/H$\beta$ line ratio when the shock velocity is small.
The absence of the [O\,{\sevensize III}]$\lambda$5007 as a result of shocks with a low velocity is
also supported by the model of \citet{1987ApJ...316..323H}. Therefore some of the candidates in the Probable SNRs category due to a week [O\,{\sevensize III}] line with a poor signal-to-noise ratio, should be reconsidered with deeper observations.
Finally, shock waves in SNRs evolve from being non-radiative to radiative where their velocity becomes lower. For Galactic SNRs, such as Cygnus Loop, proper motions of the filaments show a complicated structure and one can observe both non-radiative and radiative shocks depending on the environment in which these filaments expand \citep{2014ApJ...791...30M, 2009ApJ...702..327S}. Using optical data, we can distinguish radiative from non-radiative shocks by looking at their H$\alpha$ and [S\,{\sevensize II}] emission. While non-radiative shocks would be seen only in H$\alpha$, radiative shocks would show up in both filters. In our sample, all the SNR candidates seem to be in the radiative shock phase since they have emission in both H$\alpha$ and [S\,{\sevensize II}] emission with a signal-to-noise higher than 5.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/histo-disp.png}
\caption{
Histogram of the H$\alpha$ velocity dispersion $\sigma_v$ for the Confirmed SNRs}
\label{fig:HistoDisp}
\end{figure}
\begin{figure*}
\includegraphics[width=\textwidth]{figures/sigma_v.png}
\caption{Relations between the H$\alpha$ velocity dispersion $\sigma_v$ of the Confirmed SNRs and their emission lines ratios, size~($\Sigma$) and galactocentric distance (GCD). The parameters of the linear fit (dotted line), along with the correlation coefficient ($CR$), are indicated for each relation.}
\label{fig:Ha-sigma}
\end{figure*}
\begin{figure}
\includegraphics[width=\columnwidth]{figures/Allen1.png}
\includegraphics[width=\columnwidth]{figures/Allen2.png}
\caption{[O\,{\sevensize III}]$\lambda$5007/H$\beta$ versus [N\,{\sevensize II}]$\lambda$6583/H$\alpha$ based on
(a)~the shock-only and (b)~the shock+precursor models of \citet{2008ApJS..178...20A}
for five abundances (SMC, LMC, Dopita2005, solar, and \mbox{2 $\times$ solar},
as indicated on the plots) with the density $n = 1$~cm$^{-3}$.
Vertical lines represent different values of
the magnetic parameter, from 10$^{-4}$ to 10 $\mu$G
cm$^{3/2}$,
while horizontal lines represent the shock velocity, from 200 to 1000~km\,s$^{-1}$,
}
from the top to the bottom for the shock-only model and from the bottom to the top for the shock+precursor model.
In both models, while different values of shock velocity and magnetic field may be considered, a very low metallicity (i.e. SMC type) is excluded.
\label{fig:ModelsAllen}
\end{figure}
Figure~\ref{fig:dopita} shows curves and grids of different emission lines ratios by \cite{1984ApJ...276..653D} in the case of a weak shock velocity of 106~km\,s$^{-1}$.
The curves and grids take into account different abundances O/N/C, O/S, and O/N, and metallicities Z(O). The first plot makes use of the [O\,{\sevensize II}]$\lambda$3727 line. For comparison, SNRs identified in M81 and M82 \citep{2015ApJ...804...63L} and in M31 \citep{1999AJ....118.2775G} have been superposed on the different plots, when possible, along with our Confirmed SNRs. Our SNRs sample do not match the curves in the top plots (where a fixed abundance ratio O/S=42.8 was considered). As this is also the case for the M31 SNRs,
\citet{2015ApJ...804...63L} proposed that [N\,{\sevensize II}]/H$\alpha$ vs. [O\,{\sevensize III}]/H$\beta$ diagram may be better for the estimation of the SNR abundances. As shown in the bottom plots of Figure~\ref{fig:dopita}, this may be the case since most of the objects falls within the model grids, allowing to derive the O/N abundance and Z(O). In this case, using [N\,{\sevensize II}]/H$\alpha$ vs. [O\,{\sevensize III}]/H$\beta$, our Confirmed SNRs have an O/N abundance between 6 and 12 and Z(O) varying from 1.5$\times$10$^{-4}$ to 6$\times$10$^{-4}$. The O/N abundance seems to be higher than the central O/N obtained from \citep[][O/N $\simeq$ 4.36]{2014AJ....147..131P}. However, this value is consistent with values found using [N\,{\sevensize II}]/H$\alpha$ vs. [S\,{\sevensize II}]$\lambda$6731/H$\alpha$ (O/N between 1.5 and 6).
\begin{figure*}
\includegraphics[width=0.7\textwidth]{figures/Dopita-1.png}\\
\includegraphics[width=0.7\textwidth]{figures/Dopita-2.png}
\caption{
Emission line ratios based on the model of Dopita et al. (1984) for a shock velocity
of 106~km\,s$^{-1}$. (a) for a fixed abundance ratio O/S=42.8 and different abundances O/N/C as indicated. (b) for different ratios O/S as indicated. (c) and (d) for a fixed abundance
ratio O/S=42.8 and different values of O/N and Z(O) as indicated. SNRs from M81 and M82
(Lee et al. 2015) along with SNRs from M31 (Galarza et al. 1999) and our
Confirmed SNRs in NGC\,3344 are superposed to the curves and grids.
}
\label{fig:dopita}
\end{figure*}
Finally, other shock models by \cite{1985ApJ...298..651C} and \cite{1987ApJ...316..323H}
proposed that a value of the ratio [O\,{\sevensize III}]$\lambda$5007/H$\beta$ bellow 6 indicates shocks with complete recombination zones and values higher than 6 are expected to be found for shocks with incomplete recombination zones \citep[which means that the entire postshock cooling zone is not fully established;][]{1988ApJ...324..869R}. As we can see in the Figure~\ref{fig:BPT1}, all the Confirmed SNRs in our sample have a ratio [O\,{\sevensize III}]$\lambda$5007/H$\beta$ bellow 6 (the highest [O\,{\sevensize III}]$\lambda$5007/H$\beta$ ratio is $\sim$\,1) which means we are seeing shocks with complete recombination zones.
Figure \ref{fig:HistoDisp} presents the distribution of the velocity dispersion $\sigma_{v}$ measured with the H$\alpha$ line for the confirmed SNRs.
Values of $\sigma_{v}$ are ranging from 31 to 93 km\,s$^{-1}$ with two peaks at 35 and 55 km\,s$^{-1}$. In Figure \ref{fig:Ha-sigma}, we explored the relation between $\sigma_{v}$ and the galactocentric distance of the regions, their size $\Sigma$, and their different emission lines ratios.
The correlation coefficient for these relations are rather small. The most significant relations are found with the emission line ratios
[N\,{\sevensize II}]$\lambda$6583/H$\alpha$ and [O\,{\sevensize III}]$\lambda$5007/H$\beta$.
\subsection{SNR sizes}
A measurement of the SNR radius $\Sigma$ is taken as the FWHM of the pseudo-Voight profile fitted for each region based on their [S\,{\sevensize II}]$\lambda$6716+[S\,{\sevensize II}]$\lambda$6731 emission (\S~\ref{sec:auto}) considering that it is a good tracer for shocks. The spatial resolution obtained in this study (0.8$^{\prime\prime}$) allows us to identify SNR candidates with $\Sigma_{{\rm limit}} \ge 28.5$~pc. In our sample of Confirmed SNRs, the smallest one has $\Sigma \simeq 40$~pc, above this limit. As the size of the remnants may be an indication of its evolution \citep{2010MNRAS.407.1301B}, we have compared $\Sigma$ with the various emission lines ratios observed. In Figure~\ref{fig:sizerel}, we show the relation between $\Sigma$ and different emission line ratios: [S\,{\sevensize II}]/H$\alpha$, [N\,{\sevensize II}]/H$\alpha$, [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731, [O\,{\sevensize III}]/H$\beta$, and also [O\,{\sevensize II}]/H$\beta$. Although we can superpose a linear fit to the data, revealing a negative slope for all ratios except [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731 where a positive slope is seen, the scatter is rather large (particularly when the weaker H$\beta$ line is involved) and the probability of a significant fit (as given in each plot) is rather small. Nevertheless, the trend displayed by the slopes is interesting and rather consistent with a smaller velocity (when [O\,{\sevensize III}]$\lambda$5007 is weaker) and a lower density (when [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731 is larger) in older SNR (i.e. when they are more extended according to $\Sigma$).
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{figures/SII_Ha_size.png}
\includegraphics[width=0.7\columnwidth]{figures/NII_Ha_size.png}
\includegraphics[width=0.7\columnwidth]{figures/SII_ratio_size.png}
\includegraphics[width=0.7\columnwidth]{figures/OIII5007_Hb_size.png}
\includegraphics[width=0.7\columnwidth]{figures/OII_Hb_size.png}
\caption{Emision lines ratios for the Confirmed SNRs
as a function of $\Sigma$. The parameter $\Sigma$, a measurement of the SNR radius, corresponds to the FWHM of the SNRs pseudo-Voight profile. The dashed lines display a linear fit through the data. The fit parameters are given for each plot, along with the correlation coefficient ($CR$).
}
\label{fig:sizerel}
\end{figure}
We also looked for a relation between $\Sigma$ and the position of the remnant withing the galaxy, wondering if the galactic environment has an affect on this parameter. More precisely, we draw a plot of $\Sigma$ as a function of the galactocentric distance, but we did not find any significant relation.
We also compared the average value of $\Sigma$ for SNRs located in different structures, like the inner ring, arms, and inter arms regions, but again we did not find any important difference.
Finally, we directly investigated the galactocentric variation of the different emission line ratios. As shown in Figure~\ref{fig:GCD-relation},
a negative gradient is seen for [S\,{\sevensize II}]/H$\alpha$ and [N\,{\sevensize II}]/H$\alpha$, while the other three ratios, [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731, [O\,{\sevensize III}]/H$\beta$ and [O\,{\sevensize II}]/H$\beta$, display a positive gradient. The plot with [N\,{\sevensize II}]/H$\alpha$ has less scatter and a higher probability, and may therefore suggests a higher metallicity for the SNRs near the galaxy centre.
This result is agreement with the strong abundance gradient for H\,{\sevensize I}\ regions found by \citep{1981PASP...93..273M} and confirmed by \citep{1988PASP..100.1428V}.
\begin{figure}
\centering
\includegraphics[width=0.7\columnwidth]{figures/SII_Ha_GCD.png}
\includegraphics[width=0.7\columnwidth]{figures/NII_Ha_GCD.png}
\includegraphics[width=0.7\columnwidth]{figures/SII_ratio_GCD.png}
\includegraphics[width=0.7\columnwidth]{figures/OIII_Hb_GCD.png}
\includegraphics[width=0.7\columnwidth]{figures/OII_Hb_GCD.png}
\caption{Emission line ratios for the Confirmed SNRs
as a function of the galactocentric distance.
The dashed lines display a linear fit through the data. The fit parameters are given for each plot, along with the correlation coefficient ($CR$).
}
\label{fig:GCD-relation}
\end{figure}
\section{Conclusions}
In this paper, we used the high spectral and spatial resolution data obtained with the iFTS SITELLE at the CFHT to study a sample of SNRs in the nearby galaxy NGC\,3344. Our conclusions are as follows:
\noindent 1) We report the first identification and confirmation of a sample of SNRs in NGC\,3344. Our systematic analysis using
criteria (related to the physics and morphology of the emission regions), reveals 129~SNR candidates which are classified in three categories: Confirmed (42), Probable (45), and Less likely (42) SNRs.
\noindent 2) We present a self-consistent spectroscopic analysis, exploiting all the emission lines available with SITELLE and using the Sabbadin plots and BPT diagrams, to confirm the shock-heated nature of the ionization mechanism in the SNR candidates. Eventually, the [OI]$\lambda$6300 line should be added to this analysis, allowing to consider (i) the [OI]$\lambda$6300 emission as another tracer of SNRs and (ii) the BPT-OI diagram to complete the available list of diagrams confirming the mechanisms of ionization in the optical range.
\noindent 3) We have considered two methods to subtract the galaxy background spectrum from each emission region.
One of these methods consisted in creating a
local galaxy background (LGB) by considering a median spectrum in a small annulus centred on a region. For the other method we created a global galaxy background (GGB) for each region by considering in a ring centred on the galaxy, with a radius equal to the region galactocentric distance.
Although the LGB and GGB spectrum of a same region are similar in general (with more distinction for regions in the galaxy inner ring), we have adopted the GGB method because it may better represent, with less dispersion, the global stellar population in the disk along with the DIG component, which are quite important in this galaxy.
\noindent 4) After the background subtraction, the galaxy internal extinction seemed rather small, but this may require further investigation as the nature of the background is complex.
\noindent 5) We have compared shock models from \citet{2008ApJS..178...20A} with our emission line ratios obtained for the Confirmed SNRs. A metallicity ranging between LMC and 2$\times$solar is then prescribed for these objects. A comparison with the shock models of \citet{1984ApJ...276..653D} reveal an O/N abundance between 6 and 12 and Z(O) varying from 1.5$\times$10$^{-4}$ to 6$\times$10$^{-4}$.
\noindent 6) We do not find a strong trend between the SNR $\Sigma$ parameter, i.e. a measurement of the radius of the remnants, and their emission lines ratios, except maybe that extended ones (i.e. older ones) have a lower density (i.e. higher [S\,{\sevensize II}]$\lambda$6716/[S\,{\sevensize II}]$\lambda$6731) and a lower velocity (i.e. weaker [O\,{\sevensize III}]$\lambda$5007 lines).
There are no obvious relation between $\Sigma$ and the SNR environments in the galaxy (i.e. their galactocentric radius and position inside the inner ring, arms, or between the arms). But we find a significant gradient withing the galaxy for the line ratio [N\,{\sevensize II}]/H$\alpha$.
These correlations may indicate a metallicity gradient among
the SNR population, along with some evolutionary effect (age and shock velocity).
Although SITELLE is a powerful tool for the systematic identification and optical study of SNRs, multi-wavelength observations will also be necessary to have a complete picture of these objects, mainly to improve our understanding of their physics and evolutionary phase, to gain a detailed description of their morphology and environment, and to evaluate their impact on the galaxy evolution. A study using near-infrared data obtained recently with the [Fe\,{\sevensize II}]\,1.64\,$\mu$m filter (MOIRCS/Subaru), X-rays archived data (Chandra), and high spatial resolution optical images (HST) is ongoing to complete our work on the SNRs population in NGC\,3344 (Moumen et al. in prep.). Furthermore, the SITELLE data are being reanalyzed to describe the H\,{\sevensize II}\ regions population of NGC\,3344, along with its recent star formation history and in relation with the SNRs described here (Moumen et al. in prep.).
\section*{Acknowledgements}
This research is based on observations obtained at the Canada-France-Hawaii Telescope (CFHT) which is operated from the summit of Mauna Kea by the National Research Council of Canada, the Institut National des Sciences de l’Univers of the Centre National de la Recherche Scientifique of France, and the University of Hawaii. The observations at the Canada-France-Hawaii Telescope were performed with care and respect from the summit of Mauna Kea which is a significant cultural and historic site. The observations were obtained with SITELLE, a joint project between Universit\'e Laval, ABB-Bomem, Universit\'e de Montr\'eal and the CFHT with funding support from the Canada Foundation for Innovation (CFI), the National Sciences and Engineering Research Council of Canada (NSERC), Fonds de Recherche du Qu\'ebec - Nature et Technologies (FRQNT), and CFHT. LRN, CR, and LD are grateful to the Fonds de recherche du Qu\'ebec - Nature et Technologies (FRQNT) for individual and team financial support. CR and LD are grateful to the Natural Sciences and Engineering Research Council of Canada (NSERC). This research has made use of NASA’s Astrophysics Data System and of the VizieR catalogue access tool, CDS, Strasbourg, France.
\bibliographystyle{mnras}
|
\section{Appendix A: Derivation of Exact Athermal Eigenstates}
\label{sec: Appendix A}
\setcounter{equation}{0}
\renewcommand{\theequation}{A\arabic{equation}}
In this Appendix we show that the scar states $|\mathcal S_n\rangle$ defined in Eq.~\eqref{eq:scar} are indeed eigenstates of the Hamiltonian $H$ of Eq.~\eqref{eq:H-XY} on a $d$-dimensional hypercubic lattice. We also show that these states remain in the presence of any exchange term that preserves the original U(1) symmetry of Eq.~\eqref{eq:H-XY} and the bipartiteness of the hypercubic lattice. Finally, we comment on the existence of other exact eigenstates in the many-body spectrum of $H$. These additional exact eigenstates include another tower of athermal eigenstates at finite energy density that arise only for periodic boundary conditions and anisotropy $D=0$.
\subsection{The Scar States $|\mathcal S_n\rangle$}
To demonstrate that the states $|\mathcal S_n\rangle$ are indeed eigenstates of $H$, we work in the local $S^z$ basis, where we label the local spin states
\begin{align}
|m_i=\pm 1\rangle,|m_i=0\rangle \equiv |\pm\rangle,|0\rangle,
\end{align}
respectively. The states are then automatically eigenstates of the $h$ and $D$ terms, and it remains to check the action of the nearest-neighbor hopping term $J$ for each $n$. We do this below for the case $n=1$ in Sec.~\ref{sec:NN-exch} before proceeding to consider further-neighbor exchange processes in Sec.~\ref{sec:FN-exch}. We then consider the case $n>1$ in Sec.~\ref{sec:hardcore} and show that the presence of additional bimagnons does not spoil the hopping properties demonstrated in the preceding subsections, a fact that we attribute to the hardcore nature of the bimagnons.
\subsubsection{$|\mathcal S_1\rangle$ Nullifies Nearest-Neighbor Exchange}
\label{sec:NN-exch}
We begin by considering $d=1$, where the state
\begin{align}
\label{appeq:S_1d=1}
|\mathcal S_1\rangle=\frac{1}{\sqrt{L}}\sum^L_{j=1} e^{i \pi j}\, |\,\underset{j-1}{\underbrace{-\cdots}}+\underset{L-j}{\underbrace{-\cdots}}\,\rangle
\end{align}
(recall $V=L$ in $d=1$).
We will assume periodic boundary conditions (PBC), although we will see later that this is not necessary. We now consider the action of
\begin{align}
\label{appeq:H1}
H_1=\sum^L_{i=1}\(S^x_iS^x_{i+1}+S^y_iS^y_{i+1}\)
\end{align}
on the state~\eqref{appeq:S_1d=1}, namely
\begin{align}
H_1|\mathcal S_1\rangle\propto \sum^L_{j=1} e^{i \pi j}\bigg(|\,\underset{j-1}{\underbrace{-\cdots}}\,00\underset{L-j-1}{\underbrace{-\cdots}}\,\rangle+|\,\underset{j-2}{\underbrace{-\cdots}}\,00\,\underset{L-j}{\underbrace{-\cdots}}\,\rangle\bigg).
\end{align}
We now take advantage of PBC to shift $j\to j+1$ in the sum for the second term in parentheses above, giving
\begin{align}
\label{appeq:H1S1=0d=1}
H_1|\mathcal S_1\rangle\propto \(1+e^{i\pi}\)\sum^L_{j=1} e^{i \pi j}\, |\,\underset{j-2}{\underbrace{-\cdots}}\,00\,\underset{L-j}{\underbrace{-\cdots}}\,\rangle=0.
\end{align}
Evidently, $|\mathcal S_1\rangle$ is a zero-energy eigenstate of $H_1$. This result relies on the fact that the bimagnon in the state $\mathcal S_1$ has momentum $\pi$. Indeed, replacing $\pi$ in Eq.~\eqref{appeq:S_1d=1} by $p(2\pi/L)$ for arbitrary $p\in\mathbbm Z$ does not yield an eigenstate, even at the single-bimagnon level. This indicates that we should view the $\pi$-bimagnon as a stable bound state of two magnons, which we define as isolated ``$0$"s in a background of ``$-$"s.
Eqs.~\eqref{appeq:S_1d=1}--\eqref{appeq:H1S1=0d=1} generalize readily for $d>1$, where
\begin{align}
|\mathcal S_1\rangle=\frac{1}{\sqrt{V}}\sum_{i} e^{i \bm{r}_i\cdot \bm{\pi}}\,
\bigotimes_j
\begin{cases}
\left|+\right\rangle & j=i\\
\left|-\right\rangle & \text{otherwise}
\end{cases}
\end{align}
and
\begin{align}
\label{appeq:H1-decomp}
H_1=\sum_{\langle i,j \rangle}\(S^x_iS^x_j+S^y_iS^y_j\)=\sum^d_{a=1}H^a_1,
\end{align}
where $a=1,\dots,d$ denote the principal axes of the hypercube and $H^a_1$ contains nearest-neighbor exchange terms acting only along the principal axis $a$. It then suffices to consider the action of $H^a_1$ on $|\mathcal S_1\rangle$ for each $a=1,\dots,d$. These calculations proceed identically to the $d=1$ case, with the only modification being that the prefactor in the analog of Eq.~\eqref{appeq:H1S1=0d=1} is replaced by $1+e^{i\hat{\bm{e}}_a\cdot\bm \pi}=0$, where $\hat{\bm{e}}_a$ is the unit vector in the $a$ direction.
\subsubsection{$|\mathcal S_1\rangle$ Nullifies Certain Further-Neighbor Exchanges}
\label{sec:FN-exch}
We next consider a family of further-neighbor exchange terms,
\begin{align}
\label{appeq:Hm}
H_{\bm m}=\sum_{i,j}\delta_{j,N_i(\bm m)}\(S^x_iS^x_j+S^y_iS^y_j\),
\end{align}
where $\bm m = (m_a)^{d}_{a=1}$ is a $d$-dimensional vector with integer entries and $N_i(\bm m)$ is the site obtained by translating $\bm r_i\to\bm r_i+\bm m$. The Hamiltonian $H_{\bm m}$ can be divided into a set of terms acting along parallel lines, where each lattice site $i$ belongs to exactly one line defined as the one containing the sites $i$ and $N_i(\bm m)$. Along each such line, we again have a Hamiltonian of the form \eqref{appeq:H1} acting along an effectively one-dimensional sublattice of the hypercube. Using this decomposition of $H_{\bm m}$ and an analogous decomposition of $|\mathcal S_1\rangle$, one can compute the analog of Eq.~\eqref{appeq:H1S1=0d=1} and in particular the analogous prefactor. One finds a prefactor
\begin{align}
\label{appeq:m-crit}
1+e^{i\bm m\cdot\bm{\pi}}=1+e^{i\, \pi\sum^d_{a=1}m_a}=0 \iff \sum^d_{a=1} m_a = 1\ \text{mod}\ 2.
\end{align}
We arrive at the conclusion that any exchange term $H_{\bm m}$ of the form \eqref{appeq:Hm} on a $d$-dimensional hypercube for which $\bm m$ satisfies Eq.~\eqref{appeq:m-crit} annihilates the state $|\mathcal S_1\rangle$. An example of such an exchange term term in $d=1$ is the third-neighbor hopping term used to obtain Fig.~\ref{fig:1}.
The constraint on $\bm m$ in Eq.~\eqref{appeq:m-crit} amounts to the requirement that $H_{\bm m}$ preserves the bipartiteness of the hypercubic lattice. In other words, if we label the two sublattices of the hypercubic lattice by $\mathcal A$ and $\mathcal B$, Eq.~\eqref{appeq:m-crit} implies that $H_{\bm m}$ connects only $\mathcal A$ sites to $\mathcal B$ sites, and never two sites on the same sublattice. This is because Eq.~\eqref{appeq:m-crit} essentially states that any path in the lattice connecting any two points separated by the vector $\bm m$ must traverse an odd number of links.
Combining this observation with the fact that the bimagnons are hardcore objects (see Sec.~\ref{sec:hardcore}), we expect (but have not proven) that it is possible to generalize the construction of this paper to arbitrary bipartite graphs by replacing the factor $e^{i\bm r_i\cdot\bm\pi}$ in Eq.~\eqref{eq:Jpm} with a factor $s_i=\pm 1$, where the choice of sign is based on whether the site $i$ belongs to subgraph $\mathcal A$ or $\mathcal B$.
\subsubsection{$n>1 \!:$ The Importance of Being Frustration-Free}
\label{sec:hardcore}
Having demonstrated that the state $|\mathcal S_1\rangle$ is a zero-energy eigenstate of a family of exchange Hamiltonians [including the nearest-neighbor one appearing in Eq.~\eqref{eq:H-XY}], it remains to show that the same is true for the states $|\mathcal S_n\rangle$ for $n>1$. It suffices to show that this is the case for $d=1$ since, as we have explained in Secs.~\ref{sec:NN-exch} and \ref{sec:FN-exch}, the family of exchange Hamiltonians we consider can be viewed as acting along independent one-dimensional subsystems for general $d$.
For $d=1$, we have
\begin{align}
\label{appeq:Sn}
|\mathcal S_n\rangle=\binom{L}{n}^{-\frac{1}{2}}\sum_{i_1\neq\dots\neq i_n}(-1)^{i_1+\dots+i_n}\, \bigotimes^{L}_{j=1}
\begin{cases}
\left|+\right\rangle & j\in\{i_1,\dots,i_n\}\\
\left|-\right\rangle & \text{otherwise}
\end{cases}.
\end{align}
Importantly, $|\mathcal S_n\rangle$ contains all possible permutations of the $n$ bimagnons with weights given by an alternating sign.
We first consider the nearest-neighbor exchange Hamiltonian $H_1$, which we write as a sum of bond terms,
\begin{align}
\label{appeq: H1 local}
H_1=\sum_{i}h^1_{i,i+1}.
\end{align}
We will show that $|\mathcal S_n\rangle$ is a \textit{frustration-free} zero-energy eigenstate of $H_1$, i.e. that $h^1_{i,i+1}|\mathcal S_n\rangle=0$ for all $i$.
We proceed one bond at a time. Any spin configuration in $|\mathcal S_n\rangle$ for which the spins on sites $i,i+1$ are in the $++$ or $--$ state is automatically annihilated by $h^1_{i,i+1}$. This is a manifestation of the inertness of the vacuum state $|\Omega\rangle=|-\cdots-\rangle$ and the fact that the bimagnons are hardcore objects, i.e., two of them cannot reside on the same site. This leaves the set of configurations for which the spins on sites $i,i+1$ are in the $+-$ and $-+$ states. Because Eq.~\eqref{appeq:Sn} is an antisymmetrized sum over all configurations of $n$ $``+"s$ on a background of $``-"s$, any configuration $|\mathcal C\rangle=|L\rangle\otimes|+-\rangle\otimes|R\rangle$ has a corresponding configuration $|\tilde{\mathcal C}\rangle=|L\rangle\otimes|-+\rangle\otimes|R\rangle$ appearing in $|\mathcal S_n\rangle$ with opposite sign.
(Here, $|L\rangle$ and $|R\rangle$ are \textit{fixed} configurations of $``+"$s and $``-"$s.) Since $h^1_{i,i+1}|+-\rangle=h^1_{i,i+1}|-+\rangle=|00\rangle$, $h^1_{i,i+1}$ maps both $|\mathcal C\rangle$ and $|\tilde{\mathcal C}\rangle$ to the same configuration, $|L\rangle\otimes|00\rangle\otimes|R\rangle$. Since $|\mathcal C\rangle$ and $|\tilde{\mathcal C}\rangle$ appear in $|\mathcal S_n\rangle$ with opposite signs, $h^1_{i,i+1}$ annihilates the pair of configurations $|\mathcal C\rangle$ and $|\tilde{\mathcal C}\rangle$. It immediately follows that $H_1|\mathcal S_n\rangle=0$.
Three comments are in order. First, note that the argument above demonstrates that the eigenstates $|S_n\rangle$ are insensitive to whether periodic or open boundary conditions are imposed: the difference between the two cases is simply whether or not one includes the term $h^1_{L,1}$ in $H_1$, and this term would again annihilate $|\mathcal S_n\rangle$ by the same argument. Second, note that the argument above generalizes in a straightforward manner to any exchange term connecting sites $i$ and $i+2p+1$, i.e., any odd-neighbor hopping. In this case, one considers sites a distance $2p+1$ apart, and again the configurations with $``+"$ and $``-"$ exchanged come with opposite signs. Third, note that the frustration-free nature of these eigenstates means that they remain exact eigenstates in the presence of arbitrary bond disorder in the exchange couplings.
\subsection{Other Notable Eigenstates}
We now point out the existence of additional exact eigenstates in the spectrum of $H$. Some of these states we consider ``special," while others we do not. The designation ``special" is somewhat arbitrary---for the present purposes we take it to mean that the athermal states in question contain a finite density of excitations so that the number of states in their corresponding symmetry sector grows exponentially with system size. We reserve this designation for such states because the states with a vanishing density of excitations above the fully polarized states $|-\cdots-\rangle$ and $|+\cdots+\rangle$, which we discuss in Sec.~\ref{sec:non-special}, can be smoothly connected to low-lying excited states by applying, e.g., a sufficiently large magnetic field $h$, and thus have a small number of states in their symmetry sector.
\subsubsection{Exact But Not ``Special" Eigenstates}
\label{sec:non-special}
The Hamiltonian \eqref{eq:H-XY} admits a series of single-magnon (i.e., single-``0") excitations atop the fully polarized states $|\Omega\rangle=|-\cdots-\rangle$ and $|\Omega'\rangle=|+\cdots+\rangle$. These states and their energy eigenvalues are given by
\begin{subequations}
\begin{align}
|\Omega,\bm k\rangle&=\frac{1}{\sqrt{2V}}\sum_{i} e^{i\bm{r}_i\cdot \bm k}\, S^+_i|\Omega\rangle,
\indent
E_{\Omega,\bm k}=2\sum^d_{a=1}\cos k_a-h(V-1)+D(V-1)
\\
|\Omega',\bm k\rangle&=\frac{1}{\sqrt{2V}}\sum_{i} e^{i\bm{r}_i\cdot \bm k}\, S^-_i|\Omega'\rangle,
\indent
E_{\Omega',\bm k}=2\sum^d_{a=1}\cos k_a+h(V-1)+D(V-1),
\end{align}
\end{subequations}
where the momenta take the allowed values $k_a = \frac{2\pi}{L}\, p$ ($p=0,\dots,L-1$). Their existence is guaranteed for arbitrary $U(1)$-symmetric exchange terms, as the action of such terms is simply to propagate the magnon across the lattice.
\subsubsection{Another Tower of ``Special" Eigenstates for $d=1$}
\label{sec:special}
Intriguingly, for $d=1$ there is another tower of generically finite-energy-density eigenstates consisting of effectively independent excitations with momentum $\pi$. This tower of states is given (up to normalization) by
\begin{align}
\label{appeq:S'n}
|\mathcal S^\prime_n\rangle \propto \sum_{i_1\neq i_2\neq\dots\neq i_n}(-1)^{i_1+\dots+i_n}\, (S^+_{i_1}S^+_{i_1+1})(S^+_{i_2}S^+_{i_2+1})\dots(S^+_{i_n}S^+_{i_n+1})\, |\Omega\rangle.
\end{align}
The states $|\mathcal S^\prime_n\rangle$ contain $n$ ``bond" bimagnons, which consist of \textit{pairs} of spin-flips against a polarized background, i.e., $|-\cdots00-\cdots\rangle$.
Similarly to the tower of states discussed in the main text, there are $L+1$ such states, starting from $|\mathcal S^\prime_0\rangle=|\Omega\rangle$ and ending with $|\mathcal S^\prime_L\rangle=|\Omega'\rangle$. However, apart from these lowest- and highest-weight states, the tower~\eqref{appeq:S'n} is orthogonal to the tower~\eqref{eq:scar}.
Moreover, the states in Eq.~\eqref{appeq:S'n} are less robust than those discussed in the main text: they exist only for PBC and $D=0$ in Eq.~\eqref{eq:H-XY}. The state $|\mathcal S^\prime_n\rangle$ resides in the same total magnetization sector as $|\mathcal S_n\rangle$ and has energy
\begin{align}
E^\prime_n=h(2n-V).
\end{align}
Unlike the states $|\mathcal S_n\rangle$, the states $|\mathcal S^\prime_n\rangle$ are not frustration-free eigenstates of $H$. Nevertheless, we show below that these states nullify the nearest-neighbor hopping terms in Eq.~\eqref{eq:H-XY} when PBC are imposed, and that a finite density of momentum-$\pi$ bond bimagnons behave effectively like free particles due to destructive interference of local scattering processes.
We begin with the state $|\mathcal S^\prime_1\rangle$, which can be written as
\begin{align}
|\mathcal S^\prime_1\rangle=\frac{1}{\sqrt{L}}\sum^L_{j=1} (-1)^j\, |\,\underset{j-1}{\underbrace{-\cdots}}\,00\underset{L-j-1}{\underbrace{-\cdots}}\,\rangle.
\end{align}
The nearest-neighbor exchange Hamiltonian $H_1$ [Eq.~\eqref{appeq:H1}] acts on this state as
\begin{align}
\label{appeq:H1S'1}
H_1|\mathcal S^\prime_1\rangle
&\propto
\sum^L_{j=1} (-1)^j
\bigg(
|\,\underset{j-2}{\underbrace{-\cdots}}\,0-0\underset{L-j-1}{\underbrace{-\cdots}}\,\rangle
\!+\!
|\,\underset{j-1}{\underbrace{-\cdots}}\,0-0\underset{L-j-2}{\underbrace{-\cdots}}\,\rangle
\!+\!
|\,\underset{j-1}{\underbrace{-\cdots}}+-\underset{L-j-1}{\underbrace{-\cdots}}\,\rangle
\!+\!
|\,\underset{j-1}{\underbrace{-\cdots}}-+\underset{L-j-1}{\underbrace{-\cdots}}\,\rangle
\bigg)\nonumber\\
&=
(1-1)
\sum^L_{j=1} (-1)^j
\bigg(
|\,\underset{j-1}{\underbrace{-\cdots}}\,0-0\underset{L-j-2}{\underbrace{-\cdots}}\,\rangle
\!+\!
|\,\underset{j-1}{\underbrace{-\cdots}}+-\underset{L-j-1}{\underbrace{-\cdots}}\,\rangle
\bigg)=0,
\end{align}
where in going from the first to the second line we shifted the summation indices in the first and last terms in parentheses. This demonstrates that the bond bimagnons are stable bound states at momentum $\pi$.
Next we consider $|\mathcal S^\prime_2\rangle$, which tells us about pairwise scattering of bond bimagnons. This state can be written as
\begin{align}
|\mathcal S^\prime_2\rangle
=
\binom{L}{2}^{-\frac{1}{2}}\sum_{j_1\neq j_2} (-1)^{j_1+j_2}\, |(j_1,j_1+1),(j_2,j_2+1)\rangle,
\end{align}
where $|(j,j+1)\rangle$ denotes a bimagnon on the bond $(j,j+1)$. In considering the action of $H_1$ on $|\mathcal S^\prime_2\rangle$, there are three classes of configurations to be singled out: those for which $|j_2-j_1|\equiv r_{12}=1$, $r_{12}=2$, and $r_{12}=3$. Configurations with $r_{12}>3$ can simply be handled analogously to Eq.~\eqref{appeq:H1S'1}: at fixed $j_1$, say, the terms $j_2$ and $j_2-1$ enter with opposite signs and interfere destructively. We therefore break up the state $|\mathcal S^\prime_2\rangle$ into disjoint pieces,
\begin{align}
\label{appeq:S'2-decomp}
|\mathcal S^\prime_2\rangle=|\mathcal S^\prime_2\rangle_{\leq 3}+|\mathcal S^\prime_2\rangle_{> 3}=\binom{L}{2}^{-\frac{1}{2}}\[\sum_{|j_2-j_1|\leq 3}+\sum_{|j_2-j_1|> 3} \](-1)^{j_1+j_2}\, |(j_1,j_1+1),(j_2,j_2+1)\rangle,
\end{align}
with
\begin{align}
|\mathcal S^\prime_2\rangle_{\leq 3}\propto \sum_{j_1}\Big(-|\cdots-0+0-\cdots\rangle+|\cdots-0000-\cdots\rangle-|\cdots-00-00-\cdots\rangle\Big).
\end{align}
Next, we compute the action of $H_1$ on the above configurations (keeping track of the signs for later):
\begin{subequations}
\label{appeq:H1S'2leq3}
{\scriptsize
\begin{align}
\label{appeq:H1S'2leq3a}
-H_1|\cdots-0+0-\cdots\rangle
&=
-\bigg(|\cdots-0-+0-\cdots\rangle+|\cdots-+00-\cdots\rangle+|\cdots-00+-\cdots\rangle+|\cdots-0+-0-\cdots\rangle\bigg)\\
\begin{split}
\label{appeq:H1S'2leq3b}
+H_1|\cdots-0000-\cdots\rangle
&=
+\bigg(|\cdots-0-000-\cdots\rangle+|\cdots-000-0-\cdots\rangle+|\cdots-+-00-\cdots\rangle+|\cdots--+00-\cdots\rangle\\
&
+|\cdots-0+-0-\cdots\rangle+|\cdots-0-+0-\cdots\rangle+|\cdots-00+--\cdots\rangle+|\cdots-00-+-\cdots\rangle
\bigg)
\end{split}\\
\label{appeq:H1S'2leq3c}
-H_1|\cdots-00-00-\cdots\rangle
&=-\bigg(
|\cdots-0-0-00-\cdots\rangle
+|\cdots-00-0-0-\cdots\rangle
+|\cdots-0-000-\cdots\rangle
+|\cdots-000-0-\cdots\rangle\nonumber\\
&
+|\cdots-+--00-\cdots\rangle
+|\cdots-+-00-\cdots\rangle
+|\cdots-00-+-\cdots\rangle
+|\cdots-00--+-\cdots\rangle
\bigg).
\end{align}
}%
\end{subequations}
One immediately sees that all terms in Eq.~\eqref{appeq:H1S'2leq3a} are canceled by terms in Eq.~\eqref{appeq:H1S'2leq3b} upon summation over $j_1$ and appropriate shifting of the summation indices. Similarly, the remaining terms in Eq.~\eqref{appeq:H1S'2leq3b} are canceled by terms in Eq.~\eqref{appeq:H1S'2leq3c}, so that we are left with
\begin{align}
\begin{split}
H_1|\mathcal S^\prime_2\rangle_{\leq 3}
&=
-\sum_{i}
\bigg(
|\cdots-0-0-00-\cdots\rangle
+|\cdots-00-0-0-\cdots\rangle\\
&\qquad\qquad\qquad
+|\cdots-00--+-\cdots\rangle
+|\cdots-+--00-\cdots\rangle
\bigg)
\end{split}
\end{align}
These remaining terms cancel against terms with $d_{12}=4$ in $|\mathcal S^\prime_2\rangle_{> 3}$, which enter with the opposite sign due to the factor $(-1)^{j_1+j_2}$. The remainder of the terms in $|\mathcal S^\prime_2\rangle_{> 3}$ cancel among themselves as described above Eq.~\eqref{appeq:S'2-decomp}.
The above calculation demonstrates that the pairwise scattering of bond bimagnons interferes destructively when each bimagnon has momentum $\pi$. This property allows for the construction of the many-bond-bimagnon eigenstates $|\mathcal S^\prime_n\rangle$ despite the fact that these eigenstates are not frustration-free like the site bimagnon states $|\mathcal S_n\rangle$. PBC and translation invariance are essential for this construction, as they enable the necessary cancellations to occur. Moreover, the nearest-neighbor nature of the exchange term $H_1$ is important: longer-range hopping terms including the $J_3$ term used in the main text destroy these eigenstates.\footnote{Since the $J_3$ exchange term was used in the main text to break the hidden SU(2) symmetry of Ref.~\cite{Kitazawa03}, one might wonder whether the states $|\mathcal S^\prime_n\rangle$ are simply consequences of this symmetry. However, this is not the case because these states do not persist for $D\neq 0$, while the Hamiltonian \eqref{eq:H-XY} retains its hidden SU(2) symmetry for any $D$. Moreover, the existence of the SU(2) symmetry and $|\mathcal S^\prime_n\rangle$ require different boundary conditions.} We note that although we have not strictly proven the existence of the tower states for $n>2$, we have numerically verified their presence up to $L=10$, thus giving strong evidence for the tower based on the scattering mechanism described above.
Despite their fragility relative to their counterparts $|\mathcal S_n\rangle$ studied in the main text, the states $|\mathcal S^\prime_n\rangle$ present a number of interesting questions for future work:
\begin{enumerate}
\item How does the entanglement entropy of these states scale with $n$ and $L$?
\item Does there exist a set of operators analogous to Eq.~\eqref{eq:su(2)} in the main text that can be used to understand these states?
\item Do these states exhibit off-diagonal long-range order? That is, is there a local order parameter that yields an analogue of Eq.~\eqref{eq:ODLRO} in the main text for the states $|\mathcal S^\prime_n\rangle$?
\item Do these states have any signatures in dynamics?
\item Can analogues of these states be defined for $d>1$?
\end{enumerate}
All of these questions are worthwhile subjects for future work. Intriguingly, many of these questions are also open for the AKLT bimagnon states studied in Refs.~\cite{Moudgalya18a,Moudgalya18b}; one might speculate that studying these questions for the simpler states $|\mathcal S^\prime_n\rangle$ constructed here could shed some light on the AKLT case as well.
\section{Appendix B: Entanglement of Scar States $|\mathcal S_n\rangle$}
\setcounter{equation}{0}
\renewcommand{\theequation}{B\arabic{equation}}
In this Appendix we compute the entanglement spectrum of the scar states $|\mathcal S_n\rangle$. We also show how to extract the asymptotic expression for the entanglement entropy for the case of half filling, $m_n=0$, which is quoted in Eq.~\eqref{eq:log}.
\subsection{Calculation of Entanglement Spectrum and Entanglement Entropy}
We consider the state $|\mathcal S_n\rangle$ in a system with volume $V$, which we bipartition into regions $A$ and $B$ with volumes $V_A$ and $V_B=V-V_A$. The entanglement spectrum of the state $|\mathcal S_n\rangle$ is given by the set of eigenvalues $\lambda_k$ ($k=0,\dots,K$) of its reduced density matrix,
\begin{align}
\rho_A(n)=\text{tr}_B |\mathcal S_n\rangle\langle\mathcal S_n|=\sum^{K}_{k=0}\lambda_k\, |\lambda_k\rangle\langle\lambda_k|,
\end{align}
where $|\lambda_k\rangle$ are the associated eigenvectors. We represent the state $|\mathcal S_n\rangle$ in the local $S^z$ basis as
\begin{subequations}
\begin{align}
|\mathcal S_n\rangle=\binom{V}{n}^{-\frac{1}{2}}\sum_{i_1\neq\dots\neq i_n}\sigma(i_1,\dots,i_n)\, |i_1,\dots,i_n\rangle,
\end{align}
where $i_p=1,\dots,V$ ($p=1,\dots,n$),
\begin{align}
\sigma(i_1,\dots,i_n)=e^{i(\sum^n_{p=1}\bm{r}_p)\cdot\bm{\pi}}
\end{align}
is the alternating phase factor in Eq.~\eqref{eq:Jpm}, and
\begin{align}
|i_1,\dots,i_n\rangle=\bigotimes_j
\begin{cases}
\left|+\right\rangle & j\in\{i_1,\dots,i_n\}\\
\left|-\right\rangle & \text{otherwise}
\end{cases}
\end{align}
is a product state in the local $S^z$ basis.
\end{subequations}
Because the entanglement spectra of $\rho_A(n)$ and $\rho_A(L-n)$ are related by a ``particle-hole" transformation wherein the local $S^z$-basis states $m_i=\pm 1$ are interchanged, it suffices to consider $n\leq L/2$, which implies that
\begin{align}
K=\text{min}(n,V_A).
\end{align}
To calculate the entanglement spectrum, we bipartition the state $|\mathcal S_n\rangle$ as
\begin{align}
|\mathcal S_n\rangle
=
\sum_{\{S^z_A\},\{S^z_B\}}
M_{\{S^z_A\},\{S^z_B\}}
\left|\{S^z_A\}\right\rangle
\otimes
\left|\{S^z_B\}\right\rangle,
\end{align}
where $\{S^z_{A,B}\}$ denote $S^z$-basis configurations in subsystems $A$ and $B$, respectively. The entanglement spectrum $\{\lambda_k\}$ is then given by the eigenvalues of $M^\dagger M$ (or, equivalently, $MM^\dagger$). The form of the matrix $M$ can be read off from the state $|\mathcal S_n\rangle$ when written in the following form:
\begin{align}
|\mathcal S_n\rangle
=
\binom{V}{n}^{-\frac{1}{2}}
\!
\sum^K_{k=0}
\(\sum_{i^A_1\neq\dots\neq i^A_k}
\!\!
\sigma(i^A_1,\dots,i^A_k)
|i^A_1,\dots,i^A_k\rangle\)
\!\!
\otimes
\!\!
\(\sum_{i^B_{k+1}\neq\dots\neq i^B_{n-k}}
\!\!\!\!
\sigma(i^B_{k+1},\dots,i^B_{n-k})
|i^B_{k+1},\dots,i^B_{n-k}\rangle\),
\end{align}
where $i^{A,B}_p=1,\dots,V_{A,B}$ ($p=1,\dots,k$) are defined to reside entirely within subsystems $A$ and $B$, respectively. One then finds that the matrix $MM^\dagger$ is a symmetric matrix that decomposes into a direct sum of $K$ blocks, each of which has dimension $\binom{V_A}{k}$. Each block contributes precisely one nonzero eigenvalue, namely
\begin{subequations}
\label{appeq:S_A}
\begin{align}
\label{appeq:lambda_k}
\lambda_{k}=\frac{\binom{V_A}{k}\binom{V_B}{n-k}}{\binom{V}{n}},
\end{align}
and the entanglement entropy is given by
\begin{align}
\label{appeq:S_An}
S_A(n)=-\sum^{K}_{k=0}\lambda_{k}\ln\lambda_{k}.
\end{align}
\end{subequations}
Note that
\begin{align}
\sum^{K}_{k=0}\lambda_k = \binom{V}{n}^{-1}\sum^{K}_{k=0}\binom{V_A}{k}\binom{V_B}{n-k}=\binom{V}{n}^{-1}\sum^{n}_{k=0}\binom{V_A}{k}\binom{V_B}{n-k}=1,
\end{align}
where the second equality follows from the fact that $\binom{V_A}{k}\equiv0$ for $k>V_A$ (which is necessary only when $n>V_A$ so that $K=V_A$) and the third follows from Vandermonde's identity. Thus the eigenvalues \eqref{appeq:lambda_k} are properly normalized.
\subsection{Asymptotic Behavior of Entanglement Entropy at Half Filling}
We now show how Eq.~\eqref{eq:log} follows from Eqs.~\eqref{appeq:S_A} for $V_A=V_B=V/2$ and $n=V/2$ in the limit $V\to\infty$. After substituting $V_A=V_B=n=V/2$ into Eq.~\eqref{appeq:lambda_k}, we apply Stirling's approximation,
\begin{align}
p!\approx \sqrt{2\pi p}\ e^{p\ln p-p},
\end{align}
to obtain
\begin{subequations}
\begin{align}
\lambda_k \approx \frac{e^{V\[s\(\frac{2k}{V}\)-\ln 2\]}}{\sqrt{2\pi V}\ \frac{2k}{V}\(1-\frac{2k}{V}\)},
\end{align}
where we have introduced the Shannon entropy,
\begin{align}
s(x)=-x\ln x-(1-x)\ln(1-x),
\end{align}
\end{subequations}
and where we have assumed that both $k$ and $\frac{V}{2}-k$ are of order $V$ in anticipation of the fact that terms with $k\sim V$ predominate in the sum in Eq.~\eqref{appeq:S_An}. Next, we define $q=2k/V$ and convert the sum \eqref{appeq:S_An} to an integral over $q$, obtaining
\begin{align}
\begin{split}
S_A\(\frac{V}{2}\)
&\approx
-\frac{V}{2}\int^1_0\mathrm dq\,
\frac{e^{V\[s\(q\)-\ln 2\]}}{\sqrt{2\pi V}\ q(1-q)}
\left\{V[s(q)-\ln 2]-\ln\[\sqrt{2\pi V}\, q(1-q)\]\right\}\\
&=-\frac{V}{2}\int^1_0\mathrm dq\,
\frac{1}{\sqrt{2\pi V}\ q(1-q)}\left\{\[\partial_\alpha\, e^{\alpha V\[s\(q\)-\ln 2\]}\]_{\alpha=1}-e^{V\[s\(q\)-\ln 2\]}\ln\[\sqrt{2\pi V}\, q(1-q)\]\right\}.
\end{split}
\end{align}
We proceed to evaluate the integral by saddle point. We first find the value $q_*$ such that
\begin{align}
\partial_q\ \alpha V\[s\(q\)-\ln 2\]=\alpha V\, s'(q)=0\implies q_*=\frac{1}{2},\ s(q_*)=\ln 2,
\end{align}
where the prime symbol denotes differentiation with respect to $q$. Expanding the argument of each exponential around $q_*$ to leading order and noting that $s''(q_*)=-4$, we obtain
\begin{align}
\label{appeq:S_A-intermediate}
S_A\(\frac{V}{2}\)
&\approx
-\sqrt{\frac{2V}{\pi}}
\int^1_0\mathrm dq\,
\left\{\[\partial_\alpha\, e^{-2\alpha V\(q-\frac{1}{2}\)^2}\]_{\alpha=1}-\frac{1}{2}e^{-2 V\(q-\frac{1}{2}\)^2}\ln\(\frac{\pi V}{8}\)\right\},
\end{align}
where we have used the fact that the convexity of $s(q)$ renders the integrand sharply peaked around $q_*$.
It remains to evaluate the Gaussian integral
\begin{subequations}
\begin{align}
\int^1_0\mathrm dq\, e^{-2\alpha V\(q-\frac{1}{2}\)^2}=\sqrt{\frac{\pi}{2\alpha V}}
\end{align}
and its derivative,
\begin{align}
\partial_\alpha\int^1_0\mathrm dq\, e^{-2\alpha V\(q-\frac{1}{2}\)^2}=\partial_\alpha\sqrt{\frac{\pi}{2\alpha V}}=-\frac{1}{2}\sqrt{\frac{\pi}{2V}}\ \alpha^{-3/2},
\end{align}
\end{subequations}
which we substitute into Eq.~\eqref{appeq:S_A-intermediate} to obtain
\begin{align}
S_A\(\frac{V}{2}\)
&\approx
-\sqrt{\frac{2V}{\pi}}
\[-\frac{1}{2}\sqrt{\frac{\pi}{2V}}-\frac{1}{2}\sqrt{\frac{\pi}{2V}}\ln\(\frac{\pi V}{8}\)\]=\frac{1}{2}\[\ln\(\frac{\pi V}{8}\)+1\].
\end{align}
This is precisely Eq.~\eqref{eq:log}.
As a consistency check, note that the above large-$V$ analysis maintains the normalization of the eigenvalues $\lambda_k$. Expanding around the same saddle point, we find
\begin{align}
\sum^{V/2}_{k=0}\lambda_k\underset{V\to\infty}{\longrightarrow}\sqrt{\frac{2V}{\pi}}\int^1_0\mathrm dq\, e^{-2V\(q-\frac{1}{2}\)^2}=1,
\end{align}
as desired.
\section{Appendix C: Relationship to Embedded Hamiltonians}
\setcounter{equation}{0}
\renewcommand{\theequation}{C\arabic{equation}}
The results of this paper have an interesting connection to the method of embedded Hamiltonians developed in Ref.~\cite{Shiraishi17} and applied to quantum many-body scars (QMBS) in Refs.~\cite{Choi18,Ok19,Shiraishi19}. In this method, one chooses a set of projection operators $P_i$ acting near site $i$ which define the target space
\begin{align}
\mathcal T = \big\{|\psi\rangle \mid P_i\, |\psi\rangle = 0\ \forall\ i\big\}.
\end{align}
Next, one defines a Hamiltonian of the form
\begin{align}
\label{appeq:H-E}
H_\mathrm{E}=\sum_i P_i\, h_i\, P_i + H',
\end{align}
where $h_i$ acts near site $i$ and $[H',P_i]=0$ for all $i$. Since the first term in Eq.~\eqref{appeq:H-E} annihilates any state in $\mathcal T$, $H_\mathrm{E}$ possesses a number of eigenstates equal to $\text{dim }\mathcal T$ that are also eigenstates of $H'$. When $H'$ possesses lightly entangled eigenstates that also belong to $\mathcal T$, the Hamiltonian~\eqref{appeq:H-E} hosts strong-ETH-violating eigenstates by construction.
It is natural to suspect that the Hamiltonian $H$ in Eq.~\eqref{eq:H-XY} can be re-expressed in the form~\eqref{appeq:H-E}: the states $|\mathcal S_n\rangle$ are annihilated by the exchange term $H_1$ (as shown explicitly in Sec.~\ref{sec: Appendix A}) and are also eigenstates of the $D$ and $h$ terms, which we can identify with $H'$. We can thus rewrite Eq.~\eqref{eq:H-XY} in the form \eqref{appeq:H-E} if we can recast $H_1$ in terms of appropriate projectors $P_i$ and local Hamiltonian terms $h_i$. In Sec.~\ref{sec:P} we identify a set of such operators, but show that the resulting embedding is \textit{tautological} in the sense that the null spaces of $h_i$ and $P_i$ coincide---in other words, one can simply replace the first term in Eq.~\eqref{appeq:H-E} by $\sum_i h_i=H_1$. The embedding picture thus does not add anything to our understanding of the states $|\mathcal S_n\rangle$ for the model studied in the main text. Nevertheless, we show in Sec.~\ref{sec:algebraic} that the projectors $P_i$ have an appealing interpretation in terms of the emergent SU(2) algebra \eqref{eq:su(2)}. We also show in Sec.~\ref{sec:perturbations} that we can use these projectors to construct a class of perturbations to Eq.~\eqref{eq:H-XY} that preserve the scarred eigenstates $|\mathcal S_n\rangle$.
\subsection{Identifying the Projector}
\label{sec:P}
We now identify the desired operators $P_i$ and $h_i$.\footnote{We thank C.-J.~Lin for bringing to our attention the existence of the operator $P_{i,j}$. The present derivation and subsequent interpretation of this operator are our own.}
We write the nearest-neighbor exchange Hamiltonian $H_1$ as a sum of local nearest-neighbor bond terms,
\begin{align}
H_1 = \sum_{\langle ij\rangle}h^1_{i,j}, \indent h^1_{i,j}=\frac{1}{2}(S^+_iS^-_j+S^-_iS^+_j).
\end{align}
The local Hilbert space for the two sites $i,j$ is nine-dimensional.
In Sec.~\ref{sec:hardcore} we showed that the null space of $h^1_{i,j}$ is spanned by the three states $|\pm\pm\rangle$ and $|+-\rangle-|-+\rangle$. We want to define a projection operator $P_{i,j}$ that annihilates these three states and acts trivially on the remaining six. We claim that the operator
\begin{align}
\label{appeq:P explicit}
P_{i,j}
=
1
-\frac{3}{4}(S^z_i)^2(S^z_j)^2
+\frac{1}{8}\Big[(S^+_i)^2(S^-_j)^2+(S^-_i)^2(S^+_j)^2\Big]
-\frac{1}{4}S^z_iS^z_j
\end{align}
achieves this. First let us check the action on the six states not in the null space of $h^1_{i,j}$. We have
\begin{align}
P_{i,j}|a\rangle=|a\rangle \indent \text{for} \indent |a\rangle = |00\rangle,\ |0\pm\rangle,\ |\pm0\rangle,
\end{align}
since the non-identity terms in Eq.~\eqref{appeq:P explicit} annihilate any state with a ``0" on site $i$ or $j$. Furthermore, we have
\begin{align}
\begin{split}
\frac{1}{8}\Big[(S^+_i)^2(S^-_j)^2+(S^-_i)^2(S^+_j)^2\Big]\big(|+-\rangle+|-+\rangle\big)
&= \frac{1}{2}\big(|+-\rangle+|-+\rangle\big),\\
\Big[1
-\frac{3}{4}(S^z_i)^2(S^z_j)^2
-\frac{1}{4}S^z_iS^z_j\Big]
\big(|+-\rangle+|-+\rangle\big)
&= \bigg(1-\frac{3}{4}+\frac{1}{4}\bigg)\big(|+-\rangle+|-+\rangle\big)\\
&= \frac{1}{2}\big(|+-\rangle+|-+\rangle\big),
\end{split}
\end{align}
which delivers
\begin{align}
P_{i,j}\big(|+-\rangle+|-+\rangle\big)=\big(|+-\rangle+|-+\rangle\big).
\end{align}
Next, we check the action on the three states in the null space of $h^1_{i,j}$. First, we have
\begin{align}
P_{i,j}|b\rangle=\bigg(1-\frac{3}{4}-\frac{1}{4}\bigg)|b\rangle=0 \indent \text{for} \indent |b\rangle = |\pm\pm\rangle.
\end{align}
We then note that
\begin{align}
\begin{split}
\frac{1}{8}\Big[(S^+_i)^2(S^-_j)^2+(S^-_i)^2(S^+_j)^2\Big]\big(|+-\rangle-|-+\rangle\big)
&= -\frac{1}{2}\big(|+-\rangle-|-+\rangle\big),\\
\Big[1
-\frac{3}{4}(S^z_i)^2(S^z_j)^2
-\frac{1}{4}S^z_iS^z_j\Big]
\big(|+-\rangle-|-+\rangle\big)
&= \bigg(1-\frac{3}{4}+\frac{1}{4}\bigg)\big(|+-\rangle-|-+\rangle\big)\\
&= \frac{1}{2}\big(|+-\rangle-|-+\rangle\big),
\end{split}
\end{align}
which delivers
\begin{align}
P_{i,j}\big(|+-\rangle-|-+\rangle\big)=0.
\end{align}
Since $P_{i,j}$ and $h^{1}_{i,j}$ share the same null space, we have
\begin{align}
P_{i,j}h^{1}_{i,j} = h^{1}_{i,j}P_{i,j} = h^{1}_{i,j}.
\end{align}
We may thus trivially rewrite
\begin{align}
H_1=\sum_{\langle ij\rangle}P_{i,j}\, h^{1}_{i,j}\, P_{i,j},
\end{align}
so that Eq.~\eqref{eq:H-XY} becomes Eq.~\eqref{appeq:H-E} with $P_i=P_{i,j}$, $h_i=h^{1}_{i,j}$, and $H'=h\sum_i S^z_i+D\sum_i \left(S^{z}_i\right)^2$. Note that $[H',P_{i,j}]=0$ because $H'$ does not couple the null space of $P_{i,j}$ and its complement.
\subsection{Algebraic Interpretation}
\label{sec:algebraic}
The projector~\eqref{appeq:P explicit} has an appealing interpretation in terms of the SU(2) algebra~\eqref{eq:su(2)}. To see it, we define the local \textit{staggered} SU(2) generators
\begin{align}
J^\pm_{i}
=\frac{1}{2}\times
\begin{cases}
-(S^+_i)^2 & i\in \text{sublattice } A\\
+(S^+_i)^2 & i\in \text{sublattice } B
\end{cases}\ ,
\qquad
J^z_i = \frac{1}{2}\, S^z_i,
\end{align}
which satisfy
$\sum_i J^\pm_i = J^\pm$ and $\sum_i J^z_i=J^z$, where the latter operators are the generators in Eq.~\eqref{eq:su(2)}. In terms of these generators, Eq.~\eqref{appeq:P explicit} becomes
\begin{align}\label{appeq:P-su(2)}
P_{i,j}
=
\frac{3}{4}\bigg(1
-(S^z_i)^2(S^z_j)^2\bigg)
+\bigg(\frac{1}{4}
-\bm J_i\cdot\bm J_j\bigg),
\end{align}
where we have used the fact that sites $i$ and $j$ belong to different sublattices owing to the bipartite nature of the nearest-neighbor exchange.
The first term above vanishes for any configuration with $m_i^2=m_j^2=1$; in particular, it vanishes when acting on the scarred eigenstates $|\mathcal S_n\rangle$. The second term above projects onto the \emph{local} staggered SU(2) singlet state between sites $i$ and $j$. Thus, when acting on a state in the $m_i^2=1$ subspace in which the scarred states reside, $P_{i,j}$ annihilates any state in the \emph{global} SU(2) representation with maximal spin $j=V/2$ (a state with maximal total spin cannot contain any local singlets). Conversely, $P_{i,j}$ acts as the identity on any state with $m_{i,j}=0$ for one or both of $i$ and $j$ due to the fact that $\bm J_i\cdot\bm J_j$ annihilates such states.
The above rewriting of $P_{i,j}$ thus makes explicit the fact that the states $|\mathcal S_n\rangle$ form components of a $(V+1)$-state ``macrospin." In this way, the model studied in the main text resembles the toy model proposed in Ref.~\cite{Choi18} that has perfect QMBS. However, it is worth noting that while the projectors $P_{i,j}$ are superfluous for the model studied in the main text (as mentioned elsewhere in this Appendix), they are not for the model proposed in Ref.~\cite{Choi18}. There, the analogous projectors themselves enforce the presence of scarred eigenstates and no structure for the Hamiltonian terms analogous to $h_i$ in Eq.~\eqref{appeq:H-E} is assumed, while in the example presented here these states exist regardless of the presence or absence of the projectors $P_{i,j}$ due to the structure of the local exchange terms $h^1_{i,j}$.
\subsection{A Class of Perturbations Preserving the Scar States}
\label{sec:perturbations}
Despite the fact that the projectors defined in Eq.~\eqref{appeq:P explicit} are superfluous for the XY model defined Eq.~\eqref{eq:H-XY}, they are useful in identifying a class of perturbations to Eq.~\eqref{eq:H-XY} that preserve the scarred eigenstates $|\mathcal S_n\rangle$. This more general class of Hamiltonians takes the form
\begin{align}
\label{appeq:general}
H = H_1 + H' + \sum_{i\in A,j\in B}P_{i,j}\, h'_{i,j}\, P_{i,j},
\end{align}
where $h'_{i,j}$ is \emph{any} Hamiltonian coupling sites $i$ and $j$ and $H'$ is \emph{any} Hamiltonian that commutes with all the $P_{i,j}$ appearing in the sum. Note that the sum above is not restricted to nearest-neighbors, to emphasize that couplings of arbitrary range are allowed as long as the two sites being connected belong to different sublattices. The projector $P_{i,j}$ annihilates any state in the spin-$V/2$ representation of the algebra~\eqref{eq:su(2)} because such states cannot contain staggered-SU(2) singlets between \emph{any} two spins on different sublattices. Eq.~\eqref{appeq:general} suggests that the XY Hamiltonian $H_1$ can be viewed as a ``parent Hamiltonian" for a class of embedded-Hamiltonian models.
\section{Appendix D: Spin-$S$ Models with Exact Scarred States}
\label{sec:spin-s}
\setcounter{equation}{0}
\renewcommand{\theequation}{D\arabic{equation}}
In this Appendix we generalize the spin-1 XY model to achieve exact QMBS for microscopic degrees of freedom $\bm S_i$ with arbitrary spin $S$ [i.e., $\bm S_i\cdot\bm S_i=S(S+1)$]. The generators of the emergent SU(2) algebra \eqref{eq:su(2)} are
\begin{subequations}
\label{appeq:su(2)-s}
\begin{align}\label{appeq:su(2)-s-a}
J^{\pm}&=\sum_i J^\pm_i=\frac{1}{(2S)!}\sum_{i}e^{i\boldsymbol{r}_i\cdot\boldsymbol{\pi}}\left(S^\pm_i\right)^{2S},\\
J^z&=\sum_i J^z_i=\frac{1}{2\left((2S)!\right)^2}\sum_i \left[(1+S-S^z_i)_{2S}(1-S+S^z_i)_{2S}-(1+S+S^z_i)_{2S}(1-S-S^z_i)_{2S}\right],
\end{align}
where we have used the notation
\begin{align}\label{appeq:su(2)-s-b}
(x)_{q}=x(x+1)\dots(x+q-1)
\end{align}
\end{subequations}
for the Pochhammer symbol (or ``rising factorial").
Despite its complicated appearance, $J^z_i$ has a simple action on local $z$-basis states, namely
$J^z_i|m_i=\pm S\rangle=\pm(1/2)|m_i=\pm S\rangle$ and $J^z_i|m_i\neq\pm S\rangle=0$. When every site is in the $\pm S$ state, the spectrum of $J^z$ is that of a spin-$V/2$ degree of freedom, i.e., $-V/2,-V/2+1,\dots,V/2$. When all but $p$ sites are in the $\pm S$ state, the ``total spin" is effectively lowered by $p/2$, so that $J^z$ has the spectrum $-(V-p)/2,\dots,(V-p)/2$.
For $n=0,\dots, V$ we define the states
\begin{align}\label{appeq:scar-s}
|\mathcal{S}^{(S)}_n\rangle=\mathcal N(n)\left(J^+\right)^n|\Omega\rangle,
\end{align}
where $|\Omega\rangle=\bigotimes_i|m_i=-S\rangle$ is the all-down state and $\mathcal N(n)$ is defined below Eq.~\eqref{eq:scar} in the main text. States with different $n$ are distinguished by their $J_z$ eigenvalues
\begin{align}\label{appeq:m-s}
m^{(S)}_{n}=n-V/2.
\end{align}
Combining Eqs.~\eqref{appeq:su(2)-s}--\eqref{appeq:m-s}, one verifies that the states $|\mathcal{S}^{(S)}_n\rangle$ form a spin-$V/2$ representation of Eq.~\eqref{eq:su(2)}, i.e.
\begin{align}
\bm J\cdot \bm J\,|\mathcal S^{(S)}_n\rangle=\frac{V}{2}\left(\frac{V}{2}+1\right)|\mathcal S^{(S)}_n\rangle.
\end{align}
We find that the states $|\mathcal S^{(S)}_n\rangle$ are annihilated by the following family of Hamiltonians,
\begin{equation}\label{appeq:H-s}
H^{(S)}_p(i,j)=\left(S^+_i\right)^{2S-p}\left(S^-_j\right)^{2S-p}+\left(\frac{(2S-p)!}{p!}\right)^2 \left(S^+_i\right)^{p}\left(S^-_j\right)^{p}+{\rm H.c.},
\end{equation}
where the sites $i,j$ belong to opposite sublattices and $p=1,\dots,S$ for $S\in\mathbbm Z$ while $p=1,\dots,S-\frac{1}{2}$ for $S\in\mathbbm Z+1/2$. We note that for $S=1$ only $p=1$ is allowed and this is exactly the XY Hamiltonian $H_1$ (up to rescaling). For any $S,p$ one may add the total field $h\sum_i S^z_i$ and anisotropy $D\sum_i \left(S^z_i\right)^2$ to the Hamiltonian, giving the scarred states the energy spectrum $E^{(S)}_n=2S\, h(n-V/2)+S^2DV$.
The models defined by Eqs.~\eqref{appeq:su(2)-s}-\eqref{appeq:H-s} admit a trivial embedding structure similar to the one discussed in Sec.~\ref{sec:algebraic}. In this case, Eq.~\eqref{appeq:P-su(2)} is modified as
\begin{align}
P^{(S)}_{i,j}
=
\frac{3}{4}\bigg[1
-\big(2J^z_i\big)^2\, \big(2J^z_j\big)^2\bigg]
+\bigg(\frac{1}{4}
-\bm J_i\cdot\bm J_j\bigg),
\end{align}
where as before the first term is necessary to ensure that $P^{(S)}_{i,j}|m_{i}m_{j}\rangle=|m_{i}m_{j}\rangle$ when one or both of $m_{i,j}\neq\pm S$ and the second term projects onto the singlet of the local staggered SU(2) algebra corresponding to Eqs.~\eqref{appeq:su(2)-s} when $m_{i,j}=\pm S$.
\end{widetext}
\end{document}
|
\section{Introduction}
The link between transport related emissions, and human health, is a major issue for city municipalities worldwide. Diesel and other internal combustion engine based motor vehicles are considered to be the major culprit in this regard as they are associated with the generation of a number of harmful emissions. Apart from the link to global warming through the generation of carbon-dioxide, such vehicles are also known to produce other airborne pollutants such as nitrogen oxides, ozone, benzene, carbon monoxide, and particulate matter (PM) of varying size, all of which are considered harmful to human health. In a recent global review~\cite{chest_review}, it is stated that air pollution, in general, could be damaging every organ and every cell in the human body, showing a potential link between toxic air and skin damages, fertility, asthma and allergies to children and adults.\newline
One of the main reasons behind all this is considered to be PM emissions. PM is a generic term used for a type of pollutants that consist of a complex and varied mix of particles suspended in air. Among all the airborne pollutants PM is particularly worrying due to its ability to enter the bloodstream and reach major organs in the human body. There is rich literature documenting the link between PM and its effects on human health~\cite{valavanidis,gehring,eea2014,harvard_2006,air_survey_2014,pm_brazil_2011}. In particular, the World Health Organization reports that ``adverse health effects of PM are due to exposure over both short (hours, days) and long (months, years) terms and include respiratory and cardiovascular morbidity (aggravation of asthma, respiratory symptoms, increase in hospital admissions), as well as mortality from cardiovascular and respiratory diseases and from lung cancer''~\cite{who}. Smaller PM particles tend to be more harmful to humans compared to larger ones, as they can travel deeper into the respiratory system~\cite{who,eea2014}. Some of the health effects related to PM include oxidative stress, inflammation and early atherosclerosis. Other studies have shown that smaller particles may go into the bloodstream and thus translocate to the liver, the kidneys or the brain (see~\cite{non_exhaust} and references within). Transport related emissions are a significant contributor to airborne PM levels that harm our health. In a recent study~\cite{dementia}, it is shown that living near major roads (i.e. near emissions from vehicles) is associated with increased risk of dementia. The reduction of air quality and population exposure to harmful pollutants as a result of road passenger transportation is discussed in~\cite{toronto_2010}, in a case study in the Greater Toronto Area. The survey described in~\cite{air_survey_2014} focuses on air pollution originating from non-exhaust emissions such as brake and tyre wear, and highlights the related impact to human health as well as the significance of particulate matter reduction.\newline
Roughly speaking, three avenues are being explored worldwide in the fight against urban pollution:
\begin{itemize}
\item[(i)] outright bans on polluting vehicles and embracing zero tailpipe emission vehicles in certain city zones;
\item[(ii)] measuring air quality as a means to better informing citizens of zones of higher pollution \cite{francesco}; and
\item[(iii)] developing smart mobility devices that seek to minimise the effect of polluting devices on citizens as they transport goods and individuals in our cities \cite{annie,hermann,shaun}.
\end{itemize}
Option (i) whereby ultra-low emission zones are created by banning internal combustion engine (ICE) based vehicles in certain areas, in addition to embracing electric vehicles (EV's), has gained much traction worldwide and is being proposed for adoption in cities such as London and Dublin. Apart from the reduced tailpipe pollutants, an additional attraction of the switch from ICE to EV, is that it is beneficial from the perspective of global warming (reduced carbon dioxide), provided that the energy delivered to the EV's can be sourced in a green manner. Thus, reducing our dependency of ICE based vehicles would appear to be very beneficial; not only does the strategy achieve cleaner air but we also potentially tackle climate change through reduced production of carbon dioxide.
A major objective of this paper is to challenge the current focus on tailpipe emissions. To avoid any {\em misunderstanding}, we wholeheartedly endorse a reduced dependency on ICE based vehicles. However, the contemporary narrative is based on tailpipe emissions only, and while
it is indeed true that EV's are zero tailpipe emission vehicles, the tailpipe is only one source of PM. Thus replacing one type of vehicle fleet with another type of vehicle fleet may not result in cities with safe levels of air quality, especially if the non-tailpipe sources of PM are significant. One contribution of this work is to use data from Dublin to argue that PM levels from tyres alone may be above that which is deemed safe by the World Health Organisation (WHO). While we are by no means the first to argue that tyres are an important source of PM (see in particular the excellent report~\cite{microplastics}), we strongly believe it is important that stake-holders be reminded of this message, particularly in the context of the current transitioning from ICE-vehicles towards an electrification of the vehicle fleet. The second part of our paper describes a distributed access control mechanism that both regulates tyre-based PM generation, and provides fair access to a city zone for a set of competing vehicles.\newline
The paper is organised as follows. In Section~\ref{sec:pm_emissions} we discuss tyre-wear related PM emissions, supported by numbers from various sources, with a focus on Dublin in Ireland. We review traffic produced air-pollution mitigation measures in Section~\ref{sec:related}. Our Access Control Mechanism is presented in detail in Section~\ref{sec:acs}. We simulate our system and present results in Section~\ref{sec:simulations} and conclude the paper in Section~\ref{sec:conclusion}.
\section{Elementary calculations}\label{sec:pm_emissions}
{\bf Particulate Matter: } PM is the product of brake and tyre wear from vehicles as well as a by-product of the engine combustion process. The most common classification of particulate matter is according to size: $P\!M_{10}$ for particles with at most $10\ \mu m$ diameter, $P\!M_{2.5}$ for particles with at most $2.5\ \mu m$ diameter, and ultrafine particles which have a diameter of less than 0.1 micrometres.
Smaller PM particles tend to
be more harmful compared to larger ones as they are able to get deeper into the respiratory system
with ultrafine particles being able to get into the bloodstream and therefore translocate into vital
organs such as the liver, the kidneys and the brain.\newline
{\bf Designated safe levels of PM: } According to the WHO, for $P\!M_{2.5}$, the daily maximum deemed safe level on average is $25\ \mu g/m^3$, whereas the annual maximum permitted level is on average $10\ \mu g/m^3$. For $P\!M_{10}$, the maximum permitted levels are on average $50\ \mu g/m^3$ and $20\ \mu g/m^3$ on a daily and annual basis, respectively. In general, non-exhaust emissions (including brake and tyre wear, road surface wear and resuspension of road dust) resulting from road traffic, account for over 90\% of $P\!M_{10}$ and over 85\% of $P\!M{2.5}$ emissions from traffic~\cite{non_exhaust_electric}.\newline
{\bf Approximate guess of airborne PM in Dublin:} To parse these numbers in terms of tyre abrasion for a city with a high volume of cars, we note that in 2014, nearly 28,000 tonnes of tyre waste was managed in Ireland \cite{waste_tyres_stats}. Using publicly available data from the Central Statistics Office~\cite{cso_data} in Ireland, in 2018 approximately 540,000 private cars were continuously active in Dublin throughout the year with an average distance travelled of approximately 15,000 km per vehicle. We assume that approximately 1/3 of the vehicles (i.e. c. 170,000 vehicles) will change their tyres in a year in Dublin\footnote{A tyre is changed when it has reached a tread wear of approximately 2 mm (or 1.6 mm as is the legal minimum). This translates into approximately 35,000 km of travelled distance per vehicle; however, depending on the driving conditions, the travelled distance before a tyre is changed can vary from 10,000 km (harsh braking and acceleration, constant change of gears) to 80,000 km (perfect driving conditions and favourable road and weather).}. Depending on the type of the tyre and the road conditions, a vehicle (i.e, 4 tyres) loses $50-240 \ mg/km$ in mass~\cite{grigoratos}, which accounts for 4-6 kg of tyre mass lost before tyres are changed. By considering 4 kg of tyre mass lost per vehicle, we estimate that in Dublin, in 2018, at least 680,000 kg of tyre mass was wasted, 10\% of which goes airborne~\cite{microplastics,grigoratos} as PM. That corresponds to approximately 68,000 kg of particulate matter in a year, or 185 kg per day, in the city of Dublin. \cite{grigoratos} states that approximately 50\% of the $P\!M_{10}$ emissions (not specifically to air) fall in the $P\!M_{2.5}$ category. \cite{fausser} reports that c.90\% of airborne tyre wear particles are smaller than 1 micrometer in diameter (that is, in the $P\!M_{2.5}$ category). In~\cite{non_exhaust}, references of previous studies state that 3-7\% of tyre wear particles contribute to airborne $P\!M_{2.5}$.\newline
{\bf Rubber in road dust:} Finally, to determine the presence of tyre residuals in road dust in an urban environment, and thereby provide an experimental estimate of the relative proportion of rubber vs. non-rubber particles, road dust samples were collected from seven locations in central London: Brechin Place, Jay Mews, Exhibition Road, Bayswater Road, by Queensgate Station, Cromwell Road and Westway Roadprotect\footnote{Data collected at Imperial College London by S. Anderson, M . Mallya, H. Richardson, C. Ching.}. These locations include major thoroughfares as well as residential streets and an 'average' urban road dust sample was subsequently created by mixing these various collected samples. In order to measure the amount of tyre dust we employed thermogravimetric analysis \cite{Coats}, to measure mass change with temperature increase. The results of this analysis on the aggregated road sample are depicted in Figure \ref{fig:tga}. Tyre rubber ignites at approximately $ 700^{\circ} C$, and at this temperature the thermogravimetric curve shows a drop of $6.6\%$. This indicates that the mass fraction of rubber (or a similar material) in the collected road dust is approximately $6.6\%$.
\begin{figure}
\begin{center}
\includegraphics[scale=.3]{RCA}
\label{fig:tga}
\caption{Thermogravimetric analysis of road dust aggregated over all sites.}
\end{center}
\end{figure}
\section{Air pollution mitigation measures}\label{sec:related}
Among the early solutions to reduce traffic related air-pollution has been the application of non-thermal plasma to diesel cars~\cite{non-thermal_2003}. Similar solutions include the application of catalytic filters~\cite{diesel_filter} for reduced exhaust fumes. Such solutions however, fail to address the non-exhaust emissions from diesel and non-diesel vehicles. The potential of road sweeping and washing to reduce non-exhaust related emissions was presented in a study in the Netherlands in 2010~\cite{netherlands_2010}. The authors, although they identify non-exhaust emissions as the main source for coarse PM in urban areas, conclude that their approach does not have a significant reduction in non-exhaust emissions. The benefits of ride-sharing to the environment have been discussed in various studies, such as~\cite{ride_sharing_2010, agent_based, beijing_sharing, carpooling}. However, these studies do not take a dedicated interest to non-exhaust emissions, but rather, to fuel consumption reduction. Fuel consumption reduction has been addressed with route suggestion solutions in~\cite{truck_route_2011, time_fuel_2013,fleet_fuel_2014}, for trucks and vehicle fleets. In~\cite{time_fuel_2013}, the authors present a linear programming solution to the Time-Dependent Pollution-Routing Problem. Fleets of vehicles are re-routed depending on traffic, and speeds are recommended based on emissions, driver costs, traffic and peak hour information. As a solution, the authors introduce a departure time and speed optimization algorithm. A similar approach for optimisation of fleet size is proposed in~\cite{fleet_fuel_2014}. In the same spirit, authors in~\cite{its_2013} study a variety of measures, such as traffic control, ban of heavy duty vehicles (HDV) and speed restriction, in order to achieve reduction of traffic related emissions. Traffic control (simulated simply by reducing traffic by 20\%) and HDV banning have a significant reduction in air-pollutants (20\% to 23\%), whereas speed control exhibits increase in PM emissions, due to HDV. Last but not least, use of electric vehicles, as an alternative to diesel and petrol ones, has been suggested for the reduction of traffic related air pollutants. In a feasibility study in Canada and Italy~\cite{team_play_2016}, the use of electric cars and electric motorcycles shows a reduction in CO$_2$ emissions, however, the study ignores non-exhaust related emissions, which are as relevant to electric vehicles as they are to diesel and petrol ones~\cite{non_exhaust_electric}. Another study in the city of Dublin ~\cite{dublin_ev_2011} uses home/work commute and traffic related data to study a number of electric vehicle market penetration scenarios and evaluates the emission decrease under each of them. However, only tailpipe emissions are taken into consideration, again overlooking brake and tyre wear and other non-exhaust emissions. As opposed to the majority of works that address the reduction of road traffic related emissions, we propose a traffic control and ride-sharing scheme, that reduces the amount of cars in the streets, and therefore the tyre-related emissions, as well as other non-exhaust and exhaust emissions.
\section{Feedback-enabled Access Control}\label{sec:acs}
The pollution mitigation mechanisms discussed in the previous section, and the move from ICE to EV's that is so popular in many cities globally, is based on the assumption that the principal source of pollution is tailpipe in origin. As we have discussed in the previous section this assumption is at best only partially true and tyres, brakes, as well as road abrasion, may contribute significantly to PM generation.
Additionally, EV's are generally heavier than ICE vehicles, potentially affecting tyre wear negatively. This means that the amount of emitted non-exhaust PM might actually even be elevated for EVs. Therefore, one must look for alternative mitigation mechanisms to combat these sources of PM generation. Apart from the obvious move from private to public transport or other modes of transport such a cycling and scooters, the only real viable mechanism is to develop an access control mechanism that is based on a feedback control strategy to regulate the safe levels of PM. It is one such strategy that we now develop.\newline
Specifically, our objective is to maximise both the number of cars and people entering the city centre each day, while maintaining the tyre-generated PM emission levels significantly below the maximum permitted levels. The idea is to orchestrate an access control scheme so that it encourages ride-sharing. The access mechanism works in a simple way: at each day passengers are assigned to cars (drivers) through a matching method. Then, cars who want to have access to the city center are picked randomly using a probabilistic method that ensures fairness and privacy to each user and which is based on occupancy. See Figure~\ref{fig:Scheme}, for a visual depiction of this scheme.\newline
The rationale behind the choice of a probabilistic method instead of a deterministic one, like a water-filling algorithm, lies in the fact that the latter can be quite inefficient from the single user perspective. In order to use an access control scheme, an agent (driver or passenger) would typically buy a monthly or yearly access pass. This ticket provides them with the opportunity of competing with other users to access the city, either as a driver or as a passenger. Consider now the example of parents, that have to take their children to school outside the city centre before travelling into the city centre in the morning: even though they paid the same amount for a monthly or yearly parking ticket as everyone else, in a deterministic system they always have a greater chance of missing out the chance of having access to the center of the city, as they arrive later than everyone else. Using a probabilistic system, as the one described in \cite{QoS}, we are able to guarantee equality in regards to access for all users over the long-term period of validity of their pass, irrespective of their constraints. \newline
For this Access Control Method, we assume that the controlled region (referred to as $R$) can accommodate up to $N$ vehicles per day, decided so that the tyre-related PM emissions are kept at low levels, ensuring thus that in general PM emission levels will remain low. There are mainly two challenges to make this method work efficiently.\newline
\begin{itemize}
\item[Q1 ] {\em Compliance:} How does one make sure that users comply with the matchmaking scheme, after access has been granted? \newline
\item[Q2 ] {\em Fair access:} How does one ensure that each driver is granted access to $R$ fairly with respect to other users (for instance, keeping the amount of average access the same among all cars)?\newline
\end{itemize}
We answer these questions in detail in the following subsections.
\begin{figure}
\includegraphics[width=1\columnwidth]{Algorithm.pdf}
\caption{Block model of the access control algorithm. First cars and passengers (respectively sets $C$ and $P$) get matched by a matchmaking algorithm. Once this process is over, each car and the corresponding passengers (set $M$), are randomly picked by a probabilistic method. The resulting set of cars, A represents the cars and passengers who are granted access to the city for the day.}
\label{fig:Scheme}
\end{figure}
\subsection{Ride-Sharing Compliance}
In a Ride-Sharing scheme, as the one described above, one of the crucial elements to make the architecture work is to ensure that both drivers and passengers comply with the matchmaking system. If users are not somehow punished for negative behaviour, they might be inclined to cheat the system to maximise their own personal advantage, which in turn might lead to sub-optimal results and to a poor Quality of Service (QoS) overall. As an example, in order to increase the probability of gaining access, a driver might accept as many drivers as possible and then refuse to pick them up; on the other hand, a passenger might choose to not show up, effectively wasting time and resources (the assigned seat). In this context, on the basis of the work done in \cite{DLTAndSocial} we propose the use of a digital token as a bond, or digital deposit, to ensure that passengers and drivers comply with their respective social contract (the matchmaking system). The risk of losing a token is then the mechanism that encourages agents to comply with these social contracts. There are multiple practical ways to implement this system: a possible example could be to have each user equipped with a digital wallet and the only way to participate to the matchmaking system is to have enough tokens to use as a bond. Another way could be to link the tokens to real money, so that losing a certain amount of them would result in a real economic loss for the agent. Note that the pricing of such tokens is beyond the scope of this paper and is dealt with in \cite{DLTAndSocial}. \newline
The simple idea is that, whenever a passenger is matched with a driver, they both agree on a specific \emph{pick up} point and on a time window. Once the passenger gains access to the city center, all the agents involved 'deposit' a \emph{token} to the designed pick up point (notice that this process is repeated between each driver and passenger, therefore a driver will deposit an amount of tokens equal to the number of passengers they are carrying). Then, in order to retrieve their token each agent needs to be physically present at the pick up point, in the designed time window. If unable to do so, the agent will forfeit the possession of the token that can be retrieved by any other passenger/driver present at that time and place. To have a better understanding of this process refer to Figure \ref{fig:compliance}. \newline
\begin{figure*}
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.6\columnwidth]{BondA.pdf} & \includegraphics[width=0.6\columnwidth]{BondB.pdf} & \includegraphics[width=0.6\columnwidth]{BondC.pdf}\\
(a) & (b) & (c)
\end{tabular}
\caption{(a) Driver $D$ deposits tokens for initiating a contract that they will pick up passengers from pick up points $A$ and $B$. At the same time, the two passengers deposit their tokens for appearing at the pick-up points. (b) Driver $D$ appears at pick-up point $A$ and collects the passenger from there, therefore, both the driver and the passenger retrieve their tokens for complying with the system. (c) Similarly, tokens are retrieved by the driver and the second passenger for both appearing at pick-up point $B$. At each stage the moving tokens are represented in green.}
\label{fig:compliance}
\end{figure*}
In what follows, we propose the use of a permissioned Distributed Ledger Technology (DLT) strategy to implement the proposed access control scheme. The acronym DLT is a term that describes blockchain and a suite of related technologies. From a broad perspective, a DLT is nothing more than a ledger held in multiple places, and a mechanism for agreeing on the contents of the ledger, namely the consensus mechanism. While this technology was first discussed in Nakamoto’s white paper in 2008 \cite{Nakamoto}, the technology has been used primarily as an immutable record keeping tool that enables financial transactions based on peer-to-peer trust \cite{Puthal}. In order to reach consensus, architectures such as blockchain operate a competitive mechanism enabled via mining (Proof-of-Work), whereas architectures such as the IOTA Tangle \cite{Zheng} based on Directed Acyclic Graph (DAG) structures often operate a cooperative consensus technique. The concept of using tokens to mark specific points where conditions are to be met, perfectly conforms with a DLT-based system. In fact, it is natural to use distributed ledger transactions to update the position of the tokens and to link them to the points of interest and associated data, using transactions (this can be done, for example, using smart sensors linked to digital wallets, as shown in Figure \ref{fig:compliance}). On top of that, a DLT-based system brings a number of advantages as a byproduct of its application to the smart city domain:\newline
\begin{itemize}
\item {\em Privacy :} In DLTs, transactions are pseudo- anonymous. This is due to the cryptographic nature of the private address\protect\footnote{https://laurencetennant.com/papers/anonymity-iota.pdf}, which is less revealing than other forms of digital payments that are uniquely associated with an individual \cite{cryptoprivacy}. This does not mean that DLTs users' identities are completely anonymous, especially in architectures in which it is possible to follow the trail of transactions among addresses. At the same time though, DLT systems are pseudo-anonymous in the sense that they manage to hide the details of single users and through randomization of the address they can make it difficult for attackers to trace the transactions. Therefore, from a privacy perspective, the use of DLT is desirable in a smart mobility scenario.\newline
\item {\em Ownership :} Transactions in the DLT can be encrypted, thus allowing every issuer to maintain ownership of their own data. In the aforementioned setting, the only information required to remain public is the current ownership of the tokens, whereas auxiliary information (e.g., user quality of service, statistics on the usage of the system) can be encrypted. This information can later be monetized for the benefit for the data owner.\newline
\item {\em Microtransactions :} Due to the amount of vehicles in an urban environment, and due to the need of linking the information to real time conditions (such as traffic or pollution levels), there is the demand for a fast and large data throughput.
\end{itemize}
Furthermore, the DLT system needs to be designed in a way such that whenever a user issues a token as a bond, that same user can retrieve the token \emph{if and only if} they are present at the pick up zone at the designed time. To do so we make use of the same mechanism and architecture proposed in \cite{NarendraYale}: namely a Proof of Position (PoP), DAG-based DLT called \emph{Spatial Positioning Token} (SPToken). Unlike other DLTs, in which each user has complete freedom on how to update the ledger with transactions, the SPToken network has a regulatory policy based on the physical positions of agents. This feature allows for a number of different uses: it can be employed to prevent agents to add transactions that do not possess any relevant data (since transactions can be encrypted)\cite{NarendraYale} or, as in this specific paper, it can be used to make sure that an agent satisfies certain conditions. Therefore, as a validation mechanism, SPToken makes use of PoP to authenticate transactions. In other words, for a transaction to be authenticated, it has to carry proof that the agent was indeed at the pick up point, at the designated time. This is achieved via special nodes called \emph{Observers} (see Fig.~\ref{fig:compliance}). Each observer is linked to a physical sensor in a city and it acts as a witness for the transaction. A sensor can be a fixed piece of infrastructure, or a trusted vehicle whose position is verified. As soon as a car is granted access to $R$, each user will deposit their tokens at the designated pick up zone. As soon as an agent reaches in time their pick up point, where one or more of her tokens are available to be picked, a short range connection is established (e.g., via Bluetooth) with the observer (whose job is to authenticate the transaction) and the token is transferred back to the owner's account. Refer again to Figure \ref{fig:compliance} for a better understanding of this process. This mechanism ensures that users have to be physically present in the interested locations to be able to retrieve their bond. This further authentication step makes SPToken a permissioned DAG-based DLT (similar to permissioned blockchains \cite{Puthal}), i.e., a distributed ledger where a certain amount of trusted nodes (the observers, in this case) is responsible to maintain the consistency of the ledger (as opposed to a public one, where security is handled by a cooperative consensus mechanism \cite{DLTAndSocial}).\newline
{\bf Comment :} Before continuing, we want to stress that very often in the context of \emph{smart cities}, algorithms assume full compliance with policies that are designed to optimise the resource allocation. To assume that a human agent would not break rules, especially if an individual profit can be made, is a very strong hypothesis that if relaxed might lead the whole system to fail and to produce less than optimal results. Therefore, it is the authors' opinion that the use of a compliance system is of paramount importance in the setting described so far, if efficiency is to be achieved. The issue of compliance is often overlooked.
\subsection{Mechanism Description}
We consider now the problem of allocating a certain amount of resources (i.e., permitted number of cars) among a set of agents (i.e., drivers and passengers using the scheme). The proposed method is inspired by the algorithm presented in \cite{QoS}, appropriately adjusted to the requirements of our ride-sharing scheme.\newline
We consider the following scenario. There is a population of size $n$ of citizens participating in the scheme, who request to commute to $R$ on a daily basis. The controlled region can accommodate up to $N$ vehicles per day. We assume $n > N$ and the population could be either passengers or drivers. We assume that there is a fleet of $N' > N$ electric vehicles in the scheme that are requested by the population for access in $R$, with $n > N'$. Without loss of generality and to facilitate presentation of our mechanism, the entities \textit{driver} and \textit{car} are considered equivalent and the corresponding terms are thus used interchangeably. As already mentioned in a previous section, our method is organized in two phases : matchmaking and probabilistic access. During matchmaking, we match passengers with drivers and group them into cars. The matching can happen in a number of ways, depending on the specific requirements of those who apply the system. For example, passengers could be matched with drivers based on proximity of their departing/arriving area, or based on a preference priority ranking that drivers/passengers maintain for each other. In our simulations we take a simple approach and match passengers randomly with drivers (and subsequently with cars), as long as there are available seats in the vehicles, taking into consideration the frequency at which a particular passenger has been assigned a seat in the past. That is, if a passenger has been assigned a seat less than $50\%$ of the time, then they are given priority to take a seat in a car, otherwise, they are not given priority. After the matchmaking is complete, each car is assigned an access probability based on its occupancy records. All cars with high enough probability, are permitted access to the city center. We present the technical details of this procedure, next.\newline
In our system, we will use $k$ to denote number of days (i.e., $ k=0,1,2,3,\dots$). For ease of interpretation we assume that access is granted on a daily basis to each user, but the algorithm is not affected by this assumption. Then, $X_i(k)$ is the state variable associated with each driver; it takes the value $1$ if the $i$th driver is given access to $R$ on the $k$th day and zero otherwise. Thus, $\overline{X}_i(k)$ is the average access for the $i$th driver up to the $k$th day, defined as
\begin{equation*}
\overline{X}_i(k) = \frac{1}{k+1} \sum_{j=0}^{k} X_i(j).
\end{equation*}
In the above context, let $z_i \in [0,1]$ represent the frequency of accessing the city for a car $i$, and $f_i : [0,1] \rightarrow \mathbb{R}$ be a convex cost function associated with it, representing the car's priority during the second phase of our mechanism. In this context the shape of this function can take into account a variety of factors: the amount of money paid for the pass (e.g., premium and standard account), the amount of public transportation available in the area where this user lives or the type of vehicle driven. Following~\cite{QoS}, we are interested in solving the following shared-resource optimization problem,
\begin{eqnarray}
\label{eq:optprob}
\minimize_{z_1,\ldots, z_{N'} \in \mathbb{R}} \,\, &\sum_{i=1}^{N'} f_i (z_i) \nonumber\\
\text{subject to} \,\, & \sum_{i=1}^{N'} z_i = N,\\
& z_i\geq 0, \quad i=1,\ldots, N'.\nonumber
\end{eqnarray}
Our aim is then to control the value of the variable $X_i(k)$ (i.e., the access to $R$, at each time step) in such a way that the average access of user $i$, $\overline{X}_i(k)$, converges to the optimal value $z_i^*$, subject to $\sum_{i=1}^n X_i \approx N$ (notice that we are not requesting the algorithm to exactly match the required amount of cars, at each time step but we are instead interested in obtaining $\text{lim}_{k,\infty}\sum_{i=1}^n \overline{X}_i(k) = N$). In order to do so, the probability that at each time step car $i$ gains access to the city center (i.e., $X_i(k) = 1$) is ruled by the following equations:
\begin{equation}
\label{eq:pi}
p_i(k) \triangleq \P(X_i(k) = 1) = \Gamma(k)\ \frac{\overline{X}_i(k)}{f_i'(\overline{X}_i(k))}\dfrac{n_i(k)}{c_i},
\end{equation}
\begin{equation}
\label{eq:alg3}
\Gamma(k+1) = \Gamma(k) + \alpha \left( N - \sum_{i=1}^{N'} X_i(k) \right),
\end{equation}
where $n_i(k)$ is the number of passengers carried in car $i$ at time $k$, $c_i$ is the car's maximum capacity and $\Gamma(k)$ is a global scaling variable, dependent on the parameter $\alpha >0$, whose dynamics ensures $p_i(k) \in [0,1], \forall i,k$. Notice that, equation (\ref{eq:pi}) differs from the one proposed in \cite{QoS} by the factor $n_i(k)/c_i$: since we are interested in maximising the amount of people getting into $R$ (while maintaining the amount of users having access close to $N$), this factor ensures that a fully filled car will have higher probability to be granted access than an empty one. As a further element, notice that in a classical setting, the presence of $\Gamma(\cdot)$ requires the existence of a centralised entity to compute and broadcast this global variable to all the agents in the network. In a DLT-based system, on the other hand, where informations are stored in a public ledger, the value $\Gamma(\cdot)$ can be computed independently by each user, therefore the algorithm can be executed in a completely decentralised fashion. A discussion on the convergence of this algorithm is beyond the scope of this paper and the interested reader can refer to \cite{QoS} for further details.
\section{Simulations and Results}\label{sec:simulations}
We now present empirical results to illustrate the efficacy of the techniques presented in the previous section. In what follows we based our simulations on the recent report~\cite{canal_dublin}. We assume a city of the size of Dublin in Ireland, with population $1,100,000$ approximately, of which $50,000$ are considered drivers and about $400,000$ are daily commuting passengers\footnote{In the report~\cite{canal_dublin} it is stated that between 7-10 am, about 210,000 commuters entered the city center. We make the assumption that in the length of the day that number can potentially double and therefore consider a population of $400,000$ commuters}. Consequently, we have a fleet of $50,000$ EV's, out of which only $40,000$ are permitted in the city centre $R$ on a daily basis\footnote{In the report~\cite{canal_dublin} it is stated that between 7-10am, about 50,000 cars entered the city center. Therefore, we limit the number of drivers to that number and the number of permitted cars to slightly less than this figure.}. All users that are not granted access to the city on a $EV$, are redirected to use public transportation. In our simulations, we set $\Gamma[0] = 1$, that is, the value that the parameter $\Gamma$ takes the first day of the scheme's operation, and $\alpha = 0.0001$. We also consider an application period of 360 days, that is slightly less than a year long. For convenience, on the first day of the operation, we consider that all drivers are permitted access.\newline
The simulation results are presented in Figures~\ref{fig:simulation_city}(a) and (b). Fig.~\ref{fig:simulation_city}(a) shows the number of cars that are granted access every day. Although at the beginning the number of cars in area $R$ are above the maximum permitted number ($40,000$), due to the effect of the access control mechanism this value is quickly reduced, stabilizing around the maximum level, on average, for the rest of the application period. Notice that if the maximum number of drivers in the city center was a hard constraint, it would be sufficient to reduce $N$ to take into account the fluctuations around this value. In Fig.~\ref{fig:simulation_city}(b), we show the frequency of being granted access, per user, on average over a period of one year. The small variance indicates that each user is granted fair access to the system. Regarding the commuting (shared) passengers, every passengers is granted access more than 1/3 of the time.\newline
Figures~\ref{fig:varied_access}(a)-(e) depict the number of cars with access, when the number of maximum permitted drivers changes and all other parameters in the system remain the same. The plots depict the steady state values. We observe that in all cases, the number of cars with granted access converges to the maximum value, on average. In terms of fair access, we show in Fig.~\ref{fig:varied_access}(f) boxplots of the frequency at which \textit{each driver} is granted access to the scheme, in the length of a year, with regards to the maximum number of cars permitted in $R$. As expected, the frequency increases as the available amount of resources increases. We highlight that in all cases, the variance is very small, meaning that all drivers in the scheme are ensured fair access (i.e., all drivers are able, on average, to access the city center the same number of times).\newline
\begin{figure}
\begin{tabular}{c}
\includegraphics[scale=0.18]{cars} \\(a) \\ \includegraphics[scale=0.18]{frequency_users}\\
(b)
\end{tabular}
\caption{(a) Number of cars with granted access in the length of a year (b) Frequency of granted access per user in the scheme}\label{fig:simulation_city}
\end{figure}
\begin{figure*}
\begin{center}
\begin{tabular}{ccc}
\includegraphics[width = .7\columnwidth]{max_20_4} & \hspace*{-1cm}\includegraphics[width = .7\columnwidth]{max_35_4} & \hspace*{-1cm}\includegraphics[width = .7\columnwidth]{max_50_4}\\
(a) Max = 20k cars & \hspace*{-1cm}(b) Max = 35k cars & \hspace*{-1cm}(c) Max = 50k cars\\ \includegraphics[width = .7\columnwidth]{max_75_4} & \hspace*{-1cm}\includegraphics[width = .7\columnwidth]{max_90_4} & \hspace*{-1cm}\includegraphics[width = .7\columnwidth]{frequency_drivers}\\
(d) Max = 75k cars & \hspace*{-1cm}(e) Max = 90k cars & \hspace*{-1cm}(f)
\end{tabular}
\caption{(a)-(e) Amount of permitted cars, for varied values of maximum allowed vehicles. Steady state values depicted. (f) Frequency of granted access over a year, per driver, for different setting of number of allowed vehicles.}\label{fig:varied_access}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[scale=.18]{changing_N}
\caption{Amount of permitted cars, while changing the value of maximum allowed vehicles gradually from $N = 50k$ to $N = 80k$ and vice versa.}\label{fig:changing_N}
\end{center}
\end{figure}
Finally to prove the efficacy of our approach in a more dynamic setting, we allow the number of maximum permitted cars to vary during the year. There are many reasons that make this a realistic scenario: the city municipality might wish to increase the number of permitted vehicles for the holiday seasons, or reduce it during heat waves, for example. We simulate this setting by changing the number of allowed cars during the year and we present the results in Figure~\ref{fig:changing_N}. Here, for the first two months of the operating period (i.e., 60 days) we give access to $N = 50k$ vehicles. For the next forty days $N$ increases linearly and for in the interval (of days) $[100, 180]$ it is set at $N = 80k$ vehicles. After that, the number of permitted cars decreases linearly again until it is set to the initial value, $N = 50k$, for the rest of the operating period. As we observe in the plot, our system reliably controls the access of vehicles, maintaining the number of permitted cars on average at a stable level around the set of maximum values.\newline
Regarding pollution levels caused by the $PM_{2.5}$ pollutant coming \textit{just} from the tyre wear of vehicles, we present in Figures~\ref{fig:pm_cars_space} the amount of particulate matter, depending on two variables: number of cars permitted in a city (Fig.~\ref{fig:pm_cars_space}(a)) and the volume of road network in a city (Fig.~\ref{fig:pm_cars_space}(b)). With regards to the volume of a city's road network, we wish to estimate, very approximately, the air-space in which the airborne PM is dispersed. For this, we assume that the total mass of PM generated per hour becomes uniformly dispersed throughout a volumetric space which is determined by the street length, an average street width of 10m and effectively enclosed by an average building height of 4m. Furthermore, we assume rather simplistically that the air in this volume is continuously replenished with an equivalent volume of fresh clean air at a rate of one air change per hour, in such a way as to maintain a pollution level which remains effectively constant with time. For the length of the road network, we can compute the total length of the streets in a predefined area in a city. Note that even though $4m$ is a somewhat arbitrary number for these simulations, the basic points remain valid irrespective of this assumption; that the amount of tyre generated PM can be regulated using our access control method. To this end, and based on the above assumptions, Fig.~\ref{fig:pm_cars_space} depicts the amount of PM per $m^3$ as a function of the number of vehicles operating in a city, per possible volume of space (computed as described above). In these figures, we depict in green the levels deemed safe for human health (i.e., the ones below the maximum permitted levels) and in red the ones exceeding the annual permitted levels. These plots suggest that, with the present situation in Dublin city (that is, ~500,000 cars out of which ~170,000 change tyres every year, and a space volume of approximately $450,000,000m^3$), the levels of tyre-wear related $PM_{2.5}$ emissions are very high. However, applying an access control scheme that restricts the number of vehicles to \textit{at most} 100,000 vehicles per day, can maintain the PM levels at acceptable levels even in small size cities with relatively small volume of space.
\begin{figure}
\centering
\begin{tabular}{c}
\includegraphics[scale=0.4]{pm_cars} \\ (a) \\ \includegraphics[scale=0.4]{pm_space}\\
(b)
\end{tabular}
\caption{Levels of tyre-wear related $PM_{2.5}$ emissions per number of vehicles operating (a) and volume of space where the matter is dispersed (b)}\label{fig:pm_cars_space}
\end{figure}
\section{Conclusions}\label{sec:conclusion}
The contributions of this paper are divided into two sections. In the first one, a detailed data analysis shows suggests that a simple ban on ICE vehicles does not address the problem of non-exhaust emissions (PM from tyres, in particular). Although there have been previous studies that present such numbers for other cities, we emphasise the point that in Dublin, the PM levels from tyres alone might be above the levels that are deemed safe by WHO. This provides us with the rationale to introduce, in the second part, an access control and ride-sharing scheme to limit the amount of cars in cities and therefore maintain the amount of airborne PM within safe levels for our health. This system is designed in such a way to encourage users to comply with the matchmaking scheme and to guarantee fair access to each car. Finally, to validate the proposed algorithm, we make use of extensive simulations to show that each user receives fair access to the city centre and that the PM emissions are kept within safe boundaries. As for future lines of research we will further extend the present work by using more complex models for tyre abrasion and airborne diffusion to obtain more accurate estimates for non exhaust emissions.
|
\section{Introduction}
During the recent years getting our hands dirty of working with spectral and pseudospectral methods, when it comes to solving differential/integral equations, we feel like there is a huge lack amongst a myriad of papers published in the field of solving differential/integral equations. The instant problem to detect is that all the scientists and researchers solve problems without using others' codes and packages. Undoubtedly, having a fixed platform and inviting the scientist to develop their work can reduce the time and effort in solving these similar-based problems.
Over these years, we have known that the works of these scientists are of giant works and worth admiring, but we always ask ourself "what if we gather the works of all current scientific computer scientist in order to design a package to use easier and grownup methods for the solutions of ODEs and PDEs?".
\\
To do so, it has been attempted to provide consistent and user-friendly high-level functions that allow experimental scientists to work properly and beneficially on their equations.\\
Furthermore, users and developers can easily extend the functionality and implement new
algorithms due to the modular design functionalities and facilities, and in higher expectation, harness them in other software packages.
This package and its documentation are freely
available from \href{https://github.com/siaadfilml/SPSMAT}{This address}.
The SPSMAT toolbox consists of approximately 80
high-level and other low-level functions.
We are to say that the high-level functions give a consistent
and easy-to-use interface of the functionality and reusability to the users and developers, allowing them to do the analysis in such well-defined manners. The
low-level functions implement the core functionality and are not needed to be shown to the end-users, these functions are set and called inside other functions. In addition to handy functions, some constructive examples are given in terms of guidance and showing the handling of the functions.
\\SPSMAT has been believed to be portable and open source. At this
aim, SPSMAT has been developed using Octave, which is a kind of free MATLAB clone and runs on the
commonest operating systems, such as Unix, Linux, Windows, and Mac OS X. \\
To the best of our knowledge, SPSMAT is also the first free scientific computing package which runs on GNU/Octave platform.
\section{Features of SPSMAT}
The SPSMAT package is
self-contained; it only needs an Octave environment
with
standard toolboxes [See
\href{http://www.gnu.org/software/octave/} {Octave} ].
SPSMAT is released under the free and open source GNU GENERAL PUBLIC LICENSE (GPL)
(Version 3, 29 June 2007). The accompanying source code and the documentation of the toolbox have been enriched with numerous comments, examples, and detailed instructions for extensions.
\\
This is the first in a series of papers on implementing different numerical techniques in term of an open source
GNU Octave toolbox. The intention of this ongoing project is
to provide a rapid prototyping package
for using spectral and pseudospectral methods in solving integral and differential equations.
\\
Here, we highlighted some of the outstanding features of SPSMAT:
\begin{itemize}
\item highly modular;
the reasons beyond highly modular design are:
\begin{itemize}
\item * The combinations of the approximation techniques.
\item * The easy construction and embedding of other novel techniques; the library aims to utilize the well-known spectral and pesudospectral methods as transparent as possible to the user. A small alternation and enhancing the code to fit one's assumption is easy to handle.
\item * Having a well-defined structure for the modules also simplifies the
maintenance of the code.
\item * Their immediate application in other
packages. The modular design facilitates reuse of source code in
other software. It ignites the
collaboration with current developers and with developers
of other software packages to broaden their work and understandings.
\end{itemize}
\item Free and open source; explained in next sections.
\item Multi-platform toolbox; SPSMAT is
multi-platform: it has been extensively tested on Windows and Linux; since it is made of standard Octave files, it is expected to work on
alternative platforms as well.
\item Following "all-in-one" philosophy. By this, we tried to represent one particular function with different inputs to encompass all possible outputs and purposes.
For instance, in order to have functions for Jacobi polynomials and their derivatives, we defined a function with input 0 to show the Jacobi polynomials and with input 1 to give the first order derivative of this polynomials.
\end{itemize}
\subsection{Why numerical methods?}
After delving through different books, we know that
they concentrated on ODEs based on
linear problems.
The trouble is most ODE problems including almost
all nonlinear ones can't be solved analytically.
A numerical one, however, can be obtained by computer. The very important advantage of numerical approximations is that they can be obtained for almost any ODE. In addition to the speed of these methods, the nature of numbers helps scrutinize results
obtained numerically even apply
different analysis method to a problem and compare them with one another. On another hand, it is not easy to compare two exact formulas, but it is beneficial to compare two numbers, even as numbers
or visually in plots \cite{therifeten}. That simply can be detected as a reason for designing a numerical package rather than the analytical one.
\subsection{Why No User Interface (UI)?}
We never assume that we are familiar with everything that a user
might want to do.
A great feature of the
SPSMAT toolbox is that it does not have a Graphical User
Interface (GUI). Instead, the user is interacting directly with
the functions on the command line or in scripts and is able to mix and build his new insight in order to solve a problem.
Providing a GUI prevent the user to
mix, match, and test the SPSMAT high-level functions experimentally.
Data and code, in fact, are in the hands of the end users to reshape them suitable to their need.
\subsection{Why Open source and Octave?}
MATLAB is widely
known and used in the science communities including scientific computing. Although
MATLAB has a rich feature set and flexibility, it is relatively and ridiculously expensive. MATLAB environment is a commercial and "closed" product, thus MATLAB
kernel and libraries cannot be modified nor freely distributed.
To allow exchanging insights and contributing scientific
research, both the toolbox and the platform on which a toolbox runs must be completely free.
Mathworks like other proprietary firms has its own policies that may not help you freely use and develop some packages.
That was maybe the ultimate motive to use an open source platform. In addition to it, as what an Iranian proverb saying \textbf{ one clap is not audible}, we
required
the users extend the analyses with their own code and be exchanged between users, and between students and
their supervisors, facilitating collaboration and knowledge over Github where they can even transfer their proper contribution with us. \\Nonetheless, these are not such limitation for MATLAB users at all. They can even download the package and use it.
\subsection{Why Fractional?}
Due to the accuracy of fractional differential and integral equations in modeling various natural phenomena, fractional calculus has become the focus of many
researchers.
Fractional differential equations provide
an outstanding instrument to describe the complex phenomena in fields of viscoelasticity, electromagnetic waves, diffusion equations, and so on. Moreover, the fractional
order models of real systems are more sufficient in comparison with the integer order
cases. Therefore, the field of fractional calculus has motivated the interest of researchers
in various fields like physics, chemistry, engineering and even finance where all of their scientists and students need to work with numerical methods of solving these fractional equations. We broadly aim at this spectrum and wanted to produce a package for all of these practitioners.
\subsection{Why Jacobi?}
The classical Jacobi polynomials, have been used extensively in mathematical
analysis and practical applications, and play a pivotal role in the analysis and implementation of spectral and pseudospectral methods \cite{bhrawy2}.
It is easily proven that the Jacobi polynomials are the polynomials arising as eigenfunctions of a singular
Sturm--Liouville problem \cite{bhrawy2}.
The usual spectral method by Legendre or Chebyshev approximation, and available only for non-singular problems on rectangular bounded domains. However, the general Jacobi method can be used in a wide array of
problems \cite{bhrawy3}.\\
In addition to this, in this version, also for the simplicity's sake, we cover Jacobi polynomials which are wide-ranging polynomials encompassing Gegenbauer (ultraspherical), Chebyshev, and Legendre polynomials. But, for those who are not into these polynomials, we also offered the derived subcategories including Gegenbauer, Chebyshev (all four kinds), and Legendre polynomials.
\subsection{Why Now?}
Over the recent years, many researchers have been interested in studying the properties
of various equations and providing robust and accurate analytical and numerical
methods for solving these equations and recorded their work in their published papers: what we have seen at this moment is to see and gather their works in one single work.
Moreover, some of the high-qualified state-of-the-art methods and landmark works are done and analyzed perfectly that let us code them in the best shape that is readable and understandable, and specifically, beneficial. In addition to this, other works done on spectral methods are out-dated; like the work of Funaro/Matlab that is done in 1993 via Fortran (See \href{http://cdm.unimo.it/home/matematica/funaro.daniele/rout.htm} {Funaro package}). What these packages lack are that at that time fractional calculus, operational matrices and Lagrangian bases had not been developed like today. The other FORTRAN package, PseudoPack 2000 is implemented by Don and Costas
and is available in
[\href{http://www.cfm.brown.edu/people/wsdon/home.html}{PseudoPack package}]. In this work parallel computers capability involved to solve problems and suffer the same misfortunes we issued here. In the newest work by Trefethen in
\href{http://www.chebfun.org/docs/guide/}{Chebfun package},
a huge and broad work done to refer a tool for researchers. But the light version, and more importantly, the specification for solving differential and integral equation in terms of spectral and pseudospectral was not its first goal.
SPSMAT, lo and behold, was created to address some of these issues.
\subsection{Why Matrix method?}
As the operations in MATLAB and Octave are based on the concept of matrix, and in some extends their performance is optimal for matrices, we have given a
matirx-based method in order to reach the best action of MATLAB or Octave.
Another reason is the parallel computation that is said to be feasible and meaningful when it comes to matrix operations like multiplications and so forth. In this version, we have not considered matrix parallel computation methods, but in due course, we will work on it and report its relevant results.
\subsection{Why Spectral?}
The well-defined Spectral methods, by using computer power, have been developed rapidly through the recent years
for the numerical solutions of ordinary/partial differential/integral equations. They have been providing a satisfying accuracy compared with
other numerical methods and have wide applications in many
mathematical problems. As we have realized, the spectral methods have been developed in different ways to provide accurate solutions for linear and nonlinear ordinary/partial differential/integral equations. Additionally, spectral methods
have received considerable attention in dealing with various problems. As their most significant characteristics,
they reduce problems to those of solving a system of algebraic equations; this makes them easier for solving. Also,
they have excellent error properties and they offer exponential rates of convergence for smooth problems.
Such methods
have matured over the last 50 years and, in many cases, meet the robustness
and computational efficiency standards required in practice.
\subsection{Why Pseudospectral?}
In general, there are two
ways to construct a polynomial approximation to the solution $y(x)$:
One is to use an interpolating polynomial between the values $y(x_j)$ at
some points $x_j$. The other is to use a series expansion in terms of Lagrange or orthogonal polynomials like Jacobi polynomials. Here, the latter idea is pursued,
and both the Jacobi and Lagrange Jacobi polynomials are constructed and the difference between employing them are also represented.
In fact, the pseudospectral methods (here using Lagrange polynomials) are based on the spectral techniques. In pseudospectral methods, there are particularly two steps to achieve a numerical approximation
for a differential equation. Firstly, a proper finite or discrete representation of the
solution should be selected-- this can be done by polynomial interpolation of the solution based on
some appropriate nodes. On another hand, it is known that the Lagrange interpolation polynomial based
on equidistant points does not yield a satisfactory approximation to general smooth functions. As a matter of fact, when the number of collocation points increases, interpolant polynomials notably diverge. However, satisfying results are obtained
by relating the collocation points to the structure of classical orthogonal polynomials, such as the well-established Jacobi-Gauss-Lobatto points. That can be figured as a cogent reason for choosing these points in defining Lagrange polynomials. The use of global polynomials together with
Gaussian quadrature collocation points (here Jacobi-Gauss-Lobatto points) is known to provide accurate approximations that converge exponentially for problems
whose solutions are smooth \cite{kidderlatifi,fokkerlatifi,delkhosh,delkhosh2,div}.
\subsection{What is the prerequisite and who is this package for?}
This software package is intended for those who have a taste for solving ODEs, PDEs, integral equations and Optimal Controls. Practitioners with a not-so-deep understanding of Matlab or Octave can use this package without any difficulty. But basically this package is meant to be useful to solve different differential and integral equations; so it stipulates that the users have a grasp of Numerical analysis beforehand.
Undergraduate students taking a course in scientific computing and any relevant course can find this package as a test machine to check different things with orthogonal polynomials and develop their work with this package: this is your lightweight companion. In nutshell, whoever you are, we aim to increase your appreciation of this fundamental subject.
\section{Solving a nonlinear PDE with SPSMAT}
In this section, firstly we discuss the orthogonal Gegenbauer polynomials. The orthogonal Gegenbauer polynomial of order $N$ is defined over $[-1,1]$ \cite{bahram}:
\begin{equation}
G_{N}^{\alpha}(x)=\sum_{i=0}^{\lfloor\frac{N}{2}\rfloor}(-1)^i\frac{\Gamma(N+\alpha-i)}{\Gamma(i!(N-2i)!\Gamma(\alpha))}(2x)^{N-2i},
\end{equation}
\begin{equation}
~G_{N}^{\alpha}(-x)=(-1)^NG_{N}^{\alpha}(x),~G_{N}^{\alpha}(\pm 1 )=(\pm 1)^N\frac{\Gamma(N+2\alpha)}{\Gamma(2\alpha)\Gamma(N+1)}.
\end{equation}
Other features of Gegenbauer polynomials can be found in \cite{bahram}.\\
As said, this package helps solve partial and ordinary differential equations. To show its capability, we have considered nonlinear Fokker-Plank equation:
$$ \frac{\partial y}{\partial t} =\big[\frac{-\partial }{\partial x}A(x,t,y)+ \frac{\partial^2 }{\partial x^2}B(x,t,y)\big]y,$$
where $ A(x,t,y)= \frac{4y}{x}-\frac{x}{3}$ and $ B(x,y,t)=y $ with the initial and boundary conditions $y(x,0)=x^2$, $y(0,t)=0$ and $y(1,t)=e^t$. The exact solution is $y(x,t)=x^2e^t$ and $x\in [0,1],~t\in [0,1]$.
For solving this problem, with help of Crank-Nicolson method \cite{fokkerlatifi}, we discretized the variable $t$ and solved the equation in different iterations \cite{fokkerlatifi}.
Notice that $t_n=n\times\bigtriangleup t$ and $\bigtriangleup t=t_{n+1}-t_n$.\\
Using $\frac{\partial y}{\partial t}=\frac{y^{n+1}-y^{n}}{\bigtriangleup t}$ where $y^{n}=y(x,t_n)$ and applying Crank-Nicolson method we have got
$$ y_t=(\frac{4y}{x^2}-\frac{8y_x}{x}+\frac{1}{3}+2y_{xx})y+(\frac{x}{3}+2y_x)y_x, $$
\begin{eqnarray}
&&\frac{y^{n+1}-y^{n}}{\bigtriangleup t}=\theta (\frac{4y^{n}}{x^2}-\frac{8y_x^{n}}{x}+\frac{1}{3}+2y_{xx}^{n})y^{n+1}\nonumber\\&&+(\frac{x}{3}+2y_x^{n})y_x^{n+1}+\bigtriangleup t(1-\theta) (\frac{4y^{n}}{x^2}-\frac{8y_x^{n}}{x}+\frac{1}{3}+2y_{xx}^{n})y^{n}+(\frac{x}{3}+2y_x^{n})y_x^{n},
\end{eqnarray}
in which $y^{n+1}$ is unknown and is supposed to be determined. Simplifying the last equation we have
\begin{eqnarray}\label{eqq}
&&\bigg[1-\theta\bigtriangleup t\bigg((\frac{4y^{n}}{x^2}-\frac{8y_x^{n}}{x}+\frac{1}{3}+2y_{xx}^{n})\bigg)\bigg]y^{n+1}-\theta\bigtriangleup t(\frac{x}{3}+2y_x^{n})y_{x}^{n+1}=y^{n}+\nonumber\\&&
(1-\theta)\bigtriangleup t \bigg((\frac{4y^{n}}{x^2}-\frac{8y_x^{n}}{x}+\frac{1}{3}+2y_{xx}^{n})\bigg)y^{n} +(1-\theta)\bigtriangleup t (\frac{x}{3}+2y_x^{n})y_{x}^{n}.
\end{eqnarray}
Now by defining
$$s_1:=x\rightarrow \bigtriangleup t (\frac{4y^{n}}{x^2}-\frac{8y_x^{n}}{x}+\frac{1}{3}+2y_{xx}^{n}),$$
$$s_2:=x\rightarrow \bigtriangleup t (\frac{x}{3}+2y_x^{n}),$$
then Eq. (\ref{eqq}) will be
\begin{eqnarray}\label{eqq_}
&&\bigg(1-\theta s_1(x)\bigg)y^{n+1}-\theta s_2(x) y_{x}^{n+1}=\bigg(1+
(1-\theta)s_1(x)\bigg)y^{n} +(1-\theta)s_2(x)y_{x}^{n}.\nonumber
\end{eqnarray}
Now for solving this equation, we use an approximation expansion with Generalized Lagrange polynomials:
$y^{n+1}=\sum_{i=0}^{N}L_{i}^{u}(x)C_i^{n+1}$, where $u=2x-1$, $C_i^{n+1}$ are the unknown coefficients, and $L_{i}^{u}(x)$ are the Generalized Lagrange polynomials defined in \cite{kidderlatifi,fokkerlatifi}. Then
\begin{eqnarray}\label{eqq__}
&&\bigg(1-\theta s_1(x)\bigg)\sum_{i=0}^{N}L_{i}^{u}(x)C_i^{n+1}-\theta s_2(x) \sum_{i=0}^{N}\big(L_{i}^{u}(x)\big)'C_i^{n+1}=\nonumber\\&&
\bigg(1+
(1-\theta)s_1(x)\bigg)\sum_{i=0}^{N}L_{i}^{u}(x)C_i^{n} +(1-\theta)s_2(x)\sum_{i=0}^{N}\big(L_{i}^{u}(x)\big)'C_i^{n},
\end{eqnarray}
and with Gegenbauer Gauss Lobatto points $x_0=0$, $x_{N}=1$ and $N-1$ roots of $G_{N-1}^{\theta}(u)$ which are calculated with:
$$x_i=[0,roots(G_{N-1}^{\alpha+1}(u)),1],i=0..N.$$
\begin{center}
\begin{figure}[H]
\centering
\fbox{
\includegraphics[scale=0.46]{roots.png}}
\caption{Gegenbauer Gauss Lobatto points}
\label{piece2}
\end{figure}
\end{center}
Setting
$$v_1=[s_1(x_1),...,s_1(x_{N-1})]^T,$$
$$v_2=[s_2(x_1),...,s_2(x_{N-1})]^T,$$
$$m_1=[1,1,\dots,1]^T-\theta v_1, m_2=\theta v_2,$$
$$\hat{m_1}=[1,1,\dots,1]^T-(1-\theta) v_1, \hat{m_2}=(1-\theta) v_2,$$
and considering Eq. (\ref{eqq__}), by using Gegenbauer Gauss Lobatto points, we will get a matrix form as
\begin{equation}\label{eqqq}
A^{n+1}C^{n+1}=b^{n}
\end{equation}
in which $A^{n+1}$ is the coefficient matrix and $ C^{n+1} $ is the unknown vector.
By implementing boundary condition, $ A^{n+1} $ is rewritten as:
\[
A^{n+1}=\begin{bmatrix}
A_1\\diag(m_1)\times A_2-diag(m_2)\times A_5\\
A_3
\end{bmatrix},
b^{n}=\begin{bmatrix}
0\\diag(\hat{m_1})\times A_2+diag(\hat{m_2})\times A_5\\
e^t
\end{bmatrix}
\]
by which
\[
A_1=\begin{bmatrix}
1&0&\dots&0
\end{bmatrix},
A_3=\begin{bmatrix}
0&0&\dots&1
\end{bmatrix},
A_2=\begin{bmatrix}
0&1&0&\dots&0\\
0&0&1&\dots&0\\
\vdots\\
0&0&\dots&1&0\\
\end{bmatrix},
\]
\[
A_4=D^{(2)}[1:N-1,:], ~A_5=D^{(1)}[1:N-1,:],
\]
where $ D^{(1)} $ and $D^{(2)}$ are the Generalized Lagrange first and second order derivative matrices that are calculated in \cite{kidderlatifi,fokkerlatifi}.
\begin{center}
\begin{figure}[H]
\centering
\fbox{
\includegraphics[scale=0.46]{As.png}}
\caption{Piece of code for obtaining $A_1\dots A_5$}
\label{piece2}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[H]
\centering
\fbox{
\includegraphics[scale=0.46]{forstep.png}}
\caption{Solving $AC^{n+1}=b^n$ at different time steps.}
\label{piece2}
\end{figure}
\end{center}
Also applying initial condition $y^0=x^2$ leads to $C^0=[x_0^2,x_1^2...,x_N^2]$ .
\begin{center}
\begin{figure}[H]
\centering
\fbox{
\includegraphics[scale=0.46]{initial.png}}
\caption{Initial condition}
\label{piece2}
\end{figure}
\end{center}
The result of solving the system in Eq. (\ref{eqqq}) with SPSMAT is depicted in Fig. \ref{exa6_error}.
The completed code file contains more of these functions which are mostly discussed \href{https://github.com/siaadfilml/SPSMAT}{here}.
\begin{center}
\begin{figure}
\centering
\includegraphics[scale=0.48]{Pic_Exp6.eps}
\caption{Error illustration.$\theta=0.5$,$\alpha=0.5$, $N=7$. }
\label{exa6_error}
\end{figure}
\end{center}
\begin{acknowledgements}
In the end, we owe some researchers a thank who were at our disposal when we were in dire need. We thank profusely Mohammad Hemami and Dr. Saeed Kazem.
\end{acknowledgements}
\bibliographystyle{spmpsci}
|
\section{Introduction} \label{sec::Introduction}
Regarded as one of the most basic tools to investigate statistical properties of unsupervised data, clustering aims to group a set of objects in such a way that objects in the same cluster are more similar in some sense to each other than to those in other clusters.
Typical application possibilities are to be found reaching from categorization of tissues in medical imaging to grouping internet searching results. For instance, on PET scans, cluster analysis can distinguish between different types of tissue in a three-dimensional image for many different purposes \citep{FILIPOVYCH20112185} while in the process of intelligent grouping of the files and websites, clustering algorithms create a more relevant set of search results \citep{DiMarco2013}.
Because of their wide applications, more urgent requirements for clustering algorithms that not only maintain desirable prediction accuracy but also have high computational efficiency are raised. In the literature, a wealth of algorithms have already been proposed
such as $k$-means \citep{Macqueen67somemethods}, linkage \citep{JoeH1963Hierarchical, Sibson1973SLINK, Defays1977An}, cluster tree \citep{Stuetzle2003Estimating}, DBSCAN \citep{Ester1996A}, spectral clustering \citep{Donath1973Lower, Luxburg2007A}, and expectation-maximization for generative models \citep{Dempster1977Maximum}.
As is widely acknowledged, an open problem in cluster analysis is how to describe a conceptually and mathematically convincing definition of clusters appropriately.
In the literature, great efforts have been made to deal with this problem.
Perhaps the first definition dates back to \cite{Har1975}, which is known as the single-level density-based clustering assuming i.i.d.~data $D = (x_1, \dots, x_n)$ generated by some unknown distribution $\mathrm{P}$ that has a continuous density $h$ and the clusters of $\mathrm{P}$ are then defined to be the connected components of the level set $\{ h \geq \rho \}$ given some $\rho \geq 0$. Since then, different methods based on the estimator $\hat{h}$ and the connected components of \{$h \geq \rho\}$ have been established \citep{Cuevas1997A, Maier2012Optimal, Rigollet2006Generalization, Rinaldo2010GENERALIZED}.
Note that the single-level approach mentioned above is easily shown to have a conceptual drawback that different values of $\rho$ may lead to different (numbers of) clusters, and there is also no general rule for choosing $\rho$.
In order to address this conceptual shortcoming, another type of the clustering algorithms, namely hierarchical clustering, where the hierarchical tree structure of the connected components for different levels $\rho$ is estimated, was proposed. Within this framework, instead of choosing some $\rho$, the so-called cluster tree approach tries to consider all levels and the corresponding connected components simultaneously.
It is worth pointing out that the advantage of using cluster tree approach lies in the fact that it mainly focuses on the identification of the hierarchical tree structure of the connected components for different levels.
For this reason, in the literature, there have already been many attempts to establish their theoretical foundations.
For example, \cite{Hartigan1981Consistency} proved the consistency of
a hierarchical clustering method named single linkage merely for the one-dimensional case
which becomes a more delicate problem that it is only fractionally consistent in the high-dimensional case. To address this problem, \cite{Chaudhuri10ratesof} proposed a modified single linkage algorithm which is shown to have finite-sample convergence rates as well as lower bounds on the sample complexity under certain assumptions on $h$.
Furthermore, \cite{Kpotufe_pruningnearest} obtained similar theoretical results with an underlying $k$-NN density estimator and achieved experimental improvement by means of
a simple pruning strategy that removes connected components that artificially occur because of finite sample variability.
However, the notion of \emph{recovery} taken from \cite{Hartigan1981Consistency} falls short of only focusing on the correct estimation of the cluster tree structure and not on the estimation of the clusters itself, more details we refer to \cite{Rinaldo2010GENERALIZED}.
So far, the theoretical foundations for hierarchical clustering algorithms such as consistency and learning rates of the existing hierarchical clustering algorithms are only valid for the cluster tree structure and therefore far from being satisfactory. As a result, in this paper, we proceed with the study of single-level density-based clustering. In the literature, recently,
various results for estimating the optimal level have already been established.
First of all, \cite{steinwart2011adaptive} and \cite{steinwart2015adaptive} presented algorithms based on histogram density estimators that are able to asymptotically determine the optimal level and automatically yield a consistent estimator for the target clusters. Obviously, these algorithms are of little practical value since only the simplest possible density estimators are considered. Attempting to address this issue, \cite{sriperumbudur2012consistency} proposed a modification of the popular DBSCAN clustering algorithm.
Although consistency and optimal learning rates have been established
for this new DBSCAN-type construction,
the main difficulty in carrying out this algorithm
is that it restricts the consideration only to moving window density estimators for $\alpha$-H\"{o}lder continuous densities.
In addition, it's worth
noticing that
none of the algorithms mentioned above can be well adapted to the case where the underlying distribution possesses no split in the cluster tree.
To tackle this problem, \cite{steinwart2017adaptive} proposed an adaptive algorithm using kernel density estimators which, however, also only performs well for low-dimensional data.
In this paper, we mainly focus on clusters that are defined as the connected components of high density regions
and present an algorithm called \emph{best-scored clustering forest} which can not only guarantee consistency and attain fast convergence rates, but also enjoy satisfactory performance in various numerical experiments.
To notify, the main contributions of this paper are twofold:
\emph{(i)} Concerning with the theoretical analysis, we prove that with the help of the best-scored random forest density estimator, our proposed algorithm can ensure consistency and achieve fast convergence rates under certain assumptions for the underlying density functions and target clusters. We mention that the convergence analysis is conducted within the framework established in \cite{steinwart2015adaptive}. To be more precise,
under properly chosen hyperparameters of
the best-scored random forest density estimator \cite{hang2018best},
the consistency
of the best-scored clustering forest
can be ensured. Moreover, under some additional regularization conditions, even fast convergence rates can be achieved.
\emph{(ii)} When it comes to numerical experiments, we improve the original purely random splitting criterion by proposing an adaptive splitting method. Instead, at each step, we randomly select a sample point from the training data set and the to-be-split node is the one which this point falls in. The idea behind this procedure is that when randomly picking sample points from the whole training data set, nodes with more samples will be more likely to be chosen whereas nodes containing fewer samples are less possible to be selected. In this way,
the probability to obtain cells with sample sizes evenly distributed will be much greater. Empirical experiments further show that the adaptive/recursive method enhances the efficiency of our algorithm since it actually increases the \emph{effective} number of splits.
The rest of this paper is organized as follows: Section \ref{sec::Preliminaries} introduces some fundamental notations and definitions related to the density level sets and best-scored random forest density estimator.
Section \ref{sec::AGenericClusteringAlgorithmBasedonBRDF} is dedicated to the exposition of the generic clustering algorithm architecture.
We provide our main theoretical results and statements on the consistency and learning rates of the proposed best-scored clustering forest in Section \ref{sec::ConsistencyandRatesforRFDE-basedClustering},
where the main analysis aims to verify that our best-scored random forest could provide level set estimator that has control over both its vertical and horizontal uncertainty.
Some comments and discussions on the established theoretical results will be also presented in this section.
Numerical experiments conducted upon comparisons between best-scored clustering forest and other density-based clustering methods are given in Section \ref{sec::ExperimentalPerformance}.
All the proofs of Section \ref{sec::AGenericClusteringAlgorithmBasedonBRDF} and Section \ref{sec::ConsistencyandRatesforRFDE-basedClustering} can be found in Section \ref{sec::Proofs}. We conclude this paper with a brief discussion in the last section.
\section{Preliminaries} \label{sec::Preliminaries}
In this section, we recall several basic concepts and notations related to clusters in the first subsection while in the second subsection we briefly recall the best-scored random forest density estimation proposed recently by \cite{hang2018best}.
\subsection{Density Level Sets and Clusters} \label{subsec::DensityLevelSetsandClusters}
This subsection begins by introducing some basic notations and assumptions about density level sets and clusters.
Throughout this paper,
let $\mathcal{X} \subset \mathbb{R}^d$ be a compact and connected subset, $\mu := \lambda^d$ be the Lebesgue measure with $\mu(\mathcal{X}) > 0$.
Moreover, let $\mathrm{P}$ be a probability measure that is absolutely continuous with respect to $\mu$
and possess a bounded density $f$ with support $\mathcal{X}$.
We denote the centered hypercube of $\mathbb{R}^d$ with side length $2r$ by $B_r$ where
\begin{align*}
B_r := \{ x = (x_1, \ldots, x_d) \in \mathbb{R}^d : x_i \in [-r, r], i = 1, \ldots, d \},
\end{align*}
and the complement of $B_r$ is written by $B_r^c := \mathbb{R}^d \setminus [-r, r]^d$.
Given a set $A \subset \mathcal{X}$, we denote by $\mathring{A}$ its interior, $\bar{A}$ its closure, $\partial{A} = \bar{A} \setminus \mathring{A}$ its boundary, and
$\mathrm{diam}(A) := \sup_{x, x' \in A} \|x - x'\|_2$ its diameter.
Furthermore, for a given $x$, $d(x,A) := \inf_{x' \in A} \|x-x'\|_2$ denotes the distance between $x$ and $A$.
Given another set $B \subset \mathcal{X}$, we denote by
$A \triangle B$ the symmetric difference between $A$ and $B$.
Moreover, $\boldsymbol{1}_A$ stands for the indicator function of the set $A$.
We say that a function $f : \mathbb{R}^d \to \mathbb{R}$ is $\alpha$-H\"{o}lder continuous, if there exists a constant $c > 0$ such that
\begin{align*}
|f(x)-f(x')| \leq c \|x - x'\|_2^{\alpha},
\qquad \qquad
\alpha \in (0, 1].
\end{align*}
To mention, it can be apparently seen that $f$ is constant whenever $\alpha > 1$.
Finally, throughout this paper, we use the notation $a_n \lesssim b_n$ to denote that there exists a positive constant $c$ such that $a_n \leq c b_n$, for all $n \in \mathbb{N}$.
\subsubsection{Density Level Sets} \label{subsec::DensityLevelSets}
In order to find a notion of density level set which is topologically invariant against different choices of the density $f$ of the distribution $\mathrm{P}$, \cite{steinwart2011adaptive} proposes to define a density level set at level $\rho \geq 0$ by
\begin{align*}
M_{\rho} := \mathrm{supp} \, \mu_{\rho}
\end{align*}
where $\mathrm{supp} \, \mu_{\rho}$ stands for the support of $\mu_{\rho}$, and the measure $\mu_{\rho}$ is defined by
\begin{align*}
\mu_{\rho}(A) := \mu(A \cap \{ f \geq \rho \}),
\qquad \qquad
A \in \mathcal{B}(\mathcal{X}),
\end{align*}
where $\mathcal{B}(\mathcal{X})$ denotes the Borel $\sigma$-algebra of $\mathcal{X}$.
According to the definition, the density level set $M_\rho$ should be closed.
If the density $f$ is assumed to be $\alpha$-H\"{o}lder continuous, the above construction could be replaced by the usual $\{f\geq\rho \}$ without changing our results.
\begin{figure*}[htbp]
\begin{minipage}[t]{0.99\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/deflevelset.png}
\end{minipage}
\centering
\caption{Topologically relevant changes on set of measure zero.
Left: The thick solid lines indicate a set consisting of two connected components $A_1$ and $A_2$. The density of $\mathrm{P}$ is $f=c \boldsymbol{1}_{A_1 \cup A_2}$ with $c$ being a suitable constant, then $A_1$ and $A_2$ are the two connected components of $\{ f \geq \rho\}$ for all $\rho \in [0, c]$.
Right: This is a similar situation. The straight horizontal thin line indicates a line of measure zero connecting the two components, and the dashed lines indicate cuts of measure zero. In this case, the density of $\mathrm{P}$ is $f' = c \boldsymbol{1}_{A_1 \cup A_2 \cup A_3 \cup A_4}$, then $A_1$, $A_2$, $A_3$, and $A_4$ are the four connected components of $\{ f' \geq \rho \}$ for all $\rho \in [0, c]$.}
\label{fig:deflevelset}
\end{figure*}
Here, some important properties of the sets $M_\rho$, $\rho\geq 0 $ are useful:
\begin{enumerate}[(i)]
\item
\emph{Level Sets}.
$\overline{\{f>\rho\}} \subset M_\rho \subset\{ f \geq\rho \}$
\item
\emph{Monotonicity}.
$M_{\rho_2}\subset M_{\rho_1}$ for all $\rho_1\leq \rho_2$.
\item
\emph{Regularity}.
$\mu (M_\rho \triangle \{ f \geq\rho\})=0$.
\item
\emph{Normality}.
$\bar{M}_\rho=\dot{M}_\rho$, where $\bar{M}_\rho:=\bigcup_{\rho'>\rho}\, M_{\rho'}$ and $\dot{M}_\rho:=\bigcup_{\rho'>\rho}\, \mathring{M}_{\rho'}$.
\item
\emph{Open Level Sets}.
$\bar{M}_\rho=\{ f >\rho\}$.
\end{enumerate}
\subsubsection{Comparison of Partitions and Notations of Connectivity} \label{subsubsec::ComparisonofPartitionsandNotationsofConnectivity}
Before introducing the definition of clusters, some notions related to the connected components of level sets are in need. First of all, we give the definition that compares different partitions.
\begin{definition} \label{def::partition}
Let $A, B \subset \mathcal{X}$ be nonempty sets with $A \subset B$,
and $\mathcal{P}(A)$ and $\mathcal{P}(B)$ be partitions of $A$ and $B$, respectively.
Then $\mathcal{P}(A)$ is said to be \emph{comparable} to $\mathcal{P}(B)$, if for all $A' \in \mathcal{P}(A)$, there exists a $B' \in \mathcal{P}(B)$ such that $A'\subset B'$.
In this case, we write $\mathcal{P}(A) \sqsubset \mathcal{P}(B)$.
\end{definition}
It can be easily deduced that $\mathcal{P}(A)$ is comparable to $\mathcal{P}(B)$, if no cell $A'\in \mathcal{P}(A)$ is broken into pieces in $\mathcal{P}(B)$.
Let $\mathcal{P}_1$ and $\mathcal{P}_2$ be two partitions of $A$, then we call $\mathcal{P}_1$ is \emph{finer} than $\mathcal{P}_2$ if and only if $\mathcal{P}_1\sqsubset \mathcal{P}_2$. Moreover, as is demonstrated in \cite{steinwart2015suppA}, for two partitions $\mathcal{P}(A)$ and $\mathcal{P}(B)$ with $\mathcal{P}(A)\sqsubset\mathcal{P}(B)$, there exits a unique map $\zeta:\mathcal{P}(A)\to \mathcal{P}(B)$ such that $A'\subset \zeta(A')$ for $A'\in \mathcal{P}(A)$. We call $\zeta$ the \emph{cell relating map (CRM)} between $A$ and $B$.
Now, we give further insight into two vital examples of comparable partitions coming from connected components. Recall that an $A\subset \mathcal{X}$ is \emph{topologically connected} if, for every pair $A',A''\subset A$ of relatively closed disjoint subsets of $A$ with $A'\cup A''=A$, we have $A'=\emptyset$ or $A''=\emptyset$. The maximal connected subsets of $A$ are called the \emph{connected components} of $A$. As is widely acknowledged, these components make up a partition of $A$, and we denote it by $\mathcal{C}(A)$. Furthermore, for a closed $A\subset B$ with $|\mathcal{C}(B)|<\infty$, we have $\mathcal{C}(A) \sqsubset \mathcal{C}(B)$.
The next example describes another type of connectivity, namely \emph{$\tau$-connectivity},
which can be considered as
a discrete version of \emph{path-connectivity}. For the latter, let us fix a $\tau>0$ and $A\in \mathcal{X}$.
Then, $x,x'\in A$ are called \emph{$\tau$-connected} in $A$, if there exists $x_1,\ldots,x_n\in A$ such that $x_1=x$, $x_n=x'$ and $\|x_i-x_{i+1}\|_2<\tau$ for all $i=1,\ldots,n-1$.
Clearly, being $\tau$-connected gives an equivalence relation on $A$. To be specific, the resulting partition can be written as $\mathcal{C}_{\tau}(A)$, and we call its cells the \emph{$\tau$-connected components} of $A$.
It can be verified that, for all $A\subset B$ and $\tau>0$, we always have $C_\tau(A) \sqsubset C_\tau(B)$, see
Lemma A.2.7 in \cite{steinwart2015suppA}. In addition, if $|\mathcal{C}(A)|<\infty$, then we have $\mathcal{C}(A) = \mathcal{C}_\tau(A)$ for all sufficiently small $\tau > 0$, see Section 2.2 in \cite{steinwart2015adaptive}.
\subsubsection{Clusters} \label{subsubsec::Clusters}
Based on the concept established in the preceding subsections we now recall the definition of clusters, see also Definition 2.5 in \cite{steinwart2015suppA}.
\begin{definition}[Clusters] \label{def::ClusterProperty}
Let $\mathcal{X}\subset \mathbb{R}^d$ be a compact and connected set, and $\mathrm{P}$ be a $\mu$-absolutely continuous distribution. Then $\mathrm{P}$ \emph{can be clustered} between $\rho^*\geq 0$ and $\rho^{**}>\rho^*$, if $\mathrm{P}$ is normal and for all $\rho\in [0,\rho^{**}]$, the following three conditions are satisfied:
\begin{enumerate}[(i)]
\item We have either $|\mathcal{C}(M_\rho)| = 1$ or $|\mathcal{C} (M_\rho)| = 2$;
\item If we have $|\mathcal{C}(M_\rho)|=1$, then $\rho\leq \rho^*$;
\item If we have $|\mathcal{C}(M_\rho)|=2$, then $\rho\geq \rho^*$ and $\mathcal{C}(M_{\rho^{**}})\sqsubset \mathcal{C}(M_\rho)$.
\end{enumerate}
Using the CRMs $\zeta_{\rho}:\mathcal{C}(M_{\rho^{**}})\to \mathcal{C}(M_{\rho})$; we then define the \emph{clusters} of $\mathrm{P}$ by
\begin{align*}
A_i^*=\bigcup_{\rho\in (\rho^*,\rho^{**}]} \zeta_\rho (A_i),\quad i\in\{1,2\},
\end{align*}
where $A_1$ and $A_2$ are the two topologically connected components of $M_{\rho^{**}}$. Finally, we define
\begin{align}\label{eq::tau}
\tau^*(\varepsilon):=\frac{1}{3}d((\zeta_{\rho^*+\varepsilon}(A_1),\zeta_{\rho^*+\varepsilon}(A_2)),
\qquad \qquad
\varepsilon\in (0,\rho^{**}-\rho^{*}].
\end{align}
\end{definition}
\begin{figure*}[htbp]
\begin{minipage}[t]{0.99\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/defcluster.png}
\end{minipage}
\centering
\caption{Definition of clusters.
Left: A one-dimensional mixture of three Guassians with the optimal level $\rho^*$ and a possible choice of $\rho^{**}$. It is easily observed that the open intervals $(x_1, x_2)$ and $(x_2, x_3)$ are the two clusters of the distribution. We only have one connected component for the level $\rho_1 < \rho^*$ and the levels $\rho_3$ and $\rho_4$ are not considered in above definition.
Right: Here we have a similar situation for a mixture of three two-dimensional Gaussians drawn by contour lines. The thick solid lines indicate the levels $\rho^*$ and $\rho^{**}$, while the thin solid lines show a level in $(\rho^*, \rho^{**})$. The dashed lines correspond to a level $\rho \leq \rho^*$ and a level $\rho \geq \rho^{**}$. In this case, the clusters are the two connected components by the outer thick solid line.}
\label{fig:defcluster}
\end{figure*}
To illustrate, the above definition ensures that the level set below $\rho^*$ are connected, while there are exactly two components in the level sets for a certain range above $\rho^*$. To notify, any two level sets between this range are supposed to be comparable. As a result, the topological structure between $\rho^*$ and $\rho^{**}$ can be determined by that of $M_{\rho^{**}}$. In this manner, the connected components of $M_{\rho}$, $\rho\in (\rho^*,\rho^{**})$ can be numbered by the connected components of $M_{\rho^{**}}$. This numbering procedure can be clearly reflected from the definition of the clusters $A_i^*$ as well as that of the function $\tau^*$, which in essence measures the distance between the two connected components at level $\rho^*+\varepsilon$.
Concerning that the quantification of uncertainty of clusters is indispensable, we need to introduce for $A\subset \mathcal{X}$, $\delta>0$, the sets
\begin{align}
A^{+\delta}&:=\{x\in \mathcal{X}:d(x,A)\leq \delta\},
\nonumber\\
A^{-\delta}&:=\mathcal{X}\setminus (\mathcal{X}\setminus A)^{+\delta}.
\label{DeltaTubes}
\end{align}
In other words, $A^{+\delta}$ can be recognized as adding a $\delta$-tube to $A$, while $A^{-\delta}$ is treated as removing a $\delta$-tube from $A$. We are expected to avoid cases where the density level sets have bridges or cusps that are too thin. To be more precise, recall that for a closed $A\subset \mathbb{R}^d$, the function $\psi^*_A:(0,\infty)\to [0,\infty)$ is defined by
\begin{align*}
\psi^*_A(\delta):=\sup_{x\in A} d(x,A^{-\delta}),
\qquad \qquad
\delta > 0.
\end{align*}
Particularly, for all $\delta>0$, we have $\psi^*_A(\delta)\geq \delta$ for all $\delta>0$, and if $A^{-\delta}=\emptyset$, then $\psi^*_A(\delta)=\infty$. Consequently, according to Lemma A.4.3 in \cite{steinwart2015suppA}, for all $\delta>0$ with $A^{-\delta}\neq \emptyset$ and all $\tau>2\psi^*(\delta)$, we have
\begin{align*}
|\mathcal{C}_\tau(A^{-\delta})|\leq |\mathcal{C}(A)|,
\end{align*}
whenever $A$ is contained in some compact $\mathcal{X}\subset \mathbb{R}^d$ and $|\mathcal{C}(A)|< \infty$.
With the preceding preparations, we now come to the following definition excluding bridges and cusps which are too thin.
\begin{definition} \label{def::ThickLetSets}
Let $\mathcal{X}\subset \mathbb{R}^d$ be a compact and connected set, and $\mathrm{P}$ be a $\mu$-absolutely continuous distribution that is normal. Then we say that \emph{$\mathrm{P}$ has thick level sets of order $\gamma\in (0,1]$} up to the level $\rho^{**}>0$, if there exits constants $c_{\textit{thick}}\geq 1$ and $\delta_{\textit{thick}}\in (0,1]$ such that, for all $\delta\in (0,\delta_{\textit{thick}}]$ and $\rho\in [0,\rho^{**}]$, we have
\begin{align*}
\psi^*_{M_\rho}(\delta)
\leq c_{\textit{thick}} \delta^{\gamma}.
\end{align*}
In this case, we call $\psi(\delta):=3c_{\textit{thick}}\delta^\gamma$ the thickness function of $\mathrm{P}$.
\end{definition}
\begin{figure*}[htbp]
\begin{minipage}[t]{0.99\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/thicklevelset.png}
\end{minipage}
\centering
\caption{Thick level sets.
Left: The thick solid line presents a level set $M_{\rho}$ below or at the level $\rho^*$ and the thin solid line indicates the two clusters $B'$ and $B''$ of $M_{\rho}^{-\delta}$. Since the quadratic shape of $M_{\rho}$ around the thin bridge, the distribution has thickness of order $\gamma = 1/2$.
Right: In the same situation, the distribution has thick level sets of order $\gamma = 1$. It is worth noting that smaller $\gamma$ leads to a significantly wider separation of $B'$ and $B''$.}
\label{fig:thicklevelset}
\end{figure*}
In order to describe the distribution we wish to cluster, we now make the following assumption based on all concepts introduced so far.
\begin{assumption} \label{ass::MainDistribution}
The distribution $\mathrm{P}$ with bounded density $f$ is able to be clustered between $\rho^*$ and $\rho^{**}$. Moreover, $\mathrm{P}$ has thick level sets of order $\gamma\in (0,1]$ up to the level $\rho^{**}$. The corresponding thickness function is denoted by $\psi$ and the function defined in \eqref{eq::tau} is abbreviated as $\tau^*$.
\end{assumption}
In the case that all level sets are connected, we introduce the following assumption
to investigate the behavior of the algorithm in situations in which $\mathrm{P}$ cannot be clustered.
\begin{assumption} \label{ass::MainDistributionaddi}
Let $\mathcal{X}\subset \mathbb{R}^d$ be a compact and connected set, and $\mathrm{P}$ be a $\mu$-absolutely continuous distribution that is normal.
Assume that there exist constants $\rho_*\geq 0$, $\gamma \in (0,1]$, $c_{\textit{thick}}\geq 1$ and $\delta_{\textit{thick}}\in(0,1]$ such that for all $\rho \geq \rho_*$ and $\delta \in (0,\delta_{\textit{thick}}]$, the following conditions hold:
\begin{itemize}
\item[(i)]
$|\mathcal{C}(M_\rho)|\leq 1$.
\item[(ii)]
If $M_\rho^{-\delta}\neq \emptyset$ then $\psi_{M_\rho}^*(\delta) \leq c_{\textit{thick}} \delta^{\gamma}$.
\item[(iii)]
If $M_\rho^{-\delta}=\emptyset$, then $|\mathcal{C}_{\tau}(A)| = 1$ for all non-empty $A\subset M_\rho^{+\delta}$ and all $\tau>2c_{\textit{thick}} \delta^\gamma$.
\item[(iv)]
For each $\delta \in (0,\delta_{\textit{thick}}]$ there exists a $\rho \geq \rho_*$ with $M_{\rho}^{-\delta}=\emptyset$.
\end{itemize}
\end{assumption}
\subsection{Best-scored Random Forest Density Estimation} \label{subsec::Best-scoredRandomForestDensityEstimation}
Considering the fact that the density estimation should come first before the analysis on the level sets, we dedicate this subsection to the methodology of building an appropriate density estimator. Different from the usual histogram density estimation \citep{steinwart2015adaptive} and kernel density estimation \citep{steinwart2017adaptive}, this paper adopts a novel random forest-based density estimation strategy, namely the best-scored random forest density estimation proposed recently by \cite{hang2018best}.
\subsubsection{Purely Random Density Tree} \label{subsubsec::PurelyRandomDensityTree}
Recall that each tree in the best-scored random forest is established based on a purely random partition followed the idea of \cite{breiman2000some}. To give a clear description of one possible construction procedure of this purely random partition, we introduce the random vector $Q_i:=(L_i,R_i,S_i)$ as in \cite{hang2018best}, which represents the building mechanism at the $i$-th step. To be specific,
\begin{figure*}[htbp]
\begin{minipage}[t]{0.99\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/ap.png}
\end{minipage}
\centering
\caption{Possible construction procedures of three-split axis-parallel purely random partitions in a two-dimensional space. The first split divides the input domain, e.g.~$B_r$ into two cells $A_{1, 1}$ and $A_{1, 2}$. Then, the to-be-split cell is chosen uniformly at random, say $A_{1, 2}$, and the partition becomes $A_{1,1}$, $A_{2,1}$, $A_{2,2}$ after the second random split. Finally, we once again choose one cell uniformly at random, say $A_{1, 1}$, and the third random split leads to a partition consisting of $A_{2, 1}$, $A_{2,2}$, $A_{3, 1}$ and $A_{3, 2}$.}
\label{fig:ap}
\end{figure*}
\begin{itemize}
\item
[$L_i$] denotes the to-be-split cell at the $i$-th step chosen uniformly at random from all cells formed in the $(i-1)$-th step;
\item
[$R_i$] $\in \{1, \ldots, d\}$ stands for the dimension chosen to be split from in the $i$-th step where each dimension has the same probability to be selected, that is, $\{ R_i, i \in \mathbb{N}_{+} \}$ are i.i.d.~multinomial distributed with equal probabilities;
\item
[$S_i$] is a proportional factor standing for the ratio between the length of the newly generated cell in the $R_i$-th dimension after the $i$-th split and the length of the being-cut cell $L_i$ in the $R_i$-th dimension. We emphasize that $\{S_i, i \in \mathbb{N}_{+}\}$ are i.i.d.~drawn from the uniform distribution $\mathrm{Unif}[0, 1]$.
\end{itemize}
In this manner, the above splitting procedure leads to a so-called partition variable $Z:=(Q_1,\ldots, Q_p,\ldots) \in \mathcal{Z}$ with probability measure of $Z$ denoted by $\mathrm{P}_Z$, and any specific partition variable $Z \in \mathcal{Z}$ can be treated as a splitting criterion. Moreover, for the sake of notation clarity, we denote by $\mathcal{A}_{(Q_1, \ldots Q_p)}$ the collection of non-overlapping cells formed after conducting $p$ splits on $B_r$ following $Z$. This can be further abbreviated as $\mathcal{A}_{Z, p}$ which exactly represents a random partition on $B_r$. Accordingly, we have $\mathcal{A}_{Z, 0} := B_r$, and for certain sample $x \in B_r$, the cell where it falls is denoted by $A_{Z, p}(x)$.
In order to better characterize the purely random density tree, we give another expression of the random partition on $B_r$, which is $\mathcal{A}_{Z, p}:=\{A_j, j=1,\ldots,p\}$ where $A_j$ represents one of the resulting cells of this partition.
Based on this partition, we can build the random density tree with respect to probability measure $Q$ on $\mathbb{R}^d$, denoted as $f_{Q, Z, p}: \mathbb {R}^d \to [0, \infty)$, defined by
\begin{align*}
f_{Q, Z}(x) := f_{Q, Z, p}(x):= \sum_{j=0}^p \frac{Q(A_j)\boldsymbol{1}_{A_j}(x)}{\mu(A_j)} + \frac{Q(B_r^c)\boldsymbol{1}_{B_r^c}(x)}{\mu(B_r^c)}
\end{align*}
where unless otherwise stated, we assume that for all $A \in \mathcal{A}_{Z, p}$, the Lebesgue measure $\mu(A)>0$. In this regard, when taking $Q = \mathrm{P}$, the density tree decision rule becomes
\begin{align*}
f_{\mathrm{P}, Z}(x) = \frac{\mathrm{P}(A(x))}{\mu(A(x))}=\frac{1}{\mu(A(x))} \int_{A(x)} f(x') d\mu(x),
\qquad \qquad
x \in B_r,
\end{align*}
where $A(x) := A_j$. When taking $Q$ to be the empirical measure $D_n := \frac{1}{n}\sum_{i=1}^n \delta_{x_i}$, we obtain
\begin{align*}
D_n(A(x)) = \mathbb{E}_{D_n}\boldsymbol{1}_{A(x)}
=\frac{1}{n}\sum_{i=1}^n \delta_{x_i}(A(x))
=\frac{1}{n}\sum_{i=1}^n \boldsymbol{1}_{A(x)}(x_i),
\end{align*}
and hence the density tree turns into
\begin{align} \label{RandomDensityTree}
f_{D, Z}:=f_{D_n, Z}(x) = \frac{D_n(A(x))}{\mu(A(x))}
= \frac{1}{n \mu(A(x))} \sum_{i=1}^n \boldsymbol{1}_{A(x)}(x_i).
\end{align}
\subsubsection{Best-scored Random Density Trees and Forest} \label{subsubsec::Best-scoredRandomDensityTreesandForest}
Considering the fact that the above partitions completely make no use of the sample information, the prediction results of their ensemble forest may not be accurate enough.
In order to improve the prediction accuracy, we select one partition for tree construction out of $k$ candidates with the best density estimation performance according to certain
performance measure such as \emph{ANLL} \citep[Section 5.4]{hang2018best}.
The resulting trees are then called the \emph{best-scored random density trees}.
Now, let $f_{D,Z_t}$, $1\leq t\leq m$ be the $m$ best-scored random density tree estimators generated by the splitting criteria $Z_1,\dots,Z_m$ respectively, which is defined by
\begin{align*}
f_{D,Z_t}(x)
:=\sum^p_{j=0} \frac{D(A_{tj})1_{A_{t_j}}(x)}{\mu(A_{t_j})}+\frac{D(B_r^c)1_{B_r^c}(x)}{\mu(B_r^c)}
\end{align*}
where $\mathcal{A}_{Z_t}:=\{A_{tj},\, j=0,\dots,p\}$ is a random partition of $B_r$. Then the best-scored random density forest can be formulated by
\begin{align}\label{BRF}
f_{D,Z_{\mathrm{E}}}(x):=\frac{1}{m} \sum^m_{t=1} f_{D,Z_t}(x),
\end{align}
and its population version is denoted by $f_{\mathrm{P},Z_{\mathrm{E}}}$.
\section{A Generic Clustering Algorithm} \label{sec::AGenericClusteringAlgorithmBasedonBRDF}
In this section, we present a generic clustering algorithm, where the clusters are
estimated with the help of a generic level set estimator which can be specified later by histogram, kernel, or random forest density estimators.
To this end,
let the optimal level $\rho^*$ and the resulting clusters $A_i^*$, $i = 1, 2$ for distributions
be as in Definition \ref{def::ClusterProperty}, and the constant $\rho_*$ be as in Assumption \ref{ass::MainDistributionaddi}.
The goal of this section is to investigate whether $\rho^*$ or $\rho_*$
is possible to be estimated and $A_i^*$, $i = 1, 2$ can be clustered.
Let us first recall some more notations introduced in Section \ref{sec::Preliminaries}.
For a $\mu$-absolutely continuous distribution $\mathrm{P}$, let
the level $\rho^{**}$, the level set $M_{\rho}$, $\rho \geq 0$,
and the function $\tau^*$
be as in Definition \ref{def::ClusterProperty}.
Furthermore, for a fixed set $A$, its $\delta$-tubes $A^{-\delta}$ and $A^{+\delta}$ are defined by \eqref{DeltaTubes}.
Moreover,
concerning with the thick level sets,
the constant $\delta_{\text{thick}}$ and the function $\psi(\delta)$ are introduced
by Definition \ref{def::ThickLetSets}.
In what follows, let $(L_\rho)_{\rho\geq 0}$ always be a decreasing family of sets $L_\rho\subset \mathcal{X}$ such that
\begin{equation} \label{UncertaintyControl}
M_{\rho+\varepsilon}^{-\delta} \subset L_{\rho} \subset M_{\rho-\varepsilon}^{+\delta}
\end{equation}
holds for all $\rho \in [0, \rho^{**}]$.
The following theorem relates the component structure of a family of level sets estimators $L_\rho$, which is a decreasing family of subsets of $\mathcal{X}$, to the component structure of certain sets $M_{\rho+\varepsilon}^{-\delta}$,
more details see e.g., \cite{steinwart2015adaptive}.
\begin{theorem} \label{thr::taoCondition}
Let Assumption \ref{ass::MainDistribution} hold.
Furthermore,
for $\varepsilon^*\in (0,\rho^{**}-\rho^*)$, let
$\varepsilon \in (0,\varepsilon^*]$,
$\delta\in (0,\delta_{\text{thick}}]$,
$\tau \in (\psi(\delta),\tau^*(\varepsilon^*))$,
and $(L_\rho)_{\rho\geq 0}$ be as in \eqref{UncertaintyControl}.
Then, for all $\rho\in [0,\rho^{**}-3\varepsilon]$ and the corresponding CRMs $\zeta:\mathcal{C}_{\tau}(M_{\rho+\varepsilon}^{-\delta})\to \mathcal{C}_{\tau}(L_\rho)$, the following disjoint union holds:
\begin{align*}\label{the:1}
\mathcal{C}_\tau(L_\rho)=\zeta(\mathcal{C}_\tau(M_{\rho+\varepsilon}^{-\delta}))\cup \{B'\in \mathcal{C}_\tau(L_\rho):B'\cap L_{\rho+2\varepsilon}=\varnothing\}.
\end{align*}
\end{theorem}
From
Theorem \ref{thr::taoCondition} we see that for suitable $\varepsilon$, $\delta$, and $\tau$, all $\tau$-connected components $B'$ of $L_{\rho}$ are either contained in $\zeta(\mathcal{C}_\tau(M_{\rho+\varepsilon}^{-\delta}))$, or vanish at level $\rho+2\varepsilon$. Accordingly, carrying out these steps precisely, we obtain a generic clustering strategy shown in Algorithm \ref{alg::clustering}.
\begin{algorithm}
\caption{Estimate clusters with the help of a generic level set estimator}
\label{alg::clustering}
\KwIn{
some $\tau>0$, $\varepsilon>0$ and a start level $\rho_0\geq 0$. A decreasing family $(L_{\rho})_{\rho \geq 0}$ of subsets of X.
}
$\rho = \rho_0$\\
\Repeat
{$M\neq 1$}
{
Identify the $\tau$-connected components $B_1',\dots,B_M'$ of $L_{\rho}$ satisfying
\begin{align*}
B_i' \cap L_{\rho+2\varepsilon}\neq \emptyset
\end{align*}
$\rho=\rho+\varepsilon$
}
$\rho=\rho+2\varepsilon$\\
Identify the $\tau$-connected components $B_1',\dots,B_M'$ of $L_{\rho}$ satisfying
\begin{align*}
B_i'\cap L_{\rho+2\varepsilon}\neq \emptyset
\end{align*}
\If {$M > 1$}
{\Return {$\rho_{\textit{out}}=\rho$ and the sets $B_i'$ for $i=1,\ldots,M$.}}
\Else {\Return {$\rho_{\textit{out}}=\rho_0$ and the set $L_{\rho_0}$.}}
\KwOut{An estimator of $\rho_*$ or $\rho^*$ the corresponding clusters.}
\end{algorithm}
\begin{figure*}[htbp]
\begin{minipage}[t]{0.99\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/illualg.png}
\end{minipage}
\centering
\caption{Illustration of Algorithm \ref{alg::clustering}.
Left: The density presented by solid line has two modes on the left and a flat part on the right. A plug in approach based on a density estimator (thin solid line) with three modes is used to provide the level set estimator. The level set estimator $L_{\rho}$ satisfies \eqref{UncertaintyControl}. Only the left two components of $L_{\rho}$ do not vanish at $\rho+2\varepsilon$. Therefore, the algorithm only finds one component.
Right: We consider the same distribution at a higher level. In this case, both components of $L_\rho$ do not vanish at $\rho+2\varepsilon$ and thus the algorithm correctly identifies two connected components. }
\label{fig:illualg}
\end{figure*}
Under Assumptions \ref{ass::MainDistribution} and \ref{ass::MainDistributionaddi},
the following theorem bounds the level $\rho_{\textit{out}}$ and the components $B_i(D)$, $i = 1, 2$, and the start level $\rho_0$ and the corresponding single cluster $M_{\rho_0} =: L_0$,
respectively, which are outputs returned by Algorithm \ref{alg::clustering}.
\begin{theorem} \label{thr::LevelSetBound}
\textit{(i)}
Let Assumption \ref{ass::MainDistribution} hold.
For $\varepsilon^*\leq (\rho^{**}-\rho^*)/9$,
let $\varepsilon \in (0,\varepsilon^*]$,
$\delta \in (0, \delta_{\textit{thick}}]$,
$\tau \in (\psi(\delta),\tau^*(\varepsilon^*)]$, and
$(L_\rho)_{\rho\geq 0}$ satisfy \eqref{UncertaintyControl} for all $\rho \geq \rho_0$.
Then, for any data set $D$,
the following statements hold for Algorithm \ref{alg::clustering}:
\begin{itemize}
\item[(a)] The returned level $\rho_{\textit{out}}$ satisfies both $\rho_{\textit{out}}\in [\rho^*+2\varepsilon,\rho^*+\varepsilon^*+5\varepsilon]$ and
\begin{equation*}
\tau-\psi(\delta)<3\tau^*(\rho_{\textit{out}}-\rho^*+\varepsilon);
\end{equation*}
\item[(b)] The returned sets $B_i(D)$, $i = 1, 2$, can be ordered such that
\begin{equation} \label{equ:6}
\sum^2_{i=1}\mu(B_i(D)\triangle A_i^*)\leq 2\sum^2_{i=1}\mu(A_i^*\setminus (A_{\rho_{\textit{out}}+\varepsilon}^i)^{-\delta})+\mu(M_{\rho_{\textit{out}}-\varepsilon}^{+\delta}\setminus \{f>\rho^*\}).
\end{equation}
Here, $A_{\rho_{\textit{out}}+\varepsilon}^i\in\mathcal{C}(M_{\rho_{\textit{out}}+\varepsilon})$, $i = 1, 2$, are ordered in the sense of $A_{\rho_{\textit{out}}+\varepsilon}^i\subset A_i^*$.
\end{itemize}
\textit{(ii)}
Let Assumption \ref{ass::MainDistributionaddi} hold. Moreover,
let $\varepsilon > 0$, $\delta \in (0, \delta_{\textit{thick}}]$ be fixed, $\tau > 2 c_{\textit{thick}} \delta^\gamma$,
and $(L_{\rho})_{\rho\geq 0}$ satisfy \eqref{UncertaintyControl} for all $\rho \geq \rho_0$.
If $\rho_0 \geq \rho_*$, then Algorithm \ref{alg::clustering} returns
the start level $\rho_0$ and
the corresponding single cluster $L_0:=M_{\rho_0}$ such that
\begin{align*}
\mu(L_{\rho_0} \triangle \widehat{M}_{\rho_{*}}) \leq \mu(M_{\rho_{0}-\varepsilon}^{+\delta} \backslash \widehat{M}_{\rho_{*}})+\mu\left(\widehat{M}_{\rho_{*}} \backslash M_{\rho_{0}+\varepsilon}^{-\delta}\right)
\end{align*}
where $\widehat{M}_{\rho_{*}}:=\bigcup_{\rho>\rho_{*}} M_{\rho_{*}}$.
\end{theorem}
The above analysis is mainly illustrated on the general cases where we assume that the underlying density has already been successfully estimated. Therefore, in the following, we delve into the characteristic of components structure and other properties of clustering algorithm under the condition where the density is estimated by the forest density estimator \eqref{BRF}.
Note that one more notation is necessary for clear understanding: One way to define level set estimators with the help of the forest density estimator \eqref{BRF}
is a simple plug-in approach, which is
\begin{align*}
L_{D, \rho} := \{ f_{D,Z_{\mathrm{E}}}(x) \geq \rho \}.
\end{align*}
However, these level set estimators are too complicated to compute the $\tau$-connected components in Algorithm \ref{alg::clustering}. Instead, we take level set estimators of the form
\begin{align}\label{LevelSetEstimator}
L_{D,\rho} := \{ x \in D : f_{D,Z_{\mathrm{E}}}(x) \geq \rho \}^{+\sigma}.
\end{align}
The following theorem shows that
some kind of uncertainty control of the form \eqref{UncertaintyControl} is valid for
level set estimators of the form \eqref{LevelSetEstimator}
induced by the forest density estimator \eqref{BRF}.
\begin{theorem} \label{theorem::UncertaintyControlForest}
Let $\mathrm{P}$ be a $\mu$-absolutely continuous distribution on $\mathcal{X}$ and
$f_{D,Z_{\mathrm{E}}}(x)$ be the forest density estimator \eqref{BRF} with $\|f_{D,Z_{\mathrm{E}}}-f_{\mathrm{P},Z_{\mathrm{E}}}\|_{\infty}\leq \varepsilon$.
For any $A \in \mathcal{A}_{Z_t,p}$, $t = 1, \ldots, m$, that is, $A$ is one of the $p+1$ cells in
the $t$-th partition, there exists a constant $\delta > 0$ such that
$\mathrm{diam}(A) \leq \delta$.
Then, for all $\rho > 0$ and $\sigma \geq \delta$, there holds
\begin{align} \label{UncertaintyControlForest}
M_{\rho+\varepsilon}^{-2\sigma}
\subset L_{D,\rho}
\subset M_{\rho-\varepsilon}^{+2\sigma}.
\end{align}
\end{theorem}
Before we present the next theorem, recall that
$r$ denotes half of the side length of the centered hypercube in $\mathbb{R}^d$ and
$m$ denotes the number of trees in the best-scored random forest.
\begin{theorem}\label{thr::BoundProbability}
Let $\mathrm{P}$ be a $\mu$-absolutely continuous distribution on $\mathcal{X}$.
For $r \geq 1$, $m > 0$, $\varsigma \geq 1$, $n \geq 1$, we choose an $\varepsilon > 0$ satisfying
\begin{align} \label{BoundVarepsilon}
\varepsilon
\geq \sqrt{\|f\|_{\infty} \mathcal{E}_{\varsigma, p}/ n}
+ \mathcal{E}_{\varsigma, p} / (3n)
+ 2/n,
\end{align}
where $\mathcal{E}_{\varsigma,p}$ is defined by
\begin{align} \label{eq::CalE_n}
\mathcal{E}_{\varsigma,p} :=
128 m^2 p^{2a} \mu(B_r)^{-1} e^{2 \varsigma}
\bigl( (8 d + 1) (\log(4m) + \varsigma) + 23 \log n + 8 a d \log p \bigr).
\end{align}
Furthermore,
for $\delta \in (0,\delta_{\textit{thick}}/2]$ and $\tau > 0$, we choose a $\sigma$ with $\sigma\geq \delta$ and assume this $\sigma$ satisfying $\sigma <\delta_{\text{thick}}/2$ and $\psi(2\sigma)<\tau$.
Moreover, for each random density tree, we pick the number of splits
$p$ satisfying
\begin{align} \label{BoundSplits}
p > \bigl( 2 m d e^{\varsigma} / \delta \bigr)^{4 d / c_T}.
\end{align}
If we feed Algorithm \ref{alg::clustering} with parameters $\varepsilon$, $\tau$, $\sigma$, and $(L_{D,\rho})_{\rho\geq 0}$ as in \eqref{LevelSetEstimator},
then the following statements hold: \\
{~} \textit{(i)}
If $\mathrm{P}$ satisfies Assumption \ref{ass::MainDistribution} and
there exists an $\varepsilon^*$ satisfying
\begin{align*}
\varepsilon+\inf\{\varepsilon'\in(0,\rho^{**}-\rho^*]:\tau^*(\varepsilon')\geq \tau\}
\leq \varepsilon^*
\leq (\rho^{**}-\rho^*) / 9,
\end{align*}
then with probability $\mathrm{P}^n$ not less than $1-e^{-\varsigma}$, the following statements hold:
\begin{itemize}
\item[(a)]
The returned level $\rho_{D,\textit{out}}$ satisfies both $\rho_{D,\textit{out}}\in [\rho^*+2\varepsilon,\rho^*+\varepsilon^*+5\varepsilon]$ and
\begin{align*}
\tau-\psi(2\sigma)<3\tau^*(\rho_{D,\textit{out}}-\rho^*+\varepsilon);
\end{align*}
\item[(b)] The returned sets $B_i(D)$, $i = 1, 2$, can be ordered such that
\begin{align*}
\sum^2_{i=1}\mu(B_i(D)\triangle A_i^*)\leq 2\sum^2_{i=1}\mu(A_i^*\setminus (A^i_{\rho_{D,\textit{out}}+\varepsilon})^{-2\sigma})+\mu(M^{+2\sigma}_{\rho_{D,\textit{out}}-\varepsilon}
\setminus \{h>\rho^*\}).
\end{align*}
Here, $A^i_{\rho_{D,\textit{out}}+\varepsilon}\in \mathcal{C}(M_{\rho_{D,\textit{out}}+\varepsilon})$, $i = 1, 2$, are ordered in the sense of $A^i_{\rho_{D,\textit{out}}+\varepsilon}\in A_i^*$.
\end{itemize}
{~}\textit{(ii)}
If $\mathrm{P}$ satisfies Assumption \ref{ass::MainDistributionaddi} and $\rho_0 \geq \rho^*$, then
\begin{align*}
\mu(L_{\rho_0} \triangle \widehat{M}_{\rho_*})
\leq \mu(M^{+2\sigma}_{\rho_0-\varepsilon} \setminus \widehat{M}_{\rho_*})
+ \mu(\widehat{M}_{\rho_*} \setminus M^{-2\sigma}_{\rho_0+\varepsilon})
\end{align*}
holds with probability $\mathrm{P}^n$ not less than $1-e^{-\varsigma}$ for the returned level $\rho_0$ and the corresponding single cluster $L_0 := M_{\rho_0}$,
where $\widehat{M}_{\rho_*}:=\bigcup_{\rho>\rho_*} M_{\rho}$.
\end{theorem}
\section{Main Results} \label{sec::ConsistencyandRatesforRFDE-basedClustering}
In this section, we present main theoretical results
of our best-scored clustering forest
on the consistency as well as
convergence rates
for both the optimal level $\rho^*$ and the true clusters $A_i^*$, $i = 1, 2$, simultaneously
using the error bounds derived in Theorem \ref{thr::LevelSetBound} and Theorem \ref{thr::BoundProbability}, respectively.
We also present some comments and discussions on the obtained theoretical results.
\subsection{Consistency for Best-scored Clustering Forest}
\begin{theorem}[Consistency] \label{the::Consistency}
Let Assumption \ref{ass::MainDistribution} hold. Furthermore,
for certain constant $q \in (0, 1)$,
assume that $(\varepsilon_n)$, $(\tau_n)$, $(\delta_n)$, and $(\sigma_n)$ are strictly positive sequences converging to zero satisfying
$\varepsilon_n\geq 2/(nq)$ for sufficiently large $n$,
$\sigma_n = \delta_n$, $\psi(2\sigma_n)\leq \tau_n$.
Moreover,
let the number of splits $p_n$ satisfy
\begin{align*}
\lim_{n \to \infty} n p_n^{-2a} (\log n)^{-1} \varepsilon_n^2
& = \infty,
\\
\lim_{n \to \infty} \delta_n p_n^{c_T/(4d)}
& = \infty,
\end{align*}
where $c_T = 0.22$ and $a = 4.33$.
If we feed Algorithm \ref{alg::clustering} with parameters $\varepsilon_n$, $\sigma_n$, $(L_{D,\rho})_{\rho\geq 0}$ as in \eqref{LevelSetEstimator}, and $p_n$,
then the following statements hold:
\begin{itemize}
\item[(i)]
If $\mathrm{P}$ satisfies Assumption \ref{ass::MainDistribution}, then for all $\epsilon > 0$, the returned level $\rho_{D,\textit{out}}$ satisfies
\begin{align*}
\lim_{n\to \infty} \mathrm{P}^n \bigl(
\{D\in X^n: 0<\rho_{D,\textit{out}}-\rho^*\leq \epsilon \} \bigr) = 1.
\end{align*}
Moreover, if $\mu(\overline{A_1^*\cup A_2^*}\setminus(A_1^*\cup A_2^*))=0$, then for all $\epsilon > 0$,
the returned sets $B_i(D)$, $i = 1, 2$, satisfy
\begin{align*}
\lim_{n\to\infty} \mathrm{P}^n
\biggl( \biggl\{ D \in X^n :\sum_{i=1}^2 \mu(B_i(D) \triangle A_i^*) \leq \epsilon \biggr\} \biggr) = 1.
\end{align*}
\item[(ii)]
If $\mathrm{P}$ satisfies Assumption \ref{ass::MainDistributionaddi} and $\rho_*=0$, then for all $\epsilon > 0$, the returned level $\rho_{D,\textit{out}}$ satisfies
\begin{align*}
\lim_{n\rightarrow \infty} \mathrm{P}^n(\{D\in X^n:0<\rho_{D,\textit{out}}\leq \epsilon\} )=1.
\end{align*}
Moreover, if $\mu(\overline{\{f>0\}}\setminus \{f>0\})=0$,
then for all $\epsilon > 0$,
the returned set $L_{D,\rho_{D,\textit{out}}}$ satisfies
\begin{align*}
\lim_{n\rightarrow }\mathrm{P}^n(\{D\in X^n:\mu(L_{D,\rho_{D,\textit{out}}}\triangle \{f>0\})\leq \epsilon\})=1;
\end{align*}
\end{itemize}
\end{theorem}
\subsection{Convergence Rates for Best-scored Clustering Forest}
In this subsection, we derive the convergence rates
for both estimation problems, that is, for estimating the optimal level $\rho^*$ and the true clusters $A_i^*$, $i = 1, 2$, in our proposed algorithm separately.
\subsubsection{Convergence Rates for Estimating the Optimal Level}
In order to derive the convergence rates for estimating the optimal level $\rho^*$,
we need to make following assumption
that describes how well the clusters are separated above
$\rho^*$.
\begin{definition} \label{def::SeparationExponent}
Let Assumption \ref{ass::MainDistribution} hold. The clusters of $\mathrm{P}$
are said to have \emph{separation exponent} $\kappa \in (0,\infty]$ if there exists a constant $\underline{c}_{\textit{sep}}>0$ such that
\begin{align*}
\tau^*(\varepsilon)\geq \underline{c}_{\textit{sep}}\varepsilon^{1/\kappa}
\end{align*}
holds for all $\varepsilon\in (0,\rho^{**}-\rho^*]$.
Moreover, the separation exponent $\kappa$ is called \emph{exact} if there exists another constant $\overline{c}_{\textit{sep}}>0$ such that
\begin{align*}
\tau^*(\varepsilon)\leq \overline{c}_{\textit{sep}}\varepsilon^{1/\kappa}
\end{align*}
holds for all $\varepsilon\in (0,\rho^{**}-\rho^*]$.
\end{definition}
The \emph{separation exponent} describes how fast the connected components of the $M_\rho$ approach each other for $\rho\to \rho^*$ and a distribution having separation exponent $\kappa$ also has separation exponent $\kappa'$ for all $\kappa'<\kappa$. If the separation exponent $\kappa=\infty$, then the clusters $A_1^*$ and $A_2^*$ do not touch each other.
With the above Definition \ref{def::SeparationExponent}, we are able to establish error bounds for estimating the optimal level $\rho^*$
in the following theorem
whose proof is quite similar to that of Theorem 4.3 in \cite{steinwart2015adaptive} and hence will be omitted.
\begin{theorem} \label{the::rho_rates}
Let Assumption \ref{ass::MainDistribution} hold,
and
assume that $\mathrm{P}$ has a bounded $\mu$-density $f$ whose clusters have separation exponent $\kappa\in (0,\infty]$.
For $r \geq 1$, $m > 0$, $\varsigma \geq 1$, $n \geq 1$, we choose an $\varepsilon > 0$ satisfying
\begin{align*}
\varepsilon \geq
\sqrt{\|f\|_{\infty} \mathcal{E}_{\varsigma, p}/ n}
+ \mathcal{E}_{\varsigma, p} / (3n)
+ 2/n,
\end{align*}
with $\mathcal{E}_{\varsigma,p}$ as in \eqref{eq::CalE_n}.
Furthermore,
for $\delta \in (0,\delta_{\textit{thick}}/2]$ and $\tau > 0$, we choose a $\sigma$ with $\sigma\geq \delta$ and assume this $\sigma$ satisfying $\sigma <\delta_{\text{thick}}/2$ and $\psi(2\sigma)<\tau/2$.
Moreover, for each random density tree,
we pick
the number of splits $p$ satisfying
\begin{align*}
p > \bigl( 2 m d (K+2r)e^\varsigma / \delta \bigr)^{4 d / c_T}.
\end{align*}
Finally, suppose that
$\varepsilon^*:=\varepsilon+(\tau/\underline{c}_{\textit{sep}})^{\kappa}$ satisfies $\varepsilon^*\leq (\rho^{**}-\rho^*)/9$.
If we feed
Algorithm \ref{alg::clustering} with
parameters $\varepsilon$, $\tau$, $\sigma$, $(L_{D,\rho})_{\rho\geq 0}$
as in \eqref{LevelSetEstimator}, and $p_n$,
then the returned level $\rho_{D,\textit{out}}$ satisfies
\begin{align} \label{RhoExessError}
\varepsilon
< \rho_{D,\textit{out}}-\rho^*
\leq (\tau/\underline{c}_{\textit{sep}})^{\kappa}+6\varepsilon
\end{align}
with probability $\mathrm{P}^n$
not less than $1-e^{-\varsigma}$.
Moreover, if the separation exponent $\kappa$ is exact and $\kappa<\infty$, then we have
\begin{align}\label{equ::tau_rho}
\rho_{D,\textit{out}}-\rho^*
> \bigl( \tau / (6 \overline{c}_{\textit{sep}} ) \bigr)^{\kappa} / 4.
\end{align}
\end{theorem}
\begin{corollary}[Convergence Rates for Estimating the Optimal Level] \label{col::rho_rates}
Let Assumption \ref{ass::MainDistribution} hold and suppose that $f$ is $\alpha$-H\"{o}lder continuous with exponent $\alpha \in (0,1]$ whose clusters have separation exponent $\kappa\in (0,\infty)$.
For any $\epsilon > 0$, and all $n \geq 1$, let $(\varepsilon_n)$, $(\tau_n)$, $(\delta_n)$, and $(\sigma_n)$ be sequences with
\begin{align*}
\varepsilon_n & = \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha} \log \log n \bigr)^{\gamma\kappa/(2(\gamma\kappa+\epsilon))},
\\
\tau_n & = \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha}\log\log n \bigr)^{\gamma/2(\gamma\kappa+\epsilon)},
\\
\sigma_n & = \delta_n = \bigl( n^{- \lambda} (\log n)^{2+\lambda} \log \log \log n \bigr)^{1/(2(\gamma\kappa+\epsilon))},
\end{align*}
where $\lambda=c_T/(c_T\alpha+4ad)$, $c_T = 0.22$ and $a = 4.33$.
Moreover, we choose the number of splits as
\begin{align*}
p_n = (n/\log n)^{2d/(c_T \alpha + 4 a d)}.
\end{align*}
If we feed Algorithm \ref{alg::clustering} with parameters $\varepsilon_n$, $\tau_n$, $\sigma_n$, $(L_{D,\rho})_{\rho\geq 0}$, and $p_n$,
then for all sufficiently large n,
there exists a constant $\overline{c} \geq 1$ such that the returned level $\rho_{D,\textit{out}}$
satisfies
\begin{align*}
\mathrm{P}^n( \rho_{D,\textit{out}}-\rho^*\leq \overline{c}\varepsilon_n )\geq 1 - 1/\log n.
\end{align*}
Moreover, if the separation exponent $\kappa$ is exact, there exists another constant $\underline{c} \geq 1$ such that for all sufficiently large $n$, there holds
\begin{align*}
\mathrm{P}^n( \underline{c}\varepsilon_n\leq \rho_{D,\textit{out}}-\rho^*\leq \overline{c}\varepsilon_n) \geq 1 - 1/\log n.
\end{align*}
\end{corollary}
\subsubsection{Convergence Rates for Estimating the True Clusters}
Our next goal is to
establish learning rates for the true clusters, in other words, describing how fast
$\sum_{i=1}^2 \mu( B_i(D) \triangle A_i^*)$ goes to $0$.
On account that this is a modified level set estimation problem, we need to make some further assumptions on $\mathrm{P}$. The first definition can be considered as a one-sided variant of a well-known condition introduced by \cite[Theorem 3.6]{polonik1995measuring}.
\begin{definition}
Let $\mu$ be a measure on $\mathcal{X}$ and $\mathrm{P}$ be a distribution on $\mathcal{X}$ that has a $\mu$-density $f$. For a given level $\rho\geq 0$, we say that \emph{$\mathrm{P}$ has flatness exponent $\vartheta \in (0,\infty]$} if there exists a constant $c_{\textit{flat}}>0$ such that for all $s > 0$, we have
\begin{align} \label{ineq::flat}
\mu(\{0<f-\rho<s\})\leq (c_{\textit{flat}}s)^\vartheta.
\end{align}
\end{definition}
It can be easily observed from \eqref{ineq::flat} that the larger $\vartheta$ is, the steeper $f$ approaches $\rho$ from above. Particularly, in the case of $\vartheta = \infty$, the density $f$ is allowed to take the value $\rho$, otherwise it would be bounded away from $\rho$.
The next definition describes the roughness of the boundary of the clusters, see also Definition 4.6 in \cite{steinwart2015adaptive}.
\begin{definition} \label{def::SmoothBoundary}
Let Assumption \ref{ass::MainDistribution} hold. Given some $\alpha_0 \in(0,1]$, we say that \emph{the clusters have an $\alpha_0$-smooth boundary} if there exists a constant $c_{\textit{bound}}>0$ such that
for all $\rho \in (\rho^*,\rho^{**}]$ and
$\delta \in (0,\delta_{\textit{thick}}]$, there holds
\begin{align*}
\mu((A_\rho^i)^{+\delta}\setminus (A_\rho^i)^{-\delta})\leq c_{\textit{bound}}\delta^{\alpha_0},
\qquad
i = 1, 2,
\end{align*}
where $A_\rho^i$, $i = 1, 2$ denote the connected components of the level set $M_\rho$.
\end{definition}
Note that considering $\alpha > 1$ does not make sense in $\mathbb{R}^d$ and if $A \subset \mathbb{R}^d$ has rectifiable boundary, we always have $\alpha = 1$, see Lemma A.10.4 in \cite{steinwart2015suppA}.
Now, we summarize all the conditions on $\mathrm{P}$ needed to obtain learning rates for cluster estimation.
\begin{assumption} \label{ass::SetsAssumption}
Let Assumption \ref{ass::MainDistribution} hold.
Moreover, assume that
$\mathrm{P}$ has a bounded $\mu$-density $f$
and
a flatness exponent $\vartheta \in(0,\infty]$ at level $\rho^*$,
whose clusters have
an $\alpha_0$-smooth boundary for some $\alpha_0\in(0,1]$ and
a separation exponent $\kappa \in (0,\infty]$.
\end{assumption}
\begin{figure*}[htbp]
\begin{minipage}[t]{0.99\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/sepandflat.png}
\end{minipage}
\centering
\caption{Separation and flatness.
These two figures illustrate two possible shapes of the density $f$. The bold horizontal line indicates the set $\{\rho^*<f<\rho^*+\varepsilon\}$ and $3\tau^*(\varepsilon)$ describes the width of the valley at level $\rho^*+\varepsilon$. The value of $\varepsilon$ is chosen such that $3\tau^*(\varepsilon)$ on the right equals the value of the left. The density on the right has a narrower valley than that on the left. Therefore, $\varepsilon$ needs to be chosen larger. Moreover, it becomes more difficult to estimate the optimal level $\rho^*$ and the two clusters.}
\label{fig:sepandflat}
\end{figure*}
The following theorem provides a finite sample bound
that can be later used to describe
how well our algorithm estimates the true clusters $A_i^*$, $i = 1, 2$,
see also Theorem 4.7
in \cite{steinwart2015adaptive}.
\begin{theorem} \label{the::sets_rate}
Let Assumption \ref{ass::SetsAssumption} hold.
Furthermore, let $\varepsilon^*$ be defined as in Theorem \ref{the::rho_rates}.
For $r > 1$, $m > 0$, $\varsigma \geq 1$, $n \geq 1$,
if we feed Algorithm \ref{alg::clustering} with parameters $\varepsilon$, $\tau$, $\sigma$, $(L_{D,\rho})_{\rho\geq 0}$, and $p$
as in Theorem \ref{the::rho_rates}, then
the returned level $\rho_{D,\textit{out}}$
satisfy
inequalities \eqref{RhoExessError} and
the returned sets
$B_i(D)$, $i = 1, 2$ satisfy
\begin{align*}
\sum_{i=1}^2 \mu( B_i(D) \triangle A_i^*)
\leq \bigl( 7 c_{\textit{flat}} \varepsilon
+ c_{\textit{flat}} ( \tau/\underline{c}_{\textit{sep}})^{\kappa} \bigr)^{\vartheta}
+ 6 c_{\textit{bound}} (2 \sigma)^{\alpha_0}
\end{align*}
with probability $\mathrm{P}^n$ not less than $1-e^{-\varsigma}$.
\end{theorem}
Note that if the separation exponent $\kappa$ is exact and finite, then the inequality \eqref{equ::tau_rho} also holds for the returned level $\rho_{D,\textit{out}}$.
Moreover, if $\vartheta$ and $\kappa$ are of finite values, then the bound in Theorem \ref{the::sets_rate} behaves like
$$
\varepsilon^{\vartheta} + \tau^{\vartheta \kappa} + \delta^{\alpha_0}
$$
and the convergence rates are presented in the following corollary.
\begin{corollary}[Convergence Rates for Estimating the True Clusters] \label{col::sets_rate}
Let Assumption \ref{ass::SetsAssumption} hold.
Furthermore, for $n \geq 1$,
let
$(\varepsilon_n)$, $(\tau_n)$, $(\delta_n)$, $(\sigma_n)$, and $(p_n)$
be sequences with
\begin{align*}
\varepsilon_n & = \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha} \log \log n \bigr)^{\varrho/(2(\varrho+\vartheta))},
\\
\tau_n & = \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha} \log \log n \bigr)^{\vartheta\gamma/(2(\varrho+\vartheta))},
\\
\sigma_n & = \delta_n = \bigl( n^{- \lambda} (\log n)^{2+\lambda} \log \log \log n \bigr)^{\vartheta/(2(\varrho+\vartheta))},
\\
p_n & = (n / \log n)^{2d/(c_T\alpha+4ad)},
\end{align*}
where $\lambda=c_T/(c_T\alpha+4ad)$,
$c_T=0.22$, $a=4.33$,
and $\varrho:=\min\{\alpha_0,\vartheta\gamma\kappa\}$.
If we feed Algorithm \ref{alg::clustering} with parameters $\varepsilon_n$, $\tau_n$, $\sigma_n$, $(L_{D,\rho})_{\rho\geq 0}$ as in \eqref{LevelSetEstimator}, and $p_n$, then
there exists a constant $c \geq 1$ such that
the returned sets
$B_i(D)$, $i = 1, 2$, satisfy
\begin{align*}
\mathrm{P}^n \biggl( D : \sum_{i=1}^2 \mu(B_i(D) \triangle A_i^*)
\leq c \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha}\log\log n \bigr)^{\frac{\varrho\vartheta}{2(\varrho+\vartheta)}} \biggr)
\geq 1 - 1/\log n.
\end{align*}
\end{corollary}
\subsection{Comments and Discussions}
This subsection presents some comments and discussions on the established learning rates for estimating the optimal level $\rho^*$ and the true clusters
$A_i^*$, $i = 1, 2$.
First of all, let us compare our convergence rates for estimating the optimal level with existing convergence rates in the literature.
Corollary \ref{col::rho_rates} tells us that
for any $\epsilon > 0$,
our learning rate is of the form
\begin{align*}
n^{- \frac{\gamma \kappa}{2(\gamma\kappa+\epsilon)}
\cdot \frac{c_T \alpha}{c_T\alpha+4ad}},
\end{align*}
where $c_T=0.22$, $a=4.33$.
In contrast,
\cite{steinwart2015adaptive} has shown that
the clustering algorithm using histogram density estimator learns with the rate
\begin{align*}
n^{-\frac{\gamma\kappa}{2\gamma\kappa+d}}.
\end{align*}
Simple algebraic calculations show that
if $\epsilon$ is sufficiently small and $8 a \gamma \kappa < c_T \alpha$, then
this rate will be slower than ours.
However, if the best separation exponent $\kappa = \infty$, that is, the clusters $A_1^*$ and $A_2^*$ do not touch each other, then our learning rate becomes
\begin{align*}
n^{-\frac{c_T \alpha}{2(c_T\alpha+4ad)}}
\end{align*}
which turns out to be slower than the rate $n^{-\frac{1}{2}}$ established in \cite{steinwart2015adaptive}.
On the other hand, concerning with the learning rates for
estimating the true clusters,
Corollary \ref{col::sets_rate} shows that our algorithm learns with rate
\begin{align*}
n^{-\frac{\varrho\vartheta}{2(\varrho+\vartheta)} \cdot
\frac{c_T\alpha}{c_T\alpha+4ad}},
\end{align*}
where $c_T=0.22$ and $a=4.33$.
Obviously, this rate is strictly slower than the rate
$n^{-\frac{\varrho\vartheta}{2\varrho+\vartheta d}}$
derived by \cite{steinwart2015adaptive}. Nevertheless,
in the case of
$d \varrho \vartheta c_T\geq 2(4ad\varrho+c_T\alpha\vartheta+4\alpha d\vartheta)$,
it can be easily shown that our rate is faster than
the rate $n^{-\frac{\alpha}{2\alpha+d}}$ established in \cite{sriperumbudur2012consistency}.
Note that if Assumption \ref{ass::SetsAssumption} holds with
$\alpha_0 = 1$ and $\varrho \gamma \kappa \leq 1$, then the convergence rates for estimating $\rho^*$ and the clusters can be achieved simultaneously.
In contrast, in the case of $\varrho \gamma\kappa > 1$,
the estimation of $\rho^*$ is easier than the estimation of the level set $M_{\rho^*}$, more detailed discussion can be found in \cite{steinwart2015adaptive}.
Finally, we mention that in general, our convergence rates can be slower than other clustering algorithms due to the nature of random partition, which in turn leads to diversity and thus accuracy of our clustering algorithm.
\section{Experimental Performance} \label{sec::ExperimentalPerformance}
In this section, we first summarize the proposed best-scored clustering forest algorithm in Subsection \ref{subsec::AlgorithmConstruction}, and
discuss the model selection problem of various clustering algorithms in Subsection \ref{subsec::ExperimentalSetup}.
Then we compare our clustering algorithm with other proposals
both on synthetic data in Subsection \ref{subsec::SyntheticData}
and real data sets in Subsection \ref{subsec::RealDataAnalysis}, respectively.
\subsection{Algorithm Construction} \label{subsec::AlgorithmConstruction}
\begin{algorithm}[h]
\caption{Estimate clusters by best-scored random forest density estimation}
\label{alg::clustering1}
\SetAlgoNoLine
\KwIn{
$D=\{x_1,...x_n\}$, number of density trees $m$, some ratio $r>0$, $q>0$, some positive integer $k$, $k_N$, and $k_c$.
}
$\hat{f}(\cdot)$ $\leftarrow$ density estimate by random forest with the number of splits $\lfloor n*r \rfloor$ and the number $m$ of best-scored density trees each generated from $k$ random trees based on $\{x_1,\ldots,x_n\}$.\\
$\hat{D}\leftarrow \{x_j:\hat{f}(x_j)>\hat{f}(x)_q\}$, where $\hat{f}(x)_q$ denotes the $q$-quantile of $\{\hat{f}(x_i),i=1,\ldots,n\}$.\\
$G$ $\leftarrow$ $\varepsilon$ similarity graph on $\hat{D}$.\\
$j$ $\leftarrow$ 0\\
\Repeat
{$M=k_c$}
{
$j \leftarrow j+1$;\\
$\lambda_j\leftarrow\hat{f}(x)_{(j)}$, where $\hat{f}(x)_{(j)}$ denotes the $j$-th smallest value of $\{\hat{f}(x):x\in\hat{D}\}$;\\
$L_{\lambda_j}\leftarrow\{x_i\in \hat{D}:\hat{f}(x_i)\geq \lambda_j\}$;\\
$G_j \leftarrow $ subgraph of $G$ induced by $L_j$;\\
Identify the connected components $B_1',\dots,B_M'$ of $G_j$.
}
Allocate background points to these clusters with $k_N$-nearest neighbor classification.\\
\KwOut{$\rho_{D}^*:=\lambda_j$ and $k_c$ clusters.}
\end{algorithm}
Our proposed best-scored clustering forest algorithm
is presented in detail in Algorithm \ref{alg::clustering1}.
In order to measure the similarity between two data clusterings, we adopt the adjusted rand index (ARI) through all experiments which can be formulated as follows:
Given a set $S$ of n elements and two clusterings of these elements, namely $X=\{X_1,X_2,\ldots,X_r\}$ and $Y=\{Y_1,Y_2,\ldots,Y_s\}$, the overlap between $X$ and $Y$ can be summarized with $n_{ij}$ which stands for the number of objects in set $X_i\cap Y_j$,
$a_i=\sum^s_{j=1} n_{ij}$, and $b_i=\sum^r_{j=1} n_{ji}$.
Then the Adjusted Rand Index is defined as
\begin{align*}
\mathrm{ARI}
= \frac{\displaystyle \sum_{ij} \binom {n_{ij}} 2 - \biggl[ \sum_i \binom {a_i} 2 \sum_j \binom {b_j} 2 \biggr] \bigg/ \binom n 2}{\displaystyle \frac{1}{2} \biggl[ \sum_i \binom {a_i} 2+ \sum_j \binom {b_j} 2 \biggr] - \biggl[ \sum_i \binom {a_i} 2 \sum_j \binom {b_j} 2 \biggr] \bigg/ \binom n 2}.
\end{align*}
\begin{figure*}[htbp]
\centering
\begin{minipage}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/adaptive_partition.png}
\centering
\end{minipage}
\qquad
\begin{minipage}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/purely_random_partition.png}
\end{minipage}
\caption{Left: Adaptive method \emph{vs.} Right: Purely random method}
\label{fig:adaptiveVSpurelyrandom}
\end{figure*}
Finally, it's worth mentioning that
in order to improve the efficiency and accuracy of our clustering algorithm,
we also employ the adaptive splitting method (see Figure \ref{fig:adaptiveVSpurelyrandom})
which is proposed for the density estimation problem,
more details please refer to Section 5.1 in \cite{hang2018best}.
\subsection{Experimental Setup} \label{subsec::ExperimentalSetup}
In our experiments, we compare the clusters with true classes generated by computing the following performance measures ARI (adjusted rand index) of different approaches. We conduct comparisons among some baseline density-based methods including Fast Clustering Using Adaptive Density Peak Detection (ADP-Cluster), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), $k$-means and PDF-Cluster.
\begin{itemize}
\item
ADP-Cluster: The algorithm is built and improved upon the idea of \cite{2016Fast} by finding density peaks in a density-distance plot generated from local multivariate Gaussian density estimation. There are two leading parameters: the bandwidths of the multivariate kernel density and the number of the clusters $k$ determined automatically by validation criterion.
\item
DBSCAN: The algorithm can be traced back to \cite{Ester1996A}. It is also a density-based clustering non-parametric algorithm while it groups points that are closely packed together. The algorithm requires two parameters: $\varepsilon$ and the minimum number of points {\tt minPts} required to form a dense region.
\item
$k$-means: The only parameter in $k$-means is the number of cluster $k$. The idea goes back to \cite{Macqueen67somemethods}
and is popular for cluster analysis in data mining. It is significant to run diagnostic checks for determining the number of clusters in the data set.
\item
PDF-Cluster: The leading parameters in the algorithm are $h$ as bandwidth of kernel density estimation selected by least-square cross validation and $\lambda$ as tolerance threshold to set edges between two observations. The idea was proposed by \cite{Menardi2014An} developing a viable solution to the problem of finding connected sets in higher dimensional spaces.
\end{itemize}
To notify, more free parameters are alternative in the best-score clustering forest algorithm compared with other methods. To be specific, these free parameters include the number of density trees in the forest $m$, the ratio of number of splits for trees in the forest to the sample size $r$, the positive number $q$ for selecting low-density points as background points, the positive integer $k_N$ to allocate background points to clusters with $k$-NN classification as well as the number of clusters $k_c$.
For DBSCAN, the parameter $\varepsilon$ is picked from $0.01$ to $0.30$ by $0.01$, {\tt minPts} is default and $k$ is picked from $\{1,2,3,4,5\}$. For $k$-means, the parameter $k$ is selected from $\{2,3,4,5,6,7,8,9,10\}$. For PDF-Cluster, the parameter $\lambda$ is selected from 0.01 to 0.51 by 0.01 and for our method, the parameter $m$ is set to be 100, the ratio $r$ is selected from $\{0.05, 0.1, 0.15, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8\}$ and $\varepsilon$ is selected from $q_{\varepsilon}$-quantile of the pairwise distances $\{\|x_i-x_j\|_2,1\leq i<j \leq n\}$, where $q_\varepsilon$ is chosen from $\{ 0.01$, $0.03$, $0.05$, $0.07$, $0.09$, $0.12$, $0.15$, $0.20 \}$, the parameter $k_N$ of $k$-NN is selected from $\{1,2,5\}$ and the number of clusters $k_c$ is selected from $\{2,3,4,5,6\}$. It's worth pointing out that both DBSCAN and our method assigns only a fraction of points to clusters (the foreground points), while leaving low-density observations (background points) unlabeled. Therefore, assigning the background points to clusters can be done with $k$-NN algorithm.
In our experiment, for the algorithm with determined results, the performance is reported with the best parameter setting while for the algorithm with stochastic results, the experiment is repeated 10 times and the average performance is reported with the best parameter setting.
We simply use the Python-package scikit-learn for DBSCAN and $k$-means and R package for ADP-Cluster and PDF-Cluster.
\subsection{Synthetic Data} \label{subsec::SyntheticData}
In this subsection, we apply the density-based clustering methods mentioned above
on four artificial examples. To be specific, we simulate four two-dimensional toy datasets with different shapes of clusters:
\begin{itemize}
\item
\textbf{noisy circles}: contains a large circle containing a smaller circle with
two-dimensional noise;
\item
\textbf{varied blob}: is generated by isotropic Gaussian blobs with variant variances for clustering;
\item
\textbf{noisy moons}: is made up of two interleaving half circles adding standard deviation of Gaussian noise;
\item
\textbf{aniso-bolb}: is anisotropicly distributed, i.e., the data set is generated by anisotropic Gaussian blobs.
\end{itemize}
In order to see the scalability of these algorithms, we choose the size big enough ($n = 1500$), but not too big to avoid too long running time.
\begin{figure*}[htbp]
\begin{minipage}[t]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Images/aniso.png} \includegraphics[width=\textwidth]{Images/noise_circles.png}
\end{minipage}
\begin{minipage}[t]{0.49\textwidth}
\includegraphics[width=\textwidth]{Images/noise_moons.png} \includegraphics[width=\textwidth]{Images/varied.png}
\end{minipage}
\centering
\caption{Synthetic data. The picture at the upper left shows \textit{aniso-blob} data; the picture at the upper right shows \textit{noise moons} data; the picture at the bottom left shows \textit{noise circles} data and the picture at the bottom right shows \textit{varied} data. The clusters are distinguished by different colors.}
\label{fig:synthetic data}
\end{figure*}
\begin{table*}[h]
\setlength{\tabcolsep}{9pt}
\centering
\captionsetup{justification=centering}
\caption{\footnotesize{Average ARI (adjusted rand index) over Four simulated data sets}}
\resizebox{\textwidth}{13mm}{
\begin{tabular}{c|ccccc}
\toprule
Datasets & ADP-Cluster & DBSCAN & $k$-means & PDF-Cluster & Ours
\\
\hline
\hline
\text{aniso-blob} & 0.519693602 &\textbf{1} &0.590986466 &0.992013365 & \textbf{1}
\\
\hline
\text{noisy circles} &0.165814882 &\textbf{1} & 0.156841194 & 0.189604897 &\textbf{1}
\\
\hline
\text{noisy moons} & 0.491529298 &\textbf{1} & 0.50093108 &\textbf{1} &\textbf{1}
\\
\hline
\text{varied blob} & 0.844780818 &0.906919484& 0.824744996 &\textbf{0.939402597} & 0.936238496
\\
\bottomrule
\end{tabular}}
\begin{tablenotes}
\footnotesize
\item{*} The best results are marked in \textbf{bold}.
\end{tablenotes}
\label{tab::1}
\end{table*}
Table \ref{tab::1} reports the ARI of our clustering algorithm and other clustering methods with the best parameter setting over four toy datasets. It can be evidently observed from the Table \ref{tab::1} that our algorithm has the best ARI performances on almost all data sets, further demonstrating the effectiveness of the algorithm. Our algorithm as well as DBSCAN recognizes the correct clusters on three data sets: {\tt aniso-blob}, {\tt noisy circles}, and {\tt noisy moons}.
\subsection{Real Data Analysis} \label{subsec::RealDataAnalysis}
In our experiment, to assess the performance of various clustering methods, we evaluate the ARI among ADP-Cluster, DBSCAN, $k$-means and PDF-Cluster and our best-scored clustering forest on the following real data sets from UCI and Kaggle:
\begin{itemize}
\item
\textbf{Appendicitis}: The appendicitis data collected in the medical field was first put forward in \cite{Weiss1991Computer}.
The whole data represents $7$ medical measures taken over $106$ patients on which the class label represents if the patient has appendicitis (class label $1$) or not (class label $0$).
\item
\textbf{Customers}: The data set refers to clients of a wholesale distributor including the annual spending in monetary units on diverse product categories. This database available on UCI contains $440$ observations of dimension $8$ representing attributes such as {\tt fresh}, {\tt milk}, {\tt grocery}, {\tt frozen}, etc.
\item
\textbf{Flee-beetles}: For three species of flea-beetles: concinna, heptapotamica, and heikertingeri, the whole data set was collected with six measurements: {\tt tars1}, {\tt tars2}, {\tt head}, {\tt aede1}, {\tt aede2} and {\tt aede3}. The whole data set consists of $74$ samples.
\item
\textbf{Iris}: Regarded as one of the best known database shown in the pattern recognition literature,
Iris contains $3$ classes of $50$ instance each, where each class refers to a type of iris plant.
The learning goal is to group the iris data with four features: {\tt sepal length}, {\tt sepal width}, {\tt petal length}, and {\tt petal width} into the true classes.
\item
\textbf{Oliveoil}: Oliveoil comprises $572$ observations from oil analysis using measurements of different specimen for olive oil produced in various regions in Italy which can be further divided into three macro-areas: Centre-North, South, Sardinia.
This $8$-dimensional input data represent attributes such as {\tt palmitic}, {\tt palmitoleic}, {\tt stearic}, {\tt oleic}, {\tt linoleic}, {\tt linolenic}, {\tt arachidic}, {\tt eicosenoic}.
The learning task is to reconstruct the macro-area membership.
\item
\textbf{Wifi-localization}: The database comprising $2000$ observations was collected in indoor space by observing signal strengths of seven WiFi signals visible on a smartphone. The experiment was performed to explore how wifi signal strengths can be used to determine one of the indoor locations.
\item
\textbf{Wine}: This data set including $178$ observations are the results of a chemical analysis of wines grown in the same region in Italy but derived from three different cultivars. The analysis determined the quantities of $13$ constituents found in each of the three types of wines.
\end{itemize}
\begin{table*}[h]
\setlength{\tabcolsep}{9pt}
\centering
\captionsetup{justification=centering}
\caption{\footnotesize{Average ARI (adjusted rand index) over UCI Data Sets}}
\resizebox{\textwidth}{13mm}{
\begin{tabular}{c|ccccc}
\toprule
Datasets & ADP-Cluster & DBSCAN & $k$-means & PDF-Cluster & Ours
\\
\hline
\hline
\text{appendicitis} & \textbf{0.525568993} & 0.419405321 &0.318301412 &0.468845316 & 0.518105812
\\
\hline
\text{customers} & 0.260017768 &0.196905491& 0.400496777 & 0.012502467 &\textbf{0.625997132}\\
\hline
\text{flea} &0.879214629 & \textbf{1} & 0.957543737 &\textbf{1} & 0.973650744\\
\hline
\text{iris} & 0.568115942 & 0.61410887 & 0.716342113 &0.568115942 & \textbf{0.778123403}
\\
\hline
\text{oliveoil} & 0.572903083 & \textbf{1} & 0.627921018 & 0.865827662 & \textbf{1}
\\
\hline
\text{wifi localization} & \textbf{0.914080542} &0.869141447 & 0.314316461 &0.232882926 & 0.909102543
\\
\hline
\text{wine} & 0.817666167 & 0.847096681 & 0.622913 & 0.845786696 & \textbf{0.872752411}
\\
\bottomrule
\end{tabular}}
\begin{tablenotes}
\footnotesize
\item{*} The best results are marked in \textbf{bold}.
\end{tablenotes}
\label{tab::2}
\end{table*}
Table \ref{tab::2} summaries the ARI on the real data sets mentioned above. Careful observations will find that for most of these data sets,
the best-scored forest clustering has significantly larger ARI than other density-based clustering methods. This superiority in cluster accuracy may be attributed to both the density estimation accuracy resulted from general architecture of random forest and the advantage of the density-based clustering method to group the data into arbitrarily shaped clusters.
We mention that interested readers can further
tune the free parameters and
we believe that more accurate results could be obtained.
\section{Proofs} \label{sec::Proofs}
\begin{proof}[Proof of Theorem \ref{theorem::UncertaintyControlForest}]
\textit{(i)}
Let us first prove the inclusion $M_{\rho+\varepsilon}^{-2\sigma} \subset L_{D,\rho}$.
To this end, we fix an $x \in M_{\rho+\varepsilon}^{-2\sigma}$, then we have $x \notin (\mathbb{R}^d \setminus M_{\rho+\varepsilon})^{+2\sigma}$, that is, for all $x' \in \mathbb{R}^d \setminus M_{\rho+\varepsilon}$, we have $\|x-x'\|_2>2\sigma$. In other words, if $x' \in \mathbb{R}^d$ satisfying $\|x-x'\|_2\leq 2\sigma$, then we have $x'\in M_{\rho+\varepsilon}$.
Now we show that for all $x_i\in D$, we have
\begin{align} \label{ChaoChaoIsAPig}
f_{D,Z_{\mathrm{E}}}(x_i)\geq \rho
\quad
\text { or }
\quad
\|x-x_i\|_2> \sigma,
\end{align}
whose proof will be conducted in the following by contradiction.
Suppose that there exists a sample $x_i
\in D$ with $f_{D,Z_{\mathrm{E}}}(x_i)<\rho$ and $\|x-x_i\|_2\leq \sigma$.
If we denote $A_{Z_t}(x_i)$ as the unique cell of the partition $A_{Z_t,\mathrm{P}}$ of the $t$-th tree in the forest where $x_i$ falls,
then the assumption $\mathrm{diam}(A_{Z_t}(x_i)) \leq \delta\leq \sigma$ implies that
for any $x'\in A_{Z_t}(x_i)$, there holds
$$
\|x-x'\|_2
\leq \|x-x_i\|_2+\|x_i-x'\|_2\leq 2\sigma
$$
and consequently we have $x'\in M_{\rho+\varepsilon}$, i.e., $A_{Z_t}(x_i) \subset M_{\rho+\varepsilon}$ for $t=1,\ldots,m$. This together with the normality of $\mathrm{P}$ yields
$$
\mu(A_{Z_t}(x_i)\setminus \{f \geq \rho+\varepsilon\})\leq \mu(M_{\rho+\varepsilon}\setminus \{f \geq \rho+\varepsilon\})=0,
$$
which leads to
\begin{align*}
\mathrm{P}(A_{Z_t}(x_i))
& =\int_{A_{Z_t}(x_i)\cap \{f\geq \rho+\varepsilon\}} f d\mathrm{P}+\int_{A_{Z_t}(x_i)\setminus \{f\geq \rho+\varepsilon\}} f d\mathrm{P}
\\
& =\int_{A_{Z_t}(x_i)\cap \{f \geq \rho+\varepsilon\}} f d\mathrm{P}\geq \mu(A_{Z_t}(x_i))(\rho+\varepsilon).
\end{align*}
Consequently, we have
\begin{align} \label{UncertaintyReductio1}
f_{\mathrm{P},Z_{\mathrm{E}}}(x_i)=\frac{1}{m}\sum^m_{t=1}\frac{\mathrm{P}(A_{Z_t}(x_i))}{\mu(A_{Z_t}(x_i))}\geq \rho+\varepsilon.
\end{align}
By $f_{D,Z_{\mathrm{E}}}(x_i)<\rho$ and $\|f_{D,Z_{\mathrm{E}}}(x_i)-f_{\mathrm{P},Z_{\mathrm{E}}}(x_i)\|_{\infty}\leq \varepsilon$, we find $f_{\mathrm{P},Z_{\mathrm{E}}}(x_i)< \rho+\varepsilon$, which contradicts (\ref{UncertaintyReductio1}). Therefore, for all $x_i\in D$, we have $f_{D,Z_{\mathrm{E}}}(x_i)\geq \rho$ or $\|x-x_i\|_2> \sigma$.
Next, we show that there exist a sample $x_i \in D$ such that $\|x-x_i\|_2\leq \sigma$ by contradiction. If we denote $A_{Z_t}(x)$ as the unique cell of the partition $A_{Z_t,p}$ of the $t$-th tree in the forest where $x$ falls, then for all $x_i$, $i=1,\ldots,n$, we have $\|x-x_i\|_2>\sigma\geq \delta$, and consequently $A_{Z_t,p}\cap D=\emptyset$, $t=1,\ldots,m$. This leads to $f_{D,Z_{\mathrm{E}}}(x)=0$, which contradicts $f_{\mathrm{P},Z_{\mathrm{E}}}(x)\geq \rho+\varepsilon$ with the condition $\|f_{D,Z_{\mathrm{E}}}(x)-f_{\mathrm{P},Z_{\mathrm{E}}}(x)\|_{\infty}\leq \varepsilon$. Therefore, we conclude that there exists a sample $x_i \in D$ satisfying $\|x-x_i\|_2\leq \sigma$. This together with \eqref{ChaoChaoIsAPig} implies
$f_{D,Z_{\mathrm{E}}}(x_i)\geq \rho$, which means
$x \in L_{D,\rho}$. This finishes the proof of
$M_{\rho+\varepsilon}^{-2\sigma} \subset L_{D,\rho}$.
\textit{(ii)} To prove the second inclusion $L_{D,\rho} \subset M_{\rho-\varepsilon}^{+2\sigma}$,
let us fix an $x \in L_{D,\rho}$, then there exists $x'\in D$ satisfying
$\|f_{D,Z_{\mathrm{E}}}(x')\|_2\geq \rho$ and $\|x-x'\|_2\leq \sigma$.
Moreover, since $\|f_{D,Z_{\mathrm{E}}}(x)-f_{\mathrm{P},Z_{\mathrm{E}}}(x)\|_{\infty} \leq \varepsilon$, we have
\begin{align} \label{XiaoYuIsAPigToo}
f_{\mathrm{P},Z_{\mathrm{E}}}(x')\geq \rho-\varepsilon.
\end{align}
Now we are able to prove the inclusion $L_{D,\rho} \subset M_{\rho-\varepsilon}^{+2\sigma}$
by contradiction.
Suppose that $x\notin M_{\rho-\varepsilon}^{+2\sigma}$.
Since $\|x-x'\|_2\leq \sigma$,
then we have $B(x',\sigma)\subset \mathbb{R}^d\setminus M_{\rho-\varepsilon}$.
If $A_{Z_t}(x')$ stands for the unique cell of the partition $A_{Z_t,p}$ of the $t$-th tree in the forest where $x'$ falls, since $\mathrm{diam}(A_{Z_t}(x'))\leq \delta\leq \sigma$, we thus have $A_{Z_t}(x')\subset B(x',\sigma)\subset \mathbb{R}^d\setminus M_{\rho-\varepsilon}$. This together with the normality of $\mathrm{P}$ yields
$$
\mu(A_{Z_t}(x')\setminus \{f < \rho-\varepsilon\})\leq \mu((X \setminus M_{\rho-\varepsilon})\setminus \{f > \rho-\varepsilon\})=0,
$$
which leads to
\begin{align*}
\mathrm{P}(A_{Z_t}(x'))
& = \int_{A_{Z_t(x')}\cap \{h<\rho-\varepsilon\}} h \, d\mathrm{P} +\int_{A_{Z_t(x')} \setminus \{h<\rho-\varepsilon\}}h \, d\mathrm{P}
\\
& = \int_{A_{Z_t(x')}\cap \{h<\rho-\varepsilon\}} h \, d\mathrm{P} < \mu(A_{Z_t(x')})(\rho-\varepsilon).
\end{align*}
Consequently we have
\begin{align} \label{UncertaintyReductio2}
f_{\mathrm{P}, Z_{\mathrm{E}}}(x')=\frac{1}{m}\sum^m_{t=1}\frac{\mathrm{P}(A_{Z_t}(x'))}{\mu(A_{Z_t}(x'))}< \rho-\varepsilon
\end{align}
which contradicts \eqref{XiaoYuIsAPigToo}.
Therefore, we conclude that $x \in M_{\rho_\varepsilon}^{+2\sigma}$. This completes the proof of $L_{D,\rho} \subset M_{\rho-\varepsilon}^{+2\sigma}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thr::BoundProbability}]
The proof can be conducted by applying Theorem \ref{thr::LevelSetBound} directly and hence we need to verify its assumptions.
Let us first prove that if
$\varepsilon^*\leq (\rho^{**}-\rho^*)/9$, $\delta \in (0,\delta_{\textit{thick}}]$, $\varepsilon\in (0,\varepsilon^*]$ and $\psi(\delta)<\tau$, then we have $\tau\leq \tau^*(\varepsilon^*)$.
To this end, we define a set $E$ by
\begin{align*}
E := \{ \varepsilon' \in (0,\rho^{**}-\rho^*] : \tau^*(\varepsilon')\geq \tau \}.
\end{align*}
Obviously, we have $E \neq \emptyset$, since $\varepsilon^* < \infty$.
This implies that there exists an $\varepsilon_0 \in E$ such that $\varepsilon_0 \leq \inf E+\varepsilon\leq \varepsilon^*$. Using the monotonicity of $\tau^*$, we conclude that $\tau \leq \tau^*(\varepsilon_0) \leq \tau^*(\varepsilon^*)$.
Next, we prove that for all $\rho > 0$,
$(L_\rho)_{\rho\geq 0}$ satisfy \eqref{UncertaintyControl}
with probability not less than $1 - e^{-\varsigma}$.
For $t = 1, \dots, m$,
let the events $B_{1,t,\varepsilon}$ and $B_{2,t,\delta}$ be defined by
\begin{align}
B_{1,t,\varepsilon} & := \{ \|f_{\mathrm{D},Z_t} - f_{\mathrm{P},Z_t}\|_{\infty} \leq \varepsilon \},
\label{B1t}
\\
B_{2,t,\delta} & := \{ \forall A\in \mathcal{A}_{Z_t,p}:\mathrm{\mathrm{diam}}(A)\leq \delta \}.
\label{B2t}
\end{align}
According to Proposition 15 and Inequality (19) in \cite{hang2018best},
there hold
\begin{align*}
\mathrm{P}(B_{1,t,\varepsilon}) \geq 1 - e^{-\varsigma} / (2m)
\quad
\text{ and }
\quad
\mathrm{P}(B_{2,t,\delta}) \geq 1 - e^{-\varsigma} / (2m)
\end{align*}
for all $t = 1, \dots, m$. Moreover,
for the forest, we define the events $B_{1,\mathrm{E}}$ and $B_{2,\mathrm{E}}$ by
\begin{align}
B_{1,\mathrm{E},\varepsilon} & := \{ \| f_{\mathrm{D},Z_\mathrm{E}} - f_{\mathrm{P}, Z_\mathrm{E}} \|_{\infty} \leq \varepsilon \},
\label{B1E}
\\
B_{2,\mathrm{E},\delta} & := \{ \forall A \in \mathcal{A}_{Z_t,p} : \mathrm{diam}(A) \leq \delta, \, t=1,\dots,m \}.
\label{B2E}
\end{align}
Since the splitting criteria $Z_1,\dots,Z_m$ are i.i.d.~from $\mathrm{P}_Z$, then we have
\begin{align*}
\mathrm{P}(B_{1,\mathrm{E},\varepsilon}) \geq 1 - e^{-\varsigma} / 2
\quad
\text{ and }
\quad
\mathrm{P}(B_{2,\mathrm{E},\delta}) \geq 1 - e^{-\varsigma} / 2
\end{align*}
and consequently we obtain
\begin{align*}
\mathrm{P}(B_{1,\mathrm{E},\varepsilon} \cap B_{2,\mathrm{E},\delta})
= 1 - \mathrm{P}(B_{1,\mathrm{E},\varepsilon}^c \cup B_{2,\mathrm{E},\delta}^c)
\geq 1 - \mathrm{P}(B_{1,\mathrm{E},\varepsilon}^c)
- \mathrm{P}(B_{2,\mathrm{E},\delta}^c)
\geq 1 - e^{-\varsigma}.
\end{align*}
This proves that for all $\rho > 0$,
$(L_\rho)_{\rho\geq 0}$ satisfy \eqref{UncertaintyControl}
with probability not less than $1 - e^{-\varsigma}$
and hence all the assumptions of Theorem \ref{thr::LevelSetBound} are indeed satisfied.
\end{proof}
To prove Theorem \ref{the::Consistency} concerning with the consistency of our clustering algorithm, we need the following technical lemma.
\begin{lemma} \label{lem::trival}
Let $(a_n)$, $(b_n)$ be strictly positive sequences and $\varsigma_n$ be the solution of equation
\begin{align*}
e^{2\varsigma_n}(\varsigma_n+a_n)=b_n.
\end{align*}
If $\lim_{n\to \infty} b_n=\infty$ and $\lim_{n\to \infty}a_n/b_n=0$, then $\lim_{n\to \infty} \varsigma_n=\infty$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem::trival}]
We prove the lemma by contradiction.
To this end, we assume that $\lim_{n\to \infty}\varsigma_n\neq \infty$.
Then there exists an $M > 0$, and a subsequence of $(\varsigma_n)$ denoted by $(\varsigma_{n_k})$ such that $|\varsigma_{n_k}| < M$ hold for all $k$.
Consequently we obtain
\begin{align*}
b_{n_k}=e^{2\varsigma_{n_k}}(\varsigma_{n_k}+a_{n_k})<e^{2M}(M+a_{n_k})
\end{align*}
for all $k$. This together with the condition
$\lim_{k\to \infty} b_{n_k}=\infty$ implies that $\lim_{k\to \infty} a_{n_k}=\infty$.
Therefore, we have
\begin{align*}
\varliminf_{k\to \infty}\frac{a_{n_k}}{b_{n_k}}
\geq \varliminf_{k\to \infty} \frac{1}{e^{2M}} \cdot \frac{a_{n_k}}{M+a_{n_k}}
= \frac{1}{e^{2M}},
\end{align*}
which contradicts the condition $\lim_{n\to\infty}a_n/b_n=0$
and thus the assertion is proved.
\end{proof}
\begin{proof}[Proof of Theorem \ref{the::Consistency}]
Let the events $B_{1,t,\varepsilon}$,
$B_{2,t,\delta}$,
$B_{1,\mathrm{E},\varepsilon}$, and
$B_{2,\mathrm{E},\delta}$
be defined as in
\eqref{B1t},
\eqref{B2t},
\eqref{B1E}, and
\eqref{B2E} respectively.
According to Inequality (19) in \cite{hang2018best}, we have
\begin{align*}
\mathrm{P}^n ( B_{2,t,\delta_n} )
\geq 1 - (K + 2 r) d \delta_n^{-1} p_n^{- c_T/(4d)}
\end{align*}
and consequently we obtain
\begin{align*}
\lim_{n\to \infty} \mathrm{P}^n (B_{2,t,\delta_n}) = 1.
\end{align*}
Since $m$ is finite and splitting criteria $Z_1,\dots,Z_m$ are i.i.d.~from $\mathrm{P}_Z$, we have
\begin{align*}
\lim_{n\to \infty} \mathrm{P}^n ( B_{2,\mathrm{E},\delta_n} ) = 1.
\end{align*}
Proposition 15 in \cite{hang2018best} shows that
\begin{align*}
\mathrm{P}^n ( B_{1,\mathrm{E},\varepsilon_n} )
\geq 1 - 2 e^{-\varsigma_n}
\end{align*}
where $\varepsilon_n \geq \sqrt{\|f\|_{\infty} \mathcal{E}'_n} + \mathcal{E}'_n/3 + 2/n$ with
\begin{align} \label{eq::E_n}
\mathcal{E}'_n :=
8 n^{-1} \mu(B_r)^{-1} e^{2 \varsigma_n} p_n^{2a}
\bigl( (8 d + 1) \varsigma_n + 23 \log n + 8 a d \log p_n \bigr).
\end{align}
Obviously, there exists certain $q \in (0, 1)$ such that
$\varepsilon'_n := \sqrt{\|f\|_{\infty}\mathcal{E}'_n}+\mathcal{E}'_n/3 \geq (1 - q) \varepsilon_n$.
Next,
with the help of Lemma \ref{lem::trival},
we show that if $\varepsilon'_n \to 0$, then we have $\varsigma_n \to \infty$ with $\varsigma_n$ satisfying \eqref{eq::E_n}.
Clearly, we have
\begin{align*}
\mathcal{E}'_n = 9 \bigl( \sqrt{\|f\|_{\infty} + 4 \varepsilon'_n/3} - \sqrt{\|f\|_{\infty}} \bigr)^2 / 4.
\end{align*}
Plugging this into \eqref{eq::E_n}, we obtain
\begin{align*}
e^{2\varsigma_n}((8d+1)\varsigma_n+23\log n+8ad \log p_n)=
9 n p_n^{- 2a} \mu(B_r)
\bigl( \sqrt{\|f\|_{\infty}+ 4 \varepsilon'_n / 3} - \sqrt{\|f\|_{\infty}} \bigr)^2 / 32.
\end{align*}
Now, by setting
\begin{align*}
a_n & := (23 \log n + 8 a d \log p_n) / (8 d + 1),
\\
b_n& := 9 n p_n^{-2a} \mu(B_r) \bigl( \sqrt{\|f\|_{\infty} +
4 \varepsilon'_n / 3} - \sqrt{\|f\|_{\infty}} \bigr)^2
/ (32(8d + 1)),
\end{align*}
it can be easily verified that
there exist finite constants $c_1$, $c_2$, $c_3$, and $c_4$ such that
\begin{align*}
\lim_{n\to \infty} b_n
= \lim_{n\to \infty} c_1 n p_n^{-2a} {\varepsilon'_n}^2
\geq \lim_{n\to \infty} c_2 n p_n^{-2a} {\varepsilon_n}^2
= \infty
\end{align*}
and
\begin{align*}
\lim_{n\to \infty} a_n / b_n
=\lim_{n\to \infty} c_3 (n^{-1} \log n) p_n^{2a} {\varepsilon'_n}^{-2}
\leq \lim_{n\to \infty} c_4 (n^{-1} \log n) p_n^{2a} {\varepsilon'_n}^{-2}
= 0.
\end{align*}
Then, Lemma \ref{lem::trival} with the above $a_n$ and $b_n$
implies $\varsigma_n \to \infty$
and consequently we have
\begin{align*}
\lim_{n\to \infty} \mathrm{P}^n ( B_{1,t,\varepsilon_n} ) = 1.
\end{align*}
Since $m$ is finite and splitting criteria $Z_1,\dots,Z_m$ are i.i.d.~from $\mathrm{P}_Z$, we have
\begin{align*}
\lim_{n\to \infty} \mathrm{P}^n ( B_{1,\mathrm{E},\varepsilon_n} ) = 1
\end{align*}
which completes the proof of consistency
according to Theorem \ref{thr::LevelSetBound} and Section A.9 in \cite{steinwart2015suppA}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{col::rho_rates}]
For $n\geq 1$,
we define
$$
\varepsilon_n^*:=\varepsilon_n+(\tau_n/\underline{c}_{\textit{sep}})^\kappa.
$$
Since sequences $(\varepsilon_n)$, $(\delta_n)$ and $(\tau_n)$ converge to $0$, we have $\delta_n \in (0,\delta_{\textit{thick}}]$ and
for all sufficiently large $n$, there holds
$$
\varepsilon_n^*\leq (\rho^{**}-\rho^*)/9.
$$
Moreover, the assumed $\tau_n$ and $\sigma_n$ satisfy
\begin{align*}
\lim_{n\to\infty} \tau_n/ (2\sigma_n)^{\gamma}
= \lim_{n\to\infty} 2^{-\gamma}
\bigl( n^{\lambda(1-\alpha)}
(\log n)^{\lambda(\alpha-1)}
( \log\log n / \log \log \log n ) \bigr)^{\gamma/2(\gamma\kappa+\epsilon)}
= \infty
\end{align*}
and therefore
we have
$$
\tau_n > 3c_{\textit{thick}} (2\sigma_n)^\gamma=\psi(2\sigma_n)
$$
for all sufficiently large $n$. Set
\begin{align*}
\varsigma_n : = \log\log n,
\qquad
p_n := \bigl( n/\log n \bigr)^{\frac{2d}{c_T\alpha+4ad}},
\end{align*}
and denote $\mathcal{E}_{\varsigma_n,p}$ as in \eqref{eq::CalE_n}.
Since $\mathcal{E}_{\varsigma_n,p}\to \infty$ and $\mathcal{E}_{\varsigma_n,p}/n\to 0 $, we have
\begin{align*}
\sqrt{\|f\|_{\infty} \mathcal{E}_{\varsigma_n, p}/ n}
+ \mathcal{E}_{\varsigma_n, p} / (3n)
+ 2/n
& \sim \sqrt{\|f\|_{\infty} \mathcal{E}_{\varsigma_n, p}/ n}
\\
&\lesssim \bigl(
n^{-1} (\log n)^3 (n/\log n)^{4 a d \lambda / c_T} \bigr)^{1/2}
\\
&\lesssim \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha} \bigr)^{1/2}.
\end{align*}
Consequently, for all sufficiently large $n$, we have
\begin{align*}
\varepsilon_n
& = \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha} \log\log n \bigr)^{\gamma\kappa/(2(\gamma\kappa+\epsilon))}
\\
& \geq \bigl( n^{\lambda \alpha} (\log n)^{2+\lambda \alpha} \log\log n \bigr)^{1/2}
\\
& > \sqrt{\|f\|_{\infty} \mathcal{E}_{\varsigma_n, p}/ n}
+ \mathcal{E}_{\varsigma_n, p} / (3n)
+ 2/n
\end{align*}
and therefore condition \eqref{BoundVarepsilon} on $\varepsilon_n$ is satisfied.
Moreover, there holds
\begin{align*}
p_n^{-1} \bigl( 2 m d (K + 2 r) e^{\varsigma_n} / \delta_n \bigr)^{4 d / c_T}
& \lesssim \bigl(
n^{\lambda/2} (\log n)^{-\lambda} (\log \log n)^{-1/2} \bigr)^{4 d / c_T}
(\log n / n)^{2 d \lambda / c_T}
\\
& = (\log\log n)^{- 2d / c_T}
\end{align*}
and consequently we have
\begin{align*}
\lim_{n \to \infty}
p_n^{-1} \bigl( 2 m d (K+2r) e^{\varsigma_n} / \delta_n \bigr)^{4d/c_T} = 0.
\end{align*}
In other words, for all sufficiently large $n$, there holds
\begin{align*}
p_n > \bigl( 2 m d (K+2r) e^{\varsigma_n} / \delta_n \bigr)^{4d/c_T}
\end{align*}
and therefore condition \eqref{BoundSplits} on $p_n$ is satisfied.
Now, by applying Theorem \ref{the::rho_rates}, there exist an $n_0 \geq 1$ and a constant $\overline{c}$ such that the right-hand side of inequalities \eqref{RhoExessError} holds for $n \geq n_0$. Moreover, if $\kappa$ is exact, \eqref{equ::tau_rho} holds for all $n\geq n_0$.
\end{proof}
\begin{proof}[Proof of Corollary \ref{col::sets_rate}]
Similar as the proof of Theorem \ref{the::rho_rates}, we prove that for all sufficiently large $n$, there holds
\begin{align*}
\tau_n & \geq \psi(2\sigma_n),
\\
\varepsilon_n & > \sqrt{\|f\|_{\infty} \mathcal{E}_{\varsigma_n, p}/ n}
+ \mathcal{E}_{\varsigma_n, p} / (3n)
+ 2/n,
\\
p_n & > \bigl( 2md(K+2r)e^{\varsigma_n} / \delta_n \bigr)^{4d/c_T},
\end{align*}
with $\mathcal{E}_{\varsigma_n,p}$ as in \eqref{eq::CalE_n}, and thus
the conditions in Theorem \ref{the::sets_rate} are all satisfied.
Then, for such $n$, by applying Theorem \ref{the::sets_rate}, we obtain
\begin{align*}
\mathrm{P}^n \biggl( D : \sum^2_{i=1}
\mu(B_i(D)\triangle A_i^*)
\leq \bigl(
7 c_{\textit{flat}} \varepsilon_n
+ c_{\textit{flat}} (\tau_n/\underline{c}_{\textit{sep}})^\kappa
\bigr)^{\vartheta}
+ 6 c_{\textit{bound}}\delta_n^{\alpha_0}
\biggr)
\geq 1 - 1/\log n.
\end{align*}
Elementary calculations show that with the assumed $\varepsilon_n$,$\tau_n$, and $\delta_n$, there exists a constant $c > 0$ such that
\begin{align*}
\mathrm{P}^n \biggl( D : \sum^2_{i=1} \mu(B_i(D) \triangle A_i^*)
\leq c \bigl( n^{- \lambda \alpha} (\log n)^{2+\lambda \alpha} \log \log n \bigr)^{\frac{\varrho\vartheta}{2(\varrho+\vartheta)}} \biggr)
\geq 1 - 1 / \log n.
\end{align*}
Obviously, we have
\begin{align*}
\sum^2_{i=1} \mu(B_i(D) \triangle A_i^*)
\leq 2\mu(\mathcal{X})<\infty.
\end{align*}
Therefore, we can choose a constant $c$ large enough such that the desired inequality holds for all $n \geq 1$.
\end{proof}
\section{Conclusion} \label{sec::Conclusion}
In this paper, we present an algorithm called \emph{best-scored clustering forest} to efficiently solve the single-level density-based clustering problem. From the theoretical perspective, our main results comprise statements and complete analysis of statistical properties such as consistency and learning rates. The convergence analysis is conducted within the framework established in \cite{steinwart2015adaptive}. With the help of best-scored random forest density estimator proposed by \cite{hang2018best}, we show that consistency
of our proposed clustering algorithm
can be established
with properly chosen hyperparameters
of the density estimators and partition diameters. Moreover, we obtain fast rates of convergence for estimating the clusters under certain mild conditions on the underlying density functions and target clusters.
Last but not least, the excellence of \emph{best-scored clustering forest} was demonstrated by various numerical experiments.
On the one hand,
the new approach provides better average adjusted rand index (ARI) than other state-of-the-art methods such as ADP-Cluster, DBSCAN, $k$-means and PDF-Cluster on synthetic data, while providing average ARI that are at least comparable on several benchmark real data sets.
On the other hand,
due to the intrinsic advantage of random forest, it is to be expected that this strategy enjoys satisfactory computational efficiency by taking utmost advantage of the parallel computing.
|
\section{Introduction}
\ac{am} is increasingly applied in different fields of research like fake-news detection \citep{cabrio_five_2018} and political argumentation and network analysis\footnote{See for example the MARDY project (\url{https://www.socium.uni-bremen.de/projekte/?proj=570&print=1}, last accessed: 2019-04-15, 09:50UTC+2).}.\\
One crucial part of the \ac{am} pipeline is to segment written text into argumentative and non-argumentative units.
Recent research in the area of unit segmentation \citep{eger_neural_2017,ajjour_unit_2017} has lead to promising results with F1-scores of up to $0.90$ for in-domain segmentation \citep{eger_neural_2017}.
Nevertheless, there is still a need for more robust approaches.\\
Given the recent progress of attention-based models in \ac{nmt} \citep{bahdanau_neural_2014,vaswani_attention_2017}, this paper evaluates the effectiveness of attention for the task of argumentative unit segmentation.
The idea of the attention layers added to the recurrent network is to enable the model to prioritize those parts of the input sequence that are important for the current prediction \citep{bahdanau_neural_2014}.
This can be achieved by learning additional parameters during the training of the model.
With the additional information gained, the model learns a better internal representation which improves performance.\\
Additionally, we evaluate the impact of contextualized distributed term representations (also referred to as word embeddings hereinafter) on all our models.
The goal of word embeddings is to represent a word as a high-dimensional vector that encodes its approximate meaning.
This vector will be generated by a model trained on a language modeling task, like next-word prediction \citep{mikolov_efficient_2013}, for a given text corpus.
The approximation is based on the word's surrounding context in the train set and with that pre-defined by the chosen corpus.
Words with a similar semantic meaning should then also have similar vector representations, as measured by their distance in the vector space \citep{heuer_semantic_2015}.
Different methods to pre-compute the embeddings include word2vec \citep{mikolov_efficient_2013}, FastText \citep{bojanowski_enriching_2016} and GloVe \citep{pennington_glove:_2014}.
To make use of the capabilities of pre-trained \acp{lm}, such as BERT \citep{devlin_bert:_2018} or Flair \citep{akbik_contextual_2018}, we evaluate how well their semantic representations perform, by using contextualized word embeddings.
Those are, in contrast to previously mentioned methods, specific to the context of the word in the input sequence.
One major benefit is the fact that the time-consuming feature engineering could become obsolete since the features are implicitly encoded in the word embeddings.
Furthermore, a better semantic representation of the input could lead to better generalization capabilities of the model and, therefore, to better cross-domain performance.\\
This paper answers the following research questions, which will help to assess the importance of the attention layers and contextualized word embeddings for the argument unit segmentation task:
\begin{itemize}
\item \textbf{RQ1:} To what extent can additional attention layers help the model focus on the, for the task of unit segmentation relevant, sequence parts and how much do they influence the predictions?
\item \textbf{RQ2:} What is the impact of contextualized distributed term representations like BERT \citep{devlin_bert:_2018} and Flair \citep{akbik_contextual_2018} on the task of unit segmentation and do they improve upon pre-defined representations like GloVe?
\end{itemize}
The contributions of this paper are as follows: first, we present and evaluate new attention based architectures for the task of argumentative text segmentation.
Second, we review the effectiveness of recently proposed contextualized word embedding approaches in regard to \ac{am}.
We will continue by presenting the previous work on this specific task, followed by a description of the data set, the different architectures used and the generation of the word embeddings. Afterward, we will report the results, followed by a discussion and the limitations.
We will finish with a conclusion and an outlook on possible future work.
\section{Related Work}
\label{sec:related-work}
Attention mechanisms have long been utilized in deep neural networks.
Some of its roots are in the salient region detection for the processing of images \citep{itti_model_1998}, which takes inspiration from human perception.
The main idea is to focus the attention of the underlying network on points-of-interest in the input that are often surrounded by irrelevant parts \citep{mnih_recurrent_2014}.
This allows the model to put more weight on the important chunks.
While earlier salient detectors were task-specific, newer approaches (e.g. \citealp{mnih_recurrent_2014}) can be adapted to different tasks, like image description generation \citep{xu_show_2015}, and allow for the parameters of the attention to be tuned during the training.
These additional tasks include sequence processing and the application of such networks to different areas of \ac{nlp}.
One of the first use-cases for attention mechanisms in the field of \ac{nlp} was machine translation.
\citet{bahdanau_neural_2014} utilized the attention to improve their \ac{nmt} model.
A few years later, \citet{vaswani_attention_2017} achieved new \ac{sota} results by presenting an encoder-decoder architecture that is based on the attention mechanism, only adding a position-wise feed-forward network and normalizations in between.
\citet{devlin_bert:_2018} picked up on the encoder part of this architecture to pre-train a bidirectional \ac{lm}.
After fine-tuning, they achieved, again, a new \ac{sota} performance on different downstream \ac{nlp} tasks like Part-of-speech tagging and Questions-Answering.\\
A possible way of posing the unit segmentation as \ac{nlp} task is a token-based sequence labeling \citep{stab_argumentative_2017}.
While \citet{tobias_argument_2018} used rather simple, non-recurrent classifiers to approach this problem, others mostly applied recurrent networks to the task of unit boundary prediction.
For example, \citet{eger_neural_2017} reported different \ac{lstm} \citep{hochreiter_long_1997} architectures.
Further, \citet{ajjour_unit_2017} proposed a setup with three \acp{blstm} \citep{schuster_bidirectional_1997} in total as their best solution.
While the first two of them are fully connected and work on word embeddings and task-specific features respectively, the intention for the third is to take the output of the first two as input and learn to correct their errors.
Even though the third \ac{blstm} did not improve on the F1-score metric, it did succeed in resolving some of the wrong consecutive token predictions, without worsening the final results.\\
To the best of the authors' knowledge, the attention mechanism has not been widely utilized so far for the task of argumentative unit segmentation.
\citet{stab_cross-topic_2018} integrated the attention mechanism directly into their \ac{blstm} by calculating it at each time step $t$ to evaluate the importance of the current hidden state $h_t$.
To do that, they employed additive attention.
A similar approach has been applied by \citet{morio_end--end_2018} for a three-label classification task (claim, premise or non-argumentative).\\
In contrast to that, the approach presented in this paper uses attention as a separate layer that encodes all sequences before they are fed into a \ac{blstm}.
This might enable the recurrent part of the network to learn from better representations that are specific to the task it is trained on.
The aim is further to evaluate the possible applications of attention layers for the task of sequence segmentation and token classification.
A recurrent architecture \citep{ajjour_unit_2017} is compared to multiple modified versions that utilize the attention mechanism.\\
In order to derive a representation of the input text that better resembles the context of the input for a specific task, several approaches have been presented.
\citet{akbik_contextual_2018}, for example, pre-train a character-level \ac{blstm} to predict the next character for a given text corpus.
The pre-trained model is able to derive contextualized word embeddings by additionally utilizing the input sequence for a specific task.
This allows it to encode the previous as well as the following words of the given input sequence into the word itself.
In comparison to that, the pre-trained BERT-\ac{lm} utilizes stacked attention layers \citep{vaswani_attention_2017}.
By feeding a sequence into it and extracting the output of the last sublayer for each token, the idea is to implicitly use the attention mechanism to derive a better representation for every token.
As is the case for the \ac{lm} from \citet{akbik_contextual_2018}, the BERT embeddings are contextualized by the whole input sequence of the specific task.\\
This paper will compare the two contextualized approaches described above with the pre-defined GloVe \citep{pennington_glove:_2014} embeddings in the light of their usefulness for \ac{am}.
The goal is to encode the features necessary to detect arguments by utilizing the context of a sentence.
\section{Methodology}
This paper evaluates different machine learning architectures with added attention layers for the task of \ac{am}, and more specifically unit segmentation. The problem is framed as a multi-class token labeling task, in which each token is assigned one of three labels.
A (B) label denotes that the token is at the beginning of an argumentative unit, an (I) label that it lies inside a unit and an (O) label that the token is not part of a unit.
This framework has been applied previously for the same task \citep{stab_argumentative_2017,eger_neural_2017,ajjour_unit_2017}.\\
The architectures proposed in this section build on \citet{ajjour_unit_2017}, omitting the second \ac{blstm}, which was used to process features other than word embeddings (see section \ref{sec:features}).
They are further being modified by adding attention layers at different positions.
The goal is to reuse existing approaches and possibly enhance their ability to model long-range dependencies.
Additionally, a simpler architecture, consisting of a single \ac{blstm} paired with an attention layer, is built and evaluated with the aim of decreased complexity.\\
In order to answer the second research question, this paper reports results in combination with improved input embeddings, in order to evaluate their effectiveness and impact on the \ac{am} downstream task.\\
All models are compared to the modified re-implementation of the architecture, which is defined as the baseline architecture.
\begin{figure*}[t!]
\label{fig:architectures}
\centering
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics{figures/baseline-architecture}
\caption{\label{fig:architectures:baseline}}
\end{subfigure}
~
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics{figures/baseline2-architecture}
\caption{\label{fig:architectures:baseline2}}
\end{subfigure}
~
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics{figures/input-bilstm-architecture}
\caption{\label{fig:architectures:input-bilstm}}
\end{subfigure}
\caption{
\textit{(\subref{fig:architectures:baseline})} The original baseline architecture as reported by \citet{ajjour_unit_2017}.
\textit{(\subref{fig:architectures:baseline2})} The modified baseline architecture without the second input Bi-LSTM. The bold arrows show the positions at which the additional attention layers are added to build the \acs{baseline-i} and \acs{baseline-e} architectures.
\textit{(\subref{fig:architectures:input-bilstm})} The \acs{bilstm} architecture incorporates only one \ac{blstm}. The bold arrow shows the position at which the additional attention layer is added to build the \acs{bilstm-i} architecture.}
\end{figure*}
\subsection{Features}
\label{sec:features}
For each token, a set of three different embeddings is generated and compared regarding their capability as standalone input features.
The resulting weighted F1-score is then used as a proxy for measuring the usefulness of the generated text-representation in light of this specific downstream task.\\
In combination with the re-implemented architecture, the word vectorization approach GloVe \citep{pennington_glove:_2014}, trained on 6 billion tokens\footnote{\url{https://nlp.stanford.edu/projects/glove/}, last accessed: 2019-04-15, 09:50UTC+2.}, serves as the baseline.\\
As a first approach to enhance the performance, the GloVe embeddings are stacked with the character-based Flair embeddings \citep{akbik_contextual_2018}, which are generated by a \ac{blstm} model.
\citet{akbik_contextual_2018} argue that the resulting embeddings are contextualized, since the \ac{lm} was trained to predict the most probable next character and therefore to encode the context of the whole sequence.\\
Similar to that, we also compare contextualized BERT-embeddings as standalone features \citep{devlin_bert:_2018}.
An increased performance is expected because of the pre-training procedure of the \ac{lm}.
The BERT-\ac{lm} was trained to predict a (randomly masked) word by utilizing the context of its appearance, as well as on next sentence prediction.
Due to its \ac{sota} performance for both, token-level and sentence-level tasks, the authors of this paper argue that the derived representations are well suited for the task of unit segmentation.
Also, the representation fits the needs of the inter-token and sentence dependencies of the task.
It is expected that this enables the model to better grasp the notion or pattern of an argument.
Both contextualized embeddings are generated using the Flair library \citep{zalando_research_very_2018}.\\
Features specifically engineered for this task are not included in the input, following the argumentation of \citet{eger_neural_2017} that they will probably not be generalizable to different data sets.
\subsection{Data}
In order to evaluate the different architectures, the ``Argument annotated Essays (version 2)'' corpus (also referred to as Persuasive Essays corpus) is used \citep{stab_parsing_2017}.
It was utilized for the same task in previous literature \citep{ajjour_unit_2017,eger_neural_2017}.\\
The corpus, compiled for parsing argumentative structures in written text, consists of a random sample of 402 student essays.
The annotation scheme includes the argumentative units and the relations between them, as well as the major claim and stance of the author towards a specific topic.
The texts were annotated by non-professionals, labeling the boundary of each argumentative unit alongside the unit type.
A type can either be major-claim, claim or premise.
For the unit segmentation task, the corpus is labeled by treating major claims, claims, and premises as argumentative units\footnote{All data pre-processing scripts are available in our code repository: \url{https://gitlab.informatik.uni-bremen.de/covis1819/worth-the-attention}.}.
For comparability reasons in the evaluation process, the models are trained and tested with the train-test-split defined by \citet{stab_parsing_2017}.
\subsection{Models}
\label{sec:models}
In order to evaluate the attention mechanisms, different architectures based on previous \ac{am} literature are implemented.
The attention layer is added at different positions in the network.\\
All models were implemented using Python and the Keras framework with a TensorFlow backend.
For the self-attention and multi-head attention layers, an existing implementation is used \citep{hg_attention_2018,hg_wrapper_2018}.
The difference between the two is that the multi-head attention divides the input into multiple chunks and each head therefore works on a different vector subspace \citep{vaswani_attention_2017}, while the self-attention works on the whole input sequence.
This is supposed to allow the head to focus on specific features of the input.
In this case, the self-attention layers use additive attention, while the multi-head attention layers use scaled dot-product attention, with the latter following the implementation of \citet{vaswani_attention_2017}.
\paragraph*{Baseline re-implementation}
The baseline model from \citet{ajjour_unit_2017} uses a total of three \acp{blstm} (two of them fully connected) to assign labels to tokens (see Figure~\ref{fig:architectures:baseline}).
The re-implementation does not include the two fully connected \acp{blstm} but instead uses only a single one that works on the word embeddings (see Figure~\ref{fig:architectures:baseline2}).
Due to the fact that the second \ac{blstm} in the first layer is only used to encode the non-semantic features like Part-of-speech tags and discourse marker labels, it is omitted in the re-implementation.
Hereafter, we will refer to this model as \acs{baseline}.
Also, the batch size was increased from 8 to 64, compared to the original implementation, as a trade-off between convergence time and the model's generalization performance \citep{keskar_large-batch_2016}.
Nevertheless, this model achieves comparable scores to the ones presented in the original paper.
The slightly lower performance can probably be attributed to implementation details.
\paragraph*{\acs{baseline-i} and \acs{baseline-e}}
For both variations, the baseline architecture was used as a basis, as can be seen in Figure~\ref{fig:architectures:baseline2}.
Multi-head-attention layers are added at different positions in the network.
The number of attention heads depends on the dimension of the embedding vectors.
For the GloVe (300 features) and the BERT (3072 features) embeddings, six heads are used, while the Flair (4196 features) embeddings require four heads.
Both numbers were the largest divisor for the respective input vector size that worked inside the computational boundaries available.
In the first model, an attention layer was added before the first \ac{blstm} in an attempt to apply a relevance score directly to the tokens, in order to better capture dependencies of the input sequence.
This model will be referred to as \acs{baseline-i}.
The second variation adds the attention layer after the first and before the second \ac{blstm}, which will be called \acs{baseline-e}.
According to \citet{ajjour_unit_2017}, the latter \ac{blstm} is used to correct the errors of the first one.
The attention layer should be able to support the model in the error correction process.
In contrast to the first approach, this does not change the input data, but only works on the output of the first \ac{blstm}.
\paragraph*{\acs{bilstm} and \acs{bilstm-i}}
To decrease the complexity of the architecture, two additional models with a single \ac{blstm} are trained.
The first variant has no attention layer, while the second one utilized the same input attention described above (see Figure~\ref{fig:architectures:input-bilstm}).
They will be refered to as \acs{bilstm} and \acs{bilstm-i} respectively.
Both architectures use a self-attention mechanism instead of the above-mentioned multi-head-attention, due to better results in preliminary tests.
\section{Results}
\label{sec:results}
\newcommand\innerspacing{\rule[-2.0ex]{0pt}{0pt}}
\begin{table}[t!]
\centering
\begin{tabular}{lccc}
\specialrule{.1em}{.05em}{.05em}
Model \rule{0pt}{3.0ex} & GloVe & BERT & Flair \\ \hline
\acs{baseline} \innerspacing & 0.86 & 0.83 & \textbf{0.87} \\
\acs{baseline-i} & \textbf{0.85} & 0.68 & 0.67 \\
\acs{baseline-e} \innerspacing & 0.67 & \textbf{0.68} & 0.67 \\
\acs{bilstm} & \textbf{0.86} & 0.86 & 0.86 \\
\acs{bilstm-i} \innerspacing & \textbf{0.84} & 0.83 & 0.81 \\
\specialrule{.1em}{.05em}{.05em}
\end{tabular}
\caption{The weighted F1-scores for the \acs{baseline} and all four variations. Results are shown per variation and embedding. Each row shows the performance of one architecture with different word embeddings as input vector. The highest score for each architecture is marked in bold.}
\label{tab:results}
\end{table}
We evaluate the performance of all architectures on the Persuasive Essays data set detailed above.
The models are re-initialized after every evaluation and do not share any weights.
This allows us to answer the first research question of whether additional attention layers have a positive impact on the prediction quality.\\
To answer the second research question, we re-run each training, replacing the GloVe with BERT and Flair embeddings.
Both contextualized embedding methods are tested separately.
We contextualize the tokens on the sentence level since the BERT model \citep{google_research_tensorflow_2018} only allows for a maximum input length of 512 characters.
This makes document-level or paragraph-level embeddings impractical for the data set.\\
As a performance measure, we report the weighted F1-score instead of the macro F1-score, since it takes the imbalance of the samples per label into account.\\
For our re-implementation of the baseline, we are able to approximately reproduce the results reported by \citet{ajjour_unit_2017}.
Additionally, we can verify that there is no major change in the performance when adding a second \ac{blstm} to the network (compare results for \acs{bilstm} and \acs{baseline} in Table \ref{tab:results}).
\subsection{Attention Layers}
The results of the token classification task are presented in Table \ref{tab:results}.
Generally speaking, the added attention encodings do not improve upon the original architecture's performance, no matter at which position they are added.
Architectures with an input attention encoding, namely \acs{baseline-i} and \acs{bilstm-i}, do achieve similar performances compared to their respective baseline.
But the F1-score performance is in strong contrast to the generalization error, which is in most cases lower for the \acs{baseline} model.\\
The \acs{baseline-e} architecture on the other hand, which is supposed to help the second \ac{blstm} in the network to correct the errors made by the first one, performs worse across all tests.
For the Flair embeddings, this results in a 0.20 points performance drop in the F1-score measure.
\begin{figure*}[t!]
\centering
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[scale=0.35]{figures/baseline-loss-curve--glove300}
\caption{\label{fig:loss:glove}}
\end{subfigure}
~
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[scale=0.35]{figures/baseline-loss-curve--bert}
\caption{\label{fig:loss:bert}}
\end{subfigure}
~
\begin{subfigure}[t]{0.32\textwidth}
\centering
\includegraphics[scale=0.35]{figures/baseline-loss-curve--flair}
\caption{\label{fig:loss:flair}}
\end{subfigure}
\caption{The loss curves of the \acs{baseline} architecture using different input embeddings. \textit{(\subref{fig:loss:glove})} shows the training process of the model using the GloVe embeddings, while the model in \textit{(\subref{fig:loss:bert})} used the BERT embeddings and \textit{(\subref{fig:loss:flair})} the Flair embeddings. The bottom orange line shows the training loss, the top green line the validation loss.}
\label{fig:loss}
\end{figure*}
\subsection{Contextualized Word Embeddings}
The results for the enhanced word embedding evaluations are reported in Table \ref{tab:results}.
In some cases, the models utilizing the word embeddings generated by the BERT-\ac{lm} achieve a lower performance score than the other embeddings.
This drop is most noticeable for the \acs{baseline-i} model, while the performance for the \acs{bilstm-i} decreases only slightly.
The \acs{baseline-e} model is able to achieve results that outperform both, GloVe and Flair embeddings.\\
Compared to the GloVe vectors, the models trained on the Flair embeddings mostly lose in F1-score performance as well.
For example, the \acs{baseline-i} model drops by 0.18 points.
On the other hand, the \acs{baseline} model is able to slightly improve upon the GloVe score using the Flair embeddings, achieving a final score of 0.87, which also marks the best overall score in our testings.\\
An interesting observation is the fact that the enhanced embeddings seem to increase the generalization error (compare Figure \ref{fig:loss}).
The \acs{baseline} model trained on the GloVe embeddings for example, shows a difference in the final validation and training loss of around 0.17 and increases for the BERT and Flair embeddings to roughly 0.60 and 0.48 points, respectively.
\section{Discussion}
Given the experimental results, we discuss the resulting implications for our two research questions and conclude this section by presenting some limitations.
\subsection{Attention Layers}
Our results suggest that the attention encoding does not increase the performance of the model, as we hypothesized above.
This is true for both, the input and the error encoding.
A potential explanation is the fact that we use the attention mechanism as an additional layer to encode the input.
Other approaches, like \citet{morio_end--end_2018} or \citet{stab_cross-topic_2018}, incorporate it into the \ac{blstm} architecture and calculate the weight of the hidden states at every time step.\\
While the performance does not decrease meaningfully for the \acs{baseline-i} and \acs{bilstm-i} models (using the GloVe embeddings as features), it does for the error encoding \acs{baseline-e} model.
This drop might be explained by the vector space the attention mechanism is working on.
Due to its small size of only four features, it is unlikely that the resulting vector has a meaningful encoding.\\
A deeper inspection of the output values from the different layers in the network and how they influence the overall classification task might give more insight into the cause of the problem.
\subsection{Contextualized Word Embeddings}
For most of the tests we conduct, the contextualized embedding approaches do not improve upon the GloVe embeddings.
This is especially true for the architectures that include an attention layer, which does not seem to be able to handle the encoding of high dimensional vectors very well.
The results further suggest that the amount of neurons in the \acp{blstm} is not an issue in this case, since the \acs{baseline} model achieves comparable results across all three embeddings.\\
A potential way to improve the results of the enhanced embeddings is to contextualize them on the paragraph level.
While we contextualize them on a sentence level, the dependencies between arguments might span over multiple sentences, sometimes even a paragraph, as described by \citet{stab_parsing_2017} for the Persuasive Essays data set.
Following this reasoning, one might think that a document level contextualization makes sense and adds even more information to the embedding.
For the task of AM, however, we argue against that for two reasons.
First, argumentative units usually do not span over the whole document and it might include additional counter-arguments \citep{stab_parsing_2017}.
The contextualization would most likely cause a lot of noise and make the vector less useful.
Also, depending on the size of the document, the size of the vector might be too small to hold the contextual information of the full document.
Second, the model trained on such embeddings would probably not generalize very well.
An argumentative document can be written in different formats with different purposes, like an essay, a speech or a newspaper article.
Contextualizing the embeddings on the document level might then also encode the structure of the text and decrease the cross-domain applicability of the model.
\subsection{Limitations}
\label{sec:limitations}
The results we report and analyze above are the networks' performance as validated on the data splits provided by \citet{stab_parsing_2017}.
Due to time and resource restrictions, we evaluate the results after a single training run and perform neither an averaging over multiple runs nor any cross-validation.
Both could lead to more reliable results.
As another consequence of the above-mentioned restrictions, we are also not able to test the model's generalization capabilities on different data sets.\\
For the learning rate, we perform only a basic Bayesian hyperparameter optimization \citep{snoek_practical_2012} with four iterations per model.
These limitations are especially important for the variations of the \acs{baseline} architecture, since the performed changes to the architecture, even though rather small, entail the need for independently tuned hyperparameters.\\
Furthermore, an additional evaluation of the different contextualization levels for the embeddings could provide a clearer picture of how much the results actually improve, compared to non-contextualized methods.
\section{Conclusion}
Recent improvements in utilizing contextual information for sequence processing had a big impact on the area of \ac{nlp}, namely advances of attention architectures and contextualized word embeddings.
For example, the Transformer architecture \citep{vaswani_attention_2017} employs attention to achieve \ac{sota} scores on different \ac{nlp} tasks.
Further, the Flair model \citep{akbik_contextual_2018} incorporates character-wise context to generate enhanced word representations.\\
In this paper, we report on the usefulness of these two approaches for the task of \ac{am}.
First, we are able to show that an attention layer as additional encoding of the input does not improve upon the current \ac{sota} approach of a \ac{blstm}.
Additionally, the attention mechanism seems to fail for a low-dimensional vector space.
Second, we present the impact of contextualized word embeddings for \ac{am}.
Although the Flair embeddings slightly improve upon the performance of the GloVe embeddings for the \acs{baseline} architecture, we can not confirm any advantage over non-contextualized embeddings.
\subsection{Future Work}
A first extension of this work could be a proper hyperparameter optimization for the attention-based models.
Second, we plan to explore an attempt to fine-tune solely attention based pre-trained models like BERT \citep{devlin_bert:_2018} to domain-specific data.
Recent research by \citet{howard_universal_2018} in transfer-learning for \ac{nlp} has shown great improvement for several \ac{nlp}-downstream tasks, while reducing the needed amount of labeled training data.\\
Third, we contextualize the embeddings on the sentence level only.
According to \citet{stab_parsing_2017}, arguments can sometimes span over multiple sentences.
Therefore, the contextualization of the embeddings could be extended to a paragraph level, in order to make use of possible inter-dependencies within it.
Additionally, a fine-tuning approach of the underlying \ac{lm}s to the \ac{am} task could further enhance the embeddings.
\section*{Acknowledgments}
TODO
\section{Appendices}
\label{sec:appendix}
TODO
|
\section{Introduction}
\subsection{Motivation}
Entropy production is a central concept to quantify the irreversibility of nonequilibrium systems and dissipative
structures and as such, it is of central importance to understand many situations encountered in biology, chemistry,
physics and engineering~\cite{KondepudiPrigogineBook2007}. Within the open system paradigm, where a system exchanges
energy and entropy with its environment, entropy production $\Sigma$ is typically expressed as
\begin{equation}
\Sigma = \Delta S_S + \Delta S_\text{env} \ge 0,
\end{equation}
Here, $\Delta S_S$ ($\Delta S_\text{env}$) measures the change in thermodynamic entropy of the system (environment).
Positivity of entropy production and the second law of nonequilibrium thermodynamics become equivalent statements then.
Furthermore, if the environment is composed out of several ideal heat baths $\nu$, the change in thermodynamic entropy
of the environment can be expressed as $\Delta S_\text{env} \equiv \sum_\nu \Delta S_\nu$ with
$\Delta S_\nu = -\beta_\nu Q_\nu$. Here, $Q_\nu$ is the heat flow from bath $\nu$ \emph{into} the system and the
proportionality constant is the inverse temperature $\beta_\nu$ of the bath. Then, entropy production takes on the
familiar form
\begin{equation}\label{eq EP}
\Sigma = \Delta S_S - \sum_\nu \beta_\nu Q_\nu \ge 0.
\end{equation}
In this paper we are interested in small systems $S$, which can be dominated by fluctuations and might show quantum
effects. Also within this scenario the notion of entropy production plays an essential role, e.g., to quantify the
efficiency of molecular motors or quantum heat engines in stochastic and quantum thermodynamics (see, e.g.,
Refs.~\cite{SeifertRPP2012, KosloffEntropy2013, SchallerBook2014} for introductions). Conventionally, positivity of
entropy production is derived from an effective master or Fokker-Planck equation, which describes the dynamics of the
system after tracing out the bath under various assumptions~\cite{SeifertRPP2012, KosloffEntropy2013, SchallerBook2014,
BreuerPetruccioneBook2002}. While this reassures the thermodynamic consistency of the derived master or Fokker-Planck
equation, it is somewhat unsatisfactory as it is not clear how entropy production emerges from the underlying
microscopic (i.e., Hamiltonian) dynamics of the system and the baths.
This paper claims to provide a valid and satisfactory microscopic derivation of entropy production for a large class
of open system scenarios. Indeed, as we will review in the next section, recent progress in nonequilibrium statistical
mechanics has almost provided a satisfactory answer to this question. But one important point was left open: so far
it was impossible to find a definition for a thermodynamic state function, which, first, describes the entropy of the
system-bath composite (the `universe') and, second, whose change is equal to the entropy production. In that
respect previous derivations failed to show that entropy production measures the change in thermodynamic entropy of
the universe.
\subsection{Previous derivations of positivity of entropy production}
\label{sec previous results}
Deriving the laws of thermodynamics from an underlying microscopic (i.e., Hamiltonian) picture is a central theme of
statistical mechanics since its beginning. Here, we briefly review the approach by Esposito, Lindenberg and Van den
Broeck~\cite{EspositoLindenbergVandenBroeckNJP2010} where a microscopic (i.e., Hamiltonian) derivation of the second
law of nonequilibrium thermodynamics for a driven system in contact with multiple heat baths was provided. In case of
a single heat bath, this result was independently derived in Ref.~\cite{TakaraHasegawaDriebePLA2010}, see also
Ref.~\cite{JarzynskiJSP1999} for an early derivation in the classical case where a system acts sequentially with
different baths prepared at different temperatures. Similar work and extensions of this approach can be found in
Refs.~\cite{HasegawaEtAlPLA2010, EspositoVandenBroeckEPL2011, DeffnerJarzynskiPRX2013, ReebWolfNJP2014,
GooldPaternostroModiPRL2015, BarraSciRep2015, StrasbergEtAlPRX2017}. Furthermore, related work showed how to derive
`dissipation inequalities' on a Hamiltonian basis for a system in contact with a single heat
bath~\cite{JarzynskiPRL1997, JarzynskiPRE1997, JarzynskiJSP1999, KawaiParrondoVandenBroeckPRL2007,
VaikuntanathanJarzynskiEPL2009, ParrondoVandenBroeckKawaiNJP2009}. In this approach it is assumed that the system
starts in equilibrium and can relax back to it after the process has finished. While employing similar mathematical
manipulations, this approach is, strictly speaking, less general than the one that will be reviewed here which takes
explicitly the nonequilibrium nature of the initial and final system states into account. Also multiple baths are
commonly not treated there, but see Ref.~\cite{ParrondoVandenBroeckKawaiNJP2009}. For completeness, let us also mention
an alternative approach for cyclic Hamiltonian dynamics (i.e., where the initial and final Hamiltonian are the same)
based on the notion of complete passivity~\cite{UzdinSaarPRX2018}. For the rest of this section we will focus on
the derivation in Ref.~\cite{EspositoLindenbergVandenBroeckNJP2010}. We will also establish notation here and illustrate
the problem with this approach, which we will eventually overcome later on. We remark that we choose a quantum
mechanical notation throughout this paper, but the corresponding classical manipulations are analogous.
Within the standard paradigm of open system theory, we will assume that the dynamics of the universe are modeled by
the Hamiltonian $H_\text{tot}(\lambda_t) = H_S(\lambda_t) + H_{SB} + H_B$. Here, $H_S(\lambda_t)$ is the system
Hamiltonian with $\lambda_t$ some externally specified driving protocol (e.g., a changing electric field). Furthermore,
$H_B$ describes the bath Hamiltonian and $H_{SB}$ the interaction between system and bath. The present approach can be
generalized to the case of multiple heat baths and a driven interaction Hamiltonian, but for ease of presentation
we refrain from doing so in this section. Now, the only assumption we will add is that the initial state of the
universe is given by
\begin{equation}\label{eq initial state}
\rho_\text{tot}(0) = \rho_S(0)\otimes \pi_B,
\end{equation}
where $\rho_S(0)$ is arbitary and $\pi_B \equiv e^{-\beta H_B}/\C Z_B$
with $\C Z_B = \mbox{tr}_B\{e^{-\beta H_B}\}$ denotes the canonical equilibrium state of the bath. This assumption
is essential for the following and conventionally used in open system theory~\cite{BreuerPetruccioneBook2002}.
An extension to correlated initial states is also possible for a single heat bath~\cite{SeifertPRL2016,
StrasbergEspositoPRE2019}, see also Refs.~\cite{JarzynskiJSM2004, CampisiTalknerHaenggiPRL2009, JarzynskiPRX2017} for
related `dissipation inequalities' and Ref.~\cite{UzdinSaarPRX2018} for cyclic Hamiltonian processes. In this paper,
however, we will exclusively focus on initially decorrelated states.
Now, let us denote by $S_\text{vN}(\rho) = -\mbox{tr}\{\rho\ln\rho\}$ the von Neumann entropy of the state $\rho$. Then,
since the global dynamics are entropy-preserving, we find immediately with the help of Eq.~(\ref{eq initial state}) that
\begin{equation}\label{eq marginal entropy change}
\Delta S_\text{vN}[\rho_S(t)] + \Delta S_\text{vN}[\rho_B(t)] = I[\rho_{SB}(t)] \ge 0,
\end{equation}
where $\Delta S_\text{vN}[\rho(t)] = S_\text{vN}[\rho(t)] - S_\text{vN}[\rho(0)]$ denotes the change in von Neumann
entropy and $I[\rho_{SB}(t)] = S_\text{vN}[\rho_S(t)] + S_\text{vN}[\rho_B(t)] - S_\text{vN}[\rho_{SB}(t)] \ge 0$ is the
always positive mutual information. It is tempting to view Eq.~(\ref{eq marginal entropy change}) already as the
entropy production: at least it is always positive and given by a change in a state function, namely the sum of the local
von Neumann entropies. Indeed, if the `bath' itself is microscopically small, it makes sense to identify the heat flux
directly via $-\beta Q = \Delta S_\text{vN}[\rho_B(t)]$~\cite{BeraEtAlNC2017}. Then,
Eq.~(\ref{eq marginal entropy change}) formally takes on the conventional form of entropy production, compare with
Eq.~(\ref{eq EP}) in case of a single heat bath. Unfortunately, for a mesoscopic or macrocopic heat bath this does not
provide a satisfactory resolution as the mutual information can always be bounded by
$I(\rho_{SB}) \le 2_Q\ln\dim(\C H_S)$, where $\C H_S$ denotes the Hilbert space of the system $S$, the factor $2_Q$ is 2
for quantum systems and 1 for classical systems, and we assumed that (quite naturally) $\dim(\C H_S)\le\dim(\C H_B)$.
Thus, e.g., for a two-level system the entropy production would be bounded from above \emph{for all times} by $2_Q\ln2$,
which is clearly in general not the case. As a counterexample it suffices to consider, e.g., a driven system subjected
to a laser field which dissipates energy into its environment. The entropy production in this case should rather scale
extensively with time. This point was recently emphasized in Ref.~\cite{PtaszynskiEspositoArXiv2019}.
Therefore, one employs a second important step by noting the exact identity
$\Delta S_\text{vN}[\rho_B(t)] = \beta\Delta E_B - D[\rho_B(t)\|\pi_B]$, where
$\Delta E_B = \mbox{tr}_B\{H_B[\rho_B(t) - \rho_B(0)]\}$ is the change in bath energy and
$D[\rho\|\sigma] = \mbox{tr}\{\rho(\ln\rho-\ln\sigma)\} \ge 0$ is the always positive relative entropy. Then,
one identfies
\begin{equation}\label{eq EP open}
\tilde\Sigma \equiv \Delta S_S + \beta\Delta E_B = I[\rho_{SB}(t)] + D[\rho_B(t)\|\pi_B] \ge 0
\end{equation}
as the entropy production, here denoted by $\tilde\Sigma$ to distinguish it from our approach put forward later on.
Upon further identifying (minus) the change in bath energy with the heat flux into the system, $Q = -\Delta E_B$, one
obtains $\tilde\Sigma \equiv \Delta S_S - \beta Q$ as in Eq.~(\ref{eq EP}) for a single heat bath. It should be noted,
however, that the correct identification of heat is subtle outside the limit of a weakly coupled Markovian bath,
compare, e.g., with the discussion in the classical case~\cite{SeifertPRL2016, TalknerHaenggiPRE2016, JarzynskiPRX2017,
MillerAndersPRE2017, StrasbergEspositoPRE2017} or various approaches in quantum
thermodynamics~\cite{EspositoOchoaGalperinPRB2015, StrasbergEtAlNJP2016, BruchEtAlPRB2016, NewmanMintertNazirPRE2017,
BruchLewenkopfVonOppenPRL2018, StrasbergEtAlPRB2018, DouEtAlPRB2018, StrasbergEspositoPRE2019}.
Therefore, we will mainly use the notation $\Delta E_B$ here.
Let us summarize the picture so far: Eq.~(\ref{eq EP open}) proposes a definition of entropy production, which
\begin{enumerate}
\item[(i)] is positive for \emph{arbitary} system-bath dynamics and arbitary system and bath sizes based solely on
assumption~(\ref{eq initial state}),
\item[(ii)] has a natural information theoretic interpretation\footnote{Namely as deviation of the actual bath state
$\rho_B(t)$ from the equilibrium state $\pi_B$ as measured by the relative entropy. See also
Ref.~\cite{PtaszynskiEspositoArXiv2019} for a more specific discussion in that direction.}, and
\item[(iii)] has the conventional form as in phenomenological nonequilibrium dynamics [Eq.~(\ref{eq EP})], whereas
attention has to be paid to the point that the identification with heat is only valid in the weak coupling
limit.\footnote{In the weak coupling limit $\Delta E_B$ can be rigourously linked to the heat exchanges $Q$ (known,
e.g., from a weak coupling master equation) by use of the two point measurement approach, see
Ref.~\cite{EspositoHarbolaMukamelRMP2009}.}
\end{enumerate}
While these achievements are remarkable, they leave open the question what is the definition of thermodynamic
entropy for the universe $S_\text{univ}$, whose change equals the entropy production. If $\tilde\Sigma$ were the
correct definition of entropy production, then the thermodynamic entropy of the universe at time $t$ must be
\begin{equation}
\begin{split}
S_\text{univ}(t) &= S_\text{vN}[\rho_S(t)] + \beta\mbox{tr}_B\{H_B\rho_B(t)\} + c \\
&= S_\text{vN}[\rho_S(t)] + S_\text{vN}[\rho_B(t)] + D[\rho_B(t)\|\pi_B] + \tilde c \\
&= S_\text{vN}[\rho_S(t)] + S_\text{vN}[\rho_B(t)] + D[\rho_B(t)\|\rho_B(0)] + \tilde c
\end{split}
\end{equation}
such that $\Delta S_\text{univ}(t) = \tilde\Sigma$. Here $c$ and $\tilde c$ are arbitrary additive constants, which do
not depend on time. Obviously, the first line seems to be a rather awkward definition for thermodynamic entropy.
However, the second line reveals that this definition is identical to the local von Neumann entropies plus the relative
entropy $D[\rho_B(t)\|\pi_B]$. While looking more reasonable, this definition depends on the initial Gibbs state of the
bath (and therefore on the initial temperature), which is also unacceptable: the formal \emph{definition} of a
thermodynamic entropy should not depend on such details. For the third line we noticed that $\pi_B = \rho_B(0)$ such
that we can get rid of the formal dependence on $\pi_B$ in the definition, but unfortunately the so defined entropy
would no longer be a state function.
The goal of this paper is exactly to overcome this deficiency, while retaining the important properties (i) to
(iii) above. This requires us to put forward a different approach to the problem.
\subsection{Observational entropy}
Our notion of entropy for the universe, whose change we will identify with the entropy production, is based on the
recently introduced notion of ``observational'' (or ``coarse grained'') entropy by \u{S}afr\'anek, Deutsch and
Aguirre~\cite{SafranekDeutschAguirrePRA2019a, SafranekDeutschAguirrePRA2019b, SafranekDeutschAguirreArXiv2019}.
They claim that observational entropy provides a satisfactory generalization of standard thermodynamic
entropy to time-dependent out-of-equilibrium processes for arbitary isolated quantum and classical systems. What was
not investigated in Ref.~\cite{SafranekDeutschAguirrePRA2019a, SafranekDeutschAguirrePRA2019b,
SafranekDeutschAguirreArXiv2019}, however, is the question whether observational entropy also applies to the
open system paradigm considered here (i.e., where the isolated `system' is divided into a small system of primary
interest and the surrounding heat baths).
The basic idea of observational entropy is relatively easy explained in the classical case. Suppose we perform a set of
measurements on a thermodynamic system described by partitioning the phase space $\Gamma$ into different cells according
to some coarse-graining procedure. Thus, each coarse graining $\C C_\alpha = \{\C C_{\alpha_i}\}_i$ partitions the phase
space into non-overlapping regions corresponding to one measurement outcome, i.e., $\Gamma = \cup_i \C C_{\alpha_i}$ and
$\C C_{\alpha_i}\cap\C C_{\alpha_j} = \emptyset$ for $i\neq j$. Here, $\alpha\in\{1,\dots,M\}$ denotes the different
measurements (e.g., position of particles, energy, magnetization, etc.). The thermodynamic entropy of a system measured
in such a way is then postulated to be
\begin{equation}
S_\text{obs} \equiv -\sum_{\C C_1,\dots,\C C_M} p_{\C C_1,\dots,\C C_M}\ln\frac{p_{\C C_1,\dots,\C C_M}}{V_{\C C_1,\dots,\C C_M}},
\end{equation}
where $p_{\C C_1,\dots,\C C_M}$ denotes the probability to obtain the joint measurement outcome $(\C C_1,\dots,\C C_M)$
and $V_{\C C_1,\dots,\C C_M}$ describes the number of possible microstates associated to it.
A couple of remarks are in order. First, as also noted in Ref.~\cite{SafranekDeutschAguirrePRA2019a,
SafranekDeutschAguirrePRA2019b, SafranekDeutschAguirreArXiv2019}, not every observational entropy is thermodynamically
meaningful. Its usefulness depends crucially on the chosen observables. Second, for a `fine-grained' measurement of all
positions and momenta of all particles, we reproduce the Gibbs-Shannon entropy. On the other hand, if we measure only the
energy $E$, and if this is a conserved quantity, then we get the Boltzmann (surface but not volume!) entropy $\ln V_E$
back. Finally, a quantum extension by using projection operators in Hilbert space is possible but more complicated
as the \emph{order} of the measurements now plays a role~\cite{SafranekDeutschAguirrePRA2019a,
SafranekDeutschAguirrePRA2019b}. \u{S}afr\'anek \emph{et al.}~argue, however, that for typical situations encountered
in thermodynamics, the effect of the non-commutativity of the measurements becomes very small. In our case we will
actually choose commuting measurements, namely one measurement performed on the system Hilbert space and
one on the bath Hilbert space. The construction of observational entropy is then straightforward, see below.
\subsection{Outline}
The rest of this paper is organized as follows: In Sec.~\ref{sec general} we will present our choice of measurements,
which defines the observational entropy. Then, we will show that the change of observational entropy provides a
legitimate candidate for entropy production by verifying the points (i) to (iii) above for \emph{arbitary} open system
dynamics as in Sec.~\ref{sec previous results}. We will also quantitaively compare our definition of entropy production
with the previous approach and the case of multiple heat baths is treated in Sec.~\ref{sec multiple baths}. Additional
generally valid observations are discussed in Sec.~\ref{sec add obs}. While the results in Sec.~\ref{sec general} are
very general, they are necessarily also a bit abstract. Hence, in Sec.~\ref{sec MME} we will derive a `microcanonical
master equation' for a system weakly coupled to a Markovian bath. This master equation takes explicitly into account
changes of the bath energy. Within this important approximation we will then see that we are also able to verify all
the above properties. Finally, we will conclude our findings in Sec.~\ref{sec conclusions}.
\section{General picture}
\label{sec general}
Our definition of thermodynamic entropy for the system-bath setup requires two measurements: an arbitrary (but
fine-grained) measurement of the system described by a set of rank-1 projectors $|s\rl s|$,
$s\in\{1,\dots,\dim\C H_S\}$, and a measurement of the bath energy described by a set of projection operators
$\Pi_{E,\delta} = \sum_{E_i\in(E-\delta,E]} \Pi_{E_i,0}$. Here, $\delta$ denotes a suitable width of the measured
energy window (to be further specified in Sec.~\ref{sec ideal}) and $\Pi_{E_i,0}$ projects on a sharp energy $E_i$ such
that $H_B \Pi_{E_i,0} = E_i\Pi_{E_i,0}$ (notice that $\Pi_{E_i,0}$ is not a rank-1 projector in case of exact
degeneracies). For an arbitary system-bath state $\rho_{SB}$ the average post measurement state is given by
\begin{equation}\label{eq CPTP meas map}
\sum_{s,E} |s\rl s|\Pi_{E,\delta}\rho_{SB}\Pi_{E,\delta}|s\rl s|
= \sum_{s,E} p_{sE}|s\rl s|\otimes \rho_B(s,E).
\end{equation}
Here, $p_{sE} = \lr{s|\mbox{tr}_B\{\Pi_{E,\delta}\rho_{SB}\}|s}$ denotes the probability to obtain measurement outcome
$(s,E)$ and the state of the bath conditioned on that outcome is
$\rho_B(s,E) = \lr{s|\Pi_{E,\delta}\rho_{SB}\Pi_{E,\delta}|s}/p_{sE}$. Note that the average post measurement
state~(\ref{eq CPTP meas map}) has lost all quantum correlations, but is in general classically correlated. Now, the
observational entropy~\cite{SafranekDeutschAguirrePRA2019a, SafranekDeutschAguirrePRA2019b,
SafranekDeutschAguirreArXiv2019} becomes in this case
\begin{equation}
\boxed{
S_\text{obs} \equiv -\sum_{s,E} p_{sE} \ln\frac{p_{sE}}{V_{E,\delta}},
}
\end{equation}
where $V_{E,\delta} = \mbox{tr}_B\{\Pi_{E,\delta}\}$ is the number of microstates in the bath with respect to a given
energy window $(E-\delta,E]$.
The claim of this paper is now that $\Sigma \equiv \Delta S_\text{obs}(t) = S_\text{obs}(t) - S_\text{obs}(0)$ is the
entropy production of a system (perhaps subjected to a time-dependent driving $\lambda_t$) coupled to a single heat
bath (multiple baths are treated in Sec.~\ref{sec multiple baths}) by verifying the points (i) to (iii) from above.
As in Sec.~\ref{sec previous results} we will need an additional condition on the initial system-bath state,
which is, however, milder than Eq.~(\ref{eq initial state}). Furthermore, we remark that our approach is in some sense
close to the two-point measurement approach~\cite{EspositoHarbolaMukamelRMP2009} with the difference that we do not
have to perfectly measure the energy of the bath, see Sec.~\ref{sec ideal}.
Before verifying the points (i) to (iii) in Sec.~\ref{sec positivity EP}, we start with two general identities in
Sec.~\ref{sec two identities} followed by a discussion in Sec.~\ref{sec ideal} of how small we have to choose the width
$\delta$ in an experiment such that all theoretical claims of this paper remain true even if $\delta>0$.
\subsection{Two general identities}
\label{sec two identities}
We start by noting that for any system-bath state $\rho_{SB}$
\begin{equation}\label{eq identity 1}
S_\text{obs} = S_\text{Sh}(p_s) + S_\text{obs}^{E_B} - I(p_{sE}).
\end{equation}
Here, $S_\text{Sh}(p_s) \equiv -\sum_s p_s\ln p_s$ is the Shannon entropy of the probability distribution
$p_s = \sum_E p_{sE}$ and
\begin{equation}\label{eq obs entropy B}
S_\text{obs}^{E_B} \equiv -\sum_E p_E \ln\frac{p_E}{V_{E,\delta}}
\end{equation}
can be interpreted as the observational entropy of the bath alone with respect to an energy measurement. Finally,
$I(p_{sE}) = \sum_{s,E}p_{sE} \ln(p_{sE}/p_s p_E)$ is the classical mutual information between the measurement results
of the system state and the bath energy.
The second identity concerns only $S_\text{obs}^{E_B}$. For an arbitrary bath state $\rho_B$ let
$\rho_B(E) \equiv \Pi_{E,\delta}\rho_B\Pi_{E,\delta}/p_E$ be the post-measurement state of the bath conditioned on
outcome $E$ ignoring the measurement result $s$. This state is obtained with probability
$p_E = \sum_s p_{sE} = \mbox{tr}_B\{\Pi_{E,\delta}\rho_B\}$. Furthermore, we introduce the microcanonical equilibrium
state $\rho_\text{mic}(E) = \Pi_{E,\delta}/V_{E,\delta}$ with respect to a given energy $E$. Then, a straightforward
calculation reveals
\begin{equation}\label{eq identity 2}
S_\text{obs}^{E_B} = S_\text{vN}\left[\sum_Ep_E\rho_B(E)\right] + \sum_E p_E D[\rho_B(E)\|\rho_\text{mic}(E)],
\end{equation}
where we used the identity
\begin{equation}\label{eq theorem 1110}
S_\text{Sh}(p_E) + \sum_i p_E S_\text{vN}[\rho_B(E)] = S_\text{vN}\left[\sum_E p_E\rho_B(E)\right],
\end{equation}
which holds since the states $\rho_B(E)$ are supported on orthogonal subspaces, see Theorem~11.10 in
Ref.~\cite{NielsenChuangBook2000}. Equation~(\ref{eq identity 2}) tells us that the observational entropy of the bath
is identical to the fine-grained (von Neumann) entropy of the average post-measurement state \emph{plus} the additional
ignorance (measured by the relative entropy) due to not knowing the precise microstate of the bath for a given energy
$E$.
\subsection{Initial state and ideal measurement limit}
\label{sec ideal}
For the moment, let us consider the initial state~(\ref{eq initial state}), generalizations are discussed in
Sec.~\ref{sec add obs}. The crucial ingredient to show positivity of the second law in our approach is that, initially
at time $t=0$, the observational entropy of the bath $S_\text{obs}^{E_B}$ must coincide with the von Neumann entropy of
the average post measurement state $\sum_E p_E(0)\rho_B(E,0)$. This means that, using Eq.~(\ref{eq identity 2}), the
following expression must vanish:
\begin{equation}\label{eq difference entropies}
S_\text{obs}^{E_B}(0) - S_\text{vN}\left[\sum_E p_E(0)\rho_B(E,0)\right]
= \sum_E p_E(0) D[\rho_B(E,0)\|\rho_\text{mic}(E)]
\end{equation}
For a Gibbs state of the bath we have
\begin{equation}
p_E(0) = \pi_E \equiv \sum_{E_i\in(E-\delta,E]} \frac{e^{-\beta E_i} V_{E_i,0}}{\C Z_B}, ~~~
\rho_B(E,0) = \pi_B(E) \equiv \sum_{E_i\in(E-\delta,E]} \Pi_{E_i,0} \frac{e^{-\beta E_i}}{p_E(0)\C Z_B}.
\end{equation}
Here, the projector $\Pi_{E_i,0}$ has rank $V_{E_i,0} = \mbox{tr}_B\{\Pi_{E_i,0}\}$ (which is greater than one in case
of exact degeneracies). Using this, it becomes clear that Eq.~(\ref{eq difference entropies}) can be written as
\begin{equation}
S_\text{obs}^{E_B}(0) - S_\text{vN}(\pi_B)
= \sum_E \sum_{E_i\in(E-\delta,E]} \frac{e^{-\beta E_i}V_{E_i,0}}{Z_B}\ln\frac{e^{-\beta E_i}V_{E,\delta}}{\sum_{E_j\in(E-\delta,E]} e^{-\beta E_j} V_{E_j,0}},
\end{equation}
which vanishes in the ideal theoretical limit $\delta\rightarrow0$. For finite $\delta$ we proceed by looking at
the argument of the logarithm,
\begin{equation}
\frac{e^{-\beta E_i}V_{E,\delta}}{\sum_{E_i\in(E-\delta,E]} V_{E_i,0} e^{-\beta E_i}}
= \frac{e^{\beta\delta_i}V_{E,\delta}}{\sum_{j\in[0,\delta)} e^{\beta\delta_j}V_{E-\delta_j,0}},
\end{equation}
where we defined $E_j = E-\delta_j$. Now, if the bath is macroscopically large, we expect that we can replace
the sums by integrals and by using the mean value theorem for integration, we end up with
\begin{equation}
\frac{e^{\beta\delta_i}V_{E,\delta}}{\int_0^\delta dx e^{\beta x}V_{E-x,0}}
= \frac{e^{\beta\delta_i}V_{E,\delta}}{e^{\beta\xi}\int_0^\delta dx V_{E-x,0}}
= e^{\beta(\delta_i-\xi)} = 1 + \beta(\delta_i-\xi) + \C O[(\beta\delta)^2]
\end{equation}
with $\xi\in[0,\delta)$ and we used $\int_0^\delta dx V_{E-x,0} = V_{E,\delta}$. Thus, in the limit
\begin{equation}\label{eq condition experiment}
\boxed{
\beta\delta \ll 1
}
\end{equation}
the observational entropy practically coincides with the von Neumann entropy for a thermal state. Experimentally,
this condition has to be met in order to ensure positivity of entropy production if the bath is initially in a Gibbs
state. In the following we will assume that the energy window $\delta$ is chosen small enough such that
Eq.~(\ref{eq condition experiment}) holds and we will henceforth simply denote $V_E = V_{E,\delta}$ and
$\Pi_E = \Pi_{E,\delta}$.
\subsection{Verifying points (i) to (iii) and comparison with the previous approach for a single heat bath}
\label{sec positivity EP}
For an arbitrary initial system-bath state $\rho_{SB}(0)$ the change in observational entropy can be expressed as
\begin{equation}
\begin{split}
\Delta S_\text{obs} =&~ \Delta S_\text{Sh}[p_s(t)] + \Delta S_\text{obs}^{E_B} - \Delta I[p_{sE}(t)] \\
=&~ S_\text{Sh}[p_s(t)] - S_\text{Sh}[p_s(0)]
+ S_\text{vN}\left[\sum_Ep_E(t)\rho_B(E,t)\right] - S_\text{vN}\left[\sum_Ep_E(0)\rho_B(E,0)\right] \\
& + \sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)] - \sum_E p_E(0) D[\rho_B(E,0)\|\rho_\text{mic}(E)]
- I_{s:E}(t) + I_{s:E}(0),
\end{split}
\end{equation}
where we used Eqs.~(\ref{eq identity 1}) and~(\ref{eq identity 2}). Here, $p_s(t) = \lr{s|\rho_S(t)|s}$ and
$\rho_B(E,t) = \Pi_E\rho_B(t)\Pi_E/p_E(t)$. Next, we assume the initial state and width $\delta$ to be as described in
Sec.~\ref{sec ideal}, such that $S_\text{obs}^{E_B}(0) = S_\text{vN}(\pi_B)$. This allows us to confirm
\begin{equation}
\begin{split}
\Delta S_\text{obs} =&~ S_\text{Sh}[p_s(t)] - S_\text{Sh}[p_s(0)]
+ S_\text{vN}\left[\sum_Ep_E(t)\rho_B(E,t)\right] - S_\text{vN}(\pi_B) \\
& + \sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)] - I[p_{sE}(t)].
\end{split}
\end{equation}
Now, as the entropy is preserved during any unitary evolution, we get
$S_\text{Sh}[p_s(0)] + S_\text{vN}(\pi_B) = S_\text{vN}[\rho_{SB}(t)]$ where $\rho_{SB}(t)$ is the
time-evolved state starting from the initial state $\rho_{SB}(0) = \sum_s p_s(0)|s\rl s|\otimes\pi_B$. Writing also
$S_\text{vN}[\rho_{SB}(t)] = S_\text{vN}[\rho_S(t)] + S_\text{vN}[\rho_B(t)] - I[\rho_{SB}(t)]$, we get
\begin{equation}
\begin{split}\label{eq Delta S obs 1}
\Delta S_\text{obs} =&~ S_\text{Sh}[p_s(t)] - S_\text{vN}[\rho_S(t)]
+ S_\text{vN}\left[\sum_Ep_E(t)\rho_B(E,t)\right] - S_\text{vN}[\rho_B(t)] \\
& + \sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)] + I[\rho_{SB}(t)] - I[p_{sE}(t)].
\end{split}
\end{equation}
The positivity of $\Delta S_\text{obs}$ is now evident. First, by using that a projective measurement increases the
entropy on average, Theorem~11.9 in Ref.~\cite{NielsenChuangBook2000}, we confirm that
\begin{equation}
S_\text{Sh}[p_s(t)] \ge S_\text{vN}[\rho_S(t)], ~~~
S_\text{vN}\left[\sum_Ep_E(t)\rho_B(E,t)\right] \ge S_\text{vN}[\rho_B(t)].
\end{equation}
We expect both contributions, however, to be rather small. First, the change in system entropy due to the measurement
is at most $\ln\dim(\C H_S)$ and will likely be much smaller. Especially, we are free to choose the basis of the final
system measurement at time $t$ such that we can let it coincide with the eigenbasis of $\rho_S(t)$. Second, a large
change in bath entropy due to the final measurement requires the existence of large coherences
$\mbox{tr}_B\{\Pi_E\rho_B(t)\Pi_{E'}\} \neq 0$ between different energy sectors $E\neq E'$. This also seems very
unlikely as it would imply the existence of macrosopic Schr\"odinger cat states in the bath.
Finally, another small contribution comes from the difference in mutual information, which always obeys
\begin{equation}
\begin{split}\label{eq inequality mutual info}
I[\rho_{SB}(t)] &= D[\rho_{SB}(t)\|\rho_S(t)\otimes\rho_B(t)] \\
&\ge D\left[\sum_{s,E} p_{sE}(t) |s\rl s|\otimes\rho_\text{mic}(E)\left\|\sum_s p_s(t)|s\rl s|\otimes\sum_E p_E(t)\rho_\text{mic}(E)\right]\right. = I[p_{sE}(t)],
\end{split}
\end{equation}
This follows from monotonicity of relative entropy~\cite{UhlmannCMP1977, OhyaPetzBook1993} by noting that
$\sum_{s,E} p_{sE}(t) |s\rl s|\otimes\rho_\text{mic}(E) = \Phi\rho_{SB}(t)$ with the completely positive and
trace-preserving map defined via
\begin{equation}
\Phi\rho_{SB} \equiv
\sum_E \mbox{tr}_B\left\{|s\rl s|\Pi_E\rho_{SB}|s\rl s|\Pi_E\right\} \otimes \rho_\text{mic}(E).
\end{equation}
Thus, all together we can conclude that
\begin{equation}\label{eq second law interpretation}
\boxed{
\Delta S_\text{obs} \gtrsim \sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)] \ge 0,
}
\end{equation}
where we used the symbol ``$\gtrsim$'' to indicate that the difference between the two sites of the inequality is
expected to be rather small in the typical situation of a large heat bath. The only part, which can scale extensively
with time, is $\sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)]$. Therefore, we conclude that \emph{entropy production
arises because for a given energy $E$ of the bath we lose track of its exact microstate} compared to the maximal
uninformative microcanonical ensemble, where all microstates are assumed to be equally likely. Thus, we have confirmed
the points (i) and (ii) from above.
Next, to confirm (iii), we want to show that for a weakly coupled, macroscopic bath the change in observational bath
entropy is proportional to its change in energy. For that purpose it is indeed crucial to assume that the bath is
initially in a Gibbs state. Then, we write $p_E(t) = \pi_E + \epsilon q_E(t)$ with $\pi_E = e^{-\beta E}V_E/\C Z_B$
and $q_E(t)$ is a set of numbers such that $\sum_E q_E(t) = 0$. Now, our assumption is that $\epsilon$ is a small
parameter, i.e., the \emph{distribution of energies} in the bath remains close to the canonical probabilities
throughout the time-evolution. This should be typically justified for a weakly coupled, macroscopic bath. Notice that
this does \emph{not} imply that the entire bath state $\rho_B(t)$ is close to the Gibbs state $\pi_B$. Then, we can
write
\begin{equation}
\Delta S_\text{obs}^{E_B}
= -\sum_E [\pi_E + \epsilon q_E(t)]\ln\frac{\pi_E + \epsilon q_E(t)}{V_E} + \sum_E \pi_E \ln\frac{\pi_E}{V_E}
=\beta\Delta E_B + \C O(\epsilon^2)
\end{equation}
with the chang ein bath energy $\Delta E_B = \epsilon\sum_E E q_E(t)$. Hence,
\begin{equation}
\boxed{
\Delta S_\text{obs} = \Delta S_\text{Sh}[p_S(t)] + \Delta S_\text{obs}^{E_B} - I[p_{sE}(t)]
\approx \Delta S_\text{Sh}[p_S(t)] - \beta Q \ge 0,
}
\end{equation}
where we ignored the small contribution $I[p_{sE}(t)]$ at the end. We have also identified $Q = -\Delta E_B$, which is
justified in the limit considered here to derive $\Delta S_\text{obs}^{E_B} \approx \beta\Delta E_B$.
Thus, as a preliminary conclusion, we have shown that $\Sigma \equiv \Delta S_\text{obs}$ fulfills the three desired
properties (i) to (iii) and therefore, provides a more suitable candidate for entropy production then $\tilde\Sigma$
from Eq.~(\ref{eq EP open}) because the latter cannot be expressed as the change of a meaningful thermodynamic entropy
for the system and the bath.
Nevertheless, it is instructive to compare $\Sigma$ and $\tilde\Sigma$ quantitatively.
From Eqs.~(\ref{eq Delta S obs 1}) and~(\ref{eq EP open}) we obtain
\begin{equation}
\begin{split}\label{eq comparison help}
\Sigma - \tilde\Sigma = &~ S_\text{Sh}[p_s(t)] + S_\text{vN}\left[\sum_Ep_E(t)\rho_B(E,t)\right]
- S_\text{vN}[\rho_{SB}(t)] + \sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)] - I[p_{sE}(t)] \\
&- I[\rho_{SB}(t)] - D[\rho_B(t)\|\pi_B].
\end{split}
\end{equation}
We consider first the very last term. This becomes after a little massage
\begin{equation}
\begin{split}
D[\rho_B(t)\|\pi_B]
&= \mbox{tr}_B\left\{\rho_B(t)\left[\ln\rho_B(t) - \sum_E \Pi_E \ln\rho_\text{mic}(E)\rho_B(E)\right]\right\} \\
&= -S_\text{vN}[\rho_B(t)] - \sum_E p_E(t)\mbox{tr}_B\{\rho_B(E,t)\ln\rho_\text{mic}(E)\} - \sum_E p_E(t)\ln\pi_B(E).
\end{split}
\end{equation}
By combining this result with the second and forth term of Eq.~(\ref{eq comparison help}), we verify
\begin{equation}
S_\text{vN}\left[\sum_Ep_E(t)\rho_B(E,t)\right] + \sum_E p_E(t) D[\rho_B(E,t)\|\rho_\text{mic}(E)] - D[\rho_B(t)\|\pi_B] = S_\text{vN}[\rho_B(t)] - D[p_E(t)\|\pi_B(E)],
\end{equation}
where we also made use of Eq.~(\ref{eq theorem 1110}). Thus, we end up with the compact expression
\begin{equation}\label{eq comparison EPs 0}
\Sigma - \tilde\Sigma = S_\text{Sh}[p_s(t)] - S_\text{vN}[\rho_S(t)] - D[p_E(t)\|\pi_B(E)] - I[p_{sE}(t)].
\end{equation}
Taken together, the first two terms are non-negative (and identical to zero for classical systems), whereas each of
the remaining two terms is negative. In the weak coupling regime investigated above, we expect all terms to be rather
small. In fact, if $p_E(t) = \pi_E + \epsilon q_E(t)$ as above, then $D[p_E(t)\|\pi_B(E)] = \C O(\epsilon^2)$ such that
we can conclude $\Sigma\approx\tilde\Sigma$. Outside this regime, the only possibly unbounded term is
$D[p_E(t)\|\pi_B(E)]$ as the number of populated energy levels for a very small $\delta$ can become very large. We
therefore expect that, typically, we have
\begin{equation}\label{eq comparison EPs}
\boxed{
\tilde\Sigma \gtrsim \Sigma \ge 0.
}
\end{equation}
This means that the previous approach typically yields a larger entropy production than our novel definition.
Intuitively, this makes sense: in our approach we ideally know the entire distribution of energies in the bath whereas
in the previous approach only knowing the average energy flow to the bath is sufficient. Clearly, our approach stores
more information and hence, less entropy is produced.
\subsection{Extension to multiple heat baths}
\label{sec multiple baths}
The extension to multiple heat baths labeled by $\nu\in\{1,\dots,N\}$ is straightforward by measuring the energy of
each bath. The observational entropy is in this case
\begin{equation}
S_\text{obs} = -\sum_{s,E_1,\dots,E_N} p_{sE_1\dots E_N}\ln\frac{p_{sE_1\dots E_N}}{V_{E_1}\dots V_{E_N}}.
\end{equation}
Here, we used that the number of microstates naturally factorizes,
$V_{E_1\dots E_N} = \mbox{tr}_{B_1\dots B_N}\{\Pi_{E_1}\dots\Pi_{E_N}\} = V_{E_1}\dots V_{E_N}$, where $\Pi_{E_\nu}$
describes the projector associated to measurement outcome $E_\nu$ of bath $\nu$. Furthermore, the natural
generalization of the initial state~(\ref{eq initial state}) to multiple baths is
\begin{equation}\label{eq initial state multiple baths}
\rho_\text{tot}(0) = \rho_S(0)\otimes\pi_{B_1}\otimes\dots\otimes\pi_{B_N},
\end{equation}
which was also used in Ref.~\cite{EspositoLindenbergVandenBroeckNJP2010}. We note that every bath can have initially
a different inverse temperature, i.e., $\pi_{B_\nu} = e^{-\beta_\nu H_B^{(\nu)}}/\C Z_{B_\nu}$.
Under these circumstances (assuming that $\delta$ is chosen as in Sec.~\ref{sec ideal}) we easily confirm that the
initial observational entropy is identical to
\begin{equation}
S_\text{obs}(0) = S_\text{vN}\left[\sum_s p_s(0)|s\rl s|\otimes\pi_{B_1}\otimes\dots\otimes\pi_{B_N}\right]
= S_\text{vN}[\rho_\text{tot}(t)].
\end{equation}
Furthermore, the observational entropy at time $t$ can be split into its `local' parts and its correlations, similar
to Eq.~(\ref{eq identity 1}). Specifically,
\begin{equation}
\begin{split}
S_\text{obs}(t)
&= -\sum_s p_s(t)\ln p_s(t) - \sum_\nu \sum_{E_\nu} p_{E_\nu}(t)\ln\frac{p_{E_\nu}(t)}{V_{E_\nu}}
- \sum_{s,E_1,\dots,E_N} p_{sE_1\dots E_N}(t) \ln\frac{p_{sE_1\dots E_N}(t)}{p_s(t)p_{E_1}(t)\dots p_{E_N}(t)} \\
&\equiv S_\text{Sh}[p_s(t)] + \sum_\nu S_\text{obs}^{E_{B_\nu}}(t) - I_\text{cor}[p_{sE_1\dots E_N}(t)].
\end{split}
\end{equation}
Further use of relation~(\ref{eq identity 2}) reveals that
\begin{equation}
S_\text{obs}(t)
= S_\text{Sh}[p_S(t)] + \sum_\nu S_\text{vN}\left[\sum_{E_\nu} p_{E_\nu}\rho_{B_\nu}(E_\nu,t)\right] +
\sum_\nu \sum_{E_\nu} p_{E_\nu} D[\rho_{B_\nu}(E_\nu,t)\|\rho_\text{mic}(E_\nu)] - I_\text{cor}[p_{sE_1\dots E_N}(t)].
\end{equation}
Thus, similarly to Eq.~(\ref{eq Delta S obs 1}), the change in observational entropy can be split into a family of
terms, whose non-negativity is evident:
\begin{equation}
\begin{split}
\Delta S_\text{obs} =&~ S_\text{Sh}[p_S(t)] - S_\text{vN}[\rho_S(t)]
+ \sum_\nu\left\{S_\text{vN}\left[\sum_{E_\nu} p_{E_\nu}\rho_{B_\nu}(E_\nu,t)\right] - S_\text{vN}[\rho_{B_\nu}(t)]\right\} \\
& + \sum_\nu \sum_{E_\nu} p_{E_\nu} D[\rho_{B_\nu}(E_\nu,t)\|\rho_\text{mic}(E_\nu)]
+ I_\text{cor}[\rho_\text{tot}(t)] - I_\text{cor}[p_{sE_1\dots E_N}(t)].
\end{split}
\end{equation}
As before, we expect the non-negativity of the first line to be rather small (especially, it is exactly zero for
classical systems) such that
\begin{equation}
\Delta S_\text{obs} \gtrsim \sum_\nu \sum_{E_\nu} p_{E_\nu} D[\rho_{B_\nu}(E_\nu,t)\|\rho_\text{mic}(E_\nu)]
+ I_\text{cor}[\rho_\text{tot}(t)] - I_\text{cor}[p_{sE_1\dots E_N}(t)].
\end{equation}
Here, we introduced the notation
\begin{equation}
I_\text{cor}[\rho_\text{tot}(t)]
\equiv S_\text{vN}[\rho_S(t)] + \sum_\nu S_\text{vN}[\rho_{B_\nu}(t)] -S_\text{vN}[\rho_\text{tot}(t)]
= D[\rho_\text{tot}(t)\|\rho_S(t)\otimes\rho_{B_1}(t)\otimes\dots\otimes\rho_{B_N}(t)].
\end{equation}
Since we can also write
\begin{equation}
\begin{split}
I_\text{cor}&[p_{sE_1\dots E_N}(t)] = \\
&D\left[\sum_{s,E_1,\dots,E_N}p_{sE_1\dots E_N}(t)|s\rl s|\rho_\text{mic}(E_1)\dots\rho_\text{mic}(E_N)\left\|\sum_sp_s(t)|s\rl s|\sum_{E_1}p_{E_1}(t)\rho_\text{mic}(E_1)\dots\sum_{E_N}p_{E_N}(t)\rho_\text{mic}(E_N)\right]\right.,
\end{split}
\end{equation}
we can confirm $I_\text{cor}[\rho_\text{tot}(t)] - I_\text{cor}[p_{sE_1\dots E_N}(t)] \ge 0$ similar to
Eq.~(\ref{eq inequality mutual info}). However, in contrast to the case of a single heat bath, we can no longer expect
this contribution to be small as the different baths can become correlated. This is especially true outside the
weak coupling case. Therefore, in the case of multiple baths we expect that there are two major contributions to the
entropy production. In any case, $\Delta S_\text{obs} \ge 0$ is ensured.
Furthermore, following the same procedure as above, we can confirm in the weak coupling case that
$\Delta S_\text{obs}^{E_\nu} = \beta_\nu\Delta E_{B_\nu}$. In addition, we expect in the weak coupling limit that
$I_\text{cor}[\rho_\text{tot}(t)] \gtrsim I_\text{cor}[p_{sE_1\dots E_N}(t)] \approx
\sum_\nu I[p_{sE_\nu}(t)]$, i.e., each bath acts like a separate bath entering independently the master equation
describing the system~\cite{EspositoHarbolaMukamelRMP2009, KosloffEntropy2013, SchallerBook2014}.\footnote{A critical
discussion of this point can be found in Ref.~\cite{MitchisonPlenioNJP2018}.} Hence, we can write
$\Delta S_\text{obs}(t) \approx \Delta S_\text{Sh}[p_s(t)] - \sum_\nu\beta_\nu Q_\nu \ge 0$. This provides a
microscopic derivation of the phenomenological second law of nonequilibrium thermodynamics~(\ref{eq EP}).
\subsection{Additional observations}
\label{sec add obs}
We end this general section with a couple of interesting observations:
\emph{Observation 1.} The crucial ingredient to prove point~(i), positivity of $\Delta S_\text{obs}$, is that
Eq.~(\ref{eq difference entropies}) vanishes. This is not only true for a Gibbs state and a small enough measurement
width $\delta$. Indeed, Eq.~(\ref{eq difference entropies}) can be zero for many different initial energy
distributions $p_E(0)$ as long as the distribution of microstates within a given energy window is very close to the
microcanonical ensemble. The initial state of the bath can even contain quantum coherences between different energy
sectors as those get killed during the initial measurement.
\emph{Observation 2.} One can also choose different measurements of the bath and positivity of $\Delta S_\text{obs}$
will still hold as long as the initial observational entropy of the bath coincides with the von Neumann entropy of the
average post measurement state. In that respect the energy only seems to be an outstanding observable due to its
connection to the first law. To capture the effect of multiple conserved quantities, we can consider additional
measurements, e.g., of the energy and particle number of the bath in case of a grand-canonical reservoir.
\emph{Observation 3.} We also do not expect the initial product state assumption to be crucial. An initially correlated
system-bath state lowers the entropy production by at most $I[p_{sE}(0)]$, which is typically negligible with respect
to the positive terms appearing in Eq.~(\ref{eq second law interpretation}).
\emph{Observation 4.} Instead of taking into account correlations between the measurement results of the system state
and the bath energy, we could also neglect them in the definition of observational entropy. All the three points~(i)
to~(iii) would remain valid for the choice $\tilde S_\text{obs}(t) \equiv S_\text{Sh}[p_s(t)] + S_\text{obs}^{E_B}(t)$.
In particular, we would typically have $\Delta S_\text{obs} \approx \Delta\tilde S_\text{obs}$.
\emph{Observation 5.} Finally, we emphasize that none of our results depends on the particular form of the Hamiltonian.
Especially, the system-bath coupling could be time-dependent, $H_{SB} = H_{SB}(\lambda_t)$, and even the bath
Hamiltonian could depend on time, $H_B = H_B(\lambda_t)$.
\section{The microcanonical master equation}
\label{sec MME}
In this section we illustrate our general findings in the limit of a weakly coupled, Markovian bath. In contrast to
conventional master equations~\cite{SeifertRPP2012, KosloffEntropy2013, SchallerBook2014, BreuerPetruccioneBook2002},
we will derive a master equation describing the evolution of the system state \emph{and} the bath energies by using a
correlated projection-operator method. Such a master equation was first derived by Esposito and
Gaspard~\cite{EspositoGaspardPRE2003} and we repeat a (slightly more generalized) derivation in
Appendix~\ref{sec MME derivation}. Here, we will investigate in detail the analytical and thermodynamic properties of
this master equation, which was not done in Ref.~\cite{EspositoGaspardPRE2003}. We will call this approach the
`microcanonical master equation' (MME) in the following.
The MME is a Pauli-like rate master equation for the probabilities $p_{sE}(t)$ to find the system in state $s$ and
the energy of the bath at $E$ at time $t$. It reads
\begin{equation}\label{eq Pauli MME}
\partial_t p_{sE}(t) = \sum_{\alpha,\gamma}\sum_{s'}
\frac{2\pi}{V_E}\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}
\left[\frac{V_E}{V_{E+\epsilon_s-\epsilon_{s'}}}p_{s',E+\epsilon_s-\epsilon_{s'}}(t) - p_{sE}(t)\right].
\end{equation}
Here, the overall timescale of the dynamics is governed by the rate
$\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}$. If one assumes that the
system-bath coupling Hamiltonian reads $H_{SB} = \sum_\alpha S_\alpha\otimes B_\alpha$, where $S_\alpha$ ($B_\alpha$)
are Hermitian system (bath) operators, then $S^{ss'}_\alpha \equiv \lr{s|S_\alpha|s'}$ describes the transition
matrix elements with respect to the basis $|s\rangle$, which is assumed in this section to be the (non-degenerate)
energy eigenbasis of $H_S$. Furthermore, the function
$f_{\alpha\gamma}(E,E') \equiv \mbox{tr}_B\{\Pi_E B_\alpha\Pi_{E'} B_\gamma\}$ describes how well the energies in
the bath get redistributed. In obeys the useful relations~(\ref{eq symmetries f}).
We remark that Eq.~(\ref{eq Pauli MME}) reduces to Eq.~(42) of Ref.~\cite{EspositoGaspardPRE2003} in the limit of a
single system coupling operator $S_\alpha = \delta_{\alpha,1}S$. Multiple baths can be easily included by summing over
$\nu$ and adding this superscript to $S_\alpha$, $f$ and $V$, but we will only consider a single heat bath here. As
this equation describes the time-evolution of all energies, we will indeed find out below that the dynamics of this
equation are entropy dominated. Furthermore, we remark that driven system energies can be considered by replacing
$\epsilon_s$ by $\epsilon_s(\lambda_t)$ provided that the change of energies is slow compared to the decay of the
bath correlation functions.
\subsection{Properties}
\label{sec MME properties}
\subsubsection{Reduction to the conventional Pauli master equation}
\label{sec Pauli BMS}
As a simple crosscheck we investigate the limit in which our MME reduces to the conventional Pauli master equation
derived within the Born-Markov-secular approximation~\cite{BreuerPetruccioneBook2002}. Formally, we can write
Eq.~(\ref{eq Pauli MME}) after summing over $E$ as
\begin{equation}
\partial_t p_s(t) = \sum_E\sum_{\alpha,\gamma}\sum_{s'}
\frac{2\pi}{V_E}\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}
\left[\frac{V_E}{V_{E+\epsilon_s-\epsilon_{s'}}}p_{E+\epsilon_s-\epsilon_{s'}|s'}(t)p_{s'}(t) - p_{E|s}(t)p_s(t)\right].
\end{equation}
Here, we have introduced the conditional probability $p_{E|s}(t) \equiv p_{sE}(t)/p_s(t)$ and we will now assume that
this is approximately given by $p_{E|s}(t) \approx V_E e^{-\beta E}/\C Z_B$ for all $s$ and all times $t$. This
simplifies the expression to
\begin{equation}
\partial_t p_s(t) = \frac{2\pi}{Z_B}\sum_E\sum_{\alpha,\gamma}\sum_{s'}
\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}
\left[e^{-\beta(E+\epsilon_s-\epsilon_{s'})}p_{s'}(t) - e^{-\beta E}p_s(t)\right].
\end{equation}
We now introduce the functions
$g_{\alpha\gamma}(\epsilon_s-\epsilon_{s'}) \equiv \sum_E f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'}) \frac{e^{-\beta E}}{\C Z_B}$, which obey the symmetries
$g_{\alpha\gamma}(\epsilon_s-\epsilon_{s'}) = e^{\beta(\epsilon_s-\epsilon_{s'})} g_{\gamma\alpha}(\epsilon_{s'}-\epsilon_s)$ and $g_{\alpha\gamma}^*(\epsilon_s-\epsilon_{s'}) = g_{\gamma\alpha}(\epsilon_s-\epsilon_{s'})$.
They allows us to write
\begin{equation}\label{eq Pauli BMS}
\partial_t p_s(t) = 2\pi\sum_{\alpha,\gamma}\sum_{s'}
\Re\{S_\alpha^{ss'}S_\gamma^{s's}g_{\alpha\gamma}(\epsilon_s-\epsilon_{s'})\}
\left[e^{\beta(\epsilon_{s'}-\epsilon_s)}p_{s'}(t) - p_s(t)\right].
\end{equation}
This corresponds to the typical Pauli master equation~\cite{BreuerPetruccioneBook2002} with the rates
satisfying local detailed balance, i.e., the rate to jump from $s'$ to $s$ is enhanced by a factor
$e^{\beta(\epsilon_{s'}-\epsilon_s)}$ compared to the inverse jump rate from $s$ to $s'$ if
$\epsilon_{s'} > \epsilon_s$.
\subsubsection{Conservation of energy}
To confirm conservation of energy, we note that
\begin{equation}
\begin{split}\label{eq conservation energy help}
\sum_{s,E}\sum_{\alpha,\gamma}\sum_{s'} & (\epsilon_s+E)
\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}
\frac{1}{V_{E+\epsilon_s-\epsilon_{s'}}}p_{s',E+\epsilon_s-\epsilon_{s'}}(t) \\
&= \sum_{s,E'}\sum_{\alpha,\gamma}\sum_{s'} (\epsilon_{s'}+E')
\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E'-\epsilon_s+\epsilon_{s'},E')\}
\frac{1}{V_{E'}}p_{s'E'}(t) \\
&= \sum_{s,E}\sum_{\alpha,\gamma}\sum_{s'} (\epsilon_s+E)
\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}
\frac{1}{V_{E}}p_{sE}(t),
\end{split}
\end{equation}
where we used the symmetry relation~(\ref{eq symmetries f}) and made use of the freedom to relabel indices
within the summation. Equation~(\ref{eq conservation energy help}) holds for any fixed time and even in
presence of driving when $\epsilon_s = \epsilon_s(\lambda_t)$. In this case one confirms that
\begin{equation}\label{eq conservation law}
\frac{d}{dt}[\Delta E_S(t) + \Delta E_B(t)] \equiv \frac{d}{dt}\sum_{s,E}[\epsilon_s(\lambda_t)+E] p_{sE}(t)
= \sum_s \dot\epsilon_s(\lambda_t) p_s(t) \equiv \dot W.
\end{equation}
This is the first law of thermodynamics in presence of driving.
Interestingly, for an undriven system Eq.~(\ref{eq conservation energy help}) even implies that
\begin{equation}
\frac{d}{dt}\lr{f(\epsilon_s+E)} \equiv \frac{d}{dt}\sum_{s,E} f(\epsilon_s+E) p_{sE}(t) = 0
\end{equation}
for an arbitrary function $f(E_\text{tot})$ of the total energy $E_\text{tot} = \epsilon_s + E$. We can call this
\emph{strict energy conservation}. It essentially implies that there is only one random variable in the problem
(and not the two $\epsilon_s$ and $E$) because the distribution for $E_\text{tot}$ remains fixed for all times.
This conclusion holds, however, only in absence of driving.
\subsubsection{Steady state of the Pauli MME}
\label{sec MME steady state}
We here consider the case where $\epsilon_s$ is held fixed in time (i.e., $\dot\lambda_t = 0$) and we ask for which
state $\bar p_{sE}$ the Pauli MME~(\ref{eq Pauli MME}) evaluates to zero. We call $\bar p_{sE}$ a steady state in this
case. One point we can immediately recognize from Eq.~(\ref{eq Pauli MME}) is that every state $\bar p_{sE}$ which fulfills
\begin{equation}
\frac{V_E}{V_{E+\epsilon_s-\epsilon_{s'}}} = \frac{\bar p(s,E)}{\bar p(s',E+\epsilon_s-\epsilon_{s'})}
\end{equation}
is a steady state. However, due to the fact that the energy is strictly conserved, there are infinitely many possible
steady states; indeed even infinitely many for every initial energy $\lr{\epsilon_s+E}(0) = E_0$ depending on how
the probabilities are initially distributed. For instance, one possible steady state is the overall Gibbs state
\begin{equation}\label{eq Gibbs SB}
\bar p_{sE} = \frac{e^{-\beta\epsilon_s}}{\C Z_S} \frac{V_E e^{-\beta E}}{\C Z_B} \equiv \pi_{sE},
\end{equation}
where $\beta$ must be fixed through $E_0 = -\partial_\beta\ln(\C Z_S\C Z_B)$.
On the other hand, imagine that we start with a definite initial condition such as
$p_{sE}(t=0) = \delta_{s,0}\delta_{E,E_0}$ and we assume that the energy eigenvalues are ordered according to
$\epsilon_n > \dots > \epsilon_1 > \epsilon_0 \equiv 0$. Then, the dynamics are restricted to the following states
with energies $(\epsilon_s,E)$:
\begin{equation}
(0,E_0), (\epsilon_1,E_0-\epsilon_1), \dots, (\epsilon_{n-1},E_0-\epsilon_{n-1}), (\epsilon_n,E_0-\epsilon_n).
\end{equation}
Note that the dynamics is not restricted to jumps between nearest neighbours as one might be tempted to
think.\footnote{The precise topology of the network depends strongly on the prefactor
$\Re\{S_\alpha^{ss'}S_\gamma^{s's}f_{\alpha\gamma}(E,E+\epsilon_s-\epsilon_{s'})\}$. }
The ratio of the rates to jump from one state to another are
\begin{equation}\label{eq ratio rates MME}
\frac{\text{rate}[(\epsilon_s,E_0-\epsilon_s)\rightarrow(\epsilon_{s'},E_0-\epsilon_{s'})]}{\text{rate}[(\epsilon_{s'},E_0-\epsilon_{s'})\rightarrow(\epsilon_s,E_0-\epsilon_s)]} = \frac{V_{E_0-\epsilon_{s'}}}{V_{E_0-\epsilon_s}},
\end{equation}
which can be interpreted as a purely entropic factor. Thus, the \emph{dynamics of the MME are entropy dominated}.
Typically, one expects that $V_{E'} > V_E$ if $E'>E$. Then, the system tends to prefer low energies in order to
increase the entropy of the environment. A particularly interesting case arises if the bath behaves like an ideal heat
bath. Using Boltzmann's entropy formula, we infer that $V_E = e^{S_B(E)/k_B}$ where $S_B(E)$ is the entropy of the bath
at energy $E$. Now, the assumption of an ideal heat bath enters by invoking the standard definition of temperature,
$T^{-1} = S'_B(E)$, which allows us to derive
$\frac{V_{E_0-\epsilon_{s'}}}{V_{E_0-\epsilon_s}} = e^{\beta(\epsilon_s-\epsilon_{s'})}$. This implies that the ratio
of the rates~(\ref{eq ratio rates MME}) fulfills the conventional local detailed balance relation. One steady state of
the MME with initial condition $p_{sE}(t=0) = \delta_{s,0}\delta_{E,E_0}$ is then given by
\begin{equation}\label{eq steady state microcanonical}
\bar p_{sE} =
\left\{\begin{array}{ll}
\frac{e^{-\beta\epsilon_s}}{Z_S} \frac{1}{n+1} & \text{if } E = E_0 - \epsilon_s \\
0 & \text{otherwise} \\
\end{array}\right.
\end{equation}
That is to say, the system equilbrated to the canonical ensemble with the temperature imposed by the initial energy of
the bath which becomes equally distributed over the available phase space. This is nothing else than the equivalence of
ensembles, i.e., the reduced state of a weakly coupled subsystem is a canonical distribution if the entire system has
a fixed energy $E_0$. Note that the above state indeed fulfills $\sum_{s,E} (\epsilon_s+E)\bar p_{sE} = E_0$.
\subsection{Entropy production and observational entropy}
\label{sec ent prod obs ent}
Conventionally, the entropy production for the Pauli master equation~(\ref{eq Pauli BMS}) can be expressed as
\begin{equation}\label{eq EP Pauli BMS}
\dot{\tilde\Sigma}(t) = \frac{d}{dt} S_\text{Sh}[p_s(t)] - \beta\dot Q(t)
= -\left.\frac{\partial}{\partial t}\right|_{\lambda_t}D[p_s(t)\|\pi_s(\lambda_t)] \ge 0,
\end{equation}
where $\pi_s(\lambda_t) = e^{-\beta\epsilon_s(\lambda_t)}/\C Z_S(\lambda_t)$ is the instantaneous canonical
equilibrium state of the system at time $t$. The derivative in Eq.~(\ref{eq EP Pauli BMS}) is evaluated with respect
to a fixed $\lambda_t$ and positivity of the entropy production follows from the two facts that the dynamics are
Markovian and that $\pi_s(\lambda_t)$ is an instantaneous steady state of the dynamics, see, e.g.,
Refs.~\cite{BreuerPetruccioneBook2002, KosloffEntropy2013, StrasbergEspositoPRE2019}. Finally,
$\dot Q(t) = \sum_s \epsilon_s(\lambda_t)\partial_t p_s(t)$ is the heat flow into the system.
Similarly, also for the MME we can derive an always positive entropy production rate by considering
\begin{equation}\label{eq EP Pauli MME general}
\dot\Sigma(t) \equiv -\left.\frac{\partial}{\partial t}\right|_{\lambda_t}D[p_{sE}(t)\|\bar p_{sE}(\lambda_t)] \ge 0,
\end{equation}
where $\bar p_{sE}(\lambda_t)$ is \emph{any} admissible steady state of the Pauli MME (independent of the initial
condition), which is allowed to depend parametrically on time through $\lambda_t$. As we have a multitude of possible
steady states, see Sec.~\ref{sec MME steady state}, there are many different possible choices, each leading to a
different positive `entropy production' rate. The choice, which leads to the desired final result, turns out be the
Gibbs state from Eq.~(\ref{eq Gibbs SB}), where $\epsilon_s = \epsilon_s(\lambda_t)$ is allowed to be time-dependent.
Indeed, upon integration of Eq.~(\ref{eq EP Pauli MME general}) we have
\begin{equation}
\begin{split}
\Sigma(t) &=
\int_0^t ds\left(-\frac{d}{ds} + \dot\lambda_s\frac{\partial}{\partial\lambda_s}\right) D[p_{sE}(s)\|\pi_{sE}(\lambda_s)] \\
&= -D[p_{sE}(t)\|\pi_{sE}(\lambda_t)] + D[p_{sE}(0)\|\pi_{sE}(\lambda_0)]
- \int_0^t ds\dot\lambda_s\frac{\partial}{\partial\lambda_s} \sum_{s,E} p_{sE}(s)\ln\pi_{sE}(\lambda_s),
\end{split}
\end{equation}
where we used the chain rule
$\frac{d}{dt} = \left.\frac{\partial}{\partial t}\right|_{\lambda_t} + \dot\lambda_t\frac{\partial}{\partial\lambda_t}$.
The particular form of $\pi_{sE}(\lambda_t)$ reveals after some straightforward manipulations that
\begin{equation}
\begin{split}
\Sigma(t) =& -\sum_{s,E} p_{sE}(t)\ln\frac{p_{sE}(t)}{V_E} + \sum_{s,E} p_{sE}(0)\ln\frac{p_{sE}(0)}{V_E} \\
&+ \beta\sum_{s,E} \left\{- [\epsilon_s(\lambda_t)+E] p_{sE}(t) + [\epsilon_s(\lambda_0)+E] p_{sE}(0) +
\int_0^t ds\epsilon_s(\lambda_s)p_{sE}(s)\right\} \\
=&~ \Delta S_\text{obs}.
\end{split}
\end{equation}
Here, for the final step we used the first law~(\ref{eq conservation law}). Thus, we have confirmed that
$\Sigma(t) = \Delta S_\text{obs} \ge 0$ also follows directly from the Pauli MME.
Finally, we compare $\Sigma(t)$ with $\tilde\Sigma(t)$ obtained by integrating Eq.~(\ref{eq EP Pauli BMS}).
First, we use that we can express the heat flow within the MME approach as
\begin{equation}
-\beta Q(t) = \beta \sum_E E [p_E(t)-p_E(0)]
= \sum_E [p_E(t)- p_E(0)]\ln V_E + \Delta S_\text{Sh}[p_E(t)] + D[p_E(t)\|\pi_E],
\end{equation}
if we assume that $p_E(0) = \pi_E$. If the initial state $p_{sE}(0) = p_s(0)p_E(0)$ is furthermore decorrelated,
we can confirm that
\begin{equation}
\Sigma(t) - \tilde\Sigma(t) = -I[p_{sE}(t)] - D[p_E(t)\|\pi_E] \le 0.
\end{equation}
This result is in direct analogy to our general finding~(\ref{eq comparison EPs 0}). Thus, we have re-derived all our
general findings within the particularly imporant limit of a system weakly coupled to a Markovian bath. Notice that
we could even derive a stronger statement in that limit, namely that the \emph{rate} of entropy
production~(\ref{eq EP Pauli MME general}) is positive. This is usually not the case within the general setup of
Sec.~\ref{sec general}. Conditions which ensure the positivity of the entropy production rate
$\frac{d}{dt}S_\text{obs}(t)$ are missing; within the standard approach reviewed in Sec.~\ref{sec previous results}
answers were partially found in Refs.~\cite{StrasbergEspositoPRE2019, StrasbergTBP}.
\section{Conclusions}
\label{sec conclusions}
We have put forward a novel approach to understand and quantify entropy production in open (quantum) systems driven
arbitrarily far from equilibrium. For this purpose we constructed a suitable notion of entropy for the entire universe
(system plus bath) based on the recently introduced observational entropy from
Refs.~\cite{SafranekDeutschAguirrePRA2019a, SafranekDeutschAguirrePRA2019b, SafranekDeutschAguirreArXiv2019}.
Then, using very similar steps as in previous approaches (conservation of global von Neumann entropy, special -- but
slightly more general -- form of the initial system-bath state, see Sec.~\ref{sec ideal}), we showed that the change
in observational entropy is always positive for arbitrary dynamics, has a clear information-theoretic interpretation,
and can be linked to the standard expression~(\ref{eq EP}) in the limit of weak system-bath coupling. Therefore, our
novel notion fulfills the three minimum requirements (i) to (iii), but -- moreover and most importantly -- it fulfills
them expressed as a change in a thermodynamically meaningful definition of global entropy. Thus, we were able to
microscopically derive the statement that \emph{entropy production measures the change in thermodynamic entropy of the
universe, which is always positive.}
Quite interestingly, in the conventionally considered weak coupling limit we showed that the quantitative difference
between our and the former approach is negligible. This reassures the consistency of both, our and the former approach.
Outside the weak coupling regime, interesting difference could appear. As we have also stated the condition which needs
to be fulfilled to test this theory experimentally [Eq.~(\ref{eq condition experiment})], we have the hope that it will
be possible to measure the global entropy production in the future, for instance, in cold
atoms~\cite{LewensteinSanperaAhufingerBook2012} or in electronic nanostructures coupled to mesoscopic heat
baths~\cite{PekolaNatPhys2015}.
Finally, we believe that our results provide strong evidence that observational entropy as advertised in
Refs.~\cite{SafranekDeutschAguirrePRA2019a, SafranekDeutschAguirrePRA2019b, SafranekDeutschAguirreArXiv2019} provides
a good candidate for thermodynamic entropy of isolated out-of-equilibrium systems.
\subsection*{Acknowledgments}
I am grateful to Andreu Riera-Campeny and Andreas Winter for thoughtful comments. Also various stimulating discussions
with Massimiliano Esposito about the nature of entropy production over the years are acknowledged. I am financially
supported by the DFG (project STR 1505/2-1). I also acknowledge funding from the Spanish MINECO FIS2016-80681-P
(AEI-FEDER, UE).
|
\section{Introduction}
Let $\Omega$ be a domain in $\mathbb C^n$ and $PSH(\Omega)$ the set of plurisubharmonic (psh) functions on $\Omega$. Recall that each $\phi\in PSH(\Omega)$ satisfies the following mean-value inequality:
$$
\phi(z)\le \frac1{|S|} \int_S \phi=:\phi_S
$$
whenever $S$ is a ball or a polydisc, with center $z$. Here $|S|$ denotes the Lebesgue measure of $S$ and $\int_S$ means the Lebesgue integral. The above inequality implies $\phi\in L^1_{\rm loc}(\Omega)$ and suggests to estimate the difference $|\phi-\phi_S|$. The concept of BMO functions then enters naturally. Let ${\mathcal S}={\mathcal S}(\Omega)$ be a family of relatively compact open subsets in $\Omega$.
We say that $\phi\in L^1_{\rm loc}(\Omega)$ has \emph{bounded mean oscillation} (BMO) with respect to ${\mathcal S}$ if
$$
{\sup}_{S\in {\mathcal S}}\, MO_{S}(\phi) <\infty,\ \ \ MO_{S}(\phi):=\frac1{|S|} \int_S |\phi-\phi_S|.
$$
Let $BMO(\Omega,{\mathcal S})$ denote the set of functions which are BMO with respect to ${\mathcal S}$. When ${\mathcal S}$ is the set of balls in $\Omega$, this is the original definition of BMO functions due to John-Nirenberg \cite{JN}. A classical example of BMO functions is $\log |z|$. It is also convenient to introduce local BMO functions as follows. For an open set $\Omega_0\subset\subset \Omega$ we define ${\mathcal S}|_{\Omega_0}$ to be the sets of all $S\in {\mathcal S}$ which are relatively compact in $\Omega_0$. Let $BMO_{\rm loc}(\Omega,{\mathcal S})$ be the set of functions on $\Omega$ which belong to $BMO(\Omega_0,{\mathcal S}|_{\Omega_0})$ for every open set $\Omega_0\subset\subset \Omega$.
By using pluripotential theory, Brudnyi \cite{Brudnyi} was able to show that each psh function is locally BMO with respect to balls (see also \cite{Brudnyi1} for stronger results concerning subharmonic functions in the plane). Recently, the first author found another approach to local BMO properties of psh functions by using the Riesz decomposition theorem and some basic facts of psh functions (cf. \cite{Chen}). Benelkourchi et al. \cite{BJZ} showed that every function in the Lelong class ${\mathcal L}$ is globally BMO with respect to balls. Recall that
$$
{\mathcal L}=\left\{u\in PSH(\mathbb C^n): {\limsup}_{|z|\rightarrow \infty}\, (u(z)-\log |z|)<\infty\right\}.
$$
In this paper we propose a new and simpler approach based on the following basic observation:
\smallskip
\emph{It is easier to look at the upper oscillation instead of the mean oscillation for psh functions.}
\smallskip
To define the \emph{upper oscillation} one simply uses $\sup_S \phi$ instead of $\phi_S$:
\begin{equation}\label{eq:2}
UO_{S}(\phi):=\frac1{|S|} \int_S \left|\phi-{\sup}_S\, \phi\right|={\sup}_S\, \phi-\phi_S.
\end{equation}
Note that $-UO_S(-\phi)$ is exactly the lower oscillation introduced by Coiffman-Rochberg (cf. \cite{CR}, see also \cite{Ose13} for further properties). Since
\begin{equation}\label{eq:3}
MO_{S}(\phi)=\frac1{|S|} \int_S |\phi-\phi_S|=\frac2{|S|} \int_{\phi<\phi_S} (\phi_S-\phi) \leq 2\, UO_S(\phi),
\end{equation}
we see that bounded upper oscillation (BUO) implies BMO. One may define $BUO(\Omega,{\mathcal S})$ and $BUO_{\rm loc}(\Omega,{\mathcal S})$ analogously as the case of BMO.
Let ${\mathcal P}={\mathcal P}(\Omega)$ denote the set of relatively compact polydiscs in $\Omega$ and ${\mathcal P}_N$ the set of polydiscs $P\subset\subset \Omega$ of finite type $N$, i.e.,
$$
\max\{r_j\} \leq \min\{r_j^{1/N}\},
$$
where $N>0$ and $\{r_j\}_{1\le j\le n}$ is the\/ {\it polyradius}\/ of $P$.
Based on Harnack's inequality and convex analysis, we are able to show the following
\begin{theorem}\label{th:Main1}
\begin{enumerate}
\item $PSH(\Omega)\subset BUO_{\rm loc}(\Omega,{\mathcal P}_N) \subset BMO_{\rm loc}(\Omega,{\mathcal P}_N)$;
\item $PSH(\mathbb D^n)\nsubseteq BMO_{\rm loc}(\mathbb D^n,{\mathcal P})$ for $n\ge 2$, where $\mathbb D^n$ is the unit polydisc;
\item ${\mathcal L}\subset BUO(\mathbb C^n, \mathcal P)$; more precisely, for every $\phi \in PSH(\mathbb C^n)$ with
$$
\phi(z_1,\cdots, z_n) \leq c+\max_{1\leq j\leq n} \log (1+|z_j|), \ \ \forall \ (z_1, \cdots, z_n)\in\mathbb C^n,
$$
where $c$ is a constant, we have
$
UO_P(\phi) < 3^n
$
for all polydiscs $P$ in $\mathbb C^n$.
\end{enumerate}
\end{theorem}
For $({\rm deg\,}p)^{-1}\log |p|\in \mathcal L$ where $p$ is a complex polynomial, we even obtain a dimension-free BUO estimate with respect to all compact convex sets.
\begin{theorem}\label{th:remez1} For every non-empty compact convex set $A$ in $\mathbb C^n$, we have
$$
UO_{A} (\log |p|) \leq \gamma \cdot \deg p,
$$
for all $p\in\mathbb C[z_1, \cdots, z_n]$. Here the constant $\gamma\in (1,2)$ is determined by
$$
\gamma+\log(\gamma-1)=0.
$$
\end{theorem}
\textbf{Remark}: (i) The above estimate is sharp, in fact, there exists a line segment $A$ in $\mathbb C$ such that
$$UO_A(\log|z|)=\gamma.$$
(ii) In particular, if $A$ is a compact convex set in $\mathbb R^n \subset \mathbb C^n$ and all coefficients of $p$ are real, then we have
$$
UO_A (\log |p|) \leq \gamma \cdot \deg p <2\deg p,
$$
which is closely related the classical Remez inequality for real polynomials. Theorem \ref{th:remez1} also suggests to study the Remez inequality for \emph{complex} polynomials (see \cite{BG} and \cite{BJZ} for related results).
(iii) Notice that $
1.278<\gamma<1.279.
$
By $(\ref{eq:3})$ we have
$$
MO_A (\log |p|) \leq 2\gamma \cdot \deg p< 2.558\cdot \deg p.
$$
Such dimension-free estimate (with a slightly better constant $2+\log 2\approx 2.301$) was first obtained by Nazarov et al. \cite{Nazarov}. Our proof of Theorem \ref{th:remez1} is elementary, however.
\medskip
For $\phi\in PSH(\Omega)$ we define the (weighted) Bergman kernel by
$$
K_{\phi,\Omega}(z)=\sup\left\{|f(z)|^2: f\in {\mathcal O}(\Omega),\ \int_\Omega |f|^2 e^{-\phi}\le 1\right\}.
$$
For a vector $a=(a_1,\cdots,a_n)$ with all $a_j>0$ we set
$$
P_{r^a}:=\{z\in \mathbb C^n: |z_j|\leq r^{a_j},\,1\le j\le n\}.
$$
It was shown in \cite{Chen} that if $\phi$ is psh on the closure of the unit ball $\mathbb B^n$ and $a_0=(1,1/2,\cdots,1/2)$ then
$$
\lim_{r\rightarrow 0+} \frac{\log K_{\varepsilon\phi,\mathbb B^n}(1-r,0,\cdots,0)}{\log 1/r}=n+1-\varepsilon \cdot \lim_{r\rightarrow 0+} \frac{\sup_{z\in P_{r^{a_0}}} \phi (1+z) }{\log r}
$$
provided $\varepsilon\ll 1$, where $1+z=(1+z_1,z_2,\cdots,z_n)$. The limit in RHS of the above inequality is called the $a_0-$directional Lelong number of $\phi$ at $(1,0,\cdots,0)$ (see \cite{K94}).
Here we will present an analogous but independent result, as an application of Theorem \ref{th:Main1}.
For $\phi\in PSH({\mathbb D}^n)$ and $t\in \mathbb D^n$ we define
$$
\phi^t(z):=\phi(tz), \ \ tz:=(t_1z_1, \cdots, t_n z_n).
$$
A fundamental result of Berndtsson \cite{Bern06} implies that
$$
F(\phi): \, (t,z) \mapsto \log K_{\phi^t, \, \mathbb D^n}(z)
$$
is psh on $\mathbb D^n \times \mathbb D^n$.
\begin{theorem}\label{th:Main2}
For each $a=(a_1,\cdots,a_n)$ with all $a_j>0$, there exists a number $\varepsilon_0=\varepsilon_0(a,\phi,\Omega)$ such that
\begin{equation}\label{eq:5}
\lim_{r\to 0+} \frac{\sup_{t\in P_{r^a}} F(\varepsilon \phi)(t,0)/\varepsilon }{\log r}=\lim_{r\to 0+} \frac{\sup_{z\in P_{r^a}} \phi(z) }{\log r}
\end{equation}
holds for all $\varepsilon\le \varepsilon_0$.
\end{theorem}
Although Theorem \ref{th:Main2} makes sense only when $\phi$ is singular at the origin, it is of independent interest to study the relation between $F(\phi)$ and $\phi$ for smooth $\phi$.
\begin{theorem}\label{th:Main3}
Let $\phi$ be a smooth psh function on $\mathbb D^n$. Then
\begin{equation}\label{eq:6}
\lim_{t\to 0} \frac{\partial^2 F(\phi)}{\partial t_j \partial\bar t_k} (t,0)=
\begin{cases}
\frac12 \cdot \frac{\partial^2 \phi}{\partial z_j \partial\bar z_j} (0), \
\text{if}\ \, j=k; \\
0, \
\text{if}\ \, j\neq k.
\end{cases}
\end{equation}
In particular $F(\phi)(t,0)$ is strictly psh at $t=0$ if $\phi$ is strictly psh at $z=0$.
\end{theorem}
\textbf{Remark}: Since $F(\phi)(t,0)$ depends only $(|t_1|,\cdots, |t_n|)$, it follows from the psh property of $F(\phi)$ that
\begin{equation}\label{eq:bo}
\log\frac{e^{\phi(0)}}{\pi^n}=F(\phi)(0,0)\leq F(\phi)(t,0)=\log K_{\phi^t, \,\mathbb D^n}(0).
\end{equation}
Letting $t$ tend to $(1,\cdots, 1)$, we obtain the sharp Ohsawa--Takegoshi estimate (cf. \cite{Blocki}; see also \cite{GZ,BL}):
\begin{equation}\label{eq:7}
K_{\phi,\, \mathbb D^n}(0) \geq \frac{e^{\phi(0)}}{\pi^n}.
\end{equation}
Theorem \ref{th:Main3} suggests that one should have a better lower bound for $K_{\phi, \,\mathbb D^n}$ in case $\phi$ is \/ {\it strictly}\/ psh.
\section{An enlightening example}
To explain why BUO is easier than BMO, we will show that the upper oscillation of $\log|z|$ with respect to discs is computable. Recall that
$$
UO_{B} (\log |z|):={\sup}_B \log|z|-(\log|z|)_B
$$
for every disc $B$ in $\mathbb C$.
\begin{lemma}\label{lm:integral}
Fix $\hat{z}\in \mathbb C$ and set
$$
I(c):=\frac1{2\pi} \int_{0}^{2\pi} \log|\hat z+c e^{i\theta}| \, d\theta, \ \ c>0.
$$
Then we have
$$
I(c)=
\begin{cases}
\begin{array}{ll}
\log|\hat z| & \text{if}\ \ c\leq |\hat z| \\
\log c & \text{if}\ \ c> |\hat z|.
\end{array}
\end{cases}
$$
\end{lemma}
\begin{proof}
If $c\leq |\hat z|$ then $\log|z|$ is harmonic in the disc $\{z:|z-\hat z| < c\}$, so that $I(c)=\log|\hat z|$, in view of the mean-value equality. For $c>|\hat z|$ we may write
$$
I(c)=\frac1{2\pi}\int_{0}^{2\pi} \log|\hat z e^{i\theta}+c|\, d\theta.
$$
As $\log|z|$ is harmonic in $\{z:|z-c|< |\hat z|\}$, we get $I(c)=\log c$.
\end{proof}
\begin{proposition}
For any disc $B$ we have
$$UO_{B} (\log |z|) \leq \log\frac{\sqrt 5+1}{2} +\frac{\sqrt 5-1}{4}.$$
Moreover, the bound is sharp.
\end{proposition}
\begin{proof}
Suppose $B=\{z:|z-\hat z|<b\}$. By Lemma \ref{lm:integral} we have
$$
(\log|z|)_{B}= \log|\hat z|, \ \ \ \ \text{if}\ \ b\leq |\hat z|,
$$
and if $b>|\hat z| $ then
\begin{eqnarray*}
(\log|z|)_{B} & = & \frac{1}{\pi b^2}\int_{0}^b 2\pi c \cdot I(c)\, dc\\
& = & \log b-\frac{1}{2} \left(1-\frac{|\hat z|^2}{b^2}\right).
\end{eqnarray*}
It follows that
$$
UO_{B} (\log |z|) =
\begin{cases}
\begin{array}{ll}
\log(b+|\hat z|)-\log |\hat z| & \text{if} \ \ b\leq |\hat z|\\
\log(b+|\hat z|)-\log b+\frac{1}{2} \left(1-\frac{|\hat z|^2}{b^2}\right) & \text{if}\ \ b>|\hat z| .
\end{array}
\end{cases}
$$
If $b\leq |\hat z|$ then
$$
UO_{B} (\log |z|) =\log\left(\frac b{|\hat z|}+1\right)\leq \log 2.
$$
For $b> |\hat z|$ we set $x=|\hat z|/b$ and write $UO_{B} (\log |z|)$ as
$$
f(x)=\log(1+x)+\frac12 \cdot (1-x^2), \ \ 0<x<1.
$$
Since
$$
f'(x)=\frac1{1+x}-x,
$$
we see that $f$ is increasing on $[0, \hat x]$ and decreasing on $[\hat x, 1]$, where $
\hat x=\frac{\sqrt 5-1}{2}.
$
Notice that
$$
f(\hat x)=\log\frac{\sqrt 5+1}{2} +\frac{\sqrt 5-1}{4}.
$$
Thus $$UO_{B} (\log |z|) \leq \log\frac{\sqrt 5+1}{2} +\frac{\sqrt 5-1}{4}$$
and the equality holds if and only if
$$
\frac{|\hat z|}{b}=\frac{\sqrt 5-1}{2}.
$$
This finishes the proof.
\end{proof}
\section{Proof of Theorem \ref{th:Main1}}
\subsection{One dimensional case}
Let $\Omega$ be a domain in $\mathbb C$ and $\phi$ a subharmonic function on $\Omega$. Recall that
$$
UO_B(\phi)={\sup}_B\, \phi-\phi_B
$$
where
$
B=\{z:|z-\hat z| <r\}\subset \Omega.
$
The idea is to use \emph{Harnack's inequality} and a convexity lemma. Let us write
$$
UO_B(\phi)=I_1+I_2,
$$
where
$$
I_1={\sup}_{B} \,\phi-\phi_{\partial B}, \ \ I_2:=\phi_{\partial B}-\phi_{B},
$$
with $\phi_{\partial B}$ being the mean-value of $\phi$ over the boundary $\partial B$. For each $\tau>0$ we set
$$
\tau B=\{z:|z-\hat z|<\tau r\}.
$$
Applying Harnack's inequality to the nonpositive subharmonic function $\psi:=\phi-\sup_{B} \phi$, we get
$$
{\sup}_{\frac12 B}\, \psi = {\sup}_{\partial (\frac12 B)}\, \psi \leq \frac13 \cdot \psi_{\partial\, B},
$$
i.e.,
$$
I_1 \leq 3\left( {\sup}_{B}\, \phi-{\sup}_{\frac12 B}\, \phi\right).
$$
Here the constant $1/3$ comes from the Poisson kernel of the unit disc since
$$
\inf_{|z|=1/2} \frac{1-|z|^2}{|1-z|^2}=\frac13.
$$
The following fact explains why we need such an estimate.
\medskip
\textbf{Fact 1}:
$
J_1:=\sup_{B} \phi- \sup_{\frac12 B} \phi
$ is continuous in $\hat z$ and $r$ respectively; moreover, it is increasing with respect to $r$.
\begin{proof} Since $\sup_{B} \phi$ is a convex function of $\log r$ (see \cite{Demailly}, Corollary 5.14), it follows that $J_1$ is a continuous increasing function of $r$. The continuity of $J_1$ in $\hat{z}$ is obvious.
\end{proof}
Let $\Omega_0$ be a relatively compact open subset in $\Omega$. Let $\delta_0$ denote the distance between $\overline{\Omega_0}$ and $\partial \Omega$. By the above fact we see that if the radius $r$ of $B\subset \Omega_0$ is less than $\delta_0/2$ then
$$
I_1\le 3\, {\sup}_{\hat{z}\in \Omega_0}\, J_1(\hat{z},\delta_0/2)<\infty,
$$
and if $r\ge \delta_0/2$ then
$$
I_1 \le 3\,{\sup}_{\Omega_0} \phi -3\, {\inf}_{\hat{z}\in \Omega_0}\, {\sup}_{\{|z-\hat{z}|<\delta_0/4\}}\,\phi<\infty.
$$
To estimate $I_2$, we need the following convexity lemma which was communicated to the second author by Bo Berndtsson:
\begin{lemma}\label{lm:convexity}
Let $d\mu$ be a probability measure on a Borel measurable subset $S$ in $\mathbb R^n$ with barycenter $\hat t \in \mathbb R^n$. Let $f$ be a convex function on $\mathbb R^n$. Then
$$
\int_S f \, d\mu \geq f(\hat t).
$$
\end{lemma}
\begin{proof} Since $f$ is convex, there exists an affine function
$l$ such that $f(\hat t)=l(\hat t)$ and
$
f\geq l$ on $\mathbb R^n$,
which implies
$$
\int_S f\, d\mu \geq \int_S l\, d\mu = l(\hat t)=f(\hat t),
$$
where the first equality follows from the definition of barycenter.
\end{proof}
With
$
f(t):=\phi_{\{z:|z-\hat z|=e^t r\}}
$
we have
\begin{eqnarray*}
I_2 & = & f(0)-\frac1{\pi r^2}\int_{-\infty}^0 2\pi e^{t} r\cdot f(t) \, d(e^t r)\\
& = & f(0)-\int_{-\infty}^0 f(t) \, d (e^{2t}).
\end{eqnarray*}
Since $f(t)$ is convex and $d (e^{2t})$ is a probability measure on $(-\infty, 0)$ with barycenter at $t=-1/2$, it follows from Lemma \ref{lm:convexity} that
\begin{equation}\label{eq:2.2}
\int_{-\infty}^0 f(t) \, d (e^{2t}) \geq f(-1/2),
\end{equation}
which implies
$$
I_2 \leq J_2:= f(0)-f(-1/2).
$$
Since $f$ is convex, we get an analogous conclusion as {\bf Fact 1}:
\medskip
\textbf{Fact 2}:
$
J_2
$ is continuous in $\hat z$ and $r$ respectively; moreover, it is increasing with respect to $r$.
\medskip
By a similar argument as above, we may verify that
$$
{\sup}_{B\subset \Omega_0}\, I_2 <\infty.
$$
\subsection{High dimensional case}
The following result plays the role of {\bf Fact 1,2}.
\begin{lemma}\label{le:convex2}
Let $g(t)=g(t_1, \cdots, t_n)$ be a convex function on $(-\infty, 2)^n$ which is increasing in each variable. Then
$$
{\sup}_{t\in A_N} \left[g(t)-g(t-1)\right] \leq nN \left[g(1,\cdots, 1)-g(0)\right],
$$
where $t-1:=(t_1-1, \cdots, t_n-1)$, $N\geq 1$ and
$$
A_N:=\left\{t\in (-\infty, 0]^n: \max\{-t_j\} \leq N\min\{-t_j\}\right\}.
$$
\end{lemma}
\begin{proof}
A standard regularization process reduces to the case when $\phi$ is smooth. Set
$$
f(a)=g(t_1+a,\cdots, t_n+a):=g(t+a).
$$
We have
$$
f(0)-f(-1)=\int_{-1}^0 f'(a)\, da=\int_{-1}^0 \sum g_j(t+a)\, da
$$
where $g_j:=\frac{\partial g}{\partial t_j}$.
Notice that
$$
\sum g_j(t+a) \leq \frac{1}{\min\{-t_j-a\}} \sum (-t_j-a)g_j(t+a)
$$
and
$$
\sum (-t_j-a)g_j(-s(t+a)) =\frac{dg(-s(t+a))}{d s}
$$
is an increasing function of $s\in (-\infty, 0)$ by convexity of $g$. Thus we have
$$
\sum (-t_j-a)g_j(t+a) \leq \sum (-t_j-a)g_j(0) \leq \max\{-t_j-a\} \sum g_j(0),
$$
which implies
$$
\sum g_j(t+a) \leq \frac{\max\{-t_j-a\}}{\min\{-t_j-a\}} \sum g_j(0).
$$
For any $t\in A_N$, we have $t+a \in A_N$ (since $a\leq 0$), so that
$$
\frac{\max\{-t_j-a\}}{\min\{-t_j-a\}} \leq N.
$$
Thus
$$
g(t)-g(t-1) \leq N \sum g_j(0).
$$
Since $g$ is convex and increasing, we have
$$
g_j(0) \leq g(1,\cdots, 1)-g(0),
$$
which finishes the proof.
\end{proof}
Let $$
P:=\left\{z\in \mathbb C^n: |z_j-\hat z_j|<r_j, \ 1\leq j\leq n\right\} \subset \Omega
$$ be a polydisc of type $N$, i.e.,
$$
\max\{r_j\} \leq \min\{r_j^{1/N}\}.
$$
Similar as above, we write
$$
UO_{P} (\phi)={\sup}_P\, \phi-\phi_P =I_1+I_2,
$$
where
$$
I_1:={\sup}_P\, \phi-\phi_{\partial P}, \ \ I_2:=\phi_{\partial P}-\phi_{P},
$$
and
$$
\partial P:=\{z\in \mathbb C^n: |z_j-\hat z_j|=r_j, \ 1\leq j\leq n\}
$$
is the \emph{Shilov boundary} of $P$. Applying Harnack's inequality (see \cite{K94}, p.\,186) $n$-times, we get the following
\begin{lemma}\label{le3.3} $I_1\leq 3^n J_1$, where $J_1:=\sup_{P} \phi-\sup_{\frac12 P} \phi$.
\end{lemma}
Using \eqref{eq:2.2} repeatedly we get
\begin{lemma}\label{le3.4} $I_2\leq J_2$, where $J_2:=f(0)-f(-1/2, \cdots, -1/2)$ with
$$
f(t):=\phi_{\{z\, :\, |z_j-\hat z_j|=e^{t_j} r_j,\ 1\le j\le n\}}.
$$
\end{lemma}
Since both $\sup_{P} \phi$ and $\phi_{\partial P}$ are continuous in $\hat z_j$ and convex increasing with respect to $\log r_j$ for all $j$, it follows from Lemma \ref{le:convex2} (through a similar argument as the one-dimensional case) that
$$
{\sup}_{P\in {\mathcal P}_N|_{\Omega_0}
} \left(J_1+J_2\right) <\infty,
$$
for every open set $\Omega_0\subset\subset\Omega$, which finishes the proof of the first part of Theorem \ref{th:Main1}.
\subsection{A counterexample} For the second part of Theorem \ref{th:Main1}, we need to construct a counterexample. For the sake of simplicity, we only consider the case $n=2$. It suffices to verify the following
\begin{theorem}
\label{th:example}
Set
$
\phi(z, w):=-\sqrt{(\log|z|+\log|w|)\log|w|},
$
$z,w\in\mathbb D$. Then we have $\phi\in PSH(\mathbb D^2)$, while
$$
\sup_{0<r_1, r_2<1} \frac{1}{|\mathbb D^2_r|}\int_{\mathbb D^2_r} |\phi-\phi_{\mathbb D^2_r}| =\infty,
$$
where
$$
\mathbb D^2_r:=\left\{(z,w)\in\mathbb C^2: |z|<r_1, \, |w|<r_2 \right\}.
$$
\end{theorem}
The following lemma shows that {\bf Fact 1, 2} is no more true for general bidiscs.
\begin{lemma}
\label{le:2.5}
$f(x,y):=-\sqrt{(x+y)y}$ is convex on $ (-\infty, 0)^2$ and increasing in each variable; moreover,
\begin{equation}\label{eq:2.4}
{\sup}_{\{x, y\leq -1\}} \left[f(x,y)-f(x-1, y-1)\right]=\infty.
\end{equation}
\end{lemma}
\begin{proof}
The first conclusion follows by a straightforward calculation. For \eqref{eq:2.4} it suffices to note that
$$
f(x,-1)-f(x-1, -2)=\frac{5-x}{\sqrt{6-2x}+\sqrt{1-x}} \to \infty
$$
as $x\to-\infty$. The proof is complete.
\end{proof}
Let us first verify that $\phi \notin BUO_{\rm loc}(\mathbb D^2, \mathcal P)$.
\begin{lemma}\label{lm:NonBUO}
$\sup_{0<r_1, r_2<1} \sup_{\mathbb D^2_r}(\phi- \phi_{\mathbb D^2_r})=\infty$.
\end{lemma}
\begin{proof} With $x=\log r_1$ and $y=\log r_2$, we get
$$
{\sup}_{\mathbb D^2_r} \left(\phi- \phi_{\mathbb D^2_r}\right)= f(x,y)-\int_{-\infty}^0 \int_{-\infty}^0 f(x+t, y+s) \, de^{2t} d e^{2s}=: I(x,y).
$$
Integrate by parts with respect to $t$ and $s$ successively, we may write
$$
I(x,y)=I_1 +I_2,
$$
where
$$
I_1=\int_{-\infty}^0 \frac{x+2y+2s}{-4f(x+t, y+s)} \, de^{2s}
$$
and
$$
I_2=\int_{(-\infty, 0)^2} \frac{y+s}{-4f(x+t, y+s)} \, de^{2t} d e^{2s}.
$$
Obviously, $I_2(x,-1)$ is bounded on $(-\infty, 0]$, but $I_1(x,-1)\to \infty$ as $x\to -\infty$, from which the assertion immediately follows.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:example}] By Lemma \ref{lm:convexity} we have (still with $x=\log r_1,\,y=\log r_2$)
\begin{eqnarray*}
\phi_{\mathbb D^2_r} & = & \int_{-\infty}^0 \int_{-\infty}^0 f(x+t, y+s) \, de^{2t} d e^{2s} \\
& \geq & f(x-1/2, y-1/2)= {\sup}_{\mathbb D^2_{e^{-1/2} r}}\,\phi,
\end{eqnarray*}
which yields
\begin{eqnarray*}
\frac{1}{|\mathbb D^2_r|}\int_{\mathbb D^2_r} |\phi-\phi_{\mathbb D^2_r}|
& \geq &
\frac{1}{|\mathbb D^2_r|}\int_{\mathbb D^2_{e^{-1/2} r}} \left({\sup}_{\mathbb D^2_{e^{-1/2} r}}\,\phi - \phi\right)\\
& = & e^{-2} \left( { \sup}_{\mathbb D^2_{e^{-1/2} r}}\,\phi- \phi_{\mathbb D^2_{e^{-1/2} r}} \right).
\end{eqnarray*}
By a similar argument as Lemma \ref{lm:NonBUO}, we conclude the proof of Theorem \ref{th:example}.
\end{proof}
\subsection{Lelong class}
In this section we shall prove the third part of Theorem \ref{th:Main1}. The key ingredient is the following counterpart of Lemma \ref{le:convex2}.
\begin{lemma}\label{le:convex3}
Let $g(t)=g(t_1, \cdots, t_n)$ be a convex function on $\mathbb R^n$ which is increasing in each variable. Assume that
$$
g(t) \leq \max_{1\leq j\leq n}\{\log(1+e^{t_j})\}, \ \ \forall \ t\in\mathbb R^n.
$$
Then for every $M>0$ we have
$$
{\sup}_{t\in \mathbb R^n} \left[g(t)-g(t-M)\right] \leq M,
$$
where $t-M:=(t_1-M, \cdots, t_n-M)$.
\end{lemma}
\begin{proof} For fixed $t$, we consider the following convex increasing function
$$
f(s):=g(t_1+s, \cdots, t_n+s)
$$
on $\mathbb R$. Convexity of $f$ gives
$$
\frac{f(0)-f(-M)}M \leq \lim_{s\to \infty} \frac{f(s)-f(0)}{s}.
$$
By the assumption, we have
$$
f(0) \leq f(s) \leq \max_{1\leq j\leq n}\{\log(1+e^{t_j} e^s)\}
$$
for every $s\geq 0$, so that
$$
\lim_{s\to \infty} \frac{f(s)-f(0)}{s}\leq 1.
$$
The proof is complete.
\end{proof}
\begin{proof}[Proof of the third part of Theorem \ref{th:Main1}]
Again for any polydisc
$$
P:=\left\{z\in \mathbb C^n: |z_j-\hat z_j|<r_j, \ 1\leq j\leq n\right\},
$$
we may write
$$
UO_{P} (\phi)={\sup}_P\, \phi-\phi_P =I_1+I_2,
$$
where
$$
I_1:={\sup}_P\, \phi-\phi_{\partial P}, \ \ I_2:=\phi_{\partial P}-\phi_{P}.
$$
By Lemma \ref{le3.3} we have
$$
I_1\leq 3^n \left({\sup}_P\, \phi-{\sup}_{\frac12 P}\, \phi\right).
$$
Put
$$
P_t:=\left\{z\in \mathbb C^n: |z_j-\hat z_j|<e^{t_j}r_j, \ 1\leq j\leq n\right\}
$$
and $f_1(t):=\sup_{P_t} \phi$.
Since $\phi\in {\mathcal L}$, we know that for some constant $c_1\gg 1$ the function $f_1-c_1$ satisfies the assumption in Lemma \ref{le:convex3}, so that
$$
{\sup}_{P}\, \phi-{\sup}_{\frac12 P}\,\phi=f_1(0)-f_1(-\log 2) \leq \log 2,
$$
which in turn implies
$$
I_1\leq 3^n \log 2.
$$
Moreover, we infer from Lemma \ref{le3.4} that
$$
I_2\leq f(0)-f(-1/2, \cdots, -1/2), \ \ f(t):=\phi_{\partial P_t}.
$$
Applying Lemma \ref{le:convex3} in a similar way as above, we have
$$
I_2 \leq 1/2.
$$
Thus
$$
UO_{P} (\phi) \leq 3^n \log 2+ 1/2 < 3^n,
$$
which finishes the proof.
\end{proof}
\section{Proof of Theorem \ref{th:remez1}}
The starting point is the following
\begin{definition}[$\gamma$-constant] We shall define the constant $\gamma$ as the BUO norm of\/ $\log |z|$ on $\mathbb C$ with respect to all line segments. More precisely,
$$
\gamma:=\sup_{a\neq b\in \mathbb C} UO_{[a,b]} (\log |z|),
$$
where $[a,b]$ denotes the line segment connecting $a$ and $b$, and the upper oscillation is defined by
$$
UO_{[a,b]} (\log |z|) :=\left({\sup}_{0\leq t\leq 1}\,\log |a(1-t)+bt| \right)- \int_{0}^1 \log |a(1-t)+bt| \, dt.
$$
\end{definition}
The key step is to show the following
\begin{lemma}\label{lm:gamma}
$1<\gamma<2$ is determined by
\begin{equation}\label{eq:gamma}
\gamma+\log(\gamma-1)=0.
\end{equation}
\end{lemma}
\begin{proof}
For each pair $a, b \in \mathbb C$, we shall compute
$$
UO_{[a,b]} (\log |z|)={\sup}_{[a,b]}\, \log |z|-(\log |z|)_{[a,b]}.
$$
Since $\log |z|$ is $S^1$-invariant, by a rotation of $z$, we may assume that
$$
b \in \mathbb R,\ \ \ b>|a|.
$$
Thus
$$
{\sup}_{[a,b]}\, \log |z|=\log b
$$
is independent of $a$.
Since
\begin{eqnarray*}
(\log |z|)_{[a,b]} & = & \int_{0}^1 \log |a(1-t)+bt| \, dt\\
& \ge & \int_{0}^1 \log |{\rm Re\,}a\cdot(1-t)+bt| \, dt\\
& = & (\log |z|)_{[{\rm Re\,}a,b]}
\end{eqnarray*}
with equality holds if and only if $a\in \mathbb R$. Thus it suffices to verify \eqref{eq:gamma} for
$$
a, b\in \mathbb R, \ \ \ |a|<b.
$$
Consider $\log|z|-\log b$ instead of $\log|z|$, one may further assume that
$$
b=1, \ \ \ -1<a<1,
$$
which implies
$$
UO_{[a,1]} (\log|z|)= \log 1- (\log|z|)_{[a, 1]}=-(\log|z|)_{[a, 1]}.
$$
We divide into two case.
(i) $0\leq a<1$. Then we have
$$
-(\log|z|)_{[a, 1]}=\frac{-1}{1-a} \int_{a}^1 \log x \, dx=\frac{a\log a}{1-a}+1 \leq 1.
$$
(ii) $-1<a<0$. Then we have
$$
-(\log|z|)_{[a, 1]}=\frac{-1}{1-a} \int_{a}^1 \log |x| \, dx=\frac{a\log (-a)}{1-a}+1 >1.
$$
Thus
$$
\gamma=\sup_{-1<a<0} \frac{a\log (-a)}{1-a}+1.
$$
It suffices to verify that $\gamma$ satisfies $(\ref{eq:gamma})$.
To see this, put
$$
t^{-1}:=1-a \in (1,2)
$$
and write
$$
\frac{a\log (-a)}{1-a}= (1-t)\log t-(1-t)\log (1-t)=:f(t).
$$
Since
$$
f'(t)=t^{-1}-\log t +\log(1-t),
$$
it follows that $f'(t)=0$ if and only if
$$
t^{-1}=\log \frac{1}{t^{-1}-1},
$$
i.e.,
$$
1-a+\log(-a)=0.
$$
Thus we have
$$
\gamma=\sup_{-1<a<0} \frac{a\log (-a)}{1-a}+1= \frac{a_0\log (-a_0)}{1-a_0}+1,
$$
where $a_0$ is determined by
\begin{equation}\label{eq:a_0}
1-a_0+\log(-a_0)=0,
\end{equation}
which gives
$$
\gamma=1-a_0 \in (1,2).
$$
It is clear that $(\ref{eq:a_0})$ is equivalent to $(\ref{eq:gamma})$.
\end{proof}
Since a translation of a line segment is still a line segment, we know that $\log|z-z_0|$ and $\log|z|$ have the same line segment BUO norm. This fact can be used to estimate the line segment BUO norm of $\log|p|$ for general polynomials $p$. In fact, if we write
$$
p=a_0(z-a_1)^{n_1} \cdots (z-a_k)^{n_k},
$$
then
$$
{\sup}_{[a,b]}\, \log|p| \leq \log |a_0|+ \sum_{j=1}^k n_j {\sup}_{[a,b]}\, \log|z-a_j|
$$
and
$$
(\log|p|)_{[a,b]}= \log |a_0|+ \sum_{j=1}^k n_j (\log|z-a_j|)_{[a,b]}.
$$
Thus
\begin{eqnarray*}
UO_{[a,b]} (\log|p|) & := & {\sup}_{[a,b]}\, \log|p|-(\log|p|)_{[a,b]}\\
& \leq & \sum_{j=1}^k n_j UO_{[a,b]}(\log|z-a_j|).
\end{eqnarray*}
This combined with the fact $UO_{[a,b]}(\log|z-a_j|) \leq \gamma$ gives
\begin{equation}\label{eq:ss}
UO_{[a,b]} (\log |p|) \leq \gamma \cdot \deg p.
\end{equation}
for all polynomials $p$ and all $a,b\in\mathbb C$.
Now we may conclude the proof of Theorem \ref{th:remez1} as follows.
Since $A$ is compact, we may choose $z_0\in A$ such that
$$
|p(z_0)|={\sup}_{z\in A}\, |p(z)|.
$$
For every ray (half line), say $L$, starting from $z_0$, we see that $A\cap L$ is a line segment in view of convexity of $A$. Let $L_{\mathbb C}$ be the complex line containing $L$. Apply \eqref{eq:ss} to $p|_{L_{\mathbb C}}$, we have
$$
UO_{A\cap L} (\log |p|)= UO_{A\cap L} (\log |p|_{L_{\mathbb C}}|) \leq \gamma\deg p|_{L_{\mathbb C}}\leq \gamma \deg p,
$$
which gives
$$
UO_{A} (\log |p|) \leq \gamma \deg p
$$
since $UO_{A} (\log |p|) $ is a certain average of $UO_{A\cap L} (\log |p|)$ for all $L$ starting from $z_0$: in fact, since $z_0$ is a maximum point of $\log|p|$ on $A$ and $L$ contains $z_0$, we always have
$$
\sup_{A\cap L} \log|p|= \log|p(z_0)|,
$$
together with \eqref{eq:ss} it gives
\begin{equation}\label{eq:last}
\gamma\cdot \deg p\geq UO_{A\cap L} (\log |p|) = \log|p(z_0)| - \frac1{|A\cap L|} \int_{A\cap L} \log |p|.
\end{equation}
Thus
\begin{eqnarray*}
\int_A \log |p| & = & \int_{ S_{2n-1}} \int_{A\cap L} \log |p| \, d\mu(L)\\
& \ge & (\log|p(z_0)|-\gamma\cdot \deg p)\int_{ S_{2n-1}} |A\cap L| \, d\mu(L)\\
& = & (\log|p(z_0)|-\gamma\cdot \deg p) |A|,
\end{eqnarray*}
where $d\mu$ is a certain measure on the unit sphere $S_{2n-1}$ and we identify the set of rays $L$ starting from $z_0$ with $S_{2n-1}$. Notice that the above inequality gives
$$
\gamma\cdot \deg p \geq \log|p(z_0)| - \frac1{|A|} \int_A \log |p|=UO_A (\log|p|),
$$
from which the assertion immediately follows.
\begin{comment}
Again we will start from the following Remez inequality for $\log |z|$ on $\mathbb C$ (proof is in the appendix):
\begin{equation}\label{eq:ss2}
\sup_{[a,b]} \log|z| \leq \sup_{E} \log |z|+\log \frac{2-\delta}{\delta}, \ \ \delta:=\frac{|E|}{|b-a|}, \ \ \ \forall \ \text{compact} \ E\subset [a,b].
\end{equation}
Then we shall show that the above proposition implies the following Remez type inequality:
\begin{theorem}[The constant 26 is not optimal] For every line segment $[a,b]$ in $\mathbb C^n$ we have
$$
\sup_{[a,b]} |P| \leq \sup_{E} |P| \left(\frac{26\cdot |b-a|}{|E|}\right)^{\deg P}, \ \ \ \ \ \ \forall \ \ \text{compact} \ E\subset [a,b].
$$
\end{theorem}
\textbf{Remark}: Replace $[a,b]$ by a general convex set then we need to replace $26$ by $52 n$.
\begin{proof} Consider the restriction of $P$ to the complex line passing through $a,b$, we know it suffices to prove the one dimensional case. Up to a constant let us assume that
$$
P=(z-a_1)^{n_1} \cdots (z-a_k)^{n_k}.
$$
Then
$$
(\log |P|)_{[a,b]} =\sum_{j=1}^k n_j (\log|z-a_j|)_{[a,b]}.
$$
\end{proof}
following estimate dimension of the convex set $A$ above can be
\textbf{Remark}: Apply the above proposition to the case that $A=l$ is a line segment we get
\begin{equation}\label{eq:remez2}
\sup_l \log |P| \leq (\log |P|)_l +\gamma \deg P.
\end{equation}
By the fundamental theorem of algebra, one may write
$$
P|_{L_{\mathbb C}}=c(z-a_1)^{n_1} \cdots (z-a_k)^{n_k},
$$
where $L_{\mathbb C}$ denotes the complex line containing $l$, which gives
$$
(\log |P|)_l =\log|c|+\sum_{j=1}^k n_j (\log|z|)_{l-a_j}.
$$
Let $E$ be a compact subset of $l$ with one-dimensional Lebesgue $|E|$ satisfies
$$
1\geq \frac{|E|}{|l|}=\delta>0.
$$
We want to estimate the upper bound of
$$
(\log|z|)_{l-a_j}- (\log|z|)_{E-a_j}.
$$
Notice that there always exists a line segment $E'$ with $|E'|=|E|$ such that
$$
(\log|z|)_{E-a_j}\geq (\log|z|)_{E'-a_j}.
$$
Thus one may assume that $E$ is a line segment. Then the lemma below implies
$$
\sup_{l-a_j} \log|z| \leq \sup_{E-a_j} \log|z| +\log \frac{2-\delta}{\delta} \leq (\log |z|)_{E-a_j} +\gamma+\log \frac{2-\delta}{\delta},
$$
where the last inequality follows from Lemma \ref{le:remez1}. To summarize, we get
\begin{equation}\label{eq:remez3}
(\log |P|)_l \leq (\log |P|)_E+ \left(\gamma+\log \frac{2-\delta}{\delta}\right)\deg P.
\end{equation}
\begin{lemma} For every line segment $l$ in $\mathbb C$ we have
$$
\sup_{l} \log|z| \leq \sup_{E} \log |z|+\log \frac{2-\delta}{\delta}.
$$
for every compact subset $E$ of $l$ with $|E|=\delta |l|$.
\end{lemma}
\begin{proof} Notice that
$$
\sup_{E} \log |z|=\sup_{\hat E} \log |z|,
$$
where $\hat E$ denotes the convex hull of $E$. We know it is enough to assume that $E$ is also a line segment. Moreover, by a simple rotation around the origin, one may assume that the vertices of $l$ are $a,b$ with
$$
|a| < b.
$$
Now let $T$ be a rotation around $b$ such that
$$
T(l)\subset [-b, b].
$$
By an elementary geometry reflection, we have $|z| \geq |Tz|$ for every $|z|\leq b$, which gives
$$
\sup_{E} \log |z| \geq \sup_{T(E)} \log |z|.
$$
Thus it suffices to prove the case that $a \in \mathbb R$. Assume that
$$
E= [a',a'+\delta(b-a)]\subset [a,b].
$$
If $a\geq -\delta(b-a)/2$ then we must have
$$
\sup_E \log |z|= \log|a'+\delta(b-a)|\geq \log (a+\delta(b-a)) \geq \log b+\log \frac{\delta}{2-\delta}.
$$
Otherwise, if $a < -\delta(b-a)/2$ then
$$
\sup_E \log |z|\geq \log|a'+\delta(b-a)|\geq \log \frac{\delta(b-a)}2 \geq \log b +\log \frac{\delta}{2-\delta}.
$$
Thus our lemma follows.
\end{proof}
By \eqref{eq:remez2} and \eqref{eq:remez3}, we have the following
\begin{theorem}[weak Remez inequality] For every line segment $l$ in $\mathbb C^n$ we have
$$
\sup_{l} \log|P| \leq \sup_{E} \log |P|+\left(2\gamma+\log \frac{2-\delta}{\delta}\right)\deg P, \ \ P\in \mathbb C[z_1,\cdots, z_n],
$$
for every compact subset $E$ of $l$ with $|E|=\delta |l|$.
\end{theorem}
\textbf{Remark}: The above theorem also implies a Remez type inequality for a general compact convex set $A$. In fact, for every compact subset $E$ of $A$, by Lemma 3 in \cite{BG}, there always exists a real line $L$ passing through $z_0$ ($|z_0|=\sup_A |z|$) such that
$$
2n \frac{|L\cap E|}{|L\cap A|} \geq 2n \left(1-\left(\frac{|E|}{|A|}\right)^{\frac1{2n}}\right) \geq \frac{|E|}{|A|},
$$
thus the above theorem implies a weak version of a Remez type inequality by Yu. Brudnyi and
Ganzburg \cite{BG}.
\begin{corollary} For every compact convex set $A$ in $\mathbb C^n$ we have
$$
\sup_{A} \log|P| \leq \sup_{E} \log |P|+\left(2\gamma+\log \frac{2-\delta}{\delta}\right)\deg P, \ \ P\in \mathbb C[z_1,\cdots, z_n],
$$
for every compact subset $E$ of $A$ with $|E|=2n\delta |A|$.
\end{corollary}
\textbf{Remark}: The above corollary can be used to estimate the \emph{relative logarithmic capacity} (see \cite{BJZ} for more related results).
\end{comment}
\section{Proof of Theorem \ref{th:Main2}}
The starting point is the following
\begin{proposition}[John-Nirenberg inequality]\label{prop:J-N}
Suppose $\phi\in PSH(\Omega)$ and $\Omega_0\subset\subset\Omega$ is open. For each $a=(a_1,\cdots,a_n)$ with all $a_j>0$ there exists $\varepsilon_0=\varepsilon(a,\phi,\Omega_0,\Omega) >0$ such that
$$
\sup_{ P_{r^a}(\hat{z})\subset \Omega_0} \frac{1}{|P_{r^a}(\hat{z})|}\int_{P_{r^a}(\hat{z})} e^{-\varepsilon(\phi-\sup_{P_{r^a}(\hat{z})} \phi)} <\infty,
$$
for every $\varepsilon \leq \varepsilon_0$. Here
$$
P_{r^a}(\hat{z})=\{z\in \mathbb C^n: |z_j-\hat{z}_j|\leq r^{a_j}\}.
$$
\end{proposition}
Although the argument is fairly standard, we will provide a proof in Appendix, because the result cannot be found in literature explicitly.
\begin{lemma}\label{lm:BergmanEstimate}
Let $\psi$ be a psh function on $\Omega$ which satisfies $\sup_{\Omega} \psi <\infty$ and $\int_\Omega e^{-\psi} <\infty$. Suppose $\Omega$ is circular, i.e., $\zeta z\in \Omega$ for every $\zeta\in \mathbb C$, $|\zeta|\leq 1$, and $z\in \Omega$. Then
\begin{equation}\label{eq:BergmanEstimate}
\left(\frac{1}{|\Omega|} \int_{\Omega} e^{-(\psi-\sup_{\Omega} \psi)} \right)^{-1} \leq K_{\psi, \,\Omega}(0) \cdot |\Omega|\cdot e^{-\sup_\Omega \psi} \leq 1.
\end{equation}
\end{lemma}
\begin{proof}
The extremal property of the Bergman kernel implies that
$$
K_{\psi, \,\Omega}(0) \geq \frac{1}{\int_{\mathbb D_r} e^{-\psi}}
$$
and the first inequality in (\ref{eq:BergmanEstimate}) holds.
On the other hand, as $\Omega$ is circular, it is easy to verify that
$$
f(0)=\frac{1}{|\Omega|} \int_{\Omega} f
$$
for all $f\in {\mathcal O}(\Omega)$. Thus we have
$$
|f(0)|^2 = \left|\,\frac{1}{|\Omega|} \int_{\Omega} f\,\right|^2 \leq \frac{1}{|\Omega|} \int_{\Omega} |f|^2e^{-\psi} \cdot e^{\sup_\Omega \psi},
$$
so that the second inequality in (\ref{eq:BergmanEstimate}) also holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:Main2}]
Since
$$
\int_{\mathbb D^n} |f|^2e^{-\varepsilon \phi^t} =|t_1\cdots t_n|^{-2}\int_{\mathbb D_t^n} |f|^2e^{-\varepsilon \phi},\ \ \ \forall\,f\in {\mathcal O}(\Omega),
$$
it follows that
\begin{equation}\label{eq:2.5}
K_{\varepsilon \phi^t, \,\mathbb D^n}(z)=\frac{|\mathbb D^n_t|}{|\mathbb D^n|} \cdot K_{\varepsilon\phi, \,\mathbb D^n_t}(z),\end{equation}
where
$$
\mathbb D^n_t:=\left\{z\in {\mathbb C}^n: |z_j|<|t_j|, \ 1\le j\le n \right\}.
$$
Thus we have
$$
F(\varepsilon \phi)(t,0)=\log (|\mathbb D^n_t|\cdot K_{\varepsilon\phi, \,\mathbb D^n_t}(0))-n \log \pi.
$$
This combined with Lemma \ref{lm:BergmanEstimate} gives
$$
-\log \left(\frac{1}{|\mathbb D^n_t|} \int_{\mathbb D^n_t} e^{-\varepsilon(\phi-\sup_{\mathbb D^n_t} \phi)}\right)-n \log \pi
\leq F(\varepsilon \phi)(t,0) - \varepsilon\, {\sup}_{\mathbb D^n_t}\, \phi \leq -n \log \pi.
$$
By Proposition \ref{prop:J-N}, we conclude the proof.
\end{proof}
\section{Proof of Theorem \ref{th:Main3}}
Recall that
$$
\phi^t(z):=\phi(t_1z_1,\cdots, t_nz_n).
$$
By Proposition 2.2 in \cite{Bern17}, we have
\begin{equation}\label{eq:2.6}
\frac{\partial}{\partial t_j} K_{\phi^t, \,\mathbb D^n}(0)=\int_{\mathbb D^n} \frac{\partial \phi^t}{\partial t_j}\,|K_{\phi^t, \,\mathbb D^n}(z,0)|^2 e^{-\phi^t},
\end{equation}
where $K_{\phi^t, \,\mathbb D^n}(z,0)$ satisfies the following reproducing property
$$
f(0)=\int_{\mathbb D^n } f(z)\overline{K_{\phi^t, \,\mathbb D^n}(z,0)}e^{-\phi^t}
$$
for all $L^2$ holomorphic functions $f$ on $\mathbb D^n$. In particular, if $f=zK_{\phi^t, \,\mathbb D^n}(z,0)$ then
$$
0=\int_{\mathbb D^n} z\cdot |K_{\phi^t, \,\mathbb D^n}(z,0)|^2 e^{-\phi^t},
$$
and since
$
\frac{\partial \phi^t}{\partial t_j}|_{t=0}=z_j\phi_{z_j}(0),
$
we get
$$
\left.\int_{\mathbb D^n} \frac{\partial \phi^t}{\partial t_j}\right|_{t=0} |K_{\phi^t, \,\mathbb D^n}(z,0)|^2 e^{-\phi^t}=0
$$
for all $t\in\mathbb D^n$. Thus we may write \eqref{eq:2.6} as
$$
\frac{\partial}{\partial t_j} K_{\phi^t, \,\mathbb D^n}(0)
=\int_{\mathbb D^n} \left(\frac{\partial \phi^t}{\partial t_j}-\left.\frac{\partial \phi^t}{\partial t_j}\right|_{t=0}\right)|K_{\phi^t, \,\mathbb D^n}(z,0)|^2 e^{-\phi^t}.
$$
In particular,
$$
\left.\frac{\partial}{\partial t_j} K_{\phi^t, \,\mathbb D^n}(0)\right|_{t=0}=0.
$$
Thus we can further write \eqref{eq:2.6} as
$$
\frac{\partial}{\partial t_j} K_{\phi^t, \,\mathbb D^n}(0)-\left.\frac{\partial}{\partial t_j} K_{\phi^t, \,\mathbb D^n}(0)\right|_{t=0}=\int_{\mathbb D^n} \left(\frac{\partial \phi^t}{\partial t_j}-\left.\frac{\partial \phi^t}{\partial t_j}\right|_{t=0}\right)|K_{\phi^t, \,\mathbb D^n}(z,0)|^2 e^{-\phi^t},
$$
which implies
$$
\left.\frac{\partial^2}{\partial t_j \partial \bar t_k} K_{\phi^t, \,\mathbb D^n}(0)\right|_{t=0}=\int_{\mathbb D^n} \left.\frac{\partial^2 \phi^t}{\partial t_j\partial \bar t_k}\right|_{t=0}\cdot |K_{\phi(0), \,\mathbb D^n}(z,0)|^2 e^{-\phi(0)}.
$$
Since
$$
K_{\phi(0), \,\mathbb D^n}(z,0)=\frac{e^{\phi(0)}}{\pi^n}
$$
and
$$
\left.\frac{\partial^2 \Phi}{\partial t_j\partial \bar t_k}\right|_{t=0}=z_j\bar z_k \phi_{z_j \bar z_k}(0),
$$
we get
$$
\left.\frac{\partial^2}{\partial t_j \partial \bar t_k} K_{\phi^t, \,\mathbb D^n}(0)\right|_{t=0}=\frac{e^{\phi(0)}\phi_{z_j \bar z_k}(0)}{\pi^{2n}}\int_{\mathbb D^n} z_j \bar z_k.
$$
Notice that
$$
\int_{\mathbb D^n} z_j \bar z_k=
\begin{cases}
\begin{array}{ll}
\pi^{n}/2 & \text{if}\ \, j=k\\
0 & \text{if}\ \, j\neq k,
\end{array}
\end{cases}
$$
and
$$
\left.\frac{\partial^2}{\partial t_j \partial \bar t_k} K_{\phi^t, \,\mathbb D^n}(0)\right|_{t=0}
= K_{\phi^t, \,\mathbb D^n}(0)\cdot \left.\frac{\partial^2}{\partial t_j \partial \bar t_k} \log K_{\phi^t, \,\mathbb D^n}(0)\right|_{t=0},
$$
our assertion follows.
\section{Appendix}
In this section we provide a proof of Proposition \ref{prop:J-N}. Let us first recall a few basic facts in real-variable theory, by following Stein \cite{Stein}. A\/ {\it quasi-distance} \/ defined on ${\mathbb R}^m$ means a nonnegative continuous function $\rho$ on ${\mathbb R}^m\times
\mathbb R^m$ for which there exists a constant $c>0$ such that
\begin{enumerate}
\item $\rho(x,y)=0$ iff $x=y$;
\item $\rho(x,y)\le c\rho(y,x)$;
\item $\rho(x,y)\le c(\rho(x,z)+\rho(y,z))$.
\end{enumerate}
Given such a $\rho$, we define "balls"
$$
B(x,r):=\{y:\rho(y,x)<r\},\ \ \ r>0.
$$
One can verify that there exists a constant $c_1>1$ such that for all $x,y$ and $r$,
\begin{equation}\label{eq:App1}
B(x,r)\cap B(y,r)\neq \emptyset\ \ \ \Rightarrow \ \ \ B(y,r)\subset B(x,c_1r).
\end{equation}
In the case of Proposition \ref{prop:J-N}, we define
$$
\rho(z,w)=\max_k |z_k-w_k|^{1/a_k},\ \ \ z,w\in {\mathbb C}^n.
$$
It is easy to verify that $\rho$ is a quasi-distance on ${\mathbb C}^n$ and
$$
B(\hat{z},r)=P_{r^a}(\hat{z}),\ \ \ \hat{z}\in \mathbb C^n,\, r>0.
$$
Besides (\ref{eq:App1}), the following properties also hold for $B(\hat{z},r)$:
\begin{equation}\label{eq:App2}
|B(\hat{z},c_1r)|\le c_1^{2\sum_k a_k} \cdot |B(\hat{z},r)|=:c_2\, |B(\hat{z},r)|;
\end{equation}
\begin{equation}\label{eq:App3}
{\bigcap}_r \, \overline{B}(\hat{z},r)=\{\hat{z}\}\ \ \ {\text{and}\ \ \ } {\bigcup}_r\, B(\hat{z},r)=\mathbb C^n;
\end{equation}
\begin{equation}\label{eq:App4}
\text{For each open set\/ $U$\/ and each $r>0$, the function $\hat{z}\mapsto |B(\hat{z},r)\cap U|$ is continuous.}
\end{equation}
Fix a pair of positive constants $c^\ast$ and $c^{\ast\ast}$ with $1<c^\ast<c^{\ast\ast}$. For $B=B(\hat{z},r)$ we define $B^\ast=B(\hat{z},c^\ast r)$ and $B^{\ast\ast}=B(\hat{z},c^{\ast\ast} r)$. Then we have
\begin{lemma}[cf. \cite{Stein}, p.\,15--16]\label{lm:balls}
Choose $c^\ast=4c_1^2$ and $c^{\ast\ast}=16c_1^2$. Given a closed nonempty set $F\subset {\mathbb C}^n$, there exists a collection of balls $\{B_k\}$ such that
\begin{enumerate}
\item The $B_k$ are pairwise disjoint;
\item $\bigcup_k B_k^\ast = F^c:=\mathbb C^n\backslash F$;
\item $B_k^{\ast\ast}\cap F\neq \emptyset$ for each $k$.
\end{enumerate}
\end{lemma}
\begin{proposition}[Calder\'on-Zygmund decomposition]\label{prop:C-Z}
Let $B_0$ be a ball in $\mathbb C^n$ and $f\in L^1(B_0)$. There is a constant $c=c(c_1,c_2)>0$ such that given a positive number $\alpha$, there exists a sequences of
balls $\{B_k^\ast\}$ in $B_0$ such that
\begin{enumerate}
\item $|f(z)|\le \alpha$, for a.e. $z\in B_0\backslash\bigcup_k B_k^\ast$;
\item $
\int_{B_k^\ast} |f|\le c\alpha |B_k^\ast|,
$
for each $k$;
\item $\sum_k |B_k^\ast| \le \frac{c}{\alpha} \int_{B_0} |f|$.
\end{enumerate}
\end{proposition}
\begin{proof}
We extend $f$ to an integrable function on ${\mathbb C}^n$ by setting $f=0$ outside $B_0$.
Recall the following two types of Hardy-Littlewood maximal functions:
$$
Mf(z):=\sup_{r>0}\, \frac1{|B(z,r)|} \int_{B(z,r)} |f|,
$$
$$
\widetilde{M}f(z):=\sup_{z\in B}\, \frac1{|B|} \int_B |f|
$$
where the supremum is taken over all balls $B$ containing $z$. The relationship between $Mf$ and $\widetilde{M}f$ is as follows:
\begin{equation}\label{eq:compare}
Mf \le \widetilde{M}f\le c_2\, Mf.
\end{equation}
Notice that
$$
E_\alpha:=\left\{z\in B_0: \widetilde{M}f(z)>\alpha\right\}
$$
is an open set since $\widetilde{M}f$ is lower semicontinuous, and
$$
|E_\alpha|\le \frac{c}{\alpha} \int_{\mathbb C^n}|f| = \frac{c}{\alpha} \int_{B_0}|f|
$$
in view of (\ref{eq:compare}) and \cite{Stein}, p. 13, Theorem 1. Here and in what follows $c$ will denote a generic positive constant depending only on $c_1,c_2$. With $F:=\mathbb C^n\backslash E_\alpha$ we choose balls $\{B_k\}$, $\{B_k^\ast\}$ and $\{B_k^{\ast\ast}\}$ according to Lemma \ref{lm:balls}. Then we have
$$
{\sum}_k |B_k^\ast| \le c \, {\sum}_k |B_k| \le c\,|E_\alpha| \le \frac{c}{\alpha} \int_{B_0} |f|.
$$
Since $B_k^{\ast\ast}\cap F\neq \emptyset$ for each $k$, we have
$$
\int_{B_k^\ast} |f|\le \int_{B_k^{\ast\ast}} |f| \le \alpha |B_k^{\ast\ast}| \le c\alpha |B_k^\ast|.
$$
Finally, by (\ref{eq:compare}) and \cite{Stein}, p.\,13, Corollary, we know that $|f(z)|\le \widetilde{M}f(z)$ for a.e. $z$, from which {\it (1)} immediately follows.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:J-N}]
By Theorem \ref{th:Main1}, we know that
$$
M:=\sup_{B(\hat{z},r)\subset \Omega_0} \frac1{|B(\hat{z},r)|} \int_{B(\hat{z},r)} |\phi-\phi_{B(\hat{z},r)}|<\infty.
$$
Assume without loss of generality $M=1$. Fix a ball $B_0\subset \Omega_0$. It suffices to show
\begin{equation}\label{eq:6.6}
|\{z\in B_0: |\phi-\phi_{B_0}|>t\}|\le {\rm const}\cdot e^{-\varepsilon t}\,|B_0|,\ \ \ t>0,
\end{equation}
for certain $\varepsilon\ll 1$.
With $c$ as Proposition \ref{prop:C-Z} we choose
$$
\alpha>c>1\ge \frac1{|B_0|}\int_{B_0} |\phi-\phi_{B_0}|.
$$
Applying Proposition \ref{prop:C-Z} with
$
f=
|\phi-\phi_{B_0}|,
$
we have a sequence of balls $\{B_k^{(1)}\}$ in $B_0$ such that
$$
|\phi(z)-\phi_{B_0}|\le \alpha \ \ \ \text{a.e.}\ \ \ z\in B_0\backslash {\bigcup}_k B_k^{(1)},
$$
$$
{\sum}_k |B_k^{(1)}|\le \frac{c}{\alpha} \int_{B_0} |\phi-\phi_{B_0}|\le \frac{c}{\alpha}\,|B_0|,
$$
and
$$
|\phi_{B_k^{(1)}}-\phi_{B_0}|\le \frac1{|B_k^{(1)}|} \int_{B_k^{(1)}} |\phi-\phi_{B_0}|\le c\alpha.
$$
Applying Proposition \ref{prop:C-Z} with
$
f=
|\phi-\phi_{B_k^{(1)}}|
$
for each $k$, we obtain a sequence of balls $\{B_k^{(2)}\}$ in $\bigcup_k B_k^{(1)}$ such that
$$
{\sum}_k |B_k^{(2)}|\le \frac{c}{\alpha} \,{\sum}_k \int_{B_k^{(1)}} |\phi-\phi_{B_k^{(1)}}|\le \frac{c}{\alpha} \,{\sum}_k
|B_k^{(1)}| \le \frac{c^2}{\alpha^2} \, |B_0|
$$
and
$$
|\phi(z)-\phi_{B_k^{(1)}}|\le \alpha \ \ \ \text{a.e.}\ \ \ z\in B_k^{(1)}\backslash {\bigcup}_k B^{(2)}_k,
$$
which in turn implies
$$
|\phi(z)-\phi_{B_0}|\le 2\cdot c\alpha \ \ \ \text{a.e.}\ \ \ z\in B_0\backslash {\bigcup}_k B^{(2)}_k.
$$
Continue this process. For each $j$ there exists a sequence of balls $\{B_k^{(j)}\}$ in $\bigcup_k B_k^{(j-1)}$ such that
$$
{\sum}_k |B_k^{(j)}|\le \frac{c^j}{\alpha^j} \, |B_0|,
$$
$$
|\phi(z)-\phi_{B_0}|\le j\cdot c\alpha \ \ \ \text{a.e.}\ \ \ z\in B_0\backslash {\bigcup}_k B^{(j)}_k.
$$
Thus
$$
|\{z\in B_0:|\phi-\phi_{B_0}|>j\cdot c\alpha\}|\le {\sum}_k |B_k^{(j)}|\le \frac{c^j}{\alpha^j} \, |B_0|.
$$
For any $t$ there exists an integer $j$ such that $t\in [j\cdot c\alpha,(j+1)\cdot c\alpha)$. It follows that
$$
(c/\alpha)^j=(\alpha/c)e^{-(j+1)\log \alpha/c}\le (\alpha/c)e^{-\frac{\log \alpha/c}{c\alpha} \, t},
$$
from which \eqref{eq:6.6} immediately follows. Now we have
$$
\frac{1}{|B_0|} \int_{B_0} e^{\varepsilon |\phi-\phi_{B_0}|} = \frac{1}{|B_0|} \int_{0}^\infty e^{\varepsilon t} |\{|\phi-\phi_{B_0}|>t\}| + \frac{ |\{\phi=\phi_{B_0}\}| }{|B_0|} \leq {\rm const}+1,
$$
which gives
$$
\frac{1}{|B_0|} \int_{B_0} e^{-\varepsilon (\phi-\sup_{B_0}\phi)} \leq ({\rm const}+1) e^{\varepsilon(\sup_{B_0}\phi-\phi_{B_0})},
$$
By Theorem 1.1, $\sup_{B_0}\{\sup_{B_0}\phi-\phi_{B_0}\} <\infty$, thus Proposition \ref{prop:J-N} follows.
\end{proof}
\subsection*{Acknowledgments} The authors would like to thank Ahmed Zeriahi for bringing their attention to the reference \cite{BJZ}.
The second author would like to thank Bo Berndtsson for numerous useful discussions about the topics of this paper.
|
\section{Introduction.}
Cantor's Continuum Problem, which later became Hilbert's first Problem (see \cite{hilbert}), asks how many real numbers there are. This question got a non-answer through the discovery of the method of {\em forcing} by Paul Cohen: {\sf CH}, Cantor's Continuum Hypothesis, is independent from {\sf ZFC} (see \cite{cohen}), the standard axiom system for set theory which had been isolated by Zermelo and Fraenkel. {\sf CH} states that every uncountable set of reals has the same size as ${\mathbb R}$.
Ever since Cohen, set theorists have been searching for natural new axioms which extend {\sf ZFC} and which settle the Continuum Problem. See e.g.\ \cite{hugh-ch1}, \cite{hugh-ch2}, \cite{peter}, and the discussion in
\cite{new?}. There are two prominent such axioms which decide {\sf CH} in the negative and which in fact both prove that there are $\aleph_2$ reals: Martin's Maximum ({\sf MM}, for short) or variants thereof on the one hand (see \cite{FMS}), and Woodin's axiom $(*)$ on the other hand (see \cite{hugh}). See e.g.\ \cite{menachem}, \cite{stevo}, and \cite{moore}.
Both of these axioms may be construed as maximality principles for the theory of the structure $(H_{\omega_2};\in)$, but up to this point the relationship between {\sf MM} and $(*)$ was a bit of a mystery, which led M.\ Magidor to call $(*)$ a ``competitor'' of {\sf MM} (\cite[p.\ 18]{menachem}). Both {\sf MM} and $(*)$ are inspired by and formulated in the language of forcing, and they both have ``the same intuitive motivation: Namely, the
universe of sets is rich'' (\cite[p.\ 18]{menachem}).
This paper resolves the tension between {\sf MM} and $(*)$ by proving that ${\sf MM}^{++}$, a strengthening of {\sf MM}, actually {\em implies} $(*)$, see Theorem \ref{main_thm} below, so that {\sf MM} and $(*)$ are actually compatible with each other.
This answers \cite[Question (18) a) on p.\ 924]{hugh}, see also \cite[p.\ 846]{hugh}, \cite[Conjecture 6.8 on p.\ 19]{menachem},
and \cite[Problem 14.7]{moore}.
\section{Preliminaries.}
{\em Martin's Maximum}${}^{++}$, abbreviated by {\sf MM}${}^{++}$, see \cite{FMS} (cf.\ also \cite[Definition 2.45 (2)]{hugh}), is the statement that if ${\mathbb P}$ is a forcing which preserves stationary subsets of $\omega_1$, if $\{ D_i \colon
i<\omega_1 \}$ is a collection of dense subsets of ${\mathbb P}$, and if $\{ \tau_i
\colon i<\omega_1 \}$ is a collection of ${\mathbb P}$-names for stationary subsets
of $\omega_1$, then there is a filter $g \subset {\mathbb P}$ such that
for every $i<\omega_1$,
\begin{enumerate}
\item[(i)]
$g \cap
D_i \not= \emptyset$ and
\item[(ii)] $(\tau_i)^g = \{ \xi<\omega_1
\colon \exists p \in g \, p \Vdash_{\mathbb P} {\check \xi} \in \tau_i \}$ is stationary.
\end{enumerate}
Woodin's ${\mathbb P}_{\rm max}$ axiom (*), see \cite[Definition 5.1]{hugh},
is the statement that
\begin{enumerate}
\item[(i)]
{\sf AD} holds in $L({\mathbb R})$ and
\item[(ii)] there is some $g$
which is ${\mathbb P}_{\rm max}$-generic over $L({\mathbb R})$ such that
${\cal P}(\omega_1) \cap V \subset L({\mathbb R})[g]$.
\end{enumerate}
Already {\sf PFA}, the {\em Proper Forcing Axiom}, which is weaker than {\sf MM}${}^{++}$, implies ${\sf AD}^{L({\mathbb R})}$ and much more, see \cite{john}, \cite{stacking}, and \cite[Chapter 12]{trang-sargsyan}.
This paper produces a proof of the following result.
\begin{theorem}\label{main_thm}
Assume {\em Martin's Maximum}${}^{++}$. Then Woodin's ${\mathbb P}_{\rm max}$-axiom $(*)$ holds true.
\end{theorem}
Our key new idea is ($\Sigma$.8) on page \pageref{(C.9)} below.
P.\ Larson, see \cite{plarson} and \cite{plarson2}, has shown that ${\sf MM}^{+\omega}$ is consistent with $\lnot (*)$
relative to a supercompact limit of supercompact
cardinals.
Throughout our entire paper, ``$\omega_1$''
will {\em always} denote $\omega_1^V$, the $\omega_1$ of $V$.
Let us fix throughout this paper some $A \subset \omega_1$ such that $\omega_1^{L[A]} = \omega_1$. Let us
define $g_A$ as the set of all ${\mathbb P}_{\rm max}$ conditions $p=(N;\in,I,a)$
such that there is a generic iteration $$(N_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = N_0$ of length $\omega_1+1$ such that if we write $N_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$,\footnote{Here and elsewhere we often confuse a model with its underlying universe.} then $I^* = ({\sf NS}_{\omega_1})^V \cap N_{\omega_1}$ and $a^* = A$.
\begin{lemma}\label{folklore} \textbf{\em (Woodin)} Assume that ${\sf NS}_{\omega_1}$ is saturated and that ${\cal P}(\omega_1)^\#$ exists.
\begin{enumerate}
\item[(1)] $g_A$ is a filter.
\item[(2)] If $g_A$ is ${\mathbb P}_{\rm max}$-generic over $L({\mathbb R})$,
then ${\cal P}({\omega_1}) \subset L({\mathbb R})[g]$.
\end{enumerate}
\end{lemma}
{\sc Proof.} This routinely follows from the proof of \cite[Lemma 3.12 and Corollary 3.13]{hugh} and from \cite[Lemma 3.10]{hugh}. \hfill $\square$
\bigskip
One may also use {\sf BMM} plus ``${\sf NS}_{\omega_1}$ is precipitous'' to show that
$g_A$ is a filter, this is by the proof from \cite{cs}.
Let $\Gamma \subset \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k)$. We say that $\Gamma$ is {\em productive} iff for all $k<\omega$ and all $D \in \Gamma \cap {\cal P}({\mathbb R}^{k+2})$, if $D$ is universally Baire (see \cite{FMW}) as being witnessed by the trees $T$ and $U$ on ${}^{k+2} \omega \times {\rm OR}$, i.e., $D = p[T]$ and for all posets ${\mathbb P}$,
\begin{eqnarray}
\Vdash_{\mathbb P} \, p[U] = {\mathbb R}^{k+2} \setminus p[T]{\rm , }
\end{eqnarray}
and if
\begin{eqnarray}\label{U-tilde}
{\tilde U} = \{ (s \upharpoonright (k+1), (s(k+1),t)) \colon (s,t) \in U \}{\rm , }
\end{eqnarray}
so that $(x_0, \ldots , x_{k}) \in p[{\tilde U}]$ iff there is some $y$ such that
$(x_0, \ldots , x_{k},y) \in p[U]$,
then
there is a tree ${\tilde T}$ on ${}^{k+1} \omega \times {\rm OR}$ such that for all posets ${\mathbb P}$,
\begin{eqnarray}
\Vdash_{\mathbb P} \, p[{\tilde U}] = {\mathbb R}^{k+1} \setminus p[{\tilde T}].
\end{eqnarray}
Let us denote by $\Gamma^\infty$ the collection of all $D \in \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k)$ which are universally Baire. If $D \in \Gamma^\infty$, then there is an
unambiguous version of $D$ in any forcing extension of $V$, which as
usual we
denote by $D^*$. (\ref{U-tilde}) then means that
if $D=p[U]$ and $E = p[{\tilde U}]$, then in any forcing extension of $V$,
$E^* = \exists^{\mathbb R} \, D^*$.
If $\Gamma \subset \Gamma^\infty$ is productive and if $D \in \Gamma$, then any projective statement about $D$ is absolute between $V$ and any forcing extension of $V$,\footnote{This seems to be wrong if we just assume $\Gamma \subset \Gamma^\infty$, but the hypothesis that $\Gamma$ be productive is crossed out.} i.e., if $\varphi$ is projective, $x_1$, $\ldots$, $x_k \in {\mathbb R}$, and ${\mathbb P}$ is any poset, then
$$V \models \varphi(x_1, \ldots , x_k, D) \ \Longleftrightarrow \ \Vdash_{\mathbb P} \, \varphi({\check x}_1, \ldots , {\check x}_k, {D}^*).$$
By a theorem of Woodin,
see e.g.\ \cite[Theorem 1.2]{forcing-free}, combined with the key result of Martin and Steel
in \cite{proj-determinacy}, the pointclass $\Gamma^\infty$ is productive under the
hypothesis that there is a proper class of Woodin cardinals.
The proof from \cite{john} produces the result that under {\sf PFA}, the universe is closed under the operation $X \mapsto M_\omega^\#(X)$, which implies that every set of reals in $L({\mathbb R})$ is universally Baire and that $\bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k) \cap L({\mathbb R})$ is productive. See e.g.\ \cite[Section 3, pp.\ 187f.]{both} on the relevant argument.
Therefore, in the light of Lemma \ref{folklore},
Theorem \ref{main_thm}
follows from the following more general statement.
\begin{theorem}\label{main_thm_gen}
Let $\Gamma \subset \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k)$. Assume that
\begin{enumerate}
\item[(i)] $\Gamma = \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k) \cap L(\Gamma,{\mathbb R})$,
\item[(ii)] $\Gamma \subset \Gamma^\infty$,
\item[(iii)] $\Gamma$ is productive, and
\item[(iv)] Martin's Maximum${}^{++}$ holds true.
\end{enumerate}
Then $g_A$ is ${\mathbb P}_{\rm max}$-generic over $L(\Gamma,{\mathbb R})$.\footnote{In the presence of a proper class of Woodin cardinals, hypotheses (i), (ii), and (iv) then give $(*)_\Gamma$, see \cite[Definition 4.1]{both}.}
\end{theorem}
In the light of Lemma \ref{folklore} and \cite[Corollary 17]{FMS}, Theorem
\ref{main_thm_gen} readily follows from the following via a standard application
of ${\sf MM}^{++}$.
\begin{lemma}\label{main_claim}
Let $\Gamma \subset \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k)$. Assume that
\begin{enumerate}
\item[(i)] $\Gamma = \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k) \cap L(\Gamma,{\mathbb R})$,
\item[(ii)] $\Gamma \subset \Gamma^\infty$,
\item[(iii)] $\Gamma$ is productive, and
\item[(iv)] ${\sf NS}_{\omega_1}$ is saturated.\footnote{We could weaken this hypothesis to ``${\sf NS}_{\omega_1}$ is precipitous.''}
\end{enumerate}
Let $D \subset {\mathbb P}_{\rm max}$ be open dense, $D \in L(\Gamma,{\mathbb R})$.\footnote{By hypothesis, $D$ is then universally Baire in the codes, so that
there is an
unambiguous version of $D$ in any forcing extension of $V$, which again
we
denote by $D^*$.}
There is then a stationary set preserving forcing ${\mathbb P}$ such that in $V^{\mathbb P}$
there is some $p = (N;\in,I^*,a^*) \in D^*$ and some generic iteration $$(N_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = N_0$ of length $\omega_1+1$ such that if we write $N_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$, then $I^* = ({\sf NS}_{\omega_1})^{V^{\mathbb P}} \cap N_{\omega_1}$ and $a^* = A$.
\end{lemma}
The attentive reader will notice that we don't need the full power of ${\sf MM}^{++}$
in order to derive Theorem \ref{main_thm_gen} from Lemma \ref{main_claim},
the hypothesis that
$D$-${\sf BMM}^{++}$ holds true for all $D \in \Gamma^\infty$ would suffice, see
\cite[Definition 10.123]{hugh}. By the proof of \cite[Theorem 2.7]{david-ralf}, $(*)$ is then actually {\em equivalent} to a version of {\sf BMM}; we state this as follows.
\begin{theorem}
Let $\Gamma \subset \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k)$. Assume that
\begin{enumerate}
\item[(i)] $\Gamma = \bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k) \cap L(\Gamma,{\mathbb R})$,
\item[(ii)] $\Gamma \subset \Gamma^\infty$,
\item[(iii)] $\Gamma$ is productive, and
\item[(iv)] ${\sf NS}_{\omega_1}$ is saturated.\footnote{Again, we could weaken this hypothesis to ``${\sf NS}_{\omega_1}$ is precipitous.''}
\end{enumerate}
The following statements are then equivalent.
\begin{enumerate}
\item[(1)] $D$-${\sf BMM}^{++}$ holds true for all $D \in \Gamma$.
\item[(2)] $g_A$ is ${\mathbb P}_{\rm max}$-generic over $L(\Gamma,{\mathbb R})$.
\end{enumerate}
\end{theorem}
\begin{theorem} Assume that there is a proper class of Woodin cardinals.
The following statements are then equivalent.
\begin{enumerate}
\item[(1)] $D$-${\sf BMM}^{++}$ holds true for all
$D \in {\cal P}({\mathbb R}) \cap L({\mathbb R})$.
\item[(2)] $(*)$.
\end{enumerate}
\end{theorem}
Our next section is entirely devoted to a proof of Lemma \ref{main_claim}.
The authors thank Andreas Lietz for his comments on earlier versions of this paper.
\section{The forcing.}
Let us assume throughout the hypotheses of Lemma \ref{main_claim}.
We aim to verify its conclusion.
Let us fix $D \subset {\mathbb P}_{\rm max}$, an open dense set in $L(\Gamma,{\mathbb R})$.
By hypotheses (ii) and (iii) in the statement of Lemma \ref{main_claim} we will have that
\begin{enumerate}
\item[(D.1)] $V^{{\rm Col}(\omega,\omega_2)} \models$ ``$D^*$ is an open dense
subset of ${\mathbb P}_{\rm max}$.''
\end{enumerate}
Let us identify $D$ with a canonical set of reals coding the elements of $D$,\footnote{We will have to spell out a bit more precisely below in which way we aim to have the elements of $p[T]$ code the elements of $D$, see ($\Sigma$.5) below.} and
let $T \in V$ be a
tree on $\omega \times 2^{{\aleph_2}}$ such that
\begin{enumerate}
\item[(D.2)] $V^{{\rm Col}(\omega,\omega_2)} \models D^*=p[T]$.
\end{enumerate}
Let us write
\begin{eqnarray} \kappa = (2^{\aleph_2})^+ {\rm , }
\end{eqnarray}
so that $T \in H_\kappa$.
Let $d$ be ${\rm Col}(\kappa,\kappa)$-generic over $V$. In $V[d]$, let
$({\bar A}_\lambda \colon \lambda<\kappa)$ be a $\Diamond_\kappa$-sequence, i.e.,
for all ${\bar A} \subset \kappa$, $\{ \lambda<\kappa \colon {\bar A} \cap \lambda = {\bar A}_\lambda \}$ is stationary.
Also, let $c \colon \kappa \rightarrow H_\kappa^{V} = H_\kappa^{V[d]}$,
$c \in V[d]$, be bijective.
For $\lambda < \kappa$, let
\begin{eqnarray}
Q_\lambda = c \mbox{''} \lambda \mbox{ and } A_\lambda = c \mbox{''} {\bar A}_\lambda.
\end{eqnarray}
Let $C \subset \kappa$ be club such that for all $\lambda \in C$,
\begin{enumerate}
\item[(i)] $Q_\lambda$
is transitive,
\item[(ii)] $\{ T , ((H_{\omega_2})^V ; \in , ({\sf NS}_{\omega_1})^V ,A) \} \cup
2^{\aleph_2} \subset Q_\lambda$,
\item[(iii)] $Q_\lambda \cap {\rm OR} = \lambda$ (so that $c \upharpoonright \lambda \colon
\lambda \rightarrow Q_\lambda$ is bijective), and
\item[(iv)] $({\cal Q}_\lambda;\in) \prec
(H_\kappa;\in)$.
\end{enumerate}
In $V[d]$, for all $P$, $B \subset H_\kappa$, the set of all $\lambda \in C$ such that
$$(Q_\lambda;\in,P \cap Q_\lambda , B \cap Q_\lambda) \prec (H_\kappa;\in,P,B)$$
is club, and the set of all $\lambda \in C$ such that $B \cap Q_\lambda = A_\lambda$ is stationary, so that
\begin{enumerate}
\item[($\Diamond$)] For all $P$, $B \subset H_\kappa$ the set
$$\{ \lambda \in C \colon (Q_\lambda;\in , P \cap Q_\lambda , A_\lambda) \prec (H_\kappa;\in , P , B) \}$$
is stationary.
\end{enumerate}
We shall sometimes also write $Q_\kappa=H_\kappa$.
We shall now go ahead and
produce a stationary set preserving forcing ${\mathbb P} \in V[d]$ which adds some
$p \in D^*$ and some generic iteration $$(N_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = N_0$ such that if we write $N_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$, then $I^* = ({\sf NS}_{\omega_1})^{V[d]^{\mathbb P}} \cap N_{\omega_1}$ and $a^* = A$. As the forcing ${\rm Col}(\kappa,\kappa)$ which added $d$ is certainly stationary set preserving, this will verify Lemma \ref{main_claim}.
${\sf NS}_{\omega_1}$ is still saturated in $V[d]$ and (D.1) and (D.2) are still true in $V[d]$, so that in order to simplify our notation, we shall in what follows confuse $V[d]$ with $V$, i.e., pretend that in addition to ``${\sf NS}_{\omega_1}$ is saturated''
plus (D.1) and (D.2), ($\Diamond$) is also true in $V$.
Working under these hypotheses, we shall now
recursively define a $\subset$-increasing and continuous chain of forcings ${\mathbb P}_\lambda$ for all
$\lambda \in C \cup \{ \kappa \}$. The forcing ${\mathbb P}$ will be ${\mathbb P}_\kappa$.
Assume that $\lambda \in C \cup \{ \kappa \}$ and ${\mathbb P}_\mu$ has already
been defined in such a way that ${\mathbb P}_\mu \subset Q_\mu$
for all $\mu \in C \cap \lambda$.
We shall be interested in objects ${\mathfrak C}$ which exist in some outer model\footnote{$W$ is an outer model iff $W$ is a transitive model of {\sf ZFC} with
$W \supset V$ and which has the same ordinals as $V$; in other words, $W$ is an
outer model iff $V$ is an inner model of $W$.}
and which have the following properties.
\begin{eqnarray}\label{certificate}
{\mathfrak C} = \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle {\rm , }
\end{eqnarray}
and
\begin{enumerate}
\item[(C.1)] $M_0$, $N_0 \in {\mathbb P}_{\rm max}$,
\item[(C.2)] $x = \langle k_n \colon n<\omega \rangle$ is a real code for $N_0$ and
$\langle (k_n , \alpha_n ) \colon n<\omega \rangle \in [T]$,
\item[(C.3)] $\langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1^{N_0} \rangle \in N_0$ is a generic iteration of $M_0$ which witnesses that $N_0
< M_0$ in ${\mathbb P}_{\rm max}$,
\item[(C.4)] $\langle N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle$ is a generic iteration of $N_0$ such that if $$N_{\omega_1} = (N_{\omega_1};\in,I^*,A^*){\rm , }$$ then
$A^* = A$,\footnote{There is no requirement on $I^*$ matching the non-stationary ideal of some model in which ${\mathfrak C}$ exists.\label{no-requirement}}
\item[(C.5)] $\langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1 \rangle = \sigma_{0 \omega_1}( \langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1^{N_0}\rangle )$ and \begin{eqnarray}\label{Momega1}
M_{\omega_1} = ((H_{\omega_2})^V ; \in , ({\sf NS}_{\omega_1})^V ,A){\rm , }
\end{eqnarray}
\item[(C.6)] $K \subset {\omega_1}$,
\end{enumerate}
and for all $\delta \in K$,
\begin{enumerate}
\item[(C.7)] $\lambda_\delta<\lambda$, and if $\gamma<\delta$ is in $K$, then
$\lambda_\gamma<\lambda_\delta$ and
$X_\gamma \cup \{ \lambda_\gamma \} \subset X_\delta$, and
\item[(C.8)] $X_\delta \prec (Q_{\lambda_\delta};\in,{\mathbb P}_{\lambda_\delta},A_{\lambda_\delta})$ and $X_\delta \cap \omega_1 = \delta$.
\end{enumerate}
We need to define a language ${\cal L}$ (independently from $\lambda$)
whose formulae will be able to describe ${\mathfrak C}$ with the above properties by producing the
models $M_i$ and $N_i$, $i <\omega_1$, as term models out
of equivalence classes of terms of the form ${\dot n}$, $n<\omega$.
The language ${\cal L}$ will have the the following constants.
\begin{align*}
{\dot T} & \mbox{ \ \ \ intended to denote } T \\
x \mbox{ for every } x \in H_\kappa & \mbox{ \ \ \ intended to denote } x \mbox{ itself } \\
{\dot n} \mbox{ for every } n<\omega & \mbox{ \ \ \ as terms for elements of } M_i \mbox{ and } N_i, i < \omega_1 \\
{\dot M}_i \mbox{ for } i<\omega_1 & \mbox{ \ \ \ intended to denote } M_i \\
{\dot \pi}_{ij} \mbox{ for } i \leq j \leq \omega_1 & \mbox{ \ \ \ intended to denote } \pi_{ij} \\
{\dot {\vec M}} & \mbox{ \ \ \ intended to denote } (M_j , \pi_{jj'} \colon j \leq j' \leq
\omega_1^{N_i}) \mbox{ for } i < \omega_1 \\
{\dot N}_i \mbox{ for } i < \omega_1 & \mbox{ \ \ \ intended to denote } N_i \\
{\dot \sigma}_{ij} \mbox{ for } i \leq j < \omega_1 & \mbox{ \ \ \ intended to denote } \sigma_{ij} \\
{\dot a} & \mbox{ \ \ \ intended to denote the distinguished $a$-predicate of } M_i, N_i, i<\omega_1 \\
{\dot I}
& \mbox{ \ \ \ intended to denote the distinguished ideal of } N_i, i<\omega_1\\
{\dot X}_\delta \mbox{ for } \delta<\omega_1 & \mbox{ \ \ \ intended to denote } X_\delta.
\end{align*}
Formulae of ${\cal L}$ will be of the following form.
\begin{gather*}
\ulcorner {\dot N}_i \models \varphi(\xi_1, \ldots , \xi_k , {\dot n}_1 , \ldots , {\dot n}_\ell, {\dot a}, {\dot I} ,
{\dot M}_{j_1}, \ldots, {\dot M}_{j_m} , {\dot \pi}_{q_1 r_1} , \ldots , {\dot \pi}_{q_s r_s} , {\dot {\vec M}}) \urcorner \\
\mbox{for } i<\omega_1,
\xi_1, \ldots , \xi_k < \omega_1, n_1, \ldots , n_\ell < \omega, j_1, \ldots, j_m < \omega_1,
q_1 \leq r_1 < \omega_1, \ldots, q_s \leq r_s < \omega_1 \\
\ulcorner {\dot \pi}_{i \omega_1}({\dot n})=x \urcorner \mbox{ \ \ \ for }
i < \omega_1 \mbox{ and } x \in H_{\omega_2} \\
\ulcorner {\dot \pi}_{\omega_1 \omega_1}(x)=x \urcorner \mbox{ \ \ \ for }
x \in H_{\omega_2} \\
\ulcorner {\dot \sigma}_{ij}({\dot n})={\dot m} \urcorner \mbox{ \ \ \ for }
i \leq j < \omega_1, n, m<\omega \\
\ulcorner ({\vec u},{\vec \alpha}) \in {\dot T} \urcorner \mbox{ \ \ \
for } {\vec u} \in {}^{<\omega} \omega \mbox{ and } {\vec \alpha}
\in {}^{<\omega} (2^{\aleph_2}) \\
\ulcorner \delta \mapsto {\bar \lambda} \urcorner \mbox{ \ \ \ for } \delta<\omega_1,
{\bar \lambda}<\kappa \\
\ulcorner x \in {\dot X}_\delta \urcorner \mbox{ \ \ \ for } \delta<\omega_1, x \in H_\kappa
\end{gather*}
Let us write ${\cal L}^\lambda$ for the collection of all ${\cal L}$-formulae except for the
formulae which mention elements outside of $Q_\lambda$, i.e., except for the formulae
of the form $\ulcorner \delta \mapsto {\bar \lambda} \urcorner$
for $\delta < \omega_1$ and ${\lambda} \leq {\bar \lambda} < \kappa$ and $\ulcorner x \in {\dot X}_\delta \urcorner$ for $\delta < \omega_1$ and $x
\in H_\kappa \setminus Q_\lambda$.
We may and shall assume that ${\cal L}$ is built in a canonical way so that
${\cal L}^\lambda \subset Q_\lambda$.
We say that the objects ${\mathfrak C}$ as in
(\ref{certificate}) are {\em pre-certified} by a collection $\Sigma$ of ${\cal L}^\lambda$-formulae
if and only if (C.1) through (C.8) are satisfied by ${\mathfrak C}$
and there are surjections $e_i \colon \omega \rightarrow
N_i$
for $i<\omega_1$
such that the following hold true.
\begin{enumerate}
\item[($\Sigma$.1)]
$\ulcorner {\dot N}_i \models \varphi(\xi_1, \ldots , \xi_k , {\dot n}_1 , \ldots , {\dot n}_\ell, {\dot a}, {\dot I} ,
{\dot M}_{j_1}, \ldots, {\dot M}_{j_m} , {\dot \pi}_{q_1 r_1} , \ldots , {\dot \pi}_{q_s r_s} , {\dot {\vec M}})
\urcorner \in \Sigma$ iff \newline $i<\omega_1$,
$\xi_1$, $\ldots$, $\xi_k \leq \omega_1^{N_i}$, $n_1$, $\ldots$ , $n_\ell < \omega$, $j_1$, $\ldots$, $j_m \leq \omega_1^{N_i}$,
$q_1 \leq r_1 \leq \omega_1^{N_i}$, $\ldots$, $q_s \leq r_s \leq \omega_1^{N_i}$ and \newline
$N_i \models \varphi(\xi_1, \ldots , \xi_k , e_i(n_1) , \ldots , e_i(n_\ell), A \cap \omega_1^{N_i}, I^{N_i} , M_{j_1}, \ldots
, M_{j_m}, \pi_{q_1 r_1}, \ldots , \pi_{q_s r_s}, {\vec M})$, where $I^{N_i}$ is the distinguished ideal of $N_i$ and ${\vec M}=\langle M_j , \pi_{jj'} \colon
j \leq j' \leq \omega_1^{N_i})$,
\item[($\Sigma$.2)] $\ulcorner {\dot \pi}_{i \omega_1}({\dot n})=x \urcorner \in \Sigma$ iff $i < \omega_1$, $n<\omega$, and $\pi_{i \omega_1}(e_i(n))=x$,
\item[($\Sigma$.3)] $\ulcorner {\dot \pi}_{\omega_1 \omega_1}(x)=x \urcorner \in \Sigma$ iff $x \in H_{\omega_2}$,
\item[($\Sigma$.4)] $\ulcorner {\dot \sigma}_{ij}({\dot n})={\dot m} \urcorner \in \Sigma$ iff $i \leq j < \omega_1$, $n$, $m<\omega$, and $\sigma_{ij}(e_i(n))=e_j(m)$,
\item[($\Sigma$.5)] letting $F$ with ${\rm dom}(F)=\omega$ be the monotone enumeration
of the G\"odel numbers of all $\ulcorner {\dot N}_0 \models \varphi({\dot n}_1, \ldots , {\dot n}_\ell, {\dot a}, {\dot I})
\urcorner$ with $\ulcorner {\dot N}_0 \models \varphi({\dot n}_1, \ldots , {\dot n}_\ell, {\dot a}, {\dot I})
\urcorner \in \Sigma$, we have that $\ulcorner ({\vec u},{\vec \alpha}) \in {\dot T} \urcorner \in \Sigma$ iff
there is some $n<\omega$ such that $\langle {\vec u},{\vec \alpha} \rangle
= \langle (F(m),\alpha_m ) \colon m< n \rangle$ and $F(m)=k_m$ for all $m<n$,
\item[($\Sigma$.6)] $\ulcorner \delta \mapsto {\bar \lambda} \urcorner \in \Sigma$ iff
$\delta \in K$ and ${\bar \lambda} = \lambda_\delta$, and
\item[($\Sigma$.7)] $\ulcorner {x} \in {\dot X}_\delta \urcorner \in \Sigma$ iff
$\delta \in K$ and $x \in X_\delta$.
\end{enumerate}
We say that the objects ${\mathfrak C}$ as in
(\ref{certificate}) are {\em certified} by a collection $\Sigma$ of formulae
if and only if ${\mathfrak C}$ is pre-certified by $\Sigma$ and in addition,
\begin{enumerate}
\item[($\Sigma$.8)]\label{(C.9)} if $\delta \in K$, then
$[\Sigma]^{<\omega} \cap X_\delta \cap
E \not= \emptyset$ for every $E \subset {\mathbb P}_{\lambda_\delta}$ which is dense in ${\mathbb P}_{\lambda_\delta}$ and definable over the structure $$(Q_{\lambda_\delta}; \in , {\mathbb P}_{\lambda_\delta} , A_{\lambda_\delta} )$$ from parameters in $X_\delta$.\footnote{Equivalently,
$[\Sigma]^{<\omega} \cap
E \not= \emptyset$ for every $E \subset {\mathbb P}_{\lambda_\delta} \cap
X_\delta$ which is dense in ${\mathbb P}_{\lambda_\delta} \cap
X_\delta$ and definable over the structure $$(X_\delta; \in , {\mathbb P}_{\lambda_\delta} \cap
X_\delta , A_{\lambda_\delta} \cap X_\delta )$$ from parameters in $X_\delta$.}
\end{enumerate}
By way of definition, we call ${\mathfrak C}$ as in (\ref{certificate}) a {\em semantic
certificate} iff
there is a
collection $\Sigma$ of formulae such that ${\mathfrak C}$ is certified by $\Sigma$.
We call $\Sigma$ a {\em syntactic certificate}
iff there is a semantic certificate ${\mathfrak C}$ such that ${\mathfrak C}$ is certified by
$\Sigma$. Given a syntactic certificate $\Sigma$, there is a {\em unique} semantic
certificate ${\mathfrak C}$ such that ${\mathfrak C}$ is certified by $\Sigma$.
Even though it is obvious how to construct ${\mathfrak C}$ from $\Sigma$, in the proof of Lemma \ref{generics_give_certificates} below we \label{details} will provide details on how to derive a semantic certificate from a given $\Sigma$.
Let $\Sigma \cup p$ be a set of formulae, where $p$ is finite.
We say that $p$
is {\em certified by}
$\Sigma$ if and only if
there is some (unique) ${\mathfrak C}$ as in (\ref{certificate})
such that
${\mathfrak C}$ is certified by
$\Sigma$ and
\begin{enumerate}
\item[($\Sigma$.9)] $p \in [\Sigma]^{<\omega}$.
\end{enumerate}
We may also say that $p$ is {\em certified by} ${\mathfrak C}$ as in (\ref{certificate}) iff there is some $\Sigma$ such that ${\mathfrak C}$ and $p$ are both certified by $\Sigma$ -- and we will then also refer to $\Sigma$ as a syntactical certificate for $p$ and to ${\mathfrak C}$ as the associated semantical certificate.
We are then ready to define the forcing ${\mathbb P}_\lambda$.
We say that $p \in {\mathbb P}_\lambda$ if and only if
\begin{eqnarray}\label{in_gen_ext}
V^{{\rm Col}(\omega,\lambda)} \models \mbox{``There is a set $\Sigma$ of ${\cal L}^\lambda$-formulae such that } p \mbox{ is certified by } \Sigma \mbox{.''}
\end{eqnarray}
Let $p$ be a finite set of formulae of ${\cal L}^\lambda$. By the homogeneity of ${\rm Col}(\omega,\lambda)$, if there is some $h$ which is
${\rm Col}(\omega,\lambda)$-generic over $V$ and there is some
$\Sigma
\in V[h]$
such that $p$ is certified by $\Sigma$, then for all $h$ which are ${\rm Col}(\omega,\lambda)$-generic over $V$ there is some $\Sigma \in V[h]$ such that $p$ is certified by $\Sigma$.
It is then easy to see that $\langle {\mathbb P}_\lambda \colon \lambda \in C \cup
\{ \kappa \} \rangle$ is definable over $V$ from $\langle A_\lambda \colon
\lambda < \kappa \rangle$
and $C$, and is hence an element of $V$.
Again let $p$ be a finite set of formulae of ${\cal L}^\lambda$. By $\Sigma^1_1$ absoluteness, if there is any outer model in which there is some $\Sigma$ which
certifies $p$, then there is some $\Sigma \in V^{{\rm Col}(\omega,\lambda)}$
which certifies $p$.\footnote{In fact, if $P$ is a transitive model of {\sf KP} plus the axiom
{\em Beta} with $(Q_\lambda; \langle A_{\bar \lambda} \colon {\bar \lambda} < \lambda \rangle ) \in P$ and if $p \in {\mathbb P}_\lambda$,
then there is some $\Sigma \in P^{{\rm Col}(\omega,\lambda)}$ which certifies $p$.} This simple observation is important in the verification
that ${\mathbb P}_\lambda$ is actually non-empty, cf.\ Lemma \ref{P_is_nonempty},
and in the proof of Lemma
\ref{P_stat_pres}.
It is easy to see that \begin{enumerate}
\item[(i)] ${\mathbb P}={\mathbb P}_\kappa \subset H_\kappa$,
\item[(ii)] if ${\bar \lambda} < \lambda$ are both in $C \cup \{ \kappa \}$, then ${\mathbb P}_{\bar \lambda} \subset {\mathbb P}_\lambda$, and
\item[(iii)] if $\lambda \in C \cup \{ \kappa \}$ is a limit point of
$C \cup \{ \kappa \}$, then ${\mathbb P}_\lambda = \bigcup_{{\bar \lambda} \in C \cap \lambda} \, {\mathbb P}_{\bar \lambda}$,
\end{enumerate}
so that there is some club $D \subset C$ such that for all $\lambda \in D$,
$${\mathbb P}_\lambda = {\mathbb P} \cap Q_\lambda.$$ Hence ($\Diamond$) gives us the following.
\begin{enumerate}\label{diamond-thing}
\item[($\Diamond({\mathbb P})$)] For all $B \subset H_\kappa$ the set
$$\{ \lambda \in C \colon (Q_\lambda;\in , {\mathbb P}_\lambda , A_\lambda) \prec (H_\kappa;\in , {\mathbb P} , B) \}$$
is stationary.
\end{enumerate}
The first one of the following lemmas is entirely trivial.
\begin{lemma}\label{+++}
Let $\Sigma$ be a syntactic certificate, and let $p$, $q \in [\Sigma]^{<\omega}$. Then $p$ and $q$ are compatible conditions in ${\mathbb P}$.
\end{lemma}
\begin{lemma}\label{P_is_nonempty}
Let $\lambda \in C \cup \{ \kappa \}$. Then $\emptyset \in {\mathbb P}_\lambda$.
\end{lemma}
{\sc Proof.} See the proof of \cite[Theorem 2.8]{david-ralf}, or the proof of
\cite[Theorem 4.2]{both}. Notice that for all $\lambda \in C \cup \{ \kappa \}$,
$\emptyset \in {\mathbb P}_\lambda$ iff $\emptyset \in {\mathbb P}$.
Let
$h$ be ${\rm Col}(\omega,\omega_2)$-generic over $V$, and write $\rho=\omega_3^V=
\omega_1^{V[h]}$. Inside $V[h]$, $$((H_{\omega_2})^V;\in ({\sf NS}_{\omega_1})^V , A)$$ is a ${\mathbb P}_{\rm max}$ condition, call it $p$.
Let $q \in ({\mathbb P}_{\rm max})^{V[h]}$, $q < p$, $q \in D^*$, cf.\ (D.1).
Let $q = N_0 = (N_0; \in , I , a)$.
Let $(M_i, \pi_{ij} \colon i \leq j \leq \omega_1^{N_0}) \in N_0$ be the unique generic iteration of $p$ which witnesses $q < p$.
Let $(N_i,\sigma_{ij} \colon i \leq j \leq \rho) \in V[h]$ be a generic iteration of
$N_0$ such that $\rho=\omega_1^{N_\rho}$.\footnote{If we wished, then we could even arrange that
writing $N_\rho = (N_\rho;\in,I^*,a^*)$, we have that $I^* = ({\sf NS}_\rho)^{V[h]}
\cap N_\rho$, but this is not relevant here; cf.\ footnote \ref{no-requirement}.} Let
\begin{eqnarray}\label{it_of_p}
(M_i , \pi_{ij} \colon i \leq j \leq \rho)=
\sigma_{0 \rho}((M_i, \pi_{ij} \colon i \leq j \leq \omega_1^{N_0}))
\end{eqnarray}
We may lift (\ref{it_of_p}) to a generic iteration
\begin{eqnarray}\label{it_of_p}
(M_i^+ , \pi^+_{ij} \colon i \leq j \leq \rho)
\end{eqnarray}
of $V$. Let us write $M=M^+_\rho$ and $\pi = \pi^+_{0 \rho}$.
Let $\langle k_n , \alpha_n \colon n<\omega \rangle$ be such that $x
= \langle k_n \colon n<\omega \rangle$ is
a real code for $N_0$ \`a la ($\Sigma$.5) and $\langle (k_n , \alpha_n) \colon n<\omega \rangle \in [T]$.
We then clearly have that $\langle (k_n , \pi(\alpha_n)) \colon n<\omega \rangle \in [\pi(T)]$.
It is now easy to see that
\begin{eqnarray}\label{certificate_at_rho}
{\mathfrak C} = \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \rho \rangle ,
\langle (k_n , \pi(\alpha_n) ) \colon n<\omega \rangle ,
\langle \rangle
\end{eqnarray}
certifies $\emptyset$, construed as the empty set of $\pi({\cal L}^\kappa)$ formulae: as the third
component $\langle \rangle$ of ${\mathfrak C}$ in (\ref{certificate_at_rho}) is empty, {\em any} set of surjections $e_i \colon \omega \rightarrow N_i$, $i<\omega_1$, will induce a syntactic certificate for $\emptyset$ whose associated sematic certificate is ${\mathfrak C}$. By $\Sigma^1_1$ absoluteness, there is then some
${\mathfrak C} \in M^{{\rm Col}(\omega,\pi(\omega_2))}$ as in (\ref{certificate_at_rho}) which certifies $\emptyset$ so that
$\emptyset \in \pi({\mathbb P})$. By the elementarity of $\pi$,
then, there is some ${\mathfrak C} \in V^{{\rm Col}(\omega,\omega_2)}$
which certifies $\emptyset$, construed as the empty set of ${\cal L}^\kappa$ formulae. Hence
$\emptyset \in {\mathbb P}$.
\hfill $\square$
\begin{lemma}\label{generics_give_certificates}
Let $\lambda \in C \cup \{ \kappa \}$.
Let $g \subset {\mathbb P}_\lambda$ be a filter such that $g \cap E \not= \emptyset$ for all dense $E \subset {\mathbb P}_\lambda$ which are definable over $(Q_\lambda;\in,{\mathbb P}_\lambda)$ from elements of $Q_\lambda$. Then
$\bigcup g$
is a syntactic certificate.
\end{lemma}
{\sc Proof.}
It is obvious how to read off from $\bigcup g$ a candidate
$${\mathfrak C} = \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle$$
like in (\ref{certificate}) for
a semantical certificate for $\bigcup g$.
Let us be somewhat explicit, though. A variant of what is to come shows
how to derive ${\mathfrak C}$ from a given syntactic certificate $\Sigma$,
where ${\mathfrak C}$ is unique such that $\Sigma$ certifies ${\mathfrak C}$, cf.\ the remark on p.\ \pageref{details}.
For
$i$, $j <\omega_1$ and
$\tau$, $\sigma \in \{ {\dot n} \colon n<\omega \} \cup \omega_1$
define
\begin{align*}
\tau \sim_{i} \sigma & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau = \sigma \urcorner \in \bigcup g \\
(i,\tau) \sim_{\omega_1} (j,\sigma) & \mbox{ \ \ \ iff } i \leq j \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ij}(\tau)=\rho \urcorner , \ulcorner {\dot N}_j \models
\rho = \sigma \urcorner \} \subset \bigcup g\\
{} & \mbox{ \ \ \ \ or } j \leq i \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ji}(\sigma)=\rho \urcorner , \ulcorner {\dot N}_j \models
\rho = \tau \urcorner \} \subset \bigcup g\\
[\tau]_{i} & = \{ \sigma \colon \tau \sim_{i} \sigma \} \\
[(i,\tau)] & = \{ (j,\sigma) \colon (i,\tau) \sim_{\omega_1} (j,\sigma) \} \\
{M}_i & = \{ [\tau]_i \colon \tau \in \{ {\dot n} \colon n<\omega \} \cup \omega_1 \wedge \ulcorner {\dot N}_i \models \tau \in {\dot M}_i \urcorner \in \bigcup g \} \\
M_{\omega_1} & = (H_{\omega_2})^V \\
{N}_i & = \{ [\tau]_i \colon \tau \in \{ {\dot n} \colon n<\omega \} \cup \omega_1 \} \\
N_{\omega_1} & = \{ [i,\tau]_{\omega_1} \colon i<\omega_1 \wedge \tau \in {\dot M}_i \urcorner \in \bigcup g \} \\
[\tau]_{i} \, {\tilde \in}_i \, [\sigma]_{i} & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau \in \sigma \urcorner \in \bigcup g \\
[i,\tau] {\tilde \in}_{\omega_1} [j,\sigma] & \mbox{ \ \ \ iff }
i \leq j \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ij}(\tau)=\rho \urcorner , \ulcorner {\dot N}_j \models
\rho \in \sigma \urcorner \} \subset \bigcup g\\
{} & \mbox{ \ \ \ \ or } j \leq i \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ji}(\sigma)=\rho \urcorner , \ulcorner {\dot N}_j \models
\tau \in \rho \urcorner \} \subset \bigcup g\\
[\tau]_i \in I^{N_i} & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau \in {\dot I} \urcorner \in \bigcup g \\
[i,\tau] \in I^{N_{\omega_1}} & \mbox{ \ \ \ iff \ \ } [\tau]_i \in I^{N_i} \\
[\tau]_i \in a^{N_i} & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau \in {\dot a} \urcorner \in \bigcup g \\
[i,\tau] \in a^{N_{\omega_1}} & \mbox{ \ \ \ iff \ \ } [\tau]_i \in I^{N_i} \\
\pi_{ij}([\tau]_{i}) = [\sigma]_j & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_j \models
{\dot \pi}_{ij}(\tau)=\sigma \urcorner \in \bigcup g \\
\pi_{i \omega_1}([\tau]_{i}) = x & \mbox{ \ \ \ iff \ \ } \ulcorner
{\dot \pi}_{i \omega_1}(\tau)=x \urcorner \in \bigcup g \\
\pi_{\omega_1 \omega_1}(x) = x & \mbox{ \ \ \ iff \ \ } x \in (H_{\omega_2})^V \\
\sigma_{ij}([\tau]_{i}) = [\sigma]_j & \mbox{ \ \ \ iff \ \ } \ulcorner
{\dot \sigma}_{ij}(\tau)=\sigma \urcorner \in \bigcup g \\
\sigma_{i \omega_1}([\tau]_i) = [i,\tau] & {} \\
(k,\alpha) = (k_n^g, \alpha_n^g) & \mbox{ \ \ \ iff } \exists {\vec u} \exists
{\vec \alpha} ( \ulcorner ({\vec u} , {\vec \alpha}) \in {\dot T} \urcorner
\in \bigcup g \wedge k = {\vec u}(n) \wedge \alpha = {\vec \alpha}(n) )\\
\delta \in K^g & \mbox{ \ \ \ iff } \exists {\bar \lambda} \, \ulcorner \delta \mapsto {\bar \lambda} \urcorner \in \bigcup g\\
{\bar \lambda} = \lambda^g_\delta & \mbox{ \ \ \ iff } \delta \in K^g \wedge \ulcorner \delta \mapsto {\bar \lambda} \urcorner \in \bigcup g\\
x \in X_\delta^g & \mbox{ \ \ \ iff } \delta \in K^g \wedge \ulcorner x \in {\dot X}_\delta \urcorner \in \bigcup g
\end{align*}
We will first have that
${\tilde \in}_0$ is wellfounded and that in fact (the transitive collapse of)
the structure $N_0 = (N_0;{\tilde \in}_0,a^{N_0},I^{N_0})$ is an iterable ${\mathbb P}_{\rm max}$ condition. This is true because
straightforward density arguments give (C.2), i.e., that $\langle (k_n, \alpha_n) \colon n<\omega \rangle \in [T]$ and $\langle k_n \colon
n<\omega \rangle$ will code the theory of $N_0$ \`a la ($\Sigma$.5).
Another set of easy density arguments will give that $(N_i ,\sigma_{ij} \colon i \leq j \leq \omega_1)$
is a generic iteration of $N_0$, were we identify $N_i$ with the structure
$(N_i;{\tilde \in}_i,a^{N_i},I^{N_i})$.
To verify this, let us first show:
\begin{claim}\label{claim2.4}
For each $i < \omega_1$ and for each $\xi \leq \omega_1^{N_i}$, $[\xi]_i$ represents
$\xi$ in (the transitive collapse of the well-founded part of) the term model for $N_i$;
moreover, $a^{N_i} = A \cap \omega_1^{N_i}$. Hence $a^{N_{\omega_1}}=A$.
\end{claim}
{\sc Proof} of Claim \ref{claim2.4}. Straightforward arguments using ($\Sigma$.1) show that $[\xi]_i$ must always represent $\xi$ in $N_i$ as given by any certificate.
Claim \ref{claim2.4} then follows by straightforward density arguments. \hfill $\square$ (Claim \ref{claim2.4})
\bigskip
Similarily:
\begin{claim} Let $i<\omega_1$.
$N_{i+1}$ is generated from ${\rm ran}(\sigma_{ii+1}) \cup \{ \omega_1^{N_i} \}$
in the sense that for every $x \in N_{i+1}$ there is some function $f \in {}^{\omega_1^{N_i}}
(N_i) \cap N_i$ such that $x = \sigma_{ii+1}(f)(\omega_1^{N_i})$.
\end{claim}
\begin{claim}
Let $i<\omega_1$. $\{ X \in {\cal P}(\omega_1^{N_i}) \cap N_i \colon \omega_1^{N_i} \in
\sigma_{ii+1}(X) \}$ is generic over $N_i$ for the forcing given by the $I^{N_i}$-positive sets.
\end{claim}
\begin{claim}
Let $i \leq \omega_1$ be a limit ordinal. For every $x \in N_i$ there is some $j<i$ and some ${\bar x} \in N_j$ such that $x = \sigma_{ji}({\bar x})$.
\end{claim}
$(N_i ,\sigma_{ij} \colon i \leq j \leq \omega_1)$
is then indeed a generic iteration of $N_0$. As $N_0$ is iterable, we may and shall identify $N_i$ with its transitive collapse, so that
(C.4) holds true.
Another round of density arguments will show that
${\mathfrak C}$
satisfies (C.1), (C.3), (C.5), (C.6), and (C.7), where we identify $M_i$
with the structure $(M_i;\in,({\sf NS}_{\omega_1^{M_i}})^{M_i},A \cap \omega_1^{M_i})$.
Let us now verify (C.8) and (C.9).
As for (C.8), $X_\delta \cap \omega_1 = \delta$ for $\delta \in K$ is easy. Let $x_1$, $\ldots$, $x_k
\in X_\delta$, $\delta \in K$.
Suppose that
\begin{eqnarray}\label{in-the-model}
(Q_{\lambda_\delta};\in {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \exists v \, \varphi(v,x_1,\ldots , x_k).
\end{eqnarray}
Let $p \in g$ be such that $\{ \ulcorner x_1 \in {\dot X}_\delta \urcorner , \ldots ,
\ulcorner x_k \in {\dot X}_\delta \urcorner , \ulcorner \delta \mapsto \lambda_\delta \urcorner
\} \subset p$. Let $q \leq p$, and let
$\Sigma$ be a syntactical certificate for $q$ whose associated semantical certificate is
$${\mathfrak C}' = \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X'_\delta \colon \delta \in K' \rangle.$$
Then $\delta \in K'$ and $$\{ x_1 , \ldots , x_k \} \subset X'_{\delta}
\prec (Q_{\lambda_\delta};\in {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}){\rm , }$$
so that by (\ref{in-the-model}) we may choose some $x \in X'_{\delta}$ with
$$(Q_{\lambda_\delta};\in {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \varphi(x,x_1,\ldots , x_k).$$ Let $r = q \cup \{
\ulcorner x \in {\dot X}_{\delta} \urcorner \}$.
By density, there is then some $y \in X_\delta$ such that $$(Q_{\lambda_\delta};\in {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \varphi(y,x_1,\ldots , x_k).$$
The proof of (C.9) is similar.
Let again $\delta \in K$. Let $E \subset {\mathbb P}_{\lambda_\delta} \cap X_\delta^g$ be dense in ${\mathbb P}_{\lambda_\delta} \cap X_\delta$,
and $r \in E$ iff $r \in {\mathbb P}_{\lambda_\delta} \cap X_\delta$
and
\begin{eqnarray}\label{defn_of_E}
(Q_{\lambda_\delta};\in {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \varphi(r,x_1, \ldots , x_k).
\end{eqnarray}
Let $p \in g$ be such that $\{ \ulcorner x_1 \in {\dot X}_\delta \urcorner , \ldots ,
\ulcorner x_k \in {\dot X}_\delta \urcorner , \ulcorner \delta \mapsto \lambda_\delta \urcorner
\} \subset p$. Let $q \leq p$, and again let
$\Sigma$ be a syntactical certificate for $q$ whose associated semantical certificate is
$${\mathfrak C}' = \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X'_\delta \colon \delta \in K' \rangle.$$
Then $[\Sigma]^{<\omega} \cap X'_\delta$ has an element, say $r$, such that (\ref{defn_of_E}) holds true. Let $s = q \cup r \cup \{ \ulcorner r \in {\dot X}_\delta \urcorner \}$.
By density, then, $g \cap X_\delta \cap E \not= \emptyset$. \hfill $\square$
\begin{lemma}\label{P_stat_pres}
Let $g$ be ${\mathbb P}$-generic over $V$. Let
$${\mathfrak C} = \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle$$
be the semantic certificate associated with the syntactic certificate $\bigcup g$.
Let
$$N_{\omega_1} = (N_{\omega_1};\in,A,I^*){\rm , }$$ and let
$T \in ({\cal P}(\omega_1) \cap N_{\omega_1}) \setminus I^*$.
Then $T$ is stationary in $V[g]$.
\end{lemma}
If ${\mathfrak C}$, $I^*$, and $T$ are as in the statement of Lemma \ref{P_stat_pres},
then by Lemma \ref{generics_give_certificates} and (C.3) and (C.4) we will have that
$({\sf NS}_{\omega_1})^V = I^* \cap V$, so that the conclusion of Lemma \ref{P_stat_pres} also gives that ${\mathbb P}$ preserves the stationarity of $T$. In other words:
\begin{corollary}
${\mathbb P}$ preserves stationary subsets of $\omega_1$.
\end{corollary}
{\sc Proof} of Lemma \ref{P_stat_pres}. Let ${\dot N}_{\omega_1} \in
V^{\mathbb P}$
be a canonical name for $N_{\omega_1}$, and let ${\dot I}^* \in
V^{\mathbb P}$
be a canonical name for $I^*$.
Let ${\bar p} \in g$, ${\dot C}$, ${\dot S} \in V^{\mathbb P}$, and
$i<\omega_1$ and $n<\omega$ be such that
\begin{enumerate}
\item[(i)] $T = {\dot S}^g$,
\item[(ii)] ${\bar p} \Vdash$ ``${\dot C} \subset \omega_1$ is club,''
\item[(iii)] ${\bar p} \Vdash$ ``${\dot S} \in ({\cal P}(\omega_1) \cap {\dot N}_{\omega_1}) \setminus {\dot I}^*$,'' and
\item[(vi)] ${\bar p} \Vdash$ ``${\dot S}$ is
represented by $[i,{\dot n}]$ in the term model producing ${\dot N}_{\omega_1}$.''
\end{enumerate}
We may and shall also assume that
\begin{eqnarray}\label{right-thing}
\ulcorner {\dot N}_i \models {\dot n} \mbox{ is a subset of the first uncountable cardinal, yet } {\dot n} \notin {\dot I} \urcorner \in {\bar p}.
\end{eqnarray}
Let $p \leq {\bar p}$ be arbitrary. We aim to produce some $q \leq p$ and some $\delta<\omega_1$ such that $q \Vdash {\check \delta} \in {\dot C} \cap {\dot S}$,
see Claim \ref{claim2.10} below.
For $\xi<\omega_1$, let $$D_\xi = \{ q \leq p \colon \exists \eta \geq \xi \, (\eta < \omega_1 \wedge q \Vdash {\check \eta} \in {\dot C}) \}{\rm, }$$ so that $D_\xi$ is open dense below $p$. Let $$E = \{ (q,\eta) \in {\mathbb P} \times \omega_1 \colon
q \Vdash {\check \eta} \in {\dot C} \}.$$ Let us write $$\tau = ((D_\xi \colon \xi<\omega_1),E).$$
We may and shall identify $\tau$ with some subset of $H_\kappa$ which codes $\tau$.
By ($\Diamond({\mathbb P})$), we may pick some
$\lambda \in C$ such that $p \in {\mathbb P}_\lambda$ and
\begin{eqnarray}\label{choice_of_lambda}
(Q_\lambda;\in,{\mathbb P}_\lambda,A_\lambda) \prec (H_\kappa;\in,{\mathbb P},\tau).
\end{eqnarray}
Let $h$ be ${\rm Col}(\omega,2^{\aleph_2})$-generic over $V$, and let
$g' \in V[h]$ be a filter on ${\mathbb P}_\lambda$ such that
$p \in g'$
and $g'$ meets every dense set which is definable over $(N_\lambda;\in,{\mathbb P}_\lambda,A_\lambda)$ from parameters in $N_\lambda$.
By Lemma \ref{generics_give_certificates}, $\bigcup g'$ is a syntactic certificate for $p$,
and we may let
$$\langle M_i' , \pi_{ij}' , N_i' , \sigma_{ij}' \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n' , \alpha_n' ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X_\delta' \colon \delta \in K' \rangle$$
be the associated semantic certificate. In particular, $K' \subset \lambda$.
Let $S$ denote the subset of $\omega_1$ which is represented by $[i,{\dot n}]$ in the term model giving $N_{\omega_1}'$, so that
if $N_{\omega_1}'=(N_{\omega_1}',\in,A,I')$, then by (\ref{right-thing}),
\begin{eqnarray}\label{S_positive}
S \in ({\cal P}(\omega_1) \cap N'_{\omega_1})
\setminus I'.
\end{eqnarray}
Let us also write $\rho=\omega_1^{V[h]} = (2^{\aleph_2})^{+V}$.
Inside $V[h]$, we may extend $\langle N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle$ to a generic iteration $$\langle N'_i , \sigma'_{ij} \colon i \leq j \leq \rho \rangle$$ such that
\begin{eqnarray}
\omega_1 \in \sigma'_{\omega_1,\omega_1+1}(S).
\end{eqnarray}
This is possible as $\omega_1^{N'_{\omega_1}} = {\rm sup} \{ \omega_1^{N_j} \colon j<\omega_1 \} = \omega_1$ and by (\ref{S_positive}).
Let $$\langle M'_i , \pi'_{ij} \colon i \leq j \leq \rho \rangle = \sigma_{0,\rho}(\langle M'_i , \pi'_{ij} \colon i \leq j \leq \omega_1^{N'_0} \rangle){\rm , }$$ so that $\langle M'_i , \pi'_{ij} \colon i \leq j \leq \rho \rangle$ is an extension of $\langle M'_i , \pi'_{ij} \colon i \leq j \leq \omega_1 \rangle$.
Recalling (\ref{Momega1}), we may lift $\langle M'_i , \pi'_{ij} \colon \omega_1 \leq i \leq j \leq \rho \rangle$ to a generic iteration $$\langle M_i^+ , \pi_{ij}^+ \colon \omega_1 \leq i \leq j \leq \rho \rangle$$ of $V$. Let us write $M=M_\rho^+$ and $\pi=\pi_{\omega_1,\rho}^+$.
The key point is now that $\langle M_i' , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \rho \rangle$ may be used to extend $\pi \mbox{''} \bigcup g'$ to a
syntactic certificate
\begin{eqnarray}\label{supset}
\Sigma \supset \pi \mbox{''} \bigcup g'
\end{eqnarray}
for $\pi(p)$
in the followig manner.
Let
$K^* = K' \cup \{ \omega_1 \}$. For $\delta \in K'$, let $\lambda_\delta^* =
\pi(\lambda_\delta')$ and $X_\delta^* =
\pi \mbox{''} X_\delta'$. Also, write $\lambda_{\omega_1}^* = \pi(\lambda)$ and $X_{\omega_1}^* = \pi \mbox{''} Q_\lambda$. Let
$${\mathfrak C}^* =
\langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \rho \rangle \mbox{, }
\langle (k'_n , \pi(\alpha'_n) ) \colon n<\omega \rangle \mbox{, }
\langle \lambda_\delta^* , X_\delta^* \colon \delta \in K^* \rangle.$$
It is then straightforward to verify that ${\mathfrak C}^*$ is a semantic certificate
for $\pi(p)$, and that in fact there is some syntactic certificate $\Sigma$ as in
(\ref{supset}) such that ${\mathfrak C}^*$ is certified by $\Sigma$.
Now let $[{\dot m}]_{\omega_1+1}$ represent $\sigma'_{\omega_1 \omega_1+1}(S)$ in the term model
for $N_{\omega_1+1}'$ provided by $\Sigma$, so that\footnote{Here, ${\dot \sigma}_{i \omega_1+1}$ and ${\dot N}_{\omega_1+1}$ are terms of the language associated with $\pi({\mathbb P}_\lambda)$ and $\ulcorner {\dot \sigma}_{i \omega_1+1}({\dot n}) = {\dot m} \urcorner$ and $
\ulcorner {\dot N}_{\omega_1+1} \models \omega_1 \in {\dot m} \urcorner$ are formulae of that language.}
$$\{ \ulcorner {\dot \sigma}_{i \omega_1+1}({\dot n}) = {\dot m} \urcorner ,
\ulcorner {\dot N}_{\omega_1+1} \models \omega_1 \in {\dot m} \urcorner \}
\subset \Sigma{\rm , }$$
in other words,
\begin{eqnarray}\label{half_of_new_p}
\pi(p) \cup \{ \ulcorner {\dot \sigma}_{i \omega_1+1}({\dot n}) = {\dot m} \urcorner ,
\ulcorner {\dot N}_{\omega_1+1} \models \omega_1 \in {\dot m} \urcorner \}
\mbox{ is certified by } \Sigma .
\end{eqnarray}
Let us now define
\begin{eqnarray}\label{defn_q}
q^* = \pi(p) \cup \{
\ulcorner {\dot \sigma}_{i \omega_1+1}({\dot n}) = {\dot m} \urcorner ,
\ulcorner {\dot N}_{\omega_1+1} \models \omega_1 \in {\dot m} \urcorner ,
\ulcorner \omega_1 \mapsto \pi(\lambda) \urcorner \}.
\end{eqnarray}
We thus established the following.
\begin{claim}\label{claim_1}
$q^* \in \pi({\mathbb P})$, as being certified by $\Sigma$.
\end{claim}
The elementarity of
$\pi \colon V \rightarrow M_\rho^+$ then gives some $\delta<\omega_1$ and some $\mu<\kappa$
such that
\begin{eqnarray}
q = p \cup \{ \ulcorner {\dot \sigma}_{i \delta+1}({\dot n}) = {\dot m} \urcorner ,
\ulcorner {\dot N}_{\delta+1} \models \delta \in {\dot m} \urcorner , \ulcorner \delta \mapsto \lambda \urcorner \} \in {\mathbb P}.
\end{eqnarray}
\begin{claim}\label{claim_2}\label{claim2.10}
$q \Vdash {\check \delta} \in {\dot C} \cap {\dot S}$.
\end{claim}
{\sc Proof} of Claim \ref{claim_2}. $q \Vdash {\check \delta} \in {\dot S}$ readily follows from $\{ \ulcorner {\dot \sigma}_{i \delta+1}({\dot n}) = {\dot m} \urcorner ,
\ulcorner {\dot N}_{\delta+1} \models \delta \in {\dot m} \urcorner \} \subset q$, the fact that ${\bar p} \geq p$ forces that ${\dot S}$ is
represented by $[i,{\dot n}]$ in the term model giving ${\dot N}_{\omega_1}$, and the fact that
by Claim \ref{claim2.4}, $[\delta]_{\delta+1}$ represents $\delta$ in the model $N_{\delta+1}$
of any semantic certificate for $q$.
Let us now show that $q \Vdash {\check \delta} \in {\dot C}$.
We will in fact show that $q$ forces that ${\check \delta}$ is a limit point of ${\dot C}$.
Otherwise
there is some $r \leq q$ and some $\eta < \delta$ such that
\begin{eqnarray}\label{21}
r \Vdash {\dot C} \cap {\check \delta} \subset {\check \eta}.
\end{eqnarray}
Let
\begin{eqnarray}\label{prime}
\langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_{\bar \delta} , X'_{\bar \delta} \colon {\bar \delta} \in K' \rangle
\end{eqnarray}
certify $r$. We must have that
\begin{enumerate}
\item[(a)] $\delta \in K'$,
\item[(b)] $X'_\delta \prec (Q_\lambda; \in , {\mathbb P}_\lambda , A_\lambda)$,
\item[(c)] $X'_\delta \cap \omega_1 = \delta$,
and
\item[(d)] for some $\Sigma$ such that the objects from (\ref{prime}) are certified by $\Sigma$, $[\Sigma]^{<\omega} \cap X'_\delta \cap
E \not= \emptyset$ for every $E \subset {\mathbb P}_{\lambda}$ which is dense in ${\mathbb P}_{\lambda} \cap
X'_\delta$ and definable over the structure
$$(Q_\lambda;\in, {\mathbb P}_\lambda , A_\lambda)$$
from parameters in $X'_\delta$.
\end{enumerate}
Notice that $A_\lambda = \tau \cap Q_\lambda$, and hence $A_\lambda$ may be identified
with $((D_\xi \cap Q_\lambda \colon \xi < \omega_1), E \cap Q_\lambda )$.
As $\eta < \delta \subset X_\delta'$, $D_\eta$ is definable over
the structure $$(Q_\lambda;\in, {\mathbb P}_\lambda , A_\lambda)$$
from a parameter in $X'_\delta$. By
(\ref{choice_of_lambda}), $D_\eta \cap Q_\lambda$ is dense in ${\mathbb P}_\lambda$.
By (d) above, there is then some
$s \in [\Sigma]^{<\omega} \cap X'_\delta \cap D_\eta \cap Q_\lambda$.
By (\ref{choice_of_lambda}) again, the unique smallest $\eta' \geq \eta$ with $s \Vdash {\check \eta}' \in {\dot C}$ must be in $X'_\delta$, hence $\eta' < \delta$ by (c) above.
By Lemma \ref{+++}, $s$ is compatible with $r$.
We have reached a contradiction with (\ref{21}).
\hfill $\square$
\section{Introduction.}
\begin{comment}
Cantor's Continuum Problem, which later became Hilbert's first Problem (see \cite{hilbert}), asks how many real numbers there are. This question got a non-answer through the discovery of the method of {\em forcing} by Paul Cohen: {\sf CH}, Cantor's Continuum Hypothesis, is independent from {\sf ZFC} (see \cite{cohen}), the standard axiom system for set theory which had been isolated by Zermelo and Fraenkel. {\sf CH} states that every uncountable set of reals has the same size as ${\mathbb R}$.
Ever since Cohen, set-theorists have been searching for natural new axioms which extend {\sf ZFC} and which settle the Continuum Problem. See e.g.\ \cite{hugh-ch1}, \cite{hugh-ch2}, \cite{peter}, and the discussion in
\cite{new?}. There are two prominent such axioms which decide {\sf CH} in the negative and which in fact both prove that there are $\aleph_2$ reals: Martin's Maximum ({\sf MM}, for short) or variants thereof on the one hand (see \cite{FMS}), and Woodin's axiom $(*)$ on the other hand (see \cite{hugh}). See e.g.\ \cite{menachem}, \cite{stevo}, and \cite{moore}.
Both of these axioms may be construed as maximality principles for the theory of the structure $(H_{\omega_2};\in)$, but up to this point the relationship between {\sf MM} and $(*)$ was a bit of a mystery, which led M.\ Magidor to call $(*)$ a ``competitor'' of {\sf MM} (\cite[p.\ 18]{menachem}). Both {\sf MM} and $(*)$ are inspired by and formulated in the language of forcing, and they both have ``the same intuitive motivation: Namely, the
universe of sets is rich'' (\cite[p.\ 18]{menachem}).
This paper resolves the tension between {\sf MM} and $(*)$ by proving that ${\sf MM}^{++}$, a strengthening of {\sf MM}, actually {\em implies} $(*)$, see Theorem \ref{main_thm} below, so that {\sf MM} and $(*)$ are actually compatible with each other.
This answers \cite[Question (18) a) on p.\ 924]{hugh}, see also \cite[p.\ 846]{hugh}, \cite[Conjecture 6.8 on p.\ 19]{menachem},
and \cite[Problem 14.7]{moore}.
\end{comment}
Cantor's Continuum Problem, which later became Hilbert's first Problem (see \cite{hilbert}), asks how many real numbers there are. After
having proved his celebrated theorem according to which ${\mathbb R}$ is uncountable, i.e., $2^{\aleph_0} > \aleph_0$, see \cite{cantor}, Cantor conjectured that every uncountable set of reals has the same size as ${\mathbb R}$, i.e., $2^{\aleph_0}=\aleph_1$. This statement is known as Cantor's Continuum Hypothesis ({\sf CH}). G\"{o}del \cite{goedel} proved in the 1930's that {\sf CH} is consistent with the standard axiom system for set theory, {\sf ZFC}, by showing that {\sf CH} holds in his constructible universe $L$, the minimal transitive model of {\sf ZFC} containing all the ordinals. The axiom $V=L$, saying that the universe $V$ of all sets is simply identical with $L$, has often been rejected, however, as an undesirable minimalistic assumption about $V$. For instance, $L$ cannot have measurable cardinals by a result of Scott \cite{scott}. G\"{o}del himself believed that {\sf CH} would be shown not to follow from {\sf ZFC}, and at least for part of his life he held the view that {\sf CH} is indeed false and that actually
\begin{eqnarray}\label{-ch}
2^{\aleph_0}=\aleph_2
\end{eqnarray}
(see \cite{godel_nonCH} and
\cite[pp.\ 173ff.]{goedel-feferman}).
In 1947, G\"odel \cite{what_is_ch} wrote:
\begin{quote}
[...] one may on good reason suspect that the role of the continuum problem in set theory will be this, that it will finally lead to the discovery of new axioms which will make it possible to disprove Cantor's conjecture.
\end{quote}
As we shall now try to explain, in the light of our unifying result, Theorem \ref{main_thm}, one could make the case that
with the two axioms ${\sf MM}^{++}$ and $(*)$, natural and strong such axioms
have already been found.
Luzin \cite{luzin} proposed a related hypothesis which also refutes {\sf CH}, namely
\begin{eqnarray}\label{luzin}
2^{\aleph_0}=2^{\aleph_1}.
\end{eqnarray}
That {\sf CH} does not follow from {\sf ZFC} was confirmed by Cohen in 1963 through the discovery of the method of {\em forcing}: Every model of {\sf ZFC} can be generically extended to a model of {\sf ZFC} in which {\sf CH} fails, see \cite{cohen}. In fact, using forcing one can show that it is relatively consistent with {\sf ZFC} that the cardinality of the continuum is $\aleph_1$, $\aleph_2$, $\aleph_{155}$, $\aleph_{\omega^2 + 17}$, or $\aleph_\alpha$ for many other values of $\alpha$, see \cite{solovay}.
\subsection{New axioms.}
Ever since Cohen's work, set-theorists have been searching for natural new axioms which extend {\sf ZFC} and which settle the Continuum Problem (see e.g.\ \cite{hugh-ch1}, \cite{hugh-ch2}, \cite{peter}, and the discussion in
\cite{new?}). One family of such axioms is the hierarchy of large cardinal axioms. It was realized early on, though, that these axioms cannot settle the Continuum Problem: one can always force {\sf CH} to hold or be false by small forcing notions, and all large cardinals which exist in $V$ will retain their large cardinal properties in the respective extensions, see \cite{l-s}.
One axiom which does settle the Continuum Problem is {\sf CH} itself; after all, {\sf CH} looks natural in that it gives the least possible value to $2^{\aleph_0}$ consistent with Cantor's theorem, $2^{\aleph_0}>\aleph_0$. {\sf CH} allows the ``diagonal'' construction of objects of size $\aleph_1$ with specific combinatorial properties, e.g.\ Luzin or Sierpi\'nski sets.
In 1985, Woodin proved his $\Sigma^2_1$ absoluteness result conditioned on {\sf CH}.
Namely, if {\sf CH} holds true, there is a proper class of measurable Woodin cardinals, and $\sigma$ is a statement of the form ``There is a set of reals $X$ such that $\varphi(X, r)$,'' where $r$ is a real and $\varphi(X)$ is a formula of set theory all of whose quantifiers are restricted to
reals, such that $\sigma$ can be forced over $V$, then $\sigma$ actually holds true in $V$, see e.g.\ \cite[Theorem 4.1]{ds2}.
Over the last decade, Woodin has developed a sophisticated scenario for set theory according to which {\sf CH} is true, see e.g.\ \cite{ICM} and \cite{midrasha}.
Nevertheless, and despite the appeal of $\Sigma^2_1$ absoluteness, {\sf CH} is often regarded as
a minimalistic assumption on a par with its parent, $V=L$.
To give an illustrative example,
under {\sf CH} one can easily find sets $X$ and $Y$ of reals without endpoints which are both $\aleph_1$-dense---in the sense that every interval of points
contains exactly $\aleph_1$ many points---but such that $X$ and $Y$ are not order-isomorphic. On the other hand,
by a theorem of Baumgartner \cite{baumgartner}, given any such $X$ and $Y$, there is a
nicely behaved forcing notion which adds an order-isomorphism between $X$ and $Y$. Thus, adopting {\sf CH}
precludes the existence of sufficiently generic filters for such forcing notions---which may consistently
exist.
A dual approach to {\sf CH} is to formulate axioms stipulating the existence of objects which may possibly exist, i.e.,
to look for ``maximality principles'' expressing some form of saturation of the universe of all sets with respect to its generic extensions. Such principles are known as \emph{forcing axioms}.
Shortly after the discovery of forcing, it was realized that it is possible to iterate the process of forming generic extensions $V \subset V[g_0] \subset V[g_1] \subset \ldots \subset V[g_\alpha]$ of $V$ in any length $\alpha$ in such a way that the final model is itself a generic extension of $V$.
By ``closing off'' one may then get to final models which are in fact saturated with respect to the existence of certain (partial) generics in the way prescribed by forcing axioms.
\subsection{Forcing axioms.}
Forcing axioms are generalizations of the Baire Category Theorem. Formally, they assert the existence of sufficiently generic filters for all members of some reasonably large class of forcing notions. In a general form, given an infinite cardinal $\kappa$ and a class $\mathcal K$ of forcing notions, the forcing axiom $\FA_\kappa(\mathcal K)$ is the statement that for every $\mathbb P\in\mathcal K$ and for every collection $\mathcal D$ of dense subsets of $\mathbb P$ such that $|\mathcal D|=\kappa$ there is a filter $g$ of $\mathbb P$ which is $\mathcal D$-generic (i.e., is such that $g \cap D \neq \emptyset$ for each $D\in\mathcal D$). $\FA_\kappa(\mathcal K)$ is to be seen as a maximality principle with respect to forceability via forcing notions from $\mathcal K$: If $\FA_\kappa(\mathcal K)$ holds, then all $\Sigma_1$ statements with parameters in $H_{\kappa^+}$ that can be forced to hold by some forcing notion in $\mathcal K$ already hold in the universe.\footnote{As a matter of fact, forcing axioms of the form $\FA_\kappa(\mathcal K)$ can
be fully characterized in terms of a suitable form of $\Sigma_1$-absoluteness with respect to generic extensions via members from $\mathcal K$ (see e.g.\ \cite{hugh} or \cite[Theorem 1.3]{cs2}).} The answers to questions about $H_{\kappa^+}$ provided by forcing axioms $\FA_\kappa(\mathcal K)$ are often regarded as being natural in that $\FA_{\kappa^+}(\mathcal K)$ offers a uniformly `saturated' picture of $H_{\kappa^+}$, ruling out the type of pathological objects that one can construct when $H_{\kappa^+}$ has an artificially constrained structure.
In what follows we will consider only forcing axioms $\FA_\kappa(\mathcal K)$ for $\kappa=\omega_1$.\footnote{This is the first $\kappa$ for which $\FA_\kappa(\mathcal K)$ does not follow outright from $\ZFC$. Also, $\kappa=\omega_1$ is the only level for which we currently have a reasonably complete picture of the available forcing axioms.} The first such forcing axiom shown to be consistent was Martin's Axiom at $\omega_1$, $\MA_{\omega_1}$, see \cite{solovay-tennenbaum} and \cite{martin-solovay}. $\MA_{\omega_1}$ is $\FA_{\omega_1}(\mathcal K)$, where $\mathcal K$ is the class of partial orders $\mathbb P$ with the countable chain condition (i.e., such that there is no uncountable family of pairwise incompatible conditions in $\mathbb P$).
Over the following years, a number of generalizations of $\MA_{\omega_1}$ were isolated. {\sf PFA}, the Proper Forcing Axiom, is
$\FA_{\omega_1}(\mathcal K)$, where $\mathcal K$ is the class of partial orders $\mathbb P$ which are proper.
The following is a list of examples of natural statement which are implied by forcing axioms.\label{list!}
\begin{itemize}
\item $\MA_{\omega_1}$ implies that there are no Suslin lines (\cite{solovay-tennenbaum}).
\item $\MA_{\omega_1}$ implies that every union of $\aleph_1$-many Lebesgue null subsets of reals is Lebesgue null (\cite{martin-solovay}).
\item $\MA_{\omega_1}$ implies the existence of a non-free Whitehead group (\cite{Shelah-Whitehead}).
\item {\sf PFA} implies Baumgartner's Axiom that all $\aleph_1$-dense sets of reals are order-isomorphic (essentially \cite{baumgartner}).\footnote{But it does not follow from $\MA_{\omega_1}$ (s.\ \cite{ARS}).}
\item {\sf PFA} implies Kaplansky's conjecture (\cite{todo}).
\item {\sf PFA} implies that there is a $5$-element basis for the class of uncountable linear orders (\cite{Moore5}).
\item {\sf PFA} implies that every automorphism of the Calkin algebra of a separable Hilbert space is inner (\cite{Farah}).
\end{itemize}
This line of research culminated in the proof by Foreman-Magidor-Shelah of the consistency of \emph{Martin's Maximum}, {\sf MM}, see \cite{FMS}.
Martin's Maximum is $\FA_{\omega_1}(\mathcal K)$, where $\mathcal K$ is the class of partial orders $\mathbb P$ such that forcing with $\mathbb P$ preserves the stationarity of all stationary subsets of $\omega_1$ in $V$. {\sf MM} is provably maximal in the sense that the forcing axiom $\FA_{\omega_1}(\{\mathbb P\})$ fails for any forcing notion $\mathbb P$ destroying some stationary subset $S$ of $\omega_1$.
At the same time, {\sf MM} can be forced by means of a forcing iteration ${\mathbb P} \subset V_\kappa$, assuming that $\kappa$ is a supercompact cardinal.
The natural such forcing ${\mathbb P}$ actually produces a model of a strengthening of {\sf MM}, called ${\sf MM}^{++}$. This is the statement that if $\mathbb P$ is a forcing notion preserving stationary subsets of $\omega_1$, $\mathcal D$ is a collection of size $\aleph_1$ consisting of dense subsets of $\mathbb P$, and $\{\tau_\alpha\,:\,\alpha<\omega_1\}$ is a collection of $\mathbb P$-names for stationary subsets of $\omega_1$, then there is a filter $g \subset \mathbb P$ which is $\mathcal D$-generic and which, furthermore, interprets every $\tau_\alpha$, $\alpha<\omega_1$, as a truly stationary set in $V$ (i.e., $\{\nu<\omega_1\,:\, \exists p \in g, \, p\Vdash_{\mathbb P} {\check \nu} \in\tau_\alpha \}$ is stationary for every $\alpha<\omega_1$).
Already $\MA_{\omega_1}$ contradicts {\sf CH}, and it even proves Luzin's hypothesis (\ref{luzin}), i.e., $2^{\aleph_0}=2^{\aleph_1}$. More interestingly, {\sf MM} (in contrast to $\MA_{\omega_1}$) decides the cardinality of the continuum,
and in fact it
confirms G\"odel's conjecture
(\ref{-ch}), $2^{\aleph_0}=\aleph_2$. This is shown by producing an affirmative answer to Friedman's Problem under {\sf MM}, see \cite[Theorems 9 and 10]{FMS}.\footnote{It was later verified by Moore that already the much weaker forcing axiom {\sf BPFA} implies (\ref{-ch}), see \cite{justin}.} ${\sf MM}^{++}$ is---by its very definition and the fact that no strictly stronger forcing axiom can be consistent---a prototype maximality principle for $V$. Remarkably, the empirical evidence seems to suggest that ${\sf MM}^{++}$ provides a complete theory of the initial segment $H_{\omega_2}$ of the universe of sets, at least with respect to natural questions.
Here, $H_{\omega_2}$ is the collection of all sets which are hereditarily of size $< \aleph_2$.
\begin{comment}
Under $\ZFC$, the universe can be stratified by the collection of sets of the form $H_\kappa$, for a cardinal $\kappa$, where $H_\kappa$ denotes the class of all sets whose transitive closure is of cardinality less than $\kappa$. Each $H_\kappa$ is a set, $H_\kappa\subseteq H_{\kappa'}$ whenever $\kappa<\kappa'$, and in $\ZFC$ every set is in some $H_\kappa$. Moreover, if $\kappa$ is an infinite regular cardinal, then $H_{\kappa}$ is itself a nice model of set theory to work in, as it satisfies all $\ZFC$ axioms except, possibly, for the Power Set Axiom. $H_{\omega_2}$ is the minimal level in this stratification of $V$ in which {\sf CH} can be expressed and, furthermore, $H_{\omega_2}$ is correct about {\sf CH} (i.e., {\sf CH} holds if and only if $(H_{\omega_2}; \in)\models\CH$).
\end{comment}
\subsection{The ${\mathbb P}_{\rm max}$ axiom $(*)$.}
There is another maximality principle, though, which Magidor called a ``competitor'' of {\sf MM}, see \cite[p.\ 18]{menachem}, and which is denoted by $(*)$. Its formulation involves the notion of ${\mathbb P}_{\rm max}$, a forcing which
was isolated by W.H.\ Woodin, see \cite[Definition 4.33]{hugh} and Definition \ref{defn_P_max} below. In much the same way as {\sf MM}, $(*)$ is inspired by and formulated in the language of forcing, and they both have ``the same intuitive motivation: Namely, the
universe of sets is rich'' (\cite[p.\ 18]{menachem}).
$(*)$, introduced by Woodin in \cite[Definition 5.1]{hugh}, is the conjunction of the following two statements.
\begin{enumerate}
\item[(i)]
{\sf AD}, the Axiom of Determinacy,\footnote{See e.g.\ \cite[Chapter 12]{book}.} holds in $L({\mathbb R})$, and
\item[(ii)] there is some $g$
which is ${\mathbb P}_{\rm max}$-generic over $L({\mathbb R})$ such that
${\cal P}(\omega_1) \subset L({\mathbb R})[g]$.
\end{enumerate}
Item (i), that ${\sf AD}$ holds in ${L(\mathbb R)}$, follows from the existence of large cardinals, e.g.\ from the existence of infinitely many Woodin cardinals with a measurable cardinal above them all, see \cite[p.\ 91]{proj-determinacy}. Item (ii) is the part of $(*)$ which goes beyond assuming the existence of large cardinals. $\mathbb P_{\max}$
arose out of earlier work by Steel-Van Wesep \cite{StVW} and by Woodin \cite{hugh83}
on the size of
\boldmath$\delta$\unboldmath${}^1_2$ and the question if ${\sf NS}_{\omega_1}$, the nonstationary ideal on $\omega_1$, can be saturated.
Here, ${\sf NS}_{\omega_1}$ is called saturated iff there is no collection ${\cal A}$ of stationary subsets of $\omega_1$ of size $\aleph_2$ such that $S \cap T$ is nonstationary for all $S$, $T \in {\cal A}$, $S \not= T$.
${\mathbb P}_{\rm max}$ consists of countable transitive structures, membership in ${\mathbb P}_{\rm max}$ is uniformly $\Pi^1_2$ in the codes, and the order $<_{{\mathbb P}_{\rm max}}$ is arithmetical. ${\mathbb P}_{\rm max}$ is $\omega$-closed and homogeneous, see \cite[Lemma 4.43]{hugh}.
The fact that a forcing ${\mathbb P}$ is homogeneous means that the validity in the forcing extension of a given statement is decided in the ground model by the trivial condition in ${\mathbb P}$.
The homogeneity of ${\mathbb P}_{\rm max}$ then yields that under $(*)$, the theory of $L({\mathcal P}(\omega_1))$ becomes part of the theory of $L({\mathbb R})$ in the sense that if $\varphi$ is any sentence, then
\begin{eqnarray}\label{part-of}
L({\mathcal P}(\omega_1)) \models \varphi \mbox{ if and only if } \Vdash_{L({\mathbb R})}^{{\mathbb P}_{\rm \max}} \varphi.
\end{eqnarray}
If {\sf AD} holds in ${L(\mathbb R)}$, then there is no well-order of the reals in ${L(\mathbb R)}$ (see
e.g.\ \cite[Lemma 12.2]{book}), but if $g$ is ${\mathbb P}_{\rm max}$-generic over ${L(\mathbb R)}$, then {\sf ZFC} is true in $L({\mathbb R})[g]$ (see \cite[Theorem 4.54]{hugh}), and moreover ${\sf NS}_{\omega_1}$ is saturated in $L({\mathbb R})[g]$ (see \cite[Theorem 4.50]{hugh})
and $L({\mathbb R})[g]$ provides an effective failure of {\sf CH} in that \boldmath$\delta$\unboldmath${}^1_2 = \omega_2$ is true in $L({\mathbb R})[g]$ (see \cite[Theorem 4.53]{hugh}).
Like the ``classical'' forcing axioms culminating with ${\sf MM}^{++}$, $(*)$ is also a maximality principle. While $(*)$ implies none of the stronger forcing axioms, see e.g.\ \cite[Theorem 1.3]{mlq0002d}, it does imply ${\sf MA}_{\omega_1}$. In particular, $(*)$ implies the first three implications of ${\sf MA}_{\omega_1}$ which are listed on p.\ \pageref{list!}. As it turns out, $(*)$ also implies that every automorphism of the Calkin algebra of a separable Hilbert space is inner, see \cite{paul!}, \cite{Farah}; and, at least in conjunction with the existence of a Woodin cardinal,\footnote{This is by the proof of Theorem \ref{Wo1}.} it also implies Baumgartner's Axiom on $\aleph_1$-dense sets of reals, as this can be expressed by a $\Pi_2$ sentence over $H_{\omega_2}$, as well as the existence of a $5$-element basis for the uncountable linear orders (since, in the presence of Baumgartner's Axiom and $\MA_{\omega_1}$, the existence of such a basis follows from every Aronszajn line containing a Countryman line, see \cite{Moore5}, which again can be expressed by a $\Pi_2$ sentence over $H_{\omega_2}$).
$(*)$ implies (and is in fact
equivalent to) what is dubbed ``$\Pi_2$ maximality.''
A sentence $\sigma$ (in the language of set theory, possibly augmented with some additional predicates) is said to be $\Pi_2$ if it is of the form $\forall x\exists y\varphi(x, y)$, with $\varphi(x, y)$ being a formula with only restricted quantifiers.
There is a whole
family of interesting statements which are $\Pi_2$ in the language for the structure
$$(H_{\omega_2}; \in, \NS_{\omega_1}){\rm , }$$
see e.g.\ the discussion in \cite{ds}.
The formulation of ``$\Pi_2$ maximality'' involves the concept of $\Omega$-logic, see
\cite[Section 10.4]{hugh}; for a sentence $\sigma$ to be ``$\Omega$-consistent'' is stronger than it just being consistent in that $\sigma$ needs to be true in models which are closed under arbitrarily complicated universally Baire operations, see \cite[Definition 10.144]{hugh}. The $\Pi_2$ maximality theorem, see \cite[Theorem 10.150]{hugh}, then runs as follows.
\begin{comment}
\begin{theorem}\label{Wo0} Suppose there is a proper class of Woodin cardinals. Suppose $\sigma$ is a $\Pi_2$ sentence, $\mathbb P$ is partial order, and $$\Vdash_{\mathbb P} (H_{\omega_2}; \in, \NS_{\omega_1})\models \sigma$$ Suppose $G$ is a $\mathbb P_{\max}$-generic filter over $L(\mathbb R)^V$. Then $$(H_{\omega_2}; \in, \NS_{\omega_1})^{L(\mathbb R)[G]}\models \sigma$$
\end{theorem}
As the following extension of Theorem \ref{Wo0} demonstrates, the axiom $(\ast)$, which on the face of its formulation has nothing to do with forcing axioms, can in fact be seen, in a broad sense and at least in the presence of large cardinals, as a forcing axiom.
\end{comment}
\begin{theorem}\label{Wo1} {\em \textbf{(Woodin)}} Suppose there is a proper class of Woodin cardinals. Then the following statements are equivalent.
\begin{enumerate}
\item[(1)] $(*)$.
\item[(2)]
Let $\sigma$ be a $\Pi_2$ sentence in the language for the structure
$$(H_{\omega_2}; \in, \NS_{\omega_1}, A \colon A \in {\mathcal P}({\mathbb R}) \cap L({\mathbb R})).$$
If $\sigma$ is $\Omega$-consistent, then $\sigma$ is true.
\end{enumerate}
\end{theorem}
\begin{comment}
(In the formulation of (2) in the theorem, $A^*$ denotes the canonical interpretation of the universally Baire set of reals $A$ in the generic extension by $\mathbb P$.) Hence, in the presence of large cardinals, $(\ast)$ implies, and in fact can be characterized as, a maximality principle, relative to $\Pi_2$-sentences and to arbitrary set-forcing extensions, for natural enrichments of the structure $(H_{\omega_2}; \in)$.
It is also easy to prove that, again in the presence of large cardinals, the axiom $(\ast)$ completely decides the theory of $L(\mathcal P(\omega_1))$ modulo forcing:
\begin{theorem}\label{abs} (Woodin) Suppose there is a proper class of Woodin cardinals and suppose $(\ast)$ holds. Suppose $\sigma$ is a sentence, $A$ is a universally Baire set of reals, $\mathbb P$ is a partial order forcing $(\ast)$, and $G$ is a $\mathbb P$-generic filter over $V$. Then $$(L(\mathcal P(\omega_1)); \in, A)\models\sigma$$ if and only if $$(L(\mathcal P(\omega_1))^{V[G]}; \in, A^*)\models\sigma$$
\end{theorem}
The proof of Woodin \ref{abs} is straightforward, using the fact that $\mathbb P_{\max}$ is a definable forcing, over $L(\mathbb R)$, which is weakly homogeneous---in particular, given any sentence $\sigma$ with ground model sets as parameters, all conditions in $\mathbb P_{\max}$ decide the truth value of $\psi$ in the same way. Hence, the fact that $\sigma$ holds in the $(L(\mathcal P(\omega_1)); \in, A)$ of some generic extension of $L(\mathbb R)$ by $\mathbb P_{\max}$ is equivalent to a sentence $\psi$ holding in $(L(\mathbb R); \in, A)$. But, by a theorem of Woodin, the theory of the inner model $(L(\mathbb R); \in, A)$ is invariant under set-forcing if there is a proper class of Woodin cardinals.
\end{comment}
One specific instance of $\sigma$ in (2) of Theorem \ref{Wo1} is called $\psi_{\rm AC}$, see \cite[Definition 5.12]{hugh}. It is in spirit a
local version of an affirmative solution to Friedman's Problem. Woodin showed, see \cite[Theorem 5.14, Lemmata 5.15 and 5.18]{hugh}, that $\psi_{\rm AC}$
follows from both {\sf MM} and $(*)$ and that $\psi_{\rm AC}$ implies
G\"odel's conjecture (\ref{-ch}), i.e.\ $2^{\aleph_0}=\aleph_2$.
The homogeneity of ${\mathbb P}_{\rm max}$ gives that
in the presence of large cardinals,
$(*)$ yields a complete theory for $L({\cal P}(\omega_1))$ modulo set-forcing: by (\ref{part-of}), all set-generic extensions of $V$ in which $(*)$ holds true agree on the theory of $L({\cal P}(\omega_1))$.
\begin{comment}
We have thus seen that $(\ast)$ is, in the presence of large cardinal axioms, a natural axiom: It is a strong maximality principle for (enrichments of) the structure $(H_{\omega_2}; \in)$ under arbitrary set-forcing, it provides a complete theory for the inner model $L(\mathcal P(\omega_1))$---and therefore in particular for $H_{\omega_2}$---modulo forcing and it yields the `right' value of $2^{\aleph_0}$ from the point of view of classical forcing axioms, namely $\aleph_2$. However, i
\end{comment}
\begin{comment}
There is a strengthening of ${\sf MM}^{++}$, isolated by Viale in \cite{viale}, which has strong completeness properties modulo forcing similar to those of $(*)$. This is the axiom ${\sf MM}^{+++}$. It says that a class $\mathbb T$ of towers of ideals with certain nice structural properties is dense in the category of stationary set preserving forcings; in other words, for every stationary set preserving forcing $\mathcal P$ there is a towers $\mathcal T$ in $\mathbb T$ such that $\mathcal P$ completely embeds in $\mathcal T$ in such a way that the quotient forcing preserves stationary sets in $V^{\mathcal P}$. This axiom implies ${\sf MM}^{++}$, if $\kappa$ is an almost super-huge cardinal, then there is a partial order $\mathcal P\subseteq V_\kappa$ which forces ${\sf MM}^{++}$, and if there is a proper class of almost super-huge cardinals, then ${\sf MM}^{+++}$ is complete for the theory of the $\omega_1$-Chang model\footnote{The $\omega_1$-Chang model is the $\subseteq$-minimal transitive model of ${\sf ZF}$ containing all ordinals and closed under $\omega_1$-sequences. It can be construed as $\bigcup_{\alpha\in \Ord}L([\alpha]^{\aleph_1})$ and it includes $L(\mathcal P(\omega_1))$ as a definable submodel.} with respect to stationary set preserving partial orders forcing ${\sf MM}^{+++}$.
Thus, the completeness obtained from ${\sf MM}^{+++}$ relative to generic extensions is for an inner model larger than $L(\mathcal P(\omega_1))$, namely the $\omega_1$-Chang model, but only for forcing notions preserving stationary sets. So, from the point of completeness modulo forcing, ${\sf MM}^{+++}$ is similar to $(*)$, but also somewhat orthogonal to it.
\end{comment}
Despite its nice properties, in order for $(*)$ to be a convincing candidate for a natural axiom, it would
have to be compatible with all consistent large cardinal axioms. While $L(\mathbb R)[g]$
is trivially a model of
$(*)$, provided that $g$ is ${\mathbb P}_{\rm max}$-generic over $L(\mathbb R)$,
Scott's result \cite{scott} carries over from $L$ to
$L(\mathbb R)[g]$ and shows that $L(\mathbb R)[g]$ cannot have measurable cardinals either.
\cite{hugh} and subsequent work left open the problem whether $(*)$ would be compatible with large cardinals beyond the level of Woodin cardinals.
\subsection{Unifying forcing axioms and $(*)$.}
Prior to the current paper, the relation between classical forcing axioms like {\sf MM},
which could be forced by iterated forcing over models of {\sf ZFC} with large cardinals, and the axiom $(*)$, whose models were obtained by forcing over models satisfying
the Axiom of Determinacy,
remained a complete mystery. It had been known by a result of P.\ Larson \cite{plarson} that even ${\sf MM}^{+\omega}$, an axiom strictly between {\sf MM} and ${\sf MM}^{++}$, does not imply $(*)$.\footnote{${\sf MM}^{+\omega}$ is the strengthening of ${\sf MM}$ obtained by replacing, in the formulation of ${\sf MM}^{++}$, collections of $\aleph_1$ many names for stationary sets with collections of only countably many such names.} One can build models of ${\sf MM}^{+\omega}$ with a well-order of $H_{\omega_2}$ which is definable over $(H_{\omega_2}; \in)$ by a formula without parameters, and the existence of such a well-order is incompatible with $(*)$ by the homogeneity of ${\mathbb P}_{\rm max}$. It remained even unclear whether classical strong forcing axioms would be compatible at all with $(*)$, see
\cite[p.\ 846]{hugh}. See also \cite[Question (18) a) on p.\ 924]{hugh}, \cite[Conjecture 6.8 on p.\ 19]{menachem},
and \cite[Problem 14.7]{moore}.
This paper resolves the tension between {\sf MM} and $(*)$. We prove:
\begin{theorem}\label{main_thm}
Assume {\em Martin's Maximum}${}^{++}$. Then Woodin's ${\mathbb P}_{\rm max}$-axiom $(*)$ holds true.
\end{theorem}
In particular, {\sf MM} and $(*)$ are compatible with one another, and
$(*)$ is compatible with all consistent large cardinal axioms: If $\kappa$ is a supercompact cardinal and ${\mathbb P} \subset V_\kappa$ is the partial order from
\cite{FMS}
to force
${\sf MM}^{++}$, then by Theorem \ref{main_thm} the axiom $(*)$ holds in $V^{\mathbb P}$, and all the large cardinals of $V$ above $\kappa$ are preserved by ${\mathbb P}$.
Theorem \ref{main_thm} renders ${\sf MM}^{++}$ a particularly appealing axiom. Not only is ${\sf MM}^{++}$ a provably maximal forcing axiom providing the `right' answers to questions pertaining to $H_{\omega_2}$,\footnote{The `right' answers from a conception of the universe as being uniformly saturated with respect to forcing.
} but it follows from Theorem \ref{main_thm} that ${\sf MM}^{++}$ implies the form of $\Pi_2$ maximality for arbitrary set-forcing given by (2) of Theorem \ref{Wo1} and, moreover, that ${\sf MM}^{++}$ completely decides the theory of $L(\mathcal P(\omega_1))$ via set-forcing.
It also follows from Theorem \ref{main_thm}
that $(*)$ can be characterized, in the presence of large cardinals, by a statement which on the face of its formulation is weaker than (2) of Theorem \ref{Wo1}. We will prove the following theorem at the end of the next section.
\begin{theorem}\label{char_of_star} Suppose there is a supercompact cardinal. Then the following statements are equivalent.
\begin{enumerate}
\item[(1)] $(*)$.
\item[(2)]
Let $\sigma$ be a $\Pi_2$ sentence in the language for the structure
$$(H_{\omega_2}; \in, \NS_{\omega_1}, A \colon A \in {\mathcal P}({\mathbb R}) \cap L({\mathbb R})).$$
If there is a stationary set preserving forcing ${\mathbb P}$ such that $\sigma$
holds in $V^{\mathbb P}$, then $\sigma$ is true in $V$.
\end{enumerate}
\end{theorem}
This equivalence of $(*)$ is a variant of one which we are going to state below, see Theorem
\ref{char_star}, and which characterizes $(*)$ as a strong version of a bounded forcing axiom.
\begin{comment}
Cantor's Continuum Problem, which later became Hilbert's first Problem (see \cite{hilbert}), asks how many real numbers there are. This question got a non-answer through the discovery of the method of {\em forcing} by Paul Cohen: {\sf CH}, Cantor's Continuum Hypothesis, is independent from {\sf ZFC} (see \cite{cohen}), the standard axiom system for set theory which had been isolated by Zermelo and Fraenkel. {\sf CH} states that every uncountable set of reals has the same size as ${\mathbb R}$.
Ever since Cohen, set-theorists have been searching for natural new axioms which extend {\sf ZFC} and which settle the Continuum Problem. See e.g.\ \cite{hugh-ch1}, \cite{hugh-ch2}, \cite{peter}, and the discussion in
\cite{new?}. There are two prominent such axioms which decide {\sf CH} in the negative and which in fact both prove that there are $\aleph_2$ reals: Martin's Maximum ({\sf MM}, for short) or variants thereof on the one hand (see \cite{FMS}), and Woodin's axiom $(*)$ on the other hand (see \cite{hugh}). See e.g.\ \cite{menachem}, \cite{stevo}, and \cite{moore}.
Both of these axioms may be construed as maximality principles for the theory of the structure $(H_{\omega_2};\in)$, but up to this point the relationship between {\sf MM} and $(*)$ was a bit of a mystery, which led M.\ Magidor to call $(*)$ a ``competitor'' of {\sf MM} (\cite[p.\ 18]{menachem}). Both {\sf MM} and $(*)$ are inspired by and formulated in the language of forcing, and they both have ``the same intuitive motivation: Namely, the
universe of sets is rich'' (\cite[p.\ 18]{menachem}).
This paper resolves the tension between {\sf MM} and $(*)$ by proving that ${\sf MM}^{++}$, a strengthening of {\sf MM}, actually {\em implies} $(*)$, see Theorem \ref{main_thm} below, so that {\sf MM} and $(*)$ are actually compatible with each other.
This answers \cite[Question (18) a) on p.\ 924]{hugh}, see also \cite[p.\ 846]{hugh}, \cite[Conjecture 6.8 on p.\ 19]{menachem},
and \cite[Problem 14.7]{moore}.
\end{comment}
The authors thank Ilijas Farah, Andreas Lietz, and Matteo Viale for their comments on earlier versions of this paper. They would also like to thank the anonymous referees for their reports,
which have been
of extraordinary quality and most helpful. Finally, they would like to thank Andreas Lietz for drawing the beautiful diagrams.
The reader should have some acquaintance with forcing, determinacy, and universally Baire sets of reals. The relevant material is covered e.g.\ in \cite[Chap.\ 6 and sections 7.1, 8.1, and 12.1]{book}. Familiarity with stationary set preserving forcings and Martin's Maximum, see e.g.\ \cite{FMS} or \cite[Chap.\ 37]{jech}, and with ${\mathbb P}_{\rm max}$ forcing to the extent of say \cite[Chap.\ 4]{hugh} or \cite[Sections 1-6]{larson-handbook} would be desirable. Knowledge of forcings which are similar to the one which will be designed here and which were developed earlier e.g.\ in \cite{ronald_l_forcing}, \cite{cs}, \cite{ds}, or \cite{ds2} is by no means required or presupposed.
\section{Preliminaries.}\label{section2}
Let us first state again
{\em Martin's Maximum}${}^{++}$, abbreviated by {\sf MM}${}^{++}$ and isolated by Foreman-Magidor-Shelah \cite{FMS} (cf.\ also \cite[Definition 2.45 (2)]{hugh}).
\begin{definition}
{\sf MM}${}^{++}$ is the statement that if ${\mathbb P}$ is a forcing which preserves stationary subsets of $\omega_1$, if $\{ D_i \colon
i<\omega_1 \}$ is a collection of dense subsets of ${\mathbb P}$, and if $\{ \tau_i
\colon i<\omega_1 \}$ is a collection of ${\mathbb P}$-names for stationary subsets
of $\omega_1$, then there is a filter $g \subset {\mathbb P}$ such that
for every $i<\omega_1$,
\begin{enumerate}
\item[(i)]
$g \cap
D_i \not= \emptyset$ and
\item[(ii)] $(\tau_i)^g = \{ \xi<\omega_1
\colon \exists p \in g \, p \Vdash_{\mathbb P} {\check \xi} \in \tau_i \}$ is stationary.
\end{enumerate}
\end{definition}
The forcing ${\mathbb P}_{\rm max}$ was designed by W.\ Hugh Woodin, see \cite[Chapter 4]{hugh}, specifically \cite[Definition 4.33]{hugh}.
In order to define ${\mathbb P}_{\rm max}$, we need the notion of
``generic iterability'' of structues of the form $(M; \in, I,a)$, where $M$ is a transitive model of a sufficiently large fragment of ${\sf ZFC}$, $(M;I)$ is amenable,\footnote{I.e., $x \cap I \in M$ for all $x \in M$.} $a \subset \omega_1^M$,
and $(M;I) \models$ ``$I$ is a normal uniform ideal on $\omega_1$.'' Given an ordinal $\gamma\leq\omega_1$, $\langle \langle (M_i; \in, I_i,a_i)
\colon i \leq \gamma \rangle, \langle \pi_{i,j} \colon i \leq j \leq \gamma \rangle , \langle g_i \colon i < \gamma \rangle \rangle$ is a generic iteration of $(M; \in, I, a)$ if the following hold true.
\begin{itemize}
\item $(M_0;\in,I_0,a_0)=(M;\in,I,a)$,
\item for $i<\gamma$, $g_i$ is a $\mathcal P(\omega_1)^{M_\alpha}\setminus I_i$-generic filter over $M_i$, $M_{i+1}$ is the ultrapower of $M_i$ by $g_i$, and $\pi_{i,i+1} \colon (M_i; \in, I_i,a_i)\to (M_{i+1}; \in, I_{i+1},a_{i+1})$ is the corresponding generic elementary embedding,
\item $\pi_{i, k}=\pi_{j,k}\circ \pi_{i,j}$ for all $i \leq j \leq k$, and
\item if $\beta$ is a nonzero limit ordinal $\leq \gamma$, then $(M_\beta , (\pi_{i, \beta} \colon i<\beta))$ is the direct limit of $(M_i , \pi_{i, j}
\colon i \leq j<\beta)$.
\end{itemize}
\noindent $(M;\in,I,a)$ being generically iterable means that all models in any generic iteration of $(M; \in, I,a)$ are well-founded, irrespective of the filters $g_\alpha$ chosen at any stage $\alpha$,
see \cite[Definition 4.1]{hugh}.
Let us stress that the current paper will only consider such generic iterations rather than iterations of mice as being studied in inner model theory.
\begin{definition}\label{defn_P_max}
The conditions in ${\mathbb P}_{\rm max}$ are countable transitive models of a sufficiently large fragment of ${\sf ZFC}$ plus ${\sf MA}_{\omega_1}$ of the form $(M;\in,I, a)$, where
\begin{enumerate}
\item[(i)] $(M;I)$ is amenable
and $(M;I) \models$ ``$I$ is a normal uniform ideal on $\omega_1$,''
\item[(ii)] $a \in {\cal P}(\omega_1^M) \cap M$ and $M \models$ ``$\omega_1 =
\omega_1^{L[a,x]}$ for some real $x$,'' and
\item[(iii)] $(M;\in,I)$ is generically iterable.
\end{enumerate}
We construe ${\mathbb P}_{\rm max}$ as a partial order by declaring that $(N;\in,J,b)$ is stronger than $(M;\in,I,a)$, denoted by $(N;\in,J,b) < (M;\in,I,a)$, if and only if $(M;\in,I,a) \in N$ and inside $N$ there is a generic iteration of $(M;\in,I,a)$ of length $\omega_1^N+1$ with last model $(M^*;\in,I^*,a^*)$ such that $I^* = J \cap M^*$ and $a^* = b$.\footnote{${\mathbb P}_{\rm max}$ according to our Definition \ref{defn_P_max} is a slightly bigger poset than the one according to
\cite[Definition 4.33]{hugh}. The difference is that we weakened the requirement $I \in M$ of \cite[Definition 4.33]{hugh} to ``$(M;I)$ is amenable.'' This natural move will make $(H_{\omega_2};\in,{\sf NS}_{\omega_1},A)$ for any $A \subset \omega_1$ a ${\mathbb P}_{\rm max}$ condition in a generic extension where $H_{\omega_2}$ is countable, cf.\ (\ref{its-a-Pmax-condition!}). It is easy to see that ${\mathbb P}_{\rm max}$ according to \cite[Definition 4.33]{hugh} is dense in ${\mathbb P}_{\rm max}$ according to our Definition \ref{defn_P_max}, so that both forcing notions are forcing-equivalent.}
\end{definition}
Most of \cite{hugh} studies the effect of forcing with ${\mathbb P}_{\rm max}$ or variants thereof over a model of the Axiom of Determinacy.
Let us state again
Woodin's ${\mathbb P}_{\rm max}$ axiom $(*)$, see \cite[Definition 5.1]{hugh}.
\begin{definition}
$(*)$ says that
\begin{enumerate}
\item[(i)]
{\sf AD} holds in $L({\mathbb R})$ and
\item[(ii)] there is some $g$
which is ${\mathbb P}_{\rm max}$-generic over $L({\mathbb R})$ such that
${\cal P}(\omega_1) \subset L({\mathbb R})[g]$.
\end{enumerate}
\end{definition}
Already the {\em Proper Forcing Axiom}, {\sf PFA}, which is much weaker than {\sf MM}${}^{++}$, implies ${\sf AD}^{L({\mathbb R})}$ and much more, see \cite{john}, \cite{stacking}, and \cite[Chapter 12]{trang-sargsyan}.
The current paper produces a proof of Theorem \ref{main_thm}.
Our key new idea is ($\Sigma$.8) on page \pageref{(C.9)} below.
We try to give an overview of the proof of Theorem \ref{main_thm} at the end of this section.
Theorem \ref{main_thm} is optimal in that P.\ Larson \cite{plarson} and \cite{plarson2} has shown that ${\sf MM}^{+\omega}$ is consistent with $\lnot (*)$
relative to a supercompact limit of supercompact
cardinals.
Our proof is also
optimal
in that the forcing which we will use to verify Theorem \ref{main_thm} has size
$2^{\aleph_2}$, while Woodin has shown that ${\sf MM}^{++}$ for forcings
of size $2^{\aleph_0} = \aleph_2$ does not imply $(*)$, see \cite[Theorem 10.90]{hugh}, and it is consistent with ${\sf MM}^{++}$ that $2^{\aleph_2}=
\aleph_3$.
Throughout our entire paper, ``$\omega_1$''
will {\em always} denote $\omega_1^V$, the $\omega_1$ of $V$. We shall also make permanent use of the following.
\begin{convention}\label{convention}
Let us fix throughout this paper some $A \subset \omega_1$ such that $\omega_1^{L[A]} = \omega_1$. Let us
define $g_A$ as the set of all ${\mathbb P}_{\rm max}$ conditions $p=(N;\in,I,a)$
such that there is a generic iteration $$(N_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = N_0$ of length $\omega_1+1$ such that if we write $N_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$,\footnote{Here and elsewhere we often confuse a model with its underlying universe.} then $I^* = ({\sf NS}_{\omega_1})^V \cap N_{\omega_1}$ and $a^* = A$.
\end{convention}
In the following statement, $X^\#$ denotes the sharp of $X$. While the formal definition of a sharp (see e.g. \cite[Section 10.2]{book}) won't play any role in what follows, the reader may think of ``${\cal P}(\omega_1)^\#$ exists'' as just some extra large cardinal structure which is assumed to be present in the universe.
We are going to use now the concept of elementary substructures.
For any two models ${\cal M}$ and ${\cal N}$ with underlying universes $M$ and $N$, respectively, and with the same first order language associated to them,
${\cal M} \prec {\cal N}$ means that ${\cal M}$ is an elementary substructure of ${\cal N}$, i.e., $M \subset N$ and for all formulae $\varphi$ of that common language and
all $x_1$, $\ldots$, $x_k \in M$,
\begin{eqnarray}\label{ee}
{\cal M} \models \varphi(x_1,\ldots,x_k) \Longleftrightarrow
{\cal N} \models \varphi(x_1,\ldots,x_k).
\end{eqnarray}
\begin{lemma}\label{folklore} \textbf{\em (Woodin)} Assume ${\sf MA}_{\omega_1}$, that ${\sf NS}_{\omega_1}$ is saturated, and that ${\cal P}(\omega_1)^\#$ exists. In the notation of Convention \ref{convention}:
\begin{enumerate}
\item[(1)] $g_A$ is a filter.
\item[(2)] If $g_A$ is ${\mathbb P}_{\rm max}$-generic over $L({\mathbb R})$,
then ${\cal P}({\omega_1}) \subset L({\mathbb R})[g]$.
\end{enumerate}
\end{lemma}
{\sc Proof.} This routinely follows from the proof of \cite[Lemma 3.12 and Corollary 3.13]{hugh} and from \cite[Lemmas 3.10 and 3.14]{hugh}. Let us sketch the argument. Let us first state the following.
\begin{claim}\label{claim_2.6}
Let $\theta \geq (2^{\aleph_1})^+$ be a cardinal, let $X \prec H_\theta$ be countable with $A \in X$, and let $\sigma \colon M \cong X$ be such that $M$ is transitive. Then
\begin{enumerate}
\item[(a)] $\sigma^{-1}((H_{\omega_2};\in, {\sf NS}_{\omega_1} , A))$ is a ${\mathbb P}_{\rm max}$-condition.
\item[(b)] $\{ X \in {\cal P}(\omega_1^M) \cap M \colon \omega_1 \in \sigma(X) \}$ is $({\cal P}(\omega_1^M) \cap M) \setminus \sigma^{-1}({\sf NS})$-generic over $M$.
\item[(c)] For $i \leq \omega_1$ let $$X_i = {\rm Hull}^{H_{\theta}}(X \cup \sup \{ X_j \colon j<i \}) {\rm , }$$ let $\sigma_i \colon M_i \cong X_i$ be such that $M_i$ is transitive, and let, for $i \leq j \leq \omega_1$, $\pi_{ij} = \sigma^{-1}_j \circ \sigma_i$. Then $(M_i,\pi_{ij} \colon i \leq j \leq \omega_1)$ is a generic iteration of $$(M;\in,\sigma^{-1}({\sf NS}_{\omega_1}),A \cap \omega_1^M)$$
\item[(d)] $\sigma^{-1}((H_{\omega_2};\in, {\sf NS}_{\omega_1} , A)) \in g_A$.
\end{enumerate}
\end{claim}
{\sc Proof} of Claim \ref{claim_2.6}. (a): This is by \cite[Lemmas 3.10 and 3.14]{hugh}. (b) and (c): This is by \cite[Lemma 3.12 and Corollary 3.13]{hugh}. (d): This follows immediately from (a) and (c). \hfill $\square$ (Claim \ref{claim_2.6})
\medskip
Let us now prove Lemma \ref{folklore}.
(1): Let ${\cal N}_0 = (N;\in,I,a) \in g_A$ as being witnessed by the generic iteration
$({\cal N}_i , \sigma_{ij} \colon i \leq j \leq \omega_1)$. Let ${\cal M}_0 > {\cal N}_0$ as being witnessed by the generic iteration $({\cal M}_i , \pi_{ij} \colon i \leq j \leq \omega_1^{{\cal N}_0}) \in N$. Then $\sigma_{0 \omega_1}(({\cal M}_i , \pi_{ij} \colon i \leq j \leq \omega_1^{{\cal N}_0}))$ is easily seen to be a generic iteration of ${\cal M}_0$ which witnesses that ${\cal M}_0 \in g_A$.
Now let ${\cal N}_0^0 = (N^0;\in,I^0,a^0) \in g_A$ and ${\cal N}_0^1 = (N^1;\in,I^1,a^1) \in g_A$ as witnessed by the generic iterations
${\cal I}^0 = ({\cal N}_i^0 , \sigma_{ij}^0 \colon i \leq j \leq \omega_1)$ and
${\cal I}^1 = ({\cal N}_i^1 , \sigma_{ij}^1 \colon i \leq j \leq \omega_1)$. Let $\sigma \colon M \cong X$ be as in Claim \ref{claim_2.6} with $\{ {\cal I}^0 , {\cal I}^1 \} \subset X$. Then ${\cal N}^0$ and ${\cal N}^1$ are both weaker than $\sigma^{-1}((H_{\omega_2};\in, {\sf NS}_{\omega_1} , A)) \in {\mathbb P}_{\rm max}$, cf.\ Claim \ref{claim_2.6} (a), as witnessed by the generic iterations $\sigma^{-1}({\cal I}^0) = ({\cal N}_i^0 , \sigma_{ij}^0 \colon i \leq j \leq \omega_1^M)$ and $\sigma^{-1}({\cal I}^1) = ({\cal N}_i^1 , \sigma_{ij}^1 \colon i \leq j \leq \omega_1^M)$.
(2): Let $Z \in {\cal P}(\omega_1)$. Let $\sigma \colon M \cong X$ be as in Claim \ref{claim_2.6} with $Z \in X$. Then $\sigma^{-1}((H_{\omega_2};\in, {\sf NS}_{\omega_1} , A)) \in g_A$ by Claim \ref{claim_2.6} (d) and in fact if $(M_i,\pi_{ij} \colon i \leq j \leq \omega_1)$ is as in Claim \ref{claim_2.6} (c), then $Z \in M_{\omega_1}$, so that trivially $Z$ is also in the last iterate of $\sigma^{-1}((H_{\omega_2};\in, {\sf NS}_{\omega_1} , A))$ via the generic iteration which is the restriction of $(M_i,\pi_{ij} \colon i \leq j \leq \omega_1)$ to $\sigma^{-1}((H_{\omega_2};\in, {\sf NS}_{\omega_1} , A))$ and its images.
\hfill $\square$ (Lemma \ref{folklore})
\bigskip
Let $1 \leq k<\omega$, and let $D \in {\cal P}({\mathbb R}^k)$. We say that
$T$ is a tree on ${}^{k} \omega \times {\rm OR}$ iff $T \subset
\bigcup_{n<\omega} ({}^n \omega)^k \times {}^n {\rm OR}$ and if $(s_0,...,s_{k-1},t) \in T$ and $m<\omega$, then $$(s_0 \upharpoonright m,...,s_{k-1} \upharpoonright m, t \upharpoonright m) \in T$$ We write $$[T] = \{ (x_0,...,x_{k-1},f) \colon \forall m<\omega \, (x_0 \upharpoonright m,...,x_{k-1} \upharpoonright m,f \upharpoonright m) \in T \}$$ and $p[T]$ for the projection of $T$, i.e., $$p[T] = \{ (x_0,...,x_{k-1}) \colon \exists f \, (x_0,...,x_{k-1},f) \in [T] \}$$
\begin{definition}
{\em The trees} $T$ {\em and} $U$ {\em on} ${}^{k} \omega \times {\rm OR}$
{\em witness that}
$D$ is {\em universally Baire} iff $D = p[T]$ and for all posets ${\mathbb P}$,
\begin{eqnarray}\label{complimenting_trees}
\Vdash_{\mathbb P} \, p[U] = {\mathbb R}^{k} \setminus p[T]{\rm . }
\end{eqnarray}
$D$ is called {\em universally Baire} iff there are trees $T$ and $U$ witnessing that
$D$ is universally Baire.
We denote by $\Gamma^\infty$ the collection of all $D \in \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k)$ which are universally Baire.
\end{definition}
The concept of universally Baire set was isolated by Feng-Magidor-Woodin in \cite[Section 2]{FMW}; see also \cite[Definition 8.6]{book}.
If $D \in \Gamma^\infty$, then there is an
unambiguous version of $D$ in any forcing extension $V[g]$ of $V$, which as
usual we
denote by $D^*$ and which is equal to $p[T] \cap V[g]$ for some/all trees T and U which witness that $D$ is universally Baire. See \cite[p.\ 149f.]{book}.
We will call a pointclass consisting of universally Baire sets productive iff it is closed under complements and projections in a strong sense
and
for all $k<\omega$ and $D \in \Gamma \cap {\mathbb R}^{k+2}$,
\begin{eqnarray}\label{subtlety}
(\exists^{\mathbb R} D)^* = \{ {\vec x} \in {\mathbb R}^{k+1} \colon \exists y \in {\mathbb R} \, ({\vec x},y) \in D^* \}
\end{eqnarray}
will be true in every generic extension. The formal definition runs as follows.
\begin{definition}\label{defn-productive}
Let $\Gamma \subset \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k)$. We say that $\Gamma$ is {\em productive} iff
\begin{enumerate}
\item[(a)] $\Gamma \subset \Gamma^\infty$,
\item[(b)] for all $k<\omega$ and all $D \in \Gamma \cap {\cal P}({\mathbb R}^{k+1})$, ${\mathbb R}^{k+1} \setminus D \in \Gamma$ and if $k>0$, then
$\exists^{\mathbb R} D = \{ (x_0, \ldots , x_{k-1} ) \colon \exists x_k (x_0, \ldots , x_{k-1}, x_k) \in D \} \in \Gamma$, and
\item[(c)] for all $k<\omega$ and all $D \in \Gamma \cap {\cal P}({\mathbb R}^{k+2})$, if the trees $T$ and $U$ on ${}^{k+2} \omega \times {\rm OR}$ witness that $D$ is universally Baire
and if
\begin{eqnarray}\label{U-tilde}
{\tilde U} = \{ (s \upharpoonright (k+1), (s(k+1),t)) \colon (s,t) \in U \}{\rm , }
\end{eqnarray}
then
there is a tree ${\tilde T}$ on ${}^{k+1} \omega \times {\rm OR}$ such that for all posets ${\mathbb P}$,
\begin{eqnarray}\label{T-tilde}
\Vdash_{\mathbb P} \, p[{\tilde U}] = {\mathbb R}^{k+1} \setminus p[{\tilde T}] {\rm . }
\end{eqnarray}
\end{enumerate}
\end{definition}
\begin{comment}
With $T$, $U$, ${\tilde U}$, and ${\tilde T}$ being as in Definition \ref{defn-productive}, (\ref{U-tilde}) means that
\begin{eqnarray}
p[{\tilde U}] = \{ (x_0, \ldots , x_k) \colon \exists y \in {\mathbb R} \, (x_0, \ldots , x_k,y) \in p[U] \}
\end{eqnarray}
is true in in every generic extension of $V$, which then by the properties of $T$ and $U$ and by (\ref{T-tilde}) gives that
\begin{eqnarray}
p[{\tilde T}] = \{ (x_0, \ldots , x_k) \colon \forall y \in {\mathbb R} \, (x_0, \ldots , x_k,y) \in p[T] \}
\end{eqnarray}
is true in every generic extension of $V$. In particular,
Hence a pointclass consisting of universally Baire sets is productive iff it is closed under complements and projections in the sense of (b) of Definition \ref{defn-productive} and
for all $k<\omega$ and $D \in \Gamma \cap {\mathbb R}^{k+2}$,
\begin{eqnarray}\label{subtlety}
(\exists^{\mathbb R} D)^* = \{ {\vec x} \in {\mathbb R}^{k+1} \colon \exists y \in {\mathbb R} \, ({\vec x},y) \in D^* \}
\end{eqnarray}
will be true in every generic extension.
\end{comment}
While (c) of Definition \ref{defn-productive} canonically ensures that every productive pointclass is closed under projections, at least on the face of its definition, $\Gamma$
being productive is stronger than having that $\Gamma \subset \Gamma^\infty$ and
$\Gamma$ is closed under complements and projections (\cite[Question 3]{FMW} exactly asks if the former is really stronger than the latter).
\begin{lemma}\label{absoluteness_for_productive}
If $\Gamma$ is productive and if $D \in \Gamma$, then any projective statement about $D$ is absolute between $V$ and any forcing extension of $V$,
i.e., if $\varphi$ is projective, $x_1$, $\ldots$, $x_k \in {\mathbb R}$, and ${\mathbb P}$ is any poset, then
$$V \models \varphi(x_1, \ldots , x_k, D) \ \Longleftrightarrow \ \Vdash_{\mathbb P} \, \varphi({\check x}_1, \ldots , {\check x}_k, {D}^*).$$
\end{lemma}
Lemma \ref{absoluteness_for_productive} is shown by a trivial induction on the complexity of $\varphi$.
Let $e \colon {\mathbb R} \rightarrow {\rm HC}$ be a fixed simple coding of hereditarily countable sets by reals, see e.g.\ \cite[p.\ 179]{both}. A set $D \subset {\rm HC}$ is then called {\em universally Baire in the codes} iff the code set $\{ x \in {\mathbb R} \colon e(x) \in D \}$ of $D$ is universally Baire. If this is the case, then every forcing extension of $V$ will have its unique new version of $D$, which we denote by $D^*$. If the code set of $D$ is a member of a productive pointclass, then for every forcing ${\mathbb P}$,
\begin{eqnarray}\label{el_equivalent}
({\rm HC}; \in, D) \prec ({\rm HC}^{V^{\mathbb P}}; \in, D^*){\rm . }
\end{eqnarray}
A classical variant of Lemma \ref{absoluteness_for_productive} is Shoenfield's absoluteness theorem, see e.g.\ \cite[Corollary 7.21]{book}. It states that
if $M \subset N$ are both transitive models of a sufficiently rich fragment of ${\sf ZFC}$ such that $\omega_1^V \subset M$, then
\begin{eqnarray}\label{el_equivalent_shoenfield}
({\rm HC}^M; \in) \prec_{\undertilde{\Sigma}_1} ({\rm HC}^{N}; \in){\rm , }
\end{eqnarray}
where (\ref{el_equivalent_shoenfield}) means that (\ref{ee}) holds true with $\varphi$ restricted to $\Sigma_1$ formulae (and ${\rm HC}^M$, ${\rm HC}^N$ playing the roles of ${\mathcal M}$, ${\mathcal N}$, respectively).
\cite[Question 3]{FMW}
is concerned
with the question about the connection of, on the one hand, projective absoluteness with respect to forcing extensions and, on the other hand, having that every projective set is universally Baire (see \cite[Questions 1 and 7]{FMW}).
\begin{theorem}\label{Gammainfty-productive} \textbf{\em (Woodin)} Assume that there is a proper class of Woodin cardinals. Then $\Gamma^\infty$ is productive.
\end{theorem}
{\sc Proof.} A theorem of Woodin says that in the presence of a proper class of Woodin cardinals, every set in $\Gamma^\infty$ is weakly homogeneously Suslin,
see e.g.\ \cite[Theorem 3.3.8]{paul} and \cite[Theorem 1.2]{forcing-free}.
Every tree ${\tilde U}$ witnessing that a given set $D$ of reals is weakly homogeneously Suslin comes with a canonical tree ${\tilde T}$ for
${\mathbb R} \setminus D$ in such a way that ${\tilde U}$ and ${\tilde T}$ are connected as in (c) of Definition
\ref{defn-productive}. For the construction of ${\tilde T}$ see e.g. \cite[p.\ 455]{kanamori}. \cite[Proposition 32.6]{kanamori} formulates how the two trees are connected. The main result of Martin and Steel
from \cite{proj-determinacy} is then that Woodin cardinals may be used to show that ${\tilde T}$ is homogeneous, cf.\ \cite[Theorem 32.11]{kanamori}.
That way, it follows that $\Gamma^\infty$ is productive provided that there is a proper class of Woodin cardinals. \hfill $\square$ (Theorem \ref{Gammainfty-productive})
\bigskip
For any set $X$, $M_\omega^\#(X)$ denotes the least active $X$-mouse which has infinitely many Woodin cardinals. See \cite{PWIM}.
\begin{theorem}\label{steel} \textbf{\em (Steel)} Assume {\sf PFA}. Then the universe is closed under the operation $X \mapsto M_\omega^\#(X)$. In particular, every set of reals in $L({\mathbb R})$ is universally Baire, and $\bigcup_{k<\omega} \, {\cal P}({\mathbb R}^k) \cap L({\mathbb R})$ is productive.
\end{theorem}
{\sc Proof.} The proof from \cite{john} produces the result that under {\sf PFA}, the universe is closed under the operation $X \mapsto M_\omega^\#(X)$.
The rest is given by standard inner model theoretic arguments, see e.g.\ \cite[Section 3, pp.\ 187f.]{both}. \hfill $\square$ (Theorem \ref{steel})
\bigskip
By Lemma \ref{folklore} and Theorem \ref{steel},
Theorem \ref{main_thm}
follows from the following more general statement.
\begin{theorem}\label{main_thm_gen}
Let $\Gamma \subset \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k)$. Assume that
\begin{enumerate}
\item[(i)] $\Gamma = \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k) \cap L(\Gamma,{\mathbb R})$,
\item[(ii)] $\Gamma$ is productive, and
\item[(iii)] Martin's Maximum${}^{++}$ holds true.
\end{enumerate}
Then $g_A$ is ${\mathbb P}_{\rm max}$-generic over $L(\Gamma,{\mathbb R})$.
\end{theorem}
The abbreviation $(*)_\Gamma$ was introduced in \cite[Definition 4.1]{both} to denote a straightforward generalization of $(*)$ to larger pointclasses. For a pointclass $\Gamma \supset {\cal P}({\mathbb R}) \cap L({\mathbb R})$, $(*)_\Gamma$ is the statement that every set in $\Gamma$ is determined, and there is a filter $g \subset {\mathbb P}_{\rm max}$ which has nonempty intersection with every dense set (coded by a set) in $\Gamma$ and is such that ${\cal P}(\omega_1) \subset L({\mathbb R})[g]$.
\begin{corollary}
Assume that there is a proper class of Woodin cardinals.
Let $\Gamma \subset \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k) \cap \Gamma^\infty$. Suppose that
(i)--(iii) from the statement of Theorem \ref{main_thm_gen} are satisfied.
Then $(*)_\Gamma$ holds true.
\end{corollary}
Theorem
\ref{main_thm_gen} readily follows from the following Lemma via a standard application
of ${\sf MM}^{++}$.
\begin{lemma}\label{main_claim}
Let $\Gamma \subset \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k)$. Assume that
\begin{enumerate}
\item[(i)] $\Gamma = \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k) \cap L(\Gamma,{\mathbb R})$,
\item[(ii)] $\Gamma$ is productive, and
\item[(iii)] ${\sf NS}_{\omega_1}$ is saturated.
\end{enumerate}
Let $D \subset {\mathbb P}_{\rm max}$ be open dense, $D \in L(\Gamma,{\mathbb R})$.
With $A$ being as in Convention \ref{convention},
there is then a stationary set preserving forcing ${\mathbb P}$ of size $2^{\aleph_2}$ such that in $V^{\mathbb P}$
there is some $p = {\cal N}_0 = (N;\in,I,a) \in D^*$ and some generic iteration $$({\cal N}_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = {\cal N}_0$ of length $\omega_1+1$ such that if we write ${\cal N}_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$, then $I^* = ({\sf NS}_{\omega_1})^{V^{\mathbb P}} \cap N_{\omega_1}$ and $a^* = A$.
\end{lemma}
{\sc Proof} of Theorem \ref{main_thm_gen} from Lemma \ref{main_claim}. ${\sf MM}$ implies that ${\sf NS}_{\omega_1}$ is saturated, see \cite[Theorem 12]{FMS}. By Lemma \ref{folklore}, it remains to show that $D \cap g_A \not= \emptyset$ for every
open dense $D \subset {\mathbb P}_{\rm max}$, $D \in \Gamma$. Here, $g_A$ is as in Convention \ref{convention}.
Let us fix such $D$. The statement that there is a $p$ as in the conclusion of
Lemma \ref{main_claim}, which is tantamount to saying that there is a $p \in D \cap g_A$, is easily seen to be $\Sigma_1$ expressible over the structure
$(H_{\omega_2};\in,{\sf NS}_{\omega_1},A,D)$. By the conclusion of Lemma \ref{main_claim}, the existence of such a $p$ may be forced by a stationary set preserving forcing.
Hence by ${\sf MM}^{++}$, cf.\ \cite[Theorem 10.124]{hugh},
there is such a $p$ in $V$. \hfill $\square$ (Theorem \ref{main_thm_gen})
\bigskip
\begin{comment}
The attentive reader will notice that with virtually the same proof that is to come we could weaken the hypothesis (iii) of Lemma
\ref{main_claim} to ``${\sf NS}_{\omega_1}$ is precipitous'' at the cost of increasing the size of ${\mathbb P}$ to become $(2^{2^{\aleph_1}})^{+}$. The same remark applies to hypothesis (iii) in Theorem \ref{2.9}. We decided to present our proof working under the stronger hypothesis, though, as
doing so
simplifies the notation, and in practice the weaker versions of Lemma \ref{main_claim} and
Theorem \ref{2.9} are as interesting as their stronger counterparts.
\end{comment}
As the proof of Theorem \ref{main_thm_gen} from Lemma \ref{main_claim} shows,
we don't need the full power of ${\sf MM}^{++}$
in order to derive Theorem \ref{main_thm_gen} from Lemma \ref{main_claim}.
Instead, a bounded version of ${\sf MM}^{++}$ suffices; it may be defined as follows, see
\cite[Definition 10.123]{hugh}.
\begin{definition}
For $D \in \Gamma^\infty$, $D$-${\sf BMM}^{++}$ is the statement that for all ${\mathbb P}$ which are stationary set preserving, $$(H_{\omega_2}^V;\in,{\sf NS}_{\omega_1}^V,D) \prec_{\Sigma_1} (H_{\omega_2}^{V^{\mathbb P}};\in,{\sf NS}_{\omega_1}^{V^{\mathbb P}},D^*).$$
\end{definition}
Modulo large cardinals, $(*)$ is then actually {\em equivalent} to $D$-${\sf BMM}^{++}$ for all $D \in {\mathcal P}({\mathbb R}) \cap L({\mathbb R})$. Let us first
state a more general fact, Theorem \ref{2.9}, which gives the characterization of $(*)$, i.e.\ Theorem \ref{char_star}, as a special case.
\begin{theorem}\label{2.9}
Assume that there is a proper class of Woodin cardinals.
Let $\Gamma \subset \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k)$. Assume that
\begin{enumerate}
\item[(i)] $\Gamma = \bigcup_{1 \leq k<\omega} \, {\cal P}({\mathbb R}^k) \cap L(\Gamma,{\mathbb R})$,
\item[(ii)] $\Gamma$ is productive.
\end{enumerate}
The following statements are then equivalent, with $g_A$ being as in Convention \ref{convention}.
\begin{enumerate}
\item[(1)] $D$-${\sf BMM}^{++}$ holds true for all $D \in \Gamma$.
\item[(2)] $g_A$ is ${\mathbb P}_{\rm max}$-generic over $L(\Gamma,{\mathbb R})$.
\end{enumerate}
\end{theorem}
{\sc Proof.} (2) $\Longrightarrow$ (1): This is exactly by the proof of (A) $\Longrightarrow$ (B) of \cite[Theorem 2.7]{david-ralf}.
(1) $\Longrightarrow$ (2): We may first force ${\sf NS}_{\omega_1}$ to be saturated
by a stationary set preserving forcing, see e.g.\ \cite[Theorem 2.64]{hugh}.
The rest is then by the proof of Theorem \ref{main_thm_gen} from Lemma \ref{main_claim}
which was given above.
\hfill $\square$ (Theorem \ref{2.9})
\begin{theorem}\label{char_star} Assume that there is a proper class of Woodin cardinals.
The following statements are then equivalent.
\begin{enumerate}
\item[(1)] $D$-${\sf BMM}^{++}$ holds true for all
$D \in {\cal P}({\mathbb R}) \cap L({\mathbb R})$.
\item[(2)] $(*)$.
\end{enumerate}
\end{theorem}
Let us now give a {\sc Proof} of Theorem \ref{char_of_star} from Lemma \ref{main_claim}.
(1) $\Longrightarrow$ (2) is weaker than (1) $\Longrightarrow$ (2) of Theorem
\ref{Wo1}.
Let us now assume (2) and show (1). Fix $D \subset {\mathbb P}_{\rm max}$, any open dense set in $L({\mathbb R})$. As the statement of the theorem assumes a supercompact cardinal to exist, there is a semi-proper (and hence stationary set preserving) forcing ${\mathbb P}$ such that ${\sf MM}^{++}$ holds true in $V^{\mathbb P}$.
Inside $V^{\mathbb P}$, we will have that ${\sf MM}^{++}$ yields via
Lemma \ref{main_claim} that for all $A' \subset \omega_1$ with $\omega_1^{L[A']} =
\omega_1$ there will be some $p = (N;\in,I^*,a^*) \in D^*$ and some generic iteration $$({\cal N}_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = {\cal N}_0$ of length $\omega_1+1$ such that if we write ${\cal N}_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$, then $I^* = ({\sf NS}_{\omega_1})^{V^{\mathbb P}} \cap N_{\omega_1}$ and $a^* = A'$. This is a statement which is $\Pi_2$ over the
structure mentioned in (2) of Theorem \ref{char_of_star}.
This statement will therefore be true in $V$, which readily implies that
$g_A$ is ${\mathbb P}_{\rm max}$-generic over $L({\mathbb R})$
and ${\cal P}(\omega_1) \subset L({\mathbb R})[g_A]$, where $g_A$ is as in Convention \ref{convention}. \hfill $\square$ (Theorem \ref{char_of_star})
\bigskip
The forcing which we designed in order to produce Lemma \ref{main_claim} is a souped up version of the forcings from \cite{cs} and \cite{ds}, which are in turn variants of the ${\cal L}$-forcing of Jensen as being developed e.g.\ in \cite{ronald_l_forcing}.\footnote{One of the referees informs us that J.\ Keisler in \cite{kei71} and \cite{kei73}
developed forcings which work in a similar fashion.}
All these forcings may be construed as building uncountable models as term models of a given language, ${\cal L}$, with the forcing conditions being finite fragments of a consistent and complete ${\cal L}$-theory which will give those term models, augmented by ``side conditions'' which will guarantee that the forcing only collapses cardinals in a controlled way. Our forcing will change the cofinalities of $\omega_2$ and $\omega_3$ to $\omega$ and $\omega_1$, respectively, and it won't collapse any other cardinal
outside of the (possibly empty) half-open interval $(\omega_3,2^{\aleph_2}]$.
Let us give an outline of the proof of Lemma \ref{main_claim}.
To prove Lemma \ref{main_claim}, we aim to build a stationary set preserving forcing $\mathbb P$ which adds a generic iteration of some ${\mathbb P}_{\rm max}$-condition $(N; \in, I, a)$ coded by a real in the projection of a tree ${\tilde T}$ projecting to the set of codes for conditions in our given dense set $D$. Moreover, we want this iteration to send the distinguished set $a$ of $(N; \in, I, a)$ to $A$, and we want every $I^*$-positive set in the final model $(N^*;\in,I^*,A)$ to be a stationary subset of $\omega_1$ in $V^{\mathbb P}$. Our approach is to think of all the relevant objects -- $(N; \in, I, a)$, a branch through ${\tilde T}$ projecting to a real coding $(N; \in, I, a)$, and the generic iteration of $(N; \in, I, a)$ of length $\omega_1+1$ -- as being given by ``term models'' in a suitable language, ${\cal L}$, and add them via finite approximations.
Thus, the working parts of our forcing will be finite sets $p$ of sentences from $\mathcal L$ providing partial information about the above objects. We will require these finite bits of information $p$ to be realized in some outer model.\footnote{$W$ is an outer model iff $W$ is a transitive model of {\sf ZFC} with
$W \supset V$ and which has the same ordinals as $V$; in other words, $W$ is an
outer model iff $V$ is an inner model of $W$.}
The existence of such an outer model will be absolute to
any generic extension of $V$ via ${\rm Col}(\omega, \omega_2)$.
\begin{comment}
In order to generate these models, a natural strategy is to observe that if we force with ${\rm Col}(\omega, \omega_2)$, then $(M_0; \in, I_0, a_0):=(H_{\omega_2}; \in, {\sf NS}_{\omega_1})^V$ becomes a ${\mathbb P}_{\rm max}$-condition, which we can then iterate in $V^{{\rm Col}(\omega, \omega_2)}$, while preserving the corresponding ideal, inside a ${\mathbb P}_{\rm max}$-condition $(N; \in, J, b)$ which is coded by a real in the projection of $T$ (the reason being that, using the projectivity of $\Gamma$, we may assume that we have taken $T$ so that this is the case). Let $\mathcal I$ be this iteration of $(M_0; \in, I_0, a_0)$. Now we can keep iterating $(N; \in, J, b)$, in any way we want, in length $\omega_1^{V^{{\rm Col}(\omega, \omega_2)}}=\omega_3^V$. The generic iteration $j(\mathcal I)$ of $(M_0; \in, I_0, a_0)$---where $j:(N; \in, J)\to (N^*; \in, J^*)$ is the elementary embedding corresponding to the iteration of $(N; \in, J)$---can then be lifted to a generic iteration $j^*:V\to W$ of $V$. One can now verify that $V^{{\rm Col}(\omega, \omega_2)}$ contains the objects we intend to approximate, albeit not defined relative to the relevant parameters $T$ and $A$, but relative to $j^*(T)$ and $j^*(A)$. By absoluteness, the same is true for any extension of $W$ via ${\rm Col}(\omega, j^*(\lambda))$, and the existence of the intended objects holds in any extension of $V$ via ${\rm Col}(\omega, \lambda)$ by elementarity of $j^*$. Given our strategy, the relevant items in the above construction---i.e., not only the objects we are ultimately interested in obtaining, but also $(M_0; \in, I_0, a_0)$ and its iteration $\mathcal I$ inside $(N; \in, J, b)$, etc.---will be among the objects our forcing $\mathbb P$ will need to approximate. We will think of these objects, in the configuration outlined above, as `certificates' of some finite piece of information (described by a condition in $\mathbb P$). In other words, the natural idea here is to define our forcing as consisting of those finite sets of $\mathcal L$-sentences for which there is a `certificate' in some outer model.
The problem with the above strategy is that, although a forcing $\mathbb P$ like the one we have described would in fact add the desired objects, one would still need to show that it preserves stationary subsets of $\omega_1$ and that every positive set in the final model of the iteration being added by $\mathbb P$ is in fact stationary in that extension. The first ingredient for handling this issue is the use of a `diamond sequence' $\langle A_\alpha\,:\,\alpha\in\kappa\rangle$ on $H_\kappa$, for a high enough $\kappa$ our forcing $\mathbb P$ will be a subset of,\footnote{We will in fact be able to take $\kappa=\omega_3$.} in order to predict $\mathbb P$-names for clubs of $\omega_1$ (which will be subsets of $H_{\kappa})$. We may assume that this `diamond sequence' exists in $V$ as otherwise we can simply add it with a preliminary forcing. Now, the idea for dealing with the present problem is to also take, among the objects that make up the `certificate' of a $\mathbb P$-condition $p$, certain countable $X_\lambda\prec H_{\kappa}^V$, indexed by ordinals $\lambda\in\kappa$,\footnote{These $X_\lambda$ will be elementary submodels of $H_\kappa^V$ although they will typically not be in $V$.} as ``side conditions'' inside of which we can work in order to find some condition extending $p\cap X_\lambda$ to a condition in some given dense set of $\mathbb P$ definable from a $\mathbb P$-name $\dot C$ for a club of $\omega_1$ guessed by $A_\lambda$. If we succeed in finding $r$ in such a way that $r$ and $p$ are compatible, then we are done (s.\ the proof of Lemma \ref{P_stat_pres}). However, the problem for this is that, even though $p$ and $r$ would have certificates $\mathfrak C_p$ and $\mathfrak C_r$, respectively, it is not obvious how to find $r$ in such a way that we can amalgamate (the relevant part of) $\mathfrak C_r$ with (the relevant part of) $\mathfrak C_p$ into a certificate for $p\cup r$.
The above obstacle leads to the final main ingredient of our construction. This consists in defining our forcing $\mathbb P$ as $\mathbb P_\kappa$ for a certain $\subseteq$-increasing sequence $\vec{\mathbb P}=(\mathbb P_\lambda\,:\,\lambda\in C\cup\{\kappa\})$, for a club $C$ of $\kappa$, constructed recursively.\footnote{This increasing sequence $\vec{\mathbb P}$ of forcing notions is not an iteration, i.e., it is not the case that $\mathbb P_\lambda$ is a complete suborder of $\mathbb P_{\lambda'}$ for $\lambda<\lambda'\leq\kappa$.} Each $\mathbb P_\lambda$, for $\lambda\in C$, will be a subset of $Q_\lambda$, where $(Q_\lambda\,:\,\lambda\in C)$ is a $\subseteq$-continuous filtration of elementary submodels of $H_{\kappa}$ such that $A_\lambda\subseteq Q_\lambda$ for each $\lambda\in C$. At each stage $\lambda$ in the construction of $\vec{\mathbb P}$, we define $\mathbb P_\lambda$ by saying that a finite set $p$ of $\mathcal L\cap Q_\lambda$-sentences is in $\mathbb P_\lambda$ if and only if there is a certificate $\mathfrak C$ for it which, when intersected with each of the side conditions $X_{\delta}$, for $\delta<\lambda$, is generic over $X_\delta$, for the already defined forcing $\mathbb P_\delta$.\footnote{This version of the idea of `certificate' we finally end up with is thus our official definition of certificate.} This genericity requirement, as hinted at by the description of the construction in the previous paragraph, will be enough to guarantee that the condition $r$ found inside $X_\lambda$ in that construction will in fact be compatible with $p$ as witnessed by the same certificate $\mathfrak C$.
\end{comment}
In
$V^{{\rm Col}(\omega, \omega_2)}$, $$(H_{\omega_2}^V; \in, {\sf NS}_{\omega_1}^V,A)$$ becomes a ${\mathbb P}_{\rm max}$-condition and $p[{\tilde T}]=D^*$ is still dense, so that in $V^{{\rm Col}(\omega, \omega_2)}$
there is a ${\mathbb P}_{\rm max}$-condition $(N; \in, I, a) \in D^*$ which is
stronger than $(H_{\omega_2}^V; \in, {\sf NS}_{\omega_1}^V,A)$.
We may now iterate $(N; \in, I, a)$ in length $\omega_1^{V^{{\rm Col}(\omega, \omega_2)}}+1=\omega_3^V+1$ so as to produce $$\sigma \colon (N;\in,I,a) \rightarrow
(N^*;\in,I^*,a^*).$$
If $(M_i,\pi_{ij} \colon i \leq j \leq \omega_1^N) \in N$ is the generic iteration of $(H_{\omega_2}^V; \in, {\sf NS}_{\omega_1}^V,A)$ witnessing that $(N; \in, I, a)$ is stronger than
$(H_{\omega_2}^V; \in, {\sf NS}_{\omega_1}^V,A)$, then
$\sigma((M_i,\pi_{ij} \colon i \leq j \leq \omega_1^N)) = (M_i,\pi_{ij} \colon i \leq j \leq
\omega_3^V)$ is an extension of that iteration. We have that
$$M_{0} = (H_{\omega_2}^V; \in, {\sf NS}_{\omega_1}^V,A){\rm , }$$ and
$$\pi_{0, \omega_3^V}
\colon M_{0} \rightarrow
M_{\omega_3^V}$$ may be
lifted to a generic iteration $${\tilde \pi} \colon V\to M$$ of $V$, for a transitive $M$, such that ${\tilde \pi} \supset \pi_{0, \omega_3^V}$ and ${\tilde \pi}(M_0)=M_{\omega_3^V}$. See \cite[Lemma 3.8]{hugh}.
\bigskip
\begin{tikzpicture}
\def\x{2.1};
\def\y{0.5};
\node (pT) at (0*\x, 0*\y){$p[{\tilde T}]$};
\node (ppiT) at (0*\x+1.5, 0*\y){$p[{\tilde \pi}({\tilde T})]$};
\node (N) at (0*\x, -2*\y) {$(N;\in, I, a)$};
\node (Nast) at (2*\x, -2*\y) {$(N^*;\in, I^*, a^*)$};
\node (M0) at (-2*\x, -4*\y) {$M_0$};
\node (MN) at (0*\x, -4*\y) {$M_{\omega_1^N}$};
\node (M3) at (2*\x, -4*\y) {$M_{\omega_3^V}$};
\node (H) at (-2*\x, -6*\y) {$(H_{\omega_2}^V;\in, \mathsf{NS}_{\omega_1}^V, A)$};
\node (V) at (-2*\x, -8*\y) {$V$};
\node (Mp) at (2*\x, -8*\y) {$M$};
\path (pT)--(ppiT) node[midway]{$\subseteq$};
\path (N)--(pT) node[midway, rotate=90]{$\in$};
\draw[->] (N)--(Nast) node[midway, above]{$\sigma$};
\path (MN)--(N) node[midway, rotate=90]{$\in$};
\path (M3)--(Nast) node[midway, rotate=90]{$\in$};
\draw[->] (M0)--(MN) node[midway, above]{$\pi_{0 \omega_1^N}$};
\draw[->] (MN)--(M3) node[midway, above]{$\pi_{\omega_1^N \omega_3^V}$};
\path (H)--(M0) node[midway, rotate=90]{$=$};
\path (M3)--(Mp) node[midway, rotate=270]{$\in$};
\path (H)--(V) node[midway, rotate=270]{$\in$};
\draw[->] (V)--(Mp) node[midway, above]{${\tilde \pi}$};
\end{tikzpicture}
\bigskip
We now see that $V^{{\rm Col}(\omega, \omega_2)}$ contains objects like the ones we intend to add by our forcing---namely $(N; \in, I, a)$, a branch through ${\tilde T}$ projecting to a real coding $(N; \in, I, a)$, and the generic iteration of $(N; \in, I, a)$ of length $\omega_1+1$---albeit not defined relative to the parameters ${\tilde T}$, $H_{\omega_2}^V$, and $A$, but relative to ${\tilde \pi}({\tilde T})$, $H_{\omega_2}^M$, and ${\tilde \pi}(A)$.
The statement that such objects exist is $\Sigma_1$ in the parameters $H_{\omega_2}^M$, ${\tilde \pi}(A)$, and a Skolem hull\footnote{See Claim \ref{ex_of_T}.} of ${\tilde \pi}({\tilde T})$ of size $\aleph_2^M$, which will both be elements of ${\rm HC}^{M^{{\rm Col}(\omega,{\tilde \pi}(\omega_2))}}$.
By Shoenfield absoluteness (\ref{el_equivalent_shoenfield}), see e.g.\ \cite[Corollary 7.21]{book},
such objects will also exist in $M^{{\rm Col}(\omega, {\tilde \pi}(\omega_2))}$.
The point that $\pi_{0,\omega_3^V}$ could be lifted to ${\tilde \pi}$ is then the following.
The statement that objects like the ones we intend to add by our forcing exist in
$M^{{\rm Col}(\omega, {\tilde \pi}(\omega_2))}$
may now be pulled back via ${\tilde \pi}$. This buys us that objects like the ones we intend to add by our forcing exist in $V^{{\rm Col}(\omega, \omega_2)}$ -- and this time with the right parameters ${\tilde T}$, $H_{\omega_2}^V$, and $A$. The argument that combined lifting $\pi_{0,\omega_3^V}$ to ${\tilde \pi}$, applying Shoenfield absoluteness, and pulling back the statement of interest was crucial to arrive at the desired conclusion, viz.\ that objects like the ones we intend to add by our forcing exist in $V^{{\rm Col}(\omega, \omega_2)}$.\label{discussion_why_M} This will be our starting point for cooking up the forcing ${\mathbb P}$.
In order to prove that our forcing ${\mathbb P}$ preserves stationary subsets of $\omega_1$ we will need an argument exploiting lifting, Shoenfield absoluteness, and pulling back. In order to be able to run this argument we will need our forcing to approximate, not only the objects we are ultimately interested in obtaining, but also the iteration
$(M_i,\pi_{ij} \colon i \leq j \leq \omega_1^N) \in N$. (See footnote \ref{footnote}.)
We will think of the objects themselves, which exist in $V^{{\rm Col}(\omega,\omega_2)}$, as ``certificates'' for some finite piece of information about them. The idea is then to have our forcing consist of finite sets of $\mathcal L$-sentences for which there is a ``certificate'' in $V^{{\rm Col}(\omega,\omega_2)}$.
The problem with the above strategy is that, although a forcing $\mathbb P$ like the one we have described would in fact add the desired objects, one would still need to show that it preserves stationary subsets of $\omega_1$ and that every positive set in the final model of the iteration being added by $\mathbb P$ is in fact stationary in that extension.
Our forcing ${\mathbb P}$ will be a subset of $H_{\omega_3}$, and
one tool for taking care of these issues is the use of a diamond sequence $\langle (Q_\lambda, A_\lambda) \,:\,\lambda < \omega_3\rangle$ consisting of transitive structures\footnote{I.e., the underlying universe $Q_\lambda$ will be transitive and $A_\lambda \subset Q_\lambda$ will be a distinguished predicate of the structure.} in $H_{\omega_3}$ in order to guess $(H_{\omega_3},{\dot C})$, where ${\dot C}$ is a $\mathbb P$-name for a
club in $\omega_1$, ${\dot C} \subset H_{\omega_3}$.
That $(H_{\omega_3},{\dot C})$ be guessed means that there are stationarily many $\lambda<\omega_3$ such that $(Q_\lambda,A_\lambda)$ is an elementary substructure of $(H_{\omega_3},{\dot C})$. See $(\Diamond)$ on p.\ \pageref{defn_diamond}.
Imagine that ${\mathbb P}$ is a forcing which adds the desired objects, but which also preserves stationary subsets of $\omega_1$. Let ${\dot C}$ be a $\mathbb P$-name for a club in $\omega_1$,
${\dot C} \subset H_{\omega_3}$, and let $S \subset \omega_1$ be stationary in $V$.
Let $g$ be ${\mathbb P}$-generic
over $V$. There will be some $\lambda<\omega_3$ such that $(H_{\omega_3},{\dot C})$ is guessed by $(Q_\lambda,A_\lambda)$ and in $V[g]$ there will be some countable elementary substructure
$X$ of $(Q_\lambda,A_\lambda)$ such that, writing $X[g] = \{ \tau^g \colon \tau \in V^{\mathbb P} \cap X \}$,
\begin{enumerate}
\item[(a)] $X \cap \omega_1 \in S$, and
\item[(b)] $X[g] \cap Q_\lambda = X \cap Q_\lambda$.
\end{enumerate}
That $(Q_\lambda,A_\lambda)$ is an elementary substructure of $(H_{\omega_3},{\dot C})$ will then mean in practice that $X \cap \omega_1 \in {\dot C}^g$, so that
$X \cap \omega_1$ witnesses that $S$ is still stationary in $V[g]$.
Calling some $g$ with the property (b) ``$({\mathbb P},X)$-generic''
there is, however, no reason to expect an $X$ such that $g$ is $({\mathbb P},X)$-generic to exist in $V$ (in fact, $V$ won't have such $X$).
When defining ${\mathbb P}$, we will turn this around and have our conditions also approximate finite bits of information about such elementary substructures $X$.
Our key tool for taking care of the above issue is then to define ${\mathbb P}$ as the last forcing from a
recursively defined
$\subset$-increasing sequence $\vec{\mathbb P} = (\mathbb P_\lambda\,:\,\lambda \leq \omega_3
)$. Each $\mathbb P_\lambda$ will be a subset of $Q_\lambda$.
Hence, when defining ${\mathbb P}_{\eta}$, $\eta \leq \omega_3$,
if $\lambda<\eta$, then we already know what it means for some $g$ to be (partially) ${\mathbb P}_\lambda$-generic over $(Q_\lambda, A_\lambda)$, and if
$X$ is a countable elementary substructure of $(Q_\lambda, A_\lambda)$, then $X[g]$ may be assigned a meaningful interpretation as
$X[g] = \{ \tau^g \colon \tau \in V^{{\mathbb P}_\delta} \cap X \}$.
We will maintain that
at each stage $\eta$ in the construction of $\vec{\mathbb P}$ we define $\mathbb P_\eta$ by saying that a finite set $p$ of $\mathcal L\cap Q_\eta$-sentences is in $\mathbb P_\eta$ if and only if there is a certificate for $p$ which extends $p$ and when intersected with each of the side conditions $X_{\lambda} \prec (Q_\lambda , A_\lambda)$ (also given by the certificate), for $\lambda<\eta$, is generic over $X_\eta$ for the already defined forcing $\mathbb P_\lambda$.
The role of condition (b) above (with $X_\lambda$ replacing $X$) will be that if $A_\lambda$ codes the name of a club subset ${\dot C}$ of $\omega_1$, then some $p \in g$ must force that $X_\lambda \cap \omega_1 \in {\dot C}$, so that in the light of (a) above, $p$ also forces that $S$ has non-empty intersection with ${\dot C}$.
Our next section is entirely devoted to a proof of Lemma \ref{main_claim}.
\section{The forcing.}
Recall our Convention \ref{convention} which we are now going to make use of without further notice.
Let us assume throughout the hypotheses of Lemma \ref{main_claim}.
We aim to verify its conclusion.
Let us fix $D \subset {\mathbb P}_{\rm max}$, an open dense set in $L(\Gamma,{\mathbb R})$. The fact that $D$ is open dense may be written as $$\forall p \in {\mathbb P}_{\rm max} \, \exists q \leq p, \, q \in D \wedge \forall p \in D \, \forall q \leq p, \, q \in D.$$
By hypothesis (ii) in the statement of Lemma \ref{main_claim} (i.e., $\Gamma \subset \Gamma^\infty$ is productive), we
may apply Lemma \ref{absoluteness_for_productive} to conclude that
(\ref{el_equivalent}) on p.\ \pageref{el_equivalent} holds true with $D$ and $\mathbb P={\rm Col}(\omega,\omega_2)$, i.e., $$({\rm HC} ; \in, D) \prec ({\rm HC}^{V^{{\rm Col}(\omega,\omega_2)}} ; \in, D^*).$$ This will ensure that
\begin{enumerate}
\item[(D.1)] $V^{{\rm Col}(\omega,\omega_2)} \models$ ``$D^*$ is an open dense
subset of ${\mathbb P}_{\rm max}$.''
\end{enumerate}
Let us identify $D$ with a canonical set of reals coding the elements of $D$,\footnote{We will later have to spell out a bit more precisely in which way we aim to have the elements of $p[T]$ code the elements of $D$, see (C.2) and ($\Sigma$.5) below.} and
let ${\tilde T} \in V$ be a
tree on $\omega \times \theta$, for some ordinal $\theta$, such that
\begin{enumerate}
\item[(D.2)] $V^{{\rm Col}(\omega,\omega_2)} \models D^*=p[{\tilde T}]$.\footnote{An easy Skolem hull argument may be used to show that we might actually pick ${\tilde T} \in V$ as a tree
on $\omega \times 2^{{\aleph_2}}$. We won't need that, though, but we shall prove and make use of a related fact below, see (\ref{q-star}).}
\end{enumerate}
Let
$h$ be ${\rm Col}(\omega,\omega_2)$-generic over $V$. Inside $V[h]$,
\begin{eqnarray}\label{its-a-Pmax-condition!}
((H_{\omega_2})^V;\in, ({\sf NS}_{\omega_1})^V , A)
\end{eqnarray} is a ${\mathbb P}_{\rm max}$ condition, call it $p$.
Let $q^* \in ({\mathbb P}_{\rm max})^{V[h]}$, $q^* < p$, $q^* \in D^*$, cf.\ (D.1).
Let $q^* = (N^*; \in , I^* , a^*)$. Identifying $q^*$ with some real coding it, we have that
$q^* \in p[{\tilde T}]$, cf.\ (D.2).
\begin{claim}\label{ex_of_T}
There is a tree ${T} \in V$ on $\omega \times \omega_2$ such that
\begin{eqnarray}\label{q-star}
q^* \in p[{T}] \subset p[{\tilde T}].
\end{eqnarray}
\end{claim}
{\sc Proof} of Claim \ref{ex_of_T}.
Let $q^* = \sigma^h$, where $\sigma \in V^{{\rm Col}(\omega,\omega_2)}$. We may assume that $\sigma \in H_{\omega_3}$. Recall that ${\tilde T}$ is on $\omega \times \theta$. Let $X \in V$, $X \prec H_{\theta^+}^V$ be such that $\omega_2+1 \cup \{ \sigma , {\tilde T} \} \subset X$ and ${\rm Card}(X)=\aleph_2$. Let $\pi \colon P \cong X \prec H_{\theta^+}^V$ be such that $P$ is transitive, and write $T = \pi^{-1}({\tilde T})$. We have that $\pi(\sigma)=\sigma$, and $\pi$ lifts to ${\tilde \pi} \colon P[h] \rightarrow
H_{\theta^+}^{V[h]}$ with ${\tilde \pi}(q^*) = {\tilde \pi}(\sigma^h) = \pi(\sigma)^h = \sigma^h = q^*$. As $q^* \in p[{\tilde T}]$, the elementarity of ${\tilde \pi}$ then yields that $q^* \in p[T]$. The tree $T$ is on $\omega \times P \cap {\rm OR}$, but using a bijection of $P \cap {\rm OR}$ with $\omega_2$, we may construe it as a tree on $\omega \times \omega_2$.
\hfill $\square$ (Claim \ref{ex_of_T})
\bigskip
Let us fix $T$ as in Claim \ref{ex_of_T}.
Let us write
\begin{eqnarray} \kappa = \aleph_3 {\rm , }
\end{eqnarray}
so that $T \in H_\kappa$.
Let $d$ be ${\rm Col}(\kappa,\kappa)$-generic over $V$. In $V[d]$, let
$({\bar A}_\lambda \colon \lambda<\kappa)$ be a $\Diamond_\kappa$-sequence, i.e.,
for all ${\bar A} \subset \kappa$, $\{ \lambda<\kappa \colon {\bar A} \cap \lambda = {\bar A}_\lambda \}$ is stationary.
Also, let $c \colon \kappa \rightarrow H_\kappa^{V} = H_\kappa^{V[d]}$,
$c \in V[d]$, be bijective.
For $\lambda < \kappa$, let
\begin{eqnarray}
Q_\lambda = c \mbox{''} \lambda \ \mbox{ and } \ A_\lambda = c \mbox{''} {\bar A}_\lambda.
\end{eqnarray}
An easy closure argument will give us some club\label{defn-club} $C \subset \kappa$ such that for all $\lambda \in C$,
\begin{enumerate}
\item[(i)] $Q_\lambda$
is transitive,
\item[(ii)] $\{ T , ((H_{\omega_2})^V ; \in , ({\sf NS}_{\omega_1})^V ,A) \} \cup
(\omega_2+1) \subset Q_\lambda$,
\item[(iii)] $Q_\lambda \cap {\rm OR} = \lambda$ (so that $c \upharpoonright \lambda \colon
\lambda \rightarrow Q_\lambda$ is bijective), and
\item[(iv)] $(Q_\lambda;\in) \prec
(H_\kappa;\in)$.
\end{enumerate}
\begin{comment}
In item (iv), $(M;\in) \prec (N;\in)$ means that $(M;\in)$ is an elementary substructure of $(N;\in)$, i.e., $M \subset N$ and for all formulae $\varphi$ and
all $x_1$, $\ldots$, $x_k \in M$, $$(M;\in) \models \varphi(x_1,\ldots,x_k) \Longleftrightarrow
(N;\in) \models \varphi(x_1,\ldots,x_k).$$
\end{comment}
(ii) is true for all sufficiently large $\lambda < \kappa$, and (iv) is true for all $\lambda$ such that $Q_\lambda = c \mbox{''} \lambda$ is closed under some fixed set of Skolem functions for $H_\kappa$. As the set of $\lambda < \kappa$ with (i) and (iii) is each easily seen to be club, a club of $\lambda$ with the above properties certainly exists.
We will fix from now on some club $C \subset \kappa$ with (i) through (iv) for all $\lambda \in C$.
In $V[d]$, for all $P$, $B \subset H_\kappa$, the set of all $\lambda \in C$ such that
$$(Q_\lambda;\in,P \cap Q_\lambda , B \cap Q_\lambda) \prec (H_\kappa;\in,P,B)$$
is club, and the set of all $\lambda \in C$ such that $B \cap Q_\lambda = A_\lambda$ is stationary, so that
\begin{enumerate}
\item[($\Diamond$)] For all $P$, $B \subset H_\kappa$ the set \label{defn_diamond}
$$\{ \lambda \in C \colon (Q_\lambda;\in , P \cap Q_\lambda , A_\lambda) \prec (H_\kappa;\in , P , B) \}$$
is stationary.
\end{enumerate}
We shall sometimes also write $Q_\kappa=H_\kappa$.
Readers who are familiar with Jensen's diamond will
easily see that the principle which we refer to as ($\Diamond$) is actually equivalent to $\Diamond_\kappa$; see e.g.\ \cite[Definition 5.34]{book}. We shall use ($\Diamond$) to guess information about names for club subsets of $\omega_1$;
this will play a crucial role in the verification that our forcing preserves stationary subsets of $\omega_1$.
\subsection{The definition of the forcing.}\label{3.1}
We shall now go ahead and
produce a stationary set preserving forcing ${\mathbb P} \in V[d]$ of size $\kappa$
which adds some
$p \in D^*$ and some generic iteration $$(N_i,\sigma_{ij} \colon i \leq j \leq \omega_1)$$
of $p = N_0$ such that if we write $N_{\omega_1} = (N_{\omega_1};\in,I^*, a^*)$, then $I^* = ({\sf NS}_{\omega_1})^{V[d]^{\mathbb P}} \cap N_{\omega_1}$ and $a^* = A$. As the forcing ${\rm Col}(\kappa,\kappa)$ which added $d$ is certainly stationary set preserving, this will verify Lemma \ref{main_claim}.
${\sf NS}_{\omega_1}$ is still saturated in $V[d]$. This is true simply because
forcing with ${\rm Col}(\kappa,\kappa)$ doesn't add any sequences of elements of $V$ of length $\aleph_2$. Moreover, (D.1) and (D.2) are true with $V$ being replaced by $V[d]$, as no reals are added and ${\mathbb P}_{\rm max}$ remains unchanged. Hence, in order to simplify our notation, we shall in what follows write $V$ for $V[d]$, i.e., assume that, in addition to ``${\sf NS}_{\omega_1}$ is saturated''
plus (D.1) and (D.2), ($\Diamond$) is also true in $V$.
Working under these hypotheses, we shall now
recursively define a $\subset$-increasing and continuous chain of forcings ${\mathbb P}_\lambda$ for all
$\lambda \in C \cup \{ \kappa \}$.
The forcing ${\mathbb P}$ will be ${\mathbb P}_\kappa$. The conditions in each ${\mathbb P}_\lambda$ will be finite sets of formulae of an associated first order language, ${\cal L}^\lambda$, which will be defined below. The order of each ${\mathbb P}_\lambda$ will be just reverse inclusion, i.e.,
$q \leq_{{\mathbb P}_\lambda} p$ iff $q \supset p$ for $p$, $q \in {\mathbb P}_\lambda$.\footnote{Every ${\mathbb P}_{\lambda}$ will be designed to add certain objects by means of finite sets of formulae giving partial information on these objects. As indicated by the description at the end of Section \ref{section2}, our intended objects are mentioned by the first forcing ${\mathbb P}_{\min(C)}$; the additional objects mentioned by the latter forcings are introduced in order to ensure that the final member of the sequence, i.e., ${\mathbb P}_\kappa$,
has the desired property of preserving stationary sets and adding a correct generic iteration. (So it is natural to think of ${\mathbb P}_{\min(C)}$ as supporting the ``working part'' of the conditions in ${\mathbb P}_\kappa$, and all latter forcings as supporting also ``side conditions.'') It might therefore be useful, on a first reading of the remainder of this section, to only focus on the definition and analysis of ${\mathbb P}_{\min(C)}$---in other words, to ignore everything involving clauses (C.6) to (C.8) in the definition below of potential certificate, to also ignore
clauses ($\Sigma$.6) to ($\Sigma$.8) in the definition below of potential certificates being pre-certified,
and to skip Subsection \ref{preservation-subsection}---and to pay attention to the entirety of the present section only on a second reading.}
Assume that $\lambda \in C \cup \{ \kappa \}$ and ${\mathbb P}_\mu$ has already
been defined in such a way that ${\mathbb P}_\mu \subset Q_\mu$
for all $\mu \in C \cap \lambda$. We aim to define ${\mathbb P}_\lambda$.
\begin{convention}\label{convention_real_code}
We say that $x \subset \omega$ is a real code for $N_0 = (N_0;\in,I,a)$
if
there is some surjection $f \colon \omega \rightarrow N_0$ such that
$x$ is
the monotone enumeration
of the G\"odel numbers of all expressions of the form $\ulcorner {\dot N}_0 \models \varphi({\dot n}_1, \ldots , {\dot n}_\ell, {\dot a}, {\dot I})
\urcorner$ such that
$\varphi$ is a first order formula of the language associated to $(N_0;\in,I,a)$\footnote{I.e., the language of set theory augmented by predicates for $I$ and $a$}
and
${N}_0 \models \varphi(f(n_1), \ldots , f(n_\ell),a,I)$ holds true.
\end{convention}
We shall be interested in objects ${\mathfrak C}$ which exist in some outer model
and which have the following properties. Any ${\mathfrak C}$ will be a triple of sets which are indexed by $\omega_1$, $\omega$, and $K$, respectively, with $K$ being a subset of $\omega_1$:
\begin{eqnarray}\label{certificate}
{\mathfrak C} = \langle \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle \rangle {\rm , }
\end{eqnarray}
where
\begin{enumerate}
\item[(C.1)] $M_0$, $N_0 \in {\mathbb P}_{\rm max}$,
\item[(C.2)] $x = \langle k_n \colon n<\omega \rangle$ is a real code for $N_0 = (N_0;\in,I,a)$ in the sense of Convention \ref{convention_real_code} and
$x$ is
$\langle (k_n , \alpha_n ) \colon n<\omega \rangle \in [T]$,
\item[(C.3)] $\langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1^{N_0} \rangle \in N_0$ is a generic iteration of $M_0$ which witnesses that $N_0
< M_0$ in ${\mathbb P}_{\rm max}$,
\item[(C.4)] $\langle N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle$ is a generic iteration of $N_0$ such that if $$N_{\omega_1} = (N_{\omega_1};\in,I^*,A^*){\rm , }$$ then
$A^* = A$,\footnote{There is no requirement on $I^*$ matching the non-stationary ideal of some model in which ${\mathfrak C}$ exists.\label{no-requirement}}
\item[(C.5)] $\langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1 \rangle = \sigma_{0 \omega_1}( \langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1^{N_0}\rangle )$ and \begin{eqnarray}\label{Momega1}
M_{\omega_1} = ((H_{\omega_2})^V ; \in , ({\sf NS}_{\omega_1})^V ,A){\rm , }\footnotemark
\end{eqnarray}
\footnotetext{In particular, the distingushed ideal of $M_{\omega_1}$ is the true nonstationary ideal of $V$.}
\item[(C.6)] $K \subset {\omega_1}$,
\end{enumerate}
and for all $\delta \in K$,
\begin{enumerate}
\item[(C.7)] $\lambda_\delta \in \lambda \cap C$,\footnote{Recall that $\lambda$ is the index of the forcing ${\mathbb P}_\lambda$ which we are about to define.} and if $\gamma<\delta$ is in $K$, then
$\lambda_\gamma<\lambda_\delta$ and
$X_\gamma \cup \{ \lambda_\gamma \} \subset X_\delta$, and
\item[(C.8)] $X_\delta \prec (Q_{\lambda_\delta};\in,{\mathbb P}_{\lambda_\delta},A_{\lambda_\delta})$ and $X_\delta \cap \omega_1 = \delta$.
\end{enumerate}
For future purposes, let us refer to any object ${\mathfrak C}$ as in (\ref{certificate}) which satisfies the above properties (C.1) through (C.8) as a {\em potential certificate}. As this definition of potential certificates
has $A$, $T$, $\langle A_\mu\,:\,\mu\in C \cap \lambda\rangle$, and $\langle \mathbb P_\mu\,:\,\mu\in C\cap\lambda\rangle$ as hidden parameters, we shall also refer to them as {\em potential certificates relative to} $A$, $T$, $H_{\omega_2}^V$, $\langle A_\mu\,:\,\mu\in C \cap \lambda\rangle$, and $\langle \mathbb P_\mu\,:\,\mu\in C\cap \lambda \rangle$.
We need to define a first order language ${\cal L}$ (independently from $\lambda$)
whose formulae will be able to describe some ${\mathfrak C}$ with the above properties by producing the
models $M_i$ and $N_i$, $i <\omega_1$, as term models out
of equivalence classes of terms of the form ${\dot n}$, $n<\omega$.
The language ${\cal L}$ will have the following constants.
There will be one constant for each set in $H_\kappa$; these constants will be underlined. In addition,
there will be constants for all those objects outside of $V$ which our forcing will add; those constants will be dotted.
\begin{align*}
{\underline T} & \mbox{ \ \ \ intended to denote } T \\
{\underline x} \mbox{ for every } x \in H_\kappa & \mbox{ \ \ \ intended to denote } x \mbox{ itself } \\
{\dot n} \mbox{ for every } n<\omega & \mbox{ \ \ \ as terms for elements of } M_i \mbox{ and } N_i, i < \omega_1 \\
{\dot M}_i \mbox{ for } i<\omega_1 & \mbox{ \ \ \ intended to denote } M_i \\
{\dot \pi}_{ij} \mbox{ for } i \leq j \leq \omega_1 & \mbox{ \ \ \ intended to denote } \pi_{ij} \\
{\dot {\vec M}} & \mbox{ \ \ \ intended to denote } (M_j , \pi_{jj'} \colon j \leq j' \leq
\omega_1^{N_i}) \mbox{ for } i < \omega_1 \\
{\dot N}_i \mbox{ for } i < \omega_1 & \mbox{ \ \ \ intended to denote } N_i \\
{\dot \sigma}_{ij} \mbox{ for } i \leq j < \omega_1 & \mbox{ \ \ \ intended to denote } \sigma_{ij} \\
{\dot a} & \mbox{ \ \ \ intended to denote the distinguished $a$-predicate of } M_i, N_i, i<\omega_1 \\
{\dot I}
& \mbox{ \ \ \ intended to denote the distinguished ideal of } N_i, i<\omega_1\\
{\dot X}_\delta \mbox{ for } \delta<\omega_1 & \mbox{ \ \ \ intended to denote } X_\delta.
\end{align*}
As $T \in H_{\kappa}$, the first line above is redundant. The constants ${\dot n}$, $n<\omega$, will produce the term models $N_i$ for $i<\omega_1$; it is of course not important to use (dotted) natural numbers as these constants, the elements of any other fixed countable set (in $H_\kappa$) would be equally good.
The formulae of ${\cal L}$ will be exactly the expressions of the following form.\footnote{Again, the first order language associated with ${\mathbb P}_{\rm max}$-structures $(M;\in,I,a)$ is the language of set theory augmented by predicates for $I$ and $a$.
\begin{gather*}
\ulcorner {\dot N}_i \models \varphi({\underline \xi}_1, \ldots , {\underline \xi}_k , {\dot n}_1 , \ldots , {\dot n}_\ell, {\dot a}, {\dot I} ,
{\dot M}_{j_1}, \ldots, {\dot M}_{j_m} , {\dot \pi}_{q_1 r_1} , \ldots , {\dot \pi}_{q_s r_s} , {\dot {\vec M}}) \urcorner \\
\mbox{for } i<\omega_1,
\xi_1, \ldots , \xi_k < \omega_1, n_1, \ldots , n_\ell < \omega, j_1, \ldots, j_m < \omega_1,
q_1 \leq r_1 < \omega_1, \ldots, q_s \leq r_s < \omega_1 \\
\mbox{and for } \varphi \mbox{ being a formula of the first order language associated with ${\mathbb P}_{\rm max}$-structures,}
\end{gather*}
as well as:
\begin{align*}
\ulcorner {\dot \pi}_{i \omega_1}({\dot n})={\underline x} \urcorner & \mbox{ \ \ \ for }
i < \omega_1 \mbox{ and } x \in H_{\omega_2} \\
\ulcorner {\dot \pi}_{\omega_1 \omega_1}({\underline x})={\underline x} \urcorner & \mbox{ \ \ \ for }
x \in H_{\omega_2}\\
\ulcorner {\dot \sigma}_{ij}({\dot n})={\dot m} \urcorner & \mbox{ \ \ \ for }
i \leq j < \omega_1, n, m<\omega \\
\ulcorner ({\underline {\vec u}},{\underline {\vec \alpha}}) \in {\underline T} \urcorner & \mbox{ \ \ \
for } {\vec u} \in {}^{<\omega} \omega \mbox{ and } {\vec \alpha}
\in {}^{<\omega} \omega_2 \\
\ulcorner {\underline \delta} \mapsto {\underline \mu} \urcorner & \mbox{ \ \ \ for } \delta<\omega_1,
{\mu}<\kappa \\
\ulcorner {\underline x} \in {\dot X}_\delta \urcorner & \mbox{ \ \ \ for } \delta<\omega_1, x \in H_\kappa
\end{align*}
Let us write ${\cal L}^\lambda$ for the collection of all ${\cal L}$-formulae except for the
formulae which mention elements outside of $Q_\lambda$, i.e., except for the formulae
of the form $\ulcorner {\underline \delta} \mapsto {\underline \mu} \urcorner$
for $\delta < \omega_1$ and ${\lambda} \leq {\mu} < \kappa$ as well as $\ulcorner {\underline x} \in {\dot X}_\delta \urcorner$ for $\delta < \omega_1$ and $x
\in H_\kappa \setminus Q_\lambda$.
We may and shall assume that ${\cal L}$ is built in a canonical way so that
${\cal L}^\lambda \subset Q_\lambda$ and in fact ${\cal L}^\lambda = {\cal L} \cap Q_\lambda$.
We say that a potential certificate
\begin{eqnarray*}
{\mathfrak C} = \langle \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle \rangle
\end{eqnarray*}
as in
(\ref{certificate}) is {\em pre-certified} by a collection $\Sigma$ of ${\cal L}^\lambda$-formulae
if and only if (C.1) through (C.8) are satisfied by ${\mathfrak C}$
and there are surjections\label{the_surjections} $e_i \colon \omega \rightarrow
N_i$
for $i<\omega_1$
such that the following hold true.
\begin{enumerate}
\item[($\Sigma$.1)]
$\ulcorner {\dot N}_i \models \varphi({\underline \xi}_1, \ldots , {\underline \xi}_k , {\dot n}_1 , \ldots , {\dot n}_\ell, {\dot a}, {\dot I} ,
{\dot M}_{j_1}, \ldots, {\dot M}_{j_m} , {\dot \pi}_{q_1 r_1} , \ldots , {\dot \pi}_{q_s r_s} , {\dot {\vec M}})
\urcorner \in \Sigma$ iff
\begin{enumerate}
\item[(a)] $i<\omega_1$,
\item[(b)] $\xi_1$, $\ldots$, $\xi_k \leq \omega_1^{N_i}$,
\item[(c)] $n_1$, $\ldots$ , $n_\ell < \omega$,
\item[(d)] $j_1$, $\ldots$, $j_m \leq \omega_1^{N_i}$,
\item[(e)] $q_1 \leq r_1 \leq \omega_1^{N_i}$, $\ldots$, $q_s \leq r_s \leq \omega_1^{N_i}$, and
\end{enumerate}
$$N_i \models \varphi(\xi_1, \ldots , \xi_k , e_i(n_1) , \ldots , e_i(n_\ell), A \cap \omega_1^{N_i}, I^{N_i} , M_{j_1}, \ldots
, M_{j_m}, \pi_{q_1 r_1}, \ldots , \pi_{q_s r_s}, {\vec M}){\rm , }$$ where $I^{N_i}$ is the distinguished ideal of $N_i$ and ${\vec M}=\langle M_j , \pi_{jj'} \colon
j \leq j' \leq \omega_1^{N_i})$,
\item[($\Sigma$.2)] $\ulcorner {\dot \pi}_{i \omega_1}({\dot n})={\underline x} \urcorner \in \Sigma$ iff $i < \omega_1$, $n<\omega$, and $\pi_{i \omega_1}(e_i(n))=x$,
\item[($\Sigma$.3)] $\ulcorner {\dot \pi}_{\omega_1 \omega_1}({\underline x})={\underline x} \urcorner \in \Sigma$ iff $x \in H_{\omega_2}$,
\item[($\Sigma$.4)] $\ulcorner {\dot \sigma}_{ij}({\dot n})={\dot m} \urcorner \in \Sigma$ iff $i \leq j < \omega_1$, $n$, $m<\omega$, and $\sigma_{ij}(e_i(n))=e_j(m)$,
\item[($\Sigma$.5)] letting $F$ with ${\rm dom}(F)=\omega$ be the monotone enumeration
of the G\"odel numbers of all formulae of the form $$\ulcorner {\dot N}_0 \models \varphi({\dot n}_1, \ldots , {\dot n}_\ell, {\dot a}, {\dot I})
\urcorner$$ with $\ulcorner {\dot N}_0 \models \varphi({\dot n}_1, \ldots , {\dot n}_\ell, {\dot a}, {\dot I})
\urcorner \in \Sigma$, we have that $\ulcorner ({\underline {\vec u}},{\underline {\vec \alpha}}) \in {\underline T} \urcorner \in \Sigma$ iff
there is some $n<\omega$ such that $\langle {\vec u},{\vec \alpha} \rangle
= \langle (F(m),\alpha_m ) \colon m< n \rangle$ and $F(m)=k_m$ for all $m<n$,\footnote{Here, $\langle (k_n,\alpha_n) \colon n<\omega \rangle$ is a component of ${\mathfrak C}$.}
\item[($\Sigma$.6)] $\ulcorner {\underline \delta} \mapsto {\underline \mu} \urcorner \in \Sigma$ iff
$\delta \in K$ and ${\mu} = \lambda_\delta$, and
\item[($\Sigma$.7)] $\ulcorner {\underline x} \in {\dot X}_\delta \urcorner \in \Sigma$ iff
$\delta \in K$ and $x \in X_\delta$.
\end{enumerate}
We say that a potential certificate ${\mathfrak C}$ as in
(\ref{certificate}) is {\em certified} by a collection $\Sigma$ of formulae
if and only if ${\mathfrak C}$ is pre-certified by $\Sigma$ and, in addition,
\begin{enumerate}
\item[($\Sigma$.8)]\label{(C.9)} if $\delta \in K$, then
$[\Sigma]^{<\omega} \cap X_\delta \cap
E \not= \emptyset$ for every $E \subset {\mathbb P}_{\lambda_\delta}$ which is dense in ${\mathbb P}_{\lambda_\delta}$ and definable over the structure $$(Q_{\lambda_\delta}; \in , {\mathbb P}_{\lambda_\delta} , A_{\lambda_\delta} )$$ from parameters in $X_\delta$.\footnote{Equivalently,
$[\Sigma]^{<\omega} \cap
E \not= \emptyset$ for every $E \subset {\mathbb P}_{\lambda_\delta} \cap
X_\delta$ which is dense in ${\mathbb P}_{\lambda_\delta} \cap
X_\delta$ and definable over the structure $$(X_\delta; \in , {\mathbb P}_{\lambda_\delta} \cap
X_\delta , A_{\lambda_\delta} \cap X_\delta )$$ from parameters in $X_\delta$.}
\end{enumerate}
Item ($\Sigma$.8) is to play a crucial role in the proof that our forcing preserves stationary sets, see the proof of Lemma \ref{P_stat_pres}. If $p$ is a condition with $\ulcorner \underline{\delta} \mapsto \underline{\lambda} \urcorner \in p$, and if the predicate $A_\lambda$ guesses -- via ($\Diamond$) -- a name ${\dot C}$ for a club subset of $\omega_1$, then ($\Sigma$.8) will guarantee that $p \Vdash {\check \delta} \in {\dot C}$, see the proof of Claim \ref{claim2.10}.
\begin{definition}\label{semc} Let ${\mathfrak C}$ as in (\ref{certificate})
be a potential certificate relative to $A$, $T$, $H_{\omega_2}^V$, $\langle A_\mu\,:\,\mu\in C \cap \lambda\rangle$, and $\langle \mathbb P_\mu\,:\,\mu\in C\cap\lambda\rangle$. We call ${\mathfrak C}$ a {\em semantic
certificate relative to} $A$, $T$, $H_{\omega_2}^V$, $\langle A_\mu\,:\,\mu\in C\cap \lambda\rangle$, and $\langle \mathbb P_\mu\,:\,\mu\in C\cap\lambda\rangle$, or just a {\em semantic
certificate}, iff
there is a
collection $\Sigma$ of formulae such that ${\mathfrak C}$ is certified by $\Sigma$.
We call $\Sigma$ a {\em syntactic certificate}
iff there is a semantic certificate ${\mathfrak C}$ such that ${\mathfrak C}$ is certified by
$\Sigma$.
\end{definition}
In the proofs of Lemma \ref{P_is_nonempty} and of Claim \ref{its-a-sem-c} we will run Definition \ref{semc} inside a generic iterate $M$ of $V$. The corresponding iteration map $\pi$ will move the parameters $A$, $T$, $H_{\omega_2}^V$, $\langle A_\mu\,:\,\mu\in C\cap \lambda\rangle$, and $\langle \mathbb P_\mu\,:\,\mu\in C\cap\lambda\rangle$, so that we will talk about semantic certificates relative to the parameters $\pi(A)$, $\pi(T)$, $\pi(H_{\omega_2}^V)$, $\pi(\langle A_\mu\,:\,\mu\in C \cap \lambda\rangle)$, and $\pi(\langle \mathbb P_\mu\,:\,\mu\in C\cap\lambda\rangle)$ with the understanding that this will be an object obtained by running Definition \ref{semc} over $M$ with these shifted parameters.
Given a syntactic certificate $\Sigma$, there is a {\em unique} semantic
certificate ${\mathfrak C}$ such that ${\mathfrak C}$ is certified by $\Sigma$.
Even though it is obvious how to construct ${\mathfrak C}$ from $\Sigma$, in the proof of Lemma \ref{generics_give_certificates} below we \label{details} will provide details on how to derive a semantic certificate from a given $\Sigma$.
It is worth stressing that not every collection of ${\cal L}^\lambda$-formulae which is merely consistent is already a syntactic certificate. The requirement that the constant $x \in H_\kappa$ is to be interpreted by itself (cf.\ ($\Sigma$.2), ($\Sigma$.3), and
($\Sigma$.7)) may be restated as saying that for a consistent
${\cal L}^\lambda$-theory to be a syntactic certificate it is to be true that certain types are omitted.
Let $\Sigma$ and $p$ be sets of formulae, where $p$ is finite.
We say that $p$
is {\em certified by}
$\Sigma$ if and only if
there is some (unique) ${\mathfrak C}$ as in (\ref{certificate})
such that
${\mathfrak C}$ is certified by
$\Sigma$ and
\begin{enumerate}
\item[($\Sigma$.9)] $p \in [\Sigma]^{<\omega}$.
\end{enumerate}
We may also say that $p$ is {\em certified by} ${\mathfrak C}$ as in (\ref{certificate}) iff there is some $\Sigma$ such that ${\mathfrak C}$ and $p$ are both certified by $\Sigma$---and we will then also refer to $\Sigma$ as a syntactical certificate for $p$ and to ${\mathfrak C}$ as the associated semantic certificate.
We are then ready to define the forcing ${\mathbb P}_\lambda$.
We say that $p \in {\mathbb P}_\lambda$ if and only if
\begin{eqnarray}\label{in_gen_ext}
V^{{\rm Col}(\omega,\lambda)} \models \mbox{``There is a set $\Sigma$ of ${\cal L}^\lambda$-formulae such that } p \mbox{ is certified by } \Sigma \mbox{.''}
\end{eqnarray}
Let $p$ be a finite set of formulae of ${\cal L}^\lambda$. By the homogeneity of ${\rm Col}(\omega,\lambda)$, if there is some $h$ which is
${\rm Col}(\omega,\lambda)$-generic over $V$ and there is some
$\Sigma
\in V[h]$
such that $p$ is certified by $\Sigma$, then for all $h$ which are ${\rm Col}(\omega,\lambda)$-generic over $V$ there is some $\Sigma \in V[h]$ such that $p$ is certified by $\Sigma$.
It is then easy to see that $\langle {\mathbb P}_\lambda \colon \lambda \in C \cup
\{ \kappa \} \rangle$ is definable over $V$ from $\langle A_\lambda \colon
\lambda < \kappa \rangle$
and $C$, and is hence an element of $V$.\footnote{To remind the reader, $C$ is the club from p.\ \pageref{defn-club}.}
The following absoluteness fact will be relevant in the proofs of Lemma \ref{P_is_nonempty} and of Claim \ref{its-a-sem-c}.
\begin{lemma}\label{abs_of_ex_of_certificate}
Let $\lambda \in C \cup \{ \kappa \}$, and let $p$ be a finite set of formulae of ${\cal L}^\lambda$. If there is any outer model in which there is some $\Sigma$ which
certifies $p$, then there is some $\Sigma \in V^{{\rm Col}(\omega,\lambda)}$
which certifies $p$.
\end{lemma}
{\sc Proof.}
The statement that there is a $\Sigma$ which
certifies $p$ is $\Sigma_1$ in the parameters $A$, $T$, $H_{\omega_2}^V$, $\langle A_\mu \colon \mu \in C \cap \lambda \rangle$, and $\langle {\mathbb P}_\mu \colon \mu \in C \cap \lambda \rangle$, all of which are elements of ${\rm HC}^{V^{{\rm Col}(\omega,\lambda)}}$. Hence
by Shoenfield absoluteness (\ref{el_equivalent_shoenfield}), see \cite[Corollary 7.21]{book}, if there is any outer model in which there is some $\Sigma$ which
certifies $p$, then there is some $\Sigma \in V^{{\rm Col}(\omega,\lambda)}$
which certifies $p$.\footnote{In fact, if $P$ is a transitive model of {\sf KP} plus the axiom
{\em Beta} with $A$, $T$, $H_{\omega_2}^V$, $\langle A_\mu \colon \mu \in C \cap \lambda \rangle$, $\langle {\mathbb P}_\mu \colon \mu \in C \cap \lambda \rangle \in P$ and if $p \in {\mathbb P}_\lambda$,
then there is some $\Sigma \in P^{{\rm Col}(\omega,\lambda)}$ which certifies $p$.}\label{abs_of_cert} \hfill $\square$ (Lemma \ref{abs_of_ex_of_certificate})
\subsection{Some properties of the forcing.}
It is easy to see that \begin{enumerate}
\item[(i)] ${\mathbb P}={\mathbb P}_\kappa \subset H_\kappa$,
\item[(ii)] if ${\bar \lambda} < \lambda$ are both in $C \cup \{ \kappa \}$, then ${\mathbb P}_{\bar \lambda} \subset {\mathbb P}_\lambda$, and
\item[(iii)] if $\lambda \in C \cup \{ \kappa \}$ is a limit point of
$C \cup \{ \kappa \}$, then ${\mathbb P}_\lambda = \bigcup_{{\bar \lambda} \in C \cap \lambda} \, {\mathbb P}_{\bar \lambda}$,
\end{enumerate}
so that there is some club $D \subset C$ such that for all $\lambda \in D$,
$${\mathbb P}_\lambda = {\mathbb P} \cap Q_\lambda.$$ Hence ($\Diamond$) gives us the following.
\begin{enumerate}\label{diamond-thing}
\item[($\Diamond({\mathbb P})$)] For all $B \subset H_\kappa$ the set
$$\{ \lambda \in C \colon (Q_\lambda;\in , {\mathbb P}_\lambda , A_\lambda) \prec (H_\kappa;\in , {\mathbb P} , B) \}$$
is stationary.
\end{enumerate}
The first one of the following lemmas is entirely trivial.
\begin{lemma}\label{+++}
Let $\Sigma$ be a syntactic certificate, and let $p$, $q \in [\Sigma]^{<\omega}$. Then $p$ and $q$ are compatible conditions in ${\mathbb P}$.
\end{lemma}
\begin{lemma}\label{P_is_nonempty}
$\emptyset \in {\mathbb P}_{\min(C)}$.
\end{lemma}
{\sc Proof.} This is a simple variant of the proofs of \cite[Theorem 2.8]{david-ralf} and of
\cite[Theorem 4.2]{both}. What needs to be done is to construct a semantic/syntactic certificate (for $\emptyset$) in some outer model.
Let
$h$ be ${\rm Col}(\omega,\omega_2)$-generic over $V$.
Let $q^* = (N^*;\in,I^*, a^*) \in ({\mathbb P}_{\rm max})^{V[h]}$
be as in the paragraph preceeding (\ref{q-star}), i.e.,
$q^* \in ({\mathbb P}_{\rm max})^{V[h]}$, $q^* < (H_{\omega_2}^V; \in,
{\sf NS}_{\omega_1}^V,A)$, $q^* \in D^*$, and such that (\ref{q-star}) is true, i.e., $q^* \in p[{T}] \subset p[{\tilde T}]$, cf.\ Lemma \ref{ex_of_T}.
Let $(M_i, \pi_{ij} \colon i \leq j \leq \omega_1^{N^*}) \in N^*$ be the unique generic iteration of the $({\mathbb P}_{\rm max})^{V[h]}$-condition $(H_{\omega_2}^V; \in,
{\sf NS}_{\omega_1}^V,A)$ which witnesses that $q^*$ is stronger than this condition.
Let $(N_i,\sigma_{ij} \colon i \leq j \leq \kappa) \in V[h]$ be a generic iteration of
$N_0 = N^*$ such that $\kappa=\omega_1^{N_\kappa}$.\footnote{If we wished, we could even arrange that
writing $N_\kappa = (N_\kappa;\in,I',a')$, we have that $I' = ({\sf NS}_\kappa)^{V[h]}
\cap N_\kappa$, but this is not relevant here; cf.\ footnote \ref{no-requirement}.} Let
\begin{eqnarray}\label{it_of_p}
(M_i , \pi_{ij} \colon i \leq j \leq \kappa)=
\sigma_{0 \kappa}((M_i, \pi_{ij} \colon i \leq j \leq \omega_1^{N_0}))
\end{eqnarray}
Since $M_{0} = ((H_{\omega_2})^V; \in, ({\sf NS}_{\omega_1})^V, A)$ and $({\sf NS}_{\omega_1})^V$ is assumed to be saturated in $V$,
every maximal antichain in $V$ consisting of stationary subsets of $\omega_1$ is an element of $M_{0}$.
By \cite[Lemma 3.8]{hugh},
we may hence lift the generic ultrapower map
$\pi_{01} \colon M_{0} \rightarrow M_{1}$ to act on all of $V$, and inductively we may
lift the entire generic iteration (\ref{it_of_p})
to a generic iteration
\begin{eqnarray}\label{it_of_pV}
(M_i^+ , \pi^+_{ij} \colon i \leq j \leq \kappa)
\end{eqnarray}
of $V$ in such a way that all $M_i^+$, $i \leq \kappa$, are transitive.
Let us write $M=M^+_\kappa$ and $\pi = \pi^+_{0 \kappa}$.
Let $\langle k_n , \alpha_n \colon n<\omega \rangle$ be such that $x
= \langle k_n \colon n<\omega \rangle$ is
a real code for $N_0$ in the sense of Convention \ref{convention_real_code} and $\langle (k_n , \alpha_n) \colon n<\omega \rangle \in [T]$.
We then clearly have that $\langle (k_n , \pi(\alpha_n)) \colon n<\omega \rangle \in [\pi(T)]$.
\bigskip
\begin{tikzpicture}
\def\x{2.1};
\def\y{0.5};
\node (pT) at (0*\x, 0*\y){$p[T]$};
\node (ppiT) at (0*\x+1.5, 0*\y){$p[\pi(T)]$};
\node (N) at (0*\x, -2*\y) {$(N^*;\in, I^*, a^*)$};
\node (Nast) at (2*\x, -2*\y) {$N_\kappa$};
\node (M0) at (-2*\x, -4*\y) {$M_0$};
\node (MN) at (0*\x, -4*\y) {$M_{\omega_1^{N^*}}$};
\node (M3) at (2*\x, -4*\y) {$M_{\kappa}$};
\node (H) at (-2*\x, -6*\y) {$((H_{\omega_2})^V;\in, (\mathsf{NS}_{\omega_1})^V, A)$};
\node (V) at (-2*\x, -8*\y) {$V$};
\node (Mp) at (2*\x, -8*\y) {$M_\kappa^+$};
\node (M) at (2*\x+1, -8*\y) {$M$};
\path (pT)--(ppiT) node[midway]{$\subseteq$};
\path (N)--(pT) node[midway, rotate=90]{$\in$};
\draw[->] (N)--(Nast) node[midway, above]{$\sigma_{0\kappa}$};
\path (MN)--(N) node[midway, rotate=90]{$\in$};
\path (M3)--(Nast) node[midway, rotate=90]{$\in$};
\draw[->] (M0)--(MN) node[midway, above]{$\pi_{0 \omega_1^{N^*}}$};
\draw[->] (MN)--(M3) node[midway, above]{$\pi_{\omega_1^{N^*} \kappa}$};
\path (H)--(M0) node[midway, rotate=90]{$=$};
\path (M3)--(Mp) node[midway, rotate=270]{$\in$};
\path (H)--(V) node[midway, rotate=270]{$\in$};
\draw[->] (V)--(Mp) node[midway, above]{$\pi$};
\path (Mp)--(M) node[midway]{$=$};
\end{tikzpicture}
\bigskip
It is now easy to see that
\begin{eqnarray}\label{certificate_at_rho}
{\mathfrak C} = \langle \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \kappa \rangle ,
\langle (k_n , \pi(\alpha_n) ) \colon n<\omega \rangle ,
\langle \rangle \rangle
\end{eqnarray}
certifies $\emptyset$ relative to the parameters $\pi(A)$, $\pi(T)$, $\pi(H_{\omega_2}^V)$, $\langle \rangle$, and $\langle \rangle$,
with $\emptyset$ being construed as the empty set of $\pi({\cal L}^\kappa)$ formulae: as the third
component $\langle \rangle$ of ${\mathfrak C}$ in (\ref{certificate_at_rho}) is empty, {\em any} set of surjections $e_i \colon \omega \rightarrow N_i$, $i<\omega_1$, will induce a syntactic certificate for $\emptyset$ relative to $\pi(A)$, $\pi(T)$, $\pi(H_{\omega_2}^V)$, $\langle \rangle$, and $\langle \rangle$, whose associated semantic certificate is ${\mathfrak C}$.
The statement that there is a syntactic certificate for $\emptyset$ is $\Sigma_1$ in the parameters $H_{\omega_2}^M$, $\pi(A)$, and $\pi(T)$,
which will all be in ${\rm HC}^{M^{{\rm Col}(\omega,\pi(\omega_2))}}$. Hence
by Shoenfield absoluteness (\ref{el_equivalent_shoenfield}), see \cite[Corollary 7.21]{book}, there is then some
${\mathfrak C} \in M^{{\rm Col}(\omega,\pi(\omega_2))}$ as in (\ref{certificate_at_rho}) which certifies $\emptyset$ relative to $\pi(A)$, $\pi(T)$, $\pi(H_{\omega_2}^V)$, $\langle \rangle$, and $\langle \rangle$, so that
$\emptyset \in \pi(\mathbb P_{\min(C)})$, cf.\ Lemma \ref{abs_of_ex_of_certificate}.\footnote{Exactly in order to be able to do this we let the forcing also search for $\langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1 \rangle$ rather than just $\langle N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle$. The presence of
$\langle M_i , \pi_{ij} \colon i \leq j \leq \omega_1 \rangle$ allows us to lift $\pi_{0\kappa}$ to a map acting on all of $V$, so that we may then apply Shoenfield absoluteness and pull back the statement of interest---namely
$\emptyset \in \pi(\mathbb P_{\min(C)})$. Cf.\ the discussion on p.\ \pageref{discussion_why_M}. \label{footnote}} By the elementarity of $\pi$,
then,
$\emptyset \in {\mathbb P_{\min(C)}}$.
\hfill $\square$ (Lemma \ref{+++})
\begin{lemma}\label{generics_give_certificates}
Let $\lambda \in C \cup \{ \kappa \}$.
Let $g \subset {\mathbb P}_\lambda$ be a filter such that $g \cap E \not= \emptyset$ for all dense $E \subset {\mathbb P}_\lambda$ which are definable over $(Q_\lambda;\in,{\mathbb P}_\lambda)$ from elements of $Q_\lambda$. Then
$\bigcup g$
is a syntactic certificate.
\end{lemma}
{\sc Proof.}
Let us first describe how to read off from $\bigcup g$ a candidate
$${\mathfrak C} = \langle \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle \rangle$$
for
a semantic certificate for $\bigcup g$.
A variant of what is to come shows
how to derive ${\mathfrak C}$ from a given syntactic certificate $\Sigma$,
where ${\mathfrak C}$ is unique such that $\Sigma$ certifies ${\mathfrak C}$, cf.\ the remark on p.\ \pageref{details}.
Some of the formulas to follow simply describe the construction of a direct limit associated with the maps $\pi_{ij}$ and $\sigma_{ij}$.
For
$i$, $j <\omega_1$ and
$\tau$, $\sigma \in \{ {\dot n} \colon n<\omega \} \cup \{ {\underline \xi} \colon \xi < \omega_1 \}$
define
\begin{align*}
\tau \sim_{i} \sigma & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau = \sigma \urcorner \in \bigcup g \\
(i,\tau) \sim_{\omega_1} (j,\sigma) & \mbox{ \ \ \ iff } i \leq j \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ij}(\tau)=\rho \urcorner , \ulcorner {\dot N}_j \models
\rho = \sigma \urcorner \} \subset \bigcup g\\
{} & \mbox{ \ \ \ \ or } j \leq i \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ji}(\sigma)=\rho \urcorner , \ulcorner {\dot N}_i \models
\rho = \tau \urcorner \} \subset \bigcup g\\
[\tau]_{i} & = \{ \sigma \colon \tau \sim_{i} \sigma \} \\
[(i,\tau)] & = \{ (j,\sigma) \colon (i,\tau) \sim_{\omega_1} (j,\sigma) \} \\
{M}_i & = \{ [\tau]_i \colon \tau \in \{ {\dot n} \colon n<\omega \} \cup \{ {\underline \xi} \colon \xi < \omega_1 \} \wedge \ulcorner {\dot N}_i \models \tau \in {\dot M}_i \urcorner \in \bigcup g \} \\
M_{\omega_1} & = (H_{\omega_2})^V \\
{N}_i & = \{ [\tau]_i \colon \tau \in \{ {\dot n} \colon n<\omega \} \cup \{ {\underline \xi} \colon \xi < \omega_1 \} \} \\
N_{\omega_1} & = \{ [i,\tau] \colon i<\omega_1 \wedge \ulcorner {\dot N}_i \models \tau = \tau \urcorner \in \bigcup g \} \\
[\tau]_{i} \, {\tilde \in}_i \, [\sigma]_{i} & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau \in \sigma \urcorner \in \bigcup g \\
[i,\tau] {\tilde \in}_{\omega_1} [j,\sigma] & \mbox{ \ \ \ iff }
i \leq j \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ij}(\tau)=\rho \urcorner , \ulcorner {\dot N}_j \models
\rho \in \sigma \urcorner \} \subset \bigcup g\\
{} & \mbox{ \ \ \ \ or } j \leq i \wedge \exists
\rho \, \{ \ulcorner
{\dot \sigma}_{ji}(\sigma)=\rho \urcorner , \ulcorner {\dot N}_i \models
\tau \in \rho \urcorner \} \subset \bigcup g\\
[\tau]_i \in I^{N_i} & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau \in {\dot I} \urcorner \in \bigcup g \\
[i,\tau] \in I^{N_{\omega_1}} & \mbox{ \ \ \ iff \ \ } [\tau]_i \in I^{N_i} \\
[\tau]_i \in a^{N_i} & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_i \models \tau \in {\dot a} \urcorner \in \bigcup g \\
[i,\tau] \in a^{N_{\omega_1}} & \mbox{ \ \ \ iff \ \ } [\tau]_i \in I^{N_i} \\
\pi_{ij}([\tau]_{i}) = [\sigma]_j & \mbox{ \ \ \ iff \ \ } \ulcorner {\dot N}_j \models
{\dot \pi}_{ij}(\tau)=\sigma \urcorner \in \bigcup g \\
\pi_{i \omega_1}([\tau]_{i}) = x & \mbox{ \ \ \ iff \ \ } \ulcorner
{\dot \pi}_{i \omega_1}(\tau)={\underline x} \urcorner \in \bigcup g \\
\pi_{\omega_1 \omega_1}(x) = x & \mbox{ \ \ \ iff \ \ } x \in (H_{\omega_2})^V \\
\sigma_{ij}([\tau]_{i}) = [\sigma]_j & \mbox{ \ \ \ iff \ \ } \ulcorner
{\dot \sigma}_{ij}(\tau)=\sigma \urcorner \in \bigcup g \\
\sigma_{i \omega_1}([\tau]_i) = [i,\tau] & {} \\
(k,\alpha) = (k_n, \alpha_n) & \mbox{ \ \ \ iff } \exists {\vec u} \exists
{\vec \alpha} ( \ulcorner ({\underline {\vec u}} , {\underline {\vec \alpha}}) \in {\underline T} \urcorner
\in \bigcup g \wedge k = {\vec u}(n) \wedge \alpha = {\vec \alpha}(n) )\\
\delta \in K & \mbox{ \ \ \ iff } \exists {\mu} \, \ulcorner {\underline \delta} \mapsto {\underline \mu} \urcorner \in \bigcup g\\
{\mu} = \lambda_\delta & \mbox{ \ \ \ iff } \delta \in K \wedge \ulcorner {\underline \delta} \mapsto {\underline \mu} \urcorner \in \bigcup g\\
x \in X_\delta & \mbox{ \ \ \ iff } \delta \in K \wedge \ulcorner {\underline x} \in {\dot X}_\delta \urcorner \in \bigcup g
\end{align*}
In order to see that this all works out we have to run a few density arguments. To show that a given subset of ${\mathbb P}$ is dense we frequently
make use of Lemma \ref{+++}.
We will provide more details in some cases and fewer in others, and we are confident that the reader will be easily able to fill in the straightforward details herself in the latter cases.
Let us first observe that
${\tilde \in}_0$ is wellfounded and that in fact (the transitive collapse of)
the structure $N_0 = (N_0;{\tilde \in}_0,a^{N_0},I^{N_0})$ is an iterable ${\mathbb P}_{\rm max}$ condition. This is true because of the following.
\begin{claim}\label{N_0}
(C.2) is true, i.e., $\langle (k_n, \alpha_n) \colon n<\omega \rangle \in [T]$ and $\langle k_n \colon
n<\omega \rangle$ codes the theory of $N_0$ in the sense of Convention \ref{convention_real_code}.
\end{claim}
{\sc Proof} of Claim \ref{N_0}. Let $m<\omega$. Writing $$q_0 = \{ \ulcorner (\underline{(k_n \colon n<m)},\underline{(\alpha_n \colon n<m)}) \in {\dot T} \urcorner \} {\rm , }$$ we have that $q_0 \in g.$ If
$${\mathfrak C} = \langle \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X'_\delta \colon \delta \in K' \rangle \rangle$$ certifies
$q_0$, then $k'_n = k_n$ and $\alpha'_n = \alpha_n$ for $n<m$ by ($\Sigma.5$),
and then $$((k_n \colon n<m),(\alpha_n \colon n<m)) \in T$$ by (C.2).
This shows $\langle (k_n, \alpha_n) \colon n<\omega \rangle \in [T]$.
By ($\Sigma.5$) and (C.2), for each $k<\omega$ the sets
\begin{eqnarray*}
D^0_k & = & \{ p \in {\mathbb P} \colon \exists m \, \exists ((k_n \colon n<m),(\alpha_n \colon n<m)) \exists r \, \\
{} & {} & ( \, \ulcorner (\underline{(k_n \colon n<m)},\underline{(\alpha_n \colon n<m)}) \in {\underline T} \urcorner \in p \wedge k_r = k \\
{} & {} & \wedge \, k \mbox{ is the G\"odel number of } \ulcorner {\dot N}_0 \models \varphi({\dot n}_1 , \ldots , {\dot n}_\ell , {\dot a} , {\dot I}) \urcorner \, ) \rightarrow \\
{} & {} & \ulcorner {\dot N}_0 \models \varphi({\dot n}_1 , \ldots , {\dot n}_\ell , {\dot a} , {\dot I}) \urcorner \in p \}
\end{eqnarray*} and
\begin{eqnarray*}
D^1_k & = & \{ p \in {\mathbb P} \colon k \mbox{ is the G\"odel number of } \ulcorner {\dot N}_0 \models \varphi({\dot n}_1 , \ldots , {\dot n}_\ell , {\dot a} , {\dot I}) \urcorner \wedge \\
{} & {} & \ulcorner {\dot N}_0 \models \varphi({\dot n}_1 , \ldots , {\dot n}_\ell , {\dot a} , {\dot I}) \urcorner \in p \rightarrow \\
{} & {} & \exists m \, \exists ((k_n \colon n<m),(\alpha_n \colon n<m)) \exists r \,
\, ( \, k_r = k \wedge \\
{} & {} & \ulcorner (\underline{(k_n \colon n<m)},\underline{(\alpha_n \colon n<m)}) \in {\underline T} \urcorner \in p \, ) \}
\end{eqnarray*}
are dense in ${\mathbb P}$. This implies that $\langle k_n \colon
n<\omega \rangle$ codes the theory of $N_0$ in the sense of Convention \ref{convention_real_code}.
\hfill $\square$ (Claim \ref{N_0})
\bigskip
Another set of easy density arguments will give that $(N_i ,\sigma_{ij} \colon i \leq j \leq \omega_1)$
is a generic iteration of $N_0$, where we identify $N_i$ with the structure
$(N_i;{\tilde \in}_i,a^{N_i},I^{N_i})$.
To verify this, let us first show:
\begin{claim}\label{claim2.4}
For each $i < \omega_1$ and for each $\xi \leq \omega_1^{N_i}$, $[\underline{\xi}]_i$ represents
$\xi$ in (the transitive collapse of the well-founded part of) the term model for $N_i$;
moreover, $a^{N_i} = A \cap \omega_1^{N_i}$. Hence $a^{N_{\omega_1}}=A$.
\end{claim}
{\sc Proof} of Claim \ref{claim2.4}. The set $$D^2 = \{ p \in {\mathbb P} \colon
\exists \xi \, \ulcorner {\dot N}_i \models \underline{\xi} \mbox{ is the least uncountable cardinal } \urcorner \in p \}$$ is easily seen to be dense in ${\mathbb P}$, so that there is some (unique!) $\xi_0<\omega_1$ such that writing $$p_0 = \{ \ulcorner {\dot N}_i \models \underline{\xi}_0 \mbox{ is the least uncountable cardinal } \urcorner \} {\rm , }$$
$p_0 \in g$.
Let us now prove by induction on $\xi \leq \xi_0$
that $[\underline{\xi}]_i$ must always represent $\xi$ in (the transitive collapse of the well-founded part of) $N_i$. Fix such $\xi$.
For all $n<\omega$,
$$D_{n}^3 = \{ p \in {\mathbb P} \colon \ulcorner {\dot N}_i \models {\dot n} \in
\underline{\xi} \urcorner \in p \rightarrow \exists \underline{\zeta} < \xi \, \ulcorner {\dot N}_i \models {\dot n} =
\zeta \urcorner \in p \}$$
is dense below $p_0$. Also, for all $\zeta < \xi$,
$$D_{\zeta}^4 = \{ p \in {\mathbb P} \colon \ulcorner {\dot N}_i \models
\underline{\zeta} \in \underline{\xi} \urcorner \in p \}$$ is dense below $p_0$. This shows that if $\tau \in \{ {\dot n} \colon n<\omega \} \cup \{ \underline{\xi} \colon \xi \in \xi_0+1 \}$ is a term, then $[\tau]_i {\tilde \in}_i [\underline{\xi}]_i$ iff $[\tau]_i = [\underline{\zeta}]_i$ for some $\zeta<\xi$. Using the inductive hypothesis, this then implies that $[\underline{\xi}]_i$ represents $\xi$ in (the transitive collapse of the well-founded part of) $N_i$. In particular, $\xi_0 = \omega_1^{N_i}$.
Now if $\xi \in A \cap \omega_1^{N_i}$, then $$D^5_\xi = \{ p \in {\mathbb P} \colon \ulcorner {\dot N}_i \models \underline{\xi} \in {\dot a} \urcorner \}$$ is dense below $p_0$, and if $\xi \in \omega_1^{N_i} \setminus A$, then $$D^6_\xi = \{ p \in {\mathbb P} \colon \ulcorner {\dot N}_i \models \underline{\xi} \notin {\dot a} \urcorner \}$$ is dense below $p_0$.
Claim \ref{claim2.4} then follows. \hfill $\square$ (Claim \ref{claim2.4})
\bigskip
Similarly:
\begin{claim}\label{claim3.5} Let $i<\omega_1$.
$N_{i+1}$ is generated from ${\rm ran}(\sigma_{ii+1}) \cup \{ \omega_1^{N_i} \}$
in the sense that for every $x \in N_{i+1}$ there is some function $f \in {}^{\omega_1^{N_i}}
(N_i) \cap N_i$ such that $x = \sigma_{ii+1}(f)(\omega_1^{N_i})$.
\end{claim}
{\sc Proof} of Claim \ref{claim3.5}. Let $p_0$ be as in the proof of Claim \ref{claim2.4}, let $p \leq p_0$, and let $\Sigma \supset p$ be a syntactic certificate for $p$ with associated semantic certificate
$${\mathfrak C} = \langle \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X'_\delta \colon \delta \in K' \rangle \rangle . $$
Fix $i<\omega_1$, and let $e_i \colon \omega \rightarrow N'_i$ and $e_{i+1} \colon \omega \rightarrow N'_{i+1}$ be as on p.\ \pageref{the_surjections}. Notice that $\omega_1^{N'_i} = \omega_1^{N_i}$.
Let $n<\omega$. There must be some $f \in {}^{\omega_1^{N_i}}N_i' \cap N_i'$ with $e_{i+1}(n) = \sigma_{i i+1}'(f)(\omega_1^{N_i})$. Let $m$, $m' <\omega$ be such that $f=e_i(m)$ and
$\sigma_{i i+1}'(f)=e_{i+1}(m')$. Then
\begin{flalign*}
& p \cup \{ \ulcorner {\dot N}_{i+1} \models {\dot n} = {\dot m}'(\underline{\omega_1^{N_i}}) \urcorner , \ulcorner {\dot \sigma}_{i i+1}({\dot m})={\dot m}' \urcorner \} \leq p.
\end{flalign*}
This argument shows that the set
$$D^7_n = \{ p \in {\mathbb P} \colon
\exists m \, \exists m' \, \{ \ulcorner {\dot N}_{i+1} \models {\dot n} = {\dot m}'(\underline{\omega_1^{N_i}}) \urcorner , \ulcorner {\dot \sigma}_{i i+1}({\dot m})={\dot m}' \urcorner \} \subset p \}$$ is dense below $p_0$.
Claim \ref{claim3.5} then follows.
\hfill $\square$ (Claim \ref{claim3.5})
\begin{claim}\label{claim3.6}
Let $i<\omega_1$. $\{ X \in {\cal P}(\omega_1^{N_i}) \cap N_i \colon \omega_1^{N_i} \in
\sigma_{ii+1}(X) \}$ is generic over $N_i$ for the forcing given by the $I^{N_i}$-positive sets.
\end{claim}
{\sc Proof} of Claim \ref{claim3.6}. Let $p_0$, $p$, $\Sigma$, ${\mathfrak C}$, $e_i$, and $e_{i+1}$ be as in the previous proof. Let $n<\omega$ be such that $e_i(n)$ is a maximal antichain in $N_i'$ for the forcing given by the $I^{N_i'}$-positive sets. Let $m$, $m' <\omega$ be such that $e_i(m) \in e_i(n)$ and $\omega_1^{N_i} = \omega_1^{N_i'} \in \sigma'_{i i+1}(e_i(m)) = e_{i+1}(m')$. Then
\begin{flalign*}
& p \cup \{ \ulcorner {\dot N}_i \models {\dot m} \in {\dot n} \urcorner ,
\ulcorner {\dot \sigma}_{i i+1}({\dot m})={\dot m}' \urcorner , \ulcorner {\dot N}_{i+1} \models \underline{\omega_1^{N_i}} \in\dot m' \urcorner \} \leq p.
\end{flalign*}
This argument shows that the set
$$D^8_n = \{ p \in {\mathbb P} \colon
\exists m \, \exists m' \, \{ \ulcorner {\dot N}_i \models {\dot m} \in {\dot n} \urcorner ,
\ulcorner {\dot \sigma}_{i i+1}({\dot m})={\dot m}' \urcorner , \ulcorner {\dot N}_{i+1} \models \underline{\omega_1^{N_i}} \in {\dot m}' \urcorner \} \subset p \}$$
is dense below $p_0$.
Claim \ref{claim3.6} then follows.
\hfill $\square$ (Claim \ref{claim3.6})
\bigskip
Claims \ref{claim3.5} and \ref{claim3.6} readily imply that if $i<\omega_1$, then $N_{i+1}$ is a generic ultrapower of $N_i$.
By the next claim, direct limits are taken at limit stages:
\begin{claim}\label{claim3.7}
Let $i \leq \omega_1$ be a limit ordinal. For every $x \in N_i$ there is some $j<i$ and some $z \in N_j$ such that $x = \sigma_{ji}(z)$.
\end{claim}
{\sc Proof} of Claim \ref{claim3.7}. This is trivial for $i=\omega_1$. Now let $i<\omega_1$. Let $p$, $\Sigma$, and ${\mathfrak C}$ be as in the previous two proofs. Fix a limit ordinal $i$. For each $n<\omega$ there are $j<i$ and $n' < \omega$ with $\sigma_{ji}(e_j(n'))=e_i(n)$, where $e_j \colon \omega \rightarrow N_j'$ and $e_i \colon \omega \rightarrow N_i'$ are as on p.\ \pageref{the_surjections}. Then
\begin{flalign*}
& p \cup \{ \ulcorner {\dot \sigma}_{ji}({\dot n}')={\dot n} \urcorner \} \leq p.
\end{flalign*}
This argument shows that the set
$$D^9_n = \{ p \in {\mathbb P} \colon \exists n' \, \ulcorner {\dot \sigma}_{ji}({\dot n}')={\dot n} \urcorner \in p \}$$ is dense in ${\mathbb P}$.
Claim \ref{claim3.7} then follows.
\hfill $\square$ (Claim \ref{claim3.7})
\bigskip
$(N_i ,\sigma_{ij} \colon i \leq j \leq \omega_1)$
is then indeed a generic iteration of $N_0$. As $N_0$ is iterable, we may and shall identify $N_i$ with its transitive collapse, so that
(C.4) holds true.
Another round of density arguments will show that
${\mathfrak C}$
satisfies (C.1), (C.3), (C.5), (C.6), and (C.7), where we identify $M_i$
with the structure $(M_i;\in,({\sf NS}_{\omega_1^{M_i}})^{M_i},A \cap \omega_1^{M_i})$.
Let us now verify (C.8) and ($\Sigma$.8), without writing down the relevant dense sets any more.
As for (C.8), its second part, $X_\delta \cap \omega_1 = \delta$ for $\delta \in K$, is easy.
We will now use the Tarski-Vaught test to verify the first part of (C.8).
Let $\varphi$ be any formula, and let $x_1$, $\ldots$, $x_k
\in X_\delta$, $\delta \in K$.
Suppose that
\begin{eqnarray}\label{in-the-model}
(Q_{\lambda_\delta};\in , {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \exists v \, \varphi(v,x_1,\ldots , x_k).
\end{eqnarray}
Let $p \in g$ be such that $\{ \ulcorner {\underline x}_1 \in {\dot X}_\delta \urcorner , \ldots ,
\ulcorner \underline{x}_k \in {\dot X}_\delta \urcorner , \ulcorner \delta \mapsto \lambda_\delta \urcorner
\} \subset p$. Let $q \leq p$, and let
$\Sigma$ be a syntactical certificate for $q$ whose associated semantic certificate is
$${\mathfrak C}' =\langle \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X'_\delta \colon \delta \in K' \rangle\rangle.$$
Then $\delta \in K'$ and $$\{ x_1 , \ldots , x_k \} \subset X'_{\delta}
\prec (Q_{\lambda_\delta};\in , {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}){\rm , }$$
so that by (\ref{in-the-model}) we may choose some $x \in X'_{\delta}$ with
$$(Q_{\lambda_\delta};\in , {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \varphi(x,x_1,\ldots , x_k).$$ Let $r = q \cup \{
\ulcorner \underline{x} \in {\dot X}_{\delta} \urcorner \}$.
By density, there is then some $y \in X_\delta$ such that $$(Q_{\lambda_\delta};\in , {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \varphi(y,x_1,\ldots , x_k).$$
The proof of ($\Sigma$.8) is similar.
Let again $\delta \in K$. Let $E \subset {\mathbb P}_{\lambda_\delta} \cap X_\delta^g$ be dense in ${\mathbb P}_{\lambda_\delta} \cap X_\delta$,
and $r \in E$ iff $r \in {\mathbb P}_{\lambda_\delta} \cap X_\delta$
and
\begin{eqnarray}\label{defn_of_E}
(Q_{\lambda_\delta};\in , {\mathbb P}_{{\lambda_\delta}},
A_{{\lambda_\delta}}) \models \varphi(r,x_1, \ldots , x_k).
\end{eqnarray}
Let $p \in g$ be such that $\{ \ulcorner {\underline x}_1 \in {\dot X}_\delta \urcorner , \ldots ,
\ulcorner {\underline x}_k \in {\dot X}_\delta \urcorner , \ulcorner \delta \mapsto \lambda_\delta \urcorner
\} \subset p$. Let $q \leq p$, and again let
$\Sigma$ be a syntactical certificate for $q$ whose associated semantic certificate is
$${\mathfrak C}' = \langle \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X'_\delta \colon \delta \in K' \rangle \rangle.$$
Then $[\Sigma]^{<\omega} \cap X'_\delta$ has an element, say $r$, such that (\ref{defn_of_E}) holds true. Let $$s = q \cup r \cup \{ \ulcorner \underline{r} \in {\dot X}_\delta \urcorner \}$$
By density, then, $g \cap X_\delta \cap E \not= \emptyset$. \hfill $\square$ (Lemma \ref{generics_give_certificates})
\bigskip
Forcing with any ${\mathbb P}_\lambda$ makes $\omega_2^V$ $\omega$-cofinal, as the iteration map $\pi_{0 \omega_1}$ as being added by the generic filter maps the ordinals of the countable model $M_0$ cofinally into $\omega_2^V$. If $\lambda < \kappa$ (and $\lambda \in C$), then ${\mathbb P}_\lambda$ has size $\aleph_2$, so that by a result of S.\ Shelah, see \cite[Corollary 23.20]{jech}, ${\mathbb P}_\lambda$ will collapse $\omega_1$ to become countable. We are now going to prove that ${\mathbb P} = {\mathbb P}_\kappa$, on the other hand, does not collapse $\omega_1$ and in fact preserves stationary subsets of $\omega_1$.
\subsection{The forcing preserves stationary sets.}\label{preservation-subsection}
\begin{lemma}\label{P_stat_pres}
Let $g$ be ${\mathbb P}$-generic over $V$. Let
$${\mathfrak C} = \langle \langle M_i , \pi_{ij} , N_i , \sigma_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n , \alpha_n ) \colon n<\omega \rangle ,
\langle \lambda_\delta , X_\delta \colon \delta \in K \rangle\rangle$$
be the semantic certificate associated with the syntactic certificate $\bigcup g$.
Let
$$N_{\omega_1} = (N_{\omega_1};\in,A,I^*).$$ Then every element of
$({\cal P}(\omega_1) \cap N_{\omega_1}) \setminus I^*$ is stationary in $V[g]$.
\end{lemma}
\begin{corollary}\label{stat_sets_are_preserved}
${\mathbb P}$ preserves stationary subsets of $\omega_1$.
\end{corollary}
{\sc Proof} of Corollary \ref{stat_sets_are_preserved} from Lemma \ref{P_stat_pres}.
Let ${\mathfrak C}$
be as in the statement of Lemma \ref{P_stat_pres},
and let us write $M_i = (M_i;\in,I_i,a_i)$ and $N_i = (N_i;\in,I_i^*,a_i^*)$ for $i \leq \omega_1$.
In the light of Lemma \ref{generics_give_certificates}, by (C.3) we will have that $I_{\omega_1^{N_0}} = I_0^* \cap M_{\omega_1^{N_0}}$, so that also $I_{\omega_1} = I^*_{\omega_1} \cap M_{\omega_1}$. By
(C.5), the universe of $M_{\omega_1}$ is $(H_{\omega_2})^V$ and
$I_{\omega_1}=({\sf NS}_{\omega_1})^V$, while $I_{\omega_1}^*$ is denoted by $I^*$ in the statement of Lemma \ref{P_stat_pres}. We thus get that $({\sf NS}_{\omega_1})^V = I^* \cap V$, so that the conclusion of Lemma \ref{P_stat_pres} also gives that ${\mathbb P}$ preserves stationary subsets of $\omega_1$. \hfill $\square$(Corollary \ref{stat_sets_are_preserved})
\bigskip
{\sc Proof} of Lemma \ref{P_stat_pres}. Let ${\dot N}_{\omega_1} \in
V^{\mathbb P}$
be a canonical name for $N_{\omega_1}$, and let ${\dot I}^* \in
V^{\mathbb P}$
be a canonical name for $I^*$.
Let ${\bar p} \in g$, ${\dot C}$, ${\dot S} \in V^{\mathbb P}$, and
$i_0<\omega_1$ and $n_0<\omega$ be such that
\begin{enumerate}
\item[(i)] ${\bar p} \Vdash$ ``${\dot C} \subset \omega_1$ is club,''
\item[(ii)] ${\bar p} \Vdash$ ``${\dot S} \in ({\cal P}(\omega_1) \cap {\dot N}_{\omega_1}) \setminus {\dot I}^*$,'' and \label{ithroughiii}
\item[(iii)] ${\bar p} \Vdash$ ``${\dot S}$ is
represented by $[i_0,{\dot n_0}]$ in the term model producing ${\dot N}_{\omega_1}$.''
\end{enumerate}
We may and shall also assume that
\begin{eqnarray}\label{right-thing}
\ulcorner {\dot N}_{i_0} \models {\dot n_0} \mbox{ is a subset of the first uncountable cardinal, yet } {\dot n_0} \notin {\dot I} \urcorner \in {\bar p}{\rm , }
\end{eqnarray}
because the ${\cal L}$-formula in (\ref{right-thing}) must belong to every syntactic certificate for ${\bar p}$, as ${\bar p}$ satisfies (ii) and (iii).
Let $p \leq {\bar p}$ be arbitrary, $p \in {\mathbb P}$. We aim to produce some $q \leq p$ and some $\delta<\omega_1$ such that $q \Vdash {\check \delta} \in {\dot C} \cap {\dot S}$,
see Claim \ref{claim2.10} below.
For $\xi<\omega_1$, let $$D_\xi = \{ q \leq p \colon \exists \eta \geq \xi \, (\eta < \omega_1 \wedge q \Vdash {\check \eta} \in {\dot C}) \}{\rm, }$$ so that $D_\xi$ is open dense below $p$. Let $$E = \{ (q,\eta) \in {\mathbb P} \times \omega_1 \colon
q \Vdash {\check \eta} \in {\dot C} \}.$$ Let us write $$\tau = ((D_\xi \colon \xi<\omega_1),E).$$
We may and shall identify $\tau$ with some subset of $H_\kappa$ which codes $\tau$. Here and in what follows, $\kappa$ is still equal to $\omega_3$.
By ($\Diamond({\mathbb P})$), we may pick some
$\lambda \in C$ such that $p \in {\mathbb P}_\lambda$ and
\begin{eqnarray}\label{choice_of_lambda}
(Q_\lambda;\in,{\mathbb P}_\lambda,A_\lambda) \prec (H_\kappa;\in,{\mathbb P},\tau).
\end{eqnarray}
Let $h$ be ${\rm Col}(\omega,\omega_2)$-generic over $V$, and let
$g' \in V[h]$ be a filter on ${\mathbb P}_\lambda$ such that
$p \in g'$
and $g'$ meets every dense set which is definable over $(Q_\lambda;\in,{\mathbb P}_\lambda,A_\lambda)$ from parameters in $Q_\lambda$.
By Lemma \ref{generics_give_certificates}, $\bigcup g'$ is a syntactic certificate for $p$,
and we may let
$$\langle \langle M_i' , \pi_{ij}' , N_i' , \sigma_{ij}' \colon i \leq j \leq \omega_1 \rangle ,
\langle (k_n' , \alpha_n' ) \colon n<\omega \rangle ,
\langle \lambda'_\delta , X_\delta' \colon \delta \in K' \rangle \rangle$$
be the associated semantic certificate. In particular, $K' \subset \lambda$.
Let $S$ denote the subset of $\omega_1$ which is represented by $[i_0,{\dot n_0}]$ in the term model giving $N_{\omega_1}'$, so that
if $N_{\omega_1}'=(N_{\omega_1}',\in,A,I')$, then by (\ref{right-thing}),
\begin{eqnarray}\label{S_positive}
S \in ({\cal P}(\omega_1) \cap N'_{\omega_1})
\setminus I'.
\end{eqnarray}
Notice that $\omega_1^{V[h]} = \omega_3^V = \kappa$.
Inside $V[h]$, we may extend $\langle N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle$ to a generic iteration $$\langle N'_i , \sigma'_{ij} \colon i \leq j \leq \kappa \rangle$$ such that
\begin{eqnarray}
\omega_1 \in \sigma'_{\omega_1,\omega_1+1}(S).
\end{eqnarray}
This is possible as $\omega_1^{N'_{\omega_1}} = {\rm sup} \{ \omega_1^{N_j} \colon j<\omega_1 \} = \omega_1$ and by (\ref{S_positive}).
Let $$\langle M'_i , \pi'_{ij} \colon i \leq j \leq \kappa \rangle = \sigma_{0,\kappa}(\langle M'_i , \pi'_{ij} \colon i \leq j \leq \omega_1^{N'_0} \rangle){\rm , }$$ so that $\langle M'_i , \pi'_{ij} \colon i \leq j \leq \kappa \rangle$ is an extension of $\langle M'_i , \pi'_{ij} \colon i \leq j \leq \omega_1 \rangle$.
Since $M'_{\omega_1} = ((H_{\omega_2})^V; \in, ({\sf NS}_{\omega_1})^V, A)$, cf.\ (\ref{Momega1}), and $({\sf NS}_{\omega_1})^V$ is assumed to be saturated in $V$,
every maximal antichain in $V$ consisting of stationary subsets of $\omega_1$ is an element of $M'_{\omega_1}$.
By \cite[Lemma 3.8]{hugh}, we may hence lift the generic ultrapower map
$\pi'_{\omega_1 \omega_1+1} \colon M'_{\omega_1} \rightarrow M_{\omega_1+1}'$ to act on all of $V$, and inductively we may
lift the entire generic iteration
$\langle M'_i , \pi'_{ij} \colon \omega_1 \leq i \leq j \leq \kappa \rangle$ to a generic iteration $$\langle M_i^+ , \pi_{ij}^+ \colon \omega_1 \leq i \leq j \leq \kappa \rangle$$ of $V$ with all $M_i^+$, $\omega_1 \leq i \leq \kappa$, being transitive.
Let us write $M=M_\kappa^+$ and $\pi=\pi_{\omega_1,\kappa}^+$.
\bigskip
\begin{tikzpicture}
\def\x{1.7};
\def\y{0.6};
\def\p{'};
\node (pT) at (0*\x, 0*\y){$p[T]$};
\node (ppiT) at (0*\x+1.5, 0*\y) {$p[\pi(T)]$};
\node (N0) at (0*\x, -2*\y) {$N_0\p$};
\node (N1) at (2*\x, -2*\y) {$N_{\omega_1}\p$};
\node (Nrho) at (4*\x, -2*\y) {$N_\kappa\p$};
\node (M0) at (-2*\x, -4*\y) {$M_0\p$};
\node (MN0) at (0*\x, -4*\y) {$M_{\omega_1^{N_0}}\p$};
\node (M1) at (2*\x, -4*\y) {$M_{\omega_1}\p$};
\node (Mrho) at (4*\x, -4*\y) {$M_{\kappa}\p$};
\node (H) at (2*\x, -6*\y) {$((H_{\omega_2})^V;\in, (\mathsf{NS}_{\omega_1})^V, A)$};
\node (V) at (2*\x, -8*\y) {$V$};
\node (Mrhop) at (4*\x, -8*\y) {$M_\kappa^+$};
\node (M) at (4*\x+1, -8*\y) {$M$};
\path (pT)--(ppiT) node[midway]{$\subseteq$};
\path (N0)--(pT) node[midway, rotate=90]{$\in$};
\draw[->] (N0)--(N1) node[midway, above]{$\sigma_{0\omega_1}\p$};
\draw[->] (N1)--(Nrho) node[midway, above]{$\sigma_{\omega_1\kappa}\p$};
\path (MN0)--(N0) node[midway, rotate=90]{$\in$};
\path (M1)--(N1) node[midway, rotate=90]{$\in$};
\path (Mrho)--(Nrho) node[midway, rotate=90]{$\in$};
\draw[->] (M0)--(MN0) node[midway, above]{$\pi_{0 \omega_1^{N_0\p}}\p$};
\draw[->] (MN0)--(M1) node[midway, above]{$\pi_{\omega_1^{N_0\p} \omega_1}\p$};
\draw[->] (M1)--(Mrho) node[midway, above]{$\pi_{\omega_1 \kappa}\p$};
\path (H)--(M1) node[midway, rotate=90]{$=$};
\path (Mrho)--(Mrhop) node[midway, rotate=270]{$\in$};
\path (H)--(V) node[midway, rotate=270]{$\in$};
\draw[->] (V)--(Mrhop) node[midway, above]{$\pi$};
\path (Mrhop)--(M) node[midway]{$=$};
\end{tikzpicture}
\bigskip
The key point is now that $\langle M_i' , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \kappa \rangle$ may be used to extend $\pi \mbox{''} \bigcup g'$ to a
syntactic certificate
\begin{eqnarray}\label{supset}
\Sigma \supset \pi \mbox{''} \bigcup g'
\end{eqnarray}
for $\pi(p)$
in the following manner.
Let
$K^* = K' \cup \{ \omega_1 \}$. For $\delta \in K'$, let $\lambda_\delta^* =
\pi(\lambda_\delta')$ and $X_\delta^* =
\pi \mbox{''} X_\delta'$. Also, write $\lambda_{\omega_1}^* = \pi(\lambda)$ and $X_{\omega_1}^* = \pi \mbox{''} Q_\lambda$. Notice that $\omega_1 \in
\pi(C)$, so that $K^* \subset \pi(C)$.
Let
$${\mathfrak C}^* =
\langle \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \kappa \rangle \mbox{, }
\langle (k'_n , \pi(\alpha'_n) ) \colon n<\omega \rangle \mbox{, }
\langle \lambda_\delta^* , X_\delta^* \colon \delta \in K^* \rangle \rangle.$$
\begin{claim}\label{its-a-sem-c}
${\mathfrak C}^*$ is a semantic certificate
for $\pi(p)$ relative to the parameters $\pi(A)$, $\pi(T)$, $\pi(H_{\omega_2}^V)$, $\pi(\langle A_\mu\,:\,\mu\in C \cap \lambda\rangle)$, and $\pi(\langle \mathbb P_\mu\,:\,\mu\in C\cap\lambda\rangle)$.
\end{claim}
{\sc Proof} of Claim \ref{its-a-sem-c}: First notice that $$\langle (k'_n , \pi(\alpha'_n) ) \colon n<\omega \rangle \in
[\pi(T)]$$ Next, if $\delta \in K'$, then $$X_\delta^* = \pi \mbox{''} X_\delta' \prec
(\pi({Q}_{\lambda'_\delta});\in,\pi({\mathbb P}_{\lambda'_\delta}) , \pi(A_{\lambda'_\delta})),$$
and
$\pi \mbox{''} g' \cap X_\delta^* =
\pi \mbox{''} (g' \cap X_\delta')$; as $\bigcup g'$ is a syntactic certificate for $p$, we thus have that
$\pi \mbox{''} g' \cap X_\delta^* \cap E \not= \emptyset$ for every $E \subset
\pi({\mathbb P}_{\lambda_\delta'})$ which is dense in $\pi({\mathbb P}_{\lambda_\delta'})$ and definable over the structure $(\pi({Q}_{\lambda'_\delta}); \in , \pi({\mathbb P}_{\lambda_\delta'}) , \pi(A_{\lambda_\delta'}))$ from parameters in $X_\delta^*$. Finally, $X_{\omega_1}^*= \pi \mbox{''} Q_\lambda$ and the choice of $g'$ imply that $\pi \mbox{''} g' \cap X_{\omega_1}^* \cap E \not= \emptyset$ for every $E \subset
\pi({\mathbb P}_{\lambda})$ which is dense in $\pi({\mathbb P}_{\lambda})$ and definable over the structure $$(\pi(Q_\lambda); \in , \pi({\mathbb P}_{\lambda}) , \pi(A_{\lambda}))$$ from parameters in $X_{\omega_1}^*$.
This buys us that ${\mathfrak C}^*$ is indeed a semantic certificate
for $\pi(p)$ as an element of $\pi({\mathbb P})$, and that therefore there is some syntactic certificate $\Sigma$ as in
(\ref{supset}), relative to $\pi(A)$, $\pi(T)$, $\pi(H_{\omega_2}^V)$, $\pi(\langle A_\mu\,:\,\mu\in C \cap \lambda\rangle)$, and $\pi(\langle \mathbb P_\mu\,:\,\mu\in C\cap \lambda\rangle)$, such that ${\mathfrak C}^*$ is certified by $\Sigma$.
\hfill $\square$ (Claim \ref{its-a-sem-c})
\bigskip
Now let $[{\dot m_0}]_{\omega_1+1}$ represent $\sigma'_{\omega_1 \omega_1+1}(S)$ in the term model
for $N_{\omega_1+1}'$ provided by $\Sigma$, so that\footnote{Here, ${\dot \sigma}_{i_0 \omega_1+1}$ and ${\dot N}_{\omega_1+1}$ are terms of the language associated with $\pi({\mathbb P}_\lambda)$, and $\ulcorner {\dot \sigma}_{i_0 \omega_1+1}({\dot n_0}) = {\dot m_0} \urcorner$ and $
\ulcorner {\dot N}_{\omega_1+1} \models \underline{\omega_1} \in {\dot m_0} \urcorner$ are formulae of that language.}
$$\{ \ulcorner {\dot \sigma}_{i_0 \omega_1+1}({\dot n_0}) = {\dot m_0} \urcorner ,
\ulcorner {\dot N}_{\omega_1+1} \models \underline{\omega_1} \in {\dot m_0} \urcorner \}
\subset \Sigma{\rm , }$$
in other words,
\begin{eqnarray}\label{half_of_new_p}
\pi(p) \cup \{ \ulcorner {\dot \sigma}_{i_0 \omega_1+1}({\dot n_0}) = {\dot m_0} \urcorner ,
\ulcorner {\dot N}_{\omega_1+1} \models \underline{\omega_1} \in {\dot m_0} \urcorner \}
\mbox{ is certified by } \Sigma .
\end{eqnarray}
Let us now define
\begin{eqnarray}\label{defn_q}
q^* = \pi(p) \cup \{
\ulcorner {\dot \sigma}_{i_0 \omega_1+1}({\dot n_0}) = {\dot m_0} \urcorner ,
\ulcorner {\dot N}_{\omega_1+1} \models \underline{\omega_1} \in {\dot m_0} \urcorner ,
\ulcorner \underline{\omega_1} \mapsto \underline{\pi(\lambda)} \urcorner \}.
\end{eqnarray}
In the light of Lemma \ref{abs_of_ex_of_certificate}, we thus established the following.
\begin{claim}\label{claim_1}
$q^* \in \pi({\mathbb P})$, as being certified by $\Sigma$.
\end{claim}
The elementarity of
$\pi \colon V \rightarrow M$ then gives some $\delta<\omega_1$
such that
\begin{eqnarray}
q = p \cup \{ \ulcorner {\dot \sigma}_{i_0 \delta+1}({\dot n_0}) = {\dot m_0} \urcorner ,
\ulcorner {\dot N}_{\delta+1} \models \underline{\delta} \in {\dot m_0} \urcorner , \ulcorner \underline{\delta} \mapsto \underline{\lambda} \urcorner \} \in {\mathbb P}.
\end{eqnarray}
\begin{claim}\label{claim_2}\label{claim2.10}
$q \Vdash {\check \delta} \in {\dot C} \cap {\dot S}$.
\end{claim}
{\sc Proof} of Claim \ref{claim_2}. $q \Vdash {\check \delta} \in {\dot S}$ readily follows from $$\{ \ulcorner {\dot \sigma}_{i_0 \delta+1}({\dot n_0}) = {\dot m_0} \urcorner ,
\ulcorner {\dot N}_{\delta+1} \models \underline{\delta} \in {\dot m_0} \urcorner \} \subset q,$$ the fact that ${\bar p} \geq p$ forces that ${\dot S}$ is
represented by $[i_0,{\dot n_0}]$ in the term model giving ${\dot N}_{\omega_1}$, and the fact that
by Claim \ref{claim2.4}, $[\underline{\delta}]_{\delta+1}$ represents $\delta$ in the model $N_{\delta+1}$
of any semantic certificate for $q$ as being given by a generic which contains $q$.
Let us now show that $q \Vdash {\check \delta} \in {\dot C}$.
We will in fact show that $q$ forces that ${\check \delta}$ is a limit point of ${\dot C}$.
Otherwise
there is some $r \leq q$ and some $\eta < \delta$ such that
\begin{eqnarray}\label{21}
r \Vdash {\dot C} \cap {\check \delta} \subset {\check \eta}.
\end{eqnarray}
Suppose that $r$ is certified by $\Sigma$, so that there is some
\begin{eqnarray}\label{prime}
\langle \langle M'_i , \pi'_{ij} , N'_i , \sigma'_{ij} \colon i \leq j \leq \omega_1 \rangle ,
\langle (k'_n , \alpha'_n ) \colon n<\omega \rangle ,
\langle \lambda'_{\bar \delta} , X'_{\bar \delta} \colon {\bar \delta} \in K' \rangle \rangle
\end{eqnarray}
which is certified by $\Sigma$ and
$r \in [\Sigma]^{<\omega}$. We must have that
\begin{enumerate}
\item[(a)] $\delta \in K'$,
\item[(b)] $X'_\delta \prec (Q_\lambda; \in , {\mathbb P}_\lambda , A_\lambda)$,
\item[(c)] $X'_\delta \cap \omega_1 = \delta$,
and
\item[(d)]
$[\Sigma]^{<\omega} \cap X'_\delta \cap
E \not= \emptyset$ for every $E \subset {\mathbb P}_{\lambda}$ which is dense in ${\mathbb P}_{\lambda} \cap
X'_\delta$ and definable over the structure
$$(Q_\lambda;\in, {\mathbb P}_\lambda , A_\lambda)$$
from parameters in $X'_\delta$.
\end{enumerate}
Here, (a) is given by $\ulcorner \underline{\delta} \mapsto \underline{\lambda} \urcorner \in r$,
(b) and (c) are given by (C.8), while (d) is exactly what ($\Sigma$.8) on p.\ \pageref{(C.9)} buys us.
We have that $A_\lambda = \tau \cap Q_\lambda$, and hence $A_\lambda$ may be identified
with the ordered pair $((D_\xi \cap Q_\lambda \colon \xi < \omega_1), E \cap Q_\lambda )$.
As $\eta < \delta \subset X_\delta'$, $D_\eta$ is definable over
the structure $$(Q_\lambda;\in, {\mathbb P}_\lambda , A_\lambda)$$
from a parameter in $X'_\delta$. By
(\ref{choice_of_lambda}), $D_\eta \cap Q_\lambda$ is dense in ${\mathbb P}_\lambda$.
By (d) above, there is then some
$s \in [\Sigma]^{<\omega} \cap X'_\delta \cap D_\eta$.
By (\ref{choice_of_lambda}) again, the unique smallest $\eta' \geq \eta$ with $s \Vdash {\check \eta}' \in {\dot C}$ must be in $X'_\delta$, hence $\eta' < \delta$ by (c) above.
But now $s$ is compatible with $r$, as they are both finite subsets of the very same $\Sigma$ which certifies them (cf.\
Lemma \ref{+++}).
We have reached a contradiction with (\ref{21}).
\hfill $\square$ (Claim \ref{claim_2})
\bigskip
Now ${\dot C}$, ${\dot S}$, and ${\bar p} \in g$ were such that (i) through (iii) on p.\ \pageref{ithroughiii} hold true. We showed that the set of all $q \leq {\bar p}$ with $q \Vdash {\dot C} \cap {\dot S} \not= \emptyset$ is dense. As ${\dot C}$ was arbitrary, this buys us that ${\dot S}^g$ will be stationary in $V[g]$. But then as ${\dot S}$ was arbitrary, this means that every element of $({\mathcal P}(\omega_1) \cap N_{\omega_1}) \setminus I^*$ will be stationary in $V[g]$. \hfill $\square$ (Lemma \ref{P_stat_pres})
\section{Open questions.}
Woodin \cite{hugh} also introduced the axiom $(*)^+$ as a strengthening of $(*)$. $(*)^+$ says that there is some pointclass $\Gamma \subset {\mathcal P}({\mathbb R})$ and some filter $g \subset {\mathbb P}_{\max}$ such that
\begin{enumerate}
\item[(1)] $L(\Gamma,{\mathbb R}) \models {\sf AD}^+$,\footnote{${\sf AD}^+$ is a natural strengthening of ${\sf AD}$ which was introduced by H.\ Woodin, see e.g.\ \cite[Definition 9.6]{hugh}.}
\item[(2)] $g$ is ${\mathbb P}_{\rm max}$-generic over $L(\Gamma,{\mathbb R})$,
and
\item[(3)] ${\mathcal P}({\mathbb R}) \subset L(\Gamma,{\mathbb R})[g]$.
\end{enumerate}
See \cite[p.\ 908]{hugh}. While the main
result of the current paper gives a new twist to the question if ${\sf MM}$ is compatible with $(*)^+$, see \cite[p.\ 923, Question (15) a)]{hugh}, it also leaves this question wide open. See \cite{hugh-preprint}.
There is a strengthening of ${\sf MM}^{++}$, isolated by Viale \cite{viale}, which has strong completeness properties modulo forcing similar to those of $(*)$. This is the axiom ${\sf MM}^{+++}$. It says that a class $\mathbb T$ of towers of ideals with certain nice structural properties is dense in the category of stationary set preserving forcings; in other words, for every stationary set preserving forcing ${\mathbb P}$ there is a tower $\mathcal T$ in $\mathbb T$ such that ${\mathbb P}$ completely embeds into $\mathcal T$ in such a way that the quotient forcing preserves stationary sets in $V^{\mathbb P}$. ${\sf MM}^{+++}$ implies ${\sf MM}^{++}$, if $\kappa$ is an almost super-huge cardinal, then there is a partial order ${\mathbb P} \subset V_\kappa$ which forces ${\sf MM}^{+++}$, and if there is a proper class of almost super-huge cardinals, then ${\sf MM}^{+++}$ is complete for the theory of the $\omega_1$-Chang model\footnote{The $\omega_1$-Chang model is the $\subseteq$-minimal transitive model of ${\sf ZF}$ containing all ordinals and closed under $\omega_1$-sequences. It can be construed as $\bigcup_{\alpha\in \Ord}L([\alpha]^{\aleph_1})$ and it includes $L(\mathcal P(\omega_1))$ as a definable submodel.} with respect to stationary set preserving partial orders forcing ${\sf MM}^{+++}$.
Schindler \cite[Definition 2.10]{both} introduces ${\sf MM}^{*,++}$ as a strengthening of ${\sf MM}^{++}$ by relaxing ``forceable by a stationary set preserving forcing'' to ``honestly consistent'' in an appropriate formulation of ${\sf MM}^{++}$, see
\cite{both}.
It remains open if either of ${\sf MM}^{+++}$ or ${\sf MM}^{*,++}$ is really stronger than ${\sf MM}^{++}$.
While Viale's ${\sf MM}^{+++}$ is known to be consistent modulo
a super-huge cardinal, it is open at this point if ${\sf MM}^{*,++}$ is consistent at all relative to large cardinals.
|
\section{Introduction}
Hadron collisions at particle colliders are modelled using quantum chromodynamics (QCD) Monte Carlo (MC) simulation codes. These codes consist of several parts.
The hard scattering is modelled by simulating particles from the hadronization of partons kinematics of which are calculated using perturbation theory.
Partons from initial-state radiation (ISR) and final-state radiation (FSR) are modelled by a parton shower (PS) algorithm.
The underlying event (UE), a non-perturbative component, refers to the activity resulting from additional interactions of partons in the hadron collision.
It consists of beam-beam remnants (BBR), and the particles from the additional relatively softer parton-parton scatterings referred to as the multiple-parton interactions (MPI).
To obtain a reliable description of the UE, adjustable parameters in standard MC event generators, such as \PYTHIA8~\cite{pythia8}, \HERWIG\cite{Bellm:2015jjp}, and \SHERPA\cite{Gleisberg:2008ta}, need to be tuned to the data.
Another non-perturbative ingredient is the parton distribution function (PDF) used both in the hard partonic matrix element (ME) calculation, the PS and the MPI model. $\alpha_S$($M_Z$) is a free parameter and its value used in simulations at LO or NLO are typically different. Different strategies are adopted in different collaborations; CMS~\cite{Khachatryan:2015pea} and ATLAS~\cite{ATL-PHYS-PUB-2014-021} tunes are traditionally based on LO PDFs, \PYTHIA8~\cite{Skands:2014pea} tunes are mostly based on LO PDFs, new tunes are based on NNLO PDFs, and \HERWIG7~\cite{Gieseke:2016fpz} on NLO PDFs.
Merging schemes, such as the $MLM$~\cite{mlm}, allow the combination of predictions of jet production using ME calculations with those from PS emissions for soft and collinear parton radiation at leading-log (LL) accuracy without double counting or dead regions. Merged calculations capture some higher-order corrections with respect to the formal order of the ME calculation.
Using the same PDF set and $\alpha_S$~value in the PS-ME merged calculations in the simulation of the various components is advocated in Ref.~\cite{Cooper:2011gk}, and by the \HERWIG7 and \SHERPA Collaborations. In the \PYTHIA8 tunes produced prior to Ref. \cite{Sirunyan:2019dfx}, $\alpha_S$($M_Z$) values in general are not selected to be the same as those used in the PDFs. For example, in the Monash tune, the $\alpha_S^{FSR}$($M_Z$), set to 0.1365, is obtained by fitting standalone-\PYTHIA8 predictions to LEP event-shape measurements, $\alpha_S^{ISR}$($M_Z$) is assumed to be equal to $\alpha_S^{FSR}$($M_Z$), and the $\alpha_S$$(M_Z)$ values in hard scattering and MPI are set to 0.130~\cite{Skands:2014pea}.
In this note, we discuss the preferred values of $\alpha_S^{ISR}$($M_Z$) and $\alpha_S^{FSR}$($M_Z$) obtained from jet multiplicity, substructure and UE event measurements in $\mathrm{t}\overline{\mathrm{t}}$~ events and discuss new CMS \cite{CMS} tunes based on consistent PDF and $\alpha_S$($M_Z$) values.
\section{Revisiting Parton Shower Parameters and Tunes}
The previous default CMS \PYTHIA8 tune, CUETP8M1~\cite{Khachatryan:2015pea}, does not describe the central values of $\sqrt{s}=$ 13 TeV UE data very well~\cite{Sirunyan:2019dfx} or the UE evolution from 1.96 to 7 TeV in UE in Z+jets~\cite{Sirunyan:2017vio}. It is also observed that the predictions of the POWHEG generator \cite{powhegv23} interfaced with the \PYTHIA8 code with the CUETP8M1 tune overshoot the data for large multiplicities where jets are produced by the PS ~\cite{CMS:2016kle} for both $\sqrt{s}=$ 8 and 13 TeV data.
\subsection{Jet Multiplicity in $\mathrm{t}\overline{\mathrm{t}}$~ events and $\alpha_S^{ISR}$}
Tuning $\alpha_S^{ISR}$~using 8 TeV jet multiplicity in $\mathrm{t}\overline{\mathrm{t}}$~ events using the PS dominated region yields $\alpha_S^{ISR}$ = 0.1108$^{+0.0145}_{-0.0142}$~\cite{CMS:2016kle}. The tuned $\alpha_S^{ISR}$($M_Z$) value agrees with the PDG value of $\alpha_S^{ISR}$($M_Z$) = $0.1181\pm0.0011$ \cite{PhysRevD.98.030001} well within uncertainties. It is found that to describe the number of jet distribution in $\mathrm{t}\overline{\mathrm{t}}$~ events, the \PYTHIA8 ISR parameter {\it Rapidity Ordering} needs to be used ~\cite{Sirunyan:2019dfx}.
Rapidity ordering imposes an additional constraint on the \ensuremath{p_{\mathrm{T}}}~-ordered emissions reducing the parton emission phase space.
The parton emission probability is mainly constrained by the jet activity and the interplay between the hard and soft parts of the parton emissions.~However, it does not strongly constrain the UE. Therefore, $\alpha_S^{ISR}$($M_Z$) constrained by $\mathrm{t}\overline{\mathrm{t}}$~ jet kinematics is used as a fixed input parameter in deriving a new UE tune, \textsc{cuetp8m2t4}~\cite{CMS:2016kle}. This new tune, with a lower $\alpha_S$$^{ISR}(M_Z)$, with respect to the CUETP8M1 tunes is found to improve the description of $\mathrm{t}\overline{\mathrm{t}}$~ kinematics, as as the overall description of observables at $\sqrt{s}=8$ and 13 TeV.
\subsection{Jet substructure in $\mathrm{t}\overline{\mathrm{t}}$~ events and $\alpha_S^{FSR}$($M_Z$)}
Jet substructure observables are measured using $\mathrm{t}\overline{\mathrm{t}}$~ events in the lepton+jets channel at $\sqrt{s}=13$ TeV \cite{PhysRevD.98.092014} using the \textsc{cuetp8m2t4}~tune for the $\mathrm{t}\overline{\mathrm{t}}$~ signal and \textsc{cuetp8m1} tune for other simulated samples. Jets are studied with different substructure observables using charged and neutral particles and charged-particle-only jet constituents. Inclusive jet samples as well as samples enriched in bottom, light-quark, or gluon jets are selected.
Substructure observables at particle level are compared to NLO predictions from \POWHEG+\PYTHIA8, \HERWIG7, \SHERPA and \DIRE2 \cite{Hoche:2015sya}.
Predictions with the default tunes do not yield a good overall description of the data, in particular, for particle multiplicity-related observables.
These measurements help tune models. As a first step in this direction, the angle between the groomed subjets is used to extract $\alpha_S^{FSR}$=0.115$^{+0.015}_{-0.013}$ at LO+LL accuracy and using CMW rescaling \cite{Catani:1990rr}. This value agrees with the PDG value of $\alpha_S^{ISR}$($M_Z$) well within uncertainties. A more precise $\alpha_S^{FSR}$($M_Z$) could be obtained once calculations of top-quark decays with multiple emissions at higher order with PS at approximate NLL become available. These measurements also provide the data to improve analytical QCD calculations with higher-order corrections, for infrared-and/or collinear-safe observables.
\subsection{Underlying event in $\mathrm{t}\overline{\mathrm{t}}$~ events and $\alpha_S^{FSR}$($M_Z$)}
UE activity in $\mathrm{t}\overline{\mathrm{t}}$~ dilepton events are measured by CMS at $\sqrt{s}=13$ TeV.
The measurements are based in particle-flow reconstruction.
The UE contribution is isolated by removing charged particles associated with the decay products of the $\mathrm{t}\overline{\mathrm{t}}$~ event candidates as well as with pileup interactions from the set of reconstructed charged particles for each event.
The observables and categories chosen for the measurements enhance the sensitivity to $\mathrm{t}\overline{\mathrm{t}}$~ modeling, MPI, color reconnection (CR) and the choice of $\alpha_S$($M_Z$) in the \PYTHIA8.
Most of the comparisons indicate a fair agreement between the data and the \POWHEG + \PYTHIA8 setup with the \textsc{cuetp8m2t4} tune~\cite{CMS:2016kle},
but disfavor the setups in which MPI and CR are switched off.
The data also disfavor the default configurations in \POWHEG+\HERWIG++, \POWHEG+\HERWIG7, and \SHERPA.
It has been furthermore verified that, as expected, the choice of the next-to-leading-order matrix-element generator does not impact significantly the expected characteristics of the UE by comparing predictions from \POWHEG and \MG~\cite{Alwall:2014hca}, both interfaced with \PYTHIA8.
The UE measurements in $\mathrm{t}\overline{\mathrm{t}}$~ events test the hypothesis of universality of UE at an energy scale two times the top quark mass which is considerably higher than the ones at which UE models have been studied in detail.
The results also show that a value of $\alpha_S^{FSR}$$(M_{Z})=0.120\pm0.006$
is consistent with the data and the corresponding uncertainties
translate to a variation of the renormalization scale by a factor of $\sqrt{2}$ that has already been adopted in CMS $\mathrm{t}\overline{\mathrm{t}}$~ measurements.
\subsection{New CMS UE tunes}
A set of new 13 TeV \PYTHIA8.226 UE tunes is obtained with different choices of values of $\alpha_S$$(M_Z)$ used in the modeling of the ISR, FSR, hard scattering, and MPI, as well as the order of its evolution as a function of the four-momentum squared $Q^2$. We distinguish the new tunes according to the order of the NNPDF3.1 PDF set~\cite{Ball:2017nwa} used: LO-PDF, NLO-PDF, or NNLO-PDF. The tunes are labeled as CPi, where i=1,2,...,5. CP1 and CP2 are based on LO, CP3 on NLO, and CP4 and CP5 on NNLO NNPDF3.1 set.
We fit charged-particle and \ensuremath{p_{\mathrm{T}}}~$^{sum}$~densities, measured in transMIN\ and transMAX\ regions as a function of \ensuremath{p_{\mathrm{T}}}~$^{max}$, as well as the charged-particle multiplicity as a function of pseudorapidity $\eta$, measured by CMS at $\sqrt{s}=13$ TeV~\cite{CMS:2015zev,Khachatryan:2015jna}. In addition, we also use the charged-particle and \ensuremath{p_{\mathrm{T}}}~$^{sum}$ densities as a function of the leading charged-particle \ensuremath{p_{\mathrm{T}}}~, measured in transMIN\ and transMAX\ by CMS at $\sqrt{s}=7$ TeV~\cite{CMS:2012zxa} and by CDF at $\sqrt{s}=1.96$ TeV~\cite{Aaltonen:2015aoa}.
The parameters related to the simulation of the hadronization and beam remnants are not varied in the fits and are kept fixed to the values of the Monash tune.
Only five parameters related to the simulation of MPI, to the overlap matter distribution function~\cite{Sjostrand:1987su}, and to the amount of CR are constrained for the new CMS tunes. In all tunes, we use the MPI-based CR model~\cite{Sjostrand:2004pf}.
The overlap distribution between the two colliding protons is modeled according to a double-Gaussian functional form with the parameters \texttt{coreRadius} and \texttt{coreFraction}.
In the LO NNPDF3.1 set, $\alpha_S$($M_Z$) = 0.130, whereas for the NLO and NNLO NNPDF3.1 sets, $\alpha_S$($M_Z$) = 0.118.
Irrespective of the specific PDF used, predictions from the new tunes reproduce well the UE measurements at center-of-mass energies $\sqrt{s}=1.96$ and 7 TeV. A significant improvement in the description of UE measurements at 13 TeV is observed with respect to predictions from the previous tunes that were extracted using data at lower collision energies.
For the first time, predictions based on higher-order PDF sets are shown to give a reliable description of minimum-bias (MB) and UE measurements, with a similar level of agreement as predictions from tunes using LO PDF sets. Predictions of the new tunes agree well with the data for MB observables measured in the central ($|\eta|< 2.4$) and forward ($3.2 <|\eta|< 4.7$) regions.
The CP tunes simultaneously describe the number of charged particles produced in diffractive processes and MB collisions.
Neither the CP tunes nor the CUETP8M1 tune describe the very forward region ($-6.6<\eta<-5.2$) well. Measurements sensitive to double-parton scattering contributions are reproduced better by predictions using the LO PDF set in the UE simulation, without ISR rapidity ordering.
The UE simulation provided by the new tunes can be interfaced to higher-order and multileg matrix element generators, such as \POWHEG and \MG, without degrading the good description of UE observables. Such predictions also reproduce well observables measured in multijet final states, Drell--Yan, and top quark production processes. The central values of the normalized $\mathrm{t}\overline{\mathrm{t}}$~ cross section in bins of the number of additional jets predicted by \POWHEG~+\PYTHIA8 overestimate the data when a high value of $\alpha_S^{ISR}$($M_Z$)$\sim0.13$ is used (CMS \PYTHIA8 CP1 and CP2 tunes). Even when $\alpha_S^{ISR}$($M_Z$)=0.118 is used, the CP4 tune overestimates the data at high jet multiplicities. This is cured by the ISR rapidity ordering in the CP5 tune (as in the CUETP8M2T4 tune). Measurements of azimuthal dijet correlations are also better described when a value of $\alpha_S^{ISR}$($M_Z$)=0.118 is used in predictions obtained with \POWHEG merged with \PYTHIA8. All details about the CP tunes can be found in \cite{Sirunyan:2019dfx}.
\section{Conclusions}
Monte Carlo event modeling is studied using jet multiplicity, jet substructure, and underlying event measurements in top quark pair events.
These studies showed that the data is consistent with a lower value of strong coupling parameter for both initial and final state radiation in the parton shower when merged configurations are used.
New CMS \PYTHIA8 multipurpose tunes, aiming for a consistent description of UE and MB observables at several collision energies and a reliable prediction of the UE simulation in various processes when merged with higher-order ME calculations are presented.
These new tunes exploit Monte Carlo configurations with consistent parton distribution functions and strong coupling parameter values in the matrix element and the parton shower, at leading order (LO), next-to-leading order (NLO) and next-to-next-to-leading order (NNLO).
Predictions from \PYTHIA8 obtained with tunes based on NLO or NNLO PDFs are shown to reliably describe minimum-bias and underlying-event data with a similar level of agreement to predictions from tunes using LO PDF sets along with a wide range of different measurements.
|
\section{Introduction}
In optimization, an important issue is how to approximate the feasible region using derivatives of the function and the tangent cone of the set involved. Such needs arise in optimality conditions, constraint qualifications and stability analysis when the problem data are perturbed.
In the same way that second-order derivatives provide quadratic approximations whereas first-order derivatives only provide linear approximation to a given function, second-order tangent sets provide better approximation than tangent cones to a set at a point, in particular when the given set is not a polyhedral set or the union of finitely many polyhedral sets. As a result, the second-order tangent sets have been used successfully in second-order optimality conditions, stability analysis,
and metric subregularity (see e.g. \cite{BCS99,BR05,BS,Cons06,Con16,GM11,GJN10,JN04} {and references therein}). More recently, Gfrerer and Mordukhovich \cite{GM16} use the second-order tangent set to give an estimate of the upper curvature of a set, which is used to study the Robinson regularity of parametric constraint systems.
In optimization, one often has to deal with a feasible region in the form
$C:=\{x|\, F(x)\in \Theta\},$
where $F:\Re^n\rightarrow \Re^m$ is a second-order continuously differentiable mapping and $\Theta$ is a closed set in $\Re^m$.
By \cite[Proposition 13.13]{RW98}, under a constraint qualification, the second-order tangent set of the feasible region
$C$ can be characterized as
\begin{eqnarray}
\left .
\begin{array}{r}
d\in T_C( x) \\
w\in T^2_C( x,d)
\end{array}
\right \} \quad \Longleftrightarrow \quad \left \{
\begin{array}{l}
{\nabla} F(x) d\in T_{\Theta}(F( x)) \\
{\nabla} F( x) w+d^T {\nabla}^2 F( x)d\in T^2_{\Theta}(F( x);\nabla F( x)d),
\end{array}
\right. \label{chainruletangentset}
\end{eqnarray}
where $T_C, T_C^2$ denote the tangent cone and the second-order tangent set, respectively (see Definition \ref{sotangent}).
In the case when $\Theta=\Re_-^{m_1}\times \{0\}^{m_2}$, $m_1+m_2=m$, the system is described by inequality and equality constraints.
In this case, since the set $\Theta$ is polyhedral, the second-order tangent set of $\Re_-^{m_1}\times \{0\}^{m_2}$ is a polyhedral set, and hence the second-order tangent set of the feasible region is a system of equalities and inequalities involving the second-order derivatives of the constraint mapping $F$ (see, e.g., Bonnans and Shapiro \cite[Formula (3.81)]{BS}), provided a constraint qualification holds.
In recent years, the second-order cone programming (SOCP) has attracted much attention due to a broad range of applications in fields from engineering, control and finance to robust optimization and combinatorial optimization (see e.g., \cite{AG} for introduction to the theories and its applications).
Consider the second-order cone defined as
\[
{\cal K} := \{ (x_1,x_2) \in \Re \times \Re^{m-1} \, | \, \| x_2 \| \leq x_1 \},
\]
where $\| \cdot \|$ denotes the Euclidean norm. Bonnans and Ram\'{i}rez gave the characterization for the second-order tangent set \cite[Lemma 27]{BR05}, and using it to formulate second-order necessary and sufficient optimality conditions for nonlinear SOCPs. Since the second-order cone is not polyhedral, the second-order tangent set is not polyhedral \cite{BR05}.
In recent years, there are more and more researches on the second-order cone (SOC) complementarity system defined as
\begin{eqnarray*}
&&{\cal K} \ni G(z) \perp H(z) \in {\cal K},
\end{eqnarray*}
where $u\perp v$ means the vectors $u$ and $v$ are perpendicular, $G(z), H(z): \Re^n\rightarrow \Re^m$. One of the sources of the SOC complementarity system is the Karush-Kuhn-Tucker (KKT) optimality condition for the second-order cone programming (see e.g. \cite{AG,CT}), and the other is the equilibrium system for a Nash game where the constraints involving second-order cones (see e.g. \cite{HYF}).
We call the closed cone
$$ \Omega
:= \{(x,y) \in \Re^{2m} \, | \, {\cal K} \ni x \perp y \in {\cal K}\},$$
the SOC complementarity set (or the complementarity set associated with the second-order cone,
c.f. \cite{JLZ15}). Using the SOC complementarity set, the SOC complementarity system can be reformulated as
$ (G(z),H(z)) \in \Omega.$ Due to the existence of the complementarity condition, the SOC complementarity set is a nonconvex set. Moreover, due to the nonpolyhedral structure of the second-order cone ${\cal K}$, the SOC complementarity set is also nonpolyhedral.
Hence the SOC complementarity set is a difficult object to study in the variational analysis.
The main goal of this paper is to provide a precise formula for the second-order tangent set to the SOC complementarity set $\Omega$. The projection operator over the second-order cone $\Pi_{\cal K}(x):=\arg\min_{ x'\in {\cal K}}\|x'-x\|$ is one of our main tools in the subsequent analysis.
It is well-known that the metric projection operator $\Pi_{\cal K}(x)$ provides an alternative characterization of the SOC complementarity set:
\begin{eqnarray} \label{Omeganew}
(x,y)\in \Omega
\Longleftrightarrow \Pi_{{\cal K}}(x-y)=x .
\end{eqnarray}
The projection operator $\Pi_{\cal K}(x)$ is known to be first-order directionally differentiable (see e.g. \cite[Lemma 2]{OS08}) and the
connection between its tangent cone and its directional derivative has been {given} (see \cite{JLZ15,YZ16}): for any $(x,y)\in \Omega$,
\begin{equation}
\label{directionalderivative}
(d, w) \in T_\Omega(x,y) \Longleftrightarrow \Pi'_{\cal K}(x-y;d-w)=d.
\end{equation}
Using this connection, it has been shown that the SOC complementarity set $\Omega$ is geometrically derivable and
the exact formula for its tangent cone is given; see, e.g., \cite[Theorem 5.1]{YZ16}. Moreover, the coderivative of the projection operator $\Pi_{\cal K}$ allows us to characterize the various normal cones as in \cite[Proposition 2.1]{yzhou} and show that the SOC complementarity set is not only geometrically derivable but also directionally regular \cite[Theorem 6.1]{YZ16}.
So far by using the first-order variational analysis, it has been revealed that although the SOC complementarity set is neither a convex set nor the union of finitely many polyhedral convex sets, it enjoys certain nice properties that a convex set or the union of finitely many polyhedral convex sets has. In this paper, we continue to investigate the second-order variational properties of the SOC complementarity cone.
Our main contributions are as follows:
\begin{enumerate}
\item[$\bullet$] We derive the exact formula for the second-order directional derivative of the projection operator over second-order cone. We {further} establish the connection between the second-order tangent set and the second-order directional derivative of the projection operator: for any $(x,y)\in \Omega$ and $(d,w) \in T_\Omega(x,y)$,
\begin{equation}
\label{sdirectionalderivative}(p,q) \in T^2_\Omega ((x,y);(d,w)) \Longleftrightarrow \Pi_{\cal K}''(x-y;d-w,p-q)=p.
\end{equation}
\item[$\bullet$] We show that the SOC complementarity set is second-order directionally differentiable (see Definition \ref{BS}). Note that this
nice property is not even enjoyed by a convex set (see \cite[Example 3.31]{BS}).
\item[$\bullet$]
Using the characterization (\ref{sdirectionalderivative}) and the precise formula for the second-order directional derivative of the projection operation over the second-order cone, we derive the exact formula for the second-order tangent set of the SOC complementarity set. Compared with the usual vector complementarity set, our research shows that the task of establishing the formula of second-order tangent set to the second-order cone complementarity set, { which has nonpolyhedral and nonconvex structure}, is not trivial.
\item[$\bullet$]{ Based on the exact formula of the second-order tangent set of $\Omega$, we develop the second-order optimality conditions for the mathematical program with second-order cone complementarity constraints (SOCMPCC).}
\end{enumerate}
We organize our paper as follows. Section 2 contains the preliminaries. In Section 3, we calculate the second-order directional derivative of the projection operator over the second-order cone. Section 4 is devoted to the exact formula of the second-order tangent set to the SOC complementarity set. {The second-order optimality conditions of SOCMPCC are discussed in Section 5}.
\section{Preliminaries}
In this section, we clarify the notation and recall some background materials. First,
we denote by $\Re_+$ and $\Re_{++}$ the set of nonnegative scalars and
positive scalars respectively, i.e., $\Re_+:=\{\alpha|\ \alpha\geq 0\}$ and
$\Re_{++}:=\{\alpha|\ \alpha>0\}$. For a set ${ C}$, denote by int$C$, cl$C$, bd$C$, co$C$, $C^c$ its interior, closure, boundary, convex hull, and its complement, respectively. For a closed set $C\subseteq \Re^n$, { let
$C^\circ$ and $\sigma(\cdot|C)$
stand for the polar cone and the support function of $C$, respectively, i.e.,
$C^\circ=\{v|\, \langle v,w \rangle\leq 0,\ \forall w\in C \}$
and $\sigma(z|C)=\sup\{\langle z,x \rangle|\, x\in C\}$ for $z\in \Re^n$.} Denote by ${\rm lin}C$ the largest subspace $L$ such that $C+L\subseteq C$.
For a vector $x=(x_1,x_2)\in \Re\times \Re^{n-1}$, we denote $x^\circ$ the polar set of the set $\{x\}$ and $\hat{x}:=(x_1, -x_2)$, the reflection of vector $x$ on the $x_1$ axis. For a nonzero vector $x$, we denote by $\bar{x}:=x/\|x\|$.
{Let $ o(\lambda):\mathbb{R}_+\rightarrow \mathbb{R}^m$ stand for} a mapping with the property that $o(\lambda)/\lambda\rightarrow 0$ when $\lambda \downarrow 0$.
For a mapping $F:\mathbb{R}^n\to \mathbb{R}^m$ and vectors $x, d \in \mathbb{R}^n$, we denote by $\nabla F(x)\in \mathbb{R}^{m\times n}$ the Jacobian of $F$ at $x$,
by $\nabla^2 F(x)$ the second-order derivative of $F$ at $x$, and by $\nabla^2 F(x)(d,d)$ the quadratic form corresponding to $\nabla^2 F(x)$.
The directional derivative of $F$ at $x$ in direction $d$ is defined as
$$F'(x;d):=\lim_{{t\downarrow 0}}\frac{F(x+td) -F(x)}{t},
$$ provided that the above limit exists. If $F$ is directionally differentiable at $x$ in direction $d$, its parabolic second-order directional derivative is defined as
$$F''(x;d, w):=\lim_{{t\downarrow 0}}\frac{F(x+td+\frac{1}{2}t^2w)-F(x)-tF'(x;d)}{\frac{1}{2}t^2},
$$ provided that the above limit exists. Moreover {if the following limit exists}
$$F''(x;d, w)=\lim_{{t\downarrow 0}\atop {w'\to w}}\frac{F(x+td+\frac{1}{2}t^2w')-F(x)-tF'(x;d)}{\frac{1}{2}t^2},
$$
then $F$ is
said to be parabolical second-order directionally differentiable at $x$ in the direction $d$ in the sense of Hadamard. In general, the concept of parabolical second-order directional differentiability in the Hadamard sense is stronger than that of parabolical second-order directional differentiability. However, when $F$ is locally
Lipschitz at $x$, these two concepts coincide.
It is known that if $F$ is parabolical second-order directional differentiable in the Hadamard sense at $x$ along $d,w$, then
\[
F(x+td+\frac{1}{2}t^2w+o(t^2))=F(x)+tF'(x;d)+\frac{1}{2}t^2 F''(x;d,w)+o(t^2).
\]
\begin{definition}[Tangent Cones]\label{sotangent}
Let $S\subseteq \mathbb{R}^m$ and $x\in S$. {The regular/Clarke, inner and
(Bouligand-Severi) tangent/contingent cone} to $S$ at $x$ are defined respectively as
\begin{eqnarray*}
\widehat{T}_S(x)&:=& { \liminf\limits_{{x' \stackrel{S}{\to}x} \atop {t\downarrow 0}}\frac{S-x'}{t}=\Big\{ d\in \Re^m\,\Big|\, \forall \, t_k\downarrow 0,\, {x_k \stackrel{S}{\to}x}, \; \exists d_k\to d \
\ {\rm with}\ \ x_k+t_kd_k\in S \Big\}},\\
T^i_S(x)&:=& \liminf\limits_{t\downarrow 0}\frac{S-x}{t}
= \Big\{d\in\Re^m \, \Big| \, \forall \, t_k\downarrow 0,\; \exists d_k\to d \ \ {\rm with}
\ \ x+t_k d_k\in S\Big\},\\
T_S(x)&:=& \limsup\limits_{t\downarrow 0}\frac{S-x}{t}
= \Big\{d\in\Re^m \, \Big| \, \exists \, t_k\downarrow 0,\;d_k\to d \ \ {\rm with}
\ \ x+t_k d_k\in S\Big\}.
\end{eqnarray*}
The inner and
outer second-order tangent sets to $S$ at $x$ in direction $d$ are defined
respectively as
\begin{eqnarray*}
T_S^{i,2}(x; d) &:=& \left \{ w \in \Re^m \, \bigg | \,
{\rm dist} \left( x+t d + \frac{1}{2}t^2w, S \right) = o(t^2), \ \ t\geq 0 \right \}
,\\
T_S^2(x; d) &:= &\left \{ w \in \Re^m \, \bigg | \, \exists \ t_n \downarrow 0 \
{\rm such \ that}\ {\rm dist} \left( x+t_n d + \frac{1}{2}t_n^2w, S \right)
= o(t_n^2) \right \}.
\end{eqnarray*}
\end{definition}
While for a nonconvex set $S$, the contingent cone $T_S(x)$ may be nonconvex, it is known that the regular/Clarke tangent cone $\widehat{T}_S(x)$ is always closed and convex.
By definition, since the distance function of a convex set is convex, it is easy to see that the inner second-order tangent set is always convex when the set $S$ is convex. On the other hand, the outer second-order tangent set may be nonconvex even when the set $S$ is convex (see \cite[Example 3.35]{BS}).
Note that
$T_S^{i,2}(x; d)\subseteq T_S^{2}(x; d)$ and the outer second-order tangent set $T_S^{2}(x; d)$ needs not be a cone (it may be empty; see e.g. an example in \cite[page 592]{RW98}). If $T_S^{i,2}(x; d)=T_S^{2}(x; d)$, we simply call $T_S^{2}(x; d)$ the second-order tangent set to $S$ in direction $d$.
\begin{definition}{\cite[Definition 3.32]{BS}} \label{BS}A set $S$ is said to be second-order directionally differentiable at $x\in S$ in a direction $d\in T_S(x)$, if
$T^i_S(x)=T_S(x)$ and $T^{i,2}_S(x;d)=T_S^2(x;d)$.
\end{definition}
\begin{definition}[Normal Cones] Let $S\subseteq \mathbb{R}^m$ and $x\in S$. The regular/Fr\'{e}chet, limiting/Mordukhovich, and Clarke normal cone of $S$ at ${x}$ are defined respectively as
\begin{eqnarray*}
\widehat{{N}}_S({x}) &:=& \Big \{v\in \mathbb{R}^m\,|\
\langle v, x'-{x}\rangle \leq o(\|x'-x\|) \ \forall x'\in S \Big\},\\
N_S (x)&:=&\limsup\limits_{x' \stackrel{S}{\to} x}\widehat{N}_{S}(x')= \Big \{\lim_{k\rightarrow \infty} v_k|\ v_k \in \widehat{N}_S(x_k),\ \ x_k \stackrel{S}{\to} x \Big\},\\
N^c_S (x)&:=& clcoN_S(x).
\end{eqnarray*}
\end{definition}
\begin{lemma}[Tangent-Normal Polarity](see \cite[Theorem 6.28]{RW98}, \cite{Clarke1983}) \label{polarity}
For a closed set $S\subseteq \mathbb{R}^m$ and $x\in S$, $\widehat{T}_S(x)=(N_S(x))^\circ=(N_S^c(x))^\circ, \widehat{N}_S(x)=(T_S(x))^\circ$, $(\widehat{T}_S(x))^\circ=N_S^c(x)$.
\end{lemma}
We recall some known results concerning the second-order cone ${\cal K}$ in $\Re^m$.
The topological interior and the boundary of ${\cal K}$ are
\begin{eqnarray*}
{\rm int} {\cal K}= \{(x_1,x_2) \in \mathbb{R} \times \mathbb{R}^{m-1}|x_1>\|x_2\|\} \ \ {\rm and}\ \ {\rm bd} {\cal K}=\{(x_1,x_2) \in \mathbb{R} \times \mathbb{R}^{m-1}|x_1=\|x_2\|\},
\end{eqnarray*}
respectively.
Similar to the eigenvalue decomposition of a matrix, for any given vector $x:=(x_1,x_2)\in \Re\times\Re^{m-1}$, $x$ can be decomposed as
(see e.g \cite{FLT02})
\[
x = \lambda_1(x) u_{x}^{(1)} + \lambda_2(x) u_{x}^{(2)},
\]
where $\lambda_i(x)$ and $u_{x}^{(i)}$ for $i=1,2$ are
the spectral values and the associated spectral vectors of $x$
respectively, given by
\[
\lambda_i(x):=x_1+(-1)^i\|x_2\| \quad {\rm and}\quad
u^{(i)}_x:=\left\{\begin{array}{ll}\frac{1}{2}(1, (-1)^i \bar{x}_2) & {\rm if} \ \ x_2\neq 0,\\
\frac{1}{2}(1,(-1)^iw) & {\rm if}\ \ x_2=0,\end{array}\right.
\]
with $w$ being a fixed unit vector in $\mathbb{R}^{m-1}$.
\begin{lemma}(see e.g. \cite[Proposition 2.2]{YZ15})\label{lem2.1}
For any $x,y \in bd {\cal K}\backslash \{0\}$, the following equivalence holds:
$$x^Ty=0 \Longleftrightarrow y=k\hat{x} \mbox{ with } k=y_1/x_1>0 \Longleftrightarrow y=k\hat{x} \mbox{ with } k\in \mathbb{R}_{++}.$$
\end{lemma}
For a given real-valued function $f:\Re\to \Re$, we define the SOC function
$f^{\rm soc}: \Re^m \to \Re^m$ as
\begin{equation}
f^{\rm soc}(z) := f \left( \lambda_1(z) \right) u^{(1)}_z
+ f \left( \lambda_2(z) \right) u^{(2)}_z.\label{SOCfcn1}
\end{equation}
For $z\in \Re^m$, let $\Pi_{{\cal K}}(z)$ be the metric projection of $z$ onto ${\cal K}$. Then by \cite{FLT02}, it can be calculated as \begin{equation}
\Pi_{{\cal K}}(z)=\lambda_1(z)_+u^{(1)}_z
+ \lambda_2(z)_+ u^{(2)}_z,\label{SOCfcn2}
\end{equation}
where $\alpha_+:=\max\{\alpha,0\}$ is the nonnegative part of the number $\alpha\in \mathbb{R}$.
Hence the projection operator $\Pi_{\cal K}(\cdot)$ is an SOC function corresponds to the plus function $f(\alpha):=\alpha_+$.
\section{{Second-order directional derivative of the projection operator over the second-order cone}}
As commented in the introduction, there exists a close relationship between the second-order tangent set of the SOC complementarity set and the second-order directional derivative of the projection operator $\Pi_{\cal K}$; see (\ref{sdirectionalderivative}). Therefore, to obtain the exact formula of the second-order tangent set, we need to calculate the second-order directional derivative of the projection operator $\Pi_{\cal K}$. This task is done in this section, which is of independent interest.
For the convenience of notations, we sometime use $\Phi(x)$ instead of $\bar{x}$ to stand for ${x}/{\|x\|}$ as $x \neq 0$. It is easy to verify (see e.g. \cite[Theorem 3.1]{ZTC15}) that
$\Phi$ is second-order continuously
differentiable at $x\neq 0$ with
\begin{eqnarray*}
{\nabla} \Phi(x) &=& ( I-\bar{x}\bar{x}^T) /{\|x\|},\\
{\nabla}^2\Phi(x)(w,w) &=& -2 \frac{\bar{x}^Tw}{\|x\|^2} w
+ w^T \left (\frac{ 3 \bar{x}\bar{x}^T -I }{\|x\|^3} \right) wx \\
&=& -2\frac{\bar{x}^Tw}{\|x\|}\nabla\Phi(x)(w)-\frac{1}{\|x\|}w^T\nabla \Phi(x)w \bar{x},
\end{eqnarray*}
where $I$ is the identity matrix in $\Re^{m\times m}$.
Since the second-order cone ${\cal K}$ is a special circular cone $\mathcal{L}_{\theta}$ defined by
$$\mathcal{L}_{\theta}:=\{(x_1,x_2)\in \Re\times \Re^{n-1}| \cos\theta \|x\|\leq x_1\}$$
with $\theta=45^\circ$, the SOC function $f^{\rm soc}$ is a special case of the circular cone function $f^{\mathcal{L}_{\theta}}$ studied in \cite{ZTC15} with $\theta= 45^{\circ}$. The following result follows from \cite[Theorem 3.3]{ZTC15} immediately.
\begin{lemma} \label{second-order diff}
Suppose that $f:\Re\to \Re$. Then, the SOC function $f^{\rm soc}$ is parabolic second-order
directionally differentiable at $x$ in the Hadamard sense if and only if
$f$ is parabolic second-order directionally differentiable at $\lambda_i(x)$
in the Hadamard sense for $i=1,2$. Moreover,
\begin{description}
\item[(i)] if $x_2=0$ and $d_2=0$, then
\[
(f^{\rm soc})^{''}(x;d,w)
= f^{''} \left( x_1;d_1,w_1-\|w_2\| \right) u_w^{(1)}
+ f^{''} \left( x_1;d_1,w_1+\|w_2\| \right) u_w^{(2)};
\]
\item[(ii)] if $x_2=0$ and $d_2 \neq 0$, then
\begin{eqnarray*}
(f^{\rm soc})^{''}(x;d,w) &=& f^{''} \left( x_1;d_1-\|d_2\|,w_1-\bar{d}_2^Tw_2 \right) u_d^{(1)} \\
& & + f^{''} \left( x_1;d_1+\|d_2\|,w_1+\bar{d}_2^Tw_2 \right) u_d^{(2)} \\
& & + \frac{1}{2}\Big(f'(x_1;d_1+\|d_2\|)
- f'(x_1;d_1-\|d_2\|) \Big)
\left (
\begin{matrix}
0\\
{\nabla}{\Phi}(d_2)w_2 \end{matrix}\right );
\end{eqnarray*}
\item[(iii)] if $x_2\neq 0$, then
\begin{eqnarray*}
\lefteqn{ (f^{\rm soc})^{''}(x;d,w)}\\
&=& f^{''}\left( x_1-\|x_2\|;d_1-\bar{x}_2^Td_2,w_1
- \big[ \bar{x}_2^T w_2+d_2^T {\nabla}\Phi(x_2) d_2\big] \right)u_x^{(1)} \\
& & + f^{''} \left( x_1+\|x_2\|;d_1+\bar{x}_2^Td_2,w_1
+ \big[ \bar{x}_2^T w_2+d_2^T {\nabla}\Phi(x_2)d_2\big] \right) u_x^{(2)} \\
& & + \Big(f'(x_1+\|x_2\|;d_1+\bar{x}_2^Td_2)
- f'(x_1-\|x_2\|;d_1-\bar{x}_2^Td_2)\Big) \left ( \begin{matrix}
0\\
{\nabla}{\Phi}(x_2)d_2 \end{matrix}\right ) \\
& & + \frac{1}{2}\Big(f(x_1+\|x_2\|)-f(x_1-\|x_2\|)\Big)\left ( \begin{matrix}
0\\
\nabla {\Phi}(x_2)w_2+ {\nabla}^2{\Phi}(x_2)(d_2,d_2) \end{matrix}\right ).
\end{eqnarray*}
\end{description}
\end{lemma}
Since the projection operator $\Pi_{\cal K}(\cdot)$ is the SOC function corresponding to the plus function $f(\alpha):=\alpha_+$, we will need the second-order directional derivative of the plus function.
\begin{lemma}(see e.g. \cite{ZZX13}) \label{Lem3.4} Let $f(\alpha):=\alpha_+$ for $\alpha \in \Re$. Then $f$ is parabolic second-order directionally differentiable at $x$
in the Hadamard sense and
\begin{equation}\nonumber
f'(x;d)=\left\{
\begin{array}{ll}
d & {\rm if} \ x>0, \\
d_{+} & {\rm if} \ x=0, \\
0 & {\rm if} \ x<0,
\end{array}
\right. \ \ \ {\rm and}\
\quad f^{''}(x;d,w)=\left\{
\begin{array}{ll}
w & {\rm if} \ x>0 \ {\rm or} \ x=0, d>0, \\
0 & {\rm if} \ x<0 \ {\rm or} \ x=0, d<0, \\
w_{+} & {\rm if} \ x=d=0.
\end{array}
\right.
\end{equation}
\end{lemma}
Since in the formula of the second-order directional derivative of the projection operator, we will need the tangent cone and the second-order tangent set for the set ${\cal K}$ and its polar ${\cal K}^\circ$, for convenience we summarize their formulas in the following two lemmas.
\begin{lemma}\cite[Lemma 25 and Lemma 27]{BR05}\label{Lem3.3} For any $x\in {\cal K}$, one has\begin{eqnarray*}
T_{\cal K}(x)
&=& \left\{
\begin{array}{lll}
& \mathbb{R}^m \ &{\rm if } \ x\in {\rm int}{\cal K}; \\
& {\cal K} \ &{\rm if}\ x=0; \\
& \{ d\in \mathbb{R}^m| -d_1+\bar{x}_2^T d_2\leq 0\} \ \ & {\rm if}\ x\in {\rm bd}{\cal K}\backslash\{0\}.
\end{array}
\right.
\end{eqnarray*}For any $x\in {\cal K}$ and $d\in T_{\cal K}(x)$,
\[
T_{{\cal K}}^2(x;d)=\left\{
\begin{array}{ll}
\Re^m \ & \ {\rm if } \ d\in {\rm int}T_{\cal K}(x); \\
T_{{\cal K}}(d) &\ {\rm if }\ x=0; \\
\{w|\ w_2^Tx_2-w_1x_1\leq d_1^2-\|d_2\|^2\} & \ {\rm if } \ x\in {\rm bd}{\cal K}\backslash \{0\} \mbox{ and } d\in {\rm bd}T_{{\cal K}}(x).
\end{array}
\right.
\]
\end{lemma}
{ Applying \cite[Lemma 25 and Lemma 27]{BR05} to ${\cal K}^\circ=-{\cal K}$ yields the following result.}
{ \begin{lemma}\label{Lem3.5}
For $x\in {\cal K}^\circ$, one has
\[
T_{{\cal K}^\circ}(x)=\left\{\begin{array}{ll}
\Re^m & \ {\rm if } \ x\in {\rm int}{\cal K}^\circ;\\
{\cal K}^\circ & \ {\rm if } \ x=0;\\
\{d\in \mathbb{R}^m| d_1+\bar{x}_2^Td_2\leq 0\} & \ {\rm if } \ x\in {\rm bd}{\cal K}^\circ\backslash\{0\}.
\end{array}\right.
\]
For $x\in {\cal K}^\circ$ and $d\in T_{{\cal K}^\circ}(x)$, one has
\[
T^2_{{\cal K}^\circ}(x;d)=\left\{\begin{array}{ll}
\Re^m & \ {\rm if } \ d\in {\rm int}T_{{\cal K}^\circ}(x);\\
T_{{\cal K}^\circ}(d) & \ {\rm if } \ x=0;\\
\{w| w_2^Tx_2-w_1x_1\leq d_1^2-\|d_2\|^2\} & \ {\rm if } \ x\in {\rm bd}{\cal K}^\circ\backslash \{0\} \mbox{ and } d\in {\rm bd}T_{{\cal K}^\circ}(x).
\end{array}\right.
\]
\end{lemma}}
We are now ready to give the second-order directional derivative of the projection operator.
\begin{theorem} \label{projection-second-order diff}
The projection operator $\Pi_{{\cal K}}$ is parabolic second-order
directionally differentiable in the Hadamard sense. Moreover, for any $x,d,w\in \Re^m$, the second-order directional derivative can be calculated as in the following six cases.
\begin{description}
\item[Case (i)] $x\in {\rm int} {\cal K}$.
$\Pi_{{\cal K}}^{''}(x;d,w)=w$.
\item[Case (ii)] $x\in {\rm int}{\cal K}^\circ$.
$\Pi_{{\cal K}}^{''}(x;d,w)=0$.
\item [Case (iii)] $x=0$.
\[
\Pi_{{\cal K}}^{''}(x;d,w)=\left\{\begin{array}{ll}
w \ \ \ \ \ \ \ \ \ \ \ \ & \ {\rm if } \ d\in {\rm int}{\cal K}, \\
0 & \ {\rm if } \ d\in {\rm int}{\cal K}^\circ,\\
\frac{1}{2}\left(\begin{matrix}
w_1+\bar{d}_2^Tw_2\\
\left[w_1-\frac{d_1}{\|d_2\|} {\bar{d}_2}^Tw_2\right]\bar{d}_2+\left[1+\frac{d_1}{\|d_2\|}\right]w_2
\end{matrix}\right) & \ {\rm if } \ d\in ({\cal K}\cup{\cal K}^\circ)^c, \\
w & \ {\rm if } \ d\in {\rm bd}{\cal K}\backslash\{0\}, w\in T_{{\cal K}}(d),\\
\frac{1}{2}\left(\begin{matrix}
w_1+\bar{d}_2^Tw_2\\
2w_2+(w_1-\bar{d}_2^Tw_2)\bar{d}_2\end{matrix}\right) & \ {\rm if } \ d\in {\rm bd}{\cal K}\backslash\{0\}, w\notin T_{{\cal K}}(d), \\
0 & \ {\rm if } \ d\in {\rm bd}{\cal K}^\circ\backslash \{0\}, w\in T_{{\cal K}^\circ}(d),\\
\frac{1}{2}(w_1+\bar{d}_2^Tw_2)\left(\begin{matrix}
1 \\
\bar{d}_2
\end{matrix}\right) & \ {\rm if } \ d\in {\rm bd}{\cal K}^\circ\backslash \{0\}, w\notin T_{{\cal K}^\circ}(d),\\
\Pi_{{\cal K}}(w) & \ {\rm if } \ d=0.
\end{array} \right.
\]
\item[Case (iv)] $x\in {\rm bd}{\cal K}\backslash \{0\}$.
\begin{eqnarray*}
\lefteqn{\Pi_{{\cal K}}^{''}(x;d,w)=}\\
&& \left\{\begin{array}{l}
w \qquad \qquad \qquad \qquad \ {\rm if } \ d\in {\rm int}T_{{\cal K}}(x), \ \ \ \ \ \ \ \ \\
w \qquad \qquad \qquad \qquad \ {\rm if } \ d\in {\rm bd}T_{{\cal K}}(x), w\in T^2_{{\cal K}}(x;d), \\
\frac{1}{2}\left(\begin{matrix}
w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\\
\left[w_1-\bar{x}_2^Tw_2-\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right]\bar{x}_2+2w_2
\end{matrix}\right)\qquad \ {\rm if } \ d\in {\rm bd} T_{{\cal K}}(x), w\notin T^2_{{\cal K}}(x;d), \\
\frac{1}{2}\left(
\begin{matrix}
w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-(\bar{x}_2^Td_2)^2}{\|x_2\|}\\
\left[w_1-\bar{x}_2^Tw_2-\frac{\|d_2\|^2-3(\bar{x}_2^Td_2)^2+2d_1\bar{x}_2^Td_2}{\|x_2\|}\right]\bar{x}_2
+2w_2+2\frac{d_1-\bar{x}_2^Td_2}{\|x_2\|}d_2
\end{matrix}
\right) \quad \ {\rm if } \ d\in T_{\cal K}(x)^c.
\end{array}\right.
\end{eqnarray*}
\item[Case (v)] $x\in {\rm bd}{\cal K}^\circ\backslash \{0\}$.
\begin{eqnarray*}
\lefteqn{\Pi_{{\cal K}}^{''}(x;d,w)=}\\
&&\left\{\begin{array}{l}
0 \qquad \qquad \qquad \qquad \ {\rm if } \ d\in {\rm int}T_{{\cal K}^\circ}(x), \\
0 \qquad \qquad \qquad \qquad \ {\rm if } \ d\in {\rm bd}T_{{\cal K}^\circ}(x), w\in T^2_{{\cal K}^\circ}(x;d), \\
\frac{1}{2}\left(w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right)
\left(\begin{matrix}
1 \\ \bar{x}_2
\end{matrix}\right) \qquad \ {\rm if } \ d\in {\rm bd} T_{{\cal K}^\circ}(x), w\notin T^2_{{\cal K}^\circ}(x;d),\\
\frac{1}{2}\left(\begin{matrix}
w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-(\bar{x}_2^Td_2)^2}{\|x_2\|}\\
\left[w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-3(\bar{x}_2^Td_2)^2-2d_1\bar{x}_2^Td_2}{\|x_2\|}\right]\bar{x}_2+
2\frac{d_1+\bar{x}_2^Td_2}{\|x_2\|}d_2
\end{matrix}\right) \qquad \ {\rm if } \ d\in T_{{\cal K}^\circ}(x)^c.
\end{array}\right.
\end{eqnarray*}
\item[Case (vi)] $x\in ({\cal K}\cup{\cal K}^\circ)^c$.
\end{description}
\begin{eqnarray*}
&&\Pi_{{\cal K}}^{''}(x;d,w)=\\
&&\frac{1}{2}\left(
\begin{matrix}
w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-(\bar{x}_2^Td_2)^2}{\|x_2\|}\\
\left[w_1-\frac{x_1}{\|x_2\|}\bar{x}_2^Tw_2-\frac{x_1}{\|x_2\|^2}\big(\|d_2\|^2-3(\bar{x}_2^Td_2)^2\big)
-2d_1\frac{\bar{x}_2^Td_2}{\|x_2\|}\right]\bar{x}_2+2\frac{\|x_2\|d_1-x_1\bar{x}_2^Td_2}{\|x_2\|^2}d_2
+\left[1+\frac{x_1}{\|x_2\|}\right]w_2\end{matrix}
\right).
\end{eqnarray*}
\end{theorem}
\par\noindent {\bf Proof.}\ \
By (\ref{SOCfcn1})-(\ref{SOCfcn2}), the projection operator $\Pi_{{\cal K}}$ is the SOC function $f^{soc}$ with
$f(t):=t_+$. Applying Lemmas \ref{second-order diff} and \ref{Lem3.4} will give the parabolic second-order
directional differentiability of $\Pi_{{\cal K}}$ in the Hadamard sense and a formula for $\Pi^{''}_{\cal K}$. However in some cases the formula obtained will still involve the plus operator $(\cdot)_+$. In this theorem we aim at obtaining the exact formula as proposed. For some cases, e.g., in the cases $x\in {\rm int}{\cal K}$; $x\in {\rm int}{\cal K}^\circ$; $x=0,d\in {\rm int}{\cal K}$; $x=0,d\in {\rm int}{\cal K}^\circ$; $x=0,d=0$, we can prove the results by directly using the definition of second-order directional derivative.
In some other cases, e.g., in the cases $x=0, d\in {\rm bd}{\cal K}\backslash \{0\}$; $x=0,d\in {\rm bd}{\rm {\cal K}^\circ}\backslash \{0\}$; $x \in {\rm bd}{{\cal K}}\backslash\{0\}, d\in {\rm bd}T_{{\cal K}}(x)$; $x\in {\rm bd}{\cal K}^\circ\backslash \{0\}, d\in {\rm bd} T_{{\cal K}^\circ}(x)$, we can further use the representation of tangent cones in Lemmas \ref{Lem3.3} and \ref{Lem3.5} to obtain the proposed exact formula. For simplicity, we only prove some of the cases. The others can be obtained by following similar arguments.
\noindent {\bf Case }
$x\in {\rm int}{\cal K}$. In this case $ \Pi_{\cal K}(x)=x$, $\Pi_{\cal K}'(x;d)=d$ and $\Pi_{{\cal K}}(x+td+\frac{1}{2}t^2w)=x+td+\frac{1}{2}t^2w$ for $t>0$ sufficiently small. Hence
$$\Pi_{{\cal K}}^{''}(x;d,w):=\lim_{t\downarrow 0} \frac{ \Pi(x+td+\frac{1}{2}t^2 w)- \Pi_{\cal K}(x)-t\Pi_{\cal K}'(x;d)}{\frac{1}{2}t^2}=w. $$
\noindent {\bf Case } $x=0$ and $d\in {\rm int}{\cal K}$. In this case $\Pi_{\cal K}(x)=0$ and $ \Pi_{\cal K}'(x;d)=d$. Note that \[\Pi_{{\cal K}}(x+td+\frac{1}{2}t^2w)=\Pi_{{\cal K}}(td+\frac{1}{2}t^2w)=
td+\frac{1}{2}t^2w,\]
for $t>0$ sufficiently small. Hence $\Pi_{{\cal K}}^{''}(x;d,w)=w.$
\noindent {\bf Case } $x=0$ and $ d=0$. It is obvious that $\Pi_{\cal K}(0)=0, \Pi_{\cal K}'(0;0)=0$ and
$ \Pi_{{\cal K}}(x+td+\frac{1}{2}t^2w)= \Pi_{{\cal K}}(\frac{1}{2}t^2w)=\frac{1}{2}t^2\Pi_{{\cal K}}(w)$. Hence $\Pi_{{\cal K}}^{''}(x;d,w)=\Pi_{{\cal K}}(w)$.
\noindent {\bf Case} $x=0$ and $d\in {\rm bd}{\cal K}\backslash \{0\}$. Then $\Pi_{\cal K}(x)=0$ and $d_1=\|d_2\|\neq 0$. Directly applying Lemmas \ref{second-order diff}(ii) and \ref{Lem3.4} yield
\begin{eqnarray}
\Pi_{{\cal K}}^{''}(x;d,w)& =&
\frac{1}{2}(w_1-\bar{d}_2^Tw_2)_+\left (\begin{matrix}1\\ -\bar{d}_2\end{matrix} \right ) + \frac{1}{2}(w_1+\bar{d}_2^Tw_2)\left (\begin{matrix}1\\ \bar{d}_2\end{matrix} \right ) + \left (\begin{matrix}0\\( I-\bar{d}_2\bar{d}_2^T)w_2\end{matrix} \right). \label{eqn3-3-4-1}
\end{eqnarray}
Recall from Lemma \ref{Lem3.3} that $w\in T_{{\cal K}}(d)$ if and only if $w_1\geq \bar{d}_2^Tw_2$. It follows from (\ref{eqn3-3-4-1}) that
$$\Pi_{{\cal K}}^{''}(x;d,w)=\left \{ \begin{array}{ll}
w& \mbox{ if } w\in T_{{\cal K}}(d),\\
\frac{1}{2}\left(\begin{matrix}
w_1+\bar{d}_2^Tw_2\\
2w_2+(w_1-\bar{d}_2^Tw_2)\bar{d}_2\end{matrix}\right)
&\mbox{ if } w\not \in T_{{\cal K}}(d).\end{array} \right. $$
\noindent {\bf Case} $x \in {\rm bd}{{\cal K}}\backslash\{0\}$ and $d\in {\rm bd}T_{{\cal K}}(x)$. Then $ x_1=\|x_2\|\not =0$.
and $-d_1+\bar{x}_2^T d_2=0$. Directly applying Lemmas \ref{second-order diff}(iii) and \ref{Lem3.4} yield
\begin{eqnarray}
\lefteqn{ \Pi_{{\cal K}}^{''}(x;d,w)}\nonumber\\
&=& \frac{1}{2}\left( w_1- \left[\bar{x}_2^Tw_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right] \right)_+\left(\begin{matrix}1\\ -\bar{x}_2\end{matrix}\right) + \frac{1}{2}\left(\begin{matrix}
w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\\
\left(w_1-\bar{x}_2^Tw_2-\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right)\bar{x}_2+2w_2
\end{matrix}\right). \nonumber\\
\label{eqn3-4-2}
\end{eqnarray}
Recall from Lemma \ref{Lem3.3} that $w\in T^2_{{\cal K}}(x;d)$ if and only if $w_2^Tx_2-w_1x_1\leq d_1^2-\|d_2\|^2 $.
Hence it follows from (\ref{eqn3-4-2}) that $\Pi_{{\cal K}}^{''}(x;d,w)=w$ if $w\in T^2_{{\cal K}}(x;d)$ and
\begin{eqnarray*}
\Pi_{{\cal K}}^{''}(x;d,w)= \frac{1}{2}\left(\begin{matrix}
w_1+\bar{x}_2^Tw_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\\
\left(w_1-\bar{x}_2^Tw_2-\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right)\bar{x}_2+2w_2
\end{matrix}\right)
\end{eqnarray*}
if $w\notin T^2_{{\cal K}}(x;d)$.
\hskip .5cm $\Box$ \vskip .5cm
\section{Second-order tangent set for the SOC complementarity set}
This section is devoted to deriving the exact formula for the second-order tangent set to the SOC complementarity set. To this end, we first build its connection with the second-order directional derivative of the projection operator $\Pi_{\cal K}$, whose existence is guaranteed by virtue of Theorem \ref{projection-second-order diff}.
\begin{proposition}\label{relationship}
For any $(x,y)\in \Omega$ and
$(d,w)\in T_{\Omega}(x,y)$, one has
\[
T^{i,2}_{\Omega}\big((x,y);(d,w)\big)=T^2_{\Omega}\big((x,y);(d,w)\big)
= \big \{(p,q) \, | \, \Pi^{''}_{{\cal K}}(x-y;d-w,p-q)=p \big \}.
\]
\end{proposition}
\par\noindent {\bf Proof.}\ \
Since $T^{i,2}_{\Omega}\big((x,y);(d,w)\big)\subseteq T^{2}_{\Omega}\big((x,y);(d,w)\big)$, it suffices to show
\[
T^{2}_{\Omega}\big((x,y);(d,w)\big) \subseteq \Upsilon\big((x,y);(d,w) \big) \subseteq T^{i,2}_{\Omega}\big((x,y);(d,w)\big),
\]
where $
\Upsilon\big((x,y);(d,w) \big):=\big \{(p,q) \, | \, \Pi^{''}_{{\cal K}}(x-y;d-w,p-q)=p \big \}.
$
Let $(p,q)\in T^{2}_{\Omega}\big((x,y);(d,w)\big)$. Then by definition, there exist $t_n \downarrow 0$, $(\alpha(t_n),\beta(t_n))=o(t_n^2)$ such that
$(x,y)+t_n(d,w) + \frac{1}{2}t_n^2 (p,q)
+ (\alpha(t_n),\beta(t_n))\in \Omega.$
By the equivalence in (\ref{Omeganew}), it follows that
\begin{eqnarray*}
\lefteqn{ \Pi_{{\cal K}} \Big(x-y+t_n (d-w)+\frac{1}{2}t_n^2 (p-q)+\alpha(t_n)-\beta(t_n)\Big)} \\
&=& x+t_nd + \frac{1}{2}t_n^2p + \alpha(t_n) \\
&=& \Pi_{{\cal K}}(x-y) + t_n \Pi'_{\cal K}(x-y;d-w) + \frac{1}{2}t_n^2p + \alpha(t_n),
\end{eqnarray*}
where the last equality follows from the equivalence in (\ref{Omeganew}) and
(\ref{directionalderivative}).
Hence, $\Pi^{''}_{{\cal K}}(x-y;d-w,p-q)=p$, i.e., $(p,q)\in \Upsilon\big((x,y);(d,w) \big)$.
Now, take $(p,q)\in \Upsilon\big((x,y);(d,w) \big)$, i.e., $ \Pi^{''}_{{\cal K}}(x-y;d-w,p-q)=p$. For $t>0$, define
\[
r(t):=\Pi_{{\cal K}}(x-y+t(d-w)+\frac{1}{2}t^2(p-q))-
\Pi_{{\cal K}}(x-y)-t\Pi'_{{\cal K}}(x-y;d-w)-\frac{1}{2}t^2\Pi^{''}_{{\cal K}}(x-y;d-w,p-q).
\]
Then $r(t)=o(t^2)$ according to the second-order directional differentiability of $\Pi_{{\cal K}}$ by Theorem \ref{projection-second-order diff}. Note that
\begin{eqnarray*}
&&\Pi_{{\cal K}}\big(x-y+t(d-w)+\frac{1}{2}t^2p+r(t)-\frac{1}{2}t^2q-r(t)
\big)\\
&=&\Pi_{{\cal K}}(x-y+t(d-w)+\frac{1}{2}t^2(p-q))\\
&=& \Pi_{{\cal K}}(x-y)+t\Pi'_{{\cal K}}(x-y;d-w)+\frac{1}{2}t^2
\Pi_{{\cal K}}^{''}(x-y;d-w,p-q)+r(t)\\
&=& x+td+\frac{1}{2}t^2p+r(t),
\end{eqnarray*}
where the last equality follows from the equivalence in (\ref{Omeganew}) and
(\ref{directionalderivative}).
This together with equivalence (\ref{Omeganew}) yields
that
\[
\left(x+td+\frac{1}{2}t^2p+r(t), y + tw +
\frac{1}{2}t^2q+r(t) \right)\in \Omega.
\]
It means $(p,q)\in T^{i,2}_{\Omega}\big((x,y);(d,w)\big)$. The proof is complete.
\hskip .5cm $\Box$ \vskip .5cm
\begin{remark} The proof of equivalence (\ref{sdirectionalderivative}) in Proposition \ref{relationship} is very similar to that of equivalence (\ref{directionalderivative}) as in \cite[Proposition 5.2]{YZ16}. Note that although the equivalence (\ref{directionalderivative}) was shown in \cite[Proposition 3.1]{JLZ15}, the proof in \cite[Proposition 5.2]{YZ16} is much more concise without going over each possible cases as in \cite[Proposition 3.1]{JLZ15}. Moreover from the proof of \cite[Proposition 5.2]{YZ16}, one can {see} that the equivalence (\ref{directionalderivative}) holds for any general convex cone ${\cal K}$ as long as the projection operator $\Pi_{{\cal K}}$ satisfies the Lipschitz continuity and directional differentiability.
Similarly from the proof of Proposition \ref{relationship}, we can see that equivalence (\ref{sdirectionalderivative}) in Proposition \ref{relationship} holds for any general convex cone ${\cal K}$ whenever the projection operator $\Pi_{{\cal K}}$ satisfies the Lipschitz continuity and parabolic second-order directional differentiability in the Hadamard sense.
\end{remark}
The above result tells us that for characterizing the structure of the
second-order tangent set to $\Omega$, we need to study the expression
of the second-order directional derivative of the projection operator
$\Pi_{{\cal K}}$, which has been obtained in Theorem \ref{projection-second-order diff}. With these preparations, the explicit expression of the second-order
tangent set to $\Omega$ is given below. For convenience, we recall the formula for the tangent cone first.
\begin{lemma}\label{tangentcone}\cite[Theorem 5.1]{YZ16}
For any $(x,y)\in \Omega$,
\begin{eqnarray*}
\lefteqn{T^{i}_\Omega(x,y)=T_\Omega(x,y)}\\
&=& \left\{ (d,w) \left|
\begin{array}{lll}
& d\in \mathbb{R}^m,\ w=0, \ &{\rm if}\ x\in {\rm int}{\cal K}, \ y=0; \\
& d=0,\ w\in \mathbb{R}^m, \ &{\rm if} \ x=0, \ y\in {\rm int}{\cal K}; \\
& x_1\hat{w}-y_1d\in \mathbb{R} x, \ \ d\perp y,\ w\perp x, \ \ & {\rm if}
\ x,y \in {\rm bd}{\cal K}\backslash\{0\}; \\
& d\in T_{\cal K}(x),\ w=0 \ {\rm or}\ d\perp \hat{x}\ , w\in \mathbb{R}_{+}\hat{x}, \ \
& {\rm if}\ x\in {\rm bd}{\cal K}\backslash \{0\},\ y=0; \\
& d=0,\ w\in T_{\cal K}(y) \ {\rm or}\ d\in \mathbb{R}_{+}\hat{y}, \ w\perp \hat{y}, \ \
& {\rm if}\ x=0,\ y\in {\rm bd}{\cal K}\backslash \{0\}; \\
& d\in {\cal K},\ w\in {\cal K}, \ d\perp w, \ \ & {\rm if}\ x=0, \ y=0.
\end{array}
\right.\right\}.
\end{eqnarray*}
\end{lemma}
According to Proposition \ref{relationship} and Lemma \ref{tangentcone}, we obtain the following result.
\begin{theorem}\label{second-order-directionally}
The set $\Omega$ is second-order directionally differentiable at every $(x,y)\in \Omega$ in every direction $(d,w)\in T_{\Omega}(x,y)$.
\end{theorem}
\begin{remark}
It is well-known that for a convex set, the tangent cone and inner tangent cone coincide, but the inner and outer second-order tangent sets can be different; see \cite[Example 3.31]{BS}. Here we show that SOC complementarity set $\Omega$, although it is nonconvex, is second-order directionally differentiable, i.e., the tangent cone and inner tangent cone coincide, and the inner and outer second-order tangent sets coincide as well.
\end{remark}
The inner and outer second-order tangent set to product sets have been studied in \cite[Page 168]{BS}. Particularly, for $C:=C_1\times\cdots\times C_m$ with $C_i\in \Re^{n_i}$, at certain $x=(x_1,\dots,x_m)$ with $x_i\in C_i$, according to \cite{BS},
\[
T^{i,2}_{C}(x;d)=T^{i,2}_{C_1}(x_1;d_1)\times\dots T^{i,2}_{C_m}(x_m;d_m)
\]
and
\begin{equation}\label{R-product-1}
T^{2}_{C}(x;d)\subset T^{2}_{C_1}(x_1;d_1)\times\dots T^{2}_{C_m}(x_m;d_m).
\end{equation}
If all except at most one of $C_i$ are second-order directional differentiable, then the equality holds in (\ref{R-product-1}).
Noting that second-order cone complementarity set is second-order directional differentiable, Theorem \ref{second-order-directionally} can be then extended to the Cartesian product of finitely many second-order cone complementarity sets.
\begin{corollary}
Suppose that $\Omega_1,\cdots,\Omega_l$ are all SOC complementarity sets. Then the Cartesian product $\Omega:=\Omega_1\times \Omega_2\times \cdots \times \Omega_l$ is second-order directionally differentiable at every $(x,y)\in \Omega$ in every direction $(d,w)\in T_{\Omega}(x,y)$ and
$$T^2_{\Omega}((x,y);(d,w))=T^2_{\Omega_1}((x_1,y_1);(d_1,w_1))\times \cdots \times T^2_{\Omega_1}((x_l,y_l);(d_l,w_l)).$$
\end{corollary}
\par\noindent {\bf Proof.}\ \
Since $(d,w)\in T_{\Omega}(x,y)=T_{\Omega_1} (x_1,y_1)\times \dots \times T_{\Omega_l} (x_l,y_l)$, $(d_i,w_i)\in T_{\Omega_i}(x_i,y_i)$ for $i=1,\dots,l$. Take $(p,q)\in T^2_{\Omega}((x,y);(d,w))$. Hence
\begin{eqnarray*}
T^2_{\Omega}((x,y);(d,w)) &\subseteq & T^2_{\Omega_1}((x_1,y_1);(d_1,w_1))\times \cdots \times T^2_{\Omega_1}((x_l,y_l);(d_l,w_l)) \\
& =& T^{i,2}_{\Omega_1}((x_1,y_1);(d_1,w_1))\times \cdots \times T^{i,2}_{\Omega_l}((x_l,y_l);(d_l,w_l))\\
&= & T^{i,2}_{\Omega}((x,y);(d,w)),
\end{eqnarray*}
where the first inclusion and the second equation follows from \cite[Page 168]{BS}, and the first equation comes from Theorem \ref{second-order-directionally}.
\hskip .5cm $\Box$ \vskip .5cm
\begin{theorem} \label{formula-regular normal cone-1}
For any $(x,y)\in \Omega$ and
$(d,w)\in T_\Omega(x,y)$, the formula of the second-order tangent set for the SOC complementarity set can be described as in the following six cases.
\begin{description}
\item[Case (i)] $ x\in {\rm int} {\cal K}\ {\rm and } \ y=0$.
$ T^2_\Omega\big((x,y);(d,w)\big)=\Re^m \times \{0\}. $
\item[Case (ii)] $ x=0 \ {\rm and }\ y \in {\rm int} {\cal K}$.
$ T^2_\Omega\big((x,y);(d,w)\big)=\{0\}\times \Re^m .$
\item[Case (iii)] $x,y\in {\rm bd}{\cal K}\backslash \{0\}$.
\begin{eqnarray*}
\lefteqn{ T^2_\Omega\big((x,y);(d,w)\big)}\\
&& = \left\{(p,q)\left|
\begin{array}{ll}
p\in {\rm bd} T^2_{{\cal K}}(x;d), \ \ q\in {\rm bd}T^2_{{\cal K}}(y;w), \\
(x_1w_1-y_1d_1)\left(\frac{w_2-w_1\bar{y}_2}{y_1}-\frac{d_2-d_1\bar{x}_2}{x_1}\right)-
p_1y_2-q_1x_2=x_1q_2+y_1p_2
\end{array} \right.\right\}.
\end{eqnarray*}
\item[Case (iv)] $x\in {\rm bd}{\cal K}\backslash \{0\}$ and $y=0$.
\begin{eqnarray*}
\lefteqn{T^2_\Omega\big((x,y);(d,w)\big)} \\
&=&\left\{(p,q)\left|
\begin{array}{lll}
q=0, \ & {\rm if} \ d\in {\rm int} T_{{\cal K}}(x), \ w=0; \\
p\in T_{{\cal K}}^2(x;d),\ q=0, \ {\rm or} \ \
p\in {\rm bd} T^2_{{\cal K}}(x;d), \ q\in \Re_+\hat{x}\ & {\rm if}\ d\in {\rm bd}T_{{\cal K}}(x),\ w=0; \\
p \in {\rm bd}T^2_{{\cal K}}(x;d),\ -q_1\bar{x}_2-2\frac{w_1d_2}{\|x_2\|}
-2 \frac{d_1w_2}{\|x_2\|}=q_2, \ \ & {\rm if}\ d\perp \hat{x}, \ w\in \Re_{++}\hat{x}.
\end{array}
\right.\right\}.
\end{eqnarray*}
\item[Case (v)] $x=0$ and $y\in {\rm bd}{\cal K}\backslash \{0\}$.
\begin{eqnarray*}
\lefteqn{ T^2_\Omega\big((x,y);(d,w)\big)}\\
&=& \left\{(p,q)\left|
\begin{array}{lll}
p=0, \ & {\rm if} \ d=0,\ w\in {\rm int} T_{{\cal K}}(y); \\
p=0,\ q\in T_{{\cal K}}^2(y;w), \ {\rm or} \ p\in \Re_+\hat{y},
\ \ q\in {\rm bd} T^2_{{\cal K}}(y;w)\ & {\rm if}\ d=0,\ w\in {\rm bd}T_{{\cal K}}(y); \\
q\in {\rm bd}T^2_{{\cal K}}(y;w),\ -p_1\bar{y}_2-2\frac{w_1d_2}{\|y_2\|}
- 2\frac{d_1w_2}{\|y_2\|}=p_2, \ \ & {\rm if}\ d\in \Re_{++}\hat{y}, \ w\perp \hat{y}.
\end{array}
\right.\right\}.
\end{eqnarray*}
\item[Case (vi)] $ x=y=0$.
$ T^2_\Omega\big((x,y);(d,w)\big)= T_\Omega(d,w) .$
\end{description}
\end{theorem}
\par\noindent {\bf Proof.}\ \
By Proposition \ref{relationship}, to describe an element $(p,q)\in T_\Omega^2((x,y);(d,w))$, it suffices to
describe an element $(p,q)$ satisfying $\Pi_{\cal K}^{''}(x-y;d-w,p-q)=p$.
For simplicity, we denote by
$
z:=x-y,\xi:=d-w $ and $ \eta:=p-q.$
\medskip
\noindent
{\bf Case (i)} $x\in {\rm int}{\cal K}$ and $y=0$. Since $z=x-y\in {\rm int}{\cal K}$, by Theorem \ref{projection-second-order diff}(i), we have $\Pi_{{\cal K}}^{''}(x-y;d-w,p-q)=p-q$.
It follows that
$$\Pi_{\cal K}^{''}(x-y;d-w,p-q)=p\Longleftrightarrow q=0.$$
Hence $ T^2_\Omega\big((x,y);(d,w)\big)=\Re^m \times \{0\}. $
\medskip
\noindent
{\bf Case (ii)} $x=0$ and $y\in {\rm int}{\cal K}$. Since $z=x-y\in -{\rm int}{\cal K}$, by Theorem \ref{projection-second-order diff}(ii), we know
$
\Pi^{''}_{{\cal K}}(z;d-w,p-q)=0.$
It follows that
$$\Pi_{\cal K}^{''}(x-y;d-w,p-q)=p\Longleftrightarrow p=0.$$
Hence $ T^2_\Omega\big((x,y);(d,w)\big)=\{0\} \times \Re^m. $
\medskip
\noindent
{\bf Case (iii)} $x,y\in {\rm bd}{\cal K}\backslash \{0\}$ and $x^Ty=0$. In this case $x_1=\|x_2\|\not =0$ and by Lemma \ref{lem2.1},
\begin{equation}
z=x-y=(x_1,x_2)-k(x_1,-x_2)=((1-k)x_1,(1+k)x_2), \ \ k=y_1/x_1. \label{case3}\end{equation}
This yields $z_1+\|z_2\|=2x_1>0$ and $z_1-\|z_2\|=-2kx_1<0$, i.e., $z\in ({\cal K}\cup {\cal K}^\circ)^c$.
Then by
Theorem \ref{projection-second-order diff}(vi), $\Pi_{{\cal K}}^{''}(z;\xi,\eta)=p$ where $p=(p_1,p_2)\in \mathbb{R}\times \mathbb{R}^{m-1}$ if and only if
\begin{eqnarray}
p_1 &=& \frac{1}{2}\left(\eta_1+\bar{z}_2^T\eta_2
+\frac{\|\xi_2\|^2-(\bar{z}_2^T\xi_2)^2}{\|z_2\|}\right), \label{formula-case-3-1}
\\
p_2&=&\frac{1}{2}\left(\eta_1-\frac{z_1}{\|z_2\|}\bar{z}_2^T\eta_2
-\frac{z_1}{\|z_2\|^2}
\Big[\|\xi_2\|^2-3(\bar{z}_2^T\xi_2)^2\Big]
-2\xi_1\frac{\bar{z}_2^T\xi_2}{\|z_2\|}\right)\bar{z}_2\nonumber\\
&& +\frac{\|z_2\|\xi_1-z_1\bar{z}_2^T\xi_2}{\|z_2\|^2}\xi_2
+\frac{1}{2}\left(1+\frac{z_1}{\|z_2\|}\right)\eta_2.\label{formula-case-3-2}
\end{eqnarray}
We now try to derive an equivalent expression for (\ref{formula-case-3-1}) and (\ref{formula-case-3-2}). Since $(d,w)\in T_{\Omega}(x,y)$, according to Lemma \ref{tangentcone},
$ x \perp w$, $y\perp d$ and there exists $\beta\in \Re$ such that $x_1\hat{w}-y_1d=\beta x$, from which and $x_1=\|x_2\|\not =0$ we have
\begin{equation} \label{3-6}
w_1=kd_1+\beta, \quad w_2=-kd_2-\beta\bar{x}_2,
\end{equation}
and
\begin{equation}
\bar{x}_2^Tw_2=-w_1, \quad \bar{x}_2^Td_2=d_1.\label{eqn3-12}
\end{equation}
Note that $\bar{z}_2=\bar{x}_2$ by (\ref{case3}). Hence it follows from (\ref{3-6}) and (\ref{eqn3-12}) that
\begin{eqnarray}
&& \bar{z}_2^T\xi_2 = \bar{x}_2^T(d_2-w_2)=d_1+w_1 = (1+k)d_1 + \beta,\label{eqn3-13}\\
&&\|\xi_2\|^2=\|d_2-w_2\|^2
= \|(1+k)d_2+\beta \bar{x}_2\|^2
= (k+1)^2\|d_2\|^2+2\beta(k+1)d_1 + \beta^2,\nonumber \\
&& \xi_1 = d_1-w_1 = (1-k)d_1-\beta. \label{eqn3-19}
\end{eqnarray}
Hence (\ref{formula-case-3-1}) can be rewritten as
\begin{equation} \label{3-3old}
p_1=-q_1+\bar{x}_2^T(p_2-q_2)+\frac{x_1+y_1}{x_1^2}\Big(\|d_2\|^2-d^2_1\Big). \end{equation}
The term in front of $\bar{z}_2$ in (\ref{formula-case-3-2}) becomes
\begin{eqnarray*}
&&\frac{1}{2}\left(\eta_1-\frac{z_1}{\|z_2\|}\bar{z}_2^T\eta_2-\frac{z_1}{\|z_2\|^2}
\Big[\|\xi_2\|^2-3(\bar{z}_2^T\xi_2)^2\Big]
-2\xi_1\frac{\bar{z}_2^T\xi_2}{\|z_2\|}\right)\\
&= & \frac{1}{2}\left(\eta_1+\bar{z}_2^T\eta_2
+ \frac{\|\xi_2\|^2-(\bar{z}_2^T\xi_2)^2}{\|z_2\|}\right)
-\frac{z_1+\|z_2\|}{2\|z_2\|}\bar{z}_2^T\eta_2\\
&& +\frac{z_1+\|z_2\|}{2\|z_2\|^2}\Big[(\bar{z}_2^T\xi_2)^2-\|\xi_2\|^2\Big] +\left[\frac{z_1}{\|z_2\|^2}\big(\bar{z}_2^T\xi_2\big)^2
-\frac{\xi_1}{\|z_2\|}\big(\bar{z}_2^T\xi_2\big)\right] \\
&=& \frac{y_1p_1-x_1q_1}{x_1+y_1}+\left[\frac{x_1-y_1}{(x_1+y_1)^2}(d_1+w_1)^2
-\frac{1}{x_1+y_1}(d_1^2-w_1^2) \right] \\
&=& \frac{y_1p_1-x_1q_1}{x_1+y_1}
+2\frac{x_1w_1^2+x_1d_1w_1-y_1d_1^2-y_1d_1w_1}{(x_1+y_1)^2},
\end{eqnarray*}
where the second equality uses (\ref{case3}), (\ref{formula-case-3-1}), and (\ref{eqn3-13})-(\ref{3-3old}).
It follows from (\ref{eqn3-13}) and (\ref{eqn3-19}) that the term in front of $\xi_2$ in (\ref{formula-case-3-2}) is
\[
\frac{\xi_1}{\|z_2\|}- \frac{z_1}{\|z_2\|}\frac{\bar{z}_2^T\xi_2}{\|z_2\|}
= \frac{1}{x_1+y_1}\big(d_1-w_1\big)-\frac{x_1-y_1}{(x_1+y_1)^2}\big(d_1+w_1\big)
= 2\frac{y_1d_1-x_1w_1}{(x_1+y_1)^2}.
\]
The term in front of $\eta_2$ in (\ref{formula-case-3-2}) is
$
1/2\big(1+(z_1/\|z_2\|)\big)=x_1/(x_1+y_1).
$
Hence (\ref{formula-case-3-2}) can be rewritten as
\begin{eqnarray}
p_2 &= & \left(\frac{y_1p_1-x_1q_1}{x_1+y_1}
+2\frac{x_1w_1^2+x_1d_1w_1-y_1d_1^2-y_1d_1w_1}{(x_1+y_1)^2}\right)\bar{x}_2+2\frac{y_1d_1-x_1w_1}{(x_1+y_1)^2} (d_2-w_2)\nonumber\\
&&
+\frac{x_1}{x_1+y_1}(p_2-q_2).\label{eqn4-3-2}
\end{eqnarray}
Further notice that
$
(y_1p_1-x_1q_1)\bar{x}_2=-p_1y_2-q_1x_2
$
and
\begin{eqnarray*}
\lefteqn{2 \frac{y_1d_1-x_1w_1}{(x_1+y_1)^2}(d_2-w_2)
+2 \frac{x_1w_1^2+x_1d_1w_1-y_1d_1^2-y_1d_1w_1}{(x_1+y_1)^2}\bar{x}_2 } \nonumber\\
&=& 2 \frac{-x_1\beta}{(x_1+y_1)^2}[(1+k)d_2+\beta\bar{x}_2]
+2 \frac{x_1\beta^2+y_1d_1\beta+x_1d_1\beta}{(x_1+y_1)^2}\bar{x}_2 \nonumber \\
&=& 2\frac{\beta}{x_1+y_1}\big(-d_2+d_1\bar{x}_2\big)
= 2\frac{x_1w_1-y_1d_1}{x_1(x_1+y_1)}\big(-d_2+d_1\bar{x}_2\big),
\end{eqnarray*}
where the first and third equations use (\ref{3-6}) and the fact $y_1=kx_1$. Hence (\ref{eqn4-3-2}) can be rewritten as
\begin{eqnarray}
y_1p_2+x_1q_2&=&2 \frac{x_1w_1-y_1d_1}{x_1}(-d_2+d_1\bar{x}_2)-p_1y_2-q_1x_2 \nonumber\\
&=& (x_1w_1-y_1d_1)\left(\frac{w_2-w_1\bar{y}_2}{y_1}+\frac{-d_2+d_1\bar{x}_2}{x_1}\right)-p_1y_2-q_1x_2, \label{eqn4-3-1}
\end{eqnarray}
where the second step comes from the fact
$ (-d_2+d_1\bar{x}_2)/x_1=(w_2-w_1\bar{y}_2)/y_1$ due to (\ref{3-6}).
Hence (\ref{formula-case-3-1}) and (\ref{formula-case-3-2}) is equivalent to
(\ref{3-3old}) and (\ref{eqn4-3-1}).
Now, multiplying (\ref{eqn4-3-1}) by $\bar{x}_2^T$ and using (\ref{eqn3-12}) yields
\begin{equation} \label{eqn4-3-4}
x_1\bar{x}_2^Tq_2+y_1\bar{x}_2^Tp_2= y_1p_1-x_1q_1.
\end{equation}
Hence it follows from (\ref{3-3old}) and (\ref{eqn4-3-4}) that
\[
p_1= \left(1+\frac{y_1}{x_1}\right)\bar{x}_2^Tp_2-\frac{y_1}{x_1}p_1+\frac{x_1+y_1}{x_1^2}\big(\|d_2\|^2-d^2_1\big),
\]
i.e.,
\begin{equation}\label{eqn4-3-5} p_1=\bar{x}_2^Tp_2+\frac{1}{x_1}\big(\|d_2\|^2-d_1^2\big) \Longleftrightarrow p\in {\rm bd}T_{\cal K}^2(x; d).
\end{equation}
Since $d_1=(w_1-\beta)/k$, $d_2=-(w_2+\beta\bar{x}_2)/k$
by (\ref{3-6}), and $w_2^T\bar{x}_2=-w_1$ by (\ref{eqn3-12}), we see
\begin{equation}\label{case-3-a}
\frac{\|d_2\|^2-d_1^2}{x_1^2}
=\frac{\|w_2\|^2-w_1^2}{y_1^2}.
\end{equation}
Similarly, it follows from (\ref{3-3old}), (\ref{eqn4-3-4}), (\ref{case-3-a}), and $\bar{x}_2=-\bar{y}_2$ that
\begin{eqnarray*}
q_1=
-\frac{x_1}{y_1}q_1+\left(1+\frac{x_1}{y_1}\right)\bar{y}_2^Tq_2+\frac{x_1+y_1}{y_1^2}
\big(\|w_2\|^2-w_1^2\big),
\end{eqnarray*}
i.e.,
\begin{equation}\label{eqn4-3-6}
\frac{1}{y_1}\big(\|w_2\|^2-w_1^2\big)=q_1-\bar{y}_2^Tq_2 \Longleftrightarrow q\in {\rm bd} T_{\cal K}^2(y; w).
\end{equation}
Hence along the line
\[
\left\{\begin{array}{l} (\ref{formula-case-3-1})\\
(\ref{formula-case-3-2}) \end{array}\right. \Longleftrightarrow \left\{\begin{array}{l} (\ref{3-3old})\\
(\ref{eqn4-3-1}) \end{array}\right.\Longleftrightarrow \left\{\begin{array}{l} (\ref{eqn4-3-5}), (\ref{eqn4-3-6})\\
(\ref{eqn4-3-1}) \end{array}\right.
\]
the desired result follows.
\medskip
\noindent
{\bf Case (iv)} $x\in {\rm bd}{\cal K}\backslash \{0\}$ and $y=0$. In this case
$z=x-y=x\in {\rm bd}{\cal K}\backslash \{0\}$.
\medskip
$ {\bf (iv)}$-1. $d\in {\rm int} T_{{\cal K}}(x)$ and $w=0$.
Then $\xi=d-w=d\in {\rm int}T_{{\cal K}}(x)$. Hence $\Pi_{{\cal K}}^{''}(z;\xi,\eta)=\eta$ by Theorem \ref{projection-second-order diff}(iv). It follows that $\Pi_{{\cal K}}^{''}(x-y;d-w,p-q)=p$ if and only if $q=0$.
\medskip
$ {\bf (iv)}$-2. $d\in {\rm bd} T_{{\cal K}}(x)$ and $w=0$. Then
$\xi=d\in {\rm bd} T_{{\cal K}}(x)$.
Hence $\Pi^{''}_{{\cal K}}(z;\xi,\eta)=\Pi_{{\cal K}}^{''}(x;d,\eta)$ and by Proposition \ref{second-order diff}(iv)
\begin{eqnarray*}
\lefteqn{\Pi_{{\cal K}}^{''}(x;d,\eta)}\\
&=& \left\{\begin{array}{ll}
\eta \ \ \ \ \ \ \ \ \qquad & \mbox{if } \eta \in T^2_{{\cal K}}(x;d), \\
\frac{1}{2}\left(\begin{matrix}
\eta_1+\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\\
\left(\eta_1-\bar{x}_2^T\eta_2-\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right)\bar{x}_2+2\eta_2
\end{matrix}\right) & \mbox{if } \eta \notin T^2_{{\cal K}}(x;d).
\end{array}\right.
\end{eqnarray*}
Note that
$\eta \in T^2_{{\cal K}}(x;d)\Longleftrightarrow\eta_2^Tx_2-\eta_1x_1\leq d_1^2-\|d_2\|^2$ by Lemma \ref{Lem3.3}.
Hence $\Pi_{{\cal K}}^{''}(x;d,p-q)=p$ if and only if either $p-q\in T^2_{{\cal K}}(x;d)$ and $q=0$ or the following system holds
\begin{eqnarray}\label{Case-4-Part-2}
\left\{ \begin{array}{ll}
\eta_1-\bar{x}_2^T\eta_2-\frac{\|d_2\|^2-d_1^2}{\|x_2\|} < 0, \\
\frac{1}{2}\left(\eta_1+\left[\bar{x}_2^T\eta_2
+ \frac{\|d_2\|^2-d_1^2}{\|x_2\|} \right] \right)= p_1, \\
\frac{1}{2}\left(\eta_1-\left[\bar{x}_2^T\eta_2
+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right]\right)\bar{x}_2
+\eta_2=p_2. \end{array}\right.
\end{eqnarray}
We now further simplify the system (\ref{Case-4-Part-2}).
\begin{eqnarray*}
(\ref{Case-4-Part-2}) & \Longleftrightarrow& \left\{ \begin{array}{ll}
\eta_1-\left[\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right]< 0 ,\\ \bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} = p_1 + q_1, \\
\frac{1}{2}\big(\eta_1-p_1-q_1\big)\bar{x}_2
=q_2. \end{array}\right.
\Longleftrightarrow
\left\{ \begin{array}{ll}
\eta_1-\left[\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right]< 0, \\
\bar{x}_2^Tp_2-\bar{x}_2^Tq_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} = p_1 + q_1, \\
-q_1\bar{x}_2 = q_2.
\end{array}\right.\\
& \Longleftrightarrow&
\left\{ \begin{array}{ll}
\eta_1-\left[\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right]< 0,\\
\bar{x}_2^Tp_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} = p_1, \\
-q_1\bar{x}_2 = q_2.
\end{array}\right.
\Longleftrightarrow
\left\{ \begin{array}{ll}
q_1> 0,\\
\bar{x}_2^Tp_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} = p_1, \\
-q_1\bar{x}_2 = q_2.
\end{array}\right.\\
&\Longleftrightarrow&
\left\{ \begin{array}{ll}
q\in \Re_{++}\hat{x},\\
p\in {\rm bd}T_{\cal K}^2(x;d).
\end{array}\right.
\end{eqnarray*}
Hence either $p\in T^2_{{\cal K}}(x;d)$ and $q=0$ or $ q\in \Re_{++}\hat{x}$ and $p\in {\rm bd}T_{\cal K}^2(x;d)$.
\noindent
\medskip
$ {\bf (iv)}$-3. $d\perp \hat{x}$ and $w\in \Re_{++}\hat{x}$.
Then
$ \bar{x}_2^Td_2 = d_1$ and
$\bar{x}_2^Tw_2= -w_1$.
Hence
\begin{equation}\label{eqn09}
\xi_1-\bar{x}_2^T\xi_2= d_1-w_1-\bar{x}_2^T(d_2-w_2)=-2w_1<0,
\end{equation} which implies $\xi\in T_{{\cal K}}(x)^c$ by Lemma \ref{Lem3.3}. Thus by Theorem \ref{projection-second-order diff}(iv),
$\Pi^{''}_{{\cal K}}(z;\xi,\eta)=p$ takes the form
\begin{equation}\label{add-4-1}
\left\{\begin{array}{ll} & \frac{1}{2}\left(\eta_1+\bar{x}_2^T\eta_2
+\frac{\|\xi_2\|^2-(\bar{x}_2^T\xi_2)^2}{\|x_2\|} \right)=p_1,\\
&\frac{1}{2}\left(\eta_1-\bar{x}_2^T\eta_2-\frac{\|\xi_2\|^2-3(\bar{x}_2^T\xi_2)^2
+2\xi_1\bar{x}_2^T\xi_2}{\|x_2\|}\right)\bar{x}_2
+\eta_2+\frac{\xi_1-\bar{x}_2^T\xi_2}{\|x_2\|}\xi_2=p_2.
\end{array}\right.\end{equation}
Note that
\begin{eqnarray*}
\xi_1\bar{x}_2^T\xi_2-(\bar{x}_2^T\xi_2)^2&=& (\xi_1-\bar{x}_2^T\xi_2)\bar{x}_2^T\xi_2=-2w_1(d_1+w_1), \\
\|\xi_2\|^2-(\bar{x}_2^T\xi_2)^2& =&
\|d_2-w_2\|^2-(d_1+w_1)^2
=\|d_2\|^2-d_1^2,
\end{eqnarray*}
where we have used the fact $d\perp w$ and $w_2=-w_1\bar{x}_2$ due to $d\perp \hat{x}$ and $w\in \Re_{++}\hat{x}$.
Therefore
\begin{eqnarray}
\frac{\|\xi_2\|^2-3(\bar{x}_2^T\xi_2)^2
+2\xi_1\bar{x}_2^T\xi_2}{\|x_2\|}& =&
\frac{\|\xi_2\|^2-(\bar{x}_2^T\xi_2)^2}{\|x_2\|}+2\frac{\xi_1\bar{x}_2^T\xi_2-(\bar{x}_2^T\xi_2)^2
}{\|x_2\|}\nonumber\\
&=& \frac{\|d_2\|^2-d_1^2}{\|x_2\|}-4\frac{w_1(d_1+w_1)}{\|x_2\|}.\label{eqn4-4-2}
\end{eqnarray}
Putting (\ref{eqn09})-(\ref{eqn4-4-2}) into (\ref{add-4-1}) yields
\begin{eqnarray*}
&&\Pi_{{\cal K}}^{''}(z;\xi,\eta)=p\\
&\Longleftrightarrow& \left\{\begin{array}{ll} \label{4-2newnnn}
& \frac{1}{2}\left(\eta_1+\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} \right)=p_1 ,\\
& \left(\frac{1}{2}\eta_1-\frac{1}{2}\left[\bar{x}_2^T\eta_2+
\frac{\|d_2\|^2-d_1^2}{\|x_2\|}\right]+2\frac{w_1}{\|x_2\|}(d_1+w_1)\right)\bar{x}_2
-2\frac{w_1}{\|x_2\|}(d_2-w_2)=q_2.
\end{array}\right. \\
&\Longleftrightarrow &
\left\{\begin{array}{ll} \label{4-2new}
& \frac{1}{2}\left(\eta_1+\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} \right)=p_1, \\
& \left(-q_1+2\frac{w_1}{\|x_2\|}(d_1+w_1)\right)\bar{x}_2
-2\frac{w_1}{\|x_2\|}(d_2-w_2)=q_2.
\end{array}\right. \\
&\Longleftrightarrow &
\left\{\begin{array}{ll} \label{4-2newn}
& \frac{1}{2}\left(\eta_1+\bar{x}_2^T\eta_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|} \right)=p_1, \\
& -q_1\bar{x}_2-2\frac{d_1w_2}{\|x_2\|}-2\frac{w_1d_2}{\|x_2\|}=q_2.
\end{array}\right.\\
&\Longleftrightarrow & \left\{\begin{array}{ll} \label{4-2newnn}
& \bar{x}_2^Tp_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}=p_1, \\
& -q_1\bar{x}_2-2\frac{w_1d_2}{\|x_2\|}-2\frac{d_1w_2}{\|x_2\|}=q_2.
\end{array}\right.
\end{eqnarray*}
where the third equivalence uses the fact $w_2=-w_1\bar{x}_2$ due to $w\in \Re_{++}\hat{x}$ and the last step follows from substituting the expression for $q_2$ in the second equation into the first one to obtain
\[
\bar{x}_2^Tq_2=\bar{x}_2^T\left(-q_1\bar{x}_2-2\frac{w_1d_2}{\|x_2\|}-2\frac{d_1w_2}{\|x_2\|} \right)=-q_1.
\]
The desired result follows from noting that $p\in {\rm bd}T_{\cal K}^2(x;d)$ if and only if $ \bar{x}_2^Tp_2+\frac{\|d_2\|^2-d_1^2}{\|x_2\|}=p_1$ by virtue of Lemma \ref{Lem3.3}.
\medskip
\noindent
{\bf Case (v)} $x=0$ and $y\in {\rm bd}{\cal K}\backslash \{0\}$.
The proof is omitted, since this case is symmetric to Case (iv).
\medskip
\noindent
{\bf Case (vi)} $x=0$ and $y=0$. Since
$\Omega$ is cone, according to the definition of second-order tangent
set, we have
\[
T^2_{\Omega}\big((0,0);(d,w)\big)=T_{\Omega}(d,w).
\]
From all the above, the proof is complete.
\hskip .5cm $\Box$ \vskip .5cm
\section{Second-order optimality conditions for SOCMPCC}
In this section, as an application of the second-order tangent set for the SOC complementarity set, we consider second-order optimality conditions for the mathematical programming with second-order cone complementarity constraints
(SOCMPCC):
\begin{equation} \label{SOCMPCC}
\min \ f(x)
\quad {\rm s.t.} \ \ {\cal K} \ni G(x) \perp H(x) \in {\cal K},
\end{equation}
where $f:\Re^n\to \Re$ and $G,H:\Re^n\to \Re^m$ are second-order continuously differentiable.
For simplicity, we restrict our attention on the simpler case, i.e., ${\cal K}$ is a $m$-dimensional second-order cone. All analysis
can be easily carried over to more general cases where ${\cal K}$ is a Cartesian product of some second-order cones. SOCMPCC is an important class of optimization problems that has many applications. We refer the reader to \cite{YZ15,YZ16} and the reference within for applications and the first-order necessary optimality conditions.
Denote by $F(x):=(G(x),H(x))$. Then SOCMPCC (\ref{SOCMPCC}) can be rewritten as
\begin{equation} \label{SOCMPCCre}
\min \ f(x)
\quad {\rm s.t.} \ \ F(x) \in \Omega.
\end{equation}
For a convex set-constrained optimization problem in the form of (\ref{SOCMPCCre}) where $\Omega$ is replaced by a convex closed set $K$ (see \cite[(3.93)]{BS}), second-order optimality conditions that involve the second-order tangent set to $K$ have been developed in \cite{BCS99, BS}. In particular when the convex set $K$ is not polyhedral, the second-order tangent set to $K$ is needed in the second-order optimality conditions. However, {if set $\Omega$ in problem (\ref{SOCMPCCre}) is} nonconvex, these optimality conditions are not applicable in general. In what follows, we will establish the second-order optimality conditions
for the SOCMPCC, which is not a convex set-constrained optimization problem.
We would like to emphasize that, even if the second-order cone complementarity set is nonconvex, its tangent cone and second-order tangent set have nice properties so that some of the theories in the second-order optimality conditions for a convex set-constrained optimization problem still hold.
This observation relies heavily on the exact formula of tangent cone and second-order tangent set established in the previous section.
First we present some results needed for further analysis.
Recall that the regular tangent cone is always convex. The following result shows that the regular tangent cone to the SOC complementarity set $\Omega$ is not only convex but is a subspace.
\begin{proposition}\label{frechet-tangent-cone}
For any $(x,y)\in \Omega$,
\begin{eqnarray*}
\lefteqn{\widehat{T}_\Omega(x,y)={\rm lin} T_{\Omega}(x,y)}\\
&=&
\left\{(d,w)\left|
\begin{array}{lll}
& d\in \Re^m,\ w=0, \ & {\rm if}\ x\in {\rm int}{\cal K} \ {\rm and}\ y=0; \\
& d=0,\ w\in \Re^m, \ & {\rm if} \ x=0\ {\rm and}\ y\in {\rm int}{\cal K}; \\
& x_1\hat{w}-y_1d\in \Re x, \ \ d\perp y,\ w\perp x, \ \
& {\rm if} \ x, y\in {\rm bd}{\cal K}\backslash\{0\}; \\
& d\perp \hat{x},\ w=0, \ \ & {\rm if}\ x\in {\rm bd}{\cal K}\backslash \{0\}\ {\rm and}\ y=0; \\
& d=0, \ w\perp \hat{y}, \ \ & {\rm if}\ x=0\ {\rm and}\ y\in {\rm bd}{\cal K}\backslash \{0\}; \\
& d=0,\ w=0, \ \ & {\rm if}\ x=0 \ {\rm and} \ y=0.
\end{array}\right.\right\}.
\end{eqnarray*}
\end{proposition}
\par\noindent {\bf Proof.}\ \
{The formula of ${\rm lin} T_{\Omega}(x,y)$ is clear from that of $T_{\Omega}(x,y)$ in Lemma \ref{tangentcone}. According to the tangent-normal polarity as in Lemma \ref{polarity}, we can obtain the formula of $\widehat{T}_\Omega(x,y)$ by taking the polar of the limiting normal cone to $\Omega$ given in \cite[Theorem 5.1]{yzhou}. The obtained formula of ${\rm lin} T_{\Omega}(x,y)$ and $\widehat{T}_\Omega(x,y)$ shows that they have the same expression.} \hskip .5cm $\Box$ \vskip .5cm
{The exact formula established in Theorem \ref{formula-regular normal cone-1} and Proposition \ref{frechet-tangent-cone} immediately imply the following results.}
\begin{corollary}\label{cor-lin-tangent}
For all $(x,y)\in \Omega$,
$
T_{\Omega}(x,y)+\widehat{T}_{\Omega}(x,y)=T_{\Omega}(x,y).
$
\end{corollary}
\begin{proposition}\label{second-order-tangent-cone-add}
For $(x,y)\in \Omega$ and $(d,w)\in T_{\Omega}(x,y)$,
\[
T^2_{\Omega}\big((x,y);(d,w)\big) + \widehat{T}_{\Omega}(x,y)
= T^2_{\Omega}\big((x,y);(d,w)\big).
\]
\end{proposition}
\par\noindent {\bf Proof.}\ \
{The inclusion ``$\supseteq$" is clear, since $(0,0)\in \widehat{T}_{\Omega}(x,y)$.
For all cases except where $x, y \in {\rm bd}{\cal K}\backslash\{0\}$, it is easy to see that ``$\subseteq$" can be achieved by using the formula of $T^2_{\Omega}\big((x,y);(d,w)\big)$ given in Theorem \ref{formula-regular normal cone-1} and the formula of $\widehat{T}_\Omega(x,y)$ given in Proposition \ref{frechet-tangent-cone}. Now consider the case where $x, y \in {\rm bd}{\cal K}\backslash\{0\}$.
Let $(p,q)\in T^2_{\Omega}\big((x,y);(d,w)\big)$ and $(u,v)\in \widehat{T}_{\Omega}(x,y)$. Since $p\in {\rm bd}T_{{\cal K}}^2(x;d)$, then $\hat{x}^Tp= \|d_2\|^2-d_1^2$
by Lemma \ref{Lem3.3}. Hence $\hat{x}^T(p+u)=\hat{x}^Tp+\hat{x}^Tu= \|d_2\|^2-d_1^2$ due to the fact $u\perp \hat{x}$ (since $u\perp y$ by Proposition \ref{frechet-tangent-cone} and $y\in \Re_{++}\hat{x}$ ). This means $p+u\in {\rm bd}T_{{\cal K}}^2(x;d)$. Similarly, we can obtain $q+v\in {\rm bd}T_{{\cal K}}^2(y;w)$. Since $(u,v)\in \widehat{T}_{\Omega}(x,y)$, it follows from Proposition \ref{frechet-tangent-cone} that there exists $\tau\in \Re$ such that $x_1\hat{v}-y_1u=\tau x$. Thus
\begin{equation}\label{re-eq-a11a}
x_1v_2+y_1u_2=-\tau x_2=-\frac{x_1v_1-y_1u_1}{x_1}x_2=
-v_1x_2-u_1y_2,
\end{equation}
where the last step comes from Lemma \ref{lem2.1}. Since $x, y \in {\rm bd}{\cal K}\backslash\{0\}$, $(p,q)\in T^2_{\Omega}\big((x,y);(d,w)\big)$, it follows from Theorem \ref{formula-regular normal cone-1} that
\begin{equation}
(p,q)\in T^2_{\Omega}\big((x,y);(d,w)\big) \Longleftrightarrow \left\{ \begin{array}{l}
p\in {\rm bd}T_{{\cal K}}^2(x;d), q \in {\rm bd}T_{{\cal K}}^2(y;w) \\
\xi-p_1y_2-q_1x_2=x_1q_2+y_1p_2\end{array} \right. , \label{xi}
\end{equation}
where $ \xi:= (x_1w_1-y_1d_1)\left(\frac{{w_2-w_1\bar{y}_2}}{y_1}-\frac{d_2-d_1\bar{x}_2}{x_1}\right).
$ This, together with (\ref{re-eq-a11a}), implies
\begin{eqnarray*}
x_1(q_2+v_2)+y_1(p_2+u_2)&=& x_1q_2+y_1p_2-v_1x_2-u_1y_2\\
&=& \xi-p_1y_2-q_1x_2-v_1x_2-u_1y_2\\
&=& \xi-(p_1+u_1)y_2-(q_1+v_1)x_2.
\end{eqnarray*}
Hence together with $p+u\in {\rm bd}T_{{\cal K}}^2(x;d)$ and $q+v\in {\rm bd}T_{{\cal K}}^2(y;w)$, we have that $(p+u,q+v)\in T^2_{\Omega}\big((x,y);(d,w)\big)$ by virtue of (\ref{xi}).
\hskip .5cm $\Box$ \vskip .5cm
With these preparations, we are now ready to develop a second-order necessary optimality condition for SOCMPCCs. Define the Lagrange function as
$L(x,\lambda):= f(x)+\langle F(x), \lambda\rangle$ and the following three multiplier sets
\begin{eqnarray*}
\Lambda^c(x^*) &:=& \{\lambda \, | \, \nabla_x L(x^*,\lambda)=0,
\ \lambda\in N^c_{\Omega}(F(x^*))\} , \\
\Lambda (x^*) &:=& \{\lambda \, | \, \nabla_x L(x^*,\lambda)=0,
\ \lambda\in {N}_{\Omega}(F(x^*))\},\\
\Lambda^F(x^*) &:=& \{\lambda \, | \, \nabla_x L(x^*,\lambda)=0,
\ \lambda\in \widehat{N}_{\Omega}(F(x^*))\}.
\end{eqnarray*}
Denote by $
C(x^*):=\big\{d|\, \nabla f(x^*)d\leq 0, \nabla F(x^*)d\in T_{\Omega}(F(x^*)) \big\}
$ the critical cone.
Note that if there exists $\lambda\in \Lambda^F({x^*})$, then $C(x^*)=\big\{d|\, \nabla f(x^*)d= 0, \nabla F(x^*)d\in T_{\Omega}(F(x^*)) \big\}$.
\begin{theorem}\label{dual-conditions}
Let $x^*$ be a locally optimal solution of SOCMPCC.
Suppose that the non-degeneracy condition
\begin{equation}\label{conditions-revision}
{\nabla} F(x^*)\Re^n+{\rm lin}T_{\Omega}(F(x^*)) = \Re^{2m}
\end{equation} holds. Then
$\Lambda^c(x^*) =\Lambda(x^*)=\Lambda^F(x^*)=\{\lambda_0\}$ and
\[
\nabla^2_{xx}L(x^*,\lambda_0)(d,d) - \sigma \left( \lambda_0| T^2_{\Omega}
(F(x^*);\nabla F(x^*)d) \right) \geq 0, \quad
\forall d\in C(x^*).
\]
\end{theorem}
\par\noindent {\bf Proof.}\ \
\noindent
Step 1. We prove $\Lambda^c(x^*) =\Lambda(x^*)=\Lambda^F(x^*)=\{\lambda_0\}$.
Since
$
\nabla F(x^*)\Re^n+{\rm lin}T_{\Omega}(F(x^*))=
\nabla F(x^*)\Re^n+\widehat{T}_{\Omega}(F(x^*))=\Re^{2m},$ taking polars on the both sides of the above equation, by the rule for polar cones \cite[Corollary 11.25]{RW98} and the fact that
$(\widehat{T}_{\Omega})^{\circ}=N^c_{\Omega}$,
we have
\begin{equation} \label{clarke-normal}
\nabla F(x^*)^T\lambda=0, \ \lambda\in N^c_{\Omega}(F(x^*))
\quad \Longrightarrow \quad \lambda = 0.
\end{equation}
Suppose that $\lambda^1, \lambda^2\in \Lambda^c(x^*)$.
Then $\lambda^1-\lambda^2$ satisfies $\nabla F(x^*)^T(\lambda^1-\lambda^2)=0$
and $\lambda^1-\lambda^2\in N^c_{\Omega}(F(x^*))$ since $N^c_{\Omega}(F(x))$
is subspace (because $\widehat{T}_{\Omega}(x,y)$ is a subspace and
$N^c_{\Omega}=(\widehat{T}_{\Omega})^{\circ}$). Thus
$\lambda^1=\lambda^2$ by (\ref{clarke-normal}). This means that
$\Lambda^c(x^*)$ is a singleton. Since
$\Lambda^F(x^*) \subseteq \Lambda(x^*) \subseteq \Lambda^c(x^*)$, it remains to
show that $\Lambda^F(x^*)$ is nonempty. Since
$N_{\Omega}\subseteq N^c_{\Omega}$, the condition
(\ref{clarke-normal}) ensures
\begin{equation}
\nabla F(x^*)^T\lambda=0, \ \lambda\in N_{\Omega}(F(x^*))
\quad \Longrightarrow \quad \lambda=0,\label{MR}
\end{equation}
which in turn implies that the system $F(x)-\Omega$ is metrically regular
at $(x^*,0)$. Thus according to \cite[Theorem 4]{GO16}, Proposition \ref{frechet-tangent-cone}, and Corollary \ref{cor-lin-tangent}, we have
$\widehat{N}_S(x)=\nabla F(x)^T\widehat{N}_{\Omega}(F(x))$, where
$S:=\{x \, | \, F(x)\in \Omega\}$. As $x^*$ is a local optimal
solution of problem (\ref{SOCMPCCre}), we have
\(
0\in \nabla f(x^*)^T+\widehat{N}_{S}(x^*)
=\nabla f(x^*)^T+\nabla F(x^*)^T\widehat{N}_{\Omega}(F(x^*))
\),
which indicates that $\Lambda^F(x^*)$ is nonempty. Hence $\Lambda^c(x^*), \Lambda(x^*),\Lambda^F(x^*)$
are all singleton and coincide with each other. Let us denote the unique element by $\lambda_0$.
\medskip
\noindent
Step 2. We show that for all $d\in C(x^*)$ and for any convex subset
${\cal T}(d)$ in $T^2_{\Omega}(F(x^*);\nabla F(x^*)d)$,
$
\nabla^2_{xx}L(x^*,\lambda_0)(d,d)-\sigma \left( \lambda_0|{\cal T}(d) \right)\geq 0.
$
{The idea of the proof is inspired by the arguments in \cite[Theorem 3.1]{BCS99} and using the properties of tangent cone and second-order tangent set discussed above.} {For the sake of completeness, we give the detailed proof here.}
Consider the set
$\Gamma(d):={\rm cl}\{{\cal T}(d)+\widehat{T}_{\Omega}(F(x^*))\}$. {Since the regular tangent cone is convex, the set $\Gamma(d)$} is
closed and convex. Moreover, it follows from Proposition \ref{second-order-tangent-cone-add}
and the fact that the second-order tangent set is closed that $\Gamma(d)\subseteq T^2_{\Omega}(F(x^*);\nabla F(x^*)d)$.
Because $x^*$ is locally optimal of problem (\ref{SOCMPCCre}), {by definition of the second-order tangent cone, we can show that
$$\nabla f(x^*)w +\nabla^2f(x^*)(d,d) \geq 0, \quad \forall d\in C(x^*), w\in T^2_C\big(x^*;d
\big),$$
where $C$ denotes the feasible region of problem (\ref{SOCMPCCre}). Since (\ref{MR}) holds, by \cite[Proposition 13.13]{RW98}, the chain rule for tangent sets (\ref{chainruletangentset}) holds with $\Theta$ taken as $\Omega$.} It follows that for all $d\in C(x^*)$,
the following optimization problem
\[
\left\{
\begin{array}{ll}
\displaystyle \min_w & \nabla f(x^*)w+\nabla^2f(x^*)(d,d) \\
{\rm s.t.} & \nabla F(x^*)w + \nabla^2F(x^*)(d,d)
\in T^2_\Omega\big(F(x^*);\nabla F(x^*)d \big)
\end{array}
\right.
\]
has nonnegative optimal value.
Since $\Gamma(d)\subseteq T^2_{\Omega}(F(x^*);\nabla F(x^*)d)$, it
is clear that the following convex set constrained problem
\begin{equation} \label{convex-subproblem}
\left\{
\begin{array}{ll}
\displaystyle \min_w & \nabla f(x^*)w+\nabla^2f(x^*)(d,d) \\
{\rm s.t.} & \nabla F(x^*)w + \nabla^2F(x^*)(d,d)
\in \Gamma(d)
\end{array}
\right.
\end{equation}
has nonnegative optimal value as well. Since the optimization problem (\ref{convex-subproblem}) can be put into the form of problem \cite[(2.291)]{BS} involving an indicator function of set $\Gamma(d)$ and the dual problem of \cite[(2.291)]{BS} is in the form of \cite[(2.298)]{BS} and the conjugate function of an indicator function is the support function, the dual problem of (\ref{convex-subproblem}) is
$$ \max_{\lambda} \left \{ \inf_{w} {\cal L}(w, \lambda)
-\sigma \left( \lambda|\Gamma(d) \right) \right \},$$
where ${\cal L}(w,\lambda):=\nabla_x L(x^*,\lambda)w +\nabla_{xx}^2 L(x^*, \lambda)(d,d) $ is the Lagrange function of (\ref{convex-subproblem}).
{Note that
\[\sigma(\lambda|\Gamma(d))=\sigma\big(\lambda|{\cal T}(d)+\widehat{T}_{\Omega}(F(x^*))\big)=\sigma(\lambda|{\cal T}(d))+
\sigma\big(\lambda|\widehat{T}_{\Omega}(F(x^*))\big)=+\infty,\]
whenever
$\lambda\notin [\widehat{T}_{\Omega}(F(x))]^\circ=N^c_{\Omega}(F(x))$.}
Therefore, the dual problem of (\ref{convex-subproblem}) is
\begin{equation}\label{dual-problem}
\max_{\lambda\in \Lambda^c(x^*)} \left \{ \nabla^2_{xx}L(x^*,\lambda)(d,d)
-\sigma \left( \lambda|\Gamma(d) \right) \right \}
= \nabla^2_{xx}L(x^*,\lambda_0)(d,d)-\sigma \left( \lambda_0|\Gamma(d) \right),
\end{equation}
where the equality holds since $\Lambda^c(x^*)=\{\lambda_0\}$ by Step 1.
Since ${\rm lin}T_{\Omega}(F(x^*))=\widehat{T}_{\Omega}(F(x^*))$ by Proposition \ref{frechet-tangent-cone} and
${\rm lin}T_{\Omega}(F(x^*))$ is a subspace, we have ${\rm lin}T_{\Omega}(F(x^*))=-\widehat{T}_{\Omega}(F(x^*)).$
Hence condition (\ref{conditions-revision}) is
$\nabla F(x^*)\Re^n-\widehat{T}_{\Omega}(F(x^*))=\Re^{2m}$, which in turn implies
$\nabla F(x^*)\Re^n-\big({\cal T}(d)+\widehat{T}_{\Omega}(F(x^*))\big)=\Re^{2m}$. Hence
$\nabla F(\bar x)\Re^n-\Gamma(d)=\Re^{2m}$. So
the
Robinson's constraint qualification
(see \cite[(2.313)]{BS}) for problem (\ref{convex-subproblem}) holds. It ensures that the zero dual gap property holds (see
\cite[Theorem 2.165]{BS}). Hence the optimal value of the dual
problem (\ref{dual-problem}) is equal to the optimal value of problem (\ref{convex-subproblem}) and hence nonnegative. In addition, noting
${\cal T}(d)\subseteq \Gamma(d)$,
$\sigma(\lambda_0| {\cal T}(d))\leq \sigma(\lambda_0|\Gamma(d))$, which further implies that
\begin{equation} \label{second-order-conditions}
\nabla^2_{xx}L(x^*,\lambda_0)(d,d)-\sigma \left( \lambda_0|{\cal T}(d) \right) \geq 0.
\end{equation}
\noindent
Step 3. Note that $T^2_{\Omega}(F(x^*);\nabla F(x^*)d)=\bigcup_{a\in T^2_{\Omega}(F(x^*);\nabla F(x^*)d)}\{a\}$ is the union of
convex sets. For each $a\in T^2_{\Omega}(F(x^*);\nabla F(x^*)d)$, by (\ref{second-order-conditions}) we have
$$ \nabla^2_{xx}L(x^*,\lambda_0)(d,d)-\langle \lambda_0,a \rangle \geq 0.$$
It then yields the desired result
\[
\nabla^2_{xx}L(x^*,\lambda_0)(d,d) - \sigma \left( \lambda_0|T^2_{\Omega}(F(x^*);
\nabla F(x^*)d) \right) \geq 0.
\]
\hskip .5cm $\Box$ \vskip .5cm
\begin{remark}
The nondegeneracy condition (\ref{conditions-revision}), together with the special geometric structure of second-order cone complementarity set, can ensure not only the uniqueness of Lagrangian multiplier in Step 1, but also the zero-dual gap property between (\ref{convex-subproblem}) and (\ref{dual-problem}) in Step 2. The nondegeneracy condition, stronger than the Robinson's constraint qualification, is a generalization of linear independence constraint qualification in the conic case. We refer to \cite[Proposition 4.75]{BS} for the detailed discussion on the relationship between nondegeneracy condition and uniqueness of multiplier in the convex case.
\end{remark}
We next derive the exact formula for the support function of the second-order tangent set to the SOC complementarity set needed in applying Theorem \ref{dual-conditions}. Under the assumption of Theorem
\ref{dual-conditions} we have $C(x^*)=\big\{d|\, \nabla f(x^*)d= 0, \nabla F(x^*)d\in T_{\Omega}(F(x^*)) \big\}$. Thus $d\in C(x^*)$ if and only if $\nabla F(x^*)d\in T_{\Omega}(F(x^*))$ and $\langle \lambda_0, \nabla F(x^*)d\rangle =0$. Therefore the following results will be useful.
\begin{proposition}\label{formula-regular normal cone}
{ For $(x,y)\in \Omega$ and $(d,w)\in T_\Omega(x,y)$, take $(u,v)\in \widehat{N}_{\Omega}(x,y)$ such that $\langle (u,v),(d,w) \rangle=0$. Then }
\[
\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=\left\{\begin{array}{lll}
0, \ & {\rm if} \ x\in {\rm int} {\cal K}\ {\rm and}\ y=0; \\
0, \ & {\rm if}\ x=0 \ {\rm and}\ y \in {\rm int}{\cal K};\\
0, \ & {\rm if}\ x=0 \ {\rm and} \ y=0.
\end{array}\right\}.
\]
If $x\in {\rm bd}{\cal K}\backslash \{0\}$ and $y=0$, then
\begin{eqnarray*}
&&\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)\\
&=&\left\{\begin{array}{lll}
0, \ & {\rm if} \ d\in {\rm int}T_{{\cal K}}(x)\ {\rm and}\ w=0;
\\
-\frac{u_1}{x_1}(d_1^2-\|d_2\|^2)-2\frac{w_1d_2^Tv_2}{\|x_2\|}-2\frac{d_1w_2^Tv_2}{\|x_2\|}, \ \ & {\rm if}\ d\in {\rm bd}T_{{\cal K}}(x) \ {\rm and} \ w\in \Re_{+}\hat{x}.
\end{array}\right\}.
\end{eqnarray*}
If $x=0$ and $y\in {\rm bd}{\cal K}\backslash \{0\}$, then
\begin{eqnarray*}
&&\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)\\
&=&\left\{\begin{array}{lll}
0, \ & {\rm if} \ d=0\ {\rm and}\ w\in {\rm int}T_{{\cal K}}(y); \\
-\frac{v_1}{y_1}(w_1^2-\|w_2\|^2)-2\frac{d_1w_2^Tu_2}{\|y_2\|}
-2\frac{w_1d_2^Tu_2}{\|y_2\|}, \ \ & {\rm if}\ d\in \Re_{+}\hat{y} \ {\rm and} \ w\in {\rm bd}T_{{\cal K}}(y).
\end{array}\right\}.
\end{eqnarray*}
If $x,y\in {\rm bd}{\cal K}\backslash \{0\}$, then
\begin{eqnarray*}
&&\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)\\
&=& \frac{x_1u_1+y_1v_1}{x^2_1}\Big(\|d_2\|^2-d_1^2\Big)+\frac{x_1w_1-y_1d_1}{x_1y_1} \Big(w^Tv-d^Tu\Big).
\end{eqnarray*}
\end{proposition}
\par\noindent {\bf Proof.}\ \ {For $(x,y) \in \Omega$, take $(d,w)\in T_\Omega(x,y)$, $(p,q)\in T^2_\Omega((x,y);(d,w))$} with the exact formula given in Theorem \ref{formula-regular normal cone-1} and $(u,v) \in \widehat{N}_\Omega(x,y) $ whose exact formula can be found in \cite[Theorem 3.1]{yzhou}.
\noindent {\bf Case (i)} $x\in {\rm int}{\cal K}$ and $y=0$. In this case $u=0$ and $q=0$. Hence
$\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=
\max\{\langle (u,v),(p,q)\rangle|(p,q)\in T_\Omega^2((x,y);(d,w))\}=0$. The proof for the case of $x=0$ and $y\in {\rm int}{\cal K}$ is similar and hence we omit it.
\noindent {\bf Case (ii)} $x\in {\rm bd}{\cal K}\backslash \{0\}$ and $y=0$. Then $u\in \Re_-\hat{x}$ and $v\in \hat{x}^\circ$ by the formula of $\widehat{N}_\Omega(x,y)$.
${\bf (ii)}$-1. Suppose further that $d\in {\rm int} T_{{\cal K}}(x)\ {\rm and}\ w=0$. Then by Theorem \ref{formula-regular normal cone-1}, $q=0$. Since $0=\langle u,d \rangle+ \langle v,w \rangle =\langle u,d \rangle $, which together with the fact that $d\in {\rm int} T_{{\cal K}}(x)$ (i.e., $d^T\hat{x}>0$) implies $u=0$.
Hence $\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=\max\{\langle (u,v), (p,q)\rangle|(p,q)\in T^2_{\Omega}((x,y);(d,w))\}=0.$
${\bf (ii)}$-2. Suppose further that $d\in {\rm bd}T_{{\cal K}}(x)\ {\rm and } $ $ w=0$, then $q=0$ or $q\in \Re_+\hat{x}$ by Theorem \ref{formula-regular normal cone-1}. Hence
\begin{eqnarray*}
\sigma\big((u,v)|T^2_{\Omega}\big((x,y);(d,w)\big)\big)
&=&\max\Big\{\sigma(u| T^2_{{\cal K}}(x;d)),\sigma(u|{\rm bd}T^2_{\cal K}(x;d))+\sigma(v|\Re_+\hat{x})\Big\} \\
&=& \sigma(u| T^2_{{\cal K}}(x;d))= -\frac{u_1}{x_1}(d_1^2-\|d_2\|^2),
\end{eqnarray*}
where the second equality holds because
$\sigma(v|\Re_+\hat{x})=0$ since $v\in \hat{x}^\circ$, and the last step comes from the fact that since $u\in \Re_-\hat{x}$,
$\langle u,p \rangle=\frac{u_1}{x_1} \langle \hat{x}, p \rangle\leq \frac{u_1}{x_1}(\|d_2\|^2-d_1^2)$ for all $p\in T^2_{{\cal K}}(x;d)$ by Lemma \ref{Lem3.3}, and the maximum can be attained by letting $p=\frac{\|d_2\|^2-d_1^2}{2x_1^2}\hat{x}$.
{Now consider the case where $d\in {\rm bd}T_{{\cal K}}(x)\ {\rm and } $ and $w\in \Re_{++}\hat{x}$}. From the formula for $ {\rm bd}T_{{\cal K}}(x)$ in this case, we get $d\perp \hat{x}$. Hence $\langle v, w\rangle =\langle (u,v),(d,w) \rangle=0$ taking into the account that $u\in \Re_-\hat{x}$. It further implies that $v\perp \hat{x}$ (i.e., $v_1=\bar{x}_2^Tv_2$), because $w\in \Re_{++}\hat{x}$.
Hence
\begin{eqnarray*}\sigma\big((u,v)|T^2_{\Omega}\big((x,y);(d,w)\big)\big)
&=&\sigma(u| {\rm bd}T^2_{{\cal K}}(x;d)) +\langle v,q\rangle \\
&=&-\frac{u_1}{x_1}(d_1^2-\|d_2\|^2)+v_1q_1-q_1v_2^T\bar{x}_2-2\frac{w_1d_2^Tv_2}{\|x_2\|}-2\frac{d_1w_2^Tv_2}{\|x_2\|}\\
&=& -\frac{u_1}{x_1}(d_1^2-\|d_2\|^2)-2\frac{w_1d_2^Tv_2}{\|x_2\|}-2\frac{d_1w_2^Tv_2}{\|x_2\|}.
\end{eqnarray*}
\noindent {\bf Case (iii)} $x=0$ and $y\in {\rm bd}{\cal K}\backslash \{0\}$. The argument is similar to the above case.
\noindent {\bf Case (iv)} $x,y\in {\rm bd}{\cal K}\backslash \{0\}$. Note that in this case since $x^Ty=0$, we have $y =k \hat{x}$ with $k:=y_1/x_1$. Since $(u,v)\in \widehat{N}_{\Omega}(x,y)$ and $(d,w)\in T_{\Omega}(x,y)$, by the formula of $\widehat{N}_{\Omega}(x,y)$ and $T_{\Omega}(x,y)$ we have $v\perp y$, $d\perp y$, and there exist $\beta,\gamma\in \Re$ such that
$
\hat{u}+kv=\beta x$ and $\hat{w}-kd=rx
$. To simplify the notation, let
\[\xi:= (x_1w_1-y_1d_1)\left(\frac{w_2+w_1\bar x_2}{y_1}-\frac{d_2-d_1\bar x_2}{x_1}\right).
\]
Since $v\perp y$, $y \in \Re \hat{x}$ and $x_1=\|x_2\|\not =0$, we have $v_1-{\bar x}_2^Tv_2=0$. It follows that
\begin{equation}\label{EE}
v_2^T\xi = (x_1w_1-y_1d_1)\left(\frac{w_2^Tv_2+w_1v_1}{y_1}-\frac{d_2^Tv_2-d_1v_1}{x_1}\right)
= \frac{x_1w_1-y_1d_1}{y_1}(w^Tv-d^Tu),
\end{equation}
where in the last step we used the fact that
$d^T\hat{v}=(1/k)d^T(\beta\hat{x}-u)=-(1/k)d^Tu$ since $d\perp \hat{x}$. By the formula of $T^2_\Omega((x,y);(d,w))$ in Theorem \ref{formula-regular normal cone-1} for this case we have
\begin{equation}\label{SOT}
p\in {\rm bd} T^2_{{\cal K}}(x;d),\ \ q\in {\rm bd}T^2_{{\cal K}}(y;w),\ \
\xi-
p_1y_2-q_1x_2=x_1q_2+y_1p_2.\end{equation}
Therefore,
\begin{eqnarray*}
&&\langle u,p \rangle + \langle v,q \rangle \\
&&=\langle \hat{u},\hat{p} \rangle + \langle v,q \rangle =\langle \beta x-kv, \hat{p} \rangle + \langle v,q \rangle = \beta \langle \hat{x}, p \rangle + \langle v,q-k\hat{p} \rangle\\
&&=\beta\langle \hat{x},p \rangle +v_1 (q_1-kp_1)+v_2^T
\left( \frac{\xi}{x_1}+(kp_1-q_1)\bar x_2\right) \\
&&= \beta\langle \hat{x},p \rangle +\frac{1}{x_1}v_2^T\xi
= \frac{x_1u_1+y_1v_1}{x^2_1}\big(\|d_2\|^2-d_1^2\big)+\frac{x_1w_1-y_1d_1}{x_1y_1} (w^Tv-d^Tu),
\end{eqnarray*}
where the forth equality holds by virtue of (\ref{SOT}), the fifty equality holds because $v_1=v_2^T\bar{x}_2$ and the sixth equality holds due to (\ref{EE}) and (\ref{SOT}). The desired formula follows.
\noindent {\bf Case (v)} $x=0$ and $y=0$. In this case $ d, w \in {\cal K}, d\perp w $, $(u,v)\in \widehat{N}_{\Omega}(x,y)=(-{\cal K},-{\cal K})$ and $(p,q) \in T^2_{\Omega}\big((x,y);(d,w)\big)=T_\Omega (d,w)$.
${\bf (v)}$-1. $d=0$ and $w\in {\rm int}{\cal K}$. Since $\langle v,w \rangle =\langle (u,v), (d,w) \rangle=0$ and $v\in -{\cal K}$, we have $v=0$. Since $d=0$ and $w\in {\rm int}{\cal K}$, $(p,q)\in T^2_{\Omega}\big((x,y);(d,w)\big)=T_{\Omega}(d,w)$ implies that
$p=0$. Hence $\langle (u,v), (p,q) \rangle=0$. It follows that $\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=0$.
${\bf (v)}$-2. $d\in {\rm int}{\cal K}$ and $w=0$. It is similar to the above case.
${\bf (v)}$-3. $d,w\in {\rm bd}{\cal K}\backslash\{0\}$. Then since $\langle (u,v), (d,w) \rangle=0$ and $(u,v)\in (-{\cal K},-{\cal K})$, we have $u\in \Re_-\hat{d}=\Re_-w$ and $v\in \Re_-\hat{w}=\Re_-d$.
Since $(p,q)\in T_{\Omega}(d,w)$ and $d,w\in {\rm bd}{\cal K}\backslash\{0\}$, we have $p\perp w$ and $q\perp d$. Hence $p\perp u$ and $q\perp v$. So
$\langle (u,v), (p,q) \rangle=0$. It follows that $\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=0$.
${\bf (v)}$-4. $d=0$ and $w\in {\rm bd}{\cal K}\backslash\{0\}$. Since $\langle v,w \rangle =\langle (u,v), (d,w) \rangle=0$ and $v\in -{\cal K}$, we have $v\in \Re_-\hat{w}$.
In this case since $(p,q)\in T_{\Omega}(d,w)$ with $d=0$ and $w\in {\rm bd}{\cal K}\backslash\{0\}$, we have either $p=0$ and $q\in T_{{\cal K}}(w)$ or $p\in \Re_+\hat{w}$ and $q\perp \hat{w}$. If $p=0$ and $q\in T_{{\cal K}}(w)$ (i.e., $\hat{w}^Tq\geq 0$), then
$\langle (u,v), (p,q) \rangle=\langle v,q \rangle \leq 0$ and the maximum is $0$ which can be attained by letting $q=0$. If $p\in \Re_+\hat{w}$ and $q\perp \hat{w}$, then
$\langle (u,v), (p,q) \rangle=\langle u,p \rangle \leq 0$, where the last step is due to $u\in -{\cal K}$ and $p\in \Re_+\hat{w}\in {\cal K}$, and the maximum is $0$ which can be attained by letting $p=0$. It follows that $\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=0$.
${\bf (v)}$-5. $d\in {\rm bd}{\cal K}/\{0\}$ and $w=0$. It is similar to the above case by symmetric argument.
${\bf (v)}$-6. $d=0$ and $w=0$. In this case $(p,q)\in T_{\Omega}(d,w)=\Omega$. Since $(u,v)\in (-{\cal K},-{\cal K})$, we have
$\langle (u,v), (p,q) \rangle \leq 0$ and the maximum is $0$ which can be attained by letting $(p,q)=(0,0)$. It follows that $\sigma\big((u,v)|T^2_{\Omega}((x,y);(d,w))\big)=0$.
\hskip .5cm $\Box$ \vskip .5cm
\begin{example}
{Consider the following SOCMPCC. }
\[\begin{array}{ll}
\min & f(x):=-x_2^2+x_1-x_4 \\
{\rm s.t.} & {\cal K} \ni G(x):=(x_1,x_3-x_1,x_1-x_2) \perp (-x_2+1,x_1,x_1-x_4)=:H(x) \in {\cal K}.
\end{array}
\]
\end{example}
Since $x_1\geq x_1-x_2$ and $-x_2+1\geq 0$, we have $x_2\in [0,1]$. Hence $x_2^2\leq x_2$. Since
$-x_2+1\geq -x_1+x_4$, $-x_2+x_1-x_4\geq -1$.
Thus
$
-x_2^2+x_1-x_4\geq -x_2+x_1-x_4\geq -1.
$
Hence $x^*=(0,0,0,1)$ is an optimal solution, and $G(x^*)=(0,0,0)$ and $H(x^*)=(1,0,-1)\in {\rm bd}{\cal K}\backslash \{0\}$.
Note that
\[
\nabla G(x^*)=\begin{bmatrix}
1 & 0 & 0 & 0\\
-1 & 0 & 1 & 0 \\
1 & -1 & 0 & 0
\end{bmatrix},\ \ \ \ \ \nabla H(x^*)=\begin{bmatrix}
0 & -1 & 0 & 0\\
1 & 0 & 0 & 0\\
1 & 0 & 0 & -1
\end{bmatrix}\]
and by the formula of the tangent cone in Lemma \ref{tangentcone}, we have
\begin{equation}\label{Tconeeg} T_{\Omega}(G(x^*),H(x^*))=\left \{ (d,w)\left| \begin{array}{l}
\mbox{either } d=0, -w_1-w_3\leq 0 \\
\mbox {or } d=t(1,0,1) \mbox{ for some } t\geq 0, w_1+w_3=0
\end{array} \right. \right\}.\end{equation}
It follows that
\[ {\rm lin}T_{\Omega}(G(x^*),H(x^*))=\{((0,0,0),(\tau_1,\tau_2,-\tau_1))|\tau_1,\tau_2\in \Re\}.
\]
For any $v\in \Re^6$, take $\xi=(v_1,v_1-v_3,v_1+v_2,v_3-v_4-v_6)\in \Re^4$ and $\tau=(v_1-v_3+v_4,v_5-v_1,-(v_1-v_3+v_4))\in \Re^3$, then \[
v= \left(\begin{matrix}
\nabla G(x^*)\\
\nabla H(x^*)
\end{matrix}\right)\xi+\left(\begin{matrix}
0\\
\tau
\end{matrix}\right)\in \nabla F(x^*)\Re^4+{\rm lin}T_{\Omega}(F(x^*)).
\]
Since $v$ is arbitrarily taken from $\Re^6$, condition (\ref{conditions-revision}) holds.
The Lagrangian multiplier system is
\[\left\{\begin{array}{cc}
\left(\begin{matrix}
1 \\ 0 \\ 0 \\-1
\end{matrix}\right)+
\lambda^G_1\left(\begin{matrix}
1 \\ 0 \\ 0 \\0
\end{matrix}\right)+ \lambda^G_2\left(\begin{matrix}
-1 \\ 0 \\ 1 \\0
\end{matrix}\right) + \lambda^G_3
\left(\begin{matrix}
1 \\ -1 \\ 0 \\0
\end{matrix}\right)
+ \lambda^H_1\left(\begin{matrix}
0 \\ -1 \\ 0 \\0 \end{matrix}\right)+ \lambda^H_2\left(\begin{matrix}
1 \\ 0 \\ 0 \\0 \end{matrix}\right)+\lambda^H_3\left(\begin{matrix}
1 \\ 0 \\ 0\\ -1 \end{matrix}\right)=\left(\begin{matrix}
0 \\ 0 \\ 0 \\0
\end{matrix}\right), \\
(\lambda^G,\lambda^H)\in N_{\Omega}\big(G(x^*),H(x^*)\big).
\end{array}
\right.
\]
Since $G(x^*)=(0,0,0)$ and $H(x^*)=(1,0,-1)\in {\rm bd}{\cal K}\backslash \{0\}$, we obtain the following expression of the limiting normal cone from \cite[Theorem 5.1]{yzhou}
\begin{flalign*}
&N_{\Omega}\big(G(x^*),H(x^*)\big)\\
&=\{(u,v)|v=0 \ \ {\rm or} \ \ u_1+u_3=0, v=t(1,0,1),t\in \Re\ \ {\rm or} \ \ u_1+u_3\leq 0, v=t(1,0,1), t\leq 0\}.
\end{flalign*}
Hence the only multipliers $(\lambda^G,\lambda^H)$ satisfying the Lagrangian multiplier system is $\lambda^G=(-1,0,1)$ and $\lambda^H=(-1,0,-1)$. Note that
\begin{eqnarray*}
C(x^*)&=&\{d\in \Re^4| (d_1,d_3-d_1,d_1-d_2,-d_2,d_1,d_1-d_4)\in T_{\Omega}(G(x^*),H(x^*)), d_1\leq d_4\} \\
&=& \{d=(t,0,t,t)|t\geq 0\},
\end{eqnarray*}
where the second equality follows from (\ref{Tconeeg}).
Since $\nabla G(x^*)d=(t,0,t), \nabla H(x^*)d =(0,t,0)$ for any $d=(t,0,t,t)$ with $t\geq 0$ in $C(x^*)$,
by
Proposition \ref{formula-regular normal cone} we obtain
\[
\sigma\big((\lambda^G,\lambda^H)|T^2_\Omega(G(x^*),H(x^*);\nabla G(x^*)d,\nabla H(x^*)d)\big)=
-t^2=-d_1^2.
\]
Since $
\nabla^2_{xx} L(x, \lambda)=\nabla^2f(x)=\begin{bmatrix}
0 & 0 & 0 & 0\\
0 & -2 & 0 & 0 \\
0 & 0 & 0 & 0\\
0&0&0&0
\end{bmatrix}$, we have
\begin{equation}\label{re-eq-2}
\nabla^2_{xx}L(x^*,\lambda)(d,d)=0, \ \ \forall d\in C(x^*),
\end{equation}
and by Theorem \ref{dual-conditions}
\begin{eqnarray}
\Upsilon(x^*,\lambda)(d) &:=& \nabla^2_{xx}L(x^*,\lambda)(d,d)- \sigma\big((\lambda^G,\lambda^H)|T^2_\Omega(G(x^*),H(x^*);\nabla G(x^*)d,\nabla H(x^*)d)\big) \nonumber\\
&=& d_1^2\geq 0, \ \ \ \forall d\in C(x^*). \label{re-eq-1}
\end{eqnarray}
(\ref{re-eq-2}) and (\ref{re-eq-1}) indicate that $\nabla^2_{xx}L(x^*,\lambda)$ is positive semidefinite over $C(x^*)$ while
$\Upsilon(x^*,\lambda)$ is positive definite over $C(x^*)\backslash\{0\}$. In this example, the second-order necessary conditions involving the second-order tangent set (\ref{re-eq-1}) is stronger than the one not involving the second-order tangent set (\ref{re-eq-2}).
\section*{Acknowledgments.} {The authors are gratefully indebted to the anonymous referees for their valuable suggestions that helped us greatly improve the original presentation of the paper.}
|
\section{Introduction}\label{sec1}
In many wave propagation problems it is possible to define a preferred direction of propagation.
For example, in ocean acoustics, waves propagate primarily in the horizontal direction in an acoustic wave guide, bounded by the water surface and the ocean bottom.
Similarly, in communication engineering,
microwaves or optical waves propagate as beams through electromagnetic or optical wave guides.
Wave propagation in preferred directions is not restricted to wave guides. For example, in geophysical \rev{reflection} imaging applications,
seismic or electromagnetic waves propagate mainly in the vertical direction (downward and upward) through a laterally unbounded medium.
To account for propagation in preferred directions,
the wave equation for the full wave field can be decomposed into a set of coupled equations for waves that propagate in opposite directions along the preferred axis
(for example leftward and rightward in ocean acoustics, or downward and upward in \rev{reflection} imaging).
In the literature on electromagnetic wave propagation these oppositely propagating waves are often called ``bidirectional beams''
\citep{Hoekstra97OQE, Stralen97OQE}
whereas in the acoustic literature they are usually called ``one-way wave fields''
\citep{Claerbout71GEO, Berkhout82Book, McCoy87WM, Holberg88GP, Fishman93RS}.
In this paper we use the latter terminology.
There is a vast amount of literature on the analytical and numerical aspects of one-way wave propagation
\citep{Fishman84JMP, Fishman84JMP2, Halpern88JASA, Weston89JMP, Fishman91WM, Grimbergen98GEO2}.
A discussion of this is beyond the scope of this paper. Instead,
we concentrate on the choice of the decomposition operator and the consequences for reciprocity and representation theorems.
Decomposition of a wave field into one-way wave fields is not unique. In particular, the amplitudes of the one-way wave fields can be scaled in different ways.
In this paper we distinguish between so-called ``flux-normalised'' and ``field-normalised'' one-way wave fields.
The square of the amplitude of a flux-normalised one-way wave field is by definition the power-flux density
(or, for quantum-mechanical waves, the probability-flux density)
in the direction of preference.
Field-normalised one-way wave fields, on the other hand, are scaled such that the sum of the two oppositely propagating components equals the full wave field.
These two forms of normalisation have been briefly analysed by De Hoop
\citep{Hoop92PHD, Hoop96JMP}.
From this analysis it appeared that the operators for flux-normalised decomposition
exhibit more symmetry than the operators for field-normalised decomposition. Exploiting the symmetry of the flux-normalised decomposition operators, the author derived reciprocity
and representation theorems for flux-normalised one-way wave fields
\citep{Wapenaar96GJI1, Wapenaar96GJI2}.
The first aim of this paper is to discuss flux-normalised versus field-normalised decomposition in a systematic way.
In particular, it will be shown that reciprocity
theorems for field-normalised one-way wave fields can be derived in a similar way as those for flux-normalised one-way wave fields,
even though the operators for field-normalised decomposition exhibit less symmetry.
The second aim is to discuss representation theorems for field-normalised one-way wave fields in a systematic way. This discussion includes links to ``classical'' Kirchhoff-Helmholtz
integrals for one-way wave fields as well as to recent single-sided representations for backward propagation, used for example in Marchenko imaging
\citep{Wapenaar2014GEO}.
Despite the links to earlier results, the discussed representations are more general.
An advantage of the representations for field-normalised one-way wave fields is that a straightforward summation of the one-way wave fields gives the full wave field.
We restrict the discussion to scalar wave fields. In section \ref{sec2} we formulate a unified scalar wave equation for acoustic waves,
horizontally polarised shear waves, transverse electric and transverse magnetic EM waves
and, finally, quantum-mechanical waves.
Next, we reformulate the unified wave equation into a matrix-vector form, discuss symmetry properties of the operator matrix and use this to derive reciprocity theorems
in matrix-vector form.
In section \ref{sec4} we decompose the matrix-vector wave equation into a coupled system of equations for oppositely propagating one-way wave fields. We separately consider
flux-normalisation and field-normalisation and derive reciprocity theorems for \rev{one-way wave fields, using} both normalisations.
In section \ref{sec5} we extensively discuss representation theorems for field-normalised one-way wave fields and indicate applications. We end with conclusions in section \ref{sec6}.
\section{Unified wave equation and its symmetry properties}\label{sec2}
\subsection{Unified scalar wave equation}
Using a unified notation, wave propagation in a lossless medium
(or, for quantum-mechanical waves, in a lossless potential)
is governed by the following two equations in the space-frequency domain
\begin{eqnarray}
&&-i\omega\alpha P+\partial_jQ_j=B,\label{eqA1}\\
&&-i\omega\beta Q_j+\partial_j P=C_j.\label{eqA2}
\end{eqnarray}
\rev{Here $i$ is the imaginary unit and $\omega$ the angular frequency (in this paper we consider positive frequencies only).
Operator $\partial_j$ stands for the spatial differential operator $\partial/\partial x_j$ and Einstein's summation convention applies to repeated subscripts.}
$P({\bf x},\omega)$ and $Q_j({\bf x},\omega)$ are space- and frequency-dependent wave field quantities,
$\alpha({\bf x})$ and $\beta({\bf x})$ are real-valued space-dependent parameters, and
$B({\bf x},\omega)$ and $C_j({\bf x},\omega)$ are space- and frequency-dependent source distributions.
\rev{Parameters $\alpha$ and $\beta$ are both assumed to be positive, hence, metamaterials are not considered in this paper.}
All quantities are specified in Table 1 for different wave phenomena and are discussed in more detail below.
As indicated in the first column of Table 1, we consider 3D and 2D wave problems. For the 3D situation, ${\bf x}=(x_1,x_2,x_3)$ is the 3D Cartesian coordinate vector and
\rev{lower-case} Latin subscripts take on the values 1, 2 and 3. For the 2D situation, ${\bf x}=(x_1,x_3)$ is the 2D Cartesian coordinate vector and
\rev{lower-case} Latin subscripts take on the values 1 and 3 only.
\begin{center}
{{\noindent \it Table 1: Quantities in unified equations (\ref{eqA1}) and (\ref{eqA2}).}
\begin{tabular}
{||l||c|c||c|c||c|c||}
\hline\hline
& $P$ & $Q_j$ &$\alpha$ &$\beta$ & $B$ & $C_j$ \\
\hline\hline
1. Acoustic waves (3D) & $p$ & $v_j$ &$\kappa$ &$\rho$ & $q$ & $f_j$ \\
\hline
2. SH waves (2D) & $v_2$ & $-\tau_{2j}$ &$\rho$ &$\frac{1}{\mu}$ & $f_2$ & $2h_{2j}$ \\
\hline
3. TE waves (2D) & $E_2$ & $-\epsilon_{2jk}H_k$ &$\varepsilon$ &$\mu$ & $-J_2^{\rm e}$ & $\epsilon_{2jk}J_k^{\rm m}$ \\
\hline
4. TM waves (2D) & $H_2$ & $\epsilon_{2jk}E_k$ &$\mu$ &$\varepsilon$ & $-J_2^{\rm m}$ & $-\epsilon_{2jk}J_k^{\rm e}$ \\
\hline
5. Quantum waves (3D) & $\Psi$ & $\frac{2\hbar}{mi}\partial_j\Psi$ &$4-\frac{4V}{\hbar\omega}$ &$\frac{m}{2\hbar\omega}$ & & \\
\hline
\hline
\end{tabular}
}
\end{center}
\mbox{}\\
\rev{The unified boundary conditions at an interface between two media with different parameters state that $P$ and $n_jQ_j$ are continuous over the interface. Here $n_j$ represents the
components of the normal vector ${\bf n}=(n_1,n_2,n_3)$ at the interface for the 3D situation, or ${\bf n}=(n_1,n_3)$ for the 2D situation.}
We discuss the quantities in Table 1 in more detail. The quantities in row 1, associated to 3D acoustic wave propagation in a lossless fluid medium, are
acoustic pressure $p({\bf x},\omega)$, particle velocity $v_j({\bf x},\omega)$, compressibility $\kappa({\bf x})$, mass density $\rho({\bf x})$,
volume-injection rate density $q({\bf x},\omega)$ and external force density $f_j({\bf x},\omega)$.
For 2D horizontally polarised shear waves in a lossless solid medium, we have in row 2 horizontal particle velocity $v_2({\bf x},\omega)$, shear stress $\tau_{2j}({\bf x},\omega)$,
mass density $\rho({\bf x})$, shear modulus $\mu({\bf x})$, external horizontal force density $f_2({\bf x},\omega)$ and external shear deformation rate density $h_{2j}({\bf x},\omega)$.
Rows 3 and 4 contain the quantities for 2D electromagnetic wave propagation, with TE standing for transverse electric and TM for transverse magnetic. The quantities are
electric field strength $E_k({\bf x},\omega)$, magnetic field strength $H_k({\bf x},\omega)$, permittivity $\varepsilon({\bf x})$, permeability $\mu({\bf x})$,
external electric current density $J_k^{\rm e}({\bf x},\omega)$ and external magnetic current density $J_k^{\rm m}({\bf x},\omega)$. Furthermore, $\epsilon_{ijk}$ is
the alternating tensor (or Levi-Civita tensor), with $\epsilon_{123}=\epsilon_{312}=\epsilon_{231}=1$,
$\epsilon_{213}=\epsilon_{321}=\epsilon_{132}=-1$, and all other components being zero.
In row 5, the quantities related to 3D quantum-mechanical wave propagation are wave function $\Psi ({\bf x},\omega)$,
potential $V({\bf x})$, particle mass $m$ and $\hbar=h/2\pi$, with $h$ Planck's constant.
By eliminating $Q_j$ from equations (\ref{eqA1}) and (\ref{eqA2}) we obtain the unified scalar wave equation
\begin{eqnarray}
&&\beta\partial_j\Bigl(\frac{1}{\beta}\partial_j P\Bigr) +k^2P=\beta\partial_j\Bigl(\frac{1}{\beta}C_j\Bigr)+i\omega\beta B,\label{eqwe}
\end{eqnarray}
with wave number $k$ defined via
\begin{eqnarray}\label{eqk}
&&k^2=\alpha\beta\omega^2.
\end{eqnarray}
\subsection{Unified wave equation in matrix-vector form}
\begin{figure}
\vspace{0cm}
\centerline{\epsfysize=8 cm \epsfbox{Figure1.pdf}}
\caption{\footnotesize
\rev{Configuration with the $x_3$-direction as the preferred direction. In the lateral direction this configuration can be bounded (for wave guides) or unbounded (for example for geophysical reflection imaging applications).
For the 2D situation the configuration is a cross-section of the 3D situation for $x_2=0$.}
}\label{Fig1}
\end{figure}
\rev{We define a configuration with a preferred direction and reorganise equations (\ref{eqA1}) and (\ref{eqA2}) accordingly.}
\rev{Consider a 3D spatial domain $\mathbb{D}$, enclosed by surface ${\partial\setD}$.
This surface consists of two planar surfaces ${\partial\setD}_0$ and ${\partial\setD}_1$ perpendicular to the $x_3$-axis and a cylindrical surface ${\partial\setD}_{\rm cyl}$
with its axis parallel to the $x_3$-axis, see Figure \ref{Fig1}.
The surfaces ${\partial\setD}_0$ and ${\partial\setD}_1$ are situated
at $x_3=x_{3,0}$ and $x_3=x_{3,1}$, respectively, with $x_{3,1}> x_{3,0}$. In general these surfaces do not coincide with physical boundaries.
Surface $\mathbb{S}$ in Figure \ref{Fig1} is a cross-section of $\mathbb{D}$ at arbitrary $x_3$.
The parameters $\alpha({\bf x})$ and $\beta({\bf x})$ are piecewise continuous smoothly varying functions in $\mathbb{D}$, with discontinuous jumps only at interfaces that are perpendicular to the $x_3$-axis
(hence, $P$ and $Q_3$ are continuous over the interfaces).
In the lateral direction the domain $\mathbb{D}$ can be bounded or unbounded. When $\mathbb{D}$ is laterally bounded, the configuration in Figure \ref{Fig1} represents a wave guide.
For this situation we assume that homogeneous Dirichlet or Neumann boundary conditions apply, i.e., $P=Q_3=0$ or $n_\nu\partial_\nu P=n_\nu\partial_\nu Q_3=0$ at ${\partial\setD}_{\rm cyl}$,
where lower-case Greek subscripts take on the values 1 and 2.
When $\mathbb{D}$ is laterally unbounded (for example for reflection imaging applications), the cylindrical surface ${\partial\setD}_{\rm cyl}$ has an infinite radius and we assume that
$P$ and $Q_3$ have ``sufficient decay'' at infinity.
For the 2D situation, the configuration is a cross-section of the 3D situation for $x_2=0$ and lower-case Greek subscripts take on the value 1 only.}
We reorganise equations (\ref{eqA1}) and (\ref{eqA2}) into a matrix-vector wave equation which acknowledges \rev{the $x_3$-direction as the} direction of preference.
By eliminating the lateral components $Q_1$ and $Q_2$ (or, for 2D wave problems, the lateral component $Q_1$),
we obtain
\citep{Corones75JMAA, Ursin83GEO, Fishman84JMP, Wapenaar89Book, Hoop96JMP}
\begin{eqnarray}\label{eq2.1}
&&\partial_3{\bf q}={{\mbox{\boldmath ${\cal A}$}}}{\bf q}+{\bf d},
\end{eqnarray}
where wave vector ${\bf q}$ and source vector ${\bf d}$ are defined as
\begin{eqnarray}
&&{\bf q}=
\begin{pmatrix} P\\ Q_3\end{pmatrix},
\quad {\bf d}=
\begin{pmatrix}C_3\\ B_0\end{pmatrix}, \label{eq2.2q}
\end{eqnarray}
with
\begin{eqnarray}
&&B_0=B+\frac{1}{i\omega}\partial_\nu\frac{1}{\beta}C_\nu\label{eq2.5}
\end{eqnarray}
and operator matrix ${{\mbox{\boldmath ${\cal A}$}}}$ defined as
\begin{eqnarray}
&&\quad{{\mbox{\boldmath ${\cal A}$}}}=\begin{pmatrix} 0 & {\cal A}_{12} \\ {\cal A}_{21} & 0\end{pmatrix}, \label{eq2.2}
\end{eqnarray}
with
\begin{eqnarray}
&&{\cal A}_{12}=i\omega\beta,\label{eq2.3}\\
&&{\cal A}_{21}=i\omega\alpha-\frac{1}{i\omega}\partial_\nu\frac{1}{\beta}\partial_\nu.\label{eq2.4}
\end{eqnarray}
The notation in the right-hand side of equations (\ref{eq2.5}) and (\ref{eq2.4}) should be understood in the sense that differential operators
act on all factors to the right of it. Hence, operator $\partial_\nu\frac{1}{\beta}\partial_\nu$,
applied via equation (\ref{eq2.1}) to $P$, stands for $\partial_\nu (\frac{1}{\beta}\partial_\nu P)$.
Note that the quantities contained in the wave vector ${\bf q}$ \rev{are continuous over interfaces perpendicular to the $x_3$-axis. Moreover, these quantities }constitute the power-flux density
(or, for quantum-mechanical waves, the probability-flux density)
in the $x_3$-direction via
\begin{eqnarray}\label{eqJ}
&&j=\frac{1}{4}\{P^*Q_3+Q_3^*P\}.
\end{eqnarray}
where the asterisk denotes complex conjugation.
\subsection{Symmetry properties of the operator matrix}\label{secsymx}
We discuss the symmetry properties of the operator matrix ${{\mbox{\boldmath ${\cal A}$}}}$.
First, consider a general operator ${\cal U}$ (which can be a scalar or a matrix), containing space-dependent parameters and differential operators $\partial_\nu$.
We introduce the transpose operator ${\cal U}^t$ via the following integral property
\begin{eqnarray}\label{eq90c}
&&\int_\mathbb{S} ({\cal U}f)^tg\,{\rm d}{\bf x}_{\rm L}=\int_\mathbb{S} f^t({\cal U}^tg)\,{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
Here ${\bf x}_{\rm L}$ is the lateral coordinate vector, with ${\bf x}_{\rm L}=(x_1,x_2)$ for 3D and ${\bf x}_{\rm L}=x_1$ for 2D wave problems.
$\mathbb{S}$ denotes an integration surface perpendicular to the $x_3$-axis at arbitrary $x_3$, \rev{with edge $\partial\mathbb{S}$, see Figure \ref{Fig1}.}
The quantities $f({\bf x}_{\rm L})$ and $g({\bf x}_{\rm L})$ are space-dependent test functions (scalars or vectors).
\rev{When these functions are vectors, $f^t$ is the transpose of $f$; when they are scalars, $f^t$ is equal to $f$.}
\rev{When $\mathbb{S}$ is bounded, homogeneous Dirichlet or Neumann conditions are imposed at $\partial\mathbb{S}$.
When $\mathbb{S}$ is unbounded, $\partial\mathbb{S}$ has an infinite radius and $f({\bf x}_{\rm L})$ and $g({\bf x}_{\rm L})$ are assumed to have} sufficient decay along $\mathbb{S}$ towards infinity.
Operator ${\cal U}$ is said to be symmetric when ${\cal U}^t={\cal U}$ and skew-symmetric when ${\cal U}^t=-{\cal U}$.
For the special case that ${\cal U}=\partial_\nu$, equation (\ref{eq90c}) implies $\partial_\nu^t=-\partial_\nu$ (via integration by parts \rev{along $\mathbb{S}$}). Hence, operator $\partial_\nu$ is skew-symmetric.
We introduce the adjoint operator ${\cal U}^\dagger$ (\rev{i.e., the complex conjugate transpose of ${\cal U}$})
via the integral property
\begin{eqnarray}\label{eq90d}
&&\int_\mathbb{S} ({\cal U}f)^\dagger g\,{\rm d}{\bf x}_{\rm L}=\int_\mathbb{S} f^\dagger({\cal U}^\dagger g)\,{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
\rev{When the test functions are vectors, $f^\dagger$ is the complex conjugate transpose of $f$; when they are scalars, $f^\dagger$ is the complex conjugate of $f$.}
Operator ${\cal U}$ is said to be Hermitian (or self-adjoint) when ${\cal U}^\dagger={\cal U}$ and skew-Hermitian when ${\cal U}^\dagger=-{\cal U}$.
For the operators ${\cal A}_{12}$ and ${\cal A}_{21}$, defined in equations (\ref{eq2.3}) and (\ref{eq2.4}), we find ${\cal A}_{12}^t={\cal A}_{12}$, ${\cal A}_{21}^t={\cal A}_{21}$,
${\cal A}_{12}^\dagger=-{\cal A}_{12}$ and ${\cal A}_{21}^\dagger=-{\cal A}_{21}$.
Hence, operators ${\cal A}_{12}$ and ${\cal A}_{21}$ are symmetric and skew-Hermitian.
With these relations, we find for the operator matrix ${{\mbox{\boldmath ${\cal A}$}}}$
\begin{eqnarray}
{{\mbox{\boldmath ${\cal A}$}}}^t{\bf N}&=&-{\bf N}{{\mbox{\boldmath ${\cal A}$}}},\label{eqsym}\\
{{\mbox{\boldmath ${\cal A}$}}}^\dagger{\bf K}&=&-{\bf K}{{\mbox{\boldmath ${\cal A}$}}},\label{eqsymad}
\end{eqnarray}
with
\begin{eqnarray}\label{eq4.3NK}
&&{{\bf N}}=\begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix},
\quad {{\bf K}}=\begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}.
\end{eqnarray}
Note that, using the expressions for ${\bf q}$ and ${\bf K}$ in equations (\ref{eq2.2q}) and (\ref{eq4.3NK}), we can rewrite equation (\ref{eqJ}) for the
power-flux density
(or, for quantum-mechanical waves, the probability-flux density)
as
\begin{eqnarray}
&&j=\frac{1}{4}{\bf q}^\dagger{\bf K}{\bf q}.\label{eqjj}
\end{eqnarray}
\subsection{Reciprocity theorems}\label{sec3d}
We derive reciprocity theorems between two independent solutions of wave equation (\ref{eq2.1}) \rev{for the configuration of Figure \ref{Fig1}.}
We consider two states $A$ and $B$, characterised by wave vectors ${\bf q}_A({\bf x},\omega)$ and ${\bf q}_B({\bf x},\omega)$, obeying wave equation (\ref{eq2.1}),
with source vectors ${\bf d}_A({\bf x},\omega)$ and ${\bf d}_B({\bf x},\omega)$.
In domain $\mathbb{D}$, the parameters $\alpha$ and $\beta$, and hence the matrix operator ${{\mbox{\boldmath ${\cal A}$}}}$, are chosen the same in the two states
(outside ${\partial\setD}$ they may be different in the two states).
Consider the quantity $\partial_3\bigl({\bf q}_A^t{\bf N}{\bf q}_B\bigr)$ in domain $\mathbb{D}$.
Applying the product rule for differentiation, using equation (\ref{eq2.1}) for both states,
integrating the result over $\mathbb{D}$ and applying the theorem of Gauss yields
\begin{eqnarray}\label{eq4.10}
&&\int_\mathbb{D}\Bigl(\bigl(({{\mbox{\boldmath ${\cal A}$}}}{\bf q}_A)^t+{\bf d}_A^t\bigr){\bf N}{\bf q}_B +{\bf q}_A^t{\bf N}\bigl({{\mbox{\boldmath ${\cal A}$}}}{\bf q}_B+{\bf d}_B\bigr) \Bigr){\rm d}{\bf x}=\int_{\partial\setD}{\bf q}_A^t{\bf N}{\bf q}_B n_3{\rm d}{\bf x}.
\end{eqnarray}
Here $n_3$ is the component parallel to the $x_3$-axis of the outward pointing normal vector on ${\partial\setD}$, with $n_3=-1$ at ${\partial\setD}_0$,
$n_3=+1$ at ${\partial\setD}_1$ \rev{and $n_3=0$ at ${\partial\setD}_{\rm cyl}$, see Figure \ref{Fig1}. In the following the integral on the right-hand side is restricted to the horizontal surfaces ${\partial\setD}_0$ and ${\partial\setD}_1$, which
together are denoted by ${\setdD_{0,1}}$.} The integral on the left-hand side can be written as $\int_\mathbb{D}(\cdots){\rm d}{\bf x}=\int_{x_{3,0}}^{x_{3,1}}{\rm d}x_3\int_\mathbb{S}(\cdots){\rm d}{\bf x}_{\rm L}$.
Using equation (\ref{eq90c}) for the integral along $\mathbb{S}$ and symmetry property (\ref{eqsym}), it follows that the two terms in equation (\ref{eq4.10}) containing operator ${{\mbox{\boldmath ${\cal A}$}}}$ cancel each other.
Hence, we are left with
\begin{eqnarray}\label{eq4.1}
&&\int_\mathbb{D}\bigl({\bf d}_A^t{\bf N}{\bf q}_B +{\bf q}_A^t{\bf N}{\bf d}_B \bigr){\rm d}{\bf x}=
\int_{\setdD_{0,1}}{\bf q}_A^t{\bf N}{\bf q}_B n_3{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
This is a convolution-type reciprocity theorem
\citep{Fokkema93Book, Hoop95Book, Achenbach2003Book},
because products like ${\bf q}_A^t({\bf x},\omega){\bf N}{\bf q}_B({\bf x},\omega)$ in the frequency domain correspond to convolutions in the time domain.
A more familiar form is obtained by substituting the expressions for ${\bf q}$, ${\bf d}$ and ${\bf N}$ (equations \ref{eq2.2q} and \ref{eq4.3NK}), choosing $C_j=0$ and using equation
(\ref{eqA2}) to eliminate $Q_3$, which gives
\begin{eqnarray}\label{eq4.1comp}
&&\int_\mathbb{D}(-B_AP_B+P_A B_B){\rm d}{\bf x}=
\int_{\setdD_{0,1}}\frac{1}{i\omega\beta}\bigl(P_A \partial_3P_B-(\partial_3P_A)P_B\bigr)n_3{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
Next, consider the quantity $\partial_3\bigl({\bf q}_A^\dagger{\bf K}{\bf q}_B\bigr)$ in domain $\mathbb{D}$. Following the same steps as above, using equations (\ref{eq90d}) and (\ref{eqsymad})
instead of (\ref{eq90c}) and (\ref{eqsym}), we obtain
\begin{eqnarray}\label{eq4.2}
&&\int_\mathbb{D}\bigl({\bf d}_A^\dagger{\bf K}{\bf q}_B + {\bf q}_A^\dagger{\bf K}{\bf d}_B\bigr){\rm d}{\bf x}=
\int_{\setdD_{0,1}}{\bf q}_A^\dagger{\bf K}{\bf q}_B n_3{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
This is a correlation-type reciprocity theorem
\citep{Bojarski83JASA},
because products like ${\bf q}_A^\dagger({\bf x},\omega){\bf K}{\bf q}_B({\bf x},\omega)$ in the frequency domain correspond to correlations in the time domain.
Substituting the expressions for ${\bf q}$, ${\bf d}$ and ${\bf K}$ and choosing $C_j=0$ yields the more familiar form
\begin{eqnarray}\label{eq4.2comp}
&&\int_\mathbb{D}(B_A^*P_B+P_A^* B_B){\rm d}{\bf x}=
\int_{\setdD_{0,1}}\frac{1}{i\omega\beta}\bigl(P_A^* \partial_3P_B-(\partial_3P_A)^*P_B\bigr)n_3{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
We obtain a special case by choosing states $A$ and $B$ identical. Dropping the subscripts $A$ and $B$ in equations (\ref{eq4.2}) and (\ref{eq4.2comp})
\rev{and multiplying both sides of these equations by $\frac{1}{4}$,} gives
\begin{eqnarray}\label{eq4.3}
&&\frac{1}{4}\int_\mathbb{D}\bigl({\bf d}^\dagger{\bf K}{\bf q} + {\bf q}^\dagger{\bf K}{\bf d}\bigr){\rm d}{\bf x}=
\frac{1}{4}\int_{\setdD_{0,1}}{\bf q}^\dagger{\bf K}{\bf q} n_3{\rm d}{\bf x}_{\rm L}
\end{eqnarray}
and
\begin{eqnarray}\label{eq4.3comp}
&&\frac{1}{4}\int_\mathbb{D}(B^*P+P^* B){\rm d}{\bf x}=
\frac{1}{4}\int_{\setdD_{0,1}}\frac{1}{i\omega\beta}\bigl(P^* \partial_3P-(\partial_3P)^*P\bigr)n_3{\rm d}{\bf x}_{\rm L},
\end{eqnarray}
respectively. These equations quantify conservation of power
(or, for quantum-mechanical waves, probability).
\section{Decomposed wave equation and its symmetry properties}\label{sec4}
\subsection{General decomposition of the matrix-vector wave equation}
To facilitate the decomposition of the matrix-vector wave equation (equation \ref{eq2.1}),
we recast the operator matrix ${{\mbox{\boldmath ${\cal A}$}}}$ into a somewhat different form.
To this end we introduce an operator ${\cal H}_2$, according to
\begin{eqnarray}
{\cal H}_2&=&-i\omega\sqrt{\beta}{\cal A}_{21}\sqrt{\beta}\nonumber\\
&=& k^2+\sqrt{\beta}\partial_\nu\frac{1}{\beta}\partial_\nu\sqrt{\beta},\label{eq39}
\end{eqnarray}
with operator ${\cal A}_{21}$ defined in equation (\ref{eq2.4}) and wavenumber $k$ in equation (\ref{eqk}).
Operator ${\cal H}_2$ can be rewritten as a Helmholtz operator
\citep{Wapenaar89Book, Hoop92PHD}
\begin{eqnarray}\label{eqAp7}
&&{\cal H}_2= k_s^2 + \partial_\nu\partial_\nu,
\end{eqnarray}
with the scaled wavenumber $k_s$ defined as
\citep{Brekhovskikh60Book}
\begin{eqnarray}
&&k_s^2=k^2-\frac{3(\partial_\nu\beta)(\partial_\nu\beta)}{4\beta^2}+\frac{(\partial_\nu\partial_\nu\beta)}{2\beta}.
\end{eqnarray}
Note that ${\cal H}_2^t={\cal H}_2$ and ${\cal H}_2^\dagger={\cal H}_2$, hence operator ${\cal H}_2$ is symmetric and self-adjoint \rev{and its spectrum is real-valued (with positive and negative eigenvalues)}.
Using equation (\ref{eq39}), we rewrite operator matrix ${{\mbox{\boldmath ${\cal A}$}}}$, defined in equation (\ref{eq2.2}), as
\begin{eqnarray}
&&{{\mbox{\boldmath ${\cal A}$}}}=\begin{pmatrix} 0 & i\omega\beta\\ -\frac{1}{i\omega\sqrt{\beta}}{\cal H}_2\frac{1}{\sqrt{\beta}} & 0\end{pmatrix}.\label{eqAmod}
\end{eqnarray}
Next, we decompose this operator matrix as follows
\begin{eqnarray}\label{eqdecom1}
&&{{\mbox{\boldmath ${\cal A}$}}}={{\mbox{\boldmath ${\cal L}$}}}{{\mbox{\boldmath ${\cal H}$}}}{{\mbox{\boldmath ${\cal L}$}}}^{-1},
\end{eqnarray}
with
\begin{eqnarray}
{{\mbox{\boldmath ${\cal H}$}}}&=&\begin{pmatrix} i{\cal H}_1 & 0 \\ 0 & -i{\cal H}_1\end{pmatrix},\label{eqH1}\\
{{\mbox{\boldmath ${\cal L}$}}}&=&\begin{pmatrix} {\cal L}_1 & {\cal L}_1 \\ {\cal L}_2 & - {\cal L}_2\end{pmatrix},\label{eqL}\\
{{\mbox{\boldmath ${\cal L}$}}}^{-1}&=&\frac{1}{2}\begin{pmatrix} {\cal L}_1^{-1} & {\cal L}_2^{-1} \\ {\cal L}_1^{-1} & - {\cal L}_2^{-1}\end{pmatrix}.\label{eqLi}
\end{eqnarray}
Operators ${\cal H}_1$, ${\cal L}_1$ and ${\cal L}_2$ are pseudo-differential operators
\citep{Fishman84JMP, Fishman93RS, Wapenaar89Book, Hoop92PHD, Hoop94WM, Wapenaar96GJI1, Fishman2000JMP, Thomson2015GJI1, Thomson2015GJI2}.
The decomposition expressed by equation (\ref{eqdecom1}) is not unique, hence, different choices for operators ${\cal H}_1$, ${\cal L}_1$ and ${\cal L}_2$ are possible.
We discuss two of these choices in detail in the next two sections.
Here we derive some general relations that are independent of these choices.
By substituting equations (\ref{eqAmod}) and (\ref{eqH1}) $-$ (\ref{eqLi}) into equation (\ref{eqdecom1}) we obtain the following relations
\begin{eqnarray}
\omega\beta &=& {\cal L}_1{\cal H}_1{\cal L}_2^{-1},\label{eq34}\\
\frac{1}{\omega\sqrt{\beta}}{\cal H}_2\frac{1}{\sqrt{\beta}} &=& {\cal L}_2{\cal H}_1{\cal L}_1^{-1}.\label{eq35}
\end{eqnarray}
We introduce a decomposed field vector ${\bf p}$ and a decomposed source vector ${\bf s}$ via
\begin{eqnarray}
&&{\bf q}={{\mbox{\boldmath ${\cal L}$}}}{\bf p}, \quad {\bf p}={{\mbox{\boldmath ${\cal L}$}}}^{-1}{\bf q},\label{eqdecom2}\\
&&{\bf d}={{\mbox{\boldmath ${\cal L}$}}}{\bf s}, \quad {\bf s}={{\mbox{\boldmath ${\cal L}$}}}^{-1}{\bf d},\label{eqdecom3}
\end{eqnarray}
where
\begin{eqnarray}
&& {\bf p}=\begin{pmatrix} P^+ \\ P^- \end{pmatrix},\quad {\bf s}=\begin{pmatrix} S^+ \\ S^- \end{pmatrix}.\label{eqdecom4}
\end{eqnarray}
Substitution of equations (\ref{eqdecom1}), (\ref{eqdecom2}) and (\ref{eqdecom3}) into the matrix-vector wave equation (\ref{eq2.1}) yields
\begin{eqnarray}
&&\partial_3{\bf p}=\bigl({{\mbox{\boldmath ${\cal H}$}}}-{{\mbox{\boldmath ${\cal L}$}}}^{-1}\partial_3{{\mbox{\boldmath ${\cal L}$}}}\bigr){\bf p}+{\bf s}.\label{eqoneway}
\end{eqnarray}
Substituting equations (\ref{eqH1}) $-$ (\ref{eqLi}) and (\ref{eqdecom4}) into equation (\ref{eqoneway}) gives
\begin{eqnarray}
&&\hspace{-1.1cm}\partial_3\begin{pmatrix} P^+ \\ P^- \end{pmatrix}=\begin{pmatrix} i{\cal H}_1 & 0\\ 0 & -i{\cal H}_1\end{pmatrix}\begin{pmatrix} P^+ \\ P^- \end{pmatrix}
-\frac{1}{2}\begin{pmatrix} {\cal L}_1^{-1} & {\cal L}_2^{-1} \\ {\cal L}_1^{-1} & - {\cal L}_2^{-1}\end{pmatrix}
\begin{pmatrix} \partial_3{\cal L}_1 & \partial_3{\cal L}_1 \\ \partial_3{\cal L}_2 & - \partial_3{\cal L}_2\end{pmatrix}
\begin{pmatrix} P^+ \\ P^- \end{pmatrix}+\begin{pmatrix} S^+ \\ S^- \end{pmatrix}.\label{eqAAA26}
\end{eqnarray}
This is a system of coupled one-way wave equations. From the first term on the right-hand side it follows that the one-way wave fields
$P^+$ and $P^-$ propagate in the positive and negative $x_3$-direction, respectively. The second term on the right-hand side accounts for coupling between $P^+$ and $P^-$.
The last term on the right-hand side contains sources $S^+$ and $S^-$ which emit waves in the positive and negative $x_3$-direction, respectively.
We conclude this section by substituting equations (\ref{eqdecom2}) and (\ref{eqdecom3}) into equations (\ref{eq4.1}), (\ref{eq4.2}) and (\ref{eq4.3}).
Using equations (\ref{eq90c}) and (\ref{eq90d}) for the integration along
the lateral coordinates this yields
\begin{eqnarray}
\int_\mathbb{D}\bigl({\bf s}_A^t{{\mbox{\boldmath ${\cal L}$}}}^t{\bf N}{{\mbox{\boldmath ${\cal L}$}}}{\bf p}_B +{\bf p}_A^t{{\mbox{\boldmath ${\cal L}$}}}^t{\bf N}{{\mbox{\boldmath ${\cal L}$}}}{\bf s}_B \bigr){\rm d}{\bf x}&=&\int_{\setdD_{0,1}}{\bf p}_A^t{{\mbox{\boldmath ${\cal L}$}}}^t{\bf N}{{\mbox{\boldmath ${\cal L}$}}}{\bf p}_B n_3{\rm d}{\bf x}_{\rm L},\label{eq41}\\
\int_\mathbb{D}\bigl({\bf s}_A^\dagger{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}{\bf p}_B +{\bf p}_A^\dagger{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}{\bf s}_B \bigr){\rm d}{\bf x}&=&\int_{\setdD_{0,1}}{\bf p}_A^\dagger{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}{\bf p}_B n_3{\rm d}{\bf x}_{\rm L},\label{eq42}\\
\frac{1}{4}\int_\mathbb{D}\bigl({\bf s}^\dagger{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}{\bf p} +{\bf p}^\dagger{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}{\bf s} \bigr){\rm d}{\bf x}&=&\frac{1}{4}\int_{\setdD_{0,1}}{\bf p}^\dagger{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}{\bf p} n_3{\rm d}{\bf x}_{\rm L}.\label{eq43}
\end{eqnarray}
These equations form the basis for reciprocity theorems for the decomposed field and source vectors ${\bf p}$ and ${\bf s}$ in the next two sections.
\subsection{Flux-normalised decomposition and reciprocity theorems}
The first choice of operators ${\cal H}_1$, ${\cal L}_1$ and ${\cal L}_2$ obeying equations (\ref{eq34}) and (\ref{eq35}) is
\citep{Hoop92PHD, Hoop96JMP, Wapenaar96GJI1}
\begin{eqnarray}
{\cal H}_1&=&{\cal H}_2^{\frac{1}{2}},\\
{\cal L}_1&=&(\omega/2)^\frac{1}{2}\beta^{\frac{1}{2}}{\cal H}_1^{-\frac{1}{2}},\label{eqB19q}\\
{\cal L}_2&=&(2\omega)^{-\frac{1}{2}}\beta^{-\frac{1}{2}}{\cal H}_1^{\frac{1}{2}}.\label{eqB20q}
\end{eqnarray}
Operator ${\cal H}_1$, which is the square root of the Helmholtz operator ${\cal H}_2$, is commonly known as the square-root operator
\citep{Claerbout71GEO, Berkhout82Book, Fishman84JMP}.
Like the Helmholtz operator ${\cal H}_2$, the square-root operator ${\cal H}_1$ is a symmetric operator
\citep{Wapenaar96GJI1},
hence ${\cal H}_1^t={\cal H}_1$.
For the adjoint square-root operator we have ${\cal H}_1^\dagger=({\cal H}_1^t)^*={\cal H}_1^*$. The spectrum of ${\cal H}_1$ is real-valued for propagating waves and imaginary-valued for evanescent waves.
Hence, unlike the Helmholtz operator, the square-root operator is not self-adjoint. If we neglect evanescent waves, we may approximate the adjoint
square-root operator as ${\cal H}_1^\dagger\approx{\cal H}_1$. Similar relations hold for the square root of the square-root operator and its inverse, hence
$\bigl({\cal H}_1^{\pm \frac{1}{2}}\bigr)^t={\cal H}_1^{\pm \frac{1}{2}}$ and, neglecting evanescent waves, $\bigl({\cal H}_1^{\pm \frac{1}{2}}\bigr)^\dagger\approx{\cal H}_1^{\pm \frac{1}{2}}$.
From here onward we replace $\approx$ by $=$ when the only approximation is the negligence of evanescent waves.
Using these symmetry relations for ${\cal H}_1$ and equations (\ref{eq4.3NK}), (\ref{eqL}), (\ref{eqB19q}) and (\ref{eqB20q}), we obtain
\begin{eqnarray}
&&{{\mbox{\boldmath ${\cal L}$}}}^t{\bf N}{{\mbox{\boldmath ${\cal L}$}}}=-{\bf N}\label{eqA32f}
\end{eqnarray}
and, neglecting evanescent waves,
\begin{eqnarray}
&&{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}={\bf J},\label{eqA33f}
\end{eqnarray}
with
\begin{eqnarray}\label{eq4.3J}
&&{{\bf J}}=\begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}.
\end{eqnarray}
Hence, equations (\ref{eq41}) $-$ (\ref{eq43}) simplify to
\begin{eqnarray}
-\int_\mathbb{D}\bigl({\bf s}_A^t{\bf N}{\bf p}_B +{\bf p}_A^t{\bf N}{\bf s}_B \bigr){\rm d}{\bf x}&=&-\int_{\setdD_{0,1}}{\bf p}_A^t{\bf N}{\bf p}_B n_3{\rm d}{\bf x}_{\rm L},\\
\int_\mathbb{D}\bigl({\bf s}_A^\dagger{\bf J}{\bf p}_B +{\bf p}_A^\dagger{\bf J}{\bf s}_B \bigr){\rm d}{\bf x}&=&\int_{\setdD_{0,1}}{\bf p}_A^\dagger{\bf J}{\bf p}_B n_3{\rm d}{\bf x}_{\rm L},\\
\frac{1}{4}\int_\mathbb{D}\bigl({\bf s}^\dagger{\bf J}{\bf p} +{\bf p}^\dagger{\bf J}{\bf s} \bigr){\rm d}{\bf x}&=&\frac{1}{4}\int_{\setdD_{0,1}}{\bf p}^\dagger{\bf J}{\bf p} n_3{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
By substituting the expressions for ${\bf p}$, ${\bf s}$, ${\bf N}$ and ${\bf J}$ (equations \ref{eqdecom4}, \ref{eq4.3NK} and \ref{eq4.3J}), we obtain
\begin{eqnarray}
\int_\mathbb{D}(-S_A^+P_B^- + S_A^-P_B^+-P_A^+S_B^- +P_A^-S_B^+){\rm d}{\bf x}&=&
\int_{\setdD_{0,1}}(-P_A^+P_B^- +P_A^-P_B^+)n_3{\rm d}{\bf x}_{\rm L},\label{eq4.1deccomp}\\
\int_\mathbb{D}(S_A^{+*}P_B^+ -S_A^{-*}P_B^- + P_A^{+*}S_B^+ -P_A^{-*}S_B^-){\rm d}{\bf x}&=&
\int_{\setdD_{0,1}}(P_A^{+*}P_B^+ -P_A^{-*}P_B^-)n_3{\rm d}{\bf x}_{\rm L},\label{eq4.2deccomp}\\
\frac{1}{4}\int_\mathbb{D}(S^{+*}P^+ -S^{-*}P^- + P^{+*}S^+ -P^{-*}S^-){\rm d}{\bf x}
&=&\frac{1}{4}\int_{\setdD_{0,1}}(|P^+|^2 -|P^-|^2)n_3{\rm d}{\bf x}_{\rm L}.\label{eq4.3deccomp}
\end{eqnarray}
Note that, since the right-hand side of equation (\ref{eq4.3deccomp}) is equal to the right-hand side of equation (\ref{eq4.3comp}), it quantifies the power flux
(or the probability-flux for quantum-mechanical waves)
through the surface ${\setdD_{0,1}}$. Therefore we call $P^+$ and $P^-$ flux-normalised one-way wave fields.
Consequently, equations (\ref{eq4.1deccomp}) and (\ref{eq4.2deccomp}) are reciprocity theorems of the convolution type and correlation type,
respectively, for flux-normalised one-way wave fields.
These theorems have been derived previously
\citep{Wapenaar96GJI1}
and have found applications in advanced wave field imaging methods for active and passive data
\citep{Wapenaar2004GJI, Kumar2006JGR, Fan2006GEO, Weglein2006GEO, Slob2009IEEE, Ruigrok2012GRL, Wapenaar2014JASA, Slob2014GEO, Ravasi2016GJI, Ravasi2017GEO, Elison2018GJI, Staring2018GEO}.
\subsection{Field-normalised decomposition and reciprocity theorems}\label{sec3c}
The second choice of operators ${\cal H}_1$, ${\cal L}_1$ and ${\cal L}_2$ obeying equations (\ref{eq34}) and (\ref{eq35}) is
\citep{Wapenaar89Book}
\begin{eqnarray}
{\cal H}_1&=&\beta^\frac{1}{2}{\cal H}_2^{\frac{1}{2}}\beta^{-\frac{1}{2}},\label{eqB18qf}\\
{\cal L}_1&=&1,\label{eqB19qf}\\
{\cal L}_2&=&(\omega\beta)^{-1}{\cal H}_1.\label{eqB20qf}
\end{eqnarray}
Only the Helmholtz operator ${\cal H}_2$ is the same as in the previous section (it is defined in equation \ref{eqAp7}).
The operators ${\cal H}_1$, ${\cal L}_1$ and ${\cal L}_2$ are different from those in the previous section, but for convenience we use the same symbols.
Using ${\bf q}={{\mbox{\boldmath ${\cal L}$}}}{\bf p}$ (equation \ref{eqdecom2}) and equations (\ref{eq2.2q}), (\ref{eqL}), (\ref{eqdecom4}) and (\ref{eqB19qf}), we find
\begin{eqnarray}
&&P=P^+ + P^-,\label{eq59}
\end{eqnarray}
hence, $P^+$ and $P^-$ have the same physical dimension as the full field variable $P$ (which is defined in Table 1 for different wave phenomena).
Therefore we call $P^+$ and $P^-$ field-normalised one-way wave fields (for convenience we use the same symbols as in the previous section).
The square-root operator ${\cal H}_2^\frac{1}{2}$ is symmetric, but ${\cal H}_1$ defined in equation (\ref{eqB18qf}) is not.
From this equation it easily follows that ${\cal H}_1$, premultiplied by $\beta^{-1}$ is symmetric, hence
\begin{eqnarray}
\biggl(\frac{1}{\beta}{\cal H}_1\biggr)^t &=&\frac{1}{\beta} {\cal H}_1
\end{eqnarray}
and, neglecting evanescent waves,
\begin{eqnarray}
\biggl(\frac{1}{\beta}{\cal H}_1\biggr)^\dagger &=&
\frac{1}{\beta} {\cal H}_1.
\end{eqnarray}
Using these symmetry relations for $\frac{1}{\beta}{\cal H}_1$ and equations (\ref{eq4.3NK}), (\ref{eqL}), (\ref{eqB19qf}) and (\ref{eqB20qf}), we obtain
\begin{eqnarray}
&&{{\mbox{\boldmath ${\cal L}$}}}^t{\bf N}{{\mbox{\boldmath ${\cal L}$}}}=
\begin{pmatrix} 0 & -2{\cal L}_2 \\ 2{\cal L}_2 & 0 \end{pmatrix}=-{\bf N}\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)=-\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)^t{\bf N}
\end{eqnarray}
and, neglecting evanescent waves,
\begin{eqnarray}
&&{{\mbox{\boldmath ${\cal L}$}}}^\dagger{\bf K}{{\mbox{\boldmath ${\cal L}$}}}=
\begin{pmatrix} 2{\cal L}_2 & 0 \\ 0 & -2{\cal L}_2 \end{pmatrix}={\bf J}\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)=\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)^\dagger{\bf J}.
\end{eqnarray}
Using this in equations (\ref{eq41}) and (\ref{eq42}) yields
\begin{eqnarray}
-\int_\mathbb{D}\biggl[{\bf s}_A^t\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)^t{\bf N}{\bf p}_B +{\bf p}_A^t{\bf N}\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr){\bf s}_B \biggr]{\rm d}{\bf x}
&=&-\int_{\setdD_{0,1}}{\bf p}_A^t\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)^t{\bf N}{\bf p}_B n_3{\rm d}{\bf x}_{\rm L},\\
\int_\mathbb{D}\biggl[{\bf s}_A^\dagger\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)^\dagger{\bf J}{\bf p}_B +{\bf p}_A^\dagger{\bf J}\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr){\bf s}_B \biggr]{\rm d}{\bf x}
&=&\int_{\setdD_{0,1}}{\bf p}_A^\dagger\biggl(\frac{2}{\omega\beta}{\cal H}_1\biggr)^\dagger{\bf J}{\bf p}_B n_3{\rm d}{\bf x}_{\rm L}.
\end{eqnarray}
By substituting the expressions for ${\bf p}$, ${\bf s}$, ${\bf N}$ and ${\bf J}$ (equations \ref{eqdecom4}, \ref{eq4.3NK} and \ref{eq4.3J}), using equations (\ref{eq90c}) and (\ref{eq90d}), we obtain
\begin{eqnarray}
&&-\int_\mathbb{D}\frac{2}{\omega\beta}\bigl(({\cal H}_1S_A^+)P_B^- - ({\cal H}_1S_A^-)P_B^+ + P_A^+({\cal H}_1S_B^-) -P_A^-({\cal H}_1S_B^+) \bigr){\rm d}{\bf x}\nonumber\\
&&\hspace{6cm}=-\int_{\setdD_{0,1}}\frac{2}{\omega\beta}\bigl(({\cal H}_1P_A^+)P_B^- - ({\cal H}_1P_A^-)P_B^+\bigr) n_3{\rm d}{\bf x}_{\rm L},\label{eq66}\\
&&\int_\mathbb{D}\frac{2}{\omega\beta}\bigl(({\cal H}_1S_A^+)^*P_B^+ - ({\cal H}_1S_A^-)^*P_B^- + P_A^{+*}({\cal H}_1S_B^+) -P_A^{-*}({\cal H}_1S_B^-) \bigr){\rm d}{\bf x}\nonumber\\
&&\hspace{6cm}=\int_{\setdD_{0,1}}\frac{2}{\omega\beta}\bigl(({\cal H}_1P_A^+)^*P_B^+ - ({\cal H}_1P_A^-)^*P_B^-\bigr) n_3{\rm d}{\bf x}_{\rm L}.\label{eq67}
\end{eqnarray}
We aim to remove the operator ${\cal H}_1$ from these equations.
From equation (\ref{eqAAA26}) and (\ref{eqB19qf}) we obtain
\begin{eqnarray}
\partial_3P^+&=&+i{\cal H}_1P^+ - \frac{1}{2}({\cal L}_2^{-1}\partial_3{\cal L}_2)(P^+ - P^-)+S^+,\label{eqAAA26b}\\
\partial_3P^-&=&-i{\cal H}_1P^- + \frac{1}{2}({\cal L}_2^{-1}\partial_3{\cal L}_2)(P^+ - P^-)+S^-,\label{eqAAA26z}
\end{eqnarray}
with ${\cal L}_2$ defined in equation (\ref{eqB20qf}).
Assuming that in state $A$ the derivatives in the $x_3$-direction of the parameters $\alpha$ and $\beta$ at ${{\setdD_{0,1}}}$ vanish and there are no sources at ${\setdD_{0,1}}$,
we find from equations (\ref{eqAAA26b}) and (\ref{eqAAA26z})
\begin{eqnarray}
&&\partial_3P_A^\pm=\pmi{\cal H}_1P_A^\pm\quad\mbox{at} \,\,\,{{\setdD_{0,1}}}.\label{eqAA46}
\end{eqnarray}
Below we use this to remove ${\cal H}_1$ from the right-hand sides of equations (\ref{eq66}) and (\ref{eq67}). Next, we aim to remove ${\cal H}_1$ from the left-hand sides of these
equations.
From ${\bf s}={{\mbox{\boldmath ${\cal L}$}}}^{-1}{\bf d}$ (equation \ref{eqdecom3}) and equations (\ref{eq2.2q}), (\ref{eqLi}), (\ref{eqdecom4}), (\ref{eqB19qf}) and (\ref{eqB20qf}), we find
\begin{eqnarray}
&&S^\pm= \pm \frac{1}{2}\Bigl(\frac{1}{\omega\beta}{\cal H}_1\Bigr)^{-1}B_0+ \frac{1}{2} C_3,
\end{eqnarray}
or
\begin{eqnarray}
&&\pm\frac{2}{\omega\beta}{\cal H}_1S^\pm= B_0\pm\frac{1}{\omega\beta}{\cal H}_1C_3.
\end{eqnarray}
We define new decomposed sources $B_0^+$ and $B_0^-$, according to
\begin{eqnarray}
&&B_0^\pm=B_0\pm\frac{1}{\omega\beta}{\cal H}_1C_3=\pm\frac{2}{\omega\beta}{\cal H}_1S^\pm.\label{eqA33b}
\end{eqnarray}
Using equations (\ref{eqAA46}) and (\ref{eqA33b}) in the right- and left-hand sides of equations (\ref{eq66}) and (\ref{eq67}), we obtain
\begin{eqnarray}
\int_\mathbb{D}\bigl(-B_{0,A}^+P_B^- - B_{0,A}^-P_B^+ + P_A^+B_{0,B}^- + P_A^-B_{0,B}^+\bigr){\rm d}{\bf x}
&=&\int_{\setdD_{0,1}}\frac{-2}{i\omega\beta}\bigl((\partial_3P_A^+)P_B^- + (\partial_3P_A^-)P_B^+\bigr)n_3{\rm d}{\bf x}_{\rm L},\nonumber\\
&&\label{eq94}\\
\int_\mathbb{D}\bigl(B_{0,A}^{+*}P_B^+ + B_{0,A}^{-*}P_B^- + P_A^{+*}B_{0,B}^+ + P_A^{-*}B_{0,B}^-\bigr){\rm d}{\bf x}
&=&\int_{\setdD_{0,1}}\frac{-2}{i\omega\beta}\bigl((\partial_3P_A^+)^*P_B^+ + (\partial_3P_A^-)^*P_B^-\bigr)n_3{\rm d}{\bf x}_{\rm L}.\nonumber\\
&&\label{eq95}
\end{eqnarray}
Equations (\ref{eq94}) and (\ref{eq95}) are reciprocity theorems of the convolution type and correlation type, respectively, for field-normalised one-way wave fields.
These theorems are modifications of previously obtained results
\citep{Berkhout89GEO, Wapenaar89GEO}.
The main modification is that we applied decomposition at both sides of the equations instead of at the right-hand sides only.
Moreover, in the present derivation the condition for the validity of equation (\ref{eqAA46}) is only imposed for state $A$.
In the next section we use equations (\ref{eq94}) and (\ref{eq95}) to derive representation theorems for field-normalised one-way wave fields and we indicate applications.
\section{Field-normalised representation theorems}\label{sec5}
\subsection{Green's functions}
Representation theorems are obtained by substituting Green's functions in reciprocity theorems. Our aim is to introduce one-way Green's functions, to be used in
the reciprocity theorems for field-normalised one-way wave fields (equations \ref{eq94} and \ref{eq95}).
First, we introduce the full Green's function $G({\bf x},\bx_A,\omega)$ as a solution of the unified wave equation (\ref{eqwe})
for a unit monopole point source at $\bx_A$, with $B({\bf x},\omega)= \delta({\bf x}-\bx_A)$ and $C_j({\bf x},\omega)=0$. Hence,
\begin{eqnarray}\label{eq6}
&&\beta\partial_j\Bigl(\frac{1}{\beta}\partial_j G\Bigr) +k^2G=i\omega\beta \delta({\bf x}-\bx_A).
\end{eqnarray}
As boundary condition we impose the radiation condition (i.e., outward propagating waves at infinity).
Next, we introduce one-way Green's function as solutions of the coupled one-way equations (\ref{eqAAA26b}) and (\ref{eqAAA26z}) for a unit monopole point source at $\bx_A$.
Hence, we choose again $B({\bf x},\omega)= \delta({\bf x}-\bx_A)$ and $C_j({\bf x},\omega)=0$.
Using equations (\ref{eqA33b}) and (\ref{eq2.5}), we define decomposed sources as $B_0^\pm=B^\pm=B=\pm 2 {\cal L}_2S^\pm$, with ${\cal L}_2$ defined in equation (\ref{eqB20qf}), or
\begin{eqnarray}
&&S^\pm({\bf x},\omega)=\pm\frac{1}{2}{\cal L}_2^{-1}B^\pm({\bf x},\omega)=\pm\frac{1}{2}{\cal L}_2^{-1}B({\bf x},\omega)=\pm\frac{1}{2}{\cal L}_2^{-1}\delta({\bf x}-\bx_A).
\end{eqnarray}
We consider two sets of one-way Green's functions.
For the first set we choose a point source $S^+({\bf x},\omega) =\frac{1}{2}{\cal L}_2^{-1}B^+({\bf x},\omega)$, with $B^+({\bf x},\omega)=\delta({\bf x}-\bx_A)$,
which emits waves from $\bx_A$ in the positive $x_3$-direction,
and we set $S^-({\bf x},\omega) $ equal to zero.
Hence, for this first set, one-way equations (\ref{eqAAA26b}) and (\ref{eqAAA26z}) become
\begin{eqnarray}
\partial_3G^{+,+}&=&+i{\cal H}_1G^{+,+} - \frac{1}{2}({\cal L}_2^{-1}\partial_3{\cal L}_2)(G^{+,+} - G^{-,+})+\frac{1}{2}{\cal L}_2^{-1}\delta({\bf x}-\bx_A),\label{eqAAA26G1}\\
\partial_3G^{-,+}&=&-i{\cal H}_1G^{-,+} + \frac{1}{2}({\cal L}_2^{-1}\partial_3{\cal L}_2)(G^{+,+} - G^{-,+}).\label{eqAAA26G2}
\end{eqnarray}
Here $G^{\pm,+}$ stands for $G^{\pm,+}({\bf x},\bx_A,\omega)$. The second superscript ($+$) indicates that the source at $\bx_A$ emits waves in the positive $x_3$-direction.
The first superscript ($\pm$) denotes the propagation direction at ${\bf x}$.
For the second set of one-way Green's functions we choose a point source
$S^-({\bf x},\omega) =-\frac{1}{2}{\cal L}_2^{-1}B^-({\bf x},\omega)$, with $B^-({\bf x},\omega)=\delta({\bf x}-\bx_A)$, which emits waves from $\bx_A$ in the negative $x_3$-direction,
and we set $S^+({\bf x},\omega) $ equal to zero.
Hence, for this second set, one-way equations (\ref{eqAAA26b}) and (\ref{eqAAA26z}) become
\begin{eqnarray}
\partial_3G^{+,-}&=&+i{\cal H}_1G^{+,-} - \frac{1}{2}({\cal L}_2^{-1}\partial_3{\cal L}_2)(G^{+,-} - G^{-,-}),\label{eqAAA26G3}\\
\partial_3G^{-,-}&=&-i{\cal H}_1G^{-,-} + \frac{1}{2}({\cal L}_2^{-1}\partial_3{\cal L}_2)(G^{+,-} - G^{-,-})-\frac{1}{2}{\cal L}_2^{-1}\delta({\bf x}-\bx_A).\label{eqAAA26G4}
\end{eqnarray}
Here $G^{\pm,-}$ stands for $G^{\pm,-}({\bf x},\bx_A,\omega)$, with the second superscript ($-$) indicating that the source at $\bx_A$ emits waves in the negative $x_3$-direction.
Like for the full Green's function $G({\bf x},\bx_A,\omega)$, we impose radiation conditions for both sets of one-way Green's functions.
To find a relation between the full Green's function and the one-way Green's functions,
we evaluate $\beta\partial_3\frac{1}{\beta}\partial_3(G^{+,+}+G^{-,+}+G^{+,-}+G^{-,-})$ using
equations (\ref{eqAAA26G1}) $-$ (\ref{eqAAA26G4}),
(\ref{eq39}), (\ref{eqB18qf}) and (\ref{eqB20qf}). This gives equation (\ref{eq6}), with $G$ replaced by $G^{+,+}+G^{-,+}+G^{+,-}+G^{-,-}$.
Since the full Green's function and the one-way Green's functions obey the same radiation conditions, we thus find
\begin{eqnarray}
&&G=G^{+,+}+G^{-,+}+G^{+,-}+G^{-,-}.\label{eq106}
\end{eqnarray}
This very simple relation is a consequence of the field-normalised decomposition, introduced in section \ref{sec3c}.
\subsection{Source-receiver reciprocity}
We derive source-receiver reciprocity relations for the field-normalised one-way Green's functions introduced in the previous section.
To this end we make use of the reciprocity theorem of the convolution type for field-normalised one-way wave fields,
equation (\ref{eq94}). This theorem was derived for the configuration of Figure \ref{Fig1}, assuming that
in domain $\mathbb{D}$, the parameters $\alpha$ and $\beta$ are the same in the two states (see section \ref{sec3d}).
Outside $\mathbb{D}$ these parameters may be different in the two states. For the Green's state we choose the parameters \rev{for $x_3\le x_{3,0}$ and for $x_3\ge x_{3,1}$}
independent of the $x_3$-coordinate, according to $\alpha({\bf x}_{\rm L})$ and $\beta({\bf x}_{\rm L})$.
Hence, if we let the Green's state (with a point source at $\bx_A$ in $\mathbb{D}$) take the role of state $A$, then the condition for the validity of equation (\ref{eqAA46}) is fulfilled.
Moreover, the Green's functions are purely outward propagating at ${\setdD_{0,1}}$ (because outside $\mathbb{D}$ no scattering occurs along the $x_3$-coordinate).
Hence, $G^{+,\pm}({\bf x},\bx_A,\omega)=0$ at ${\partial\setD}_0$ and $G^{-,\pm}({\bf x},\bx_A,\omega)=0$ at ${\partial\setD}_1$.
We let a second Green's state (with a point source at $\bx_B$ in $\mathbb{D}$ and the same parameters $\alpha$ and $\beta$ \rev{as in state $A$}, inside as well as outside $\mathbb{D}$) take the role of state $B$.
Hence, $G^{+,\pm}({\bf x},\bx_B,\omega)=0$ at ${\partial\setD}_0$ and $G^{-,\pm}({\bf x},\bx_B,\omega)=0$ at ${\partial\setD}_1$.
With only outward propagating waves at ${\setdD_{0,1}}$, the surface integral on the right-hand side of equation (\ref{eq94}) vanishes.
Hence, taking into account that $B_0^\pm=B^\pm$ (since $C_j=0$), equation (\ref{eq94}) simplifies to
\begin{eqnarray}\label{eq4.1g}
&&\int_{\mathbb{D}}\bigl (-B_A^+P_B^- - B_A^-P_B^+ + P_A^+B_B^- + P_A^-B_B^+\bigr){\rm d}{\bf x}=0.
\end{eqnarray}
First, we consider sources emitting waves in the positive $x_3$-direction in both Green's states, hence $B_A^+=\delta({\bf x}-{\bf x}_A)$, $B_A^-=0$, $P_A^\pm=G^{\pm,+}({\bf x},{\bf x}_A,\omega)$, $B_B^+=\delta({\bf x}-{\bf x}_B)$, $B_B^-=0$ and $P_B^\pm=G^{\pm,+}({\bf x},{\bf x}_B,\omega)$. Substituting this into equation (\ref{eq4.1g}) yields
\begin{eqnarray}
&&G^{-,+}({\bf x}_B,{\bf x}_A,\omega)=G^{-,+}({\bf x}_A,{\bf x}_B,\omega),\label{eqA50}
\end{eqnarray}
see Figure \ref{Fig2}(a).
Next, we replace the source in state $B$ by one emitting waves in the negative $x_3$-direction, hence
$B_B^+=0$, $B_B^-=\delta({\bf x}-{\bf x}_B)$ and $P_B^\pm=G^{\pm,-}({\bf x},{\bf x}_B,\omega)$.
This gives
\begin{eqnarray}
&&G^{+,+}({\bf x}_B,{\bf x}_A,\omega)=G^{-,-}({\bf x}_A,{\bf x}_B,\omega),\label{eqA51}
\end{eqnarray}
see Figure \ref{Fig2}(b).
By replacing also the source in state $A$ by one emitting waves in the negative $x_3$-direction, according to
$B_A^+=0$, $B_A^-=\delta({\bf x}-{\bf x}_A)$ and $P_A^\pm=G^{\pm,-}({\bf x},{\bf x}_A,\omega)$,
we obtain
\begin{eqnarray}
&&G^{+,-}({\bf x}_B,{\bf x}_A,\omega)=G^{+,-}({\bf x}_A,{\bf x}_B,\omega),\label{eqA52}
\end{eqnarray}
see Figure \ref{Fig2}(c). Finally, changing the source in state $B$ back to the one emitting waves in the positive $x_3$-direction yields
\begin{eqnarray}
&&G^{-,-}({\bf x}_B,{\bf x}_A,\omega)=G^{+,+}({\bf x}_A,{\bf x}_B,\omega),\label{eqA53}
\end{eqnarray}
see Figure \ref{Fig2}(d).
Source-receiver reciprocity relations similar to equations (\ref{eqA50}) $-$ (\ref{eqA53})
were previously derived for flux-normalised one-way Green's functions
\citep{Wapenaar96GJI2},
except that two of those relations involve a change of sign when interchanging the source and the receiver.
The absence of sign changes in equations (\ref{eqA50}) $-$ (\ref{eqA53})
is due to the definition of $B_0^\pm$ in equation (\ref{eqA33b}).
Moreover, unlike the flux-normalised reciprocity relations, the field-normalised source-receiver reciprocity relations of equations
(\ref{eqA50}) $-$ (\ref{eqA53}) have a very straightforward relation with the
well-known source-receiver reciprocity relation for the full Green's function.
By separately summing the left- and right-hand sides of equations (\ref{eqA50}) $-$ (\ref{eqA53}) and using equation (\ref{eq106}),
we simply obtain
\begin{eqnarray}
&&G({\bf x}_B,{\bf x}_A,\omega)=G({\bf x}_A,{\bf x}_B,\omega).\label{eqA54}
\end{eqnarray}
\begin{figure}
\vspace{0cm}
\centerline{\epsfysize=7.5 cm \epsfbox{Figure2.pdf}}
\caption{\footnotesize Visualisation of the source-receiver reciprocity relations for the field-normalised one-way Green's functions,
formulated by equations (\ref{eqA50}) $-$ (\ref{eqA53}). The ``rays'' in this and subsequent figures are strong simplifications of the complete one-way wave fields,
which include primary and multiple scattering.
}\label{Fig2}
\end{figure}
\subsection{Kirchhoff-Helmholtz integrals for forward propagation}
We derive Kirchhoff-Helmholtz integrals of the convolution type for field-normalised one-way wave fields.
For state $B$ we consider the decomposed actual field, with sources only outside $\mathbb{D}$, hence,
$B_{0,B}^\pm=0$ in $\mathbb{D}$ and $P_B^\pm=P^\pm({\bf x},\omega)$.
The parameters $\alpha$ and $\beta$ are the actual parameters inside as well as outside $\mathbb{D}$.
For state $A$ we choose the Green's state with a unit point source at $\bx_A$ in $\mathbb{D}$.
The parameters $\alpha$ and $\beta$ in $\mathbb{D}$ are the same as those in state $B$, but \rev{for $x_3\le x_{3,0}$ and for $x_3\ge x_{3,1}$} they are chosen independent of the $x_3$-coordinate.
Hence, the condition for the validity of equation (\ref{eqAA46}) is fulfilled.
First, we consider a source in state $A$ which emits waves in the positive $x_3$-direction, hence
$B_A^+=\delta({\bf x}-\bx_A)$, $B_A^-=0$ and $P_A^\pm=G^{\pm,+}({\bf x},\bx_A,\omega)$.
Substituting all this into equation (\ref{eq94}) (with $B_{0,A}^\pm=B_A^\pm$) gives
\begin{eqnarray}
&&P^-(\bx_A,\omega)=
\int_{\setdD_{0,1}}\frac{2}{i\omega\beta({\bf x})}\bigl((\partial_3G^{+,+} ({\bf x},\bx_A,\omega))P^-({\bf x},\omega)
+ (\partial_3G^{-,+} ({\bf x},\bx_A,\omega))P^+ ({\bf x},\omega)\bigr)n_3{\rm d}{\bf x}_{\rm L}.\nonumber\\
&&\label{eq4.1compqGaa}
\end{eqnarray}
Next, we replace the source in state $A$ by one which emits waves in the negative $x_3$-direction, hence
$B_A^+=0$, $B_A^-=\delta({\bf x}-\bx_A)$ and $P_A^\pm=G^{\pm,-}({\bf x},\bx_A,\omega)$.
Equation (\ref{eq94}) thus gives
\begin{eqnarray}
&&P^+(\bx_A,\omega)=
\int_{\setdD_{0,1}}\frac{2}{i\omega\beta({\bf x})}\bigl((\partial_3G^{+,-} ({\bf x},\bx_A,\omega))P^-({\bf x},\omega)
+ (\partial_3G^{-,-} ({\bf x},\bx_A,\omega))P^+ ({\bf x},\omega)\bigr)n_3{\rm d}{\bf x}_{\rm L}.\nonumber\\
&&\label{eq4.1compqGbb}
\end{eqnarray}
Recall that ${\setdD_{0,1}}$ consists of ${\partial\setD}_0$ (with $n_3=-1$) and ${\partial\setD}_1$ (with $n_3=+1$), see Figure \ref{Fig1}.
Since $G^{+,\pm}({\bf x},\bx_A,\omega)=0$ at ${\partial\setD}_0$ and $G^{-,\pm}({\bf x},\bx_A,\omega)=0$ at ${\partial\setD}_1$
(because outside $\mathbb{D}$ no scattering occurs along the $x_3$-coordinate in state $A$), the first term under the integral in
equations (\ref{eq4.1compqGaa}) and (\ref{eq4.1compqGbb}) gives a contribution only at ${\partial\setD}_1$ and the second term only at ${\partial\setD}_0$.
Hence,
\begin{eqnarray}
P^\pm(\bx_A,\omega)&=&
\int_{{\partial\setD}_0}\frac{-2}{i\omega\beta({\bf x})}(\partial_3G^{-,\mp} ({\bf x},\bx_A,\omega))P^+({\bf x},\omega) {\rm d}{\bf x}_{\rm L}\nonumber\\
&+&\int_{{\partial\setD}_1}\frac{2}{i\omega\beta({\bf x})} (\partial_3G^{+,\mp} ({\bf x},\bx_A,\omega))P^- ({\bf x},\omega)\bigr){\rm d}{\bf x}_{\rm L}.\label{eq4.1compqGcc}
\end{eqnarray}
Note that there is no contribution from $P^- ({\bf x},\omega)$ at ${\partial\setD}_0$ nor from $P^+ ({\bf x},\omega)$ at ${\partial\setD}_1$,
see Figure \ref{Fig3}.
We conclude this section by considering a special case. Suppose the source of the actual field (state $B$) is located at $\bx_B$ in the half-space $x_3<x_{3,0}$.
Then, by taking $x_{3,1}\to\infty$, the field $P^-$ at ${\partial\setD}_1$ vanishes. This leaves the single-sided representation
\begin{eqnarray}
P^\pm(\bx_A,\bx_B,\omega)&=&
\int_{{\partial\setD}_0}\frac{-2}{i\omega\beta({\bf x})}(\partial_3G^{-,\mp} ({\bf x},\bx_A,\omega))P^+({\bf x},\bx_B,\omega) {\rm d}{\bf x}_{\rm L}.\label{eq4.1compqGee}
\end{eqnarray}
Note that we included the source coordinate vector $\bx_B$ in the argument list of $P^\pm(\bx_A,\bx_B,\omega)$.
This representation is an extension of a previously derived result
\citep{Berkhout89GEO},
in which the fields were decomposed at ${\partial\setD}_0$ but not at $\bx_A$.
It describes forward propagation of the one-way field $P^+({\bf x},\bx_B,\omega)$ from the surface ${\partial\setD}_0$ to $\bx_A$ (with $\bx_A$ and $\bx_B$ defined at opposite sides of ${\partial\setD}_0$).
In the following two sections we discuss representations for backward propagation of one-way wave fields.
\begin{figure}
\vspace{0cm}
\centerline{\epsfysize=8 cm \epsfbox{Figure3.pdf}}
\vspace{-1.8cm}
\caption{\footnotesize Visualisation of the different terms in the field-normalised one-way Kirchhoff-Helmholtz integral for forward propagation,
formulated by equation (\ref{eq4.1compqGcc}).
The solid Green's functions contribute to $P^+(\bx_A,\omega)$, the dashed Green's functions to $P^-(\bx_A,\omega)$.
}\label{Fig3}
\end{figure}
\subsection{Kirchhoff-Helmholtz integrals for backward propagation (double-sided)}\label{sec4d}
We derive Kirchhoff-Helmholtz integrals of the correlation type for field-normalised one-way wave fields.
For state $B$ we consider the decomposed actual field, with a point source at $\bx_B$ and source spectrum $s(\omega)$.
The parameters $\alpha$ and $\beta$ are the actual parameters inside as well as outside $\mathbb{D}$.
For state $A$ we choose the Green's state with a unit point source at $\bx_A$ in $\mathbb{D}$.
The parameters $\alpha$ and $\beta$ in $\mathbb{D}$ are the same as those in state $B$, but \rev{for $x_3\le x_{3,0}$ and for $x_3\ge x_{3,1}$} they are chosen independent of the $x_3$-coordinate.
Hence, the condition for the validity of equation (\ref{eqAA46}) is fulfilled.
First, we consider sources emitting waves in the positive $x_3$-direction in both states, hence
$B_A^+=\delta({\bf x}-{\bf x}_A)$, $B_A^-=0$, $P_A^\pm=G^{\pm,+}({\bf x},{\bf x}_A,\omega)$,
$B_B^+=\delta({\bf x}-{\bf x}_B)s(\omega)$, $B_B^-=0$ and $P_B^\pm=P^{\pm,+}({\bf x},{\bf x}_B,\omega)$.
Substituting this into equation (\ref{eq95})
(with $B_{0,A}^\pm=B_A^\pm$ and $B_{0,B}^\pm=B_B^\pm$)
gives
\begin{eqnarray}\label{eq104a-alt}
&&P^{+,+}({\bf x}_A,{\bf x}_B,\omega)+\chi(\bx_B)\{G^{+,+}({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)=\\
&&\int_{\setdD_{0,1}}\frac{-2}{i\omega\beta({\bf x})}\bigl(\{\partial_3G^{+,+} ({\bf x},\bx_A,\omega)\}^*P^{+,+}({\bf x},\bx_B,\omega)
+\{\partial_3G^{-,+} ({\bf x},\bx_A,\omega)\}^*P^{-,+}({\bf x},\bx_B,\omega)\bigr)n_3{\rm d}{\bf x}_{\rm L}\nonumber,
\end{eqnarray}
where $\chi$ is the characteristic function of the domain $\mathbb{D}$. It is defined as
\begin{eqnarray}\label{eqC3.2D}
&&\chi(\bx_B)=
\begin{cases}
1, &\text{for } \bx_B\text{ in }{\mathbb{D}}, \\
\frac{1}{2}, &\text{for } \bx_B\text{ on }{{\setdD_{0,1}}},\\
0, &\text{for } \bx_B\text{ outside }{\mathbb{D}}.
\end{cases}
\end{eqnarray}
Since $G^{+,+}({\bf x},\bx_A,\omega)=0$ at ${\partial\setD}_0$ and $G^{-,+}({\bf x},\bx_A,\omega)=0$ at ${\partial\setD}_1$
(because outside $\mathbb{D}$ no scattering occurs along the $x_3$-coordinate in state $A$), the first term under the integral in
equation (\ref{eq104a-alt}) gives a contribution only at ${\partial\setD}_1$ and the second term only at ${\partial\setD}_0$.
Hence,
\begin{eqnarray}\label{eq104a-altb}
&&P^{+,+}({\bf x}_A,{\bf x}_B,\omega)+\chi(\bx_B)\{G^{+,+}({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)=\\
&&\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{-,+} ({\bf x},\bx_A,\omega)\}^*P^{-,+}({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\nonumber\\
&&-\int_{{\partial\setD}_1}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{+,+} ({\bf x},\bx_A,\omega)\}^*P^{+,+}({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\nonumber.
\end{eqnarray}
Next, we replace the source in state $B$ by one emitting waves in the negative $x_3$-direction, hence
$B_B^+=0$, $B_B^-=\delta({\bf x}-{\bf x}_B)s(\omega)$ and $P_B^\pm=P^{\pm,-}({\bf x},{\bf x}_B,\omega)$.
This gives
\begin{eqnarray}\label{eq104b-altb}
&&P^{+,-}({\bf x}_A,{\bf x}_B,\omega)+\chi(\bx_B)\{G^{-,+}({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)=\\
&&\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{-,+} ({\bf x},\bx_A,\omega)\}^*P^{-,-}({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\nonumber\\
&&-\int_{{\partial\setD}_1}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{+,+} ({\bf x},\bx_A,\omega)\}^*P^{+,-}({\bf x},\bx_B,\omega) {\rm d}{\bf x}_{\rm L}\nonumber.
\end{eqnarray}
By replacing also the source in state $A$ by one emitting waves in the negative $x_3$-direction, according to
$B_A^+=0$, $B_A^-=\delta({\bf x}-{\bf x}_A)$, $P_A^\pm=G^{\pm,-}({\bf x},{\bf x}_A,\omega)$,
we obtain
\begin{eqnarray}\label{eq104d-altb}
&&P^{-,-}({\bf x}_A,{\bf x}_B,\omega)+\chi(\bx_B)\{G^{-,-}({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)=\\
&&\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{-,-} ({\bf x},\bx_A,\omega)\}^*P^{-,-}({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\nonumber\\
&&-\int_{{\partial\setD}_1}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{+,-} ({\bf x},\bx_A,\omega)\}^*P^{+,-}({\bf x},\bx_B,\omega) {\rm d}{\bf x}_{\rm L}\nonumber.
\end{eqnarray}
Finally, changing the source in state $B$ back to the one emitting waves in the positive $x_3$-direction yields
\begin{eqnarray}\label{eq104c-altb}
&&P^{-,+}({\bf x}_A,{\bf x}_B,\omega)+\chi(\bx_B)\{G^{+,-}({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)=\\
&&\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{-,-} ({\bf x},\bx_A,\omega)\}^*P^{-,+}({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\nonumber\\
&&-\int_{{\partial\setD}_1}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^{+,-} ({\bf x},\bx_A,\omega)\}^*P^{+,+}({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\nonumber.
\end{eqnarray}
Equation (\ref{eq104d-altb}) is an extension of a previously derived result
\citep{Wapenaar89GEO},
in which the fields were decomposed at ${\setdD_{0,1}}$ but not at $\bx_A$ and $\bx_B$.
Equations (\ref{eq104a-altb}), (\ref{eq104b-altb}) and (\ref{eq104c-altb}) are further variations.
Equation (\ref{eq104c-altb}) is visualised in Figure \ref{Fig4a}.
Together, these equations describe backward propagation of the one-way wave fields $P^{-,\pm} ({\bf x},\bx_B,\omega)$ from ${\partial\setD}_0$ and $P^{+,\pm} ({\bf x},\bx_B,\omega)$
from ${\partial\setD}_1$ to $\bx_A$.
Except for some special cases, the integrals along ${\partial\setD}_1$ do not vanish by taking $x_{3,1}\to\infty$.
Hence, unlike the forward propagation representation (\ref{eq4.1compqGcc}), the double-sided backward propagation representations (\ref{eq104a-altb}) $-$ (\ref{eq104c-altb})
in general do not simplify to single-sided representations. In the next section we discuss an alternative method to derive single-sided representations for backward propagation.
\begin{figure}
\vspace{0cm}
\centerline{\epsfysize=8 cm \epsfbox{Figure4.pdf}}
\vspace{-2.8cm}
\caption{\footnotesize Visualisation of the different terms in the field-normalised one-way Kirchhoff-Helmholtz integral for backward propagation,
formulated by equation (\ref{eq104c-altb}).
}\label{Fig4a}
\end{figure}
We conclude this section by considering a special case.
Suppose that
in state $B$ the parameters $\alpha$ and $\beta$ are the same as in state $A$ not only in $\mathbb{D}$ but also outside $\mathbb{D}$.
Then $P^{\pm,\pm}({\bf x},\bx_B,\omega)=G^{\pm,\pm}({\bf x},\bx_B,\omega)s(\omega)$
for all ${\bf x}$. Substituting this into representations (\ref{eq104a-altb}) $-$ (\ref{eq104c-altb}),
summing the left- and right-hand sides of these representations separately and dividing both sides by $s(\omega)$,
using equations (\ref{eq106}) and (\ref{eqA54}) and assuming that $\bx_B$ is located in $\mathbb{D}$, we obtain
\begin{eqnarray}\label{eq104acbd-altb}
G_{\rm h}({\bf x}_A,{\bf x}_B,\omega)
&=&\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^- ({\bf x},\bx_A,\omega)\}^*G^-({\bf x},\bx_B,\omega){\rm d}{\bf x}_{\rm L}\\
&-&\int_{{\partial\setD}_1}\frac{2}{i\omega\beta({\bf x})}\{\partial_3G^+ ({\bf x},\bx_A,\omega)\}^*G^+({\bf x},\bx_B,\omega) {\rm d}{\bf x}_{\rm L}\nonumber,
\end{eqnarray}
where the so-called homogeneous Green's function $G_{\rm h}({\bf x}_A,{\bf x}_B,\omega)$ is defined as
\begin{eqnarray}\label{eq100}
&&G_{\rm h}({\bf x}_A,{\bf x}_B,\omega)=G({\bf x}_A,{\bf x}_B,\omega)+G^*({\bf x}_A,{\bf x}_B,\omega)=2\Re \{G({\bf x}_A,{\bf x}_B,\omega)\}
\end{eqnarray}
(with $\Re$ denoting the real part),
and where $G^\pm({\bf x},\bx_A,\omega)=G^{\pm,+}({\bf x},\bx_A,\omega)+G^{\pm,-}({\bf x},\bx_A,\omega)$ (and a similar expression for $G^\pm({\bf x},\bx_B,\omega)$).
Equation (\ref{eq104acbd-altb}) is akin to the well-known representation for the homogeneous Green's function
\citep{Porter70JOSA, Oristaglio89IP},
but with decomposed Green's functions under the integrals.
The simple relation between representations (\ref{eq104a-altb}) $-$ (\ref{eq104c-altb}) on the one hand and the homogeneous Green's function representation (\ref{eq104acbd-altb})
on the other hand is a consequence of the field-normalised decomposition, introduced in section \ref{sec3c}.
\subsection{Kirchhoff-Helmholtz integrals for backward propagation (single-sided)}\label{sec4e}
The complex-conjugated Green's functions $\{\partial_3G^{\pm,\pm}({\bf x},\bx_A,\omega)\}^*$ under the integrals in equations (\ref{eq104a-altb}) $-$ (\ref{eq104c-altb})
can be seen as focusing functions, which focus the wave fields $P^{\pm,\pm}({\bf x},\bx_B,\omega)$ onto a focal point $\bx_A$.
However, this focusing process requires that these wave fields are available at two boundaries ${\partial\setD}_0$ and ${\partial\setD}_1$, enclosing the focal point $\bx_A$.
Here we discuss single-sided field-normalised
focusing functions $f_1^\pm({\bf x},\bx_A,\omega)$ and we use these in modifications of reciprocity theorems (\ref{eq94}) and (\ref{eq95}) to derive
single-sided Kirchhoff-Helmholtz integrals for backward propagation.
\begin{figure}
\vspace{0cm}
\centerline{\epsfysize=8 cm \epsfbox{Figure5.pdf}}
\vspace{-1.8cm}
\caption{\footnotesize
Configuration for the derivation of the single-sided Kirchhoff-Helmholtz integrals for backward propagation.
}\label{Fig4}
\end{figure}
We start by defining a new domain $\mathbb{D}_A$, enclosed by two surfaces ${\partial\setD}_0$ and ${\partial\setD}_A$ perpendicular to the $x_3$-axis at $x_3=x_{3,0}$ and $x_3=x_{3,A}$,
respectively, with $x_{3,A}> x_{3,0}$, see Figure \ref{Fig4}. Hence, ${\partial\setD}_A$ is chosen such that it contains the focal point $\bx_A$.
The two surfaces ${\partial\setD}_0$ and ${\partial\setD}_A$ are together denoted by ${\setdD_{0,A}}$.
The focusing functions $f_1^\pm({\bf x},\bx_A,\omega)$, which will play the role of state $A$ in the reciprocity theorems, obey the one-way wave equations (\ref{eqAAA26b}) and
(\ref{eqAAA26z}) (but without the source terms $S^\pm$), with parameters $\alpha$ and $\beta$ in $\mathbb{D}_A$ equal to those in the actual state $B$,
and independent of the $x_3$-coordinate \rev{for $x_3\le x_{3,0}$ and for $x_3\ge x_{3,A}$}. Hence, the condition for the validity of equation (\ref{eqAA46}) is fulfilled.
Analogous to equation (\ref{eq59}), the field-normalised focusing functions $f_1^\pm({\bf x},\bx_A,\omega)$ are related to the full focusing function $f_1({\bf x},\bx_A,\omega)$, according
to
\begin{eqnarray}\label{eq101}
&&f_1({\bf x},\bx_A,\omega)=f_1^+({\bf x},\bx_A,\omega)+f_1^-({\bf x},\bx_A,\omega).
\end{eqnarray}
The focusing function $f_1^+({\bf x},\bx_A,\omega)$ is incident to the domain $\mathbb{D}_A$ from the half-space $x_3<x_{3,0}$ (see Figure \ref{Fig4}). It propagates and scatters in the
inhomogeneous domain $\mathbb{D}_A$, focuses at $\bx_A$ on surface ${\partial\setD}_A$ and continues as $f_1^+({\bf x},\bx_A,\omega)$ in the half-space $x_3>x_{3,A}$.
The back-scattered field leaves $\mathbb{D}_A$ via surface ${\partial\setD}_0$ and continues as $f_1^-({\bf x},\bx_A,\omega)$ in the half-space $x_3<x_{3,0}$.
The focusing conditions at the focal plane ${\partial\setD}_A$ are
\citep{Wapenaar2014GEO}
\begin{eqnarray}
&&[\partial_3f_1^+({\bf x},{\bf x}_A,\omega)]_{x_3=x_{3,A}}=\frac{1}{2} i\omega\beta({\bf x}_A)\delta({\bf x}_{\rm L}-{\bf x}_{{\rm L},A}),\label{eqAA47}\\
&&[\partial_3f_1^-({\bf x},{\bf x}_A,\omega)]_{x_3=x_{3,A}}=0.\label{eqAA48}
\end{eqnarray}
Here ${\bf x}_{{\rm L},A}$ denotes the lateral coordinates of $\bx_A$. The operators $\partial_3$ and the factor $\frac{1}{2} i\omega\beta({\bf x}_A)$ are
not necessary to define the focusing conditions but are chosen for later convenience.
To avoid instability, evanescent waves are excluded from the focusing functions. This implies that
the delta function in equation (\ref{eqAA47}) should be interpreted as a spatially band-limited delta function. Note that the sifting property of the delta function,
$h({\bf x}_{{\rm L},A})=\int_\mathbb{S}\delta({\bf x}_{\rm L}-{\bf x}_{{\rm L},A})h({\bf x}_{\rm L}){\rm d}{\bf x}_{\rm L}$, remains valid for a spatially band-limited delta function, assuming $h({\bf x}_{\rm L})$ is also spatially band-limited.
We now derive single-sided Kirchhoff-Helmholtz integrals for backward propagation.
We consider the reciprocity theorems for field-normalised one-way wave fields (equations \ref{eq94} and \ref{eq95}), with $\mathbb{D}$ \rev{and ${\setdD_{0,1}}$ replaced by $\mathbb{D}_A$ and ${\setdD_{0,A}}$,
respectively.}
For state $A$ we consider the focusing functions discussed above, hence, $B_A^+({\bf x},\omega)=B_A^-({\bf x},\omega)=0$ and $P_A^\pm({\bf x},\omega)=f_1^\pm({\bf x},{\bf x}_A,\omega)$.
For state $B$ we consider the decomposed actual field, with a point source at $\bx_B$ in the half-space $x_3>x_{3,0}$ and source spectrum $s(\omega)$.
The parameters $\alpha$ and $\beta$ in state $B$ are the actual parameters inside as well as outside ${\setdD_{0,A}}$.
First, we consider a source in state $B$ which emits waves in the positive $x_3$-direction, hence
$B_B^+({\bf x},\omega)=\delta({\bf x}-{\bf x}_B)s(\omega)$, $B_B^-({\bf x},\omega)=0$
and $P_B^\pm({\bf x},\omega)=P^{\pm,+}({\bf x},{\bf x}_B,\omega)$. Substituting all this
into equations (\ref{eq94}) and (\ref{eq95}) (with $B_0^\pm=B^\pm$),
using equations (\ref{eqAA47}) and (\ref{eqAA48}) in the integrals along ${\partial\setD}_A$, gives
\begin{eqnarray}
&&P^{-,+}({\bf x}_A,{\bf x}_B,\omega)+\chi_A(\bx_B)f_1^-({\bf x}_B,{\bf x}_A,\omega)s(\omega)\label{eqAA60-alt}\\
&&=\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\bigl((\partial_3f_1^+({\bf x},{\bf x}_A,\omega))P^{-,+}({\bf x},{\bf x}_B,\omega)+
(\partial_3f_1^-({\bf x},{\bf x}_A,\omega))P^{+,+}({\bf x},{\bf x}_B,\omega)\bigr){\rm d}{\bf x}_{\rm L}\nonumber
\end{eqnarray}
and
\begin{eqnarray}
&&P^{+,+}({\bf x}_A,{\bf x}_B,\omega)-\chi_A(\bx_B)\{f_1^+({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)\label{eqAA61-alt}\\
&&=\int_{{\partial\setD}_0}\frac{-2}{i\omega\beta({\bf x})}\bigl(\{\partial_3f_1^+({\bf x},{\bf x}_A,\omega)\}^*P^{+,+}({\bf x},{\bf x}_B,\omega)+\{\partial_3f_1^-({\bf x},{\bf x}_A,\omega)\}^*P^{-,+}({\bf x},{\bf x}_B,\omega)\bigr){\rm d}{\bf x}_{\rm L},\nonumber
\end{eqnarray}
where $\chi_A$ is the characteristic function of the domain $\setD_A$. It is defined by equation (\ref{eqC3.2D}), with $\mathbb{D}$ \rev{and ${\setdD_{0,1}}$ replaced by $\mathbb{D}_A$ and ${\setdD_{0,A}}$, respectively.}
Next, we replace the source in state $B$ by one which emits waves in the negative $x_3$-direction, hence
$B_B^+({\bf x},\omega)=0$, $B_B^-({\bf x},\omega)=\delta({\bf x}-{\bf x}_B)s(\omega)$
and $P_B^\pm({\bf x},\omega)=P^{\pm,-}({\bf x},{\bf x}_B,\omega)$. This gives
\begin{eqnarray}
&&P^{-,-}({\bf x}_A,{\bf x}_B,\omega)+\chi_A(\bx_B)f_1^+({\bf x}_B,{\bf x}_A,\omega)s(\omega)\label{eqAA62-alt}\\
&&=\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\bigl((\partial_3f_1^+({\bf x},{\bf x}_A,\omega))P^{-,-}({\bf x},{\bf x}_B,\omega)+(\partial_3f_1^-({\bf x},{\bf x}_A,\omega))P^{+,-}({\bf x},{\bf x}_B,\omega)\bigr){\rm d}{\bf x}_{\rm L}\nonumber
\end{eqnarray}
and
\begin{eqnarray}
&&P^{+,-}({\bf x}_A,{\bf x}_B,\omega)-\chi_A(\bx_B)\{f_1^-({\bf x}_B,{\bf x}_A,\omega)\}^*s(\omega)\label{eqAA63-alt}\\
&&=\int_{{\partial\setD}_0}\frac{-2}{i\omega\beta({\bf x})}\bigl(\{\partial_3f_1^+({\bf x},{\bf x}_A,\omega)\}^*P^{+,-}({\bf x},{\bf x}_B,\omega)+\{\partial_3f_1^-({\bf x},{\bf x}_A,\omega)\}^*P^{-,-}({\bf x},{\bf x}_B,\omega)\bigr){\rm d}{\bf x}_{\rm L}.\nonumber
\end{eqnarray}
Equations (\ref{eqAA60-alt}) $-$ (\ref{eqAA63-alt}) are single-sided representations for backward propagation of the one-way wave fields $P^{\pm,\pm}({\bf x},{\bf x}_B,\omega)$
from ${\partial\setD}_0$ to $\bx_A$.
Similar results have been previously obtained
\citep{Wapenaar2017GP2, Neut2017JASA},
but without decomposition at $\bx_B$. An advantage of these equations over equations (\ref{eq104a-altb}) $-$ (\ref{eq104c-altb}) is that the backward propagated fields
$P^{\pm,\pm}({\bf x}_A,{\bf x}_B,\omega)$ are expressed entirely in terms of integrals along the surface ${\partial\setD}_0$.
Single-sided representations containing the field-normalised focusing functions $f_1^\pm({\bf x},{\bf x}_A,\omega)$ find applications for example in reflection imaging methods which
account for multiple scattering. In these methods, the focusing functions are
retrieved from the reflection response at the surface ${\partial\setD}_0$, using the Marchenko method
\citep{Wapenaar2014GEO, Singh2017GEO2, Costa2018GJI, Wapenaar2019SE}.
We conclude this section by considering a special case.
Suppose that in state $B$ the parameters $\alpha$ and $\beta$ are the same as in state $A$ throughout space.
Then $P^{\pm,\pm}({\bf x},\bx_B,\omega)=G^{\pm,\pm}({\bf x},\bx_B,\omega)s(\omega)$
for all ${\bf x}$. Moreover, $P^{+,\pm}({\bf x},\bx_B,\omega)=0$ for ${\bf x}$ at ${\partial\setD}_0$.
Substituting this into representations (\ref{eqAA60-alt}) $-$ (\ref{eqAA63-alt}),
summing the left- and right-hand sides of these representations separately, dividing both sides by $s(\omega)$
and using equation (\ref{eq101}), we obtain
\begin{eqnarray}
&&G({\bf x}_A,{\bf x}_B,\omega)+\chi_A(\bx_B)2i\Im\{f_1({\bf x}_B,{\bf x}_A,\omega)\}\nonumber\\
&&\hspace{0.5cm}=\int_{{\partial\setD}_0}\frac{2}{i\omega\beta({\bf x})}\partial_3\bigl(f_1^+({\bf x},{\bf x}_A,\omega)-\{f_1^-({\bf x},{\bf x}_A,\omega)\}^*\bigr)G^-({\bf x},{\bf x}_B,\omega){\rm d}{\bf x}_{\rm L}\label{eqAA64-alt}
\end{eqnarray}
(with $\Im$ denoting the imaginary part), where $G^-({\bf x},\bx_B,\omega)=G^{-,+}({\bf x},\bx_B,\omega)+G^{-,-}({\bf x},\bx_B,\omega)$.
Taking the real part of both sides gives
\begin{eqnarray}
&&G_{\rm h}({\bf x}_A,{\bf x}_B,\omega)=\Re\int_{{\partial\setD}_0}\frac{4}{i\omega\beta({\bf x})}\partial_3\bigl(f_1^+({\bf x},{\bf x}_A,\omega)-
\{f_1^-({\bf x},{\bf x}_A,\omega)\}^*\bigr)G^-({\bf x},{\bf x}_B,\omega){\rm d}{\bf x},\nonumber\\&&\label{eqAA65}
\end{eqnarray}
where $G_{\rm h}({\bf x}_A,{\bf x}_B,\omega)$ is the homogeneous Green's function, defined in equation (\ref{eq100}).
Unlike in equation (\ref{eq104acbd-altb}), here the homogeneous Green's function is represented by a single integral along the surface ${\partial\setD}_0$, containing field normalised one-way focusing
and Green's functions.
\section{Conclusions}\label{sec6}
We have considered flux-normalised and field-normalised decomposition of scalar wave fields into coupled one-way wave fields. The operators for field-normalised decomposition
exhibit less symmetry than those for flux-normalised decomposition. Nevertheless, we have shown that reciprocity theorems can be derived for field-normalised one-way wave fields
in a similar way as those for flux-normalised one-way wave fields. An additional condition for the reciprocity theorems for field-normalised one-way wave fields
is that in one of the states the derivatives in the $x_3$-direction of the parameters $\alpha$ and $\beta$ vanish at the boundary of the considered domain.
This condition is easily fulfilled when one of the states is a Green's function or a focusing function, for which the parameters $\alpha$ and $\beta$ can be freely chosen at and outside the
boundary of the domain.
We have used the reciprocity theorems for field-normalised one-way wave fields as a starting point for deriving
representation theorems for field-normalised one-way wave fields in a systematic way. We obtained representations for forward and for backward propagation of one-way wave fields.
These representations account for multiple scattering.
Whereas the Kirchhoff-Helmholtz integrals for forward propagation can be easily transformed into single-sided representations, this transformation is less straightforward for
the Kirchhoff-Helmholtz integrals for backward propagation. By replacing the Green's functions by focusing functions we obtained single-sided representations for backward propagation
of field-normalised one-way wave fields. These representations are particularly useful to retrieve wave fields in the interior of a domain in situations
where measurements can be carried out only at a single surface. An important application is reflection imaging, accounting for multiple scattering.
\section*{Data Availability}
No datasets have been used for this study.
\section*{Conflicts of Interest}
The author declares that he has no competing interests.
\section*{Acknowledgements}
\rev{I thank an anonymous reviewer for the constructive review, which helped to improve the paper.}
This work has received funding from the European Union's Horizon 2020 research and innovation programme: European Research Council (grant agreement 742703).
|
\section{Introduction}
Machine learning, coupled with plentiful data, promises an approach to
the problem of constructing the large knowledge bases needed for AI.
Whereas traditional knowledge engineering by hand, as exemplified by the CYC
project~\cite{lenat95}, proved difficult to scale, machine
learning holds the promise of producing a large and consistently interpreted
knowledge base. Of course, any kind of inductive learning faces the danger of
incorrect generalization, and thus such knowledge must use a semantics that is
weaker than classical logic. Valiant~\shortcite{valiant00} proposed
{\em PAC-Semantics} as a semantics for classical logics that is liberal enough
to tolerate the imperfect rules produced by models of machine
learning~\cite{valiant84}. Subsequently, Valiant~\shortcite{valiant06} also
demonstrated that a large knowledge base can be soundly learned from a
reasonable size data set. Michael and Valiant~\shortcite{mv08} demonstrated
the use of such knowledge bases on a sentence completion task, and
a system using this approach~\cite{im16} was tied for top performance in the
first Winograd Schema Challenge competition~\cite{dmo17}.
Although Valiant~\shortcite{valiant06} showed that there is no {\em statistical}
barrier to learning a large knowledge base, {\em computational} issues from
representing and accessing such a large knowledge base may still arise. One way
of avoiding these issues was proposed by Juba~\shortcite{juba13}, building on
Khardon and Roth's {\em learning to reason} approach~\cite{kr97}: instead of
representing a knowledge base explicitly, Juba decides a query using the data
directly, and guarantees that the result is sufficient to distinguish queries
that have low validity from queries with small proofs using knowledge that
{\em could} have been learned from the data. Thus, the query is answered using
a knowledge base that is only {\em implicitly} learned. Crucially, this approach
applies to settings where some attributes are not observed in the examples used
for learning, and therefore some reasoning may be required. These attributes
may still be mentioned in the background knowledge and query. For example, we
may observe medical test results, but not whether a given patient
actually has a given disease.
A drawback of Juba's
approach, however, is that it provides no explicit representation of {\em what}
knowledge could have been learned to support the query. It only provides a set
of proofs for specific examples from the data set. This may not be adequately
interpretable for human oversight of the system; if possible, we would like to
inspect the knowledge that is being used to provide answers to our queries.
Moreover, there are applications for which we are more interested in what
knowledge could have been used to derive the conclusion than we are in the
conclusion itself. For example, such algorithms might be applied to learn a
screening rule for fraud detection as follows: Given a definition of
behavior that is legitimate and a set of example transaction histories that are
known to be legitimate, we could seek to learn what properties (if any) of the
observed portions of these transactions can be used to guarantee that they are
legitimate by using the definition of legitimate behavior as our query. We can
then check whether or not these learned properties are observed to hold on future
transactions to decide whether or not they warrant further inspection. Note that
in this case the query is presumed to hold, and the interesting part is what
properties we can discover to justify the query. We will discuss a similar but
more easily formalized application to learning input filters later.
In this work, we show how for a large class of {\em oblivious backward search}
algorithms for reasoning, we {\em can} explicitly identify rules that suffice
to answer queries. Thus, we explicitly identify a sufficient set of relevant
rules from this ``implicit'' knowledge base in a goal-driven fashion.
{\em Obliviousness} means that the only effect of the knowledge base on the
search is to terminate branches early when the subgoal is already present in the
knowledge base. We observe that this suffices to learn such rules for logics
such as chaining and treelike resolution where there are natural oblivious (or
nearly oblivious) algorithms, e.g., DPLL-like algorithms~\cite{dp60,dll62}.
This is achieved by deeming subgoals to be successful by adding them to the
knowledge base when they are supported by the data; if the search is obvlivious,
then we obtain the same results as if those formulas had belonged to the
knowledge base all along.
Our algorithms resemble algorithms arising in inductive logic programming
(ILP)~\cite{mdr94}, in particular work by Muggleton and Buntine~\shortcite{mb88}
that constructed rules for resolution; although Muggleton~\shortcite{muggleton91}
anticipated that a connection to PAC-Learning should exist, ILP learning theory
is quite different. The main distinction here is conceptual: ILP treats the
input examples as {\em defining} a domain for which we seek to synthesize a
description. By contrast, in PAC-Semantics, we are only seeking to bound the
probability our formulas are true with respect to some probability distribution
$D$ over valuations. Relatedly, the examples are simply drawn from $D$,
and thus only statistically representative of it. We thus do not have complete
knowledge of $D$ nor do we even have access to the valuation of every formula in
any given example.
Although this bounding of probability is similar to a probability
logic (e.g., as discussed by Nilsson~\shortcite{nilsson86} for propositional
logic or Halpern~\shortcite{Halpern90} for first-order), we stress that the
logical languages we use are simple Boolean logics such as chaining and
resolution, and the probability bounds appear only in our semantics.
\section{Problem formulation}
We now describe the framework we use for learning and reasoning from
partial examples (interpretations).
First, we describe learning from partial examples,
and give the key definition of concealment that captures when
a credulous strategy for learning from partial
examples will succeed. (Skeptical learning is considered later.) We then
define a family of backward
search algorithms, the family of algorithms for which we will be able to
introduce query-driven learning. Since our guarantees will have the form of
ensuring that these reasoning algorithms are as successful as if they had
started their search with the learned knowledge given up-front, we will need
a technical condition that the search algorithms are not too sensitive to the
contents of the knowledge base they are given up-front---that is, ``oblivious''
to the knowledge base. Indeed, this definition will guarantee that we can
introduce learned knowledge as the search proceeds without harming its
performance, which is our main strategy.
\subsection{Learning and reasoning in PAC-Semantics}
Following Valiant~\shortcite{valiant00}, we
will describe our logic in terms of the linear threshold connective (a common
generalization of the usual AND and OR connectives). The linear threshold
connective has the advantage that
it can capture softened
versions of AND and OR.
\begin{definition}[Threshold connective]
A {\em threshold connective} for a list of $k$ formulas $\varphi_1,\ldots,\varphi_k$
is given by a list of $k+1$ integers, $c_1,\ldots,c_k, b$.
The formula $[\sum_{i=1}^kc_i\varphi_i\geq b]$ is interpreted as follows:
given a Boolean interpretation for the $k$ formulas, the connective is true if
$\sum_{i:\varphi_i=1}c_i\geq b$.
\end{definition}
\noindent
A threshold connective expresses a $k$-ary AND connective by
taking the $c_i=1$, and $b=k$, and expresses a $k$-ary OR by taking
$c_1,\ldots,c_k, b=1$. Negation corresponds to $c_i<0$.
Although we could have taken the weights to be real-valued, on account of the
$\varphi_i$'s essentially taking values from $\{0,1\}$, any real-valued threshold
connective has an equivalent integer connective. We use integers as they are
simpler to represent and reason about.
\begin{example}\label{threshold-ex}
Suppose we have one, unary relation symbol $R$ and
six elements in our domain, $x_1,x_2,x_3,x_4,x_5$ and $x_6$. Now, an example of
a formula using a threshold connective is
$$[5R(x_1)+R(x_2)+R(x_3)+R(x_4)+R(x_5)-R(x_6)\geq 4],$$ i.e., with
formulas $\psi_i=R(x_i)$, weights $c_1=5$, $c_6=-1$, and $c_2,\ldots,c_5=1$, and
a threshold of $b=4$.
\end{example}
We will assume a finite domain, and hence a finite (but possibly large) number
of ground atomic formulas, $N$. Thus, our setting is essentially propositional.
We will formulate our exposition in terms of a simplified first-order
language without functions, in which free variables are taken to be universally
quantified; these quantified formulas are of course equivalent to a
quantifier-free (propositional) formula given by an AND (expressed by a
threshold connective) over copies of the formula with each possible binding of
the free variables. This is an example of the standard
{\em ``propositionalization''} transformation, and a suitable family of formulas
for these purposes is described by Valiant~\shortcite{valiant00}.
Nevertheless, this representation captures standard settings of (function-free)
resolution and chaining with Horn KBs.
The main feature of PAC-Semantics is a probability distribution $D$ on
interpretations of the relation symbols, i.e., assignments of truth values to
their groundings. Equivalently, we take each ground atomic formula as a
Boolean-valued random variable. We stress that we do not
assume independence (or any other relationship) between these random variables.
Given an interpretation drawn from $D$, the semantics of a formula are then
defined classically. $KB\models\varphi$ denotes that $\varphi$ is a
{\em (classically) valid formula given the knowledge base (set of formulas)
$KB$}.
We denote the
truth value of a formula $\varphi$ under an interpretation $x$ as $\varphi|_x$.
We may view an interpretation as a Boolean vector
indexed by the set of ground atomic formulas.
Following Valiant~\shortcite{valiant00}, we refer to interpretations of the
ground atomic formulas as {\em scenes}.
\begin{definition}[$(1-\epsilon)$-valid]
A sentence (i.e., formula with no free variables) $\varphi$ is said to be
{\em $(1-\epsilon)$-valid under $D$} if the
probability that $\varphi$ evaluates to true under an
interpretation drawn from $D$ is at least $1-\epsilon$.
If $\epsilon=0$, we say that $\varphi$ is {\em perfectly} valid.
\end{definition}
Now, an {\em obscured scene} is a {\em partial} interpretation of the
ground atomic formulas of the logic:
\begin{definition}[Obscured scene]
A {\em obscured scene} $\rho$ is a mapping taking ground atomic formulas to
$\{0,1,*\}$ where * denotes an ``unknown'' value. We say that an obscured scene
$\rho$ is {\em consistent} with an interpretation if whenever $\rho$ assigns an
ground atomic formula a value other than $*$, the interpretation agrees with
$\rho$.
\end{definition}
We need obscured scenes because frequently our knowledge base will refer to
atomic formulas that are not observed in the data we use for learning. Sometimes
these unobserved atomic formulas take the form of properties we wish to reason
about or predict with learned rules.
Following Rubin~\shortcite{rubin76} and Michael~\shortcite{michael10},
we suppose that a ``masking process'' takes interpretations drawn from
the distribution $D$ and hides some ground atomic formulas, producing
obscured scenes.
\begin{definition}[Masking process]\label{def-mask}
A {\em mask} is a function mapping interpretations to obscured scenes that are
consistent with the respective interpretations. A {\em masking process} $M$ is a mask-valued random
variable (i.e., a random function). We denote the probability distribution over
obscured scenes obtained by applying a masking process $M$ to a distribution $D$
over interpretations by $M(D)$.
\end{definition}
Some natural examples of masking processes that don't use the full expressive
power of the formalism are the following.
\begin{example}\label{mask-ex1}
Consider masking processes
that always produce a mask $m$ that hides the values of a subset of the ground
atomic formulas, and never hide the rest. Such masking processes capture the
information available in a (learning-driven) program analysis application where
the examples specify an input and nothing else. The hidden formulas would then
encode the omitted trace of the program's execution. It also captures the
information available in typical statistical studies in which a subset of the
attributes of sampled members of a population are (reliably) recorded, and
the rest are omitted from the data.
\end{example}
\begin{example}\label{mask-ex2}
Consider a masking process that independently tosses a fair coin
to decide whether or not to hide the value of each ground atomic formula in a
given scene. So $M$ produces $m$ by sampling a set $S$ at random by tossing a
fair coin for each ground atomic formula, and then the corresponding $m$ is of
the type described in Example~\ref{mask-ex1} -- it hides the ground atomic
formulas in the random set $S$ and no others. This captures a setting where,
due to noise corrupting a transmission, only portions of a scene can be decoded.
\end{example}
These examples do not use the ability of a masking process to optionally hide an
atom depending on the underlying truth value, but our definition allows this.
\begin{example}\label{mask-ex3}
Consider a setting where in a survey, participants are allowed to decline to
answer a question. Naturally, one might find that when participants have (for
example) atypically high or low income, or where they possess minority political
opinions, or suffer from certain kinds of diseases, they might be less inclined
to provide an answer. Thus, in such a setting, the survey responses would be
more naturally modeled by a masking process in which, given that the sampled
member of the population falls into one of these categories, the probability of
the value being obscured is much higher than otherwise. Indeed, the model also
allows for the decision to mask to depend on more than one attribute of the
example (note that $m(x)$ is allowed to depend on the entire interpretation $x$)
-- e.g., $m(x)$ may omit the value of the formulas encoding the political
inclinations only if they are relatively inconsistent with some other attributes
of the respondent encoded by $x$.
\end{example}
For a given ground atomic formula $\alpha$, a PAC-Learning
algorithm could be used to learn a formula $\varphi$ that
predicts $\alpha$, in which case the formula
$[\varphi\equiv\alpha]$ is $(1-\epsilon)$-valid with probability
$1-\delta$ over the random example scenes (for $\delta$ given to the algorithm):
if we may take the truth value of $\alpha$ as a {\em label} and the truth values
of the ground atomic formulas as {\em attributes} for the example scenes, then
PAC-Learning uses such examples to produce precisely such a formula $\varphi$ as
output.
Using such a rule, we could hope to infer the value of $\alpha$ in examples in
which it is obscured. Such an approach was proposed by Valiant~\shortcite
{valiant00}, and we will return to it later.
Following Juba~\shortcite{juba13}, we will consider the following operation
that uses obscured scenes to partially evaluate quantifier-free
formulas defined using linear threshold connectives. Again, these could
have been obtained from first-order formulas by propositionalization. Note that
the recursive definition corresponds to a linear-time algorithm for computing
these partially evaluated formulas:
\begin{definition}[Partial evaluation and witnessing]
Given an obscured scene $\rho$ and a quantifier-free formula $\varphi$, the
{\em partial evaluation of $\varphi$ under $\rho$,} denoted $\varphi|_\rho$, is
recursively defined as follows; when the partial evaluation produces a Boolean
constant, we say that the formula is {\em witnessed}:
\begin{compactitem}
\item
A ground atomic formula $\varphi$ is replaced by its value under $\rho$
(i.e., it is witnessed) unless this value is *, in which case it remains $\varphi$.
\item If $\varphi=\neg\psi$ and $\psi$ is not witnessed in $\rho$, then $\varphi|_\rho=\neg(\psi|_\rho)$; otherwise, $\varphi|_\rho$ is witnessed to be $\neg(\psi|_\rho)$.
\item
For $\varphi=[\sum_{i=1}^kc_i\psi_i\geq b]$,
\begin{compactitem}
\item $\varphi$ is witnessed true if
$\sum_{i:\psi_i\mathrm{\ witnessed\ true}}c_i
+\sum_{i:\psi_i\mathrm{\ not\ witnessed}}\min\{0,c_i\}\geq b,$
\item $\varphi$ is witnessed false if
$\sum_{i:\psi_i\mathrm{\ witnessed\ true}}c_i
+\sum_{i:\psi_i\mathrm{\ not\ witnessed}}\max\{0,c_i\}<b,$
\item and otherwise,
supposing that $\psi_1,\ldots,\psi_\ell$ are witnessed in $\rho$ (and
$\psi_{\ell+1},\ldots,\psi_k$ are not witnessed),
$\varphi|_\rho$ is
$[\sum_{i=\ell+1}^kc_i(\psi_i|_\rho)\geq d]$
where $d=b-\sum_{i:\psi_i|_\rho=1}c_i$.
\end{compactitem}
\end{compactitem}
\end{definition}
\begin{example}
Continuing Example~\ref{threshold-ex}, in any partial scene in which $R(x_1)$ is witnessed true, we find that
$\sum_{i:\psi_i\ \text{witnessed true}}c_i+\sum_{i\psi_i\ \text{not witnessed}}\min\{c_i,0\}$ is at least $5-1=4$, so the formula will be witnessed true. Likewise,
if $R(x_2),\ldots,R(x_5)$ are true and $R(x_6)$ is false, then the formula is
again witnessed true. But, if $R(x_1)$ is false and $R(x_6)$ is true, then the
formula is witnessed false, as it is if $R(x_1)$ is false and any of $R(x_2),
\ldots,R(x_5)$ are false. Finally, the formula is not witnessed if, for example,
$R(x_1)$ and $R(x_6)$ are both false, and any proper subset of $R(x_2),\ldots,
R(x_5)$ are true.
\end{example}
Following Michael~\shortcite{michael10}, we
consider learning from example obscured scenes, provided that the value of
$\alpha$ is not obscured in too many of the examples.
Essentially we distinguish formulas that are {\em
perfectly} valid under the unknown distribution from those that are not even
$(1-\epsilon)$-valid for some given $\epsilon>0$. The main finding is that the
learnability of such rules is controlled by the
probability of observing counterexamples to flawed rules under the masking
process.
\begin{definition}[Concealment]\label{def-conceal}
We say that a masking process $M$ is (at most) {\em $(1-\eta)$-concealing}
with respect to a set of formulas $\mathcal{C}$ and a distribution over
interpretations $D$ if
\[
\forall\varphi\in\mathcal{C} \Pr_{x\in D,m\in M}[\ \varphi\mathrm{\ witnessed\ on\ }m(x)\ |\ \varphi|_x=0\ ]\geq\eta.
\]
\end{definition}
We observe that the degree of concealment of a family of formulas depends
on the family of formulas, the distribution over scenes, and the masking process.
\begin{example}
In the masking process of Example~\ref{mask-ex1}, in which a fixed subset of the
ground atomic formulas is never hidden and the rest are always hidden, the degree
of concealment depends on whether or not the formula in question is ever
falsified on the distribution, and if so, which ground atomic formulas in the
scene it refers to. Generally, for formulas that only refer to the observed
ground atomic formulas, the masking process is 0-concealing (i.e., $\eta=1$,
no discounting) but for formulas that only refer to the unobserved ground
atomic formulas, the masking process is 1-concealing ($\eta=0$) and learning is,
strictly speaking, impossible. It may still be possible to infer the values of
these attributes by reasoning, however, e.g., in the program analysis example we
may be able to use the program code to infer the values of the program state from
an example input.
\end{example}
\begin{example}
In the case of the masking process of Example~\ref{mask-ex2}, where the ground
atomic formulas are hidden uniformly at random, the degree of concealment may be
bounded by the number of distinct ground atomic formulas: if we observe all $k$
of the ground atomic formulas we certainly observe the formula being falsified,
and this occurs with probability $1/2^k$. Thus, the masking process in this case
would be $(1-1/2^k)$-concealing ($\eta=1/2^k$). In this case, we can take the
number of ground atomic formulas that appear in a formula as a measure of its
complexity. Some natural fragments of logics, such as bounded-width
resolution, limit the number of atomic formulas that may appear in the lines of
a proof, and would thus control the degree of concealment for lines of the proof
for this masking process.
\end{example}
\noindent
In the statement of our main theorem, we include the size of the input query
formula as a parameter, and we suppose that the degree of concealment may depend
on this parameter. This is because it is sometimes possible to bound the number
of distinct formulas that can appear in proofs by the number of proofs, which
may be bounded similar Lemma~\ref{space-count}, below. The number of
proofs may sometimes depend, in turn, on the number and complexity of the
premises which are encoded in the query -- consider for example, proofs with a
bounded number of lines. We have included this parameterization to facilitate
the application of our theorem in such situations.
Bounded concealment justifies a ``credulous'' learning strategy: rules are
satisfactory as long as we do not observe counterexamples to them.
\begin{proposition}[Theorem 2.2 of \cite{michael10}]
\label{conceal-ce}
For any distribution $D$ over interpretations, masking process $M$, and class
$\mathcal{C}$ of formulas,
if $M$ is $(1-\eta)$-concealing with respect to $\mathcal{C}$ and $D$
and $\varphi\in\mathcal{C}$ is not $(1-\epsilon)$-valid, then
$\Pr_{\rho\in M(D)}[\varphi|_\rho=0]\geq \eta\epsilon$.
Conversely, if $M$ is not $(1-\eta)$-concealing with respect to $\mathcal{C}$ and
$D$ then there exists $\varphi\in\mathcal{C}$ such that
$\Pr_{\rho\in M(D)}[\varphi|_\rho=0]\leq \eta\Pr_{x\in D}[\varphi|_x=0]$
(where, note, $\varphi$ is $(1-\Pr_{x\in D}[\varphi|_x=0])$-valid).
\end{proposition}
That is, the degree of concealment controls the discounting of the probability
of observing counterexamples to formulas from $\mathcal{C}$. We have actually modified
the definitions slightly from the original version by
including the distribution in the definition of concealment, and moreover, by
using a notion of witnessed evaluation (as opposed to the value merely being
{\em determined} by the obscured scene) but the proof is similar.
\iffalse{
\begin{proof}
The main observation is that by definition,
\[
\Pr_{\rho\in M(D)}[\varphi|_\rho=0]
=\Pr_{\substack{x\in D,\\m\in M}}[\varphi\mathrm{\ witnessed\ on\ }m(x)|\varphi|_x=0]
\Pr_{x\in D}[\varphi_x=0].
\]
If $\varphi$ is not $(1-\epsilon)$-valid, then
$\Pr_{x\in D}[\varphi|_x=0]\geq\epsilon$;
if $M$ is $(1-\eta)$-concealing, then
\[
\Pr_{\substack{x\in D,\\m\in M}}[\varphi\mathrm{\ witnessed\ on\ }m(x)|\varphi|_x=0]
\geq\eta
\]
and so the first part of the claim is immediate. For the second part, suppose
that the conclusion does not hold and for all $\varphi\in\mathcal{C}$
\[
\Pr_{\rho\in M(D)}[\varphi|_\rho=0]>\eta\Pr_{x\in D}[\varphi|_x=0].
\]
Then the above calculation implies that for all $\varphi\in\mathcal{C}$,
\[
\Pr_{\substack{x\in D,\\m\in M}}[\varphi\mathrm{\ witnessed\ on\ }m(x)|\varphi|_x=0]>
\eta
\]
so $M$ is $(1-\eta)$-concealing with respect to $D$ and $\mathcal{C}$.
\end{proof}
}\fi
In the above formulation of PAC-Learning using equivalence rules, this means
that for any incorrect hypothesis in our representation class, the value of
$\alpha$ should be witnessed (by the masking process) on an example where the
hypothesis is incorrect with probability at least $\eta$. It turns out that for
many classes of interest, learning is still possible as long as the masking
process features an $\eta$ bounded away from zero.
\iffalse{
For some applications in which we wish to screen out an adversary such as
fraud or attack detection, we desire a more ``skeptical'' learning strategy that
does not rely on the assumption of bounded concealment. In such cases, we may
use the following notion of ``testability'' from Juba~\shortcite{juba13}:
\begin{definition}[Testable]\label{def-testable}
We say that a formula $\varphi$ is {\em $(1-\epsilon)$-testable} with respect
to the distribution over partial interpretations $M(D)$ if
$\Pr_{\rho\in M(D)}[\varphi\mathrm{\ witnessed\ true\ on\ }\rho]\geq 1-\epsilon.$
\end{definition}
Note that if $\varphi$ is $(1-\epsilon)$-testable with respect to $M(D)$, it must
be $(1-\epsilon)$-valid with respect to $D$.
}\fi
In this work, we are interested in reasoning problems, in which we only
consider proofs of bounded complexity. For simplicity our formulation is again
presented in terms of ground formulas, which could have been obtained by
propositionalization. Formally, we consider the following
family of problems:
\begin{definition}[Search problem]
Fix a logic, and let $\mathcal{P}$ be a set of proofs in the logic (e.g., a fragment
of bounded complexity).
The {\em search problem for $\mathcal{P}$} is then the following promise
problem: given as input a formula $\varphi$ with no free variables and a set
of formulas $KB$ such that either there is a proof of $\varphi$ in $\mathcal{P}$ from
$KB$ or else $KB\not\models\varphi$, return such a proof in the former case, and
return ``Fail'' in the latter.
\end{definition}
Our first example of such a fragment is a mild generalization of the usual
forward-chaining system that was presented by Valiant~\shortcite{valiant00}. It is
designed to utilize rules of the form $[\varphi\equiv\alpha]$ in which $\alpha$ is
an atomic formula, i.e., of the sort obtained from PAC-Learning algorithms.
The main inference rule is {\em chaining}: given a formula of the form $[\varphi
\equiv\alpha]$ in which $\alpha$ is a ground atomic formula, and a consistent
set of literals (atomic formulas or their negations)
$\{\ell_1,\ldots,\ell_k\}$ such that for the obscured
scene $\rho$ that satisfies $\ell_1,\ldots,\ell_k$ (and leaves every ground
atomic formula not appearing in this list unassigned) $\varphi|_\rho\in\{0,1\}$, if
$\varphi|_\rho=0$, infer $\neg\alpha$, and otherwise (i.e., if $\varphi|_\rho=1$)
infer $\alpha$.
In chaining, it is typical to distinguish ground atomic
formulas (``facts'') and ``rules'' of the form $[\varphi\equiv\alpha]$; in some
formulations, we suppose that only the facts appear as lines of the proof and
take the rules to be rules of inference. This limits the complexity of the
proofs that may appear in the fragment: they are just ordered lists of facts.
In particular, recall that one often considers augmenting the axioms of a logic
with a set of additional formulas -- {\em ``hypotheses''} -- that capture a
specific domain or a specific scene one wishes to reason about. Typically,
these hypotheses are the contents of the knowledge base. In the case of
forward-chaining, for example, often these hypotheses are restricted to be just
a set of (additional) facts. We will suppose in particular that the set of
proofs $\mathcal{P}$ parameterizing our search problem may restrict the formulas that
may be used as hypotheses in this way.
\subsection{A model of backward search algorithms}
Informally, ``backward'' (goal-directed) algorithms start with a goal query
and repeatedly generate sets of subgoal queries such that if all of the subgoal
queries succeed -- i.e., if proofs can be found for all of these subgoal queries
-- then the algorithm can construct a proof of the goal query. Although this
informal description suggests a recursive algorithm, it is
highly desirable to cache the results of subgoal queries when they are
answered---beyond the obvious time-efficiency improvements, it is often possible
for a na\"{\i}ve recursive algorithm to get stuck following a circular sequence
of subgoals. Thus, our model of such algorithms will be stated in terms of a
graph indicating the dependency structure among the (sub)goals considered by the
algorithm.
This graph would conventionally be an ``AND-OR'' graph:
a query would be associated with an OR node, with edges to various alternative
lists of subgoal queries, represented by AND nodes with edges to the OR nodes
corresponding to the queries in the list. The success of the algorithm
corresponds to the goal query node evaluating to `true' in the natural
interpretation of such a graph when one more generally associates success at
finding a proof of a (subgoal) query with a node evaluating to `true.'
Given our interest in using rules expressed using linear threshold connectives,
though, we will find it natural and convenient to replace the AND nodes with
more general ``linear threshold'' nodes, expressing that success at the node is
achieved if an appropriate subset of the subgoals are successful.
\begin{definition}[Subgoal dependency graph]
A {\em subgoal dependency graph} is a (possibly infinite) directed graph $G$ in
which the vertices are either {\em query nodes} labeled with a formula or are
labeled with an integer {\em threshold} and have outgoing edges labeled
by integer {\em weights}. A {\em partial subgoal dependency graph} also contains,
for each threshold vertex, weights $w_+$ and $w_-$. Given sets of
{\em successful} nodes $S$ and {\em unsuccessful} nodes $U$ in the partial
graph $G'$, each node $v$ is considered {\em successful using $S,U$} if there is
an acyclic subgraph of $G'$ featuring $v$ as the (unique) source with sinks from
$S\cup U$ such that $v$ has a path to every sink and at every non-query node, the sum of the weights on the
outgoing edges to vertices in $S$ plus $w_-$ and the negative weights on outgoing edges to vertices outside $S$ or $U$ is at least the threshold. Similarly, $v$ is
{\em unsuccessful} if the sum of the weights on outgoing edges to vertices in $S$
plus $w_+$ and the positive weights of edges to vertices outside $S$ and $U$ is
less than its threshold.
\end{definition}
\noindent
Our rule for determining when a vertex is successful or unsuccessful match
our rules for witnessing connectives true and false, respectively. The weights
$w_+$ and $w_-$ will allow us to determine witnessing with (unwitnessed)
unrepresented vertices. They represent the total weight of unrepresented
vertices with positive coefficients and negative coefficients, respectively;
note that the definition of witnessing uses one to establish a formula is
witnessed true and the other to establish that it is witnessed false. Our
sucessful vertices will intuitively represent either a provable query, or an
applicable inference.
The backward search algorithm is now a meta-algorithm (Algorithm~\ref
{generic-backward-alg}) parameterized by three
sub-algorithms. One algorithm, {\tt GENERATE}, generates the subgoal dependency
graph, and another, {\tt EXPLORE}, chooses edges in the dependency graph to
explore (as long as the algorithm is not done). The third algorithm, {\tt TEST},
generates a proof of the query if enough of the subgoal dependency graph has been
revealed so that the original goal vertex was successful (given the axioms and a
knowledge base as successful vertices), or else indicates that the search is not
successful yet.
Thus, the algorithm explores the subgoal dependency graph (starting from the
original query) until an appropriate collection of successful subgoals is
discovered or the search algorithm gives up.
\begin{algorithm}[t]
\DontPrintSemicolon
\SetKwInOut{Input}{input}\SetKwInOut{Output}{output}
\iffalse{
that, given a finite subgoal dependency graph
$G$ with a subset of vertices marked ``unexplored'' and a set of hypotheses $H$,
chooses an unexplored vertex $v\in G$ or outputs ``FAIL.''
Algorithm {\tt GENERATE} that, given a subgoal dependency graph $G$ and a vertex
$v\in G$ either returns ``fully explored'' or returns a subgoal dependency graph
$G'$ such that $G$ is a subgraph of $G'$, obtained by eliminating one edge and
the disconnected vertex (if one exists).
Algorithm {\tt TEST} that, given a subgoal dependency graph $G$, a query formula
$\varphi$ labeling some vertex of $G$, and a set of hypotheses $H$, either
returns a proof of $\varphi$ from $H$, or if the vertex labeled by $\varphi$ is
not successful using $H$ and the axioms of the logic, returns ``FAIL.''}
}\fi
\Input{Query formula $\varphi$, set of formulas $KB$}
\Begin{
\If{$\varphi\in KB$}{
\Return{Trivial proof of $\varphi$}
}
$G\leftarrow$ vertex labeled by $\varphi$, marked ``unexplored''\\
\While{{\tt TEST}$(G,\varphi,KB)=$ FAIL}{
\If{$v\gets${\tt EXPLORE}$(G,\varphi,KB)$ is not FAIL}{
\If{$G'\gets${\tt GENERATE}$(G,v)$ is not ``fully explored''}{
\If{$G'$ contains a vertex not in $G$, not labeled with $\psi\in KB$}{
Mark the new vertex ``unexplored.''
}
$G\leftarrow G'$
}
\lElse{
Remove ``unexplored'' mark from $v$
}
}
\lElse{
\Return{{\em Fail}}
}
}
\Return{{\tt TEST}$(G,\varphi,KB)$}
}
\caption{Backward search meta-algorithm}\label{generic-backward-alg}
\end{algorithm}
\begin{definition}[Backward search algorithm]
\label{bsearch-def}
A {\em backward search algorithm} is given by an instantiation
of Algorithm~\ref{generic-backward-alg} with three algorithms:
\begin{compactitem}
\item
An algorithm {\tt EXPLORE} that, given a finite partial subgoal dependency
graph $G$ with a source labeled by $\varphi$ and a subset of vertices marked
``unexplored,'' and set of formulas $KB$, chooses an unexplored vertex $v\in G$
or outputs ``FAIL.''
\item
An algorithm {\tt GENERATE} that, given a partial subgoal dependency graph
$G$ and a vertex $v\in G$ either returns ``fully explored'' or returns a subgoal
dependency graph $G'$ that extends $G$ by adding one new edge starting from $v$,
possibly to a new vertex, and reducing $w_+$ or $w_-$ by its weight if it is
positive or negative, respectively.
\item
An algorithm {\tt TEST} that, given a partial subgoal dependency graph $G$,
a query formula $\varphi$ labeling some vertex of $G$, and a set of formulas
$KB$, either returns a proof of $\varphi$ from $KB$, or if the vertex labeled by
$\varphi$ is not successful using $KB$ and the axioms of the logic,
returns ``FAIL.''
\end{compactitem}
\end{definition}
A key property possessed by
instantiations of the backward-search paradigm is that the graph generation
algorithm is often {\em oblivious} to which queries are successful, the
algorithm exploring the graph only ``terminates early'' when it encounters a
vertex labeled by a successful query, and the algorithm recovering the proof
depends only on the portion of the subgoal dependency graph revealed thus far
(and the formulas appearing on query vertices appearing in it).
We will restrict our attention to such {\em oblivious} algorithms.
\begin{definition}[Oblivious backward search algorithm]
\label{oblivious-search-def}
We say that a backward search algorithm is {\em oblivious} if:
\begin{compactenum}
\item For any partial subgoal dependency graph $G$, query $\varphi$ (contained as
a label in $G$), and set of formulas $\Phi$ not appearing as labels
of query nodes in $G$, $\mathtt{TEST}(G,\varphi,KB)=\mathtt{TEST}(G,\varphi,KB
\cup\Phi)$.
\item For any query $\varphi$ and sets of formulas $KB$ and $\Phi$,
the execution of the algorithm on input $\varphi$ and $KB\cup\Phi$ differs from
the execution on input $\varphi$ and $KB$ only in that the sequence of vertices
proposed by {\tt EXPLORE} on input $\varphi$ and $KB\cup\Phi$ is the subsequence
of vertices proposed on input $\varphi$ and $KB$ that omits (skips) exploring
vertices in the sequence that are successful from $KB\cup\Phi$ in the partial
subgoal dependency graph used by the algorithm in the corresponding step.
\end{compactenum}
\end{definition}
\subsubsection{An example: oblivious backward chaining}
We briefly note that standard backward-chaining algorithms (for Horn KBs on
ground atomic formulas) are
oblivious backward search algorithms in the sense of Definition~\ref
{oblivious-search-def}: recall that a {\em Horn clause} is a formula of the
form $[\alpha_1\wedge\cdots\wedge\alpha_k]\Rightarrow\alpha_{k+1}$ where each
$\alpha_i$ is a ground atomic formula. $\alpha_{k+1}$ is the {\em head} whereas
$\alpha_1\wedge\cdots\wedge\alpha_k$ is the {\em body}. The knowledge base then
consists of such clauses and a set of ground atomic formulas. The query is a
conjunction of ground atomic formulas, represented as a threshold vertex with a
threshold equal to the number of atomic formulas in the conjunction.
The oblivious backward chaining algorithm works as follows:
{\tt EXPLORE} performs a depth-first search of the subgoal dependency graph,
terminating a search early only when it encounters an atomic formula in $KB$.
{\tt GENERATE}, on the other hand, when given a vertex corresponding to a ground
atomic formula, returns an edge to a threshold vertex corresponding to the next
clause in $KB$ with the given atomic formula appearing in the head (in some
fixed ordering), with a threshold equal to the number of atomic formulas in the
body; when given a vertex corresponding to one of the clauses of $KB$ (or the
goal conjunction), it returns an edge to the vertex labeled with the next atomic
formula in the body (also in some fixed ordering) of weight $1$. Finally,
{\tt TEST} uses a dynamic programming algorithm to check if the query is
successful from $KB$ and return a chaining proof if it is. The running
time is bounded by a polynomial in the size of $KB$: it runs
for a linear number of iterations, and {\tt TEST} may take quadratic time on
each iteration.
\begin{example}
We now illustrate the backward search algorithm for chaining of Horn rules.
Let's consider the following simple domain, concerning fragile objects. The
relations are $fragile(x)$, $broken(x)$, $hard(x)$, $crushed(x)$, and
$hit(x,y)$. We have a KB containing rules
$[crushed(x)\wedge fragile(x)]\Rightarrow broken(x)$ and
$[hit(x,y)\wedge fragile(x)\wedge hard(y)]\Rightarrow broken(x)$.
Let's suppose that the domain of objects consists of $sculpture, crate, floor,$
and $sidewalk$. As an example, we might indicate that a fragile
sculpture is crushed -- $fragile(sculpture)$ and $crush(sculpture)$ are given --
and we wish to know if $broken(sculpture)$ holds. The full subgoal dependency
graph is depicted in Figure~\ref{bs-graph-fig}.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=.8\textwidth]{horn-bs-graph}
\caption{A (AND-OR) subgoal dependency graph for the query $broken(sculpture)$
with backward chaining using our Horn KB. The threshold nodes are
labeled with the corresponding groundings of the rules of the KB. ANDs
are represented by arcs connecting edges.}\label{bs-graph-fig}
\end{center}
\end{figure*}
A short execution of the backward search algorithm starts with the query node,
$broken(sculpture)$. {\tt TEST} on $broken(sculpture)$ determines that this fact
is not given, so the algorithm invokes {\tt EXPLORE} which determines that the
vertex $broken(sculpture)$ is not yet fully explored. So, the algorithm invokes
{\tt GENERATE} on $broken(sculpture)$ which begins generating the rules that
could produce $broken(x)$ in the head with $sculpture$ substituted for $x$. For
simplicity, let's suppose that it first considers the rule $[crushed(x)\wedge
fragile(x)]\Rightarrow broken(x)$, which is such a rule. This threshold vertex
represents an AND on two conjuncts, so it has a threshold of $2$. As the graph
is partial, it has values $w_+=2$ since both of the conjuncts (yet to be
generated) have a weight of $1>0$ and $w_-=0$ since this threshold formula does
not use negative weights. Note that none of the nodes are yet witnessed, so
{\tt TEST} will continue to the next iteration.
{\tt EXPLORE} would, if it is a depth-first exploration, choose the new rule to
explore. So it would invoke {\tt GENERATE} on the rule, which would first yield
the subgoal node $crushed(sculpture)$ and reduce the value of $w_+$ for the
threshold by $1$ (since one of these nodes is now represented in the graph).
Note that $crushed(sculpture)$ is in the KB, so this vertex will not be marked
``unexplored.'' Furthermore, this node is successful, since it is in the KB.
But, the rule that uses $crushed(sculpture)$ is not yet witnessed since it also
needs $fragile(sculpture)$, so {\tt TEST} will continue to the next iteration.
Now, since the new node is not marked ``unexplored,'' the depth-first
{\tt EXPLORE} would return to the rule, and {\tt GENERATE} next generates the
subgoal $fragile(sculpture)$, which is also in the KB, and reducing $w_+$ (at
the rule's vertex) by $1$ to $0$. At this point, {\tt TEST} will discover that
$crushed(sculpture)$ and $fragile(sculpture)$ suffice to witness the linear
threshold corresponding to $[crushed(sculpture)\wedge fragile(sculpture)]$, so
this node is successful, which in turn witnesses that the original goal
$broken(sculpture)$ is successful. Thus, {\tt TEST} returns the chaining proof
of the query
\begin{compactenum}
\item $crushed(sculpture)$ (hypothesis)
\item $fragile(sculpture)$ (hypothesis)
\item $broken(sculpture)$ (chaining, 1 \& 2, $[crushed(x)\wedge fragile(x)]\Rightarrow broken(x)/x=sculpture$)
\end{compactenum}
Needless to say, the execution would be rather longer if the search had first
started exploring the various domain substitutions for $y$ in the rule $[hit(x,y)
\wedge fragile(x)\wedge hard(y)]\Rightarrow broken(x)$. A breadth-first search
of the graph would have been similarly longer. Note that there are no rules with
$hit(x,y)$ or $hard(x)$ in the head, so although these nodes will initially be
marked ``unexplored'' (in contrast to, say, $fragile(sculpture)$), when
{\tt GENERATE} considers these nodes it will immediately report that they are
``fully explored.'' Thus, after exploring these rules, the search will eventually
return to exploring $[crushed(x)\wedge fragile(x)]\Rightarrow broken(x)$ and
succeed as described above, even in these cases.
Now, suppose that the sculpture hits the floor, so $hit(sculpture, floor)$
holds instead of $crushed(sculpture)$. Now, we are {\em not} given $hard(floor)$
(the floor may be carpeted) so there is no proof of $broken(sculpture)$. In this
case, the backward search algorithm will generate the entire subgoal dependency
graph, before determining that the query is not provable, and terminate with
Fail.
\end{example}
\section{Query-driven learning in backward search}
\begin{algorithm}[t]
\DontPrintSemicolon
\SetKwInOut{Input}{input}\SetKwInOut{Output}{output}
\Input{Query formula $\varphi$, set of formulas $KB$, list of
obscured scenes $\rho_1\ldots,\rho_m$}
\Begin{
$G\leftarrow$ vertex labeled by $\varphi$, marked ``unexplored''\\
\While{{\tt TEST}$(G,\varphi,KB)=$ FAIL}{
\If{$v\gets${\tt EXPLORE}$(G,\varphi,KB)$ is not FAIL}{
\If{$G'\gets${\tt GENERATE}$(G,v)$ is not ``fully explored''}{
\lIf{$G'$ contains a query vertex labeled by a formula $h$ not in $G$
and for no $\rho_i$ is $h|_{\rho_i}=0$}{
$KB\leftarrow KB\cup\{h\}$
}
\If{$G'$ contains a vertex not in $G$, not labeled with $\psi\in KB$}{
Mark the new vertex ``unexplored.''
}
$G\leftarrow G'$
}
\lElse{
Remove ``unexplored'' mark from $v$
}
}
\lElse{
\Return{{\em Fail}}
}
}
\Return{{\tt TEST}$(G,\varphi,KB)$}
}
\caption{Backward search with query-driven learning}
\label{mod-bw-alg}
\end{algorithm}
The relative blindness of oblivious algorithms to the contents of the knowledge
base allows us to add new members as the search proceeds, and
obtain the same result as if we had started with them. As some algorithms may
consider families of formulas that scale with the size of the query, in our
theorem we will parameterize our bounds on the proof size $B$ and degree of
concealment $\eta$ by the size of the query $\ell$ (in bits). For example,
adding clauses to a query for resolution generally increases the variety of
clauses that may be derived. For larger
families, we expect that the bounds grow weaker. We will use $|\varphi|$ for a
formula $\varphi$ to denote its representation size (in bits), and $|KB|$ to
denote the representation size of the $KB$, again in bits. We assume the $KB$
is represented in a way that ensures $|KB\cup H|\leq |KB|+|H|$, e.g., if it is
represented as a string in which the elements are separated by special symbols,
and terminated by another symbol.
\begin{theorem
\label{explicit-search-thm}
Let $\mathcal{P}$ be a set of proofs such that proofs of queries of length $\ell$
only have proofs with $B(\ell)$-bit encodings in $\mathcal{P}$ (in some fixed encoding
scheme). Suppose there is an oblivious backward search algorithm for the
search problem for $\mathcal{P}$ that on input
$\varphi$ and $KB$ over $N$ ground atomic formulas,
runs in time $T(N,|\varphi|,|KB|)$ (for
a function $T$ that is monotone increasing in $|KB|$).
Let $D$ be a distribution over scenes and $M$ be a masking process that is at
most $(1-\eta(\ell))$-concealing for the set of formulas $\Phi$ that may be
hypotheses in proofs of formulas of length $\ell$ in $\mathcal{P}$.
Then for any $\delta,\epsilon\in (0,1)$, on input $\varphi$ and $KB$ and
$\Theta(\frac{1}{\epsilon\eta(|\varphi|)}(B(|\varphi|)+\log 1/\delta))$ example
obscured scenes, Algorithm~\ref{mod-bw-alg} using the same {\tt EXPLORE},
{\tt GENERATE}, and {\tt TEST} as the given algorithm runs in time
$O(\frac{B(|\varphi|)}{\epsilon\eta(|\varphi|)}(B(|\varphi|)+\log 1/\delta)
T(N,|\varphi|,C))$ (for $C=|KB|+T(N,|\varphi|,|KB|)$),
and with probability $1-\delta$
\begin{compactitem}
\item
returns ``Fail'' if $[KB\Rightarrow\varphi]$ is not $(1-\epsilon)$-valid with respect to $D$, or
\item
returns a proof of $\varphi$ from $KB\cup H'$ for a set of formulas
$H'=\{h'_1,\ldots,h'_k\}$ such that $h'_1\wedge\cdots\wedge h'_k$ is
$(1-\epsilon)$-valid if there exists a set of perfectly valid formulas $H$
such that there is a proof of $\varphi$ from $KB\cup H$ in $\mathcal{P}$.
\end{compactitem}
\end{theorem}
We remark that the cases are {\em not} exhaustive, for two reasons. First, for
many logics, the fragments for which proof search can be considered tractable
are not complete. Second, it may be that the query formula is indeed
$(1-\epsilon)$-valid, but that there is not a set of formulas that we can learn
in support of it. For example, in chaining it may be that we have $p\to r$ and
$q \to r$, and we never observe $r$ but we do observe either $p$ or $q$ to be
true, half of the time each. So, we know $r$ is always true, but since neither
$p$ nor $q$ is consistently true, neither one can be learned.
\begin{proof}
First suppose that there is a set of perfectly valid formulas $H$ such
that there is a proof of the query $\varphi$ from $KB\cup H$ in $\mathcal{P}$. Since
we have assumed that the given backward search algorithm solves the
search problem for $\mathcal{P}$, on input $\varphi$ and $KB\cup H$, the given
backward search algorithm would return a proof of $\varphi$ from $KB\cup H$.
Suppose this occurs after $t^*$ iterations.
Now, consider the sets $KB'_t$ and subgoal dependency graphs $G'_t$ used on
each respective iteration $t$ by Algorithm~\ref{mod-bw-alg}. If we consider the
runs of the given algorithm using $KB'_t$, we see that since it is assumed to
be oblivious, on each step up to $t$, it proposes the same vertex to explore as
Algorithm~\ref{mod-bw-alg} (i.e., it only omits the vertices that are successful
from $KB'_t$) and thus generates the subgoal dependency graph $G'_t$ on the
$t$th step. Furthermore, we note that since $KB'_t$ contains the subset of $KB
\cup H$ that appears as labels in $G'_t$, every vertex of $G'_t$ that is not
deemed successful from $KB'_t$ is likewise not successful from $KB\cup H$ in
the graph generated on the $t$th step on input $\varphi$ and $KB\cup H$.
Therefore, if Algorithm~\ref{mod-bw-alg} runs for $t^*$ iterations, $\varphi$ is
successful from $KB'_{t^*}$. Moreover, as {\tt EXPLORE} continues to propose the
vertices of $G'_t$ not successful from $KB'_t$ on each step (since it is
oblivious) until the vertex labeled by $\varphi$ is successful, another
unsuccessful vertex must exist. Therefore the algorithm can only terminate
before $t^*$ iterations if {\tt TEST} returns a proof and so either way
Algorithm~\ref{mod-bw-alg} returns a proof of $\varphi$ in $\mathcal{P}$ from some
$KB'_t$.
The running time is bounded as follows: we see that the algorithm runs for no
more iterations than before, and the final size of $KB'_t$ is at most
$|KB|$ plus the algorithm's running time, since the algorithm creates vertices
representing each formula added to $KB'_t$. Ignoring the time to test the
proposed vertices for membership in $KB'_t$, the running time may be at most
$T(N,|\varphi|,|KB'_t|)\leq T(N,|\varphi|,|KB|+T(N,|\varphi|,|KB|))$. Now, each
formula tested appears as a premise in some proof in $\mathcal{P}$ and therefore has
a representation of size at most $B(|\varphi|)$. Since we can determine the
witnessed value of a formula on a given obscured scene in linear time
the time bound follows.
We now argue that with probability $1-\delta$ over the example obscured scenes
provided to the algorithm as input, any proof that could be returned by
Algorithm~\ref{mod-bw-alg} uses a knowledge base $KB\cup H'$ such that
the formula $h'_1\wedge\cdots\wedge h'_k$ (for $H'=\{h'_1,\ldots,h'_k\}$) is
$(1-\epsilon)$-valid. This will establish the theorem as the existence of such a
proof guarantees that the query $\varphi$ is $(1-\epsilon)$-valid, so if the first
case holds, the algorithm cannot produce a proof except with probability
$\delta$; and likewise, in the second case, the proof returned by the algorithm
is satisfactory with probability $1-\delta$.
To this end, we note that since $M$ is assumed to be $(1-
\eta(|\varphi|))$-concealing with respect to the premises that may appear on any
proof of $\varphi$ in $\mathcal{P}$, for any proof using a set of premises that is
{\em not} $(1-\epsilon)$-valid, Proposition~\ref{conceal-ce} shows that each
example produces an obscured scene $\rho$ for which $h'_i|_\rho=0$ for some
$h'_i$ used as a premise with probability at least $\epsilon\eta(|\varphi|)$.
Therefore, in a sample of $\Omega(\frac{1}{\epsilon\eta(|\varphi|)}(B(|\varphi|)+
\log 1/\delta))$ independent obscured scenes, the probability that no premise of
such a proof has $h'_i|_\rho=0$ for any $\rho$ in the sample is at most
$\delta\cdot 2^{-B(|\varphi|)}$; as the proofs in $\mathcal{P}$ for $\varphi$ have
encodings of at most $B(|\varphi|)$ bits, a union bound over these proofs gives
that the overall probability of some proof having no $\rho$ in the sample for
which $h'_i|_\rho=0$ for some premise in the proof is at most $\delta$.
Now, as the algorithm only returns proofs using premises contained in the sets
$KB'_t$ which in particular do not contain formulas $h'$ such that
$h'|_\rho=0$ for any $\rho$ in the sample, we see that with probability $1-
\delta$, the algorithm does not return a proof with a set of premises that is
not $(1-\epsilon)$-valid, as needed.
\end{proof}
\subsubsection{Example: backward chaining}
If we represent chaining proofs by the sequence of inferred ground atomic
formulas (using $\log N$ bits for each), then since any proof needs only write
down an atomic formula at most once, we can use the bound $B=N\log N$ in the
statement of Theorem~\ref{explicit-search-thm}. (The size of the query $\ell$ is
always the length of a single ground atomic formula, $\log N$ bits.)
So, after applying the transformation depicted in Algorithm~\ref{mod-bw-alg} to
the standard backward-chaining algorithm, Theorem~\ref
{explicit-search-thm} establishes that this modified backward chaining algorithm
finds chaining proofs of queries using not only ground atomic formulas from the
explicitly given $KB$, but also additional formulas that are almost always true.
The modified algorithm automatically supplements a given $KB$ with any such
additional ground atomic formulas that suffice to complete some chaining proof
of a query if one exists, provided further that the masking process has bounded
concealment with respect to ground atomic formulas---meaning here, the masking
process leaves the value of each ground atomic formula present with some bounded
probability when it is false.
\iffalse{
We also note that a similar backward chaining algorithm could be used with
the kind of formulas explicitly learned in Valiant's model~\cite
{valiant00}---rules of the form $\phi\equiv\tau$ for a literal $\tau$ and $\phi$
of the form of a linear threshold function over literals. These explicitly
learned rules could augment an explicitly designed KB.
}\fi
Chaining does not feature a rule of inference that allows new rules to be
derived, so chaining algorithms never need to consider rules outside $KB$.
Hence, the proposed generic transformation does not learn such rules.
Other logics such as resolution, which allow richer kinds of formulas to be
derived, yield more interesting query-driven learning.
In principle one could also consider a variant of backward-chaining in which
rules are learned as well. The difficulty with such a variant lies in
controlling the complexity of the search.
\subsection{Skeptical query-driven learning}
Theorem~\ref{explicit-search-thm} relies on an assumption of bounded concealment,
and uses a credulous learning strategy of searching for hypotheses for which we
do not possess counterexamples. A more conservative strategy would replace the
condition ``for no $\rho_i$ is $h|_{\rho_i}=0$'' in Algorithm~\ref{mod-bw-alg}
with the condition ``for all $\rho_i$, $h|_{\rho_i}=1$.'' An $h$ that is
witnessed true with probability 1
will pass this condition. Moreover, if $h|_{\rho_i}=1$ for $\rho_i=
m_i(x_i)$ (where $m_i$ is drawn from $M$ and $x_i$ is drawn from $D$), then
$h|_{x_i}=1$ as well. Thus, the probability that an $h$ that is {\em not}
$(1-\epsilon)$-valid with respect to $D$ passes this test for $m$ examples is
less than $(1-\epsilon)^m$. We can use this observation in place of Proposition
\ref{conceal-ce} to finish an analogous proof, given that we are searching for
an
$H$
that is always {\em witnessed}
rather than one that is merely perfectly
{\em valid}. We leave further details to an interested reader.
\subsubsection{Example application: learning input filters}
We note that our transformation for skeptical learning of rules could be applied to
static program analysis algorithms to automate the generation of sound and
approximately complete input filters. For example, the SIFT system~\cite{lskr14}
is based on a set of sound transformation rules for analyzing integer overflow
errors in a given program. Its associated static analysis algorithm uses the
knowledge base given by these transformation rules and the program code to
generate symbolic expressions that are propagated backwards from integer
operations that might produce overflows, until they refer only to program inputs.
The condition expressed by these expressions may then be used to filter out
inputs that do not satisfy the condition. The soundness of SIFT's transformation
rules ensures that no input that passes this condition generates an integer
overflow error.
We can interpret SIFT as taking the safety property of ``no integer overflows
occur at the given point in the program'' as a query, and seeking a proof of this
query using a property of the input, together with the transformation rules and
program code. Note that once SIFT generates expressions that refer only to input
values, the conditions they express are witnessed against example inputs, if they
are $(1-\epsilon)$-valid for inputs in practice. Thus, we can apply the transformation of the
skeptical variant of Algorithm~\ref{mod-bw-alg} to the static analysis
algorithm used by SIFT, and we would obtain an algorithm with a similar
termination condition, with the added requirement that the condition found must
be witnessed on a set of given examples.
Thus, the distinction between the approach taken by Long et al.~and our
transformation is that SIFT has no guarantee that it produces a rule that is
testable on real inputs.\footnote{%
Nevertheless, it has been empirically demonstrated that for a certain piece of
software, SIFT produces rules that are satisfied by real inputs with high
probability on a natural distribution~\cite{jmlsr15}.} SIFT does not use any training data, and simply
terminates once it finds a rule that refers only to the input; thus, while this
rule is sound -- inputs that satisfy it provably do not generate integer overflow
errors -- it has no guarantee of completeness, approximate or otherwise. Indeed,
SIFT conservatively considers all possible execution paths and relatedly uses
some hard-coded limits on, for example the number of loop iterations it considers
in search of loop invariants, to ensure termination. If this limit is exceeded
because the loop potentially relies on an unbounded number of values (for
example), then SIFT simply fails to find a condition. Thus, there is scope for a
backtracking variant of SIFT's static analysis algorithm to obtain greater
completeness by iteratively refining a symbolic condition. Under our
transformation, we would then obtain an algorithm that searches for a condition
on the inputs that empirically satisfies witnessing for a large fraction of a
training set of benign inputs.
\subsection{Query-driven learning in treelike resolution}
\label{explicit-discovery}
Recall that {\em resolution} is a logic that operates on {\em clauses}
(ORs of literals), using two kinds of inference rules: {\em cut} and (optionally)
{\em weakening}.
Cut takes two clauses $C=C'\vee \alpha$ and $D=D'\vee \neg \alpha$
and produces a clause of the form $C'\vee D'$.
(More general variants use substitutions to unify distinct atomic formuals
$\alpha$ and $\neg\alpha'$.) Weakening, by contrast, simply adds new literals
to the clause. Resolution is typically used to prove a DNF (an OR of ANDs) by
deriving an (unsatisfiable) empty clause from its negation, a CNF.
DPLL~\cite{dp60,dll62} is another example of such a goal-directed search
algorithm. In particular, bounded variants of DPLL that (efficiently) solve the
proof search problem for space-bounded treelike resolution are known~\cite
{kullmann99,et01}. This is the fragment of resolution refutations that can be
derived while {\em (i)} storing at most $s$ clauses in memory simultaneously and
{\em (ii)} ``forgetting'' a clause as soon as it is used in a proof step (so
that it must be derived again if it is needed again).
This is a second example of an algorithm into which we
can introduce query-driven explicit learning along the lines of Theorem~\ref
{explicit-search-thm}. The algorithm will be more interesting because it will
discover a CNF that suffices to complete the proof out of an exponentially
large (in terms of the number of ground atomic formulas) set of possible such
formulas.
Unfortunately, we cannot apply Theorem~\ref{explicit-search-thm} directly, as
the standard algorithm is not actually {\em oblivious} in our strict sense. The
difficulty is that the base case of the recursive algorithm involves a search
for a hypothesis clause for which we can derive the current clause via weakening.
This ``looking ahead'' into the hypothesis set means that the algorithm may not
engage in exactly the same search pattern if we modify the hypothesis set during
the search. Nevertheless, the use is innocuous enough that essentially the same
technique can be used to give a variant of the algorithm that learns an explicit
set of clauses from the data.
\begin{algorithm}[t]
\DontPrintSemicolon
\SetKwFunction{LSS}{Learn+SearchSpace}
\SetKwInOut{Input}{input}\SetKwInOut{Output}{output}
\Input{CNF $\varphi$, integer space bound $s\geq 1$, current clause $C$,
list of obscured scenes $\rho^{(1)},\ldots,\rho^{(m)}$.}
$\LSS(\varphi,s,C,(\rho^{(1)},\ldots,\rho^{(m)}))$
\Begin{
\If{No $\rho^{(i)}$ has $C|_{\rho^{(i)}}=0$}{
\Return{A proof asserting $C$ (from $H$).}}
\ElseIf{$C$ is a superset of some clause $C'$ of $\varphi$}{
\Return{The weakening derivation of $C$ from $C'$.}}
\ElseIf{$s>1$}{
\ForEach{Literal $\ell$ such that neither $\ell$ nor $\bar{\ell}$ is in $C$}{
\If{$\Pi_1\leftarrow$\LSS$(\varphi,s-1,C\vee \ell,(\rho^{(i)}:\ell|_{\rho^{(i)}}
\neq 1))$
does not return {\em none}}{
\If{$\Pi_2\leftarrow$\LSS$(\varphi,s,C\vee\bar{\ell},(\rho^{(i)}:
\ell|_{\rho^{(i)}}\neq 0))$ does not return {\em none}}{
\Return{Derivation of $C$ from $\Pi_1$ and $\Pi_2$}
}
\lElse{
\Return{{\em none}}
}
}
}
}
\Return{{\em none}}
}
\caption{Space-bounded resolution with learning}\label{mod-search-space}
\end{algorithm}
Our analysis of explicit query-driven learning (following
Theorem~\ref{explicit-search-thm}) still requires a bound on the lengths of the
proofs in terms of the space (and number of ground atomic formulas).
We will need the observation that DPLL-like algorithms produce {\em
normal} resolution proofs:
\begin{definition}[Normal]
We will say that a resolution proof is {\em normal} if in its corresponding DAG:
\begin{compactenum}
\item All outgoing edges from Cut nodes are directed to Cut nodes.
\item The clauses labeling any path to the sink from a Cut node contain literals
using every variable along the path.
\item A given variable is used in at most one cut step and at most one weakening
step along every path from a source to a Cut node.
\end{compactenum}
\end{definition}
\noindent
Our bound is now given in the following lemma, slightly modified from the work of
Ehrenfeucht and Haussler~\shortcite{eh89}.
\begin{lemma}[Lemma 1, Ehrenfeucht and Haussler 1989]\
\label{space-count}
Let $k$ be the number of nodes in a space-$s$ normal treelike resolution
proof over $N$ ground atomic formulas where $N\geq s\geq 1$. Then,
\begin{compactenum}
\item $2^s-1\leq k\leq 2(eN/(s-1))^{s-1}$
where $e$ is the base of the natural logarithm.
\item There are at most $(8N)^{(eN/(s-1))^{s-1}}$ space-$s$ normal
treelike proofs that do not use weakening.
\end{compactenum}
\end{lemma}
\iffalse{
\begin{proof}
For the first inequality of the first part, we argue by induction on $s$ that
the smallest proof requiring space $s$ is a complete binary tree of height $s$.
Naturally, this is trivial for space $1$, and for larger $s$ we use the
characterization of Proposition~\ref{optimal-space}: if the final step of the
derivation does involve two clauses requiring space $s-1$, then one of the
clauses derived already required space $s$ (and so there would be a smaller
proof). Thus, given that these minimal derivations are complete binary trees of
height $s-1$, the derivation is a complete binary tree of height $s$.
For the second inequality of the first part, we will argue by induction on
$N$ and $s$ that the number of leaves of a space-$s$ derivation involving the
cut rule on up to $N$ distinct ground atomic formulas is at most $\sum_{i=0}^{s-
1}{N\choose i}$. For $s=1$ this is immediate (for all $N$) and likewise for
$N=0$ (for all $s$). Now, for larger $N$ and $s$, we note that since the proof
is normal, if an atomic formula is used in an application of the cut rule in the
final step of the proof, i.e., at the root, that atomic formula appears in
every clause in the proof, and the remainder of the derivation invokes the cut
rule on at most $N-1$ distinct ground atomic formulas. Moreover, by the
characterization of Proposition~\ref{optimal-space} again, at most one of the
sub-derivations requires space-$s$ (and the other requires at most space-$s-1$).
Therefore, by the induction hypothesis, the number of leaves in the overall
derivation is at most
\[
\sum_{i=0}^{s-1}{N-1\choose i}+\sum_{i=0}^{s-2}{N-1\choose i}
={N-1\choose 0}+\sum_{i=1}^{s-1}\left[{N-1\choose i}+{N-1\choose i-1}\right]
=\sum_{i=0}^{s-1}{N\choose i}
\]
establishing the claim for $N$ and $s$. Towards the final bound, we note that
$\sum_{i=0}^{s-1}{N\choose i}\leq (eN/(s-1))^{s-1}$: in slightly more detail,
since $s-1<N$,
\begin{align*}
\left(\frac{s-1}{N}\right)^{s-1}\sum_{i=0}^{s-1}{N\choose i}&\leq
\sum_{i=0}^{s-1}\left(\frac{s-1}{N}\right)^i{N\choose i}\\
&=\left(1+\frac{s-1}{N}\right)^N\\
&<e^{s-1}
\end{align*}
where we can divide both sides by $((s-1)/N)^{s-1}$. The second inequality now
follows since the number of internal nodes in a binary tree is one less than the
number of leaves.
For the second part, we use the fact that a normal treelike proof that does not
use weakening is completely determined by the variables used in the invocation
of the cut rule (and whether the left or right child contains the positive
occurrence). Thus, it follows from the first part that we can bound the
number of possible proofs by counting the number of labellings of binary trees
with at most $L=(eN/(s-1))^{s-1}$ leaves. There are $\frac{1}{2i-1}{2i-1\choose
i}$ (unlabeled) binary trees with $i$ leaves, and so there are at most
\[
\sum_{i=0}^L\frac{(2N)^{i-1}}{2i-1}{2i-1\choose i}
\leq (2N)^{L-1}\sum_{i=1}^L{2L-1\choose i}\leq (2N)^{L-1}2^{2L-1}<(8N)^L
\]
possible proofs, establishing the claim.
\end{proof}
}\fi
\begin{theorem
\label{space-explicit-thm}
Let a clause $C$ and a (KB) CNF $\varphi$ be given. Suppose the examples are
drawn from a masking process that is $(1-\eta)$-concealing
with respect to CNFs of size $(eN/s-1)^{s-1}$ for the distribution $D$;
suppose further that $\varphi$ is perfectly valid with respect to $D$ and there
exists some other perfectly valid CNF $H$ for which there is a space-$s$
treelike resolution proof of $C$ from $\varphi\wedge H$. Then, Algorithm~\ref
{mod-search-space} run on $\varphi$ and $C$ with parameter $s$ on a sample of size
$\Theta((N^{s-1}\log N+\log\frac{1}{\delta})\frac{1}{\eta\epsilon})$ runs in time
$O(\frac{N^{2(s-1)}|\varphi|}{\eta\epsilon}(N^{s-1}\log N+\log\frac{1}{\delta}))$
and returns a proof of
$C$ from $\varphi\wedge H'$ for some CNF $H'$ of size $O(N^{s-1})$ that is
$(1-\epsilon)$-valid with respect to $D$ with probability $1-\delta$.
Similarly, if $[\varphi\Rightarrow C]$ is not $(1-\epsilon)$-valid, Algorithm~\ref
{mod-search-space} rejects in the same time bound using the same number of
examples.
\end{theorem}
\begin{proof}
First, consider the (possibly exponential size) formula $\tilde{H}$
consisting of all clauses that are consistent with the obscured scenes
$\rho^{(1)},\ldots,\rho^{(m)}$. Noting we don't need to use weakening for
clauses from $\tilde{H}$ since any clause that would be derived by weakening
is also in $\tilde{H}$, the standard analysis of algorithms for space-bounded
treelike resolution~\cite{kullmann99} establish that Algorithm~\ref
{mod-search-space} finds a normal space-$s$ treelike resolution proof of the
input $C$ from $\varphi\wedge\tilde{H}$ and runs in time
$O(m|\varphi|N^{2(s-1)})$ on a sample of size $m$. When a $1$-valid space-$s$
treelike proof is assumed to exist, it is in particular consistent with every
sample with probability $1$, so the algorithm outputs some proof in this case.
It remains only to show that, for a
sample of size $\Theta(\frac{1}{\eta\epsilon}(N^{(s-1)}\log N+\log\frac{1}
{\delta}))$, any such proof has its leaves labeled with a $(1-\epsilon)$-valid CNF
with probability at least $1-\delta$ (and so in particular, the algorithm cannot
return a proof if $[\varphi\Rightarrow C]$ is not $(1-\epsilon)$-valid).
Consider any space-$s$ normal treelike resolution proof that has
leaves that are {\em not} labeled by a $(1-\epsilon)$-valid CNF. Since, by part
1 of Lemma~\ref{space-count}, this CNF has at most $O(N^{s-1})$ clauses,
Proposition~\ref{conceal-ce} shows that each example produces a counterexample
to this CNF with probability at least $\eta\epsilon$. Thus, in a sample of size
$\Omega(\frac{1}{\eta\epsilon}(N^{s-1}\log N+\log\frac{1}{\delta}))$, the
probability that this CNF is consistent with the sample is at most $\delta N^{-
\Omega(N^{s-1})}$. Now, by part 2 of Lemma~\ref{space-count}, there are at
most $N^{O(N^{s-1})}$ possible proofs (up to weakening steps), where we notice
that Algorithm~\ref{mod-search-space} introduces weakening steps iff the clause
is consistent with some clause of the (fixed) input CNF $\varphi$. Therefore, the
algorithm indeed considers at most $N^{O(N^{s-1})}$ distinct proofs, so for
a suitable choice of constant in the sample size, a union bound gives the
probability that Algorithm~\ref{mod-search-space} encounters some proof from a
CNF that is {\em not} $(1-\epsilon)$-valid but consistent with the sample is at
most $\delta$. Since the algorithm only outputs proofs that are consistent
with the sample, we therefore find that any proof it outputs is derived from a
CNF that is $(1-\epsilon)$-valid with respect to $D$ with probability at least
$1-\delta$, establishing the claim.
\end{proof}
Lemma~\ref{space-count} also shows treelike proofs of size $k$ are necessarily
space-$\log(k+1)$, so this gives a quasipolynomial time and sample complexity
algorithm for general treelike proofs.
\section{Directions for future work}
One main direction for future work concerns a slightly relaxed variant of the
skeptical learning strategy, in which we try to learn rules that are
simultaneously witnessed true with maximum frequency, which may be less than
$1$.
What makes this formulation challenging is that it cannot be achieved by just
seeking that the formulas at the individual nodes are each witnessed with some
probability $1-\epsilon$: it matters which $\epsilon$-fraction of example scenes
are not witnessed across the various rules,
since we are interested in how many examples in total fail to witness at least
one of the hypotheses used in the proof. This formulation seems to be a much
harder computational problem, so it makes sense to ask what kind of approximate
solutions can be obtained.
Another direction for work concerns relational reasoning algorithms. The usual
algorithms for backward search in relational reasoning generate substitutions
during the search, rather than generating grounded versions of the rules as we
do here. Although this creation of substitutions may create a ``non-oblivious''
search (as viewed on the ground atomic formulas), we suspect that it may again
be innocuous enough that our main theorem will still hold, as in the example of
weakening in resolution. Such algorithms would be much more efficient, of
course.
The main difference is that we now need to identify unifiers against the set of
(all) example scenes (in addition to formulas in the KB).
A related question, along the lines of the above but perhaps more ambitious is,
can we learn universally quantified expressions in infinite or open domains?
Again, our current method generates ground expressions, and so we can only
consider domains with a reasonable number of elements. But, the credulous
bounded
concealment definition at least raises the possibility that we might be able
to infer a universally quantified statement based merely on the lack of observed
counterexamples. Of course, such inferences lean heavily on the bounded
concealment assumption.
A final direction is the development of further applications of such
algorithms. In addition to the natural application of the
algorithm in extracting knowledge that is suitable for human inspection in query
driven learning, we have identified two domains in which the kind
of sound and approximately complete rules our method generates would provide a
useful filtering criterion. It is natural to ask if there are any others.
\section*{Acknowledgements}
This work was supported in part by ONR grant number N000141210358 while the
author was affiliated with Harvard University, and by NSF Award CCF-1718380.
This manuscript was prepared in part while the author was visiting the Simons
Institute for the Theory of Computing and Google.
\bibliographystyle{named}
|
\section{Introduction}
The behavior of an average automobile driver is based on numerous factors. In this work, we focus on one aspect of behavior: the speed of travel. We specifically focus on the free-flow speed, which is the average vehicle speed along a roadway when there is no congestion or adverse weather conditions. The free-flow speed is used in a wide variety of planning and regulatory contexts. Standard practice for estimating free-flow speeds relies on the availability of many road features, such as lane widths and curvature. While some of these features are easy to gather, others require expensive surveying equipment and expert annotators, restricting their availability. As a result, in many states, roads are assigned the default speed limit of 35 mph for urban areas and 55 mph for rural areas, which can often be inconsistent with actual operational speeds and lead to undesired driver behavior.
To automate the costly process of estimating free-flow speeds, we propose a deep convolutional neural network (CNN) that takes as input the overhead imagery of the roadway and coarse-grained road features, including the posted speed limit, the functional classification, and the type of area. As output, our CNN generates a probability distribution over integer free-flow speeds. For training data, we use aggregated data collected from real-world drivers. At inference time, since the input features are all easily obtainable, it should be possible to rapidly, and inexpensively, estimate free-flow speeds over large spatial regions.
Using a large-scale evaluation dataset, we find that using only image features result in the poorest performance, but that combining imagery and road features result in significantly better performance than road features alone.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{network}
\caption{Network architecture for predicting free-flow speeds.}
\label{fig:network}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=\textwidth]{distribution}
\caption{Histograms of various road features in our dataset.}
\label{fig:distribution}
\end{figure*}
\section{Related Work}
Different studies have been proposed to estimate and map properties of the visual world using overhead images. Several authors have proposed different deep learning based approaches for vehicle detection~\cite{sakla2017deep,sommer2017fast} and road extraction~\cite{zhang2018road, bastani2018roadtracer, alshehhi2017hierarchical, mattyus2017deeproadmapper} from aerial images. Salem et al.~\cite{salem2018soundscape} introduced an approach for mapping soundscapes of geographic regions using overhead imagery. Greenwell et al.~\cite{greenwell2018objects} proposed a model that is capable of predicting object histograms from overhead imagery. Several works have addressed the problem of speed estimation. Shuai et al.~\cite{hua2018vehicle} introduced an approach to estimate the vehicle speed in traffic videos. Most similar to our work, Song et al.~\cite{song2018farsa} proposed a model for road safety estimation based on the usRAP Star Rating Protocol. While this star rating is based on approximately 60 road safety features~\cite{harwood2010validation}, their network works directly on ground level panorama images. We propose a new method that instead uses overhead imagery with the addition of auxiliary features.
\section{Approach}
\label{sec:approach}
We utilize a CNN architecture to estimate the free-flow speed of a given road segment. The neural network uses both aerial imagery and relevant road features as input, and outputs a probability mass function over $K$ possible free-flow speeds. We begin by describing the dataset that we use, followed by a more detailed description of the proposed network architecture.
\subsection{Dataset}
\label{ssec:dataset}
Our free-flow speed dataset is obtained through HERE Technologies and further annotated to incorporate coarse-grained road feature data needed for training. To calculate the free-flow speed for a particular road segment, driving speeds of vehicles were monitored during the year 2014. As free-flow speed refers to the speed that a driver can achieve without traffic congestion, we only consider data during non-holiday weekday periods from 9am to 3pm. The speeds are then averaged to obtain the ground-truth for each road segment. We rounded each ground-truth speed to the nearest integer to obtain a discrete label for training. This results in $K=79$ unique free-flow speeds.
For each road segment, we obtain an aerial image through the National Agriculture Imagery Program (NAIP) data. Each image is centered at the beginning of the segment, and covers an aerial view of $400\times 400$ m\textsuperscript{2} of land. We orient the images such that the direction of travel is up, and we reshape it to $224\times 224$. In total, our dataset consists of \num{1335132} road segments and their corresponding aerial images.
\figref{distribution} shows the distribution of the free-flow speeds along with road features used for training. The drastic difference between the distribution of free-flow speeds and speed limits is evident here. In fact, over half of the roads have free-flow speeds significantly slower than the speed limits (more than 10-miles-per-hour). We will examine a few roads with such differences in Section~\ref{ssec:qual_eval}, using predictions obtained from our model.
We split our dataset into training and testing sets. To ensure the sets are disjoint, we partition the dataset by location. The dataset is divided into counties, eight of which are selected for evaluation. The resulting test set is roughly $7\%$ of the total data, and contains a variety of roads with different area types, terrain, and free-flow speeds. A validation set ($1\%$ of the training set) is reserved for model selection.
\subsection{Network Architecture}
\label{ssec:network_arch}
The backbone of our architecture is the Xception~\cite{chollet2016xception} model, which is used to extract high-level features from the aerial image. The resulting feature vector of length $2048$ is passed to a dense layer with output size $512$, denoted as $S$, as shown in~\figref{network}.
In addition to aerial imagery, the network is trained using three integer features related to driving speed: area type, functional classification, and posted speed limit. The three features are concatenated with $S$ before feeding the combined feature vector into
the final dense layer with output size $K$. We denote the output of the final layer as $\hat{y}$.
\subsection{Loss Function}
\label{ssec:loss}
We formulate the task of predicting the free-flow speed of a road segment as a multiclass classification problem with ground-truth free-flow speed label $l$. With the output of the network defined in \ref{ssec:network_arch}, we compute the training loss as the standard cross-entropy loss between the predicted distribution $\hat{y}$ and the target distribution $y$, as shown in the following equation:
\begin{equation}
L=-\frac{1}{N}\sum_{i=1}^N\sum_{k=1}^K y_{i,k}\log \hat{y}_{i,k}
\end{equation}
where $N$ is the number of training examples.
\subsection{Implementation Details}
\label{ssec:imple_details}
Our model is implemented using Tensorflow. The network is optimized using the Adam optimizer with default parameters. We initialize the Xception network with weights pre-trained on ImageNet for image classification. During training, we freeze the weights of the Xception network and optimize only the last two dense layers with learning rate $0.001$. We decay the learning rate exponentially by a factor of 10 every 5 epochs. {\em ReLU} activation layers are used throughout the network except for the last layer, which uses the Softmax activation instead. We apply $L_2$ regularization to the two dense layers with scale $0.00005$. We train with batch size 16, for a total of 15 epochs.
\begin{table}[]
\centering
\caption{Within-5 accuracy for each method.}
\begin{tabular}{lc}
\toprule
\multicolumn{1}{c}{Method} & \multicolumn{1}{c}{Within-5 Acc.} \\ \hline
Imagery Only & 37.60 \\
Road Features Only & 40.07 \\
Combined & \textbf{49.86} \\
\bottomrule
\end{tabular}
\label{tab:eval}
\end{table}
\begin{figure}
\centering
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/18_55_13_11473.jpg}
\caption{Free-flow: 18, Limit: 55}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/17_55_15_6015.jpg}
\caption{Free-flow: 17, Limit: 55}
\end{subfigure}
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/27_15_22_1533.jpg}
\caption{Free-flow: 27, Limit: 15}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/67_55_66_38641.jpg}
\caption{Free-flow: 67, Limit: 55}
\end{subfigure}
\caption{Examples of discrepancies in the predicted free-flow speed and the speed limit of roads in miles per hour. Predictions obtained from the combined feature model.}
\label{fig:discrep}
\end{figure}
\begin{figure}[t!]
\centering
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/50_16_15_26688.jpg}
\caption*{method A: 50, method B: 16}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/31_42_47_64489.jpg}
\caption*{method A: 31, method B: 42}
\end{subfigure}
\hfill
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/46_15_20_32612.jpg}
\caption*{method A: 46, method B: 15}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/24_50_24_20339.jpg}
\caption*{method A: 24, method B: 50}
\end{subfigure}
\hfill
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/30_15_14_68180.jpg}
\caption*{method A: 30, method B: 15}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/24_40_22_19983.jpg}
\caption*{method A: 24, method B: 40}
\end{subfigure}
\caption{Free-flow speed predictions in miles per hour from two methods, where method A is the road feature only model and method B is the combined feature model.}
\label{fig:versus}
\end{figure}
\section{Evaluation}
\label{sec:eval}
Using the dataset described in \ref{ssec:dataset}, we trained our proposed model along with two other variations: imagery-only model and road-features-only model. We conducted quantitative and qualitative evaluation on our reserved test set.
\subsection{Quantitative Analysis}
\label{ssec:quant_eval}
We aim to discover the effect of each input modality in predicting free-flow speeds. We calculate the accuracy of the three trained models (overhead imagery only, road features only, and both) on the test set. \tabref{eval} displays the within-5 accuracy for each method, where we consider a prediction to be positive if it is within five-miles-per-hour of the ground-truth speed. As we can see, road feature data is better than aerial imagery for predicting free-flow speeds, but we obtain the best performance when we combine both modalities.
\subsection{Qualitative Evaluation}
\label{ssec:qual_eval}
Free-flow speed should be similar to the speed limit for a given road, but that is often not the case. We are interested in finding roads where the predicted free-flow speed is drastically different than the speed limit. \figref{discrep} shows such examples. The top two roads have free-flow speeds much lower than the speed limits, most likely due to the curvature and number of intersections present. The bottom two roads have free-flow speeds much higher than the speed limits, which is understandable since the roads are straight, wide, and without congestion. Detection of these discrepancies is very useful, as traffic engineers can quickly filter through millions of roads to identify ones that need to have their speed limits re-evaluated.
We also compared predictions from our top two models: one trained on both aerial imagery and road features, and the other trained only on road features. \figref{versus} shows the images that had drastically different predictions from the two models. The images in the right column received higher free-flow speed predictions from the combined model, where the images in the left column received higher free-flow speed predictions from the feature only model. These differences further support the importance of image features for prediction, as imagery can provide fine-grained information about the road segment. For the left three roads, it is natural for people to drive slower in residential areas or at intersections; for the right three, straight roads and ones parallel to interstates allow drivers to achieve greater speeds. Fine-grained labels such as presence of intersections, population density, and curvature of the road are hard to obtain, but they can easily be inferred from aerial imagery.
\section{Conclusion}
\label{sec:conclusion}
We introduced a method for estimating the free-flow speed of a road segment, which is important for understanding driver behavior. We demonstrated that a combination of aerial imagery and related road features as input is best for prediction, since it obtained higher accuracy than models trained on either features alone. We also performed qualitative evaluation, and obtained insights on the effect of input modalities along with the relationship between free-flow speeds and speed limits. We hope to extend this work and include ground-level imagery as an additional input, since information such as roadside hazards may not be visible in aerial imagery.
\section*{Acknowledgements}
We thank Mei Chen and Xu Zhang, both from the Univ.\ of Kentucky Dept.\ of Civil Engineering, for their assistance in obtaining and understanding this dataset. We also acknowledge the support of NSF CAREER (IIS-1553116).
{\small
\bibliographystyle{IEEEbib}
\section{Introduction}
The behavior of an average automobile driver is based on numerous factors. In this work, we focus on one aspect of behavior: the speed of travel. We specifically focus on the free-flow speed, which is the average vehicle speed along a roadway when there is no congestion or adverse weather conditions. The free-flow speed is used in a wide variety of planning and regulatory contexts. Standard practice for estimating free-flow speeds relies on the availability of many road features, such as lane widths and curvature. While some of these features are easy to gather, others require expensive surveying equipment and expert annotators, restricting their availability. As a result, in many states, roads are assigned the default speed limit of 35 mph for urban areas and 55 mph for rural areas, which can often be inconsistent with actual operational speeds and lead to undesired driver behavior.
To automate the costly process of estimating free-flow speeds, we propose a deep convolutional neural network (CNN) that takes as input the overhead imagery of the roadway and coarse-grained road features, including the posted speed limit, the functional classification, and the type of area. As output, our CNN generates a probability distribution over integer free-flow speeds. For training data, we use aggregated data collected from real-world drivers. At inference time, since the input features are all easily obtainable, it should be possible to rapidly, and inexpensively, estimate free-flow speeds over large spatial regions.
Using a large-scale evaluation dataset, we find that using only image features result in the poorest performance, but that combining imagery and road features result in significantly better performance than road features alone.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{network}
\caption{Network architecture for predicting free-flow speeds.}
\label{fig:network}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=\textwidth]{distribution}
\caption{Histograms of various road features in our dataset.}
\label{fig:distribution}
\end{figure*}
\section{Related Work}
Different studies have been proposed to estimate and map properties of the visual world using overhead images. Several authors have proposed different deep learning based approaches for vehicle detection~\cite{sakla2017deep,sommer2017fast} and road extraction~\cite{zhang2018road, bastani2018roadtracer, alshehhi2017hierarchical, mattyus2017deeproadmapper} from aerial images. Salem et al.~\cite{salem2018soundscape} introduced an approach for mapping soundscapes of geographic regions using overhead imagery. Greenwell et al.~\cite{greenwell2018objects} proposed a model that is capable of predicting object histograms from overhead imagery. Several works have addressed the problem of speed estimation. Shuai et al.~\cite{hua2018vehicle} introduced an approach to estimate the vehicle speed in traffic videos. Most similar to our work, Song et al.~\cite{song2018farsa} proposed a model for road safety estimation based on the usRAP Star Rating Protocol. While this star rating is based on approximately 60 road safety features~\cite{harwood2010validation}, their network works directly on ground level panorama images. We propose a new method that instead uses overhead imagery with the addition of auxiliary features.
\section{Approach}
\label{sec:approach}
We utilize a CNN architecture to estimate the free-flow speed of a given road segment. The neural network uses both aerial imagery and relevant road features as input, and outputs a probability mass function over $K$ possible free-flow speeds. We begin by describing the dataset that we use, followed by a more detailed description of the proposed network architecture.
\subsection{Dataset}
\label{ssec:dataset}
Our free-flow speed dataset is obtained through HERE Technologies and further annotated to incorporate coarse-grained road feature data needed for training. To calculate the free-flow speed for a particular road segment, driving speeds of vehicles were monitored during the year 2014. As free-flow speed refers to the speed that a driver can achieve without traffic congestion, we only consider data during non-holiday weekday periods from 9am to 3pm. The speeds are then averaged to obtain the ground-truth for each road segment. We rounded each ground-truth speed to the nearest integer to obtain a discrete label for training. This results in $K=79$ unique free-flow speeds.
For each road segment, we obtain an aerial image through the National Agriculture Imagery Program (NAIP) data. Each image is centered at the beginning of the segment, and covers an aerial view of $400\times 400$ m\textsuperscript{2} of land. We orient the images such that the direction of travel is up, and we reshape it to $224\times 224$. In total, our dataset consists of \num{1335132} road segments and their corresponding aerial images.
\figref{distribution} shows the distribution of the free-flow speeds along with road features used for training. The drastic difference between the distribution of free-flow speeds and speed limits is evident here. In fact, over half of the roads have free-flow speeds significantly slower than the speed limits (more than 10-miles-per-hour). We will examine a few roads with such differences in Section~\ref{ssec:qual_eval}, using predictions obtained from our model.
We split our dataset into training and testing sets. To ensure the sets are disjoint, we partition the dataset by location. The dataset is divided into counties, eight of which are selected for evaluation. The resulting test set is roughly $7\%$ of the total data, and contains a variety of roads with different area types, terrain, and free-flow speeds. A validation set ($1\%$ of the training set) is reserved for model selection.
\subsection{Network Architecture}
\label{ssec:network_arch}
The backbone of our architecture is the Xception~\cite{chollet2016xception} model, which is used to extract high-level features from the aerial image. The resulting feature vector of length $2048$ is passed to a dense layer with output size $512$, denoted as $S$, as shown in~\figref{network}.
In addition to aerial imagery, the network is trained using three integer features related to driving speed: area type, functional classification, and posted speed limit. The three features are concatenated with $S$ before feeding the combined feature vector into
the final dense layer with output size $K$. We denote the output of the final layer as $\hat{y}$.
\subsection{Loss Function}
\label{ssec:loss}
We formulate the task of predicting the free-flow speed of a road segment as a multiclass classification problem with ground-truth free-flow speed label $l$. With the output of the network defined in \ref{ssec:network_arch}, we compute the training loss as the standard cross-entropy loss between the predicted distribution $\hat{y}$ and the target distribution $y$, as shown in the following equation:
\begin{equation}
L=-\frac{1}{N}\sum_{i=1}^N\sum_{k=1}^K y_{i,k}\log \hat{y}_{i,k}
\end{equation}
where $N$ is the number of training examples.
\subsection{Implementation Details}
\label{ssec:imple_details}
Our model is implemented using Tensorflow. The network is optimized using the Adam optimizer with default parameters. We initialize the Xception network with weights pre-trained on ImageNet for image classification. During training, we freeze the weights of the Xception network and optimize only the last two dense layers with learning rate $0.001$. We decay the learning rate exponentially by a factor of 10 every 5 epochs. {\em ReLU} activation layers are used throughout the network except for the last layer, which uses the Softmax activation instead. We apply $L_2$ regularization to the two dense layers with scale $0.00005$. We train with batch size 16, for a total of 15 epochs.
\begin{table}[]
\centering
\caption{Within-5 accuracy for each method.}
\begin{tabular}{lc}
\toprule
\multicolumn{1}{c}{Method} & \multicolumn{1}{c}{Within-5 Acc.} \\ \hline
Imagery Only & 37.60 \\
Road Features Only & 40.07 \\
Combined & \textbf{49.86} \\
\bottomrule
\end{tabular}
\label{tab:eval}
\end{table}
\begin{figure}
\centering
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/18_55_13_11473.jpg}
\caption{Free-flow: 18, Limit: 55}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/17_55_15_6015.jpg}
\caption{Free-flow: 17, Limit: 55}
\end{subfigure}
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/27_15_22_1533.jpg}
\caption{Free-flow: 27, Limit: 15}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{discrepencies/67_55_66_38641.jpg}
\caption{Free-flow: 67, Limit: 55}
\end{subfigure}
\caption{Examples of discrepancies in the predicted free-flow speed and the speed limit of roads in miles per hour. Predictions obtained from the combined feature model.}
\label{fig:discrep}
\end{figure}
\begin{figure}[t!]
\centering
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/50_16_15_26688.jpg}
\caption*{method A: 50, method B: 16}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/31_42_47_64489.jpg}
\caption*{method A: 31, method B: 42}
\end{subfigure}
\hfill
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/46_15_20_32612.jpg}
\caption*{method A: 46, method B: 15}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/24_50_24_20339.jpg}
\caption*{method A: 24, method B: 50}
\end{subfigure}
\hfill
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/30_15_14_68180.jpg}
\caption*{method A: 30, method B: 15}
\end{subfigure}%
\begin{subfigure}{.45\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{versus/24_40_22_19983.jpg}
\caption*{method A: 24, method B: 40}
\end{subfigure}
\caption{Free-flow speed predictions in miles per hour from two methods, where method A is the road feature only model and method B is the combined feature model.}
\label{fig:versus}
\end{figure}
\section{Evaluation}
\label{sec:eval}
Using the dataset described in \ref{ssec:dataset}, we trained our proposed model along with two other variations: imagery-only model and road-features-only model. We conducted quantitative and qualitative evaluation on our reserved test set.
\subsection{Quantitative Analysis}
\label{ssec:quant_eval}
We aim to discover the effect of each input modality in predicting free-flow speeds. We calculate the accuracy of the three trained models (overhead imagery only, road features only, and both) on the test set. \tabref{eval} displays the within-5 accuracy for each method, where we consider a prediction to be positive if it is within five-miles-per-hour of the ground-truth speed. As we can see, road feature data is better than aerial imagery for predicting free-flow speeds, but we obtain the best performance when we combine both modalities.
\subsection{Qualitative Evaluation}
\label{ssec:qual_eval}
Free-flow speed should be similar to the speed limit for a given road, but that is often not the case. We are interested in finding roads where the predicted free-flow speed is drastically different than the speed limit. \figref{discrep} shows such examples. The top two roads have free-flow speeds much lower than the speed limits, most likely due to the curvature and number of intersections present. The bottom two roads have free-flow speeds much higher than the speed limits, which is understandable since the roads are straight, wide, and without congestion. Detection of these discrepancies is very useful, as traffic engineers can quickly filter through millions of roads to identify ones that need to have their speed limits re-evaluated.
We also compared predictions from our top two models: one trained on both aerial imagery and road features, and the other trained only on road features. \figref{versus} shows the images that had drastically different predictions from the two models. The images in the right column received higher free-flow speed predictions from the combined model, where the images in the left column received higher free-flow speed predictions from the feature only model. These differences further support the importance of image features for prediction, as imagery can provide fine-grained information about the road segment. For the left three roads, it is natural for people to drive slower in residential areas or at intersections; for the right three, straight roads and ones parallel to interstates allow drivers to achieve greater speeds. Fine-grained labels such as presence of intersections, population density, and curvature of the road are hard to obtain, but they can easily be inferred from aerial imagery.
\section{Conclusion}
\label{sec:conclusion}
We introduced a method for estimating the free-flow speed of a road segment, which is important for understanding driver behavior. We demonstrated that a combination of aerial imagery and related road features as input is best for prediction, since it obtained higher accuracy than models trained on either features alone. We also performed qualitative evaluation, and obtained insights on the effect of input modalities along with the relationship between free-flow speeds and speed limits. We hope to extend this work and include ground-level imagery as an additional input, since information such as roadside hazards may not be visible in aerial imagery.
\section*{Acknowledgements}
We thank Mei Chen and Xu Zhang, both from the Univ.\ of Kentucky Dept.\ of Civil Engineering, for their assistance in obtaining and understanding this dataset. We also acknowledge the support of NSF CAREER (IIS-1553116).
{\small
\bibliographystyle{IEEEbib}
|
\subsubsection{#1}\vspace{-3\baselineskip}\color{black}\medskip{\noindent \bf \thesubsubsection. #1.}}
\newcommand{\myparagraph}[1]{\needspace{1\baselineskip}\medskip\noindent {\it #1.}}
\newcommand{\myindentedparagraph}[1]{\needspace{1\baselineskip}\medskip \hangindent=11pt \hangafter=0 \noindent{\it #1.}}
\newcommand{\myparagraphtc}[1]{\needspace{1\baselineskip}\medskip\noindent {\it #1.}\addcontentsline{toc}{subsubsection}{\qquad\qquad\quad#1}}
\section{Introduction}
Modern society has witnessed the appearance of heterogeneous wireless services. These services demand different facilities and operate their own networks respectively, which leads to the high cost and slows down the process of deploying new wireless services. Radio over Fiber (RoF) is put forward as a universal platform that connects multiple radio frequency (RF) signals from different wireless access networks. By placing RF signals on optical carriers, RoF system transmits them through optical fibers without changing their radio formats. It takes advantage of high data-rates, low loss and zero interference, but heavily relies on the deployment of fibers, which may not be available in places like rural areas \cite{Wake2010}.
Free Space Optical (FSO) communication becomes a promising alternative when fibers are not available. With similar advantages as optical fiber communication, it also enjoys license free, easy and inexpensive setup \cite{Chang2006, Andrews2005}. Wireless FSO links get rid of physical restriction of fiber deployment, and is able to transmit RF signals through free space. Therefore, so-called Radio on Free Space Optics (RoFSO) system has been developed recently \cite{Kazaura2010}. However, RoFSO can be seriously affected by FSO channel characteristics, such as weather, turbulence, etc. Different models are proposed for the FSO channel and various techniques are developed to reduce its influence \cite{Andrews2005, Nistazakis2009, Gao2017, Gao2018}.
To improve the performance, Dense Wavelength Division Multiplexing (DWDM) RoFSO system has been developed, as a means of transmitting multiple RF signals simultaneously. It makes it feasible to employ Wavelength Division Multiplexing (WDM) in RoFSO \cite{Bekkali2009, Ciaramella2009}. At the same time, adaptive transmission based on the channel state information (CSI) is proposed to help mitigate channel effects for FSO \cite{Karimi2012, Liu2009} and RoFSO systems \cite{Kim2011, Zhou2015}.
The problem considered in this paper is the optimal power allocation for adaptive WDM transmission in RoFSO systems. According to CSI of all wavelength links, different powers are assigned to different wavelengths to maximize the objective function, subject to power limitation constraints necessary for safe implementation of RoFSO systems.
The problem is challenging not only because it is both non-convex and constrained, but also because the mathematical system model or estimated CSI may not be accurate in practice. Some model-based algorithms have been developed to handle similar problems \cite{Kim2011,Zhou2015}. These algorithms both employ relaxations to find inexact solutions and are computationally expensive to implement \cite{Zhou2015}. The inherent difficulty makes the application of machine learning appealing, due to both their low complexity and potential for model-free implementation. Deep learning in particular has been applied for resource allocation problems in wireless RF domain in both supervised \cite{sun2017learning} and unsupervised \cite{xu2017deep, eisen2019learning} manners. Such approaches have not yet been explored in FSO or RoFSO systems.
This paper develops two algorithms to solve power allocation for WDM RoFSO. We first formulate the optimal design problem and introduce the RoFSO system model (Section \ref{sec_problem}). We present the Stochastic Dual Gradient algorithm to solve the problem exactly using the idea of strong duality in \cite{Ribeiro2012} (Section \ref{sec_sdg}). This approach is limited in practice as it is dependent upon system models and requires more computational expense. As a model-free and low complexity alternative, we leverage machine learning techniques in the Primal-Dual Deep Learning algorithm (Section \ref{sec_pddl}). In particular, Deep Neural Networks (DNNs) are used to parameterize the power allocation policy, which are trained with a primal-dual method to solve the resulting constrained learning problem. A model-free implementation is employed using the policy gradient method for cases in which system models are inaccurate or unknown. The strong performance of both algorithms are shown by numerical simulations (Section \ref{sec_numerical_results}).
\section{Problem Formulation}\label{sec_problem}
Radio on Free Space Optics (RoFSO), as a universal platform for heterogeneous wireless services, can transmit RF signals through FSO links in optical networks. The developed Dense Wavelength Division Multiplexing (DWDM) RoFSO system enables the simultaneous transmission of multiple RF signals with WDM technique to increase transmission capacity. Specifically, multimedia RF signals are accessed into RoFSO system and placed on multiple optical wavelength carriers with optoelectronic devices, and then transmitted into free space. At the receiver, optical signals are received through FSO channels, and transferred back to RF signals for users.
The adaptive transmission is considered when allocating powers to wavelength channels in RoFSO. Based on the channel state information (CSI), different powers are assigned to different wavelengths to maximize the objective function. The exact objective function can be adjusted according to specific situations.
Assume there are $m$ optical wavelengths carrying different transmissions, and each of them are non-overlapping with enough spacing. The CSI is represented by the vector $\bbh = [h_1,...,h_m]$, where each $h_i (i=1,...,m)$ donates the CSI of $i$-th wavelength channel. The allocated power to signal transmitted on the $i$-th wavelength is based upon observed CSI $\bbh$ via a power allocation policy $P_i(\bbh)$. Given the collection of power allocations $\bbP(\bbh) = [P_1(\bbh),...,P_m(\bbh)]$ and current CSI $\bbh$, a channel capacity of $C_i(\bbP(\bbh),\bbh)$ is achieved on the $i$-th wavelength. Note that FSO channel is considered as a fading process with channel coherence time on the order of milliseconds, so we can assume an ergodic and i.i.d block fading process. Since the instantaneous channel capacity tends to vary fast, a long term average $\mathbb{E}_\bbh[C_i(\bbP(\bbh),\bbh)]$ is the more meaningful metric to consider. Additionally, because different wireless services accessed into RoFSO may have different priorities, we consider the weight vector $\bbomega = [\omega_1,...,\omega_m] \ge 0$ to represent such priorities.
There are two natural constraints to be considered in RoFSO power allocation. The first is the expected total power limitation $P_T$ for the FSO base station:
\begin{equation}
\mathbb{E}_\textbf{h} \left[ \sum_{i=1}^m P_i(\bbh) \right] \le P_T.
\end{equation}
The second is motivated by the eye safety concern in optical transmissions. Specifically, we set a peak power $P_S$ that can be allocated on any single wavelength so that the beam is not dangerous for human eyes in its propagation:
\begin{equation}
0 \le P_i(\bbh) \le P_S, i=1,...,m.
\end{equation}
Together, we formulate the optimal power allocation for adaptive WDM transmission in RoFSO systems as the following statistical optimization problem:
\begin{alignat}{3} \label{eq_problem111}
\mathbb{P}:= & \max_{\bbP(\bbh)} \ && \sum_{i=1}^m \omega_i \mathbb{E}_\bbh \left[C_i(\bbP(\bbh),\bbh)\right] , \\
& \st \ && \mathbb{E}_\textbf{h} \left[ \sum_{i=1}^m P_i(\bbh) \right] \le P_T, \nonumber \\
& \ && 0 \le P_i(\bbh) \le P_S, i=1,...,m. \nonumbe
\end{alignat}
Note that the above problem is formulated without any specific system model. In the proceeding subsection, we discuss channel and capacity models commonly used in the study of RoFSO systems. We then present an exact algorithm to solve \eqref{eq_problem111} in Section \ref{sec_sdg} that relies on such model information, as well as a deep-leraning based alternative algorithm in Section \ref{sec_pddl} that does not.
\subsection{System Model} \label{sec_models}
To mathematically study the FSO channel and RoFSO system, some theoretical models have been put forward in previous researches. For the FSO channel, its effects mainly consist of two parts: the attenuation $h_a$ and the turbulence $h_t$ \cite{Zhou20151}.
The attenuation fading term $h_a$ can be expressed by
\begin{equation}
\begin{split}
&h_a = A(d,\lambda)e^{-\alpha d},\\
&A(d,\lambda)=\frac{A_{TX}A_{RX}}{(d\lambda)^2},
\end{split}
\end{equation}
where $\alpha$ is the attenuation coefficient; $d$ is the transmission distance; $\lambda$ is the wavelength; $A_{TX}$ is the aperture area of transmitter, and $A_{RX}$ is the aperture area of receiver.
As for the turbulence, we use the well-known Log-normal distribution to model the fading term $h_t$, which is considered to be accurate under weak-to-moderate turbulence. Without loss of generality, we can also use other distributions like Gamma-gamma distribution according to different turbulence conditions \cite{Nistazakis2009, Uysal2006}.
The FSO channel can then be modelled as
\begin{equation}
\begin{split}
y = h_a h_t x + n,
\end{split}
\end{equation}
in which $y$ is the received signal; $x$ is the transmitted signal, and $n$ represents the additive Gaussian noise. Therefore, the channel gain (referred as CSI) under this model is expressed by
\begin{equation}
\begin{split}
h = \frac{\vert h_a h_t \vert^2}{N_0}.
\end{split}
\end{equation}
In terms of the RoFSO system with APD photo detector, its performance is commonly evaluated by Carrier to Noise Ratio (CNR), which is modelled as \cite{Dat2009}
\begin{equation}
\begin{split}
CNR = \frac{\frac{1}{2}(OMI\cdot m_prP_r)^2}{RIN\cdot (rP_r)^2+2em_p^{2+F}rP_r+\frac{4KT}{R_f}},
\end{split}
\end{equation}
where $OMI$ donates the optical modulation index; $RIN$ donates the relative intensity noise; $m_p$ is the photodiode gain; $r$ is the photodiode responsivity; $e$ is the electric charge; $F$ is the excess noise factor; $K$ is the Boltzmann's constant; $T$ is the temperature; $R_f$ is the photodiode resistance, and $P_r = Ph$ is the received power at the detector.
With this specific RoFSO system model, the capacity of $i$-th wavelength channel with allocated power $\bbP$ and CSI $\bbh$ can be expressed by
\begin{equation} \label{eq_capacity}
\begin{split}
&C_i(\bbP,\bbh)= \log \left( 1+ CNR_i(\bbP,\bbh) \right) \\
&=\log \left( 1+ CNR_i(P_i,h_i) \right)\\
&= \log \left( 1+ \frac{\frac{1}{2}(OMI\cdot m_prP_ih_i)^2}{RIN \cdot (rP_ih_i)^2+2em_p^{2+F}rP_ih_i+\frac{4KT}{R_f}} \right).
\end{split}
\end{equation}
Note that system parameters are assumed to be same for all wavelength channels.
\section{Stochastic Dual Gradient Algorithm}\label{sec_sdg}
Solving the above optimization problem \eqref{eq_problem111} is challenging due to its non-concave capacity function, functional optimization complexity and the existence of constraints. We first address these challenges by establishing a null duality gap property of \eqref{eq_problem111} and subsequently presenting the Stochastic Dual Gradient (SDG) algorithm to solve. First, for the development of SDG algorithm, we assume that models given in Section \ref{sec_models} are accurate, i.e. the capacity function $C_i(\bbP(\bbh),\bbh)$ in \eqref{eq_problem111} can be computed as in \eqref{eq_capacity}.
With two constraints in \eqref{eq_problem111}, it is natural to think about working in the dual domain. Let $\mathcal{P} = [0, P_S]^m$ represent the space satisfying the eye safety concern, and introduce the dual variable $\lambda \ge 0$. The Lagrangian of the problem is given by
\begin{equation} \label{eq_lagran}
\begin{split}
\mathcal{L}(\bbP(\bbh),\lambda) & = \sum_{i=1}^m \omega_i \mathbb{E}_\bbh \left[ \log \left( 1+ CNR_i\left( P_i(\bbh), h_i \right) \right) \right] \\
&+ \lambda \left( P_T - \mathbb{E}_\textbf{h} \left[ \sum_{i=1}^m P_i(\bbh) \right] \right).
\end{split}
\end{equation}
The dual function is then defined as
\begin{equation}
\begin{split}
\mathcal{D}(\lambda) & = \max_{\bbP(\bbh)\in \mathcal{P}} \mathcal{L}(\bbP(\bbh),\lambda).
\end{split}
\end{equation}
Its corresponding dual problem is to find $\lambda^*$ that minimizes the dual function
\begin{equation} \label{eq_dualprob}
\begin{split}
\mathbb{D} = \min_{\lambda \ge 0} \mathcal{D}(\lambda)=\min_{\lambda \ge 0} \max_{\bbP(\bbh)\in \mathcal{P}} \mathcal{L}(\bbP(\bbh),\lambda).
\end{split}
\end{equation}
However, the objective function is non-concave and complicated due to the term $CNR_i(P_i(\bbh), h_i)$, which leads it to be a non-convex optimization problem. Solving it in the dual domain then seems to be impossible in principle. Nevertheless note that the key reason here to make the dual method impractical is not the non-convex property but the existence of duality gap indeed, which indicates the loss of optimality if using the dual method. In other words, as long as we can show that this problem does have null duality gap, it is then feasible to be solved in the dual domain.
Observe that the non-concave objective function is actually inside the expectation expression. We then give the following Theorem 1 according to \cite{Ribeiro2012} to show its null duality gap:
\begin{theorem}
Assume $\mathbb{P}$ and $\mathbb{D}$ donate the optimal solution value of the primal problem \eqref{eq_problem111} and its corresponding dual problem \eqref{eq_dualprob}. If there exists a feasible point $\bbP_0$ satisfying all constraints with strict inequality, and the probability distribution of CSI $\bbh$ contains no point of positive probability, then the duality gap is null:
\begin{equation}
\begin{split}
\mathbb{P}=\mathbb{D}.
\end{split}
\end{equation}
\end{theorem}
The problem in our case satisfies all conditions of Theorem 1 such that the duality gap is null even if it is a non-convex optimization problem. Then we can directly develop the dual methodology to solve \eqref{eq_problem111} by solving \eqref{eq_dualprob} without any relaxation.
The SDG algorithm is put forward based on the above analysis, which iteratively searches for the optimal dual variable $\lambda^*$ from initial $\lambda^0$ and use $\lambda^*$ to compute the optimal power allocation $\bbP^*(\bbh)$. Specifically, at each iteration $k$, SDG consists of two steps:
(1) \emph{Primal variable update.} For the given $\lambda^{k}$ from iteration $k-1$ and CSI $\bbh$, we update the primal variable by maximizing the Lagrangian:
\begin{equation} \label{eq_priup}
\begin{split}
\bbP^{k+1}(\bbh) &= \argmax_{\bbP(\bbh) \in \mathcal{P}} \mathcal{L}\left(\bbP(\bbh),\lambda^{k}\right)\\
&=\argmax_{\bbP(\bbh) \in \mathcal{P}} \sum_{i=1}^m \omega_i \mathbb{E}_\bbh \left[ \log \left( 1+ CNR_i\left( P_i(\bbh), h_i \right) \right) \right] \\
&\quad \quad \quad+ \lambda^{k} \left( P_T - \mathbb{E}_\textbf{h} \left[ \sum_{i=1}^m P_i(\bbh) \right] \right).
\end{split}
\end{equation}
Furthermore, both the objective function and constraints separate the use of $P_1(\bbh),...,P_m(\bbh)$ and $h_1,...,h_m$, with no coupling between them. \eqref{eq_priup} can be simplified to
\begin{equation}\label{eq_primalup}
\begin{split}
&P_{i}^{k+1}(\bbh)\\
&= \argmax_{P_i(\bbh) \in [0,P_S]} \omega_i \log \left( 1+ CNR_i\left( P_i(\bbh), h_i \right) \right)- \lambda^{k} P_i(\bbh).
\end{split}
\end{equation}
(2) \emph{Dual variable update.} With $\bbP^{k+1}(\bbh)$ gotten from step (1), we then perform a dual descent method to get $\lambda^{k+1}$:
\begin{equation}
\begin{split}
\lambda^{k+1} &= \left[ \lambda^{k} - \eta^k \nabla_{\lambda} \mathcal{L}(\bm{\theta}^{k+1},\lambda^k) \right]_+ \\
&=\left[ \lambda^{k}-\eta^k \left( P_T - \mathbb{E}_\textbf{h} \left[ \sum_{i=1}^m P_i^{k+1}(\bbh) \right] \right) \right]_+,
\end{split}
\end{equation}
where $\eta^k$ is the stepsize of $\lambda$ at iteration $k$, and $[\cdot]_+$ is due to the non-negativity of $\lambda$. The expectation $\mathbb{E}_{\bbh}[\cdot]$ is computed by the stochastic method with $S$ samples of $\bbh$.
By repeating the above two steps recursively, as $k$ increases, $\lambda^k$ converges to the optimal value $\lambda^*$, and the optimal allocated power of $i$-th wavelength channel $P^*_{i}(\bbh)$ is given by
\begin{equation}
\begin{split}
&P^*_{i}(\bbh)\\
&= \argmax_{P_i(\bbh) \in [0,P_S]} \omega_i \log \left( 1+ CNR_i\left( P_i(\bbh), h_i \right) \right)- \lambda^* P_i(\bbh).
\end{split}
\end{equation}
With the knowledge of accurate system model \eqref{eq_capacity} and SDG algorithm, we can solve the problem \eqref{eq_problem111} perfectly. However, in practice, there exists several problems to discuss:
1. SDG algorithm heavily depends on the exact system model, which means that we need the accurate knowledge of model \eqref{eq_capacity} to perform this algorithm. However, due to the complexity of RoFSO systems, such models may not be accurate in practice.
2. CSI $\bbh$ needs to be estimated at the receiver and feedback to the transmitter. However, the feedback estimated $\widehat{\bbh}$ used in SDG has errors with real $\bbh$ used in the objective capacity function $C_i(\bbP,\bbh)$, which degrades the performance of SDG.
3. In step (1) of SDG, there is not a closed-form solution to the maximization problem to get optimal $P_{i}^{k+1}(\bbh)$, and thus requires time to numerically solve it for each iteration.
These three problems inspires the use of model-free and low-complexity learning algorithms to solve the power allocation problem.
\section{Primal-Dual Deep Learning Algorithm}\label{sec_pddl}
To handle the above limitations of SDG algorithm, we develop the model-free Primal-Dual Deep Learning (PDDL) algorithm, which does not directly use system models but only observed capacity and CSI values. Note that our optimization problem \eqref{eq_problem111} shares the same structure with statistical learning problem. This inspires us to introduce a parametrization $\bm{\theta}\in \mathbb{R}^q$ to represent the power allocation policy $\bbP(\bbh)$ by
\begin{equation} \label{eq_dnn}
\begin{split}
\bbP(\bbh) = \bbPhi(\bbh, \bm{\theta}).
\end{split}
\end{equation}
Substitute \eqref{eq_dnn} into the problem \eqref{eq_problem111}, our purpose then becomes to learn an optimal function $\bbPhi^*(\bbh, \bm{\theta}^*)$ with optimal parametrization $\bm{\theta}^*$, which outputs allocated powers $\bbP^*$ that maximize the objective function.
In terms of the parametrization, a good choice of $\bbPhi(\bbh, \bm{\theta})$ should provide an accurate approximation for almost any function by changing its parameters $\bm{\theta}$, which can greatly improve the learning performance. Deep Neural Networks (DNNs), widely used in modern machine learning problems, are known to exhibit such strong function approximation ability almost perfectly \cite{hornik1991approximation}. Thus, DNN is a good candidate to be used here. We briefly introduce the architecture of DNN. Assume there are $L$ layers in DNN with $n_1,...n_L$ donating the number of layer units respectively. Each layer is comprised of two parts: linear transform matrix $\bbPi_l \in \mathbb{R}^{n_{l}\times n_{l-1}}$ and non-linear operator $\sigma_l$. The output of $l$-th layer $\bbx_{l} \in \mathbb{R}^{n_{l}}$ can then be obtained by its input $\bbx_{l-1} \in \mathbb{R}^{n_{l-1}}$:
\begin{equation}
\begin{split}
\bbx_{l} = \sigma_l\left( \bbPi_l \bbx_{l-1} \right).
\end{split}
\end{equation}
Note that the input of DNN $\bbx_0$ is the CSI $\bbh$, and the parametrization $\bm{\theta}$ is the matrices $\{ \bbPi_l \}_{i=1,...,L}$. As for the non-linear operator $\sigma$, various functions can be used, such as ReLu or sigmoid. Besides, note that $\bbtheta$ should belong to the set $\Theta = \{\bbtheta \vert \bbPhi(\bbh, \bbtheta)\in \mathcal{P} \}$ to satisfy the eye safety concern.
Similar as \eqref{eq_lagran}, the Lagrangian here can be expressed by
\begin{equation}
\begin{split}
\mathcal{L}(\bm{\theta},\lambda) & = \sum_{i=1}^m \omega_i \mathbb{E}_\bbh \left[ C_i(\bbPhi(\bbh, \bm{\theta}),\bbh) \right] \\
&+ \lambda \left( P_T - \mathbb{E}_\textbf{h} \left[ \sum_{i=1}^m \Phi_i(\bbh, \bm{\theta}) \right] \right).
\end{split}
\end{equation}
And its corresponding dual problem becomes
\begin{equation} \label{eq_dual1}
\begin{split}
\mathbb{D}_{\bm{\theta}} = \min_{\lambda \ge 0} \mathcal{D_{\bm{\theta}}}(\lambda)=\min_{\lambda \ge 0} \max_{\bm{\theta}\in \Theta} \mathcal{L}(\bm{\theta},\lambda).
\end{split}
\end{equation}
For the above min-max problem with sufficient dense DNN parametrization $\bbtheta$, the duality gap between $\mathbb{P}$ and $\mathbb{D}_{\bbtheta}$ is proportional to the function approximation ability of DNN \cite{eisen2019learning}. Therefore, due to the strong ability of DNN, the duality gap is nearly null. We then develop the PDDL learning algorithm based on \eqref{eq_dual1}, which updates primal variable $\bm{\theta}$ and dual variable $\lambda$ simultaneously at every iteration using first order gradients. The ultimate purpose is to search for a local stationary point $(\bbtheta^*, \lambda^*)$ which satisfies KKT conditions. Specifically, at each iteration $k$, we follow two steps:
(1) \emph{Primal variable update.} For a given $\lambda^{k}$ from iteration $k-1$ and CSI $\bbh$, we update the primal variable $\bm{\theta}$ by
\begin{equation} \label{eq_priup1}
\begin{split}
\bm{\theta}^{k+1} &= \bm{\theta}^{k} + \delta^k \nabla_{\bm{\theta}} \mathcal{L}(\bm{\theta}^{k},\lambda^{k})\\
& = \bm{\theta}^{k} + \delta^k \nabla_{\bm{\theta}} \mathbb{E}_\bbh \left[ \sum_{i=1}^m \omega_i C_i( \bbPhi(\bbh, \bm{\theta}^{k}),\bbh) \right. \\
& \left. + \lambda^{k} \left( P_T - \sum_{i=1}^m \Phi_i(\bbh, \bm{\theta}^{k}) \right) \right],
\end{split}
\end{equation}
where $\delta_k$ is the stepsize of $\bm{\theta}$ at iteration $k$, and the last equation is because of the linearity of the expectation.
(2) \emph{Dual variable update.} Once we get $\bm{\theta}^{k+1}$, the dual variable $\lambda$ is updated by a similar way
\begin{equation} \label{eq_dualup1}
\begin{split}
\lambda^{k+1} &=\left[ \lambda^{k} - \eta^k \left( P_T -\mathbb{E}_\bbh \left[ \sum_{i=1}^m \Phi_i(\bbh, \bm{\theta}^{k+1}) \right] \right) \right]_+.
\end{split}
\end{equation}
Observed from \eqref{eq_priup1}, the update of primal variable requires not only computing the gradient of capacity function $C_i( \bbPhi(\bbh, \bm{\theta}^k),\bbh)$, but also taking expectation $\mathbb{E}_\bbh[\cdot]$ of this gradient w.r.t. the distribution of $\bbh$. Either of them may be hard to know in practice, which makes the above algorithm useless. However, so-called policy gradient method used in reinforcement learning provides a good solution for these problems. It can be used to calculate the gradient for functions with the form of $\mathbb{E}_\bbh [f(\bbPhi(\bbh,\bm{\theta}),\bbh)]$, where $f$ is an unknown function. Actually, it calculates a stochastic and model-free approximation for $\nabla_{\bm{\theta}} \mathbb{E}_\bbh [f(\bbPhi(\bbh,\bm{\theta}),\bbh)]$ \cite{Sutton2000}.
In policy gradient method, the power allocation policy $\bbPhi(\bbh, \bm{\theta})$ is considered to be drawn from a distribution with a delta density function $\pi_{\bbh,\bm{\theta}}(\bbP)=\delta(\bbP - \bbPhi(\bbh, \bm{\theta}))$, and then we can rewrite
\begin{equation} \label{eq_grad}
\begin{split}
\nabla_{\bm{\theta}} \mathbb{E}_\bbh [f(\bbPhi(\bbh,\bm{\theta}),\bbh)] = \mathbb{E}_{\bbh,\bbP}[f(\bbP,\bbh) \nabla_{\bm{\theta}} \log \pi_{\bbh,\bm{\theta}}(\bbP)],
\end{split}
\end{equation}
in which $\bbP$ is a random realization drawn from the distribution $\pi_{\bbh,\bm{\theta}}(\bbP)$. However, calculating $\nabla_{\bm{\theta}} \log \pi_{\bbh,\bm{\theta}}(\bbP)$ of a delta density function still requires the knowledge of $f$. To handle this problem, the delta function can be approximated by Gaussian distribution centered around $\bbPhi(\bbh,\bm{\theta})$. And its mean and variance are given by the output features of DNN. Then we can estimate $\nabla_{\bm{\theta}} \mathbb{E}_\bbh [f(\bbPhi(\bbh,\bm{\theta}),\bbh)]$ by using \eqref{eq_grad} without knowing $f$. In addition, we take $S$ samples and average them when computing $\mathbb{E}_{\bbh,\bbP}[\cdot]$ to reduce the stochastic error:
\begin{equation}
\begin{split} \label{eq_policy}
\widetilde{\nabla_{\bm{\theta}}} \mathbb{E}_\bbh [f(\bbPhi(\bbh,\bm{\theta}),\bbh)] = \frac{1}{S}\sum_{j = 1}^S f(\bbP_j,\bbh_j) \nabla_{\bm{\theta}} \log \pi_{\bbh_j,\bm{\theta}}(\bbP_j),
\end{split}
\end{equation}
where $\bbh_j$ is a sampled CSI and $\bbP_j=[P_{j,1},...,P_{j,m}]$ is a corresponding realization drawn from the distribution $\pi_{\bbh_j,\bm{\theta}}(\bbP)$. So with \eqref{eq_policy}, we can compute the gradient in step (1) by
\begin{equation} \label{eq_polgrad}
\begin{split}
&\widetilde{\nabla_{\bm{\theta}}} \mathcal{L}(\bm{\theta},\lambda)\\
&=\widetilde{\nabla_{\bm{\theta}}} \mathbb{E}_\bbh \left[ \sum_{i=1}^m \omega_i C_i( \bbPhi(\bbh, \bm{\theta}),\bbh) \right. \left. + \lambda \left( P_T - \sum_{i=1}^m \Phi_i(\bbh, \bm{\theta}) \right) \right] \\
&= \frac{1}{S}\sum_{j = 1}^S \left\{ \left[ \sum_{i=1}^m \omega_i C_i( \bbP_j,\bbh_j) \right. \right.\\
& \left. \left. + \lambda \left( P_T - \sum_{i=1}^m P_{j,i} \right) \right] \nabla_{\bm{\theta}} \log \pi_{\bbh_j,\bm{\theta}}(\bbP_j) \right\}.
\end{split}
\end{equation}
Therefore, the primal variable update of PDDL algorithm can be completed by using \eqref{eq_polgrad} without any knowledge of system model $C_i( \bbP,\bbh)$ or CSI distribution but only their observations, which makes PDDL model-free. By replacing $\nabla_{\bm{\theta}} \mathcal{L}(\bm{\theta}^k,\lambda^k)$ with $\widetilde{ \nabla_{\bm{\theta}}} \mathcal{L}(\bm{\theta}^k,\lambda^k)$ in \eqref{eq_priup1}, PDDL is summarized in the following Algorithm 1.
{\linespread{0.9}
\begin{algorithm}[b] \begin{algorithmic}[1]
\STATE \textbf{Input:} Initial primal and dual variables $\bbtheta^0, \bblambda^0$
\FOR [main loop]{$k = 0,1,2,\hdots$}
\STATE Draw CSI samples $\{ \bbh \}$ of batch size $S$, and get their corresponding $\{ \bbP \}$ according to $\text{DNN}_{\bbtheta^k}$, $\pi_{\bbh_j,\bm{\theta}^k}(\bbP)$
\STATE Obtain observations of capacity values $C_i( \bbP,\bbh)$ at current samples of step 3
\STATE Compute the policy gradient $\widetilde{ \nabla_{\bbtheta}} \mathcal{L}(\bbtheta^k, \lambda^k)$ by \eqref{eq_polgrad}
\STATE Update the primal variable by \eqref{eq_priup1} \\
$\bm{\theta}^{k+1} = \bm{\theta}^{k} + \delta^k \widetilde{ \nabla_{\bm{\theta}}} \mathcal{L}(\bm{\theta}^k,\lambda^k) \nonumber $
\STATE Update the dual variable by \eqref{eq_dualup1}\\
$\lambda^{k+1} = \left[ \lambda^{k} - \eta^k \nabla_{\lambda} \mathcal{L}(\bm{\theta}^{k+1},\lambda^k) \right]_+ \nonumber $
\ENDFOR
\end{algorithmic}
\caption{Primal-Dual Deep Learning Algorithm}\label{alg:learning} \end{algorithm}}
\section{Simulation Results}\label{sec_numerical_results}
In this section, we perform numerical simulations to exhibit the performance of SDG and PDDL algorithms, and show their validity by comparing with the average equal power allocation policy. Wavelengths in 1520nm-1595nm are used in simulations with 5nm guard band between adjacent wavelengths.
Note that although our PDDL learning algorithm is model-free, we are doing numerical simulations not physical experiments. CSI samples $\{\bbh \}$ and their corresponding channel capacities $\{ C_i( \bbPhi(\bbh, \bm{\theta}),\bbh) \}$ cannot be observed here. We then still use the system model to compute them, but in reality we can directly get them from the real system in experiments without the need of any theoretical model. In addition, due to the separable use of $P_i$ and $h_i$ in both objective function and constraints, we construct $m$ independent DNNs for $m$ wavelength channels, and each DNN has three hidden layers with 20, 10 and 5 units respectively. ReLU function is utilized as the non-linear operator $\sigma$. Furthermore, the truncated Gaussian distribution is used as power policy distribution $\pi_{\bbh,\bbtheta}(\bbP)$ in policy gradient method, which constrains generated powers $\bbP$ inside $\mathcal{P}$ to satisfy the eye safety concern. The outputs of DNNs are used as means and standard deviations of $\pi_{\bbh,\bbtheta}(\bbP)$.
\begin{figure}[t]
\centering
\includegraphics[width=0.47\linewidth, height=1\textheight, keepaspectratio]{1-eps-converted-to.pdf}
\quad
\includegraphics[width=0.47\linewidth, height=1\textheight, keepaspectratio]{2-eps-converted-to.pdf}
\quad
\caption{The objective function value (left) and the constraint function value (right) over learning iterations of three policies for $8$ wavelength multiplexing.}\label{fig_simple_results}
\end{figure}
Fig. 1 shows the performance of three policies for $8$ wavelength multiplexing. Weights $\bbomega$ are drawn randomly from $0$ to $1$, and other default parameters are set as: $P_T = 1.2W$; $P_S = 0.3W$; $m_p=5$; $OMI=15\%$; $r=0.8$; $RIN=-140dB/Hz$; $T=300K$; transmitter aperture diameter $D_{tx}=0.05m$; receiver aperture diameter $D_{rx} = 0.1m$; $d=1000m$. Note that these parameter values are taken as an example to show our algorithms' performance, which can be adjusted based on specific systems and experiments. It can be seen from the left figure that the objective function values achieved by SDG and PDDL learning algorithms converge as iteration increases, and the performance of them outperforms the equal power policy. Similarly, the right figure plots the constraint function values with the increasing of iteration. The values eventually converge to $0$ for both of our algorithms, which indicates the feasibility of their optimal solutions. Besides, note that the model-based SDG that solves the problem exactly exhibits the best performance, which matches with our analysis. On the other hand, the objective value achieved by model-free PDDL converges closely to that of SDG, which validates the near perfect performance of PDDL. Moreover, PDDL can be used without any knowledge of system models, which is particularly useful when FSO system models are unknown, inaccurate, or too complicated to deal with, while SDG cannot handle such situations. Additionally, SDG requires to numerically solve a local maximization problem \eqref{eq_primalup} for every $\lambda^k$ and $\bbh$. Though it is not too hard since it is one-dimensional and with no constraint, SDG is still computationally more expensive than PDDL.
\begin{figure}[t]
\centering
\includegraphics[width=0.47\linewidth, height=1\textheight, keepaspectratio]{3-eps-converted-to.pdf}
\quad
\includegraphics[width=0.47\linewidth, height=1\textheight, keepaspectratio]{4-eps-converted-to.pdf}
\quad
\caption{The objective function value over learning iterations of three policies for $16$ wavelength multiplexing with $P_T = 2.4W, P_S = 0.3W$ (left) and $P_T = 4.0W, P_S = 0.5W$ (right).}\label{fig_simple_results}
\end{figure}
In the left figure of Fig. 2, we depict the performance of three policies for $16$ wavelength multiplexing with $P_T = 2.4W, P_S = 0.3W$. Results show both SDG and PDDL learning algorithms perform well for larger WDM systems, and the advantage of PDDL compared to the equal power policy becomes bigger. The right figure plots their performance for $16$ wavelength multiplexing with larger power settings $P_T = 4W, P_S = 0.5W$, which means there is more space for algorithms to manipulate powers. We can see that PDDL performs better in this case and converges roughly the same value as the exact solution found by SDG.
\section{Conclusion} \label{sec_con}
This paper investigates the challenging problem of optimal power allocation for WDM transmission in RoFSO systems. Two algorithms are developed to adaptively assign powers to different wavelength channels based on CSI. By showing the null duality gap, we first present the model-based Stochastic Dual Gradient algorithm, which is able to solve the problem exactly but heavily relies on the system model and CSI estimation accuracy. The model-free Primal-Dual Deep Learning algorithm is then developed to overcome the shortcomings of SDG. Specifically, it parameterizes the power allocation policy with Deep Neural Networks and learns optimal parameter values by updating primal and dual variables simultaneously. Policy gradient method is applied to compute updating gradients without using the knowledge of system or channel models. Numerical simulations are performed to show that both of our algorithms outperform the equal power policy. The model-free PDDL learning algorithm presented in this paper has wide applications for problems in FSO networks and communications, where FSO systems are sophisticated to model and turbulent channels are complicated to estimate.
\bibliographystyle{IEEEbib}
|
\section{Introduction} \label{sec:intro}
Until the advent of large space missions capable of spatially resolving rocky planets from their host stars, the only terrestrial exoplanets that will be spectroscopically accessible will be those that orbit nearby, mid-to-late M dwarfs \citep{NASESS(2018)}. Transiting examples of such planets are particularly advantageous, as they allow the unambiguous determination of masses, radii, mean densities and surface gravities, and permit their atmospheres to be probed with the technique of transmission spectroscopy. Yet, even with the large apertures of upcoming facilities, such studies will be photon starved: it may be possible to search for molecular oxygen in the atmospheres of terrestrial exoplanets with the upcoming cohort of ground-based giant, segmented mirror telescopes (GSMTs), but studies indicate that even marginal detections will be feasible only for stars within 15~parsecs and no larger than 0.3~R$_{\Sun}$ \citep{Snellen(2013),Rodler(2014),LopezMorales(2019)}. The eagerly awaited {\it James Webb Space Telescope} ({\it JWST}) may also be able to detect key molecules such as water, methane, and carbon dioxide in the atmospheres of terrestrial exoplanets, but again demands parent stars that are similarly nearby, and small \citep{Morley(2017)}.
Within 15~parsecs, there are 411 M dwarfs with masses between 0.3 and 0.1 \mdot, and perhaps an additional 60 systems between 0.1 M$_\Sun$ and the main-sequence cut-off \citep{Winters(2018AAS),Winters(2019a),Gagliuffi(2019)}. How many transiting terrestrial worlds might we expect within this sample of stars? \citet{Dressing(2015)} analyzed the data from the {\it Kepler} mission and found that, on average, M dwarfs host 2.5 planets smaller than 4~R$_\Earth$ with periods less than 200~days. Considering only planets with radii between $1.0-1.5$~R$_\Earth$ and periods less than 50~days, they found a mean number of planets per M dwarf of 0.56. Importantly, these stellar primaries were typically early M dwarfs, roughly twice as massive as the mid-to-late M dwarfs required to enable the atmospheric studies described above. Although efforts are underway to use K2 data to determine the rate of planet occurrence for the less massive M dwarfs \citep[e.g.,][]{Dressing(2019)}, it is currently an open question whether they host small planets with the same frequency as their more massive counterparts.
For stars less massive than 0.3~M$_\Sun$ and within 15~parsecs, four families of transiting, terrestrial planets are known: GJ~1132bc \citep{Berta-Thompson(2015),Bonfils(2018)}; LHS~1140bc \citep{Dittmann(2017a),Ment(2019)}; TRAPPIST-1bcdefgh \citep{Gillon(2016),Gillon(2017),Grimm(2018)}; and LHS~3844b \citep{Vanderspek(2019)}. Yet, the closest of these lies at 12~parsecs, for which {\it JWST} and the GSMTs may still be at pains to access. Thus, there is great interest within the community to identify even closer examples of such systems.
The $Transiting ~Exoplanet ~Survey ~Satellite$ \citep[\tess; ][]{Ricker(2015)} mission is now one year into its 2-year prime mission to scan most of the sky in search of the transiting planets that are most amenable to follow-up study.
We report here the detection with \tess ~data of the second closest known transiting exoplanet system, \system ~(TIC~98796344, TOI~455), and the nearest one for which a terrestrial planet transits a low-mass star. The planet is 6.9~parsecs away, and orbits one member of a stellar triplet. Multi-star systems present numerous challenges that sometimes deter planet hunters: astrometric perturbations from stellar companions at small separations can hinder the measurement of the trigonometric parallax of the system; the presence of bound companions can result in trends in the radial velocities of a star that can mask the signals of planets; and, light contamination from close stellar companions in the photometry of a host star can result in an underestimated planet radius \citep{Ciardi(2015),Furlan(2017),Hirsch(2017)}. Yet, these complications are also opportunities to measure the stellar orbits and investigate the potential formation scenarios for the planets that are found within; indeed all of these features are present in the system that is the subject of our study. We present here the discovery of the planet and a description of the host star system. We first provide a detailed portrait of the host star system in \S \ref{sec:host_system}. We then detail the observations in \S \ref{sec:data}. In \S \ref{sec:analysis}, we present our analysis of the data. Finally, in \S \ref{sec:disc} we discuss the implications of this planet and the opportunity it presents for characterization of its atmosphere.
\section{Description of the Host Stellar System}\label{sec:host_system}
The host system, LTT~1445ABC \citep{Luyten(1957),Luyten(1980a)}, is a nearby, hierarchical trio of mid-to-late M dwarfs. \citet{Rossiter(1955)} is the first observer to have noted relative astrometry for LTT~1445ABC using visual micrometry. In two observations made near the beginning of 1944 (specifically, Besselian years 1943.960 and 1944.027), the primary star was measured to have a separation from the B component of 3\farcs03 and 3\farcs51. In those same observations, the BC subsystem was measured to have a separation of approximately 1\farcs3. Since then, the separation of the primary relative to the subsystem has increased to a maximum value of 7\farcs706 in 2003 \citep{Dieterich(2012)}, and is now apparently decreasing, with the most recent value of 7\farcs10 obtained in 2017, according to data available in the Washington Double Star (WDS) Catalog\footnote{\url{https://www.usno.navy.mil/USNO/astrometry/optical-IR-prod/wds/WDS}} \citep{Mason(2009b)}. In contrast, for much of the time since 1944, the BC subsystem has been on a trajectory of decreasing separation; however, the most recent speckle observations appear to show that this trend has now
reversed, and the separation is growing larger. As shown in Figure \ref{fig:finders}, the three components are visible in an archival $Hubble ~Space ~Telescope$ ~($HST$) NICMOS image (left panel), but the B and C components are blended in our ground-based image from MEarth-South (right panel).
As reported by \citet{Henry(2018)} with over eighteen years of RECONS\footnote{REsearch Consortium On Nearby Stars; \url{www.recons.org}.} astrometry data, the position of the primary star exhibits an astrometric perturbation due to the presence of the BC pair. We describe a preliminary orbit for the BC pair below in \S \ref{subsec:bc_orbit}.
\begin{figure*}
\centering
\includegraphics[scale=.70,angle=0]{figure_1.pdf}
\caption{LTT~1445. Left: \hbox{$HST$} ~NICMOS image in the F110W filter taken in 2003; Right: MEarth image taken in 2019. North is up, East is left. The BC pair are blended in the ground-based MEarth image. We note that all three components fall on one 21\arcsec ~square \tess ~pixel. \label{fig:finders}}
\end{figure*}
Because of the nearly equal brightnesses of the A and blended BC components of the system, there has been confusion in the literature regarding their designations. Convention dictates that the primary component is the brightest in the $V-$band and is therefore the most massive of all stars in a multiple system\footnote{A rare exception is the case of a red dwarf--white dwarf pair where the two stars may have equal fluxes or equal masses, but not both at a given wavelength.}. The confusion appears to have originated with Luyten, who may not have been aware of Rossiter's work on this system. The Luyten Two Tenths (LTT) Catalogue \citep{Luyten(1957)} lists only one entry for the system, but the New Luyten Two Tenths (NLTT) Catalogue \citep{Luyten(1980a)} lists two entries with identical coordinates and proper motions. The brighter star (m$_R$ $=$ 11.1 mag, m$_{pg}$ $=$ 12.7 mag) is noted as LP~771-95; the fainter star (m$_R$ $=$ 11.8 mag, m$_{pg}$ $=$ 13.5 mag) is noted as LP~771-96. The entry for LP~771-96 includes the note, `Comp. to 95, 121\arcdeg, 4\farcs5', which implies that the secondary is the southeast component. The WDS entry for this result has been edited so that the position angle between the primary and secondary is 301\arcdeg, presumably to bring it into agreement with results from Rossiter who had already resolved the brighter component into a stellar pair. For the remainder of this paper, we refer to the stellar system as LTT~1445ABC, where A is the southeast component and BC is the blended northwest component.
While LTT~1445A\footnote{Other names: TIC~98796344, TOI~455, L~730-18, BD-17~588A, RST~2292A, WDS~J03019-1633A, 2MASS~J03015142-1635356, Gaia~DR2~5153091836072107136.} is brighter than LTT~1445BC in the $V$ filter, the opposite is true in the $K$ filter. Thus, if one were unaware of the binarity of the B component, mass estimates calculated using a mass-luminosity relation (MLR) will be discrepant, depending which filter relation is used: the A component will be more massive with a $V-$band MLR calculation, while the B component will be more massive with a $K-$band MLR calculation. However, we know that LTT~1445BC\footnote{Other names: TIC~98796342, BD-17~588B, RST~2292BC, WDS~J03019-1633B, 2MASS~J03015107-1635306, Gaia~DR2~5153091836072107008.} is a sub-arcsecond binary with both optical and infrared delta-magnitudes ($\Delta$mag) reported in the literature. Once the photometry of the BC components is deblended into their individual photometric magnitudes, their overluminosity in the infrared is resolved and the A component is the brightest and most massive star in the system in both the optical and infrared filters.
We calculated the \tess ~magnitudes, $T$, for all three components from ($I_{\rm KC} - K_{\rm s}$) colors\footnote{Henceforth, we omit the subscripts on these filters. The central wavelengths are 8075 \AA ~and 2.159 $\mu$m ~for the $I$ and $K$ filters, respectively.} using relations appropriate for M dwarfs developed by Guillermo Torres (private communication). The transformation is valid for M dwarfs with near-solar metallicity (-1.0 $\leq$ [Fe/H] $\leq$ 0.5) and has a residual scatter of 0.013 mag. We provide the relation here:
\vspace{0.3cm}
$T = I - 1.2457 + 1.54056*(I-K) - 0.49790*(I-K)^2 + 0.04539*(I-K)^3$
\vspace{0.3cm}
The transformation from ($I-K$) to $T$ for the primary star was straightforward. For the secondary pair, we first deblended the $I-$ and $K-$band magnitudes using $\Delta$mags reported in the literature: $\Delta$$I$ of 0.66$\pm$0.07 mag by \citet{Henry(2006)} and $\Delta$$K$ of 0.52$\pm$0.03 mag, which we calculated by averaging the values of $\Delta$F207M and $\Delta$F222M by \citet{Dieterich(2012)}. We also deblended the $V-$ and $R-$ band photometry. We list the stellar system parameters in Table \ref{tab:system_info}.
\begin{deluxetable*}{lccccc
\tablecaption{System Parameters for LTT~1445ABC \label{tab:system_info}}
\tablecolumns{6}
\tablenum{1}
\tablewidth{0pt}
\tablehead{
\colhead{Parameter} &
\colhead{A} &
\colhead{BC} &
\colhead{B - Deblended} &
\colhead{C - Deblended} &
\colhead{Reference}
}
\startdata
RA (2000.0) (hh:mm:ss) & 03:01:51.39 & 03:01:51.04 & \nodata & \nodata & 3,3 \\
Decl. (2000.0) (dd:mm:ss) & $-$16:35:36.1 & $-$16:35:31.1 & \nodata & \nodata & 3,3 \\
Proper Motion Mag. (mas yr$^{-1}$) & 456.5$\pm$0.2 & 479.4$\pm$0.3 & \nodata & \nodata & 2,3 \\
Proper Motion PA (deg) & 234.0$\pm$0.07 & 234.1$\pm$0.08 & \nodata & \nodata & 2,3 \\
Parallax (mas) & 145.55$\pm$0.08 & 142.57$\pm$2.03 & \nodata & \nodata & 2,3;3\\
$T$ (mag) & 8.88$\pm$0.02 & 8.80$\pm$0.02J & 9.27$\pm$0.07 & 9.92$\pm$0.07 &1,1,1,1 \\
$V_{\rm J}$ (mag) & 11.22$\pm$0.02 & 11.37$\pm$0.03J & 11.78$\pm$0.09 & 12.64$\pm$0.09 &3,3,1,1 \\
$R_{\rm KC}$ (mag) & 10.07$\pm$0.02 & 10.13$\pm$0.02J & 10.57$\pm$0.04 & 11.32$\pm$0.04 &3,3,1,1 \\
$I_{\rm KC}$ (mag) & 8.66$\pm$0.02 & 8.58$\pm$0.02J & 9.05$\pm$0.07 & 9.71$\pm$0.07 &3,3,1,1 \\
$J$ (mag) & 7.29$\pm$0.02 & 7.11$\pm$0.02J & \nodata & \nodata & 4,4 \\
$H$ (mag) & 6.77$\pm$0.04 & 6.56$\pm$0.02J & \nodata & \nodata & 4,4 \\
$K_{\rm S}$ (mag) & 6.50$\pm$0.02 & 6.29$\pm$0.02J & 6.81$\pm$0.04 & 7.33$\pm$±0.04 & 4,4 \\
Mass (\mdot) & 0.257$\pm$0.014 & \nodata & 0.215$\pm$0.014 & 0.161$\pm$0.014 & 1,1,1 \\
Radius (R$_{\odot}$) & 0.268$\pm$0.027 & \nodata & 0.236$\pm$0.027 & 0.197$\pm$0.027 & 1,1,1 \\
\enddata
\tablecomments{`J' indicates that the listed parameter is `joint' and includes both the B and C components.}
\tablerefs{
(1) this work;
(2) \citet{GaiaDR2(2018)};
(3) \citet{Henry(2018)};
(4) \citet{Skrutskie(2006)}.
}
\end{deluxetable*}
\subsection{Orbit Calculation of LTT~1445BC}
\label{subsec:bc_orbit}
We used the high-resolution astrometry of observations appearing in the Fourth Interferometric Catalog\footnote{\url{https://www.usno.navy.mil/USNO/astrometry/optical-IR-prod/wds/int4}} (FIC), plus the addition of previously unpublished data using the Differential Speckle Survey Instrument (DSSI) speckle camera \citep{Horch(2009)} at the WIYN telescope in 2012, to compute a preliminary visual orbit for the BC subsystem. Table \ref{tab:bc_obs} lists the observations used in the orbit calculation, including previously published DSSI observations of this pair \citep{Horch(2015b),Horch(2017)}. Only data from 2003 to the present were used; data before this did not use high-resolution techniques, and therefore the astrometry would generally be of lower precision. We used the method of \citet{MacKnight(2004)}, which takes as input an upper limit and a lower limit for each of the seven orbital elements and first computes a grid search to find the elements within the ranges selected that minimize the squared observed-minus-predicted differences in the secondary's position. After those elements are found, a second calculation is done to refine those orbital elements using the downhill simplex algorithm. Uncertainties in orbital elements are estimated by adding Gaussian random deviations of the expected astrometric uncertainty to all of the observed position angles and separations and recomputing the orbit many times. This yields a sample distribution for each orbital element; the uncertainty is calculated to be the standard deviation of the distribution in each case.
A reanalysis of the 2014 DSSI data indicates that the quadrant of the secondary in those images is ambiguous, a situation that can arise in speckle imaging when observing fainter targets. Using the position angle values shown in the FIC results in an orbit that is highly eccentric ($e=0.9$) and implies a mass sum for the BC subsystem of $0.63 \pm 0.28$ solar masses. On the other hand, if one reverses the quadrant of those observations by adding 180 degrees to the position angle, the derived parameters indicate that the pair has a somewhat eccentric, edge-on orbit with a period of roughly 36 years. While either orbit is possible, at this point we judge the the latter as more likely because the residuals to the orbit fit are significantly smaller. We list the orbital elements derived in that case in Table \ref{tab:bc_orbit}. Because the data span only 11 years (roughly one-third of the orbital period shown in Table 3), this results in large uncertainties in some of the orbital elements, particularly the semi-major axis and the period. High-quality orbital elements are not likely to be obtained for another decade, when observations will hopefully be available to clearly show the orbital progress. We show the visual orbit we have calculated in Figure \ref{fig:BC_orbit}. We show the ephemeris predictions and residuals for all observations used in the orbit calculation in Table \ref{tab:bc_eph}.
\begin{deluxetable*}{llllrrccc}
\tabletypesize{\scriptsize}
\tablenum{2}
\tablecaption{Astrometry and Photometry for Observations Used in the Orbit Calculation for LTT~1445BC \label{tab:bc_obs}}
\tablehead{
\colhead{Date} &
\colhead{$\theta$} & \colhead{$\rho$} &
\colhead{$\Delta m$} &
\colhead{$\lambda$} &
\colhead{$\Delta \lambda$} &
\colhead{Tel. Dia.} &
\colhead{Technique} &
\colhead{Reference}\\
\colhead{} &
\colhead{($^{\circ}$)} &
\colhead{(${\prime \prime}$)} &
\colhead{(mag)} &
\colhead{(nm)} &
\colhead{(nm)} &
\colhead{(m)} &&
}
\startdata
2003.4620 & 138.1 & 1.344 & 0.56 & 1797 & 68 & 2.4 & $HST$ NICMOS & 2 \\
2008.7675 & 137.7 & 0.7305 & 1.4 & 551 & 22 & 4.1 & Speckle & 6 \\
2010.594 & 138.41 & 0.41 & 0.52 & 2150 & 320 & 3.0 & AO & 5 \\
2012.0963 & 141.8 & 0.1812 & 2.16 & 692 & 40 & 3.5 & Speckle & 4 \\
2012.0963 & 140.0 & 0.1777 & 1.80 & 880 & 50 & 3.5 & Speckle & 4 \\
2012.7516 & 146.5 & 0.0710 & 1.05 & 692 & 40 & 3.5 & Speckle & 1 \\
2012.7516 & 141.4 & 0.0694 & 0.87 & 880 & 50 & 3.5 & Speckle & 1 \\
2014.7557 & 316.1\tablenotemark{*} & 0.2838 & 1.47 & 692 & 40 & 4.3 & Speckle & 3 \\
2014.7557 & 316.6\tablenotemark{*} & 0.2824 & 1.03 & 880 & 50 & 4.3 & Speckle & 3 \\
\enddata
\tablenotetext{*}{Quadrant ambiguous; the position angle here has been changed by 180 degrees relative to the original result.}
\tablerefs{
(1) This paper;
(2) \citet{Dieterich(2012)};
(3) \citet{Horch(2015b)};
(4) \citet{Horch(2017)};
(5) \citet{Rodriguez(2015)};
(6) \citet{Tokovinin(2010a)}.
}
\end{deluxetable*}
\begin{figure}
\includegraphics[scale=.55,angle=0]{rst2292_orbit4.pdf}
\caption{Preliminary astrometric orbit of LTT~1445C relative to LTT~1445B. Data points appearing in the Fourth Interferometric Catalog are shown with open circles. The three pairs of observations obtained with DSSI are shown as filled circles, with red indicating the 692nm filter and black indicating the 880nm filter. The cross indicates the position of LTT~1445B. North is up; East is left. \label{fig:BC_orbit}}
\end{figure}
Despite the very tentative nature of this orbit, the inclination is already very well constrained and the speckle data points have residuals that are fairly consistent with the known measurement precision of the technique; typical measurement uncertainties would be in the range of 2-3 mas in separation, and 0.5-1 degree in position angle at the telescope used, for example. Using the {\it Gaia} DR2 parallax reported for the primary star and the orbital results for period and semi-major axis, we derive a total mass for the BC subsystem of $0.39\pm0.09$ \mdot. The uncertainty in the mass is calculated from the independent trials of the orbit described above. Each trial gives a period and semi-major axis, and each is used, together with a randomly-chosen parallax deviate from a Gaussian distribution of the same mean of the Gaia parallax and standard deviation of the Gaia parallax uncertainty. The standard deviation of these mass values is our estimate of the mass uncertainty. This result is consistent with the sum of the masses estimated from the mass-$M_K$ relation \citep{Benedict(2016)} using the deblended $K-$band photometry (0.221$\pm$0.014 $+$ 0.165$\pm$0.014 \mdot $=$ 0.386$\pm$0.020 \mdot). Using the total mass of the three stellar components and an average angular separation of 5\arcsec ~(corresponding to 34 AU), we estimate the period of the A-BC orbit to be roughly 250 years. No further analysis of this orbit has been done.
Photometric information on the BC pair is difficult to interpret at this stage. \citet{Henry(2006)} noted a decrease in brightness of 0.3 magnitudes in a blended image taken in 1999. This was one of a series of observations taken as a part of the RECONS parallax effort, and those authors suggested that they had viewed a potential eclipsing event. The \citet{Tokovinin(2010a)} measure from 2008 notes the magnitude difference obtained is of reduced quality. In the recent sequence of speckle measures with the DSSI speckle instrument, including the measures presented here from October 2012, there are three measures of the magnitude difference in both 692nm and 880nm filters, but they have large scatter. Computing the average and standard error from these measurements, we obtain $\Delta692nm = 1.56 \pm 0.32$ and $\Delta880nm = 1.23 \pm 0.29$. These values carry larger uncertainty than expected; more observations are warranted.
\begin{deluxetable}{lc}
\tabletypesize{\small}
\tablecaption{Preliminary Orbital Parameters for LTT~1445BC \label{tab:bc_orbit}}
\tablecolumns{2}
\tablenum{3}
\tablehead{
\colhead{Parameter} &
\colhead{Value}
}
\startdata
Orbital Period (years) & 36.2$\pm$5.3 \\
Semi-major axis (arcseconds) & 1.159$\pm$0.076 \\
Inclination (degrees) & 89.64$\pm$0.13 \\
$\Omega$ (degrees) & 137.63$\pm$0.19 \\
T$_0$ (Besselian year) & 2019.2$\pm$1.7 \\
Eccentricity & 0.50$\pm$0.11 \\
$\omega$ (degrees) & 209$\pm$13 \\
\enddata
\end{deluxetable}
\begin{deluxetable}{lrrrr}
\tablewidth{0pt}
\tabletypesize{\scriptsize}
\tablenum{4}
\tablecaption{Calculated Orbital Ephemerides and Residuals for LTT~1445BC\label{tab:bc_eph}}
\tablehead{
\colhead{Date} &
\colhead{$\theta_{\rm eph}$} & \colhead{$\rho_{\rm eph}$} &
\colhead{$\Delta \theta$} &
\colhead{$\Delta \rho$}\\
\colhead{} &
\colhead{(deg)} &
\colhead{(${\prime \prime}$)} &
\colhead{(deg)} &
\colhead{(${\prime \prime}$)}
}
\startdata
2003.4620 & 137.9 & 1.3452 & 0.2 & -0.0012\\
2008.7675 & 138.3 & 0.7166 & -0.6 & 0.0139 \\
2010.594 & 138.7 & 0.4320 & -0.3 & -0.0220\\
2012.0963 & 140.1 & 0.1786 & 1.7 & 0.0026\\
2012.0963 & 140.1 & 0.1786 & -0.1 & -0.0009\\
2012.7516 & 144.2 & 0.0648 & 2.3 & 0.0062\\
2012.7516 & 144.2 & 0.0648 & -2.8 & 0.0046\\
2014.7557 & 316.4 & 0.2821 & -0.3 & 0.0017\\
2014.7557 & 316.4 & 0.2821 & 0.2 & 0.0003\\
\enddata
\end{deluxetable}
\section{Observations} \label{sec:data}
\subsection{TESS}
LTT~1445A and BC were observed by \tess ~in Sector four from UT 2018 October 19 to November 15, in spacecraft orbits 15 and 16. The observations were acquired with CCD 4 on Camera 2. We included this system in our \tess ~Guest Investigator program (PI Winters; G011231) target list to gather short-cadence (two-minute) data of the volume-complete sample of mid-to-late M dwarfs within 15 parsecs. LTT~1445A and BC were also included in the \tess ~Input Catalog (TIC) and Candidate Target List (CTL) \citep{Stassun(2018)} via the Cool Dwarf Sample \citep{Muirhead(2018)}.
The two-minute cadence data were reduced with the NASA Ames Science Processing Operations Center (SPOC) pipeline \citep{Jenkins(2015),Jenkins(2016)} that was repurposed from the $Kepler$ reduction pipeline \citep{Jenkins(2010)}. A planetary candidate with radius 1.4$\pm$0.4 \rearth ~was detected based on four transits to have a period of 5.4 days and a transit depth of 2498$\pm$168 ppm with a signal-to-noise ratio of 15.4.
As noted in the data release notes\footnote{\url{ https://archive.stsci.edu/missions/tess/doc/tess_drn/tess_sector_04_drn05_v04.pdf}} for sector 4, this was the first sector to benefit from the improved Attitude Control System algorithm, which reduced the pointing jitter of the spacecraft by an order of magnitude over the pointing errors evident in data from sectors 1-3. Two anomalies were noted in sector 4. An incorrect guide star table was initially used; when the correct guide star table was uploaded, the spacecraft pointing shifted by 4\arcsec. All cameras show a maximum attitude residual of about 0.45 pixels that decreases to roughly 0.2 pixels once the guide star catalog was updated. In addition, communications between the spacecraft and instrument ceased for roughly 64 hours (between times 2458418.54 and 2458421.21), during which time no telemetry or data were collected.
The \tess ~light curve, shown in Figure \ref{fig:gp_lc}, shows various types of stellar variability, such as flares and rotational modulation due to spots. We estimate a rotation period of 1.4~days from the \tess ~light curve (described in more detail in \S \ref{subsubsec:exoplanet}), which we suspect originates from either the B or C component, based on the activity indicator measurements for A and BC from our TRES spectra (described below). The duty cycle of the flares is fairly low, and those that occur are rather weak, with an increase in brightness on the order of 4\% (roughly 40 mmag); however, one large flare was detected, as seen in the top panel of Figure \ref{fig:gp_lc}, which was also reported by \citet{Howard(2019)}. While we do not know from which star the flares originate, previous work indicates that rapidly rotating stars flare more frequently. Thus, we suspect that the flares, too, come from either or both of the B or C components.
\begin{figure*}
\centering
\includegraphics[scale=.70,angle=0]{Full_model_LC_paper_20190618.pdf}
\caption{The \tess ~PDCSAP light curve. The top panel shows evidence of flares and rotational modulation due to stellar spots, likely on either the B or C component. The solid (crimson) line indicates the fit to the modulation. The middle panel shows the residual data after the removal with Gaussian Processes (GP) of the stellar variability; the planetary transit model is overplotted (solid crimson line) and the number for each transit is indicated. The dotted line indicates the third transit, which was not included in our fit. The bottom panel illustrates the light curve with planetary transit and stellar variability removed. The description of the flare rejection and GP fitting are given in Section \S \ref{subsubsec:exoplanet}. \label{fig:gp_lc}}
\end{figure*}
\subsection{Photometric Follow-Up with MEarth-South}
\label{subsec:mearth}
The large \tess ~pixel size of 21\arcsec ~means that the two nearly equal-luminosity sources, A and the blended BC, are included in the \tess ~aperture. Therefore it was necessary to determine from which star the transit signal originated. We obtained follow-up observations with MEarth-South for this purpose. One transit ingress of \planet ~was observed using four telescopes of the MEarth-South array at Cerro Tololo Inter-American Observatory (CTIO), Chile, on UT 2019 February 16. Exposure times were 4 seconds, with a total of 1445 data points gathered over 3 hours at airmasses 1.2 to 3.0 starting immediately at twilight until the target set. Due to the combination of the short exposure time and high airmass, particularly at transit ingress and thereafter, these data show very high levels of noise due to atmospheric scintillation. Two of the telescopes used in this observation had shutters stuck in the open position, and as a result the images were smeared during readout, but this does not appear to affect the resulting differential photometry.
The FWHM of the stellar images ranged from approximately 2.0 to 3.5 pixels, with a plate scale of $0.84$ arcsec pixel$^{-1}$. These observations resolve A from BC, but the wings of the point spread functions are still mildly overlapping and required specialized reduction procedures. In order to mitigate the influence of aperture positioning errors, the global astrometric solutions for the images were used for aperture placement (e.g., as described by \citealt{Irwin(2007)}), rather than individually determining the location of each star from the individual images. Undersized extraction apertures with radii of 4.2 pixels were used for the A and BC components, where the aperture size was chosen to prevent overlap of the apertures. Due to the lack of useful comparison stars elsewhere on the images, BC was used as the comparison source to derive differential photometry of A. The resulting light curve is shown in Figure \ref{fig:mearth}. It is compatible with the transits detected by \tess, although with low signal-to-noise, and suggests the transits originate from A. While we acknowledge that this detection is marginal, we provide further confirmation that the planet orbits the primary component in the system in \S \ref{sec:analysis}.
\begin{figure}
\includegraphics[scale=.40,angle=0]{lcfit2_mearth.pdf}
\caption{The partial transit event of \planet ~(top panel) observed by MEarth-South, suggesting that the transit comes from \star. The data have been phase-folded using the final transit parameters given in Table \ref{tab:pparams}. The individual data points are indicated in gray; black points indicate the data binned by 2.5 minutes; the red line illustrates the fit to the transit event, computed by taking the mean of the MCMC posterior parameters from Table \ref{tab:pparams}. Error bars on the individual data points have been omitted for clarity, but are large due to the high airmass of the observation. Residuals to the fit are shown in the bottom panel. \label{fig:mearth}}
\end{figure}
\subsection{Reconnaissance Data}
We used high-resolution data we had previously acquired as part of our ongoing, nearby M dwarf binary surveys \citep{Winters(2019a)} to confirm that the host star has no additional stellar or brown dwarf companions at separations less than 50 AU. We also investigated its rotational broadening and measured the equivalent width of H$\alpha$. In addition, we used roughly two decades of RECONS data from the CTIO / SMARTS 0.9m telescope to explore the system's long-term variability.
\subsubsection{High-Resolution Speckle Imaging: DSSI}
As part of our all-sky speckle survey of 1000 nearby M dwarfs for stellar companions (Winters, in prep), we observed \star ~on UT 2016 January 18 using DSSI on the Gemini-North 8.1m telescope. One thousand 40 millisecond exposures were taken in two filters (centered at 692 nm and 880 nm) simultaneously. This group of frames was followed by a similar set of exposures for a bright, unresolved calibration star at close proximity on the sky to the science star. The data reduction and analysis were conducted as described by \citet{Horch(2017)}.
We show the contrast curves for the 692 nm (left panel) and 880 nm (right panel) filters in Figure \ref{fig:speckle_ccs}. As illustrated in the left panel, no companions to \star ~were detected with $\Delta$692nm less than 5.06 mag at separations 0\farcs2 - 1\farcs2 (corresponding to projected linear separations of 1.4 - 8.2 AU) from \star. Additionally, no companions were detected with $\Delta$880nm less than 7 mag at separations larger than 0\farcs6 (roughly 4.1 AU). An L2 V spectral type, what we consider to be the `end of stars' \citep{Dieterich(2014)}, has M$_I$ of roughly 16.0 mag. The M$_I$ for \star ~is 9.48 mag, placing constraints to M$_I$ of 16.48 mag on the presence of a companion at separations greater than 0\farcs6, beyond the end of the M dwarf sequence and into the brown dwarf regime. This is in agreement with results by \citet{Dieterich(2012)} who observed the system with \hbox{$HST$} ~(NICMOS) and did not detect a stellar or brown dwarf companion to the primary star.
\begin{figure*}
\minipage{0.50\textwidth}
\includegraphics[scale=.45,angle=0]{LP771-095a.pdf}
\endminipage\hfill
\minipage{0.50\textwidth}
\includegraphics[scale=.45,angle=0]{LP771-095b.pdf}
\endminipage\hfill
\caption{Contrast curves of \star ~from DSSI on the Gemini-North 8.1m in the 692 nm (left panel) and 880 nm (right panel) filters. Open squares represent the positions of local maxima in the reconstructed image and points represent local minima (where the absolute value of the minimum is used). The red line represents the 5-$\sigma$ line as a function of separation. The lack of points below the red line illustrates that the primary component of LTT~1445 has no companions with $\Delta$mag less than 4.55 mag at separations 0\farcs2 -- 1\farcs2 (corresponding to projected linear separations of 1.4 -- 8.2 AU) from \star. \label{fig:speckle_ccs}}
\end{figure*}
\subsubsection{High-Resolution Spectroscopy: TRES}
As part of our all-sky spectroscopic survey of mid-to-late M dwarfs within 15 pc, we acquired multiple high-resolution spectra of both the A and BC components with the Tillinghast Reflector Echelle Spectrograph (TRES) on the FLWO 1.5m Reflector. TRES is a high-throughput, cross-dispersed, fiber-fed, echelle spectrograph with a resolving power of roughly $R = 44\,000$ when using the medium fiber ($2\farcs3$ diameter) and a passband of 310-910~nm. The observations span UT 2017 February 03 to 2018 January 24, with four spectra of the primary and three of the BC pair acquired. We integrated for 120-600 and 120-180 seconds and achieved signal-to-noise ratios of 16-25 and 16-21 for A and BC, respectively, at 715 nm. We used the methods described in \citet{Winters(2018)} for our analysis.
The equivalent width of the H$\alpha$ line and the magnitude of the rotational broadening allows differentiation between the two resolved stellar components. The primary star exhibits H$\alpha$ in absorption (we measure equivalent widths of 0.14$\pm$0.01\AA, 0.23$\pm$0.02\AA, 0.19$\pm$0.01\AA, and 0.25$\pm$0.01\AA), while the BC pair shows H$\alpha$ in emission (we measure equivalent widths of \mbox{-1.16}$\pm$0.03\AA, \mbox{-1.42}$\pm$0.03\AA, and \mbox{-1.52}$\pm$0.03\AA). We see negligible rotational broadening ($v$sin$i$) for the primary star, which allows us to place an upper limit of 3.4 km \hbox{s$^{-1}$} ~(half the spectral resolution of TRES) on the rotational broadening of A; we measure a $v$sin$i$ of 4.4$\pm$3.4 km s$^{-1}$ for the blended secondary-tertiary pair. While we cannot rule out that the photometric modulation seen in the \tess ~light curve data comes from A, it is more likely to come from BC because we detect both H$\alpha$ in emission and rotational broadening in the spectra of those blended components. We use the use the relation $P_{rot} * v \sin i = 2 \pi R \sin i$, the 1.4-day rotation period, and the estimated radii for each component from Table \ref{tab:system_info}, to estimate rotational broadening of 9.7, 8.7, and 7.2 km \hbox{s$^{-1}$} ~for the primary, secondary, and tertiary components, respectively, assuming each star is viewed edge-on.
We did not detect a second pair of lines in the spectra of the primary component that would indicate the presence of an additional stellar companion to A. Using the parameters from the preliminary astrometric orbit in Table \ref{tab:bc_orbit}, we calculate an expected radial velocity semi-amplitude of B due to C of 3.3 km \hbox{s$^{-1}$}.
\subsubsection{CTIO / SMARTS 0.9m}
RECONS has been astrometrically and photometrically monitoring this system for over nineteen years at the CTIO / SMARTS 0.9m telescope. In fact, the first trigonometric parallaxes for A and BC were reported in \citet{Henry(2006)}. These data in the $V$ filter permit the investigation of the long-term variability of the system. Following the methods described in \citet{Hosey(2015)}, we note that, over the nineteen years of monitoring, the A component varies by 13.4 millimagnitudes (mmag) while the BC pair vary by 13.0 mmag. While we do not consider either of these results indicative of a clearly variable source, for which we require an overall variability of 20 mmag, the levels for both are roughly twice that of the least variable M dwarfs at 7 mmag. We conclude that there are spots present on both A and the BC pair, but they do not change in coverage by an amount that alters the emergent flux in $V$ by more than 2\%. These results are in agreement with the moderate, short-term variability we detect in the \tess ~light curve data.
\subsection{HARPS}
\label{subsec:harps}
To constrain the mass of the transiting object, we acquired five new spectra of the primary star with the High Accuracy Radial Velocity Planet Searcher (HARPS) spectrograph \citep{Mayor(2003)} on the La Silla 3.6m telescope before the target was no longer observable. This totalled fourteen spectra when combined with the nine existing spectra in the ESO HARPS archive. HARPS is a fiber-fed echelle spectrograph with resolving power $R=$ 115,000 and a wavelength range of 378-691 nm. Spectra of LTT~1445A were integrated over 900 sec, except for the first two spectra acquired on 2003 December 15 and 2004 November 29, where the exposure times were 572 sec and 772 sec, respectively. The signal-to-noise ratio ranges between 37.8 and 72.5 at 650 nm, with an average of 55. For these spectra, this translates into radial velocity uncertainties \citep[computed following][and the procedure described below]{Bouchy(2001)} ranging between 0.9 m \hbox{s$^{-1}$} ~and 1.7 m \hbox{s$^{-1}$}, with an average uncertainty of 1.2 m \hbox{s$^{-1}$}.
Radial velocities (RVs) were derived by a $\chi^2$-minimization, or so-called template matching. We briefly summarize here the implementation of the process, which is described in more detail in \citet{Astudillo-Defru(2015)}. We first used the RVs from the HARPS pipeline \citep{Lovis(2007)}, in combination with the barycentric corrections, to shift all spectra of \star ~to the solar system barycentric reference frame. We constructed an initial stellar template by computing the median of the shifted and stacked spectra. A telluric template was constructed by Doppler-shifting the observed spectra of \star ~to the laboratory rest frame, which aligns the telluric absorption features, and computing the median. An improved stellar template was then constructed with the known telluric lines removed; that is, the template is a true spectrum of the star itself with improved signal-to-noise ratio. Due to the low number of spectra and to prevent auto-correlation between the stellar template and the spectrum being analyzed, contrary to \citet{Astudillo-Defru(2015)}, we computed the stellar template for each epoch but discarded the spectrum under analysis. New RVs were then derived by minimizing the $\chi^2$ of the residuals between the observed spectra and the stellar template. We list the RVs in Table \ref{tab:harps_rvs}.
The RVs of A exhibit a long-term drift due to the presence of the BC stellar pair, as shown in the bottom panel of Figure \ref{fig:harps_rvs}. This is in agreement with the astrometric perturbation of A due to BC mentioned above in \S \ref{sec:host_system}. These precise radial-velocity data also allow us to rule out the presence of any other nearby stellar or brown dwarf object orbiting the primary star. We note that observations will continue when the system is again observable to provide a robust mass for the planet.
\begin{deluxetable}{lcc}
\tabletypesize{\small}
\tablecaption{HARPS Radial Velocities for LTT~1445A \label{tab:harps_rvs}}
\tablecolumns{3}
\tablenum{5}
\tablehead{
\colhead{BJD\tablenotemark{a}} &
\colhead{$v_{\rm rad}$\tablenotemark{b}} &
\colhead{$\sigma$} \\
\colhead{(days)} &
\colhead{(${\rm km\ s^{-1}}$)} &
\colhead{(${\rm km\ s^{-1}}$)}
}
\startdata
2452988.689447 & -5.38629 & 0.00174\\
2453338.681357 & -5.39562 & 0.00116\\
2454078.640473 & -5.40353 & 0.00105\\
2454080.657212 & -5.40513 & 0.00108\\
2454292.913009 & -5.41529 & 0.00091\\
2454316.868711 & -5.41280 & 0.00112\\
2455042.915163 & -5.42530 & 0.00088\\
2455997.500569 & -5.43569 & 0.00137\\
2456237.681670 & -5.43910 & 0.00130\\
2458546.500296 & -5.45774 & 0.00131\\
2458547.501338 & -5.46353 & 0.00116\\
2458548.500633 & -5.46085 & 0.00136\\
2458555.506869 & -5.45518 & 0.00129\\
2458556.505906 & -5.45354 & 0.00116\\
\enddata
\tablenotetext{a}{Barycentric Julian Date of mid-exposure, in the TDB time-system.}
\tablenotetext{b}{Barycentric radial velocity.}
\end{deluxetable}
\section{Analysis}
\label{sec:analysis}
\subsection{Additional Confirmation of \star ~as the Host Star} \label{subsec:centroid}
We explored a number of other methods to determine from which star the transit originates.
Because our system has high proper motion, we were able to investigate whether the host star had moved on top of a background star which could be the source of the transit signal. We compared the position of the primary star in digitally-scanned SuperCOSMOS \citep{Hambly(2001)} archival POSS-1 images (taken 1953.93) with the star's position at the time of the \tess ~observations (we chose November 1 -- 2018.83 -- the effective mid-point of sector 4). The system had moved 29\farcs6 since the POSS-1 image was taken. No background star was seen at the current position of \star ~in the POSS-1 image, which has a magnitude limit of roughly 19.5 mag at 620-670 nm.
The data validation (DV) report from the SPOC pipeline provides a number of tests to aid in analyzing planet candidates. Two of the tests in the DV report use a difference image to analyze the centroid shift that occurs during the transit event to determine from which star the transit signal originates. The difference image provides the flux value for each pixel by taking the difference between the mean out-of-transit flux value and the mean in-transit value. Therefore, the star that is the planet host will have residual flux in the difference image. A centroid is then determined for the difference image. One analysis compares the difference image centroid to the expected position based on the TIC coordinates, while a second analysis compares the centroid of the difference image to the out-of-transit centroid. For \system, the difference image centroid corresponds to LTT~1445A in both instances; however, the magnitudes of the two centroid shifts in the DV report disagree by more than a factor of three (an offset of 2\farcs969$\pm$0\farcs440 for the TIC coordinate centroid, compared to an offset of 10\farcs130$\pm$0\farcs414 for the out-of-transit centroid), so we investigated further. In contrast to the way this same analysis was conducted for \kepler ~objects of interest, the TIC coordinates are corrected for proper motion. Therefore, in crowded fields, the \tess ~centroid offset measurements with respect to the TIC coordinates are generally more reliable than the out-of-transit centroid. But as noted by \citet{Twicken(2010),Stumpe(2014),Twicken(2018)}, the out-of-transit centroid is subject to crowding and can lock on to a star that is not the target. That is, in fact, the case for \system: the out-of-transit centroid position corresponds to that of the BC pair instead of the A component, resulting in an overestimated centroid offset in the DV report. Adjusting the 10\farcs130 out-of-transit centroid offset by 7\farcs10, the most recently measured separation of A and BC, gives an offset of 3\farcs03, in agreement with the 2\farcs969 offset from the TIC coordinate centroid.
In addition to the centroid shift tests, the candidate transit signature passes all the other diagnostic tests intended to flag false positives. These tests include the odd/even transit depth test, the weak secondary test, the ghost diagnostic test (which often flags background eclipsing binaries or scattered light features), and the statistical bootstrap test (false alarm probability $<$3e-16).
The field is sparse. The TFOP-TESS entry for this system reports ten contaminating sources, but only three additional sources are listed in the DV report and are shown to lie within the target mask. One is the nearby, physically bound pair which we have discussed above (LTT~1445BC, TIC 98796342). The other two sources are faint, with reported $T$ magnitudes of 15.032 and 15.991 at angular separations of 104\farcs51 and 120\farcs16 for TIC IDs 98796341 and 98796339, respectively. Our MEarth observations produced light curves for these two stars which confirm that they are not the source of the transit.
Based on the above analyses, and in combination with our marginal ground-based MEarth detection, we are confident that the planet candidate is transiting the primary star in the system.
\subsection{Host Star Parameters}
\label{subsec:star_params}
We use the methods appropriate for M dwarfs previously used by our group \citep{Berta-Thompson(2015),Dittmann(2017a),Ment(2019)} to determine the stellar parameters of the planet candidate host, which we then used as priors for the light curve modeling described below. We estimate the mass of the host star using the mass-luminosity relation in the $K-$band by \citet{Benedict(2016)} to be 0.258$\pm$0.014 \mdot. The relation in $M_K$ has been found to be less sensitive to metallicity than the $M_V$ relation \citep{Henry(1993),Delfosse(2000),Benedict(2016)}. For comparison, the estimated mass from the $M_V$-band relation is 0.251$\pm$0.023 \mdot. We then use single-star mass-radius relations \citep{Boyajian(2012)} to find a stellar radius of 0.268$\pm$0.027 R$_{\odot}$. We calculate the bolometric correction in $K$ using the prescription in \citet[erratum]{Mann(2015)} to be 2.70$\pm$0.04 mag, resulting in a bolometric luminosity for \star ~of \mbox{0.0079$\pm$0.0003 L$_{\odot}$}. We calculate the correction in $V$ from \citet{Pecaut(2013)} to be -2.06$\pm$0.04 mag\footnote{We assume the uncertainty on the bolometric correction in $V$ is that of the ($V-K$) color.}, resulting in a bolometric luminosity of \mbox{0.0082$\pm$0.0004 L$_{\odot}$}. We adopt the mean of the two bolometric luminosities. From the Stefan-Boltzmann Law, we find an effective temperature $T_{\rm eff}$ of 3337$\pm$150 K. As a comparison, we also used the relations in \citet{Mann(2015)} to determine an effective temperature of 3332$\pm$77 K for \star, in agreement with the $T_{\rm eff}$ derived from the Stefan-Boltzmann Law. We adopt the [Fe/H] of -0.34$\pm$0.08 from \citet{Neves(2014)}, which is measured from HARPS data.
\subsection{Light Curve Modeling}
Because of the complex nature of the light curve, we used a combination of {\sc exoplanet} \citep{Foreman-Mackey(2017)} and {\sc EXOFASTv2} \citep{Eastman(2017)} for our light curve modeling. The purpose of using {\sc exoplanet} was to fit and remove the photometric modulation in the \tess ~light curve using Gaussian Processes (GP) while preserving the planetary transit signal. We then used {\sc EXOFASTv2}, which does not currently have GP capability, to simultaneously fit the de-trended transit data from {\sc exoplanet} and the HARPS RV data. {\sc exoplanet} also has the capability to fit RV data, but we did not get convergence when including the HARPS data in the model. We elected not to include the third transit in our analysis because the light curve baseline showed a strong slope at egress; we note that this transit was also omitted from the results in the SPOC DV report.
If there are other objects in the \tess ~aperture with TIC identifiers, the SPOC pipeline calculates a dilution (i.e., contamination) factor and performs a correction to the final light curve, as noted in the $Kepler$ manual \citep{Thompson(2016)}. The correction for this system, contained in the keyword `CROWDSAP' in the light curve file, is 0.485. However, the \tess ~magnitudes of 8.88 and 8.80 that we calculate for each component are slightly different from the magnitudes in the TIC and CTL (for TIC 98796344(A), $T =$ 8.64 mag; for TIC 98796342(BC), $T =$ 8.55 mag) because of our M-dwarf-specific relation; thus, our calculated dilution factor will be slightly different, as well. From our \tess ~magnitudes, we calculated the flux for each component, from which we determined the dilution ($f_A$ / ($f_A$ $+$ $f_B$ $+$ $f_C$)) to be 0.480$\pm$0.013.
\subsubsection{{\sc exoplanet}}
\label{subsubsec:exoplanet}
For the first part of the light curve analysis, we used the python package {\sc exoplanet} \citep{exoplanet:exoplanet}. {\sc Exoplanet} employs probabilistic methods to model exoplanet transit and radial velocity data sets. It has the additional capability to incorporate Gaussian Processes (GP) with {\sc celerite} \citep{Foreman-Mackey(2017)} and limb-darkened light curves with {\sc starry} \citep{Luger(2018)}. We used the SPOC-generated Pre-Search Data Conditioning Simple Aperture Photometry (PDCSAP) light curve \citep{Smith(2012),Stumpe(2014)}, corrected with our calculated dilution factor. Before fitting, we removed positive outliers (flares) deviating by more than 3.0-$\sigma$ above the median absolute deviation of the PDCSAP light curve. To model the planetary transit, we used a limb-darkened transit model and a Keplerian orbit. The stellar variability, as well as any other possible systematics, are modeled with a GP.
We parameterized the model by the radius of the star in solar units $R_{*}$, mass of the star in solar units $M_{*}$, time of transit $T_0$ in days, orbital period $P$ in days, transit impact parameter $b$, eccentricity $e$, and argument of periastron $\omega$. This is used as the input for the light curve modeler, {\sc starry}, which computes a limb-darkened light curve, with parameters for quadratic limb-darkening coefficients $u_1$ and $u_2$. In addition to the limb-darkening coefficients, we parameterized the {\sc starry} light curve by the model mentioned above, the radius of the planet $R_{p}$, the times for which the light curve is to be evaluated, the exposure time of each observation, which in our case is 120 seconds, and the mean of the stellar flux, $F_{*_{\mu}}$ in parts per thousand (ppt). We performed a Box-Least-Squares (BLS) periodogram analysis on the PDCSAP light curve in order to estimate $P$, $T_0$, and the transit depth $\delta$. These estimates were used to inform the priors for $P$ and $T_0$. We used the transit depth as a constraint on the Gaussian prior placed on the radius of the planet. The priors are summarized in Table \ref{tab:PLpriors}.
\begin{deluxetable}{lccc}
\tabletypesize{\small}
\tablecaption{\sc{exoplanet} Planetary Orbit Parameters \label{tab:PLpriors}}
\tablecolumns{4}
\tablenum{6}
\tablehead{
\colhead{Parameter} &
\colhead{Prior\tablenotemark{a,b}} &
\colhead{Value} &
\colhead{Bound}
}
\startdata
$R_{*}$ ($R_{\odot}$) & Gaussian & $\mu =$ 0.268 & $\sigma$ = 0.013 \\
$M_{*}$ ($M_{\odot}$) & Gaussian & $\mu =$ 0.258 & $\sigma$ = 0.014 \\
$F_{*_\mu}$ (ppt) & Gaussian & $\mu =$ 0.0 & $\sigma$ = 10.0 \\
$u_1$ & Flat & 0.0 & (0.0 - 1.0) \\
$u_2$ & Flat & 0.0 & (0.0 - 1.0) \\
$T_0$ (days) & Gaussian & $\mu =$ -0.448 & $\sigma$ = 0.5 \\
$\rm{log}~P$ (days) & Gaussian & $\mu =$ ~\rm{log}$~5.358$ & $\sigma$ = 0.5 \\
$b$ & Uniform & 0.5 & (0.0 - 1.0) \\
$e$ & Beta & $\alpha =$ 0.876, $\beta =$ 3.03 & (0.0 - 1.0) \\
$\omega$ & Uniform & 0.0 & ($-\pi$ - $+\pi$)\\
$\rm{log}~R_p$ & Gaussian & $\mu =$ 0.5($\rm{log}~\delta + \rm{log}~R_{*}$) & $\sigma$ = 1.0\\
\enddata
\tablenotetext{a}{Where $\mu$ and $\sigma$ parameterize the Gaussian distribution.}
\tablenotetext{b}{Where $\alpha$, and $\beta$ parameterize the Beta distribution.}
\end{deluxetable}
The GP kernel is the sum of two simple harmonic oscillators shown in Equations \ref{eq:sho1} and \ref{eq:sho2}:
\begin{equation}
\label{eq:sho1}
SHO_1(\omega_{GP}) = \sqrt{\frac{2}{\pi}} \frac{S_1\,\omega_1^4}
{(\omega_{GP}^2-{\omega_1}^2)^2 + {\omega_1}^2\,\omega_{GP}^2/Q_1^2}
\end{equation}
and
\begin{equation}
\label{eq:sho2}
SHO_2 (\omega_{GP}) = \sqrt{\frac{2}{\pi}} \frac{S_2\,\omega_2^4}
{(\omega_{GP}^2-{\omega_2}^2)^2 + {\omega_2}^2\,\omega_{GP}^2/Q_2^2}
\end{equation}
where,
\begin{equation}
Q_1 = 0.5 + \tau_3 + \tau_4
\end{equation}
\begin{equation}
\omega_1 = \frac{4\pi Q_1}{\tau_2 \sqrt{4Q_1^2 - 1}}
\end{equation}
\begin{equation}
S_1 = \frac{\tau_1}{\omega_1 Q_1}
\end{equation}
\begin{equation}
Q_2 = 0.5 + \tau_3
\end{equation}
\begin{equation}
\omega_2 = \frac{8\pi Q_2}{\tau_2 \sqrt{4Q_2^2 - 1}}
\end{equation}
\begin{equation}
S_2 = \frac{\tau_5~\tau_1}{\omega_2 Q_2}
\end{equation}
This is an appropriate kernel for data that are quasi-periodic in nature, such as the observed rotational modulation in the light curve of LTT~1445ABC. The hyper-parameters for this GP are the amplitude of variability $\tau_1$, the primary period of the variability $\tau_2$, the quality factor $\tau_3$, the difference between the quality factors of the first and second modes of the two oscillators $\tau_4$, the fractional amplitude of the secondary mode to the primary mode $\tau_5$, and a jitter term added to account for excess white noise $\tau_6$. We placed a uniform prior on $\tau_5$ and Gaussian priors on $\tau_1$, $\tau_2$, $\tau_3$, $\tau_4$, and $\tau_6$. The mean ($\mu$) and standard deviation ($\sigma$) value we set for the Gaussian prior on $\tau_1$, $\tau_2$, $\tau_3$, $\tau_4$, and $\tau_6$ are shown in Table \ref{tab:GPpriors}. For the Gaussian prior on $\tau_2$, the primary period of variability, we estimated the mean, equal to 1.4 days, using a Lomb-Scargle periodogram of the light curve with the transits masked out.
\begin{deluxetable}{lccc}
\tabletypesize{\small}
\tablecaption{Gaussian Process Hyper-parameters \label{tab:GPpriors}}
\tablecolumns{4}
\tablenum{7}
\tablehead{
\colhead{Hyper-parameters} &
\colhead{Prior} &
\colhead{Value} &
\colhead{Bound}
}
\startdata
$\rm{log}~\tau_1$ (ppt) & Gaussian & $\mu =${\rm{log}~var(Flux)} & $\sigma$ = 5.0 \\
$\rm{log}~\tau_2$ (days) & Gaussian & $\mu = ${\rm{log}(1.4)} & $\sigma$ = 1.0 \\
$\rm{log}~\tau_3$ & Gaussian & $\mu = ${~\rm{log}(-5.0)} & $\sigma$ = 1.0 \\
$\rm{log}~\tau_4$ & Gaussian & $\mu = ${~\rm{log}(-5.0)} & $\sigma$ = 2.0\\
$\tau_5$ & Uniform & 0.0 & (0.0 - 1.0) \\
$\rm{log}~\tau_6$ (ppt) & Gaussian & $\mu = ${~\rm{log}~var(Flux)} & $\sigma$ = 10.0 \\
\enddata
\end{deluxetable}
We implemented the GP-transit model using {\sc PyMC3} \citep{exoplanet:pymc3}. Before sampling the model, we compute an initial guess of the parameters using a built in optimizer from {\sc exoplanet}. The optimization finds the maximum a posteriori solution, which is used as starting values for the sampler. After an initial burn-in of 14,000 steps, 12,000 steps were drawn from the posterior. {\sc PyMC3} provides useful convergence diagnostics such as the Gelman-Rubin statistic and the number of effective samples. For each parameter in our joint GP-transit model, the Gelman-Rubin statistic was within 0.001 of 1.000 and the number of effective samples was above 4000. We show the results of our fit in Figure \ref{fig:gp_lc}.
\subsubsection{{\sc EXOFASTv2}}
We used the output light curve data from {\sc exoplanet} with the stellar variability removed as input for the software package {\sc exofastv2} \citep{Eastman(2013),Eastman(2017)}. {\sc exofastv2} is a suite of {\sc IDL} routines that simultaneously fits exoplanetary transit and radial velocity data using a differential Markov Chain Monte Carlo (MCMC) code.
Because we derived the stellar parameters as described in \S \ref{subsec:star_params}, we did not include a spectral energy distribution in the fit, and we disabled the default MIST stellar evolutionary models that use isochrones to constrain the stellar parameters. We placed Gaussian priors on the mass, radius, effective temperature, and metallicity that were equal to the uncertainties noted in \S \ref{subsec:star_params}. The quadratic limb darkening coefficients were constrained by the \tess ~data and penalized for straying from the values predicted by the \citet{Claret(2017)} limb darkening tables at a given log{\it g}, $T_{\rm eff}$, and [Fe/H], as is standard within EXOFASTv2. While the atmospheric models used to derive the limb darkening tables are questionable for low-mass stars such as \star, the impact is likely to be negligible due to the low precision of the \tess ~lightcurve.
As noted in \S \ref{subsec:harps}, the fifteen years of HARPS RVs exhibit a drift due to the presence of the stellar BC pair, so we included terms for the slope and quadratic curve of the RVs. Additionally, we modified the default that searches in logarithmic RV semi-amplitude space to a linear option because of the few RV measurements available. We allowed eccentricity to be a free parameter, but excluded values where eccentricity was greater than $1 - 3 R_*/a$. Tides would theoretically be expected to exclude such high eccentricities because the tidal circularization timescale is very short \citep{Adams(2006)}. In addition, the excluded eccentricities are at negative RV semi-amplitudes, which omits non-physical masses for the planet. In order to allow for the propagation of the uncertainty in the dilution which was corrected in the {\sc exoplanet} fit, we included a prior on the correction to the dilution that was a Gaussian centered on zero with $\sigma = 0.013$. We required the number of independent draws to be greater than 1000 and and determined that, with a Gelman-Rubin statistic of 1.0107 in the worst case, the chains were well-mixed.
We find a period of \pplanet ~days, radius of \rplanet ~\rearth, mass of \mplanet ~\mearth, and equilibrium temperature of \hbox{$433^{+28}_{-27}$} ~K for \planet. The equilibrium temperature assumes an albedo of zero with perfect redistribution. We show the individual transits, along with the phase-folded transit in Figure \ref{fig:transits_tess}. In Figure \ref{fig:harps_rvs}, we show the best-fit model to the RV data, which we acknowledge is marginal. The uncertainties illustrated are the RV uncertainties from Table \ref{tab:harps_rvs} added in quadrature with the fitted RV jitter. The uncertainty on the planet's mass is largely due to the sparse RV coverage, so we place a 3-$\sigma$ upper limit on the RV semi-amplitude K, planet mass, and planet density. The 3-$\sigma$ upper limit is where 99.73\% of all links of all chains, after discarding the burn-in, are smaller than the quoted value. Listed in Table \ref{tab:pparams} are the 68\% confidence values from the {\sc exofastv2} fit. In addition we list the 3-$\sigma$ upper limits for K and its derived parameters.
To confirm that the two light curve fitting packages produce the same result, we also ran {\sc exofastv2} on the GP-corrected transit data without including the radial velocity data. The transit-only results from {\sc exofastv2} and {\sc exoplanet} are consistent with each other, within the errors. We also compared results from a fit to the detrended, whitened (i.e., the data validation time series) \tess ~light curve to our {\sc exoplanet} $+$ {\sc exofastv2} fit results and found good agreement.
\begin{figure}
\includegraphics[scale=.80,angle=0]{TOI455_mcmc_transit1.pdf}
\includegraphics[scale=.80,angle=0]{TOI455_mcmc_transit2.pdf}
\caption{The transits of \planet ~from \tess ~data, fit simultaneously with the RV data using {\sc exofastv2}. The individual transits are shown in the top panel, where the numbers in parentheses correspond to the transit numbers in Figure \ref{fig:gp_lc}; the phase-folded transit is shown in the bottom panel. \label{fig:transits_tess}}
\end{figure}
\begin{figure}
\includegraphics[scale=.80,angle=0]{TOI455_mcmc_rv1.pdf}
\includegraphics[scale=.80,angle=0]{TOI455_mcmc_rv2.pdf}
\caption{The preliminary spectroscopic orbit of \planet ~from HARPS data, fit simultaneously with the \tess ~transit data using {\sc exofastv2}. The orbit and its residuals are shown in the top panel, while the velocity drift of the primary star due to the stellar BC components is evident in the bottom panel, along with the residuals. \label{fig:harps_rvs}}
\end{figure}
\section{Discussion}
\label{sec:disc}
To summarize, we have presented the discovery of \planet, which resides in a host system composed of three mid-to-late M dwarfs at 6.9 pc. The planet has a radius \rplanet ~\rearth, an orbital period of \pplanet ~days, and an equilibrium temperature of \hbox{$433^{+28}_{-27}$} ~K; we place a 3-$\sigma$ upper mass limit of \muplanet ~\mearth ~on the planet. We have also presented a detailed view of the host system, which includes a preliminary orbit for the bound BC stellar pair that shows it to be in an eccentric and edge-on configuration.
The planet is an S-type (satellite) planet, meaning that it orbits one component of a stellar multiple system \citep{Dvorak(1982)}. Most of the literature on planets in multiple star systems considers planets in binary systems with solar-type primary stars \citep{Wang(2014a),Wang(2014b),Winn(2015),Kraus(2016),Eggenberger(2010),Matson(2018)}. Partly, this is because attention has only recently shifted to M dwarfs as planet hosts. Because stellar multiplicity is known to be a decreasing function of primary mass \citep{Duchene(2013)}, M dwarfs have a smaller stellar multiplicity rate than more massive stars: \citet{Winters(2019a)} estimate it to be 26.8$\pm$1.4\%, in agreement with recent results of 26$\pm$3\% and 28.6$^{+2.7}_{-3.1}$\% from \citet{Duchene(2013),Ward-Duong(2015)}, respectively. As a result, there are fewer systems to discuss, even if every M dwarf system has a planet. Theoretical work has shown that planets in binary systems are in stable configurations if their separations from their host stars are less than one-third the distance to a gravitationally-bound companion \citep{Holman(1999)}. The 0.038 AU separation of \planet ~is significantly less than one-third the 21 AU separation between A and BC (the smallest separation measured between A and B by Rossiter in 1943). Therefore, the planet is likely in a dynamically stable orbit, even given the presence of two other stars in the system.
M dwarfs are thought to be more likely than more massive stars to host multiple planets in compact orbits of less than roughly ten days \citep[ respectively, for early and mid-to-late M dwarfs]{Ballard(2019),Muirhead(2015)}. LTT~1445A may very well host multiple planets, but with only one sector of data from \tess, we have not yet seen indications of other planets in the system.
The alignment of the three stellar components and the edge-on orbit of the BC pair and planet is suggestive of the co-planarity of the system. Including LTT~1445ABC, there are six `pure' M dwarf triples known within 10~pc.\footnote{We do not include GJ~799/803 or GJ~569ABC. We consider the primary of the GJ~799/803 system to be a late K-type star, based on the M dwarf color cuts described by \citet{Winters(2019a)}, and we consider one of the companions to GJ~569A to be a brown dwarf, again consistent with \citet{Winters(2019a)}. Further, we do not include M dwarf triples that are subsets of higher-order multiples.} Among these are the very tight systems LTT~12352 and GJ~866 that have separations much smaller than LTT~1445. The remaining three systems --- GJ~2005, GJ~1230, and GJ~1245 --- all have archival \hbox{$HST$} ~NICMOS images available. Remarkably, none appear to exhibit the possible co-planarity we observe in \system. The presence of a transiting planet in this system raises the possibility that the entire system is co-planar, which has intriguing implications for planet formation scenarios.
Young stars form in often densely clustered environments with separations between the stars on the order of hundreds of astronomical units. Dynamical evolution of the cluster could result in capture, leading to binaries with wide separations; however, these stellar components would likely not be co-planar. A more likely formation mechanism is the quasi-static fragmentation of the circumbinary disk as the protostellar core is collapsing, where conservation of angular momentum leads to the formation of an accretion disk \citep{Stahler(2005)}. Such a system would possess both circumbinary and circumstellar disks. At apastron, a bound stellar pair would disrupt the circumbinary disk, which would tend towards increased eccentricity in the orbit of the stellar pair. Truncation of any circumstellar disks, and the cessation of accretion onto the primary's disk could truncate both the stellar mass and the masses of any planets around the primary. Given the possible co-planarity of the LTT~1445 system, the low masses of the stellar components, the eccentricity of the orbit of the BC stellar pair, and the presence of the terrestrial planet, it is likely that the system formed from the early fragmentation of an individual protostellar core, and not from later dynamical evolution within the young stellar cluster in which the stars formed.
Previous work leads us to expect that \planet ~is rocky in composition. Planets with radii less than 1.4 R$_{\oplus}$ are typically terrestrial in nature \citep{Rogers(2015),Dressing(2015a)}. An Earth-like density for \planet ~in a circular orbit implies a planetary mass of 2.5 \mearth ~and an RV semi-amplitude of 2.3 m \hbox{s$^{-1}$}. We expect the mass of the planet to be below \muplanet ~\mearth, based on our HARPS data. Additional precise RV observations in the near future will provide a robust mass for the planet; this will allow refinement of the planet's surface gravity, which will serve as an input for atmospheric models.
The planet is not in the habitable zone (HZ) of its star. The conservative inner and outer HZ boundaries for a 1~M$_{\oplus}$ planet around a star with $T_{\rm eff}$ of 3335~K are 0.093 and 0.182 AU, respectively, as calculated from the relations in \citet{Kopparapu(2013),Kopparapu(2014)}. The planet receives 5.6 times the Sun's irradiance from its host star, as its orbital distance brings it closer to \star ~than the annulus of its HZ.
LTT~1445Ab is the nearest planet known to transit an M dwarf and is the most spectroscopically-accessible, terrestrial planet detected to-date. LTT~1445Ab is a prime target for atmospheric studies due to its large transit depth ($\delta$ $=$ 0.2\%) and bright host star ($V_J =$ 11.22, $I_{KC} =$ 8.66, $K_s =$ 6.50 mag). It is also the second nearest known transiting planetary system to Earth. Currently, the nearest transiting planetary system is HD~219134bc \citep{Motalebi(2015),Gillon(2017)}, at a distance of 6.5 pc ($\pi$ $=$ 153.08$\pm$0.09 mas; \citealt{GaiaDR2(2018)}); but while the host star is brighter than \star ~($K_s$ $=$ 3.25 mag), the planetary transit depths are nearly an order of magnitude smaller ($\delta$ $=$ 0.036\% for HD~219134b). In contrast, the multi-planet system TRAPPIST-1 at 12.6 pc exhibits planetary transit depths that are, on average, a factor of three larger than that of \planet, but the host star is faint ($V_J =$ 18.75, $I_{KC} =$ 14.10, $K_s =$ 10.30 mag).
There are many opportunities for follow-up studies of \planet. For instance, LTT~1445ABC is particularly favorable for ground-based observations to study the planet's atmosphere, as the blended BC pair may provide a valuable calibration source with the same spectral type as that of the primary star, although its utility as a comparison source may be limited if it is the source of the variation and flaring seen in \tess. \star ~is small enough and bright enough that we can collect enough photons in space for transmission spectroscopy. From \citet{Kempton(2018)}, we calculate a transmission spectroscopy metric (TSM) of 40 for \planet, where the TSM is the expected signal-to-noise ratio of transmission spectroscopy observations with {\it JWST}. This TSM is factors of four and three better than the TSMs of 9.15 and 13.7 for LHS~1140b and TRAPPIST-1f, respectively.
Based on the known occurrence rates of planets orbiting M dwarfs, it is unlikely that we will detect a small planet more favorable for atmospheric characterization than \planet. There are 55 mid-to-late M dwarf primaries closer than \star. The transit probability of \star ~is $1/30$. Assuming that all such stars have rocky planets, then we expect to find roughly one as amenable to follow-up study as \planet. However, \citet{Dressing(2015)} estimate the rate of occurrence of planets less than 1.5 \rearth ~to be 0.43 for orbital periods less than 10 days, in which case we expect that this is indeed the best one. A subtlety is that the \citet{Dressing(2015)} result pertains to early-type M dwarfs; the occurrence rates for mid-to-late M dwarfs may be higher, and thus we may find one or two more planets like \planet, if we are lucky.
\startlongtable
\begin{deluxetable*}{lcc}
\tablecaption{Median values and 68\% confidence interval for \planet \label{tab:pparams}}
\tablenum{8}
\tablehead{\colhead{~~~Parameter} &
\colhead{Units} &
\multicolumn{1}{c}{Values}}
\startdata
\smallskip\\\multicolumn{2}{l}{Stellar Parameters:}& A \smallskip\\
~~~~$M_*$\dotfill &Mass (\msun)\dotfill &$0.256\pm0.014$\\
~~~~$R_*$\dotfill &Radius (\rsun)\dotfill &$0.276^{+0.024}_{-0.019}$\\
~~~~$\rho_*$\dotfill &Density (g cm$^{-3}$)\dotfill &$17.3\pm3.9$\\
~~~~$\log{g}$\dotfill &Surface gravity (cgs)\dotfill &$4.967^{+0.061}_{-0.075}$\\
\smallskip\\\multicolumn{2}{l}{Planetary Parameters:}&b\smallskip\\
~~~~$T_0$\dotfill &Optimal conjunction Time (\bjdtdb)\dotfill &$2458423.42629\pm^{+0.00044}_{-0.00045}$\\
~~~~$P$\dotfill &Period (days)\dotfill &$5.35882^{+0.00030}_{-0.00031}$\\
~~~~$T_{14}$\dotfill &Total transit duration (days)\dotfill &$0.0574\pm0.0011$\\
~~~~$R_P/R_*$\dotfill &Radius of planet in stellar radii \dotfill &$0.0458^{+0.0012}_{-0.0011}$\\
~~~~$a/R_*$\dotfill &Semi-major axis in stellar radii \dotfill &$29.6^{+2.6}_{-2.5}$\\
~~~~$b$\dotfill &Transit Impact parameter \dotfill &$0.29^{+0.23}_{-0.20}$\\
~~~~$R_P$\dotfill &Radius (\rearth)\dotfill &$1.38^{+0.13}_{-0.12}$\\
~~~~$a$\dotfill &Semi-major axis (AU)\dotfill &$0.03807^{+0.00068}_{-0.00071}$\\
~~~~$i$\dotfill &Inclination (degrees)\dotfill &$89.40^{+0.41}_{-0.46}$\\
~~~~$e$\dotfill &Eccentricity \dotfill &$0.19^{+0.35}_{-0.14}$\\
~~~~$\omega_*$\dotfill &Argument of Periastron (degrees)\dotfill &$-139^{+120}_{-76}$\\
~~~~$T_{eq}$\dotfill &Equilibrium temperature (K)\dotfill &$433^{+28}_{-27}$\\
~~~~$K$\dotfill &RV semi-amplitude (m \hbox{s$^{-1}$})\dotfill &$2.1^{+1.6}_{-2.0}$\\
~~~~$K$\dotfill &RV semi-amplitude (m \hbox{s$^{-1}$})\dotfill &$<9.3$ [$3-\sigma$ upper limit]\\
~~~~$M_P$\dotfill &Mass (\mearth)\dotfill &$2.2^{+1.7}_{-2.1}$\\
~~~~$M_P$\dotfill &Mass (\mearth)\dotfill &$<8.4$ [$3-\sigma$ upper limit]\\
~~~~$\rho_{P}$\dotfill &Density (g cm$^{-3}$)\dotfill &$4.4^{+4.0}_{-4.2}$\\
~~~~$\rho_P$\dotfill &Density (g cm$^{-3}$)\dotfill &$<22$ [$3-\sigma$ upper limit]\\
~~~~$\log{g_P}$\dotfill &Surface gravity (cgs) \dotfill &$3.11^{+0.22}_{-0.34}$\\
~~~~$\log{g_P}$\dotfill &Surface gravity (cgs) \dotfill &$<3.7$ [$3-\sigma$ upper limit] \\
\smallskip\\\multicolumn{2}{l}{Wavelength Parameters:}&$TESS$\smallskip\\
~~~~$u_{1}$\dotfill &linear limb-darkening coeff \dotfill &$0.195\pm0.030$\\
~~~~$u_{2}$\dotfill &quadratic limb-darkening coeff \dotfill &$0.427\pm0.027$\\
~~~~$A_D$\dotfill &Dilution from neighboring stars$^{*}$ \dotfill &$0.000\pm0.013$\\
\smallskip\\\multicolumn{2}{l}{Telescope Parameters:}&HARPS\smallskip\\
~~~~$\gamma$\dotfill &Systemic RV (m \hbox{s$^{-1}$})\dotfill &$-5432.3\pm2.1$\\
~~~~$\dot{\gamma}$\dotfill &RV slope (m \hbox{s$^{-1}$} day$^{-1}$)\dotfill &$-0.01275\pm-0.00053$\\
~~~~$\ddot{\gamma}$\dotfill &RV quadratic term (m \hbox{s$^{-1}$} day$^{-2}$)\dotfill &$0.00000124\pm0.00000040$\\
~~~~$\sigma_J$\dotfill &RV Jitter (m \hbox{s$^{-1}$})\dotfill &$3.25^{+1.2}_{-0.80}$\\
\enddata
\tablecomments{$^{*}$fixed parameter}
\end{deluxetable*}
\vspace{5mm}
\begin{center}
\large
Acknowledgments
\end{center}
\normalsize
The authors thank the anonymous referee for their prompt response and for their comments and suggestions. We are indebted to Guillermo Torres for developing the M dwarf-specific relations from which we calculated the \tess ~magnitudes. We thank Mark Everett for his role in obtaining the DSSI speckle data. We are extremely grateful to Matthew Payne, Chelsea Huang, Joseph Rodriguez, Samuel Quinn, and Jacob Bean for illuminating conversations and suggestions that helped improve the analysis and interpretation of the results presented here.
The MEarth Team gratefully acknowledges funding from the David and Lucille Packard Fellowship for Science and Engineering (awarded to D.C.). This material is based upon work supported by the National Science Foundation under grants AST-0807690, AST-1109468, AST-1004488 (Alan T. Waterman Award), and AST-1616624. This work is made possible by a grant from the John Templeton Foundation. The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the John Templeton Foundation. Funding for the \tess ~mission is provided by NASA's Science Mission directorate. We acknowledge the use of public \tess ~Alert data from pipelines at the \tess ~Science Office and at the \tess ~Science Processing Operations Center. Resources supporting this work were provided by the NASA High-End Computing (HEC) Program through the NASA Advanced Supercomputing (NAS) Division at Ames Research Center for the production of the SPOC data products. HARPS observations were made with European Space Observatory (ESO) Telescopes at the La Silla Paranal Observatory under program IDs 072.C-0488, 183.C-0437, and 1102.C-0339. This research has made use of the Washington Double Star Catalog, maintained at the U.S. Naval Observatory. This work has made use of data from the European Space Agency (ESA) mission {\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia} Data Processing and Analysis Consortium (DPAC, \url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the {\it Gaia} Multilateral Agreement.
AAM and HD-L are supported by NSF Graduate Research Fellowship grants DGE1745303 and DGE1144152, respectively. NA-D acknowledges the support of FONDECYT project 3180063. EPH gratefully acknowledges funding from NSF grant AST-1517824. The RECONS team is indebted to long-term support from the NSF, most recently under grant AST-1715551. Work by JNW was partly supported by the Heising-Simons Foundation. Support for JKT and the acquistion of the speckle data was provided by NASA through Hubble Fellowship grant HST-HF2-51399.001 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5-26555. XD is supported by the French National Research Agency in the framework of the Investissements d’Avenir program (ANR-15-IDEX-02), through the funding of the "Origin of Life" project of the Universit\'e Grenoble-Alpes. CDD acknowledges support from the \tess ~Guest Investigator Program through grant 80NSSC18K1583.
\vspace{5mm}
\facilities{TESS}, {MEarth}, {Gemini:North (DSSI)}, {FLWO:1.5m (TRES)}, {ESO:3.6m (HARPS)}, {CTIO:0.9m}
\software{{\sc celerite} \citep{Foreman-Mackey(2017)}, {\sc exofastv2} \citep{Eastman(2013),Eastman(2017)}, {\sc exoplanet} \citep{exoplanet:exoplanet}, IDL, IRAF, {\sc PYMC3} \citep{exoplanet:pymc3}, {\sc python}, {\sc starry} \citep{exoplanet:luger18}}
This research made use of {\sc exoplanet} \citep{exoplanet:exoplanet} and its dependencies \citep{exoplanet:astropy13, exoplanet:astropy18, exoplanet:exoplanet, exoplanet:foremanmackey17, exoplanet:foremanmackey18, exoplanet:kipping13, exoplanet:luger18, exoplanet:pymc3, exoplanet:theano}.
\bibliographystyle{aasjournal}
|
\section{Introduction}
Nearest neighbor search (NNS) within semantic embeddings (a.k.a., vector similarity search) has become a common practice in ubiquitous eCommerce applications including neural ranking model based text search \cite{Brenner2018ete, magnani2019neural}, content-based image retrieval \cite{yang2017visual, mu2018towards}, collaborative filtering \cite{deshpande2004item}, large-scale product categorization \cite{hu2018best}, fraud detection \cite{raghava2017predicting}, etc. While vector similarity search is capable of substantially boosting search relevancy by understanding customers' intents more semantically, it presents a major challenge: how to conduct nearest neighbor search among millions or even billions of high-dimensional vectors in a real-time and cost-effective manner.
The fundamental trade-off between search latency and cost-effectiveness would naturally classify nearest neighbor search solutions into two broad categories.
\paragraph{NNS solutions implemented in main memory.}
This type of NNS solutions has been extensively studied and explored in the field of information retrieval (IR). As a result, the majority of those widely used ones (e.g., Spotify's Annoy \cite{Github:annoy}, Facebook's FAISS \cite{johnson2017billion} and Microsoft's SPTAG \cite{wang2012scalable, ChenW18}) in nowadays software market fall into this category.
\paragraph{NNS solutions implemented in secondary memory.}
In contrast, the second type of NNS solutions are delivered only recently by active efforts from both academia and industry \cite{lux2013visual,rygl2017semantic,ruuvzivckaflexible,amato2018large,mu2018towards, mu2019empowering} to empower full-text search engines (e.g., Elasticsearch and Solr) with the capability of finding nearest neighbors. By leveraging inverted-index-based information retrieval systems and cutting-edge engineering designs from these full-text search engines, such full-text search engine based solutions are capable of economically reducing RAM consumption \cite{amato2018large}, coherently supporting multi-modal search \cite{mu2018towards} and being extremely well-prepared for production deployment \cite{rygl2017semantic}. However, some of the critical performance questions have not been quantitatively answered in literature:
\begin{itemize}
\item how much RAM could these full-text search based solutions save?
\item how much search latency would these solutions sacrifice in order to reduce RAM consumption?
\end{itemize}
In this paper, we will shed light on the above questions through a case study on the task of nearest neighbor search in Hamming space (i.e., the space of binary codes). This task is an extremely important subclass of NNS, as learning and representing textual, visual and acoustic data
with compact and semantic binary vectors is a pretty mature
technology and common practice in nowadays IR systems. In particular, eBay recently builds its whole visual search system \cite{yang2017visual} upon finding nearest neighbors within binary embeddings generated through deep neural network models.
We choose one representative solution of each category--FAISS (Facebook AI Similarity Search) from Facebook's AI Research Lab \cite{johnson2017billion} and FENSHSES (Fast Exact Neighbor Search in Hamming Space on Elasticsearch) from the search and catalog teams at Walmart Labs \cite{mu2019empowering, mu2019empowering2}--to evaluate their performances in finding nearest neighbors within binary codes.
\section{FAISS \lowercase{vs.} FENSHSES}
We will compare performances of FAISS and FENSHSES from three key perspectives: time spent in data indexing, search latency and RAM consumption.
\paragraph{Data generation.} Our dataset $\mathcal B$ is generated using 2.8 million images selected from Walmart.com's home catalog through pHash \cite{klinger2010phash, christoph2010implementation}--one of the most effective perceptual hash schemes in generating fingerprints for multimedia files (e.g. images, audios and videos)--with number of bits $m\in \set{64, 256, 1024, 4096}$ respectively. Note that vector similarity search based on pHash has been widely used in a variety of visual tasks including forensic image recognition \cite{peter2012privacy}, duplicate image detection \cite{chaudhuri2018smart} and copyright protection \cite{mehta2019decentralised}, etc.
\paragraph{Settings.}
For FAISS, we use its binary flat index with five threads. For a fair comparison, we accordingly deploy FENSHSES by creating its Elasticsearch index with five shards and zero replica. The rest of configurations are left as their default and suggested values. Both FAISS and FENSHSES are set up and tested on the same Microsoft Azure virtual machine.
\paragraph{Speed in indexing.} During the indexing phase, FAISS indexes the data into main memory (i.e., RAM), while FENSHSES indexes the data into secondary memory (e.g., hard disk). As a consequence, FAISS is much faster than FENSHSES in terms of data indexing (see Table \ref{tab: speed_in_indexing}). But on the other hand, whenever the process is killed and needs a restart, FAISS has to go through this procedure again to re-index data into RAM, while FENSHSES could unaffectedly use its built index on hard disk without re-indexing.
\begin{table}[h!]
\begin{tabular}{|C{1.6cm}|C{1.6cm}|C{1.5cm}|C{1.5cm}|}
\hline
\textbf{\# of Bits} & \textbf{FAISS (sec.)} & \textbf{FENSHSES (sec.)}
\\
\hline
\hline
{64} & 18.5 & 75.5 \\
{256} & 37.7 & 140.2 \\
{1024} & 111.9 & 369.5 \\
{4096} & 397.3 & 1300.9 \\
\hline
\end{tabular}
\caption{{\bf Indexing time consumption.} {\normalfont FAISS is about four times faster than FENSHSES in creating the index for nearest neighbor search. }}\label{tab: speed_in_indexing}
\end{table}
\paragraph{Search latency.} We randomly select $10,000$ binary codes from $\mathcal B$ to act as query codes. For each query code $\bm q$, we instruct FAISS and FENSHSES to find all $r$-neighbors of $\bm q$ in $\mathcal B$, namely \begin{flalign}
B_H(\bm q, r) := \set{\bm b \in \mathcal B \;\vert\; d_H(\bm b, \bm q) \le r},
\end{flalign}
where $d_H(\bm b, \bm q):=\sum_{i=1}^m \mathbbm 1_{\set{b_i \neq q_i}}$ denotes the Hamming distance between binary code $\bm b$ and $\bm q$, and the Hamming radius $r\ge 0$. As shown in Table 2, FENSHSES is quite competitive for small radius $r$. This is because FENSHSES fully leverages Elasticsearch's inverted index to first conduct a sub-code filtering to only consider a subset of $\mathcal B$ for Hamming distance computation, which is most effective for small $r$. In contrast, FAISS scans every binary code in $\mathcal B$, so its search latency is almost invariant with respect to $r$. For applications (e.g., near-duplicate image detection and visual search) where we care most about nearest neighbors within a small radius, FENSHSES could be in a more favorable position than FAISS.
\paragraph{RAM consumption.}
Since FAISS is implemented in main memory, its RAM consumption undoubtedly rises along with the increase in the size of dataset $\mathcal B$, as shown in Table \ref{tab: RAM}. In contrast, by leveraging the highly optimized disk-based index mechanics behind full-text search engines, FENSHSES consumes a much smaller amount of RAM when conducting nearest neighbor search. This property makes FENSHSES more cost-effective and thus more suitable especially to big-data applications.
\begin{table}[h!]
\begin{tabular}{|C{1.6cm}|C{1.6cm}|C{1.5cm}|C{1.5cm}|}
\hline
\textbf{\# of Bits} & \textbf{$r$} & \textbf{FAISS (ms)} & \textbf{FENSHSES (ms)}
\\
\hline
\hline
\multirow{3}{*}{64} & 3 & 34.0 & \textbf{5.8} \\
& 7& 37.0 & \textbf{25.7} \\
& 11& \textbf{42.7} & 117.7 \\
\hline
\multirow{3}{*}{256} & 15 & 42.9 & \textbf{7.8} \\
& 31& 42.8 & \textbf{22.5} \\
& 47& \textbf{45.4} & 77.7 \\
\hline
\multirow{3}{*}{1024} & 63 &79.2 & \textbf{31.6} \\
& 127& \textbf{81.9} & 89.9 \\
& 191& \textbf{90.4} & 250.0 \\
\hline
\multirow{3}{*}{4096} & 255 & 222.7& \textbf{134.2} \\
& 511& \textbf{223.2} & 612.5 \\
& 767& \textbf{223.3} & 1797.5 \\
\hline
\end{tabular}
\caption{{\bf Search latency.} {\normalfont FENSHSES is quite competitive for $r$-neighbor search when the Hamming distance $r$ is small, while the performance of FAISS is pretty robust with respect to $r$. This provides FAISS and FENSHSES different edges for the task of NNS.} \label{tab: search_latency}}
\end{table}
\begin{table}[h!]
\begin{tabular}{|C{1.6cm}|C{1.6cm}|C{1.5cm}|C{1.5cm}|}
\hline
\textbf{\# of Bits} & \textbf{$r$} & \textbf{FAISS (GB)} & \textbf{FENSHSES (GB)}
\\
\hline
\hline
\multirow{3}{*}{64} & 3 & 2.2 & 1.6 \\
& 7& 2.2 & 1.6\\
& 11& 2.2 & 1.6 \\
\hline
\multirow{3}{*}{256} & 15 & 2.3 & 1.6 \\
& 31& 2.3 & 1.6 \\
& 47& 2.3 & 1.6\\
\hline
\multirow{3}{*}{1024} & 63 & 2.9 & 1.6 \\
& 127& 2.9 & 1.6 \\
& 191& 2.9 & 1.6 \\
\hline
\multirow{3}{*}{4096} & 255 & 4.9 & 1.6 \\
& 511& 4.9 & 1.6 \\
& 767& 4.9 & 1.6 \\
\hline
\end{tabular}
\caption{{\bf Main memory (RAM) consumption.} {\normalfont The RAM consumed by FAISS substantially grows with the increase in the size of dataset $\mathcal B$. In contrast, FENSHSES consumes a constant amount of RAM, which is much smaller than the one consumed by FAISS.}} \label{tab: RAM}
\end{table}
\section{Conclusion}
In this case study, we compare FAISS and FENSHSES for the task of nearest neighbor search in Hamming space. By evaluating their performances in terms of speed in data indexing, search latency and RAM consumption, we hope practitioners could now better understand the pros and cons of the main memory based NNS solutions and the secondary memory based ones, and thus make their best choices accordingly (at least in NNS systems within binary cods). In the future, we will compare FAISS and FENSHSES under a wider range of applications; and moreover, we will also go beyond Hamming space to evaluate vector similarity search systems for general NNS problems.
\section*{Acknowledgement}
We are grateful to three anonymous reviewers for their helpful
suggestions and comments that substantially improve the paper. CM would like to thank Jun Zhao and Guang Yang for insightful discussions on FENSHSES. BY would like to thank Alessandro Magnani for helpful discussions on pHash and its related applications, and Zuzar Nafar for his support on this study.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\label{sec:introduction}
Is it possible to constrain the primordial black hole (PBH) density such that it cannot contribute to the entire dark matter density over its viable mass range? Answering this question will have important implications for the search of the identity of dark matter and inflationary dynamics which can give rise to PBHs\,\cite{1966AZh....43..758Z, Hawking:1971ei, Carr:1974nx, Meszaros:1975ef, Carr:1975qj, Khlopov:2008qy}. In this Letter, we take one step toward answering this question. We show that combining the observation that light PBHs can produce $e^\pm$ pairs via evaporation\,\cite{Boudaud:2018hqb} with the fact that an intense 511 keV gamma-ray line has been observed in the Galactic Center\,\cite{Knodlseder:2005yq, Milne:2006ad, Siegert:2019tus, Prantzos:2010wi, Siegert:2015knp, Churazov:2010wy, Jean:2005af, Kierans:2019pkh} can efficiently constrain PBHs in a mass range which cannot yet be constrained by any other technique. The morphology of the 511 keV gamma-ray line (it has a bulge and a disk component) is such that primordial black holes, acting as the dark matter, cannot explain the entire emission. We do not yet know the source of these low-energy astrophysical positrons; therefore an understanding of the underlying astrophysical source(s)\,\cite{Lingenfelter:2009kx, Higdon:2007fu, Panther:2018xvc, Alexis:2014rba, Panther:2019fre} can further improve our constraints.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fDM_monochromatic.pdf}
\caption{Upper limits on the fraction of dark matter which can be composed of primordial black holes for monochromatic mass distribution. The black lines show the limits derived in this Letter. These limits depend on the dark matter density near the Galactic Center (NFW and isothermal) and the propagation of low-energy positrons. While deriving the limit for the ``3 kpc" (``1.5 kpc") constraint, we assume that low-energy positrons can travel about a kpc (100 pc) before annihilating. The upper limits from Voyager 1 (two lines denote the propagation and background uncertainties)\,\cite{Boudaud:2018hqb}, Planck\,\cite{Clark:2016nst}, and gamma-ray observatories\,\cite{Carr:2016hva, Carr:2009jm} are in red dotted lines.}
\label{fig:fDM monochromatic}
\end{figure}
The identity of dark matter is one of the most enduring mysteries of physics. Numerous astrophysical and cosmological observations give an irrefutable indication of the presence of dark matter, yet an absence of its microphysical understanding drives a great deal of research. A large number of dark matter candidates have been proposed in the literature and these range in masses from $\sim$ 10$^{-22}$ eV to a few hundred $M_\odot$. PBHs are one of the oldest candidates of dark matter and their abundance has been studied in a number of ways. The various constraints on PBHs arise from evaporation (and the subsequent detection of Standard Model particles), capture on astronomical bodies, lensing observations, dynamics of galaxies, gravitational wave observations, and accretion\,\cite{Katz:2018zrn, Graham:2015apa, Montero-Camacho:2019jte, Carr:2009jm, Takhistov:2017nmt, Niikura:2017zjd, Griest:2013esa, Stocker:2018avm, Allsman:2000kg, Tisserand:2006zx, Wyrzykowski:2011tr, Hektor:2018qqw, Clark:2018ghm, Zumalacarregui:2017qqd, Mediavilla:2017bok, Brandt:2016aco, Koushiappas:2017chw, Monroy-Rodriguez:2014ula, Wang:2016ana, Clesse:2016ajp, Ali-Haimoud:2016mbv, Poulin:2017bwe, Bernal:2017vvn, Sugiyama:2019dgt}. The recent detection of gravitational waves from binary black hole mergers\,\cite{Abbott:2016blz, LIGOScientific:2018jsj} have rekindled an interest in the contribution of PBHs to the dark matter energy density\,\cite{Bird:2016dcv, Clesse:2016vqa, Sasaki:2016jop}. This has led to a detailed reanalysis of older constraints\,\cite{Katz:2018zrn, Blum:2016cjs, Ali-Haimoud:2016mbv, Ali-Haimoud:2016mbv, Poulin:2017bwe}, research into new ways to constrain PBHs (for e.g., lensing of fast radio bursts\,\cite{Munoz:2016tmg, Laha:2018zav} and other techniques\,\cite{Jung:2019fcs, Bai:2018bej}), and the study of spinning PBHs\,\cite{Arbey:2019mbc, Arbey:2019jmj, Arbey:2019vqx}. A detailed study of older constraints has shown that there are viable regions of parameter space where primordial black holes can satisfy the entire dark matter density\,\cite{Katz:2018zrn, Carr:2016drx}. Ref.\,\cite{Boudaud:2018hqb} pointed out that PBHs with masses $\lesssim$ 10$^{14}$ kg can produce $e^\pm$ pairs via Hawking radiation and such a process can be constrained via the observations of the Voyager 1 satellite\,\cite{2013Sci...341..150S, 2016ApJ...831...18C}. This naturally leads us to wonder about the fate of $e^\pm$ pairs produced via PBH evaporation in the Galactic Center. PBHs are much more numerous in the Galactic Center than in the Solar circle (where the Voyager 1 observations were made), and thus a stronger bound can be expected if there is an appropriate observable. Thus, we are led to the question: is there an observable in the Galactic Center which points to the fact that $e^\pm$ pairs are copiously present there?
The answer is yes. There is a smoking gun signature which indicates that there is a huge reservoir of low-energy positrons near the Galactic Center. For many decades, an enduring astrophysical mystery is the observation of 511 keV gamma-ray line at the Galactic Center (see Ref.\,\cite{Prantzos:2010wi} for a historical account). This gamma-ray line has been observed by a number of observatories and a detailed study has been made by the SPI/ INTEGRAL observatory. Despite the intense scrutiny of this signal, we do not yet know the origin of this signal. Many viable astrophysical models (i.e., models which do not require a dark matter origin of the 511 keV signal) have been proposed\,\cite{Totani:2006zx, Bandyopadhyay:2008ts, Wang:2005cqa, Milne:2001zs, Bartels:2018eyb, Fuller:2018ttb, Alexis:2014rba, Crocker:2016zzt, Venter:2015gga, Prantzos:2005pz, Weidenspointner:2008zz, Bisnovatyi-Kogan:2016dgr}, although none are confirmed to be the source of these low energy positrons. A detailed morphological study of this signal and its absence in the dwarf galaxies\,\cite{Siegert:2016ijv} indicate that this it is not produced via dark matter interactions\,\cite{Wilkinson:2016gsy} (see, however, \cite{Farzan:2017hol} for a particle dark matter model which can explain the signal). Earlier studies trying to connect PBHs and the Galactic Center 511 keV line can be found in Refs.\,\cite{1980AA....81..263O, okeke1980primary, 1991ApJ...371..447M, Bambi:2008kx}. Thus, any astrophysical source (present in the Galactic Center/ Galactic bulge) which produces low-energy positrons can be constrained via this observation. The fact that low mass PBHs can produce positrons in large quantities leads us to expect that this observation can be a stringent constraint on the PBH abundance.
The upper limits on the PBH density assuming a monochromatic mass function derived in this work are shown in Fig.\,\ref{fig:fDM monochromatic} by the black solid and dashed lines. Various black lines indicate the dependence of this upper limit on the underlying astrophysical parameters. A part of this parameter space is already probed by gamma-ray observations, cosmic microwave background (CMB) observations, and Voyager 1 measurements. Our constraints are stronger than these and probe a new mass range of PBHs. Our constraints close a part of a mass window where PBHs could have contributed to the entire dark matter energy density of the Universe. Our technique introduces a new electromagnetic probe of PBHs beyond what has already been discussed in the literature\,\cite{Ali-Haimoud:2019khd}.
\section{Formalism}
\label{sec:formalism}
In natural units, the temperature of a black hole of mass $M_{\rm BH}$ is\,\cite{Hawking:1974rv, Hawking:1974sw}
\begin{eqnarray}
T_{\rm BH} = \dfrac{1}{8\pi G_N M_{\rm BH}} = 1.06 \, \left(\dfrac{10^{10} \, {\rm kg}}{M_{\rm BH}} \right) \, {\rm GeV} \, ,
\label{eq:black hole temperature}
\end{eqnarray}
where $G_N$ denotes Newton's gravitational constant. The temperature of the black hole also dictates the rate at which particles are produced via evaporation. The energy spectrum of these particles follows the distribution\,\cite{MacGibbon:1990zk}
\begin{eqnarray}
\dfrac{dN_s}{dE} = \dfrac{\Gamma_{\rm s}}{2\pi} \, \int dt \, \dfrac{1}{{\rm exp}(E/T_{\rm BH}) - (-1)^{\rm 2s}} \, ,
\label{eq:evaporation rate}
\end{eqnarray}
where the dimensionless absorption probability is denoted by $\Gamma_{\rm s}$, s denotes the particle spin, and $E$ denotes the energy of the emitted particle\,\cite{MacGibbon:1990zk, Page:1976df, Page:1976ki, Page:1977um}. Since the positrons in our case if interest are semi-relativistic to non-relativistic, we use the full formula of the dimensionless absorption probability as in eqn.\,6 of Ref.\,\cite{MacGibbon:1990zk}. The values of the absorption cross-section $\sigma_s$ is taken from Fig.\,1 of Ref.\,\cite{Page:1977um}. We also take into account the factor of 2 for the two chiralities of the positron in the full formula of the dimensionless absorption probability. Since a black hole loses mass via evaporation, $T_{\rm BH}$ is a function of time. In our calculation, we will use the observed positron injection luminosity of over one year, and the mass loss during this time is negligible for the black hole masses that we consider. As such, $T_{\rm BH}$ will be a constant for a given black hole mass in our Letter.
The Galactic Center 511 keV gamma-ray line has been observed for a few decades, and its origin has remained unknown throughout. Recent attempts at measuring the Doppler shifts have also not led to the identification of the source\,\cite{Siegert:2019tus} (note that a similar search technique has also been proposed for the 3.5 keV line\,\cite{Speckhard:2015eva, Powell:2016zbo}). The observed flux of this gamma-ray line indicates that the rate of positron annihilation at the Galactic Center is $\sim$ 6.3 $\times$ 10$^{50}$ per year\,\cite{Prantzos:2010wi}. In order to respect the constraints due to the continuum gamma-ray emission measurement, the positrons must be injected at an energy $\lesssim$ 3 MeV\,\cite{Beacom:2005qv}.
\begin{figure*}
\centering
\includegraphics[angle=0.0,width=0.49\textwidth]{fDM_lognormal.pdf}
\includegraphics[angle=0.0,width=0.49\textwidth]{fDM_powerlaw.pdf}
\caption{Upper limits on the fraction of dark matter which can be composed of primordial black holes for extended mass distribution. {\bf Left panel:} We use the log-normal mass distribution with four choices of $\sigma$: 0.1, 0.5, 1, and 2. The constraints from the Voyager 1 observation is shown as a dotted red line. When $\sigma$ = 2, Voyager 1 excludes the entire parameter space shown in the figure. Our limits are shown by the black lines. The solid (dashed) black line corresponds to the ``NFW 3 kpc" (``Iso 1.5 kpc") assumption. {\bf Right panel:} We use the power law mass distribution with two different values of the index, $\gamma$ = $\pm$ 1 and $M_{\rm min}$ = 4 $\times$ 10$^{11}$ kg.}
\label{fig:fDM extended mass distributions}
\vspace{0.1cm}
\end{figure*}
The constraint on the dark matter fraction of PBHs can easily be calculated in the following way. We first assume that all PBHs have a common mass, i.e., they follow a monochromatic mass distribution. The number of $e^\pm$ pairs injected per unit time, per unit energy, and per unit volume is given by\,\cite{Boudaud:2018hqb}
\begin{eqnarray}
Q(E, r) = \dfrac{\rho_{\rm DM}(r)}{M_{\rm PBH}} \, \dfrac{d^2N_e}{dE dt} \,,
\label{eq:positrons injected per unit time, energy, and volume}
\end{eqnarray}
where $\rho_{\rm DM} (r)$ denotes the dark matter density as a function of the distance from the Galactic Center, $r$. The mass of the PBH is denoted by $M_{\rm PBH}$. In this situation, the upper limit on the fraction of dark matter in the form of PBHs is
\begin{eqnarray}
f_{\rm DM} \lesssim \dfrac{6.3 \times 10^{50}}{\dfrac{\int dV \rho_{\rm DM}(r)}{M_{\rm PBH}} \, \int dE \dfrac{d^2N_e}{dt \, dE} \times 1\, {\rm year}} \, ,
\label{eq:fDM monochromatic upper limit}
\end{eqnarray}
where $dV$ denotes a differential volume element in our Galaxy, and where we use the positron injection luminosity criteria to derive our limits\,\cite{Fuller:2018ttb}. Note that we simply require that the positron injection luminosity is less than what is measured by SPI/ INTEGRAL. This is the most conservative option: if we had assumed the contribution of some astrophysical source(s) to this gamma-ray observation, then the upper limit on the positron injection luminosity would have been much lower. The flux of this line in the bulge and the central source of our Galaxy is (0.96 $\pm$ 0.07) $\times$ 10$^{-3}$ photons cm$^{-2}$ s$^{-1}$ and (0.08 $\pm$ 0.02) $\times$ 10$^{-3}$ photons cm$^{-2}$ s$^{-1}$, respectively\,\cite{Bartels:2018eyb}. Thus, our limits can be improved by more than an order of magnitude if one identifies the class(es) of astrophysical source(s) which is (are) contributing to this emission in the Galactic bulge.
One source of uncertainty to the upper limit presented in this Letter is the dark matter density profile near the Galactic Center. We bracket this source of uncertainty by assuming two different dark matter profiles: Navarro-Frenk-White (NFW) and isothermal. Owing to the cusp present in the NFW profile, the constraints derived using this input is stronger. We use the forms of the dark matter profiles used in Ref.\,\cite{Ng:2013xha}. An additional uncertainty arises from the positron propagation distance in the Galactic Center. The extent of the bulge component of the 511 keV gamma-ray emission is $\lesssim$ 1.5 kpc from the Galactic Center\,\cite{Fuller:2018ttb}. Non-relativistic positrons can travel a small distance (few hundred parsecs) before annihilation; thus, we assume that all positrons produced from PBHs within 1.5 kpc annihilate in order to derive our upper limits. However, a clear understanding of positron propagation in the Galactic Center region is still lacking, and there are scenarios via which positrons can be advected up to a distance of $\sim$ 1 kpc\,\cite{Higdon:2007fu, Panther:2018xvc}. Ref.\,\cite{Higdon:2007fu} finds that 80\% of the positrons born within 3.5 kpc of the Galactic Center annihilate within 1.5 kpc of the Galactic Center. In order to include this propagation uncertainty, we assume that 80\% of the positrons from PBHs within a 3 kpc radius of the Galactic Center can annihilate within the inner 1.5 kpc. Our calculations find that the strongest (weakest) upper limits are derived when assuming a NFW (isothermal) profile and including all the PBHs within a 3 kpc (1.5 kpc) radius of the Galactic Center. These cases bracket the uncertainties due to the dark matter halo and positron propagation. For all of the cases that we consider, our main conclusions remain true: our limits are more stringent than existing bounds and probe part of the mass range where PBH density was completely unconstrained from previous studies. The upper limit of the energy integral in Eqn.\,\ref{eq:fDM monochromatic upper limit} is determined by the diffuse gamma-ray data which are produced due to in-flight annihilation of the positrons. Following Ref.\,\cite{Beacom:2005qv}, we take this upper limit to be 3 MeV. Since the overwhelming majority of the positrons come to rest before undergoing annihilation, we take the lower limit of the energy integral to be the positron mass.
Various inflationary models also predict an extended mass distribution of PBHs\,\cite{Carr:1975qj, Clesse:2015wea, Carr:2017edp, Hawking:1982ga, Hawking:1987bn, Matsuda:2005ez, Berezin:1982ur, Pi:2017gih, Bhaumik:2019tvl, Escriva:2019phb}. It is important to study these models in detail as the possibility of whether PBHs contribute significantly to the dark matter density depend substantially on the underlying mass distribution\,\cite{Carr:2016drx, Carr:2017jsz, Green:2016xgy, Kuhnel:2017pwq, Bellomo:2017zsr}. Traditionally, the limits on the PBH density were presented only for monochromatic mass distributions, however, there has been recent interest in converting these limits for various extended mass distributions of PBHs. We consider two different classes of extended mass distributions: a log-normal distribution and a power law mass distribution. A log-normal mass distribution (with $\mu$ and $\sigma$ denoting the mean and the standard deviation of the logarithm of the mass distribution respectively) is defined as
\begin{eqnarray}
\dfrac{d\Phi_{\rm PBH}}{dM_{\rm PBH}} = \dfrac{e^{-\dfrac{\rm ln ^2 \, (M_{PBH}/ \mu)}{2 \sigma^2}}}{\sqrt{2\pi} \sigma M_{\rm PBH}} \, .
\label{eq:log-normal distribution}
\end{eqnarray}
A power law mass distribution is parametrized by the power law index, $\gamma$, the maximum, $M_{\rm max}$, and the minimum value, $M_{\rm min}$ of the mass distribution:
\begin{eqnarray}
\dfrac{d\Phi_{\rm PBH}}{dM_{\rm PBH}} = \dfrac{\gamma}{M_{\rm max}^\gamma - M_{\rm min}^\gamma} \, \dfrac{1}{M_{\rm PBH}^{1 - \gamma}}\, ,
\label{eq:power law distribution}
\end{eqnarray}
for $M_{\rm PBH}$ $\in$ $[M_{\rm min}, M_{\rm max}]$ and $\gamma$ $\neq$ 0. In order to derive the limits for these extended mass distributions, we convolve the denominator in Eqn.\,\ref{eq:fDM monochromatic upper limit} with the extended mass distribution:
\begin{eqnarray}
f_{\rm DM} \lesssim \dfrac{6.3 \times 10^{50}/ (1\, {\rm year})}{\int dM_{\rm PBH} \, \dfrac{d\Phi_{\rm PBH}}{dM_{\rm PBH}} \dfrac{\int dV \rho_{\rm DM}(r)}{M_{\rm PBH}} \, \int dE \dfrac{d^2N_e}{dt \, dE}}\,.
\label{eq:fDM extended mass distribution}
\end{eqnarray}
\section{Results}
\label{sec:results}
The results of our Letter is shown in Figs.\,\ref{fig:fDM monochromatic} and \ref{fig:fDM extended mass distributions} for the monochromatic mass function and the extended mass function respectively. The upper limits for the monochromatic mass function are shown for two different dark matter profiles (the NFW and isothermal profiles) and for two different sizes of the spherical region which hosts the PBHs. These bracket the uncertainty due to the dark matter profile and positron propagation. Our constraints are derived by simply assuming that the positrons from PBHs do not overproduce the positron injection luminosity. As such, our derived constraints are maximally conservative and robust.
The other constraints shown in the paper are from the observations of Voyager 1\,\cite{Boudaud:2018hqb}, cosmic microwave background\,\cite{Clark:2016nst}, and gamma-rays\,\cite{Carr:2016hva, Carr:2009jm}. The constraints from the Galactic gamma-ray background and the cosmic microwave background is relatively robust with respect to the underlying parameters. The constraints using the Voyager 1 observations are dependent on the background modeling and cosmic ray propagation uncertainties. However, even the least stringent bound from Voyager 1 observations is most constraining at low PBH masses. Depending on the underlying astrophysical parameters, our constraints are the strongest in the mass range $\sim$ 10$^{13}$ kg to $\sim$ 10$^{14}$ kg and probe completely new parameter space.
We also show our derived upper limits for the log-normal mass distribution of PBHs in the left panel of Fig.\,\ref{fig:fDM extended mass distributions}. In order to compare our results with the Voyager 1 observations, we use the same values of the log normal mass distribution. When $\sigma$ = 0.1, our derived constraints are stronger than those derived in Ref.\,\cite{Boudaud:2018hqb} for $\mu$ $\gtrsim$ (1 -- 2) $\times$ 10$^{13}$ kg. When $\sigma$ = 0.5, our derived constraints are stronger than those derived in Ref.\,\cite{Boudaud:2018hqb} for $\mu$ $\gtrsim$ (2 -- 4) $\times$ 10$^{13}$ kg. When $\sigma$ = 1, our derived constraints are better than the ones in Ref.\,\cite{Boudaud:2018hqb} only for the NFW 3 kpc case for $\mu$ $\gtrsim$ 10$^{14}$ kg. The constraints derived in Ref.\,\cite{Boudaud:2018hqb} are better than our constraints in the entire parameter space when $\sigma$ = 2. This illustrates the complementarity of our constraints with that using the Voyager 1 observations. Even for the log-normal mass distribution, we find that our technique probes new parts of the parameter space.
Our derived upper limits for the power law distribution is shown in the right panel of Fig.\,\ref{fig:fDM extended mass distributions}. We assume $\gamma$ = $\pm$ 1 for the two cases that we consider. We assume $M_{\rm min}$ = 4 $\times$ 10$^{11}$ kg and vary $M_{\rm max}$ to derive our upper limits. Depending on the dark matter profile and the volume under consideration, the upper limits on $f_{\rm DM}$ are less than 10\% when $M_{\rm max}$ varies between 10$^{12}$ kg to 10$^{14}$ kg. In this case, we do not show the limits from the Voyager 1 observations as these were not presented in Ref.\,\cite{Boudaud:2018hqb}.
\section{Conclusions}
\label{sec:conclusions}
The direct detection of gravitational waves have started vigorous activities in the field of gravitational wave astroparticle physics. These observations have prompted many new ideas (see, for example, \cite{Kopp:2018jom}). This discovery has also led to renewed interest in PBHs as a dark matter candidate. Remarkably, it has been pointed out by several authors that there are regions of the parameter space where PBHs can contribute to the entire dark matter energy density of the Universe. In this Letter, we introduce such a technique, which can probe a region of parameter space which cannot be probed by any other observation. We combine the theoretical insight that low mass PBHs evaporate to produce $e^\pm$ pairs with the SPI/ INTEGRAL observations of the Galactic Center 511 keV gamma-ray line to present the strongest constraint on the PBH density over a wide range in the PBH masses. We simply require that the positron injection luminosity due to PBH evaporation does not exceed that observed to explain the gamma-ray line observation. Our constraints depend on the dark matter density and the propagation of low-energy positrons. Considering all astrophysical uncertainties, we find that our constraints are the most stringent over a wide range in the PBH masses and probe a new part of the parameter space. An identification of the astrophysical source(s) responsible for the 511 keV gamma-ray emission can further improve our constraints by more than an order of magnitude.
{\bf Note:} Ref.\,\cite{PhysRevLett.123.251102}, studying a similar topic, appeared in arXiv while this work was in preparation. The authors use some different inputs (for e.g., in the upper limit of the volume integration) and our results are comparable. Both the papers use different inputs and agree on the general message of this technique: this is a new way to probe PBHs in a interesting mass range.
\section*{Acknowledgments}
We thank Jeremy Auffinger, John F.\,Beacom, Mathieu Boudaud, Marco Cirelli, Joachim Kopp, and Toby Opferkuch for discussions and comments. We especially thank Anupam Ray for extensive discussions. We thank Jane MacGibbon for pointing out an artefact due to the pdf'ing of her paper and for alerting us to relevant older literature. We are especially grateful to Don Page for providing us the data of his paper. R.L. acknowledges the support provided by the CERN Theory group.
\bibliographystyle{JHEP}
|
\setcounter{equation}{0}\Section{\setcounter{equation}{0}\Section}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{example}[theorem]{Example}
\newtheorem{claim}[theorem]{Claim}
\newtheorem{assertion}{Assertion}[section]
\newtheorem{remark}{Remark}
\DeclareMathOperator{\cat}{cat}
\newcommand\R{\mathbb R}
\newcommand\mbox{${\rm{I\!N}}$}{\mathbb N}
\newcommand\D{\mathcal{D}}
\newcommand\mbox{${\rm{C\hspace{-1.8mm}\rule{0.3mm}{2.8mm}}}$}{\mathbb C}
\begin{document}
\title[ Existence of solution for a class of problem in whole .......]
{Existence of solution for a class of problem in whole $\mathbb{R}^N$ without the Ambrosetti-Rabinowitz condition }
\author{Claudianor O. Alves}
\author{Marco A. S. Souto }
\address[Claudianor O. Alves and Marco A.S. Souto]{\newline\indent Unidade Acad\^emica de Matem\'atica
\newline\indent
Universidade Federal de Campina Grande,
\newline\indent
58429-970, Campina Grande - PB - Brazil}
\email{\href{mailto:<EMAIL>, <EMAIL>}{<EMAIL>, <EMAIL>}}
\pretolerance10000
\begin{abstract} In this paper we study the existence of solution for a class of elliptic problem in whole $\mathbb{R}^N$ without the well known Ambrosetti-Rabinowitz condition. Here, we do not assume any monotonicity condition on $f(s)/s$ for $s>0$.
\end{abstract}
\thanks{Claudianor Alves was partially supported by CNPq/Brazil Proc. 304804/2017-7 ; Marco A.S. Souto was partially
supported by CNPq/Brazil Proc. 306082/2017-9 }
\subjclass[2010]{Primary 35A15; Secondary 35B38; 35J20 }
\keywords{Variational methods; Critical points; Superlinear problems; Elliptic equations}
\maketitle
\setcounter{equation}{0}\Section{Introduction}
In the last decades a lot of authors have studied the existence of solution for elliptic problems of the form
$$
\left\{
\begin{array}{l}
-\Delta{u}=f(x,u), \quad \mbox{in} \quad \Omega \\
u=0, \quad \mbox{on} \quad \partial \Omega,
\end{array}
\right.
\leqno{(P)}
$$
where $\Omega \subset \mathbb{R}^N$ is a smooth bounded domain and $f$ is a continuous function with subcritical growth satisfying some technical conditions. In general, the main conditions on $f$ are the following: \\
$$
\displaystyle \lim_{s \to 0}\frac{f(s)}{s}=0. \leqno{(f_1)}
$$
\noindent There is $q \in (2,2^*)$, where $2^*=\frac{2N}{N-2}$, if $N \geq 3$ or $q \in (2,+\infty)$ if $N=1,2$ such that
$$
\displaystyle \limsup_{|s| \to +\infty }\frac{|f(s)|}{|s|^{q-1}}<+\infty. \leqno{(f_2)}
$$
\noindent (Ambrosetti-Rabinowitz condition) There are $\theta >2$ and $M>0$ such that
$$
0<\theta F(s) \leq f(s)s, \quad \quad \mbox{for} \quad |s| \geq M, \leqno{(f_3)}
$$
where $F(s)=\displaystyle \int_{0}^{s}f(t)\,dt$.
By using the Mountain Pass Theorem due to Ambrosetti and Rabinowitz \cite{AR}, it is possible to prove that the energy functional associated with $(P)$, given by
$$
I(u)=\frac{1}{2}\int_{\Omega}|\nabla u|^{2}\,dx-\int_{\Omega}F(u)\,dx, \quad \forall u \in H_{0}^{1}(\Omega),
$$
satisfies the well known $(PS)$ condition. Here, assumption $(f_3)$ permits to prove, of a very easy way, that all $(PS)$ sequences are bounded. However, in the last years, we have observed that in a lot papers, the authors have used a more weak condition that $(f_3)$, more precisely, instead of $(f_3)$ the following conditions are assumed:
$$
\frac{F(s)}{s^2} \to +\infty \quad \mbox{as} \quad |s| \to +\infty. \leqno{(f_4)} \\
$$
and
$$
\mbox{The function} \quad \frac{f(s)}{|s|} \quad \mbox{ is an increasing function of} \quad s \in \mathbb{R} \setminus \{0\}. \leqno{(f_5)}
$$
The condition $(f_5)$ is sometimes replaced by
$$
\mathcal{F}(s)=f(s)s-2F(s) \quad \mbox{is increasing for} \quad \mbox{for} \quad s >0, \, \mbox{and decreasing for} \, s<0. \leqno{(f_5)'}
$$
The literature is large for problems without Ambrosseti-Rabinowitz condition, we would like to cite the papers by Costa and Magalh\~aes \cite{Costa}, Jeanjean and Tanaka \cite{JeanjeanTanka}, Liu and Wang \cite{Liu-Wang}, Miyagaki and Souto \cite{Miyagaki-Souto}, Schechter and Zou \cite{SZ}, Struwe and Tarantello \cite{ST}, Wang and Wei \cite{Wang-Wei}, Zhou \cite{Zhou} and their references. In general, the main tool used in the above mentioned papers is Mountain Pass Theorem with Cerami's condition found in Bartolo, Benci and Fortunato \cite{BBF}.
Here, we would like point out that in the seminal paper \cite{Costa}, Costa and Magalh\~aes established the existence of solution for $(P)$ without Ambrosseti-Rabinowitz condition by supposing, among others conditions,
$$
\displaystyle \liminf_{|s| \to +\infty }\frac{f(s)s-F(s)}{|s|^{\mu}} \geq a>0, \quad \leqno{(f_6)}
$$
with $\mu > \frac{N}{2}(q-2)$ and $q$ as in $(f_2)$. Since that $\Omega$ is a bounded domain, the authors were able to show that $(f_2)$ and $(f_6)$ combine to give the boundedness of $(PS)$ sequences for the energy functional associated with $(P)$, which is a key point to prove $(PS)$ condition. In that paper, an interesting point is that the authors did not assume any monotonicity condition on $\frac{f(s)}{s}$ or $\mathcal{F}(s)=f(s)s-2F(s)$ for $s>0$.
The existence of solution for elliptic problem in whole $\mathbb{R}^N$ like
$$
\left\{
\begin{array}{l}
-\Delta{u}+V(x)u=f(u), \quad \mbox{in} \quad \mathbb{R}^N, \\
u \in H^{1}(\mathbb{R}^N)
\end{array}
\right.
\leqno{(P_1)}
$$
without Ambrosseti-Rabinowitz condition also have been considered in some papers, see for example, Jeanjean \cite{Jeanjean}, Liu \cite{Liu1} and their references. In \cite{Jeanjean}, Jeanjean has proved a very interesting Abstract Theorem that permits to work with a large class of problem without Ambrosetti-Rabinowitz condition in bounded or unbounded domain. In that paper the author assumes $V=1$ and that there is $D \in [1, +\infty)$ such that
\begin{equation} \label{ZE}
Z(s) \leq DZ(t), \quad \mbox{for} \quad 0 \leq s \leq t,
\end{equation}
where $Z(s)=\frac{1}{2}f(s)s-F(s)$ for $s \geq 0$. In \cite{Liu1}, Liu used essentially the same arguments explored in \cite{Miyagaki-Souto}, \cite{SZ} and \cite{ST} by supposing that $V$ is $\mathbb{Z}^N$-periodic or
$$
V(x)<V_\infty=\lim_{|x| \to +\infty}V(x).
$$
Related to the function $f$, Liu also assumed a condition like (\ref{ZE}).
Motivated by the above references, the present paper is concerned with the existence of solution for $(P_1)$ without Ambrosseti-Rabinowitz condition. Here, $f:\R \to \mathbb R$ is continuous function that satisfies $(f_1)$ and the following conditions: \\
\noindent There are $p\in\left(2,\frac{2N+4}{N}\right)$ and $C>0$ such that
$$
|f(s)|\leq C (1+ |s|^{p-1}) \quad \mbox{for all} \quad s \in \mathbb R; \leqno{(f_7)}
$$
\noindent For each $s_o>0$ there is a $\sigma_o>0$ such that
$$
\mathcal{F}(s)=sf(s)-2 F(s)\geq \sigma s^2,\mbox { for all } |s|\geq s_o. \leqno{(f_8)}
$$
Here, it is important to recall that $(f_8)$ is weaker than $(f_3)$.
\vspace{0.5 cm}
Related to the $V$, we assume that it is continuous and belongs to one of the following classes: \\
\noindent {\bf Class 1:} \, $V$ is a periodic function
\begin{itemize}
\item[$(V_1)$] \, $V$ is $\mathbb{Z}^N$ periodic continuous function, that is,
$$
V(x+y)=V(x), \quad \forall x \in \mathbb{R}^N \quad \mbox{and} \quad \forall y \in \mathbb{Z}^N.
$$
\item[$(V_2)$] \, $\displaystyle \inf_{x \in \mathbb{R}^N}V(x)=V_o>0$.
\end{itemize}
\noindent {\bf Class 2:} \, $V$ is coercive :
\begin{itemize}
\item[$(V_3)$]
$$V(x) \to +\infty \quad \mbox{as} \quad |x| \to +\infty. $$
\end{itemize}
\noindent {\bf Class 3:} \, $V$ is a Barstch \& Wang type potential, that is,
\begin{itemize}
\item[$(V_4)$] \, $V$ is a continuous function of the form
$$
V(x)=1+\lambda W(x),
$$
where $\lambda >0$ and $W$ is a nonnegative function.
\end{itemize}
Our first result is related to the case where the Ambrosetti-Rabinowiz condition only holds at infinity.
\begin{theorem} \label{T0} Assume $(f_1)-(f_3)$ and $(f_5)'$. Suppose that $V$ belongs to Class 1,2 or 3, then, $(P_1)$ has a ground state solution. When $V$ belongs to Class 3, the existence of solution holds for large $\lambda$.
\end{theorem}
The Theorem \ref{T0} is a crucial step to prove our second result that establishes the existence of solution for $(P_1)$ without Ambrosetti-Rabinowiz condition and it has the following statement
\begin{theorem} \label{T1} Assume $(f_1),(f_7)-(f_8)$ and that $V$ belongs to Class 1,2 or 3. Then, $(P_1)$ has a ground state solution. When $V$ belongs to Class 3, the existence of solution holds for large $\lambda$.
\end{theorem}
As a consequence of Theorem \ref{T1}, we can consider a more general class of problems like
$$
\left\{
\begin{array}{l}
-\Delta{u}+V(x)u=f(u) + \beta h(u) , \quad \mbox{in} \quad \mathbb{R}^N,\\
u \in H^{1}(\mathbb{R}^N),
\end{array}
\right.
\leqno{(P_2)}
$$
where $\beta >0$ and $V,f$ satisfy the conditions of Theorem \ref{T1}. Related to the $h$, we assume that it is continuous and satisfies:
\begin{itemize}
\item[$(h_1)$] $ \displaystyle{\lim_{s\to 0}\frac{h(s)}{s}=0}$ ;
\item[$(h_2)$] There are $C>0$ and $r \in (2,2^*)$ such that $|h(s)|\leq C(1+ |s|^{r-1})$ for all $s\in \mathbb R$;
\item[$(h_3)$] For each $s_o>0$, there is a $\sigma_o>0$ such that $$\mathcal{H}(s)=sh(s)-2 H(s)\geq \sigma_o s^2,\mbox { for all } |s|\geq s_o.$$
\item[$(h_4)$] There is $\Upsilon>0$ such that
$$
0< \frac{h(\Upsilon)}{f(\Upsilon)}=\max\left\{\frac{h(t)}{f(t)}\,:\, t \in [0,\Upsilon]\right\} \quad \mbox{and} \quad \frac{h(\Upsilon)}{f(\Upsilon)}=\min\left\{\frac{h(t)}{f(t)}\,:\, t \in [\Upsilon,+\infty)\right\}.
$$
\end{itemize}
We would like point out that $(h_4)$ holds, for example, if $h(s)/f(s)$ is a nondecreasing function for $s>0$.
Our third theorem is the following
\begin{theorem} \label{T2} Assume $(f_1),(f_7)-(f_8)$, $(h_1)-(h_4)$ and that $V$ belongs to Class 1,2 or 3. Then, setting $\beta^*=\frac{f(\Upsilon)}{h(\Upsilon)}$, there is $\Upsilon^*>0$ such that $(P_2)$ has a nontrivial solution for all $\beta \in [0, \beta^*)$ and $\Upsilon \in [\Upsilon^*, +\infty)$. As in Theorem 1, the existence of solution for Class 3 holds when $\lambda$ is large.
\end{theorem}
The plan of the paper is as follows: Section 2 deals with Theorem \ref{T0}, while Sections 3 and 4 are devoted to the proofs of Theorems \ref{T1} and \ref{T2} respectively. In Section 5, we point out that if $\displaystyle{\frac{h(s)}{f(s)} \to +\infty}$, we also have an existence result replacing $(h_3)$ by a weaker assumption, namely: \, For each $s_o>0$, there is a $\sigma_o>0$ such that
$$
\mathcal{H}(s)=sh(s)-2 H(s)\geq -\sigma_o s^2,\mbox { for all } |s|\geq s_o. \leqno{(h'_3)}
$$
\setcounter{equation}{0}\Section{Proof of Theorem \ref{T0}}
In this section, we will show the existence of solution for a class of auxiliary problem, which proves Theorem \ref{T0} and will be used in the proof of Theorem \ref{T1}. Specifically, we will study the existence of solution for the following class of problem
$$
\left\{
\begin{array}{l}
-\Delta{u}+V(x)u=g(u), \quad \mbox{in} \quad \mathbb{R}^N, \\
u \in H^{1}(\mathbb{R}^N),
\end{array}
\right.
\leqno{(AP)}
$$
where $V$ belongs to Class 1,2 or 3, and $g:\mathbb{R} \to \mathbb{R}$ is a continuous function verifying:
\begin{itemize}
\item[$(g_1)$] $ \displaystyle{\lim_{s\to 0}\frac{g(s)}{s}=0}$ ;
\item[$(g_2)$] There are $C>0$ and $p \in\left(2,\frac{2N}{N-2}\right)$ such that $|g(s)|\leq C (1+ |s|^{p-1})$ for all $s\in \mathbb R$ ;
\item[$(g_3)$] There are $\theta>2$ and $M>0$ such that $sg(s)-\theta G(s)>0$ for all $|s|\geq M$ ;
\item[$(g_4)$] $\mathcal{G}(s)=sg(s)-2 G(s)>0$ for all $0<|s|\leq M$.
\end{itemize}
\vspace{0.5 cm}
Observe that $(f_1) -(f_3)$ and $(f_5)'$ imply $(g_1)-(g_4)$.
\vskip .5 cm
In the sequel, we mention some facts that involve conditions $(g_1)-(g_4)$. It is easy to check that $(g_1)-(g_2)$ imply in the inequality below
\begin{equation}\label{eq00}
|G(s)|\leq c |s|+ \frac cp|s|^{p}, \quad \forall s\in \mathbb R.
\end{equation}
The condition $(g_3)$ is the usual Ambrosetti-Rabinowitz at infinity, however we would like point out that we are not assuming this condition near the origin, which is usually assumed in whole $\mathbb{R}$, when we are considering elliptic problem in whole $\mathbb{R}^N$. Moreover, we do not assume any monotonicity condition on $\displaystyle{\frac {g(s)}s} $, which is in general assumed in a lot of papers without the Ambrosetti-Rabinowitz condition. Finally, we recall that $(g_3)$ is satisfied if
\begin{equation}\label{eq000}
\lim_{|s|\to +\infty}\frac{g(s)}{|s|^{p-2}s}=1, \quad \mbox{for some} \quad p>2.
\end{equation}
From $(g_1)$, there exists $s_o>0$ such that
\begin{equation}\label{eq01}
|g(s)s - \theta G(s)|\leq \frac {(\theta-2)V_o |s|^2}{4\theta}, \quad \mbox { for all }|s|\leq s_o.
\end{equation}
After these remarks, we are able to prove the existence of solution for $(AP)$.
\subsection{The variational approach} In this subsection, related to the function $V$, we assume only condition $(V_2)$. From now on, we set
$$E=\left\{u\in D^{1,2}(\mathbb{R}^N);\,\,\int_{\mathbb{R}^N}(|\nabla u|^2+V(x)|u|^2)\,dx<+\infty\right\}$$
endowed with the norm
$$\|u\|=\left(\int_{\mathbb{R}^N}(|\nabla u|^2+V(x)|u|^2)\,dx\right)^{\frac{1}{2}}$$
and functional $J:E\to\R$ by
$$
J(u)=\frac{1}{2}\|u\|^2 \, dx-\int_{\mathbb{R}^N} G(u)\,dx.
$$
From $(g_1)-(g_3)$, $J$ is well defined in $E$ and $J \in C^{1}(E, \mathbb{R})$ with
$$
J'(u)v=\int_{\mathbb{R}^N}(\nabla u \nabla v + uv)\,dx - \int_{\mathbb{R}^N}g(u)v\,dx, \quad \forall u,v \in E.
$$
Moreover, it is very easy to check that $J$ also satisfies the mountain pass geometry. In what follows, we denote by $c>0$ the mountain pass level associated with $J$, that is,
\begin{equation} \label{MPL}
c=\inf_{\gamma \in \Gamma}J(\gamma(t))>0
\end{equation}
where
$$
\Gamma =\{\gamma\in C([0,1],E):\gamma(0)=0\mbox{ and } J(\gamma(1))< 0\}.
$$
Associated with $c$, we have a Cerami sequence $(v_n) \subset E$, that is,
\begin{equation} \label{C1}
J(v_n) \to c \quad \mbox{and} \quad (1+\|v_n\|)\|J'(v_n)\| \to 0 \quad \mbox{as} \quad n \to +\infty. \,\,\, ( \mbox{See \cite[Theorem 5.46]{MMP}} )
\end{equation}
\begin{proposition} The sequence $(v_n)$ is bounded in $E$.
\end{proposition}
\begin{proof} As $(v_n)$ is a Cerami sequence, we have
$$
J(v_n)=c+o_n(1) \quad \mbox{and} \quad J'(v_n)v_n=o_n(1).
$$
For each $n$, we set
\[
\mathcal{A}_n=\{x\in \mathbb R^N: s_o\leq|v_n(x)|\leq M\},
\]
where $s_o$ was fixed in (\ref{eq01}). From $(g_4)$,
\[
\sigma_o|\mathcal{A}_n|\leq \int_{\mathcal{A}_n}\mathcal{G}(v_n)\,dx\leq J(v_n)-\frac 12 J'(v_n)v_n=c+o_n(1),
\]
where $\sigma_o=\inf\{\mathcal{G}(s): |s|\in[s_o,M]\}>0$ and $|\mathcal{A}_n|$ denotes the Lebesgue's measure of $\mathcal{A}_n$. This inequality implies that $(|\mathcal{A}_n|)$ is a bounded sequence. In the sequel, we consider the sets
$$
\mathcal{C}_n=\{x\in \mathbb R^N: |v_n(x)|< s_o\}
$$
and
$$
\mathcal{B}_n=\{x\in \mathbb R^N: |v_n(x)|> M\}.
$$
From (\ref{eq01}),
\begin{eqnarray*}
\int_{\mathcal{C}_n}[v_ng(v_n)-\theta G(v_n)]dx\geq -\frac {(\theta-2)V_o }{4}\int_{\mathcal{C}_n} |v_n|^2dx\geq-\frac {(\theta-2) }{4}\|v_n\|^2, \quad \forall n \in \mathbb{N}
\end{eqnarray*}
and by $(g_3)$,
\[
\int_{\mathcal{B}_n}[v_ng(v_n)-\theta G(v_n)]dx\geq 0, \quad \forall n \in \mathbb{N}.
\]
On the other hand, $(g_3)$ gives
\begin{eqnarray*}
\int_{\mathcal{A}_n}[v_ng(v_n)-\theta G(v_n)]dx\geq -|\mathcal{A}_n| M_o, \quad \forall n \in \mathbb{N},
\end{eqnarray*}
for some $M_0>0$. The above analysis ensure that
$$
\begin{array}{l}
c+o_n(1)= \displaystyle J(v_n)-\frac 1 \theta J'(v_n)v_n=\frac {(\theta-2) }{2\theta}\|v_n\|^2 +\frac 1\theta \int_{\mathbb R^N}[v_ng(v_n)-\theta G(v_n)]dx \\
\mbox{} \\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \geq \displaystyle \frac {(\theta-2) }{4\theta}\|v_n\|^2-M_o|\mathcal{A}_n|, \quad \forall n \in \mathbb{N}.
\end{array}
$$
Since $(|\mathcal{A}_n|)$ is a bounded sequence, it follows that $(v_n)$ is bounded in $E$.
\end{proof}
\subsection{Existence of solution for $(AP)$: The periodic case}
In this section we assume that $V$ verifies $(V_1)-(V_2)$. By Subsection 2.1, we know that there is a bounded Cerami sequence $(v_n)$ associated with the mountain pass level $c$, that is,
$$
J(v_n) \to c \quad \mbox{and} \quad (1+\|v_n\|)\|J'(v_n)\| \to 0 \quad \mbox{as} \quad n \to +\infty.
$$
From boundedness of $(v_n)$, we deduce that $(v_n)$ is a $(PS)_c$ sequence, that is
$$
J(v_n) \to c \quad \mbox{and} \quad J'(v_n) \to 0 \quad \mbox{as} \quad n \to +\infty.
$$
Moreover, we can assume that for some subsequence, there is $u \in E$ such that
$$
v_n \rightharpoonup v \quad \mbox{in} \quad E,
$$
$$
v_n \to v \quad \mbox{in} \quad L^{t}(B_R(0)), \quad \forall t \in [1,2^{*}) \quad \mbox{and} \quad \forall R>0 ,
$$
and
$$
v_n(x) \to v(x) \quad \mbox{a.e in} \quad \mathbb{R}^N.
$$
Since $J$ is invariant by translation, by Lions \cite{Lions}, we can assume that $v \not= 0$. Hence, $v$ is a nontrivial critical point for $J$, and so, $v$ is a nontrivial solution for $(AP)$. The reader can see more details of how we can use \cite{Lions} in \cite{ACM}, \cite{ZElati}, \cite{ZR} and \cite{Szulkin1}.
\subsection{The existence of solution for $(AP)$ : The Coercive case}
In this case, it is well known that the following compact embedding holds
\begin{equation} \label{compacidade}
E \hookrightarrow L^{t}(\mathbb{R}^N), \quad \forall t \in (2,2^*) \,\, (\,\, \mbox{See \cite{Costa2}} \,\, ).
\end{equation}
This compact embedding together with the boundedness of $(PS)$ sequences permit to prove that $J$ verifies the well known (PS) condition, and so, the mountain pass level $c$ is a positive critical value for $(AP)$. This prove that $(AP)$ has a nontrivial solution.
We would like to point out that the same conclusion holds if $(V_3)$ is replaced by
$$
\mbox{For all} \quad K>0 \quad \mbox{we have that} \quad |\left\{x \in \mathbb{R}^N\,:\, V(x) \leq K \right\}|<\infty, \,\, (\,\, \mbox{See \cite{Bartsch-Wang1} } \,\, ) \leqno{(V_4)}
$$
where $|A|$ denotes the Lebesgue's measure of a mensurable set $A \subset \mathbb{R}^N$. The last condition also implies in the compact embedding (\ref{compacidade}).
Finally, we would like point out that if $V$ is radially symmetric, the compactness (\ref{compacidade}) also holds in $H_{rad}^{1}(\mathbb{R}^N)$, for more details see Willem \cite[Corollary 1.26]{Willem}. The existence of solution follows as in \cite[Section 1.6]{Willem}, where the above compact embedding (\ref{compacidade}) and the Principle of symmetric criticality due to Palais apply an important role in the arguments.
\subsection{Existence of solution for (AP): $V$ is a Barstch \& Wang type potential}
The approach explored in the previous subsection can be also used to study the existence of solution when the potential $V$ is of the form
$$
V(x)=1+\lambda W(x), \quad \forall x \in \mathbb{R}^N, \,\, (\,\, \mbox{See \cite{BW2} } \,\, )
$$
where $\lambda >0$ and $W:\mathbb{R}^N \to \mathbb{R}$ is a nonnegative continuous function that satisfies
$$
\mbox{There is} \quad K>0 \quad \mbox{such that} \quad |\left\{x \in \mathbb{R}^N\,:\, V(x) \leq K \right\}|<+\infty. \leqno{(W_1)}
$$
Arguing as \cite{BW2}, there is $\alpha>0$, which is independent of $\lambda$, such that
\begin{equation} \label{EBW1}
\int_{\mathbb{R}^N}|v_n|^{p}\,dx \geq \alpha, \quad \forall n \in \mathbb{N}, \,\, (\,\, \mbox{See \cite[Lemma 2.4]{BW2} } \,\, ) .
\end{equation}
Moreover, it is possible to prove that there are $R>0$ independent of $\lambda>0$, and $\lambda^*>0$ such that
\begin{equation} \label{EBW2}
\limsup_{n \to +\infty}\int_{B^{c}_R(0)}|v_n|^{p}\,dx \leq \frac{\alpha}{2}, \quad \forall n \in \mathbb{N} \quad \mbox{and} \quad \lambda \geq \lambda^*. \,\, (\,\, \mbox{See \cite[Lemma 2.5]{BW2} } \,\, )
\end{equation}
From (\ref{EBW1})-(\ref{EBW2}),
$$
\int_{B_R(0)}|v|^{p}\,dx \geq \frac{\alpha}{2}>0, \quad \forall \lambda \geq \lambda^*,
$$
showing that $v$ is nontrivial. Hence $(AP)$ has a nontrivial solution for $\lambda$ large enough.
\vskip .5 cm
Before concluding this section, the reader is invited to see that the results showed in this section prove the Theorem \ref{T0}.
\setcounter{equation}{0}\Section{Proof of Theorem \ref{T1}}
In this section we deal with the proof of Theorem \ref{T1}, and the results obtained in Section 2 will be crucial in our approach.
In what follows, we consider the auxiliary problem
$$
\left\{
\begin{array}{l}
-\Delta u+ V(x)u= f(u) +\varepsilon|u|^{q-2}u \,\,\,\,\mbox{in }\,\,\,\,\mathbb{R}^N, \\
u \in E.
\end{array}
\right.
\leqno{(AP)_\varepsilon}
$$
where $\displaystyle{q=\frac {2N+4}N}$ and $\varepsilon >0$.
It is easy to check that $g_\varepsilon(s)=f(s) +\varepsilon|s|^{q-2}u$ satisfies $(g_1)-(g_4)$ for $\theta=q$. Hence, from the previous sections, for fixed $\varepsilon>0$, there is a nontrivial solution $u_\varepsilon$ of $(AP)_\varepsilon$ such that
$$
J_\varepsilon(u_\varepsilon)=c_\varepsilon \quad \mbox{and} \quad J'_\varepsilon(u_\varepsilon)=0
$$
where
$$
J_\varepsilon(u)=\frac{1}{2}\|u\|^2 \, dx-\int_{\mathbb{R}^N} G_\varepsilon(u)\,dx, \quad \forall u \in E,
$$
and $c_\varepsilon$ is the mountain pass level associated with $J_\varepsilon$, that is,
$$
0<c_\varepsilon=\inf_{\gamma \in \Gamma}J_\varepsilon(\gamma(t)).
$$
Since $J_\varepsilon(v)\leq J_0(v)$ for all $v\in E$, we have $c_\varepsilon\leq c_0$ for all $ \varepsilon >0$. Furthermore, it is possible to prove that there is $\alpha>0$, which is independent of $\varepsilon \in [0,1]$, such that
\begin{equation} \label{limitab}
c_\varepsilon \geq \alpha, \quad \forall \varepsilon \in [0,1].
\end{equation}
As an immediate consequence of the above inequality, we have
\begin{equation} \label{limitab*}
\|u_\varepsilon \| \not\to 0 \quad \mbox{as} \quad \varepsilon \to 0.
\end{equation}
The next result establishes an important estimate for $\{u_\varepsilon\}$ in $E$, which is a key point in our approach.
\begin{proposition} \label{LIMITACAO} There is $\varepsilon_0>0$ such that $\{u_\varepsilon\}$ is bounded in $E$ for all $\varepsilon \in [0,\varepsilon_0]$.
\end{proposition}
\begin{proof}
For each $\varepsilon >0$, we have $\mathcal{G}_\varepsilon \geq \mathcal{F}$. Thus, by $(f_8)$,
$$
\sigma\int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^2dx\leq \int_{\mathbb R^N}\mathcal{G}_\varepsilon(u_\varepsilon)\,dx= 2J_\epsilon(u_\varepsilon)- J_\varepsilon'(u_\varepsilon)u_\varepsilon= 2c_\varepsilon \leq 2c_0.
$$
Arguing as (\ref{eq01}), there exists $s_o>0$, which is independent of $\varepsilon >0$ small, such that
\begin{equation}\label{eq02}
G_\varepsilon(s)\leq \frac {V_o |s|^2}{4}\mbox { for all }|s|\leq s_o.
\end{equation}
On the other hand, from $(f_7)$, for each $\varepsilon>0$, there is $C_\varepsilon >0$ verifying
\begin{equation}\label{eq03}
G_\varepsilon(s)\leq \varepsilon |s|^q+2C_\varepsilon |s|^2\mbox { for all }|s|\geq s_o.
\end{equation}
So, setting
$$
\mathcal{C}_\varepsilon=\{x\in \mathbb R^N: |u_\varepsilon(x)|< s_o\},
$$
and using (\ref{eq02}) and (\ref{eq03}), we get
$$
\begin{array}{l}
\frac 12\|u_\varepsilon\|^2=c_\varepsilon+ \displaystyle \int_{\mathbb R^N}G_\varepsilon (u_\varepsilon)dx \leq c_\varepsilon+\int_{\mathcal{C}_\varepsilon}G_\varepsilon(u_\varepsilon)dx +\varepsilon\int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^qdx+ 2C_\varepsilon\int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^2dx\\
\mbox{}\\
\hspace{5 cm} \leq c_0+\frac{C_\varepsilon c_0}{\sigma}+\frac 14 \displaystyle \int_{\mathcal{C}_\varepsilon}V(x)|u_\varepsilon|^2dx +\varepsilon\int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^qdx,
\end{array}
$$
that is,
\begin{eqnarray*}
\frac 14\|u_\varepsilon\|^2\leq \frac{(C_\varepsilon+\sigma)c_0}{\sigma}+\varepsilon \int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^qdx.
\end{eqnarray*}
Recalling that
\begin{eqnarray*}
\int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^qdx\leq \left( \int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^{2^*}dx\right)^\frac{N-2}{N}\left( \int_{\{|u_\varepsilon|\geq s_o\}}|u_\varepsilon|^{\frac{N(q-2)}{2}}dx\right)^\frac{2}{N}\\\leq \left( \frac{c_0}{\sigma}\right)^\frac{2}{N}\|u_\varepsilon\|_{2^*}^2\leq \left( \frac{c_0}{\sigma}\right)^\frac{2}{N}S^{-1}\|u_\varepsilon\|^2,
\end{eqnarray*}
and fixing $\varepsilon_0=\frac{S}{8}(\frac{\sigma}{c_0})^{\frac{2}{N}}$, we derive that
\[
\frac 18\|u_\varepsilon\|^2\leq \frac{(C_\varepsilon +\sigma)c_0}{\sigma}, \quad \forall \varepsilon \in (0,\varepsilon_0].
\]
This proves the boundedness of $\{u_\varepsilon\}$ for $\varepsilon \in (0,\varepsilon_0]$ in $E$.
\end{proof}
In the sequel, for each $n \in \mathbb{N}$, we denote by $u_n$ the solution of $(AP)_{\frac{1}{n}}$, that is,
$$
-\Delta u_n+ V(x)u_n= f(u_n) +\frac{1}{n}|u_n|^{q-2}u_n \,\,\,\,\mbox{in }\,\,\,\,\mathbb{R}^N. \leqno{(AP)_{\frac{1}{n}}}
$$
By Proposition \ref{LIMITACAO}, the sequence $(u_n)$ is bounded in $E$, hence for some subsequence, there is $u \in E$ such that
$$
u_n \rightharpoonup u \quad \mbox{in} \quad E,
$$
$$
u_n \to u \quad \mbox{in} \quad L^{t}(B_R(0)), \quad \forall t \in [1,2^{*}) \quad \mbox{and} \quad \forall R>0,
$$
and
$$
u_n(x) \to u(x) \quad \mbox{a.e in} \quad \mathbb{R}^N.
$$
From this, for each $\phi \in E$,
$$
\int_{\mathbb{R}^N}\nabla u_n \nabla v \,dx + \int_{\mathbb{R}^N}V(x)u_n\phi\, dx= \int_{\mathbb{R}^N}f(u_n)\phi \, dx+ \frac{1}{n}\int_{\mathbb{R}^N}|u_n|^{q-2}u_n\phi\,dx, \quad \forall n \in \mathbb{N}.
$$
Then taking the limit of $n \to +\infty$, we find
$$
\int_{\mathbb{R}^N}\nabla u \nabla v \,dx + \int_{\mathbb{R}^N}V(x)u \phi\, dx= \int_{\mathbb{R}^N}f(u)\phi\,dx, \quad \forall \phi \in E,
$$
showing that $u$ is a solution of $(P_1)$. Our goal is proving that $u$ is a nontrivial solution. Have this in mind, firstly we prove the result below
\begin{lemma} \label{naopode} $ u_n \not\to 0$ in $L^{p}(\mathbb{R}^N)$, where $p$ was given in $(f_7)$.
\end{lemma}
\begin{proof} Since $u_n$ is a solution of $(AP)_{\frac{1}{n}}$, it follows that
$$
\|u_n\|^{2}=\int_{\mathbb{R}^N}f(u_n)u_n\,dx+\frac{1}{n}\int_{\mathbb{R}^N}|u_n|^{q}\,dx, \quad \forall n \in \mathbb{N}.
$$
By $(f_1)$ and $(f_7)$,
$$
|f(t)t| \leq \frac{V_o}{2}|t|^{2}+C|t|^{p}, \quad \forall t \in \mathbb{R},
$$
for some positive constant $C$. Consequently,
$$
\|u_n\|^{2} \leq C \int_{\mathbb{R}^N}|u_n|^{p}\,dx+\frac{1}{n}\int_{\mathbb{R}^N}|u_n|^{q}\,dx, \quad \forall n \in \mathbb{N},
$$
for some positive constant $C$. Using the fact that $(u_n)$ is bounded in $E$, we have that $(u_n)$ is also bounded in $L^{q}(\mathbb{R}^N)$. Thus, supposing by contradiction that $u_n
\to 0$ in $L^{p}(\mathbb{R}^N)$, we obtain
$$
u_n \to 0 \quad \mbox{in} \quad E,
$$
which contradicts (\ref{limitab*}).
\end{proof}
\noindent {\bf Conclusion of the proof of Theorem \ref{T1}} \\
Now, we can argue as in Subsections 2.2, 2.3 and 2.4 to deduce that $u \not= 0$. For example, in the periodic case, the Lions' result together with Lemma \ref{naopode} ensures that we can assume that $u \not= 0$. For the others cases, we argue exactly as in Section 2.
\setcounter{equation}{0}\Section{Proof of Theorem \ref{T2}}
In what follows, we consider the function $h_\Upsilon:\mathbb{R} \to \mathbb{R}$ given by
$$
h_\Upsilon(t)=
\left\{
\begin{array}{l}
h(t), \quad t \in [0,\Upsilon], \\
\mbox{}\\
\frac{h(\Upsilon)}{f(\Upsilon)}f(t), \quad t \geq \Upsilon,
\end{array}
\right.
$$
and $f_\Upsilon:\mathbb{R} \to \mathbb{R}$ by
$$
f_\Upsilon(t)=
\left\{
\begin{array}{l}
0, \quad t \in (-\infty,0], \\
f(t)+\beta h_\Upsilon(t), \quad t \geq 0.
\end{array}
\right.
$$
Using the above notations, our intention is proving the existence of a nontrivial solution $u_{\beta}$ with $\|u_{\beta}\|_{\infty}<\Upsilon$ for the following auxiliary problem
$$
\left\{
\begin{array}{l}
-\Delta{u}+V(x)u=f_\Upsilon(u), \quad \mbox{in} \quad \mathbb{R}^N, \\
u \in E.
\end{array}
\right.
\leqno{(AP)_{\beta,\Upsilon}}
$$
Hereafter, we fix $\beta^*=\frac{f(\Upsilon)}{h(\Upsilon)}$ and $\beta \in [0, \beta^*]$. A simple computation shows that $f_\Upsilon$ satisfies $(f_1),(f_7)-(f_8)$, more precisely,
\begin{itemize}
\item[$(i)$] $ \displaystyle{\lim_{s\to 0}\frac{f_\Upsilon(s)}{s}=0}$ ;
\item[$(ii)$] There are $C>0$ and $p\in\left(2,\frac{2N+4}{N}\right)$ such that $|f_\Upsilon(s)|\leq C (1+ |s|^{p-1})$ for all $s\in \mathbb R$;
\item[$(iii)$] For each $s_o>0$, there is $\sigma>0$ such that $$\mathcal{F}_\Upsilon(s)=sf_\Upsilon(s)-2 F_\Upsilon(s)\geq \sigma s^2,\mbox { for all } s\geq s_o,$$
where $F_\Upsilon(s)=\int_{0}^{s}f_\Upsilon(t)\,dt$.
\end{itemize}
Indeed, we will check only $(ii)-(iii)$, because $(i)$ is immediate. By $(h_4)$,
$$
h_\Upsilon(s) \leq \frac{h(\Upsilon)}{f(\Upsilon)}f(s), \quad \forall s \geq 0,
$$
leading to
\begin{equation} \label{NOVA}
H_\Upsilon(s)=\int_{0}^{s}h_\Upsilon(t)\,dt \leq \frac{h(\Upsilon)}{f(\Upsilon)}F(s), \quad \forall s \geq 0,
\end{equation}
and
$$
|f_\Upsilon(s)| \leq 2|f(s)|\leq C(1+|s|^{p-1}), \quad \mbox{for} \quad p\in\left(2,\frac{2N+4}{N}\right).
$$
On the other hand, if $s \in [0,\Upsilon]$, we have
\begin{equation} \label{NOVA1}
\mathcal{F}_\Upsilon(s)= sf(s)-2 F(s)+\beta(sh(s)-2 H(s)).
\end{equation}
Now, if $s \geq \Upsilon$,
$$
\mathcal{F}_\Upsilon(s)=sf(s)-2 F(s)+\beta\frac{h(\Upsilon)}{f(\Upsilon)}(sf(s)-2F(s))+H(\Upsilon)-\frac{h(\Upsilon)}{f(\Upsilon)}F(\Upsilon)
$$
and so, by (\ref{NOVA}),
$$
\mathcal{F}_\Upsilon(s) \geq \left(1+ \beta\frac{h(\Upsilon)}{f(\Upsilon)}\right)(sf(s)-2 F(s)) .
$$
As a consequence of the above analysis,
$$
\mathcal{F}_\Upsilon(s) \geq \sigma s^{2} \quad \mbox{for} \quad s \geq s_o.
$$
From this, we can apply Theorem \ref{T1} to get a solution $u_\beta \in E$ of $(AP)_{\beta,\Upsilon}$. Our next step is proving that there is $\Upsilon^*>0$ such that
$ \|u_\beta\|_\infty \leq \Upsilon^*$ for $\Upsilon \geq \Upsilon^*$, because this estimate permits to conclude that $u_\beta$ is a solution of $(P_2)$ for $\Upsilon \geq \Upsilon^*$, which shows Theorem \ref{T2}. However, the existence of $\Upsilon^*$ follows from the following fact: If $I_\Upsilon:E \to \mathbb{R}$ denotes the energy functional associated with $(AP)_{\beta,\Upsilon}$, that is,
$$
I_\Upsilon(u)=\frac{1}{2}\|u\|^2 \, dx-\int_{\mathbb{R}^N} F_\Upsilon(u)\,dx, \quad \forall u \in E,
$$
we have
$$
I_\Upsilon(u_\beta) \leq J_0(u), \quad \forall u \in E \quad \mbox{and} \quad \beta \geq 0.
$$
Hence,
$$
I_\Upsilon(u_\beta) \leq c_0, \quad \forall \Upsilon \geq 0,
$$
where $c_0$ denotes the mountain pass level associated with the functional $J_0$ ( see Section 3). Since $I'_\Upsilon(u_\beta)=0$, we can argue as in Section 3 to show that there is $k>0$, which is independent of $\Upsilon$ and $\beta$, such that
$$
\|u_\beta\| \leq k, \quad \forall \Upsilon >0 \quad \mbox{and} \quad \forall \beta \in [0,\beta^*].
$$
Recalling that by $(ii)$,
$$
|f_\Upsilon(s)|\leq C (1+ |s|^{p-1}) \quad \forall s\in \mathbb R,
$$
where the constant $C$ does not depend on $\beta$ and $\Upsilon$, the bootstrap argument found \cite[Proposition 2.15]{Rab} ensures that there is $K_*>0$, which is independent of $\beta$ and $\Upsilon$ such that
$$
\|u_\beta\|_\infty \leq K, \quad \forall \Upsilon >0 \quad \mbox{and} \quad \beta \in [0,\beta^*].
$$
From this, if $\Upsilon \geq \Upsilon^*=K$, it follows that $u_\beta$ is a solution for $(P_2)$, finishing the proof of Theorem \ref{T2}.
\setcounter{equation}{0}\Section{Final comments}
In this section we would like point out that if $\displaystyle{\frac{h(s)}{f(s)} \to +\infty}$ as $s \to +\infty$, we can replace $(h_3)$ by
\begin{itemize}
\item[$(h'_3)$] For each $s_o>0$, there is a $\sigma_o>0$ such that $$\mathcal{H}(s)=sh(s)-2 H(s)\geq -\sigma_o s^2,\mbox { for all } |s|\geq s_o.$$
\end{itemize}
Indeed, using this assumption in (\ref{NOVA1}), we get
$$
\mathcal{F}_\Upsilon(s) \geq (\sigma -\beta \sigma_o)s^2, \quad \forall s \in [0,\Upsilon].
$$
If $\displaystyle{\frac{h(s)}{f(s)} \to +\infty}$ as $s \to +\infty$, we derive that $\beta^* \to 0$ as $\Upsilon^* \to +\infty$. Hence,
$$
(\sigma -\beta \sigma_o)>0, \quad \forall \beta \in (0, \beta^*),
$$
for $\Upsilon^*$ large enough. From this, Theorem \ref{T2} still holds for $\Upsilon^*$ large enough.
|
\section{Introduction}
In recent years, ferrimagnetic systems, like yttrium iron garnet (YIG), become an active and important platform for the study of strong interaction between light and matter, owing to their high spin density and low damping rate. Magnons, as collective excitations of a large number of spins, can strongly couple to cavity microwave photons~\cite{Strong1,Strong2,Strong3,Strong4,Strong5,Strong6,Tobar2,Tobar3} leading to cavity-magnon polaritons. The strong coherent interaction allows one to observe many interesting phenomena in cavity-magnon systems, such as the exceptional point~\cite{You17NC}, remote manipulation of spin currents~\cite{Bai}, bistability~\cite{You18PRL}, etc.
An intriguing direction would be hybrid systems based on magnonics~\cite{NakaRev}. The coupling of magnons with a variety of different systems, either continuous variable or discrete variable, provides great opportunities for the study of many interesting topics: e.g., coupling magnons to a superconducting qubit~\cite{NakaSci15} allows one to resolve magnon number states~\cite{Naka17SA}, coupling magnons to both optical and microwave photons~\cite{Tang16PRL} allows one to coherently convert optical and microwave photons~\cite{Naka16PRB}, and coupling magnons to the vibrational mode of a YIG sphere~\cite{Tang16SA} allows one to prepare magnon-photon-phonon entangled states~\cite{Jie18PRL} and magnon/phonon squeezed states~\cite{Jie19PRA}. In the last case, magnons and phonons are coupled via nonlinear magnetostrictive interaction, which is of radiation pressure type, and therefore many known results in the well-developed field of cavity optomechanics~\cite{omRMP} are expected to occur in the new field of cavity magnomechanics, such as magnomechanically induced transparency~\cite{Tang16SA}, and magnomechanical cooling and entanglement~\cite{Jie18PRL}. Many other interesting phenomena have been explored in cavity magnomechanics, such as slow light effect~\cite{Wu19}, phonon lasing~\cite{Li19}, and parity-time-related phenomena~\cite{Liu19,Sun19}.
In this paper, we focus on quantum effect in cavity-magnon systems and present a scheme to entangle two magnon modes in two macroscopic YIG spheres. Recently, several proposals have been put forward, using different mechanisms, for preparing entangled states of two magnon modes, either in ferrimagnetic YIG spheres~\cite{Jie19,Zhedong,Jaya} or in an antiferromagnetic system~\cite{Yung}. The two spheres are placed inside two microwave cavities, and in each cavity a magnon mode couples to a cavity mode via linear beamsplitter interaction. It is well known that such interaction will not generate any entanglement. One approach is to introduce necessary nonlinearity into the system, either from the magnetostrictive interaction~\cite{Jie19} or from the Kerr effect~\cite{Zhedong}. Another approach is to inject external quantum resource, e.g., squeezed vacuum field~\cite{Jaya}, into the system. The present scheme follows the latter approach: we drive the two cavities with a two-mode squeezed vacuum microwave field. The idea is to transfer the quantum correlation shared by the two driving fields to the two magnon modes via the linear cavity-magnon coupling. We show that this quantum state transfer occurs with high efficiency provided that the cavity and magnon modes are resonant with the driving fields, and the coupling rate $g$ and the cavity (magnon) decay rate $\kappa_a$ ($\kappa_m$) satisfy $g\gg \kappa_a \gg \kappa_m$, which has been realized, e.g., in the experiments~\cite{Strong2,Strong3,Strong6}. The two magnon modes are entangled in the steady state, and the entanglement increases with the squeezing of the input two-mode squeezed field and survives up to hundreds of milliKelvin with experimentally accessible two-mode squeezed source.
The remainder of the paper is constructed as follows. In Sec.~\ref{Model}, we introduce the system, provide its Hamiltonian and the corresponding quantum Langevin equations (QLEs), and show in detail how to solve the QLEs and calculate the entanglement. In Sec.~\ref{Numeri}, we provide numerical results and show optimal parameter regimes where large magnon entanglement can be obtained, and in Sec.~\ref{Analy}, we provide analytical solutions at the optimal conditions and analyse in more depth the topic. Finally, we draw our conclusions in Sec.~\ref{Conc}.
\begin{figure}[t]\label{fig1}
\hskip-2.0cm\includegraphics[width=0.4\linewidth]{fig1.pdf}
\caption{Sketch of the system. A double cavity-magnon system where the two cavities are driven by a two-mode squeezed vacuum microwave field. Two YIG spheres are respectively placed inside the cavities near the maximum magnetic fields of the cavity modes and simultaneously in uniform bias magnetic fields. The bias magnetic fields and the magnetic fields of the cavity modes are along the $z$ and $x$ directions, respectively.}
\end{figure}
\section{The Model}\label{Model}
The system under study, sketched in Fig. 1, consists of two microwave cavities and two magnon modes in two YIG spheres, which are respectively placed inside the cavities near the maximum magnetic fields of the cavity modes and simultaneously in uniform bias magnetic fields, which excite the magnon modes in the spheres and couple them to the cavity modes. The magnons are quasiparticles and are embodied by a collective excitation of a large number of spins in YIG spheres. The system of two YIG spheres has been used to study magnon dark modes~\cite{Tang15NC}, high-order exceptional points~\cite{You19PRB}, and entanglement properties between two magnon modes~\cite{Jie19,Zhedong,Jaya}. In each cavity, the magnon mode couples to the cavity mode via the magnetic dipole interaction, and this coupling can be very strong~\cite{Strong1,Strong2,Strong3,Strong4,Strong5,Strong6,Tobar2,Tobar3} thanks to the high spin density of YIG. The two cavities are driven by a two-mode squeezed vacuum microwave field. We consider the size of the YIG spheres is much smaller than the microwave wavelengths, and hence the effect of radiation pressure can be safely neglected. The Hamiltonian of the system reads
\begin{equation}
{\cal H}/\hbar =\underset{j=1,2}{\sum }\Big\{ \omega _{a_{j}}a_{j}^{\dag }a_{j}+\omega_{m_{j}}m_{j}^{\dag }m_{j}+g_{j} \big( a_{j}m_{j}^{\dag }+a_{j}^{\dag }m_{j} \big) \Big\},
\label{Hamilt}
\end{equation}
where $a_{j}$ ($a_{j}^{\dag }$) and $m_{j}$ ($m_{j}^{\dag }$) are annihilation (creation) operators of the $j$th cavity and magnon modes, respectively, and we have $\big[ O,O^{\dag }\big] =1$ ($O\,{=}\,a_j, m_j$). $\omega _{a_{j}}$ ($\omega _{m_{j}}$) is the resonance frequency of the $j$th cavity mode (magnon mode). The frequency of the magnon mode $\omega _{m_{j}}$ is determined by the external bias magnetic field $H_j$ and the gyromagnetic ratio $\gamma$ via $\omega _{m_{j}}=\gamma H_j$, and thus can be flexibly adjusted, and $g_{j}$ is the coupling rate between the $j$th\ cavity and magnon modes.
In the frame rotating at the frequency $\omega _{_{j}}$, i.e., the frequency of the $j$th mode of the input two-mode squeezed field, the QLEs of the system are given by
\begin{eqnarray} \label{QLEs}
\overset{\cdot }{a}_{j} &=&-(\kappa _{a_{j}}+i\Delta_{a_{j}})a_{j}-ig_{j}m_{j}+\sqrt{2\kappa _{a_{j}}}S_{j}^{in}, \\
\overset{\cdot }{m}_{j} &=&-(\kappa _{m_{j}}+i\Delta_{m_{j}})m_{j}-ig_{j}a_{j}+\sqrt{2\kappa _{m_{j}}}m_{j}^{in}, \notag
\end{eqnarray}
where $\kappa _{a_{j}}$ ($\kappa _{m_{j}}$) is the decay rate of the $j$th cavity mode (magnon mode), $\Delta _{a_{j}}=\omega_{a_{j}}-\omega _{_{j}},\Delta _{m_{j}}=\omega _{m_{j}}-\omega _{_{j}}$, and $S_{j}^{in}$ ($m_{j}^{in}$) is the input noise operator for the $j$th cavity mode (magnon mode). The two cavity input noise operators $S_j^{in}$ are quantum correlated due to the injection of the two-mode squeezed field, and have the following correlations in time domain
\begin{eqnarray}
\left\langle S_{j}^{in\dag}(t)S_{j}^{in}(t^{\prime }) ; S_{j}^{in}(t)S_{j}^{in\dag }(t^{\prime }) \right\rangle &=&( N ; N+1) \delta(t-t^{\prime}),\,\,\,\, (j=1,2) \label{NoiseCF1} \\
\left\langle S_{j}^{in}(t)S_{k}^{in}(t^{\prime }) ; S_{j}^{in\dag}(t)S_{k}^{in\dag }(t^{\prime }) \right\rangle &=&( M;M^{\ast })
\delta (t-t^{\prime }),\,\,\,\, (j\neq k=1,2) \notag
\end{eqnarray}%
with $N=\sinh ^{2}r$, $M=e^{i\theta }\sinh r\cosh r$, where $r$ and $\theta$ are the squeezing parameter and phase of the two-mode squeezed vacuum field, which could be generated by a Josephson parametric amplifier (JPA)~\cite{sqzMW1}, by a Josephson mixer~\cite{sqzMW2}, or by the combination of a JPA and a microwave beamsplitter~\cite{sqzMW3,sqzMW4}. The magnon input noise operators $m_j^{in}$ are zero mean and correlated as follows
\begin{equation}
\left\langle m_{j}^{in\dag}(t)m_{j}^{in}(t^{\prime }) ; m_{j}^{in}(t)m_{j}^{in\dag }(t^{\prime }) \right\rangle =(N_{m_{j}}; N_{m_{j}}+1)\delta(t-t^{\prime }), \,\,\,\, (j=1,2)
\end{equation}
where $N_{m_{j}}=[\exp (\frac{\hbar \omega _{m_{j}}}{k_{B}T})-1]^{-1}$ is the equilibrium mean thermal magnon number of the $j$th mode, with $T$ the environmental temperature and $k_B$ the Boltzmann constant.
Since we are interested in the quantum correlation properties of the two magnon modes, we focus on the dynamic of the quantum fluctuations of the system. The fluctuations of the system are described by the following QLEs
\begin{eqnarray}
\delta \overset{\cdot }{a}_{j} &=&-(\kappa _{a_{j}}+i\Delta _{a_{j}})\delta a_{j}-ig_{j}\delta m_{j}+\sqrt{2\kappa _{a_{j}}}S_{j}^{in}, \label{QLEs2} \\
\delta \overset{\cdot }{m}_{j} &=&-(\kappa _{m_{j}}+i\Delta _{m_{j}})\delta m_{j}-ig_{j}\delta a_{j}+\sqrt{2\kappa _{m_{j}}}m_{j}^{in}. \notag
\end{eqnarray}
The above QLEs can be written in the quadrature form, with quadrature fluctuations defined as $\delta X_{j}=(\delta a_{j}+\delta a_{j}^{\dag })/\sqrt{2},\delta Y_{j}=i(\delta a_{j}^{\dag }-\delta a_{j})/\sqrt{2},\delta x_{j}=(\delta m_{j}+\delta m_{j}^{\dag })/\sqrt{2},\delta y_{j}=i(\delta m_{j}^{\dag }-\delta m_{j})/\sqrt{2}$ (similar definition for input noises $X_{j}^{in}, Y_{j}^{in}$ and $x_{j}^{in}, y_{j}^{in}$), which are
\begin{eqnarray}
\delta \overset{\cdot }{X}_{j} &=&-\kappa _{a_{j}}\delta X_{j}+\Delta
_{a_{j}}\delta Y_{j}+g_{j}\delta y_{j}+\sqrt{2\kappa _{a_{j}}}X_{j}^{in},
\label{QELs3} \\
\delta \overset{\cdot }{Y}_{j} &=&-\kappa _{a_{j}}\delta Y_{j}-\Delta
_{a_{j}}\delta X_{j}-g_{j}\delta x_{j}+\sqrt{2\kappa _{a_{j}}}Y_{j}^{in},
\notag \\
\delta \overset{\cdot }{x}_{j} &=&-\kappa _{m_{j}}\delta x_{j}+\Delta
_{m_{j}}\delta y_{j}+g_{j}\delta Y_{j}+\sqrt{2\kappa _{m_{j}}}x_{j}^{in},
\notag \\
\delta \overset{\cdot }{y}_{j} &=&-\kappa _{m_{j}}\delta y_{j}-\Delta
_{m_{j}}\delta x_{j}-g_{j}\delta X_{j}+\sqrt{2\kappa _{m_{j}}}y_{j}^{in}.
\notag
\end{eqnarray}%
They can be cast in the matrix form
\begin{equation}
\overset{\cdot }{u}(t)=Au(t)+n(t), \label{MatrixForm}
\end{equation}
where $u(t)=[\delta X_{1},\delta Y_{1},\delta X_{2},\delta Y_{2},\delta x_{1,}\delta y_{1},\delta x_{2},\delta y_{2}]^{T}$, $A$\ is the drift matrix
\begin{equation}
A=\left(
\begin{array}{cccccccc}
-\kappa _{a_{1}} & \Delta _{a_{1}} & 0 & 0 & 0 & g_{1} & 0 & 0 \\
-\Delta _{a_{1}} & -\kappa _{a_{1}} & 0 & 0 & -g_{1} & 0 & 0 & 0 \\
0 & 0 & -\kappa _{a_{2}} & \Delta _{a_{2}} & 0 & 0 & 0 & g_{2} \\
0 & 0 & -\Delta _{a_{2}} & -\kappa _{a_{2}} & 0 & 0 & -g_{2} & 0 \\
0 & g_{1} & 0 & 0 & -\kappa _{m_{1}} & \Delta _{m_{1}} & 0 & 0 \\
-g_{1} & 0 & 0 & 0 & -\Delta _{m_{1}} & -\kappa _{m_{1}} & 0 & 0 \\
0 & 0 & 0 & g_{2} & 0 & 0 & -\kappa _{m_{2}} & \Delta _{m_{2}} \\
0 & 0 & -g_{2} & 0 & 0 & 0 & -\Delta _{m_{2}} & -\kappa _{m_{2}}
\end{array}
\right) , \label{driftmatrix}
\end{equation}
and $n(t)=[\sqrt{2\kappa _{a_{1}}}X_{1}^{in},\sqrt{2\kappa _{a_{1}}}Y_{1}^{in}, \sqrt{2\kappa _{a_{2}}}X_{2}^{in},\sqrt{2\kappa _{a_{2}}}Y_{2}^{in},\sqrt{2\kappa _{m_{1}}}x_{1}^{in},\sqrt{2\kappa _{m_{1}}}y_{1}^{in},\sqrt{2\kappa_{m_{2}}}x_{2}^{in},\sqrt{2\kappa _{m_{2}}}y_{2}^{in}]^{T}$. Since the dynamics of the system is linear and the input noises are Gaussian, the dynamical map of the system preserves the Gaussian nature of any input state. The steady state of quantum fluctuations of the system is therefore a continuous-variable four-mode Gaussian state, which is completely characterized by an $8\times 8$ covariance matrix (CM) $V$, defined as $V_{ij}(t)=\langle u_{i}(t)u_{j}(t^{\prime })+u_{j}(t^{\prime })u_{i}(t) \rangle/2$. When the system is stable, $t \to \infty $, the solution of $V$ can be obtained by directly solving the Lyapunov equation~\cite{DV07,Hahn}
\begin{equation}
AV+VA=-D, \label{LyapEq}
\end{equation}
where $D$ is the diffuse matrix defined by $D_{ij}\delta (t-t^{\prime})=\langle n_{i}(t) n_{j}(t^{\prime })+n_{j}(t^{\prime })n_{i}(t) \rangle/2$. It can be written in the form of direct sum $D=D_{a}\oplus D_{m}$, where $D_{a}$ is related to the cavity modes, given by
\begin{equation}
D_{a}=\left(
\begin{array}{cccc}
\kappa _{a_{1}}(2N+1) & 0 & \sqrt{\kappa _{a_{1}}\kappa _{a_{2}}}(M+M^{\ast
}) & i\sqrt{\kappa _{a_{1}}\kappa _{a_{2}}}(M^{\ast }-M) \\
0 & \kappa _{a_{1}}(2N+1) & i\sqrt{\kappa _{a_{1}}\kappa _{a_{2}}}(M^{\ast
}-M) & -\sqrt{\kappa _{a_{1}}\kappa _{a_{2}}}(M+M^{\ast }) \\
\sqrt{\kappa _{a_{1}}\kappa _{a_{2}}}(M+M^{\ast }) & i\sqrt{\kappa
_{a_{1}}\kappa _{a_{2}}}(M^{\ast }-M) & \kappa _{a_{2}}(2N+1) & 0 \\
i\sqrt{\kappa _{a_{1}}\kappa _{a_{2}}}(M^{\ast }-M) & -\sqrt{\kappa
_{a_{1}}\kappa _{a_{2}}}(M+M^{\ast }) & 0 & \kappa _{a_{2}}(2N+1)
\end{array}
\right) ,
\end{equation}
and $D_{m}$ is associated with the magnon modes, i.e., $D_{m}={\rm diag} \big[ \kappa _{m_{1}}(2N_{m_{1}}+1),\kappa_{m_{1}}(2N_{m_{1}}+1),\kappa_{m_{2}}(2N_{m_{2}}+1),\kappa_{m_{2}}(2N_{m_{2}}+1) \big]$.
Once the CM of the system is achieved, one can then calculate the degree of entanglement between the two magnon modes. We adopt the logarithmic negativity~\cite{LogNeg,GAJPA} to quantify the entanglement, which is defined as
\begin{equation}
E_N \equiv \max \big\{ 0, \, -\ln2\tilde\nu_- \big\},
\end{equation}
where $\tilde\nu_-\,\,{=}\,\min{\rm eig}|i\Omega_2\tilde{V}_{mm}|$ (with the symplectic matrix $\Omega_2=\oplus^2_{j=1} \! i\sigma_y$ and the $y$-Pauli matrix $\sigma_y$) is the minimum symplectic eigenvalue of the CM $\tilde{V}_{mm}={\cal P}_{1|2}{V_{mm}}{\cal P}_{1|2}$, where $V_{mm}$ is the $4\times 4$ CM of the two magnon modes, obtained by removing in $V$ the rows and columns of the two cavity modes, and ${\cal P}_{1|2}={\rm diag}(1,-1,1,1)$ is the matrix that implements partial transposition at the level of CMs~\cite{Simon}.
\begin{figure}[t]
\hskip-0.07cm\includegraphics[width=0.98\linewidth]{fig2.pdf}
\caption{Density plot of the entanglement $E_{mm}$ between two magnon modes versus (a) $\Delta _{a_1}$ and $\Delta _{m_1}$,
with $\Delta _{a_2}=\Delta _{m_2}=0$, $r=1$, and $T=100$ mK; (b) $\Delta _{a_2}$ and $\Delta _{m_2}$, with $\Delta _{a_1}=\Delta _{m_1}=0$, $r=1$, and $T=100$ mK; (c) $r$ and $T$, with $\Delta _{a_{1,2}}=\Delta _{m_{1,2}}=0$. See text for the other parameters.}
\end{figure}
\section{Numerical results of magnon entanglement}
\label{Numeri}
In Fig. 2, we show the entanglement of two magnon modes versus various system parameters. Figure 2(a) and (b) show that the optimal situation for magnon entanglement is that in each cavity the cavity and magnon modes are resonant with the driving field, i.e., $\Delta _{a_j}=\Delta _{m_j}=0$ ($j=1,2$). This is consistent with the results in the study of transferring single-mode squeezing from microwave field to the magnon mode~\cite{Jie19PRA}. Physically this is easy to understand: owing to the linear cavity-magnon coupling, the resonant situation most efficiently transfers the quantum correlation from the input fields to the two magnon modes. Figure 2(c) shows that the entanglement increases with the squeezing of the input two-mode squeezed field and decreases with the temperature. Note that we have assumed the bandwidths of the input squeezed fields are larger than the cavity linewidths. In Ref.~\cite{sqzMW2}, a two-mode squeezed field with logarithmic negativity $E_N = 0.8$ (corresponding to squeezing $r=0.4$)~\cite{Note} and bandwidth of 12.5 MHz has been produced. With this we can achieve magnon entanglement $E_{mm}= 0.6$ at $T=100$ mK, and the entanglement survives up to $\sim$0.8 K. We have employed in Fig. 2 experimentally feasible parameters \cite{Strong2}: $\omega_{a_1}/2\pi =10$ GHz, $\kappa_a/2\pi =5$ MHz, $\kappa_{m}=\kappa _a/5$, $g_1=g_2=5\kappa _{a}$, and an optimal phase $\theta =0$. In Ref.~\cite{Strong2}, a YIG sphere with a diameter of 0.5 mm was used, which contains more than $10^{17}$ spins. Therefore, we consider that the magnon modes are at macroscopic scale and the entangled states of them can be referred to as macroscopic quantum states. Note that, for simplicity, we have taken equal cavity (magnon) decay rates, $\kappa_{a_{1}} = \kappa_{a_{2}} = \kappa_a$ ($\kappa_{m_{1}} = \kappa_{m_{2}} = \kappa_m$). However, the results obtained in this paper can be straightforwardly extended to the general case of unequal decay rates. The entanglement is in the steady state guaranteed by the negative eigenvalues (real parts) of the drift matrix $A$. Actually, the steady state is always guaranteed for realistic nonzero decay rates due to the specific form of the drift matrix.
\begin{figure}[b]
\hskip-0.8cm\includegraphics[width=0.7\linewidth]{fig3.pdf}
\caption{(a) Density plot of the entanglement $E_{mm}$ between two magnon modes versus $r$ and $g_2/g_1$, with $g_1=5\kappa _{a}$. The region between two dashed lines denotes $E_{mm}>0$. (b) $E_{mm}/E_{aa}$ versus squeezing $r$ for various couplings $g_1=g_2=0.5 \kappa_a, \kappa_a, 2\kappa_a$, corresponding to solid lines from bottom to top, respectively. The rest of the parameters are $\omega_{a_1}/2\pi =10$ GHz, $\Delta _{a_{1,2}}=\Delta _{m_{1,2}}=0$, $\kappa_a/2\pi =5$ MHz, $\kappa_{m}=\kappa _a/5$, $\theta =0$, and $T=100$ mK.}
\end{figure}
It would be interesting to investigate the effectiveness of the present scheme against the mismatch of the two couplings. In a similar proposal~\cite{Jaya}, a {\it single-mode} squeezed field is injected into one cavity to entangle two magnon modes. Specifically, the squeezed field is used to squeeze one collective quadrature of two magnon modes to violate specific inequalities thus demonstrating magnon entanglement. In there, identical coupling strengths are preferred to effectively implement the proposal. In contrast, our scheme uses a different mechanism: two magnon modes get entangled due to the quantum correlation transferred from a {\it two-mode} squeezed field, and this would overcome the limitation on the couplings. Indeed, as shown in Fig. 3(a), considerable entanglement is generated in a wide range of mismatch of the two couplings, and the situation of smaller squeezing $r$ is more tolerant to the mismatch. We also explore the entanglement transfer efficiency, reflected by the ratio $E_{mm}/E_{aa}$, from the two cavity modes to the two magnon modes. In Fig. 3(b), we plot $E_{mm}/E_{aa}$ as a function of squeezing $r$ for three cases of the couplings $g_1=g_2=0.5 \kappa_a$, $\kappa_a$, and 2$\kappa_a$. It is evident that in order to transfer the entanglement with high efficiency strong coupling $g_{1,2} > \kappa_a$ should be used. The fact that coupling strength as large as double cavity decay rate yields about 90\% transfer efficiency makes our scheme quite promising.
\section{Analytical solutions at optimal conditions}
\label{Analy}
In the preceding section, we have numerically shown optimal parameter regimes for the generation of sizable magnon entanglement. When the cavity and magnon modes are resonant with the input fields and the two couplings are strong and take close values, large magnon entanglement can be achieved which increases with the squeezing of the input fields. The entanglement is in the steady state and robust against environmental temperature. In this section, we explore more deeply the problem by providing analytical solutions under the above optimal conditions, where the steady-state CMs take relatively simple expressions.
The two cavity modes get entangled due to the injection of the two-mode squeezed vacuum field, which shapes the noise properties of quantum fluctuations of the cavity fields, i.e., the cavity input noise operators become quantum correlated. This can be clearly seen in the CM of the two cavity modes by setting the couplings $g_1=g_2=0$ (the magnons get decoupled) and $\Delta _{a_1}=\Delta _{a_2}=0$, which takes the following form
\begin{equation}\label{Vaa}
V_{aa} =\frac{1}{2}
\begin{pmatrix}
\cosh 2r & 0 & \sinh 2r & 0 \\
0 & \cosh 2r & 0 & - \sinh 2r \\
\sinh 2r & 0 & \cosh 2r & 0 \\
0 & - \sinh 2r & 0 & \cosh 2r \\
\end{pmatrix} ,
\end{equation}
which is independent of cavity decay $\kappa_a$ and is exactly the CM of the two-mode squeezed vacuum state with squeezing $r$~\cite{GAJPA}. The logarithmic
negativity of such a state is
\begin{equation}
E_{aa} = \max \Big\{ 0, -{\rm ln} \big( \,{\rm min} \big\{ |\cosh r - \sinh r |^2 , \, |\cosh r + \sinh r |^2 \big\} \, \big) \Big\} ,
\end{equation}
which increases with the squeezing $r$, as shown in Fig. 4.
\begin{figure}\label{fig4}
\hskip-1.0cm\includegraphics[width=0.4\linewidth]{fig4.pdf}
\caption{Logarithmic negativity of the two cavity modes versus squeezing $r$ with $\Delta _{a_1}=\Delta _{a_2}=0$, $\theta=0$, and $g_1=g_2=0$.}
\end{figure}
The cavity-cavity entanglement is then partially transferred to the two magnon modes when the two couplings (beamsplitter type) are switched on, $g_1=g_2 = g >0$. The stationary CM of the two magnon modes can be achieved, which is
\begin{equation}\label{Vmm}
V_{mm} =\frac{1}{2(1{+}a)(a{+}b^2)}
\begin{pmatrix}
a (1{+}a{+}b^2) {+} b^2 \cosh 2r & 0 & - b^2 \sinh 2r & 0 \\
0 & a (1{+}a{+}b^2) {+} b^2 \cosh 2r & 0 & b^2 \sinh 2r \\
- b^2 \sinh 2r & 0 & a (1{+}a{+}b^2) {+} b^2 \cosh 2r & 0 \\
0 & b^2 \sinh 2r & 0 & a (1{+}a{+}b^2) {+} b^2 \cosh 2r \\
\end{pmatrix} ,
\end{equation}
where $a=\kappa_m/\kappa_a$, $b=g/\kappa_a$, and we have assumed $N_m \simeq 0$, which is the case at low temperature $T <100$ mK for magnon frequencies $\omega_{m_j} \sim 10$ GHz. The logarithmic negativity of such a state is, however, too lengthy to be reported here. In Fig. 5, we show both the steady-state cavity entanglement and magnon entanglement as a function of $\kappa_m/\kappa_a$ and $g/\kappa_a$. It is clear that small magnon decay rates and large coupling rates, $\kappa_m \ll \kappa_a \ll g$, are preferred for obtaining large magnon entanglement, and as the couplings increase, the entanglement is gradually transferred from the two cavity modes to the two magnon modes. This is very much alike to the case of transferring single-mode squeezing from the cavity to the magnon mode~\cite{Jie19PRA}, where one quadrature of the magnon mode is optimally squeezed with large coupling and small magnon decay rate, $g \gg \kappa_a \gg \kappa_m$.
\begin{figure}[t]\label{fig5}
\includegraphics[width=0.8\linewidth]{fig5.pdf}
\caption{Density plot of the entanglement (a) $E_{aa}$ of two cavity modes, (b) $E_{mm}$ of two magnon modes, versus $\kappa_m/\kappa_a$ and $g/\kappa_a$, with $\Delta _{a_{1,2}}=\Delta _{m_{1,2}}=0$, $\theta=0$, and $r=1$.}
\end{figure}
We note that in the steady state in each cavity the cavity and magnon modes never get entangled as their interaction is linear and is of beamsplitter type, $H_{int} = g \big( a m^{\dag }+a^{\dag }m \big)$~\cite{Jie18PRL,Jie19}. This is confirmed by the zero value logarithmic negativity. The CM of the cavity-magnon system in each cavity is given by
\begin{equation}\label{Vam}
V_{am} =\frac{1}{2(1{+}a)(a{+}b^2)}
\begin{pmatrix}
a b^2 {+} (a{+}a^2{+}b^2) \cosh 2r & 0 & 0 & - 2 a b \sinh^2 r \\
0 & a b^2 {+} (a{+}a^2{+}b^2) \cosh 2r & 2 a b \sinh^2 r & 0 \\
0 & 2 a b \sinh^2 r & a (1{+}a{+}b^2) {+} b^2 \cosh 2r & 0 \\
- 2 a b \sinh^2 r & 0 & 0 & a (1{+}a{+}b^2) {+} b^2 \cosh 2r \\
\end{pmatrix} ,
\end{equation}
and the logarithmic negativity can be written as $E_{am} \,{=} \,\max \big\{ 0, \cal{N} \big\} $, where $\cal{N}$ is a long expression and always nonpositive, as shown in Fig. 6, implying that the cavity and magnon modes are separable.
\begin{figure}
\includegraphics[width=0.4\linewidth]{fig6.pdf}
\caption{Density plot of $\cal{N}$ as a function of $\kappa_m/\kappa_a$ and $g/\kappa_a$. The parameters are as in Fig. 5.}
\end{figure}
\section{Conclusions}
\label{Conc}
We have presented a scheme to prepare entangled states of two magnon modes in two massive YIG spheres via transferring quantum correlations from a two-mode squeezed microwave field. We have shown optimal parameter regimes for achieving strong magnon entanglement, and in particular, studied the effectiveness of the scheme towards the mismatch of two cavity-magnon couplings and analysed the entanglement transfer efficiency. Large coupling rates and small magnon decay rates with respect to cavity decay rates are preferred for the entanglement. We have shown that, with experimentally accessible two-mode squeezed source, strong magnon entanglement could be realized which survives up to hundreds of milliKelvin. Macroscopic entangled states of magnon modes are not only useful for fundamental studies of quantum-to-classical transition, decoherence theories at macroscopic scale~\cite{Bassi}, but can also be applied to quantum information processing based on magnonic systems~\cite{NakaRev} as valuable resources.
\begin{acknowledgments}
This research was supported by National Key Research and Development Program of China (Grants No. 2017YFA0304200 and No. 2017YFA0304202), National Natural Science Foundation of China (Grant No. 11674284), Zhejiang Provincial Natural Science Foundation of China (Grant No. LD18A040001), and the Fundamental Research Funds for the Center Universities (No. 2019FZA3005).
\end{acknowledgments}
|
\section{Introduction}
\begin{figure}[h]
\begin{minipage}{0.55\textwidth}
Polynomial interpolation of Chebyshev nodes is one of the most data-efficient approaches for modelling a noise-free, $n$-differentiable function on the interval $[-1, 1]$\cite{interpolationintro}. The error between the approximating polynomial $P_{n-1}$ of degree at most $n-1$ and the true function $f$ is described by:
\begin{equation}
\label{chebyerror}
|f(x)-P_{n-1}(x)| \leq \frac{1}{2^{n-1}n!} \max_{\xi\in[-1,1]}|f^{(n)}(\xi)|
\end{equation}
The primary difficulty of applying this to neural networks is the treatment of inputs that may be noisy or outside of $[-1, 1]$. We solve this by learning the support of polynomials that map inputs to outputs. For every input unit $x_i$, we learn the positions of $n+1$ points such that the polynomial $P_n$ that forms from these points has output $P_n(x_i)$ that minimizes the final loss function. We fix the $x$-positions of the support points to be the Chebyshev nodes and learn the $y$-positions. Figure \ref{cifar-visualization} shows an example of this mechanism.
\end{minipage}\hspace{.04\textwidth}
\begin{minipage}{0.48\textwidth}
\centering
\includegraphics[scale=0.42]{images/cifar10-32n.png}
\caption{
Chebyshev-Lagrange activations before \textit{(top row)} and after \textit{(bottom row)} recieving backpropagation for 100 epochs of training on CIFAR-10. We show the activations for the first \textit{(left column)} and second \textit{(right column)} elements of the last linear layer of a modified ResNet-32.
}
\label{cifar-visualization}
\end{minipage}
\end{figure}
This idea has been explored with piece-wise polynomial activations where the model learns the weights for the Lagrangian basis functions of Chebyshev-Lobatto nodes \cite{piecewisepolynomial}, which is equivalent to learning the $y$-positions of the nodes. Apart from applying further piece-wise polynomials, there was no additional mechanism for handling inputs outside of $[-1,1]$. Therefore, the main novelty of our work is the linear extension of the polynomial from either end\footnote{Although we can linearly scale the Chebyshev nodes into an arbitrary range of $[a, b]$, it is generally unknown what range will definitely cover all inputs for any task. We also demonstrate in the results that if the input is not restrained within this stable region, the model quickly experiences exploding gradients. Therefore, we simply provide solutions for handling inputs outside of $[-1, 1]$.} using regression or extrapolation, which we find to be key in improving results and attaining competitive performance with modern architectures on a variety of tasks.
We rationalize the basic Chebyshev-Lagrange method and our variations in Sections \ref{background} and \ref{implementation}, then describe the selection process for the best variation using synthetic datasets and demonstrate its practical value on real datasets in Sections \ref{datasets-and-arch} and \ref{results}, and finally offer explanations for its versatility and performance and discuss future work in Section \ref{discussion}.
\subsection{Background}\label{background}
Several existing architectures allude to the idea that parameterized piece-wise polynomial activations can improve the data efficiency of ReLU-based neural networks \cite{prelu,maxout,polynomialnn}. PReLU is a modified ReLU where the output slope of negative inputs is learned. It is claimed to be the reason for surpassing human performance on the task of classifying of ImageNet 2012 \cite{prelu}. Maxout is a parameterized activation function where the output is the maximum of a linear transformation \cite{maxout}. It performed better on various image classification tasks than similar architectures that used ReLU \cite{maxout}. In particular, maxout differed from ReLU in that it could learn to approximate simple piece-wise polynomials per hidden unit \cite{maxout}. We build upon these observations and explicitly implement polynomials with degrees greater than one, with the goal of complementing and competing with existing neural architectures in terms of data efficiency on a variety of tasks.
Weighted Chebyshev polynomials (WCP) and Lagrangian interpolation of Chebyshev nodes (CL) are two related ways of constructing polynomials with good interpolative properties within a chosen interval. We refer the reader to Levy \cite{interpolationintro} for introductory material on these methods. Original work in artificial neurons that used WCP demonstrated improved convergence speed \cite{chebyuniversal}, simplification of neural network operations \cite{chebysigmoid} and successful application in spectral graph convolutions \cite{graphcnn, structcnn, multigraphnet}. \cite{graphcnn} implemented spectral graph convolutions using a filter operation defined by the sum of weighted Chebyshev polynomials at the normalized graph Laplacian matrix $\Tilde{L}$ of all training data: $f(x) := \sum^n_{k=0} \theta_kT_k(\Tilde{L})x$, where $\theta_k$ are the parameters and $T_k$ are the Chebyshev basis polynomials defined by the recurrence relation $T_0(x) = 1,\; T_1(x) = x,\; T_i(x) = 2xT_{i-1}(x)-T_{i-2}(x)$ for $i\geq2$. Chebyshev polynomial modelling is an appropriate method here because the values of a normalized Laplacian matrix are contained in the interval $[-1, 1]$ and are constant per graph configuration. On the other hand, little work has been done on incorporating CL into neural networks. The main advantage of CL over WCP is that the former can be expressed in closed form and thus benefits from parallelized tensor operations. We will also explore variations of CL that take advantage of the straightforward gradients and nodes of CL for linear extrapolation.
\section{Methods}
\label{methods}
\subsection{Implementation of the Chebyshev-Lagrange activation functions}\label{implementation}
We examine several variants of CL. $\tanh$ and prototype cosine-similarity \cite{deepconsensus} are two ways of compressing input spaces onto the interval $[-1, 1]$. Whereas the former computes this element-wise, the latter returns a vector representing the cosine-similarity of the entire input vector with each of its prototype parameters \cite{deepconsensus}. We refer to these as `restrictive' of the inputs to the polynomial. We also experiment with non-restrictive methods that linearly extrapolate (\textit{CL-extrapolate}) or regress (\textit{CL-regression}) the polynomial outside of $[-1, 1]$. They are expressed as:
\begin{equation}
\label{chebyshevlagrange-formula}
\sigma(v_i) =
\begin{cases}
m_{(-1)}v_i+(y_{n+1}-m_{(-1)}x_{n+1}) & v_i < -1 \\
P_n(v_i) & -1 \leq v_i \leq 1 \\
m_{(+1)}v_i + (y_1-m_{(+1)}x_1) & v_i > 1, \\
\end{cases}
\end{equation}
where $v_i$ is the input unit, $m_{(-1)}$ and $m_{(+1)}$ are the slopes of the linear pieces and $P_n$ is a polynomial of degree $n$ defined by the Chebyshev nodes $x_1, x_2, \ldots, x_{n+1}$ and parameters $y_1, y_2, \ldots, y_{n+1}$. We choose to join the linear portions to the central polynomial. Table \ref{slope-difference} outlines the difference in linear slopes for extrapolation and regression. Table \ref{prototype-variants} summarizes all variants. We describe the parameters and implementation of Chebyshev-Lagrange activations in appendices \ref{appendix-chebyparams} and \ref{appendix-lagrangecalc} because they are not the main novelty. Since the backpropagation signal to $y_i$ is not directly proportional to the output of the activation, we aim for simplicity and choose to initialize all $y$-positions to zero as Figure \ref{cifar-visualization} shows.
\begin{table}[h]
\centering
\caption{Linear slope implementations for non-restrictive, piece-wise variants of Chebyshev-Lagrange activations (CL) at inputs less than ($m_{(-1)})$) or greater than $1$ ($m_{(+1)})$. For \textit{CL-regression}, we must also choose the hyperparameter $k$ for the number of nodes from either ends of the polynomial to use in regression. The index ordering of regression is reversed because $x_1$ is closest to $+1$ and $x_{n+1}$ is closest to $-1$ for Chebyshev nodes. Appendix \ref{appendix-lagrangegrad} shows an example of how to precompute the gradients of the Lagrangian bases on tensors.
}
\label{slope-difference}
\begin{tabular}{c|c|c}
\hline
Model & $m_{(-1)}$ & $m_{(+1)}$ \\
\hline
Extrapolation & $P_n'(-1)$ & $P_n'(+1)$ \\
Regression & $\text{Cov}(x_{n-k+1...n}, y_{n-k+1...n})/\text{Var}(x_{n-k+1...n})$ & $\text{Cov}(x_{1...k}, y_{1...k})/\text{Var}(x_{1...k})$ \\
\hline
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{
Variations of the basic Chebyshev-Lagrange (CL) activation. The output is treated differently if the inputs are outside $[-1, 1]$ for piece-wise variants. We choose $n=3$ as the maximum degree of all CL polynomials and $k=2$ for the number of regression nodes.
}
\label{prototype-variants}
\begin{tabular}{ccc}
\hline
Variant & Polynomial input & Output \\
\hline
Weighted Chebyshev polynomials & $\mathbb{R}$ & Smooth \\
CL & $\mathbb{R}$ & Smooth \\
$\tanh$-CL & $[-1, 1]$ & Smooth \\
Prototype cosine-similarity \cite{deepconsensus} CL & $[-1, 1]$ & Smooth \\
\textit{CL-regression} & $\mathbb{R}$ & Piece-wise continuous \\
\textit{CL-extrapolate} & $\mathbb{R}$ & Piece-wise continuous \\
\hline
\end{tabular}
\end{table}
\subsection{Datasets and architectures}\label{datasets-and-arch}
\paragraph{Synthetic datasets.} To select the best CL variants from Table \ref{prototype-variants}, we train and test on 7 artificially generated datasets with different nonlinear input-output relationships. To generate data, the columns of matrices with random numbers sampled uniformly on $[-1, 1]$ combine according to the recipes listed in Table \ref{synthetic-description} to produce the target outputs of each dataset, which then receives Gaussian noise of 0.01 or 0.04 standard deviations to simulate field observations. We produce 2000 data points for $N=1000$ training and testing sets. See Appendix \ref{appendix-samplesynthetic} for sample visualizations of these datasets.
\begin{table}[h]
\centering
\caption{Recipes of the target output for each synthetic dataset. Subscripts of $x$ indicate the column index of the uniform random matrix input, which has exactly the number of required columns.}
\begin{tabular}{c|c|c}
\hline
Dataset & Input type & Output recipe \\
\hline
Pendulum & Smooth & $-x_1x_2\sin(2\pi x_0)$ \\
Arrhenius & Smooth & $x_1e^{-x_2x_0/4}$\\
Gravity & Smooth & $x_1x_2x_3/(0.2+x_0^2)$\\
Sigmoid & Smooth & $2x_1/(1+e^{-10x_2(x_0-x_3+0.5)})+x_4-0.5$ \\
PReLU & Continuous & if $x_0<0$ then $0.1x_0x_1$ else $x_0x_2$\\
Jump & Discontinuous & if $x_0 < x_1-3/4$ then $4x_2x_0$ else $0.1x_3((4x_2x_0)-x_2/2)$\\
Step & Discontinuous & for $t$ in $[-0.8, -0.4, 0, 0.4, 0.8]$, return $t$ if $x_0 < t$. If done, return $0.8$.\\
\hline
\end{tabular}
\label{synthetic-description}
\end{table}
The architectures are 32-width residual networks that use the different activations and properties described in Table \ref{synthetic-architectures}.
Weights of all fully-connected and convolution layers in this paper are initialized using the He uniform distribution though it likely benefits ReLU \cite{prelu} more than CL.
We train these variants for 300 epochs on each dataset using L1-loss, stochastic gradient descent (SGD) with batch size 32, 0.01 learning rate, 0.99 momentum, $1\times10^{-6}$ weight decay, cosine annealing learning rate scheduling \cite{cosine-anneal} and report the test performance in root mean square error (RMSE). Note that we choose L1-loss instead of mean squared error because the latter required careful selection of learning rates for datasets involving exponentials (Arrhenius, Sigmoid), large gaps in discontinuities (Jump) or inverses (Gravity) to avoid exploding gradients. We repeat all experiments across 10 random seeds.
\begin{table}[h]
\centering
\caption{
Chebyshev-Lagrange (CL) architecture summary for the synthetic experiments. The base architecture is a 32-width residual network using fully connected layers. We include weighted Chebyshev polynomials (WCP) as a baseline. All CL activations have maximal degree $n=3$.
The rows are grouped by the type of activation: controls, non-restrictive inputs, restrictive inputs and non-restrictive inputs with piecewise outputs. ReLU and $\tanh$ are vanilla activations. Cubic computes the third power of the input. Prototype cosine-similarity \cite{deepconsensus} (PCS) and $\tanh$ are two different ways of compressing inputs onto $[-1, 1]$ before passing it to CL (PCS-CL and $\tanh$-CL). \textit{CL-extrapolate} and \textit{CL-regression} compute linear extrapolation and regression from the two ends of the polynomial for any inputs outside of $[-1, 1]$.
}
\label{synthetic-architectures}
\begin{tabular}{ccccc}
\hline
Activation & Residual blocks & Layers/Block & Parameters \\
\hline
ReLU & 3 & 1 & 3300 \\
ReLU (2x depth) & 6 & 1 & 6500 \\
ReLU (2x layers) & 3 & 2 & 6500 \\
$\tanh$ & 3 & 1 & 3300 \\
Cubic & 3 & 1 & 3300 \\
\hline
CL & 3 & 1 & 3800 \\
WCP & 3 & 1 & 3800 \\
\hline
PCS-CL & 3 & 1 & 6785 \\
$\tanh$-CL & 3 & 1 & 3800 \\
\hline
\textit{CL-regression} & 3 & 1 & 3800 \\
\textit{CL-extrapolate} & 3 & 1 & 3800 \\
\hline
\end{tabular}
\end{table}
\paragraph{DementiaBank.}\label{methods-dementiabank} DementiaBank is a collection of 551 audio recordings of 210 subjects in various stages of cognitive decline \cite{dementiabank}. We filter for subjects with dementia or no diagnosis and count 178 recordings with the dementia label and 229 with no diagnosis. Per recording, we extract the 480 minimally-correlated features specified in \cite{heterogeneousdata}. As demonstrated in \cite{fraser-linguistic-alzheimers} and \cite{multiview-emb}, feature selection is critical for good performance on this dataset. For feature selection, we use the FamousPeople dataset, which is a collection of 543 audio recordings of 32 subjects in various stages of cognitive decline \cite{heterogeneousdata}. We replicate \cite{heterogeneousdata} by computing and normalizing the same 480 features used in DementiaBank for the recordings. We follow \cite{heterogeneousdata} and use the Boruta algorithm \cite{boruta} on FamousPeople to select the 66 final features listed in Appendix \ref{appendix-dementiabank-borutafeatures}. These features are used for DementiaBank classification.
DementiaBank classification experiments investigate the effect of replacing ReLU or $\tanh$ with the best-performing CL variant from the synthetic experiments. The base network is a 2-block, 6-layer residual network \cite{resnet}, with widths of 32 or 64 to control for hyperparameter optimization. As per \cite{fraser-linguistic-alzheimers, heterogeneousdata, normativedata}, the training and testing sets are produced from random 10-fold cross validation such that each testing fold does not contain samples of subjects existing in the training set. The models train for 300 epochs with cross entropy loss, SGD with batch size 32, 0.01 learning rate, 0.9 momentum, $1\times10^{-4}$ weight decay, cosine annealing learning rate scheduling \cite{cosine-anneal} and we report the average validation accuracy, sensitivity, specificity and F1 score for the 10-fold cross validation. Since we repeat the experiments across 30 random seeds, this amounts to 300 random cross validation folds.
\paragraph{MNIST and CIFAR-10.}\label{methods-imageclass} MNIST contains 60,000 training and 10,000 testing samples of black and white handwritten digits in $28\times28$ pixel arrays \cite{mnist}. CIFAR-10 contains 50,000 training and 10,000 testing samples of RGB images in $3\times32\times32$ pixel arrays \cite{cifar10}. The typical task for both datasets is to map the pixel inputs to one of the ten labelled classes. For MNIST, we pad images to $32\times32$ pixels and augment with 2-pixel translations. For CIFAR-10, we apply the typical data augmentation techniques of random 4-pixel translations and horizontal flips \cite{resnet, shake-shake}. The same architecture described in \cite{shake-shake} is used for both datasets, but we choose 14 layers instead of 26 and progressing widths of 16, 32, 64 and 128 instead of the original progression of 16, 64, 128 and 256. We also train for 100 instead 300 epochs. These architectural and training changes are meant to reduce the computation time of repeating all experiments across 3 random seeds and we expect lower final scores as a result. Shake-shake regularization is disabled
for MNIST since we obtain no discernible benefit with it. We replicate these experiments using ReLU and \textit{CL-extrapolate} as the last 8 convolutional activations.
Appendix \ref{appendix-architectures} provides more details about the architectures.
\section{Results}
\label{results}
\subsection{Synthetic datasets}
Table \ref{synthetic-0.01} shows the RMSE mean and standard deviations for 10 repeated trials of each model in Table \ref{synthetic-architectures} on each dataset in Table \ref{synthetic-description} with 0.01 standard deviations of Gaussian noise. Given the same data points, models using Chebyshev-Lagrange variants that compute linear extrapolation (\textit{CL-extrapolate}) show similar or notable improvement of interpolation accuracy over all other classes of activation, even when the underlying functions are discontinuous (i.e., Jump and Step). Despite limiting the raw input in $[-1, 1]$, the non-restrictive Cubic, CL and WCP activations succumb to exploding gradients on the majority of these datasets.
\begin{table}[h]
\centering
\caption{Root mean squared error (RMSE) for the smooth (\ref{synthetic-smooth-0.01}) and non-smooth (\ref{synthetic-nonsmooth-0.01}) synthetic datasets with 0.01 standard deviations of Gaussian noise. NaN indicates counts of gradient explosion. See Table \ref{synthetic-description} and \ref{synthetic-architectures} for descriptions of the datasets and models. Excluding rows containing any NaN, bold font marks the two most data-efficient methods while red font marks the four least data-efficient methods.}
\label{synthetic-0.01}
\subfloat[Smooth functions]{
\label{synthetic-smooth-0.01}
\begin{tabular}{c|c|c|c|c}
\hline
Model & Pendulum & Arrhenius & Gravity & Sigmoid \\
\hline
ReLU & $\badscore{0.1518\pm0.0297}$ & $\badscore{0.0054\pm0.0003}$ & $\badscore{0.098\pm0.069}$ & $\badscore{0.065\pm0.010}$ \\
ReLU (2x depth) & $\badscore{0.1941\pm0.0207}$ & $\badscore{0.0052\pm0.0004}$ & $\badscore{0.187\pm0.115}$ & $\badscore{0.058\pm0.007}$ \\
ReLU (2x layers) & $\badscore{0.1597\pm0.0203}$ & $\badscore{0.0048\pm0.0003}$ & $\badscore{0.062\pm0.049}$ & $\badscore{0.043\pm0.005}$\\
$\tanh$ & $0.0327\pm0.0102$ & $0.0036\pm0.0005$ & $\badscore{0.075\pm0.059}$ & $\badscore{0.111\pm0.013}$\\
Cubic & (10/10 NaN) & (4/10 NaN) & (10/10 NaN) & (10/10 NaN) \\
\hline
CL & (8/10 NaN) & $0.0029\pm0.0007$ & (10/10 NaN) & (10/10 NaN) \\
WCP & (9/10 NaN) & $0.0032\pm0.0010$ & (10/10 NaN) & (10/10 NaN) \\
\hline
PCS-CL & $0.0348\pm0.0021$ & $0.0045\pm0.002$ & $\mathbf{0.022\pm0.002}$ & $0.035\pm0.010$ \\
$\tanh$-CL & $\badscore{0.1642\pm0.0749}$ & $\badscore{0.0056\pm0.0015}$ & $0.048\pm0.009$ & $0.041\pm0.011$ \\
\hline
\textit{CL-regression} & $\mathbf{0.0208\pm0.0022}$ & $\mathbf{0.0035\pm0.0002}$ & $0.042\pm0.003$ & $\mathbf{0.027\pm0.004}$ \\
\textit{CL-extrapolate} & $\mathbf{0.0113\pm0.0006}$ & $\mathbf{0.0030\pm0.0002}$ & $\mathbf{0.022\pm0.002}$ & $\mathbf{0.019\pm0.004}$ \\
\hline
\end{tabular}
}
\subfloat[Non-smooth functions]{
\label{synthetic-nonsmooth-0.01}
\begin{tabular}{c|c|c|c}
\hline
Model & Jump & PReLU & Step \\
\hline
ReLU & $\badscore{0.17\pm0.15}$ & $\badscore{0.0063\pm0.0006}$ & $\badscore{0.040\pm0.010}$ \\
ReLU (2x depth) & $\badscore{0.17\pm0.12}$ & $\badscore{0.0079\pm0.0053}$ & $0.035\pm0.010$\\
ReLU (2x layers) & $\badscore{0.12\pm0.03}$ & $0.0058\pm0.0004$ & $\mathbf{0.030\pm0.013}$\\
$\tanh$ & $\mathbf{0.09\pm0.03}$ & $\badscore{0.0085\pm0.0041}$ & $\badscore{0.050\pm0.005}$\\
Cubic & (10/10 NaN) & (9/10 NaN) & (7/10 NaN) \\
\hline
CL & (10/10 NaN) & $0.0071\pm0.0005$ & $0.109\pm0.018$\\
WCP & (10/10 NaN) & $0.0066\pm0.0006$ & $0.091\pm0.025$\\
\hline
PCS-CL & $0.11\pm0.3$ & $\badscore{0.0060\pm0.0005}$ & $\badscore{0.069\pm0.007}$\\
$\tanh$-CL & $\badscore{0.12\pm0.03}$ & $\badscore{0.0060\pm0.0008}$ & $\badscore{0.088\pm0.018}$\\
\hline
\textit{CL-regression} & $\mathbf{0.08\pm0.02}$ & $\mathbf{0.0045\pm0.0003}$ & $\mathbf{0.020\pm0.001}$ \\
\textit{CL-extrapolate} & $\mathbf{0.09\pm0.02}$ & $\mathbf{0.0040\pm0.0003}$ & $\mathbf{0.030\pm0.002}$ \\
\hline
\end{tabular}
}
\end{table}
In order to test if these particular CL variants do better than the other activations only in cases of small noise pertubation, we increase the Gaussian noise standard deviation to 0.04. Table \ref{synthetic-0.04-samples} in Appendix \ref{appendix-syntheticnoisy} visualizes the magnitude of perturbation this has on the inputs from $[-1, 1]$. Table \ref{synthetic-0.04} in Appendix \ref{appendix-syntheticnoisy} shows that \textit{CL-extrapolate} continues to perform best for smooth functions and achieves identical performance to the best vanilla activation variants on non-smooth functions, while \textit{CL-regression} degrades in performance with the increased noise. Non-restrictive activations continue to experience high frequencies of exploding gradients. Although PCS-CL achieved the lowest RMSE on $2/7$ datasets, it performed inconsistently on datasets with less noise. Therefore, we selected \textit{CL-extrapolate} to investigate in the following experiments on real datasets.
\subsection{Dementia detection}
Table \ref{chebyeffect-dementiabank} shows the consistent improvement in various metrics of binary classification from replacing ReLU or $\tanh$ with \textit{CL-extrapolate}. Table \ref{dementiabank-globalcomparison} indicates that these results are, to the best of our knowledge, the state-of-the-art for this task and dataset. Note that results from \cite{dementiabank-cnn} are excluded since their dataset is different, as indicated by a higher 1-class accuracy of 79.8\%.
\begin{table}[h]
\centering
\caption{
The effect of replacing ReLU or $\tanh$ with extrapolated Chebyshev-Lagrange (CL) activations on the DementiaBank healthy-versus-dementia classification scores (\%). Dotted lines separate the controls from the experimental condition for each architecture class. The parameter counts of width 32 and 64 ReLU/$\tanh$ are 6.5K and 21K respectively, and increase to 7.0K and 22K for CL. 1-class represents a model that predicts the majority class (i.e. healthy condition) only. Results represent means and standard deviations of 30 random initializations and sets of 10-fold cross validation. Lilliefors test for diagnosing non-normal distribution of the $\tanh$ and \textit{CL-extrapolate} results returned $p > 0.10$ for either distributions, validating the Student's $t$-test with no assumption of equal variance to compute the significance in their differences of means. $^{**}$ denotes $p < 0.01$ and $^{***}$ denotes $p < 0.001$.
}
\label{chebyeffect-dementiabank}
\begin{tabular}{l|c|c|c|c|c}
\hline
Activation & Width & Accuracy & Sensitivity & Specificity & Micro-F1 \\
\hline
1-class & \multirow{4}{*}{32} & $56.3\pm 0.0$ & $0.0\pm0.0$ & $100.0\pm0.0$ & $0.0\pm0.0$\\
ReLU & & $82.5\pm1.2$ & $79.4\pm1.5$ & $84.9\pm1.8$ & $79.9\pm1.3$\\
$\tanh$ & & $82.7\pm1.2$ & $79.1\pm1.3$ & $85.4\pm1.5$ & $80.0\pm1.3$\\
\hdashline
\textit{CL-extrapolate} & & $\mathbf{84.7\pm 0.8^{***}}$ & $\mathbf{80.7\pm 1.2^{***}}$ & $87.7\pm1.0^{***}$ & $\mathbf{82.1\pm0.9^{***}}$\\
\hline
1-class & \multirow{4}{*}{64} & $56.3\pm 0.0$ & $0.0\pm0.0$ & $100.0\pm0.0$ & $0.0\pm0.0$ \\
ReLU & & $82.9\pm1.0$ & $79.6\pm1.4$ & $85.4\pm1.2$ & $80.3\pm1.2$ \\
$\tanh$ & & $83.1\pm1.1$ & $79.7\pm1.8$ & $85.7\pm1.3$ & $80.5\pm1.3$\\
\hdashline
\textit{CL-extrapolate} & & $\mathbf{84.6\pm 0.9^{***}}$ & $\mathbf{80.7\pm 1.3^{**}}$ & $87.6\pm1.0^{***}$ & $\mathbf{82.0\pm1.1^{***}}$\\
\hline
\end{tabular}
\end{table}
\begin{table}[h]
\caption{
A comparison of 32-width, 2-block, 6-layer residual networks that use extrapolated Chebyshev-Lagrange (CL-ex) activations with previously reported results on DementiaBank healthy-versus-dementia classification scores (\%). CL-ex results for accuracy, micro-F1 and macro-F1 show means and standard deviations of 30 random initializations and sets of 10-fold cross validation. As per \cite{multiview-emb}, we also show the best macro-F1 across 10 validation folds for a single seed.
}
\label{dementiabank-globalcomparison}
\centering
\begin{tabular}{l|c|c|c|c}
\hline
Model & Accuracy & Micro-F1 & Macro-F1 & Best macro-F1 \\
\hline
1-layer neural net \cite{heterogeneousdata} & -- & $75.9$ & -- & -- \\
TCN \cite{transductiveconsensus} & $75-77$ & -- & -- & --\\
Random forest (RF) \cite{normativedata} & -- & $79.6$ & -- & --\\
SVM \cite{normativedata} & -- & $80.6$ & -- & -- \\
Multiview embed. + RF \cite{multiview-emb} & -- & -- & -- & $82.4\pm5.2$ \\
Custom features + CL-ex ResNet & $84.7\pm 0.8$ & $82.1\pm0.9$ & $81.8\pm1.1$ & $84.0\pm7.5$ \\
\hline
\end{tabular}
\end{table}
\subsection{Image classification}
Table \ref{image-results} shows MNIST and CIFAR-10 results when we replace ReLU with \textit{CL-extrapolate} in the last 8 convolutional layers of a 14 layer residual network. No significant change is seen.
\begin{table}[h]
\centering
\caption{Results for image classification. Accuracy shows mean and standard deviations for 3 random initializations. Since normality of the two distributions is unknown, we use the Kruskal-Wallis H-test on the results of extrapolated Chebyshev-Lagrange versus ReLU to determine the p-value. }
\label{image-results}
\subfloat[MNIST]{
\begin{tabular}{c|c|c}
\hline
Model & Accuracy & P-value \\
\hline
ReLU & $99.69\pm0.04$ & \multirow{2}{*}{0.38} \\
\textit{CL-extrapolate} & $99.66\pm0.03$ & \\
\hline
\end{tabular}
\label{mnist-results}
}\quad
\subfloat[CIFAR-10]{
\begin{tabular}{c|c|c}
\hline
Model & Accuracy & P-value \\
\hline
ReLU & $93.45\pm0.10$ & \multirow{2}{*}{0.83} \\
\textit{CL-extrapolate} & $93.40\pm0.29$ & \\
\hline
\end{tabular}
\label{cifar-results}
}
\end{table}
\begin{figure}[h]
\begin{minipage}[t]{0.33\textwidth}\vspace{0pt}
Figure \ref{hist-cifar} shows the distribution of inputs and the shape of select activations in the trained model. Similar results are seen for both MNIST and CIFAR-10. Appendix \ref{appendix-relu-vis} shows corresponding plots of the original ReLU network for reference. The distribution of inputs for ReLU or \textit{CL-extrapolate} appear to be bell-shaped and centered close to the origin for the majority of units, although there is also a minority of \textit{CL-extrapolate} inputs which form bi-modal distributions. The ranges of inputs appear consistent between the two activations. The typical hidden unit of shallow layers cover intervals in $\pm10$ to $\pm30$, while deeper layers cover intervals around $\pm5$.
Since the range of input for shallow layers is significantly greater than the range covered by the polynomial, we observe \textit{CL-extrapolate} adopting shapes that emphasize piece-wise linearity. Figure \ref{hist-shallow} shows \textit{CL-extrapolate} learning various configurations of V-shape or PReLU activations. The densest part of the distribution is still within the polynomial region. For deeper layers, we see greater use of the polynomial region and smoother outputs, as illustrated in Figure \ref{hist-deep}.
\strut\end{minipage}\hspace{.02\textwidth}
\begin{minipage}[t]{0.65\textwidth}\vspace{0pt}
\subfloat[Activations in the 6th convolutional layer.]{
\includegraphics[scale=0.21]{images/cifar10-cl-shallow.png}
\label{hist-shallow}
}\\
\subfloat[Activations in the 13th convolutional layer.]{
\includegraphics[scale=0.21]{images/cifar10-cl-deep.png}
\label{hist-deep}
}
\caption{
Sample plots of activations and histograms of their inputs for the first 5 hidden units at the 6th (\ref{hist-shallow}) and 13th (\ref{hist-deep}) layers of the model trained on CIFAR-10. The $x$-axis range shows the minimum and maximum input value per element.
}
\label{hist-cifar}
\end{minipage}
\end{figure}
\section{Discussion}
\label{discussion}
The identical performance of the models used in image classification does not seem to be caused by \textit{CL-extrapolate} behaving identically to ReLU, as shown in Figure \ref{hist-cifar}. We can interpret the mechanism of ReLU as reporting the one-sided variance of its input distribution (Appendix \ref{appendix-relu-vis}). On the other hand, \textit{CL-extrapolate} can learn to report one or two-sided variances, induce and report bimodal distributions from its input network, and produce smoothly interpolated outputs (Figure \ref{hist-cifar}). These properties allow for models to adapt to both smooth and non-smooth functions, which may be the reason for their superior performance over similar variants or controls on the synthetic dataset experiments (Table \ref{synthetic-0.01}). Consistently low RMSE on the various non-linear relationships of the synthetic datasets indicates that \textit{CL-extrapolate} reduces overfitting and learns the true relationship between inputs and outputs better than the other tested activations. This implies that models can improve in data efficiency by replacing their activation functions with \textit{CL-extrapolate}. We replace ReLU or $\tanh$ with \textit{CL-extrapolate} in two different architectures trained on the task of classifying dementia diagnosis from speech and observe significantly closer performance of these models to the diagnoses of real physicians (Table \ref{chebyeffect-dementiabank}). These benefits come for free in respect to the acquisition of data, which may be of use for other machine learning tasks where data is slow or difficult to obtain.
Given this, we consider why \textit{CL-extrapolate} has limited effect on image classification (Table \ref{image-results}). One likely explanation is that the combination of noise level and function characteristic in image classification negates the interpolation benefits of \textit{CL-extrapolate}. Lagrangian interpolation of Chebyshev nodes minimizes the error bound between the true, smooth function and the polynomial of noise-free observations. When the true function is not smooth, the error bound described in equation \ref{chebyerror} no longer holds as $\max_{\xi\in[-1,1]}\left|f^{(n)}(\xi)\right| $ is not defined. However, we also observed that models that use \textit{CL-extrapolate} continue to enjoy superior interpolation even on non-smooth datasets (Table \ref{synthetic-0.01}). We observe reduced impact of \textit{CL-extrapolate} on non-smooth function modelling only with higher levels of noise contamination (Appendix \ref{appendix-samplesynthetic}, Table \ref{synthetic-0.04}). This suggests that we should use \textit{CL-extrapolate} when the input-output relationship is suspected to be smooth or the data has low levels of noise. Other options include finding ways of inducing the input and output into an artificial, smooth relationship, or applying algorithms that reduce noise, such as employing feature selection as we did in the DementiaBank experiment (Section \ref{methods-dementiabank}). An interesting experiment would be to use \textit{CL-extrapolate} to map intermediate features of deep models trained on large datasets to target outputs, assuming that these features have reduced noise compared to the raw input.
Costs for using \textit{CL-extrapolate} include increased memory and computation complexity. The implementation and gradient equations of \textit{CL-extrapolate} in Appendices \ref{appendix-lagrangecalc} and \ref{appendix-lagrangegrad} clearly show respective complexities of $O(n^2)$ and $O(n^3)$ of the hyperparameter choice $n$ for the maximum polynomial degree, which is multiplied to the base complexity of the network. We find $n=3$ to be sufficient for our experiments, but the current implementation may be impractical for the lowest convolutional layers of modern image classifiers or with much larger choices of $n$ (i.e., $n=100,1000...$).
Future work should examine the theoretical properties of these activations on generalizing the outcomes of the synthetic experiments. In particular, it is not clear why \textit{CL-extrapolate} showed superior performance on non-smooth datasets with low noise, or if there exist sets of problems where certain CL variants are likely to consistently out-perform ReLU or $\tanh$. Further research should investigate how noise influences the degradation of performance for these activations and regularization techniques to reduce the degradation rate.
\section{Conclusion}
We implement a new piecewise continuous activation function composed of the Lagrangian interpolation of parameterized Chebyshev nodes on $[-1, 1]$ and extrapolated linear components outside of this range. We observe consistent and better interpolation results compared to ReLU, $\tanh$ and other activation functions on a variety of smooth and non-smooth synthetic experiments. Significant improvement is seen in DementiaBank, at no cost to MNIST or CIFAR-10 classification. We show how the proposed activations are more versatile than ReLU, which appear to explain their superior performance.
\clearpage
\bibliographystyle{ieee}
|
\subsection{Acknowledgments}
Y. M. and Y.-Z. Z. especially thank Prof. Francesco Buscemi for numerous advice and encouragement throughout the project.
We thank Jun Zhang, Wen-Yuan Wang, Yan-Lin Tang, Ping Xu, Leonardo Guerini, and Qinghe Mao for valuable and illuminating discussions.
After submission, we became aware that a similar experiment was performed using a different type of system in \cite{Graffitti2019}. The scenario considered in our work can be further generalized to the semi-quantum prepare-and-measure scenario \cite{Guerini2019}.
This work has been supported by
the National Key R\&D Program of China (2017YFA0303903),
the Chinese Academy of Science,
the National Fundamental Research Program,
the National Natural Science Foundation of China (Grant No.~61875182,
No.~11575174, No.~11874346, and No.~11574297),
Anhui Initiative in Quantum Information Technologies,
and Fundamental Research Funds for the Central Universities (WK2340000083).
\section{The RS-MDI certification of non-EB channels}
In this section, we introduce the details of realistic source based measurement-device-independent (RS-MDI) approach proposed and verified in the main text, including the experimental bound for all EB channels, the decoy-state method to exclude multiphoton contributions, and the exact form of payoff functions used in the experiment. We begin by introducing the result of semi-quantum signaling game (SQSG), which was meticulously proposed in the Rosset-Buscemi-Liang's paper \cite{rosset2018resource} (see Fig.1 of the main text). When the referee can perfectly prepare an information complete set of quantum states, a positive value of the average payoff
\begin{equation}
\label{eq:SM-payoff1}
I_\mathcal{N}=\sum_{x,y,b}\wp(b,x,y)P_\mathcal{N}\left(b|\xi_x,\psi_y\right),
\end{equation}
suggests the ability of the channel $\mathcal{N}$ to convey and maintain entanglement. Here,
$P_\mathcal{N}\left(b|\xi_x,\psi_y\right)$ is the probability of Abby obtaining $b$ given the referee asks quantum questions $\xi_x$ and $\psi_y$, and $\wp(b,x,y)$ is the payoff function assigned by the referee.
Theoretically, if $\wp(b,x,y)$, $\xi_x$, and $\psi_y$ are properly selected, then for all EB channels, the average payoff is no more than $0$. Thus, a positive value of Eq.~(\ref{eq:SM-payoff1}) certifies that the channel $\cal N$ under test is non-EB. In this work, we further let $\wp(b\neq 0,x,y)=0$ and only consider
\begin{equation}
\label{eq:SM-payoff}
I_\mathcal{N}=\sum_{x,y}\wp(0,x,y)P_\mathcal{N}\left(b|\xi_x,\psi_y\right).
\end{equation}
\subsection{The experimental bound}\label{practical-bound}
The ideal SQSG assumes that quantum states are perfectly prepared and form an information complete set. In practice, state preparation unavoidably involves flaws and errors due to realistic devices. Also, information complete sets may not be accessible due to limitations on physical systems and devices. In fact, few quantum states may be sufficient for the certification of some quantum channels. Such problems can be solved by applying an experimental bound.
To prove Eq. (2) in the main text, notice that the EB channel ${\cal N}_{\textrm{EB}}$ is in general a measure-and-prepare
channel, i.e.,
\[
{\cal N}_{\textrm{EB}}\left(\rho\right)=\sum_{k}{\rm tr}\left[E_{k}\rho\right]\gamma_{k},
\]
where $\left\{ E_{k}|E_{k}\geqslant0,\sum_{k}E_{k}=\mathbb{I}\right\} $
is a set of POVMs and $\left\{ \gamma_{k}\right\} $ is a set of quantum
states. In the RS-MDI certification, we relax sets of $\left\{\xi_x\right\}$ and $\left\{\psi_y\right\}$ to be sets of arbitrary states. The maximal average payoff for all EB channels ${\cal N}_{\textrm{EB}}$, given the payoff function $\wp\left(0,x,y\right)$, can then be bounded by
\begin{equation}
\begin{aligned}
C_{\textrm{EB}} & =\max_{{\cal N}_{\textrm{EB}}}I_{{\cal N}_{\textrm{EB}}}=\max_{{\cal N}_{\textrm{EB}}}\sum_{x,y}\wp\left(0,x,y\right)P_{{\cal N}_{\textrm{EB}}}\left(0|\xi_{x},\psi_{y}\right)\\
& =\max_{{\cal N}_{\textrm{EB}},M}\sum_{x,y}\wp\left(0,x,y\right){\rm tr}\left[{\cal N}_{\textrm{EB}}\left(\xi_{x}\right)\otimes\psi_{y}M\right]\\
& =\max_{E_{k},F_{k}}\sum_{x,y,k}\wp\left(0,x,y\right){\rm tr}\left[E_{k}\xi_{x}\right]{\rm tr}\left[\gamma_{k}\otimes\psi_{y}M\right]\\
& =\max_{E_{k},F_{k}}{\rm tr}\left[W\sum_{k}E_{k}^{T}\otimes F\left(k\right)^{T}\right],
\end{aligned}
\end{equation}
where $M$ and $F\left(k\right)={\rm tr}_{1}\left[\gamma_{k}\otimes\mathbb{I}M\right]$
are POVMs acting on ${\cal N}_{\textrm{EB}}\left(\xi_{x}\right)\otimes\psi_{y}$
and $\psi_{y}$, respectively. Here, we denote
\begin{equation}
W=\sum_{xy}\wp\left(0,x,y\right)\xi_x^T\otimes\psi_y^T.
\end{equation}
When $\xi_x$ and $\psi_y$ are from information complete sets, and $\wp\left(0,x,y\right)$ are chosen properly, $-W$ can be constructed as an EW for the Choi state of a quantum channel \cite{rosset2018resource,Wilde2013}. In this case, the bound $C_{\textrm{EB}}$ is exactly
$0$ as $E_{k}^{T}\otimes F\left(k\right)^{T}$ are separable positive
operators and ${\rm tr}\left[WE_{k}^{T}\otimes F\left(k\right)^{T}\right]\leqslant0$
holds for all $k$. For arbitrary sets of quantum states $\xi_x $ and $\psi_y$, this bound can also be analytically derived as
\begin{equation}
C_{\textrm{EB}}^{\textrm{exp}} =\max_{E_{k},F_{k}}{\rm tr}\left[\tilde{W}\sum_{k}E_{k}^{T}\otimes F\left(k\right)^{T}\right] =d^{2}\max_{\omega_{\textrm{sep}}}{\rm tr}\left[\tilde{W}\omega_{\textrm{sep}}\right],
\label{eq:expbound}
\end{equation}
where $\omega_{\textrm{sep}}$ is a separable state.
To see this, notice that $\sum_k E_k ={\mathbb I}$ and $0\leqslant F\left(k\right)\leqslant{\mathbb I}$, $\sum_{k}E_{k}^{T}\otimes F\left(k\right)^{T}$ satisfies
\begin{equation}
\sum_{k}E_{k}^{T}\otimes F\left(k\right)^{T} =d^{2}\sum_{k}\frac{{\rm tr}\left[E_{k}^{T}\right]}{d}\frac{E_{k}^{T}}{{\rm tr}\left[E_{k}^{T}\right]}\otimes\frac{F\left(k\right)^{T}}{d},
\end{equation}
where $E_k^T/{\rm tr}\left[E_k^T\right]$ and $F\left(k\right)^T/d$ can be viewed as quantum states and unnormalized quantum states, respectively.
\subsection{Evaluation of single-photon detections}
Considering the multiphotons in real photon sources, we apply the decoy-state technique to weak coherent pulses (WCPs), such that the single-photon detection events can be efficiently evaluated \cite{hwang2003quantum,wang2005beating,lo2005decoy}.
Generally, the quantum states of phase-randomized WCPs can be written in the Fock basis as
\begin{equation}
\rho_{\alpha}=e^{-\alpha}\sum_{n=0}^\infty\frac{\alpha^{n}}{n!}\left|n\right\rangle \left\langle n\right|,
\end{equation}
where $\alpha$ is the mean photon number per pulse and $n$ is the photon number. When pulses $\xi_{x}$ and $\psi_{y}$ are prepared with intensities $\alpha_{\xi}$ and $\alpha_{\psi}$, respectively, the probability to obtain the result $b$ of a jointly measurement can be written as
\begin{equation}
\label{eq:gain}
Q_{b,\xi_{x},\psi_{y}}^{\alpha_{\xi}\alpha_{\psi}}=e^{-\alpha_{\xi}-\alpha_{\psi}}\sum_{n,m=0}^{\infty}\frac{\alpha^{n}_{\xi} \alpha^{m}_{\psi}}{n! m!}Y^{nm}_{b,\xi_{x},\psi_{y}}.
\end{equation}
Here, $Q_{b,\xi_{x},\psi_{y}}^{\alpha_{\xi}\alpha_{\psi}}$ is the ratio of the number of detection events to the number of emitted pulse pairs in $\xi_{x}$ and $\psi_{y}$ with mean photon numbers $\alpha_{\xi}$ and $\alpha_{\psi}$, respectively. Consequently, $Y^{nm}_{b,\xi_{x},\psi_{y}}$ is the conditional probability of detection events $b$ given that $n$-photon and $m$-photon pulses are emitted in $\xi_{x}$ and $\psi_{y}$, respectively \cite{mdiqkd,Xu2013practical}.
The decoy-state method is applied when $\xi_x$ and $\psi_y$ are randomly prepared with different intensities.
We use three types of mean photon per pulse values, i.e., $\alpha_{\xi},\alpha_{\psi}\in\left\{ \mu,\nu,\omega | \mu>\nu>\omega=0 \right\}$, and obtain seven gains of $Q_{b,\xi_{x},\psi_{y}}^{\alpha_{\xi_x}\alpha_{\psi_y}}$ with $\{\alpha_{\xi}\alpha_{\psi}\}$ $\in$ $\{\mu\mu, \nu\nu, \mu\omega, \omega\mu, \nu\omega, \omega\nu, \omega\omega\}$. Let $J_{1}$ and $J_{2}$ be ($\xi_{x}$, $\psi_{y}$, and $b$ are omitted for simplicity),
\begin{equation}
\begin{aligned}
J_{1} & = Q_{\nu\nu}e^{2\nu}+Q_{\omega\omega}-Q_{\nu\omega}e^{\nu}-Q_{\omega\nu}e^{\nu},\\
J_{2} & = Q_{\mu\mu}e^{2\mu}+Q_{\omega\omega}-Q_{\mu\omega}e^{\mu}-Q_{\omega\mu}e^{\mu}.
\end{aligned}
\end{equation}
We calculate the equation $\mu^{3}J_{1}-\nu^{3}J_{2}$, where $Y^{n,0},Y^{0,m}$ and $Y^{1,2},Y^{2,1}$ can be canceled, and obtain
\begin{equation}
Y^{11}= \frac{\mu^{3}J_{1}-\nu^{3}J_{2}}{\mu^{2}\nu^{2}\left(\mu-\nu\right)} -\frac{1}{\mu^{2}\nu^{2}\left(\mu-\nu\right)}\sum_{n,m \geqslant 2}\frac{\mu^{3}\nu^{n+m}-\nu^{3}\mu^{n+m}}{n!m!}Y^{nm}.
\end{equation}
Since $Y^{n,m}\in[0,1]$ and $\mu>\nu$, the lower bound $Y^{1,1,L}$ and upper bound $Y^{1,1,U}$ can be written as
\begin{equation}
\begin{aligned}
Y^{11,L}= & \frac{\mu^{3}J_{1}-\nu^{3}J_{2}}{\mu^{2}\nu^{2}\left(\mu-\nu\right)},\\
Y^{11,U}= & Y^{11,L} -\frac{\mu^3 \left( e^\nu-1-\nu \right)^2 - \nu^3 \left( e^\mu-1-\mu \right)^2}{\mu^{2}\nu^{2}\left(\mu-\nu\right)},
\end{aligned}
\end{equation}
respectively.
\subsection{The 6-state and 4-state average payoff}
In this work, we consider two kinds of average payoff with different sets of input states.
\subsubsection{The 6-state average payoff}
For the first kind, we use an information complete set of states, i.e., eigenstates of three Pauli matrices denoted as $\xi_{x},\psi_{y}$ $\in\left\{ \left|0\right\rangle _{Z},\left|1\right\rangle _{Z},\allowbreak \left|0\right\rangle _{X},\left|1\right\rangle _{X},\left|0\right\rangle _{Y},\left|1\right\rangle _{Y} \right\}$, respectively.
The corresponding payoff function is
\begin{equation}
\label{six-state-payoff}
\wp\left(0,\xi_{x},\psi_{y}\right)=
\begin{cases}
\frac{1}{4}, & \left(\xi_{x},\psi_{y}\right) \text{ anti-correlated in } X \text{ or } Z;\\
-\frac{1}{4}, & \left(\xi_{x},\psi_{y}\right) \text{ correlated in } X \text{ or } Z;\\
-\frac{1}{2}, & \left(\xi_{x},\psi_{y}\right)\text{ correlated in } Y ;\\
0, & \text{otherwise}.
\end{cases}
\end{equation}
Here, ``anti-correlated'' represents $\left(0,1\right)$ or $\left(1,0\right)$ in the respective bases and ``correlated'' represents $\left(0,0\right)$ or $\left(1,1\right)$ in the respective bases.
By replacing $P_{\cal N}(b|\xi_x,\psi_y)$ with $Y^{1,1,L}$ or $Y^{1,1,U}$ according to the sign of $\wp(0,\xi_x,\psi_y)$, the lower bound of the payoff value $I_{\cal N}$ can be written as
\begin{equation}
\begin{aligned}
I_{\cal N} \geqslant I_{\cal N}^L= & -\frac{1}{4} Y^{11,U}_{0_Z 0_Z} + \frac{1}{4} Y^{11,L}_{0_Z 1_Z}
+ \frac{1}{4} Y^{11,L}_{1_Z 0_Z} -\frac{1}{4} Y^{11,U}_{1_Z 1_Z}\\
& -\frac{1}{4} Y^{11,U}_{0_X 0_X} + \frac{1}{4} Y^{11,L}_{0_X 1_X}
+ \frac{1}{4} Y^{11,L}_{1_X 0_X} -\frac{1}{4} Y^{11,U}_{1_X 1_X}\\
&-\frac{1}{2} Y^{11,U}_{0_Y 0_Y} -\frac{1}{2} Y^{11,U}_{1_Y 1_Y}.
\end{aligned}
\end{equation}
When six states are perfectly prepared, the average payoff from Eq. (\ref{six-state-payoff}) has the maximal value $0.5$, achieved by the identity channel.
The corresponding ideal EB bound is $0$, which can also be proven by Eq. (\ref{eq:expbound}).
In our experiment, based on the assumption that the referee has full knowledge of his states, he can use state tomography to determine the exact density matrices of the prepared states (as shown in Sec. \ref{state-tomo}).
Then, the experimental EB bound is numerically calculated as $0.047$, higher than the ideal bound of $0$.
Also, the maximal value of $I_{\cal N}$ is calculated as 0.451, which is slightly lower than $0.5$ of the ideal case due to inaccurate state preparations.
\subsubsection{The 4-state average payoff}
For the second kind, we use an information \textit{incomplete} set of states, i.e., eigenstates of the Pauli matrices $Z$ and $Y$, denoted as $\xi_{x},\psi_{y}$ $\in\left\{ \left|0\right\rangle _{Z},\left|1\right\rangle _{Z},\allowbreak \left|0\right\rangle _{Y},\left|1\right\rangle _{Y} \right\}$, respectively.
The corresponding payoff function is
\begin{equation}
\label{four-state-payoff}
\wp\left(0,\xi_{x},\psi_{y}\right)=
\begin{cases}
\frac{1}{4}, & \left(\xi_{x},\psi_{y}\right) \text{ anti-correlated in } Z;\\
-\frac{1}{4}, & \left(\xi_{x},\psi_{y}\right) \text{ correlated in } Z;\\
-\frac{1}{2}, & \left(\xi_{x},\psi_{y}\right)\text{ correlated in } Y ;\\
0, & \text{otherwise}.
\end{cases}
\end{equation}
According to the sign of $\wp(0,\xi_x,\psi_y)$ in Eq. (\ref{four-state-payoff}), the lower bound of the payoff value $I_{\cal N}$ can be written as
\begin{equation}
\begin{aligned}
I_{\cal N} \geqslant I_{\cal N}^L= & -\frac{1}{4} Y^{11,U}_{0_Z 0_Z} + \frac{1}{4} Y^{11,L}_{0_Z 1_Z}
+ \frac{1}{4} Y^{11,L}_{1_Z 0_Z} -\frac{1}{4} Y^{11,U}_{1_Z 1_Z}\\
&-\frac{1}{2} Y^{11,U}_{0_Y 0_Y} -\frac{1}{2} Y^{11,U}_{1_Y 1_Y}.
\end{aligned}
\end{equation}
When four states are perfectly prepared, the average payoff from Eq. (\ref{four-state-payoff}) has the maximal value $0.25$, and the ideal EB bound is $0$.
In our experiment, the experimental EB bound can be calculated as $0.025$, and the maximal value is $0.226$.
\section{Experiment details}
In this section, we introduce the basic techniques of the experiments, including feedback systems, the design and implementation of the decoherence channel, secure key rates using the noise-added fiber channel, and the quantum state tomography.
\subsection{Setup details and feedback systems}
Our experimental setup is composed of three portions: state preparation, detection, and the channel to be tested.
For state preparation, DFB1 (DFB2) at wavelength of $1550.12$ nm is used and directly modulated to emit pulses of $37.5$ MHz repetition rate, which are narrowed to $2.5$ ns at FWHM using an IM.
Time-bin states are created using an AMZI that separates the pulses at $6.5$ ns, and the basis of $Z$ or $X$($Y$) is chosen with the following IM.
Phase states are created using the FR, PM, and CIR, where the optical pulses travel through the PM twice for lower modulation voltage.
In addition, another IM is used to adjust the intensity difference between pulses of $Z$ and $X$($Y$) bases.
All IMs and PMs for preparing $\xi_x$ and $\psi_y$ are controlled by independent random numbers.
The pulses are lowered down to single-photon level with an EVOA, and are filtered with a $100$ GHz narrow pass-band filter for spectral noise.
In addition, we adopt feedback systems for phase, optical intensity, wavelength, and bias voltage
to ensure the stability of the entire system.
In order to obtain high interference visibility in the $X$ and $Y$ basis at the Bell-state measurement (BSM), a phase reference needs to be established between the two state preparation modules, i.e., the phase difference between the two asymmetric Mach-Zehnder interferometers (AMZI) should be $2\pi k$, with $k$ a positive integer. Pulses of wavelength $1550.12$ nm and frequency of $25$ MHz are sent from DFB3 through the two AMZI in sequence, as shown in Fig. 2 in the main text. Here, the pulses can be detected in three time-windows, where only those in the second time window show interference. A gated InGaAs single-photon detector (SPD) is used to detected the interference photons, and the $2\pi$ phase drift of the AMZIs is measured to be $\sim8$ minutes. For the experiment, a real-time phase feedback system is built. A fiber phase shifter is placed on the short arm of AMZI1 to maintain the interference steady at destructive interference, i.e., the interference photons are monitored with the SPD to remain minimum photon count. Thus, the phase reference between $\xi_{x}$ and $\psi_{y}$ is established.
For optimal BSM, firstly the arrival time of both $\xi_{x}$ and $\psi_{y}$ pulses are calibrated at the BSM site. To compensate the difference in arrival time to the BSM for pulses $\mathcal{N}(\xi_{x})$ and $\psi_{y}$, DFB1 and DFB2 alternatively send pulses to the SNSPDs.
Based on arrival time difference between the two pulses, and using a programmable delay chip to adjust the pulse delays, $\mathcal{N}(\xi_{x})$ and $\psi_{y}$ pulses are precisely overlapped at the BS.
The detection efficiency is $27\%$ and the dark count rate is $50$ counts per second for each SNSPD.
To achieve optimal trade-off between effective coincidence count rates and eliminating imperfect phase encodings at the pulse edges of $X$ and $Y$ basis, the effective BSM time window is set to $~85\%$. Secondly, the wavelengths of both $\xi_{x}$ and $\psi_{y}$ are required to be almost the same. Here, the temperature of DFB2 is scanned, and the Hong-Ou-Mandel dip is measured. The temperature is set to the value where minimum coincidence count of $55\%$ occurs. Thus, the wavelengths are optimal. Thirdly, considering the fluctuations of the source and bias voltages of IMs, the intensities of $\xi_{x}$ and $\psi_{y}$ pulses are calibrated every $\sim600$ seconds.
The clock reference for the entire system is established by sending synchronization laser pulses at $1570$ nm of $500$ kHz through a separate fiber to each of $\xi_x$ and $\psi_y$ state preparation modules.
Then, they are detected through a photodiode and the system repetition rate of $37.5$ MHz is generated.
With this configuration, we optimized pulse modulation from the modulators such that fast real-time data acquisition is realized.
\subsection{Realization of the decoherence channel through the Sagnac interferometer}
As discussed in the main text, the decoherence channel
preserves populations of the state in the $Z$ basis, while the coherence between them, i.e., the relative phases, is suppressed.
To realize such a channel, we design the Sagnac interferometer (SI) with a phase modulator (PM) to randomly eliminate the first or second time-bin when states in $X$ and $Y$ bases are prepared.
Initially, four input states $\left| 0 \right\rangle_X$, $\left| 1 \right\rangle_X$, $\left| 0 \right\rangle_Y$, $\left| 1 \right\rangle_Y$ are encoded in the phase between first and second time-bins, written as $|\sqrt{\alpha/2}\rangle ^{s} |e^{i\phi}\sqrt{\alpha/2}\rangle ^{l}$ with $\phi = 0,\pi,\pi/2$, $3\pi/2$ for simplicity. Here, $s$ and $l$ represent the short and long arm of the AMZI, respectively.
As shown in Fig.~2b of the main text, when entering the SI, $\xi_x$ is firstly split into two paths $a$ and $b$ by the beam splitter (BS), with path $a$ a shorter arrival time to the PM than that of path $b$. Then, $\xi_x$ is split into four pulses.
By precisely adjusting the length of paths $a$ and $b$, we set the arrival time for the pulses of two paths in a difference of $13.3$ ns. Now, the time-bins in the sequence of $|\sqrt{\alpha}/2\rangle _{a}^{s},|e^{i\phi}\sqrt{\alpha}/2\rangle _{a}^{l},|\sqrt{\alpha}/2\rangle _{b}^{s},|e^{i\phi}\sqrt{\alpha}/2\rangle _{b}^{l}$ are modulated by the PM, adding relative phases $\theta_{a}^{s},\theta_{a}^{l},\theta_{b}^{s},\theta_{b}^{l}$. Below we write the exact process for $\xi_x$ to transmit through the SI,
\begin{small}
\begin{equation*}
\begin{aligned}
\left|\sqrt{\frac{\alpha}{2}}\right\rangle ^{s}\left|e^{i\phi}\sqrt{\frac{\alpha}{2}}\right\rangle ^{l}
&\overset{BS}{\rightarrow}\left|i\frac{\sqrt{\alpha}}{2}\right\rangle _{a}^{s}\left|ie^{i\phi}\frac{\sqrt{\alpha}}{2}\right\rangle _{a}^{l}\left|\frac{\sqrt{\alpha}}{2}\right\rangle _{b}^{s}\left|e^{i\phi}\frac{\sqrt{\alpha}}{2}\right\rangle _{b}^{l}
\overset{PM}{\rightarrow}\left|ie^{i\theta_{a}^{r}}\frac{\sqrt{\alpha}}{2}\right\rangle _{a}^{s}\left|ie^{i\left(\phi+\theta_{a}^{s}\right)}\frac{\sqrt{\alpha}}{2}\right\rangle _{a}^{l}\left|e^{i\theta_{b}^{r}}\frac{\sqrt{\alpha}}{2}\right\rangle _{b}^{s}\left|e^{i\left(\phi+\theta_{b}^{s}\right)}\frac{\sqrt{\alpha}}{2}\right\rangle _{b}^{l}\\
&\overset{BS}{\rightarrow}\left| \left(e^{i\theta_{b}^{r}}-e^{i\theta_{a}^{r}} \right)\frac{\sqrt{\alpha}}{2\sqrt{2}}\right\rangle _{c}^{s}
\left| e^{i\phi} \left(e^{i\theta_{b}^{s}}-e^{i\theta_{a}^{s}} \right) \frac{\sqrt{\alpha}}{2\sqrt{2}}\right\rangle_{c}^{l}
\left|i \left(e^{i\theta_{a}^{r}}+e^{i\theta_{b}^{r}} \right) \frac{\sqrt{\alpha}}{2\sqrt{2}}\right\rangle_{d}^{s}
\left|i e^{i\phi} \left(e^{i\theta_{a}^{s}}+e^{i\theta_{b}^{s}} \right)\frac{\sqrt{\alpha}}{2\sqrt{2}}\right\rangle_{d}^{l}.
\end{aligned}
\end{equation*}
\end{small}
In the experiment, port $d$ is used for the input to the BSM. By adjusting the modulation voltages on the PM for $\theta_{a}^{s},\theta_{a}^{l},\theta_{b}^{s},\theta_{b}^{l}$, the channel can either model the identity channel or the fully decoherence channel.
Precisely, the identity channel is realized when the PM is turned off, i.e., adding phases $0,0,0,0$ in the same sequence. In this case, the state remains the same and is output in port $d$.
As for the fully decoherence channel, we randomly add phases in the sequence $0,0,0,\pi$ or $0,0,\pi,0$ with the same probability, such that the second or first time-bin is eliminated, respectively. The output of port $d$ is thus in the form of $|\sqrt{\alpha/2}\rangle^s$ or
$|\sqrt{\alpha/2}\rangle^l$
(up to an overall phase), respectively.
By using an independent random number string to control the PM, the identity channel and fully decoherence channel can be realized with probabilities $1-\gamma $ and $\gamma$, respectively. In this manner, decoherence channel of the form
\begin{equation}
\mathcal{D}_{\gamma}(\rho)=(1-\gamma)\rho+\gamma(|0\rangle\langle0|_{Z} \rho_{00}+|1\rangle \langle1|_{Z}\rho_{11}),
\end{equation}
can be constructed.
\subsection{Secure key rates using the noise-added fiber channel}
Since effects from noise on actual channels are extremely complicated and related to specific channel structures, application environment, detection process etc., for simplicity and without loss of generality, we design the additive noise in a fiber channel by combining photons of a continuous-wave (CW) source with the WCPs of $\xi_x$ into the untrusted measurement, as shown in Fig.~2c of the main text.
We adjust the intensity of the CW, such that different strengths of noise can be modeled.
We use the ratio of CW intensity to $\xi_x$ pulse intensity per second, denoted by $\beta$, to describe the strength of noise.
Since directly analyzing the effect of noise photons is extremely complicated, here, we use the secure key rate of quantum key distribution to confirm the correct certification of the tested channel against noise.
We suppose that $\xi_x$ and $\psi_y$ are prepared by two distinct users, namely Alice and Bob, and the BSM is performed by an untrusted third party Charlie in the usual measurement-device-independent quantum key distribution scenario. For both Alice and Bob, pulses of $Z$ basis with intensity $\mu$ are used for key generation, and pulses of $X$ basis are used for error estimation. The key rate for measurement-device-independent quantum key distribution is calculated by \cite{mdiqkd},
\begin{equation}\label{eq:keyrate}
R \geq Q^{11}_{Z,\mu\mu}(1 - H(e^{11}_X)) - Q^{\mu\mu}_Z f H(E^{\mu\mu}_Z).
\end{equation}
Here, $f$ is the error correction inefficiency factor, and $H(x) = -x \textrm{log}_2 x -(1-x) \textrm{log}_2(1-x)$ is the binary entropy function. $Q^{11}_{Z,\mu\mu}$, $Q^{\mu\mu}_Z$, $E^{\mu\mu}_Z$, and $e^{11}_X$ are the single photon gain, the total gain, the quantum bit error rate in the $Z$ basis, and the quantum bit error rate in the $X$ basis when $\xi_x$ and $\psi_y$ are both of single photons, respectively \cite{mdiqkd}. The experimental key rates in Fig.~4 of the main text are obtained by taking the measured $Q^{\mu\mu}_Z$ and $E^{\mu\mu}_Z$ into Eq.~(\ref{eq:keyrate}). The theoretical key rate (dotted) line is obtained by simulating Eq.~(\ref{eq:keyrate}) with our experimental parameters, where noise photons from the CW are treated as dark counts \cite{Mao2018QKD}.
\subsection{Robustness against flawed state preparation}\label{state-tomo}
As discussed in Sec. \ref{practical-bound}, to avoid falsely witnessing non-EB channels, the experimental bound is determined by the full knowledge of quantum questions. The referee can achieve this by performing state tomography of $\xi_{x}$ and $\psi_{y}$.
Generally, the density matrix of an arbitrary qubit state is in the form of
\begin{equation}
\rho = \frac{\mathbb{I}+\left\langle X\right\rangle X +\left\langle Y\right\rangle Y +\left\langle Z\right\rangle Z }{2},
\end{equation}
where $X$, $Y$, $Z$ are the three Pauli matrices, and $\left\langle X\right\rangle$, $\left\langle Y\right\rangle$, $\left\langle Z\right\rangle$ are expectation values when measuring the corresponding Pauli matrices.
In this experiment, the value $\left\langle Z\right\rangle$ is evaluated by a time-bin intensity ratio $r$, i.e., the ratio of the photon counts in the first time-bin window to that in the second.
To obtain $\left\langle X\right\rangle$ for each $\xi_{x}$ and $\psi_{y}$, the corresponding pulses prepared by one AMZI are sent through the other AMZI.
Photon pulses can be detected at both output ports of the final BS, where they are observed in three consecutive time windows using a gated SPD.
Interference is shown in the second time window, where the phase differs $\pi$ for two ports.
By using the photon counts of three time window in both BS outputs, values of $\left\langle X\right\rangle$ can be obtained.
For the value of $\left\langle Y\right\rangle$, we adjust the phase shifter of AMZI1 (see Fig.~2a of the main text), such that two output ports of the final BS correspond to the projection onto relative phases $\pi/2$ and $3\pi/2$ \cite{Sun2016}.
The constructed density matrices of $\tilde{\xi}_x$ and $\tilde{\psi}_y$ are taking into Eq.~(\ref{eq:expbound}), such that the experimental bound for all EB channels are determined. Here, we show the tomography results in Fig. \ref{fig:tomo} and list fidelity \cite{nielsen2002quantum} of our input states in Table \ref{tab:fidelity}.
\begin{figure}\centering
\includegraphics[width=\linewidth]{fig5fin}
\caption{\label{fig:tomo}The tomography results of quantum states used in this experiment.
In each figure, four blue bars represent the real part of four entries in the density matrix, while the red bars represent the imaginary part. Based on this description of the quantum states, the referee can determine the bound $C_{EB}$ of all EB channels.}
\end{figure}
\begin{table}[h]
\centering
\caption{Fidelity of input quantum states}
\begin{tabular}{c|cccccc}
\hline
\text{Fidelity}(\%) & $\left|0\right\rangle_Z$ & $\left|1\right\rangle_Z$ & $\left|0\right\rangle_X$ & $\left|1\right\rangle_X$ & $\left|0\right\rangle_Y$ & $\left|1\right\rangle_Y$ \\
\hline $\xi_{x}$ & 99.7 & 99.2 & 98.4 & 96.9 & 97.7 & 93.6\\
$\psi_{y}$ & 99.5 & 99.1 & 99.3 & 95.7 & 98.6 & 94.3\\
\hline
\end{tabular}
\label{tab:fidelity}
\end{table}
|
\section{Introduction}
Since the work of Young, Laplace, and Gauss, multiphase flow at the continuum scale has been modeled almost exclusively by the Young-Laplace (YL) law, imposed as a boundary condition for the Navier-Stokes (NS) equations at the fluid-fluid interface. In this work, we propose an alternative continuum description of the multiphase flow in the form of an integral of a molecular-force-like function with support $\varepsilon$ added to the Navier-Stokes momentum conservation equation.
At the nanoscale, multiphase flow is traditionally described by the equations of molecular dynamics (MD). MD simulations show a sharp density and pressure drop across the fluid-fluid interface in a region of around ten nanometers \cite{Masuda2011, Nakamura2011}. Outside of this region the pressure satisfies the YL law, i.e., the pressure difference across the interface is linearly proportional to the interface curvature.
Another molecular-scale feature of the fluid-fluid interface is the surface tension dependence on the curvature radius for curvature radii smaller than 10 nanometers \cite{Kashchiev2003}. As the curvature radius increases, surface tension asymptotically approaches its ``macroscopic'' value $\sigma_0$. Therefore, one can conclude that the NS-YL model is suitable for describing interfacial dynamics on scales larger than 10 nanometers given that the thermal fluctuations are properly accounted for \cite{Landau1987}. Comparisons between continuum and MD simulations have shown disagreement below a droplet diameter of 36 nanometers, with drastic differences at 10 nanometers \cite{Bardia2016}.
{ One example of a physical system that cannot be described by the YL law are nanobubbles. Bulk nanobubbles, also known as ultrafine bubbles, show extraordinarily long-term stability compared to that which would be expected from the Young-Laplace Law with macroscopic surface tension \cite{Jadhav2020}, up to three months in a laboratory setting \cite{Michailidi2020}. This long-term stability leads to a number of applications, for example long term stability of bulk nanobubbles is an advantage when used as an ultrasound contrast agents over ultrasound contrast agents with larger bubbles, which have a limited half-life \cite{Perera2017}. Additionally, nanobubbles have a higher surface area than macro-scale bubbles for the same volume fraction of bubbles, and change the physiochemical properties of the fluid-bubble system \cite{Ohgaki2010, Liu2013}. Bulk nanobubbles show strong promise for biomedical applications in ultrasound \cite{Hernandez2018, Abenojar2019}, radiofrequency ablation of tumors \cite{Perera2014}, and for mediating drug delivery \cite{Hernandez2017}, as well as water and waste treatment \cite{Agarwal2011, Uchida2011, Temesgen2017, Wang2019, Hu2018, Atkinson2019}, prevention of membrane and surface fouling and cleaning \cite{Chen2009, Zhu2016, Ghadimkhani2016}, froth flotation \cite{Fan2010}, and plant and animal growth \cite{Zhou2019, Ebina2013}. Numerical simulations of a large number of nanobubbles over large timescales are essential for these and other important applications.
}
On the molecular scale, surface tension results from the broken symmetry in the molecular interactions near the interface, i.e., the molecular forces acting between like molecules differ from forces acting between unlike molecules. In a similar manner, the molecular-like-forces generate surface tension in the non-local model.
Many interfacial phenomena are multiscale in nature. For example,
modeling colliding macroscale droplets requires resolving a large range of relevant length scales and topology changes that occur at the droplet interfaces and a thin fluid film forming between two colliding droplets, and is another example where the YL law fails under certain capillary numbers \cite{Jiang2007}. The interfaces of droplets become locally flat as they approach each other, so the surface tension force due to the YL law becomes zero.
The film of surrounding fluid forms between droplets and reaches $100-1000\AA$ before rupturing. This film drains as the droplets come closer to each other, eventually rupturing due to the inter-molecular van der Waals forces, which are absent in the macroscale models \cite{Jiang2007}.
In traditional local front capturing Level Set and volume-of-fluid methods, droplet coalescence will occur if the droplets are separated by less than about one grid point, a phenomenon known as numerical (or artificial) coalescence \cite{Coyajee2009}. Simulations of binary head-on collisions by \cite{Pan2005} using the Level Set method were unable to capture the small deformation regime of bubble merging because they could not resolve the film drainage and rupturing. The Coupled Level-Set/Volume-of-Fluid (CLSVOF) method was proposed to overcome numerical coalescence, at the expense of prohibiting even physically-feasible coalescence \cite{Sussman2000}. In the simulations by \cite{Pan2008}, coalescence occurred at timing determined by experiments or by a van der Waals force with augmented range such that the length scale was large enough to be resolved by the simulations, implemented by adding a surface force at the interface. This study, and others, showed that the behavior of the droplets after coalescence is sensitive to the timing of the front rupture, and experimental data may not be available in all cases to determine the coalescence time \cite{Pan2008, Mason2012}. To resolve this challenge, sub-grid-scale (SGS) models have been proposed, where a semi-analytical model for the thin film dynamics is coupled with the local model at the droplet lengthscale to determine if and when the front will rupture \cite{Mason2012, Liu2018}. This results in a predictive method, removing the necessity of prescribing the time for the fronts to merge. SGS models were considered in the CLSVOF method by \cite{Coyajee2009} and \cite{Kwakkel2013} and in the front tracking method by \cite{Tryggvason2010}. These models require coupling knowledge of the dynamics of the nano-scale gap with the macro-scale bubble dynamics as well as computationally expensive periodic searching through the domain to find interfaces close to collision \cite{Chan2018}. Another approach by \cite{Jiang2007} included van der Waals forces in the momentum conservation equation in addition to the YL surface tension force.
This approach was limited to the head-on collisions with equal sized droplets.
To demonstrate the multiscale nature of the non-local model, we present a semi-analytical steady-state solution for the fluid pressure across a fluid-fluid interface. This solution shows nanoscale behavior for the radius of curvature smaller than $3.5\varepsilon$ and macroscopic behavior (i.e., the solution follows the YL law) for the radius of curvature larger than $3.5\varepsilon$.
Using the non-local model, we perform numerical simulations of droplets under dynamic conditions, including a rising droplet, a droplet in a shear flow, and two colliding droplets in a shear flow, and compare results with standard NS model subject to the YL boundary condition at the fluid-fluid interface implemented via the Conservative Level Set (CLS) method. We find good agreement with the CLS method for a rising macroscopic droplet and a droplet in a shear flow. For a small capillary number (Ca = 0.24), we find that the non-local model predicts a microscopic droplet in shear flow to form ``ears''. The CLS method predicts the same (macroscopic) shape of the droplet regardless of the droplet size.
Finally, for colliding droplets in shear flow we find that the non-local model converges (with respect to the grid size) to the correct behavior, including (depending on capillary number) sliding, coalescing, and temporary bridging (coalescing and then separating) of two droplets. On the other hand, in our simulations the CLS method results are highly grid-size dependent.
\section{Non-local surface tension model}
We consider flow of two incompressible Newtonian fluids, denoted $\alpha$ and $\beta$ in a fixed domain $\Omega = \Omega_\alpha(t) \cup \Omega_\beta(t)$. The macroscopic (hydrodynamic) model for two-phase flow includes the continuity equation for phase $i = \alpha$ or $\beta$
\begin{equation}
\nabla \cdot \mathbf{u}_i = 0, \quad\mathbf{x}\in\Omega_i(t),
\end{equation}
and the momentum conservation equation
\begin{equation}
\rho_i \left( \frac{\partial \mathbf{u}_i }{\partial t} +\mathbf{u}_i \cdot \nabla \mathbf{u}_i \right) = -\nabla P_i + \nabla \cdot \boldsymbol{\tau}_i + \rho_i \mathbf{g}, \quad\mathbf{x}\in\Omega_i(t),
\label{Eq-Mom}
\end{equation}
subject to the YL boundary condition at the fluid-fluid interface
\begin{equation}\label{Young-Laplace}
(P_\alpha-P_\beta)\mathbf{n}= (\boldsymbol{\tau}_\alpha-\boldsymbol{\tau}_\beta)\cdot\mathbf{n} +\kappa \sigma_0 \mathbf{n},
\quad\mathbf{x}\in\Omega_\alpha \cap \Omega_\beta,
\end{equation}
where the subscript $i$ denotes phase $i$, $\rho_i$ is the density, $\mathbf{u}_i$ is the velocity,
$P_i$ is the pressure, $\mathbf{g}$ is the gravitational acceleration, $\tau_i = [\mu_i (\nabla \mathbf{u}_i + \nabla \mathbf{u}_i^T)]$ is the viscous stress tensor with the dynamic viscosity $\mu_i$, $\sigma_0$ is the macroscopic surface tension, and $\mathbf{n}$ is the normal vector.
To simplify numerical treatment of these equations, it is common to replace eqs. \ref{Eq-Mom} and \ref{Young-Laplace} with \cite{Brackbill1992}
\begin{equation}
\rho \left( \frac{\partial \mathbf{u} }{\partial t} +\mathbf{u} \cdot \nabla \mathbf{u} \right) = -\nabla P + \nabla \cdot \boldsymbol{\tau} + \rho \mathbf{g} + \mathbf{F}, \quad\mathbf{x}\in\Omega,
\label{Eq-Mom-Brackbill}
\end{equation}
where the surface force $\mathbf{F}$ is given by the YL law as
\begin{equation}
\mathbf{F} = \sigma_0 \kappa \hat{\bf{n}}
\label{eq:YLlaw}
\end{equation}
where $\kappa$ is the interface curvature.
We introduce the color function $\phi$:
\begin{equation}
\phi(\mathbf{x}) = \left\{ \begin{array}{ll}
0, & \mathbf{x} \in \Omega_\alpha, \\
1, & \mathbf{x} \in \Omega_\beta.
\end{array}
\right.
\label{equ:color_func}
\end{equation}
The color function is advected with the velocity $\mathbf{u}$ as $\frac {\partial \phi}{\partial t} + \mathbf{u} \cdot \nabla \phi = 0.$
Then, from \cite{Brackbill1992}, the force
\begin{equation}
\mathbf{F} =\sigma_0 \left(-\nabla \cdot \frac{\nabla \phi}{|\nabla \phi|}\right) \nabla \phi
\label{eq:YLlawvolume}
\end{equation}
gives the same total force as eq. \ref{eq:YLlaw}, but spread over the interface width.
Here, we propose to replace the YL definition of the surface tension force from eq. \ref{eq:YLlaw} with the non-local model
\begin{equation}
\mathbf{F} = -\int_\Omega s(\mathbf{x}, \mathbf{y}) f_\varepsilon(|\mathbf{x}-\mathbf{y}|) \frac{\mathbf{x}-\mathbf{y}}{|\mathbf{x}-\mathbf{y}|} \; d\mathbf{y}, \; \; \mathbf{x} \in \Omega,
\label{eq:Nonlocal_force}
\end{equation}
where $ f_\varepsilon(|\mathbf{x}-\mathbf{y}|)$ is the force shape function and $s(\mathbf{x}, \mathbf{y})$ is the force strength. We obtain eq. \ref{eq:Nonlocal_force} as the continuous limit of the so-called pairwise surface tension force that is used in multiphase Smoothed Particle Hydrodynamics \cite{Tartakovsky2005,Tartakovsky2016}, a fully Lagrangian particle method.
The force strength is given by
\begin{equation}
s(\mathbf{x}, \mathbf{y})= \left\{ \begin{array}{ll}
s_{\alpha\alpha}, & \mathbf{x} \in \Omega_\alpha, \; \mathbf{y} \in \Omega_\alpha, \\
s_{\alpha\beta}, & \mathbf{x} \in \Omega_\alpha, \; \mathbf{y} \in \Omega_\beta, \\
s_{\beta\beta}, & \mathbf{x} \in \Omega_\beta, \; \mathbf{y} \in \Omega_\beta. \\
\end{array}
\right.
\end{equation}
To ensure that $\sigma_0$ is positive, the coefficients must satisfy $ s_{\alpha\alpha} + s_{\beta\beta} > 2s_{\alpha\beta}$. For convenience, we take
$ s_{\alpha\alpha} = s_{\beta\beta} = 10^k s_{\alpha\beta}$ with $k = 3$. Then, $ s_{\alpha\alpha}$ and $ s_{\beta\beta}$ can be found as a function of $\sigma_0$:
\begin{equation}\label{s11}
s_{\alpha \alpha } = {s}_{ \beta \beta} =\frac1{2(1-10^{-k})} \frac{\sigma_0}{ \lambda},
\end{equation}
where
$ \lambda = \frac18 \pi \int \limits_0^\infty z^4
f_\varepsilon(z)dz
$
and
$
\lambda = \frac13 \int \limits_0^\infty z^3 f_\varepsilon(z)dz
$
in three and two spatial dimensions, respectively.
As in molecular dynamics, $f_\varepsilon(|\mathbf{x}-\mathbf{y}|)$ must be repulsive for small $|\mathbf{x}-\mathbf{y}|$, attractive for large $|\mathbf{x}-\mathbf{y}|$, and, for computational efficiency, should have compact support $h=O(\varepsilon)$ or become negligibly small for $|\mathbf{x}-\mathbf{y}| \geq h$.
Several forms of $f_\varepsilon$ have been proposed \cite{Tartakovsky2016}. Here, we use
\begin{equation}
f_\varepsilon(|\mathbf{x}-\mathbf{y}|) = |\mathbf{x}-\mathbf{y}| \left[-A e^{-\frac{|\mathbf{x}-\mathbf{y}|^2}{2\varepsilon_0^2}}+ e^{-\frac{|\mathbf{x}-\mathbf{y}|^2}{2\varepsilon^2}}\right]
\label{eq:forceshapefunction}
\end{equation}
and $\lambda = \frac{1}{2}\sqrt{2 \pi}(-A \varepsilon_0^5 + \varepsilon^5).$
A fundamental difference between the non-local model in eq. \ref{eq:Nonlocal_force} and the YL law is that the former has an internal length scale $\varepsilon$, while the latter does not have any internal length scale. Because of this, the YL law predicts the same behavior (for the same dimensionless numbers) regardless of the problem's length scale. In the following, we obtain an analytical solution for pressure that demonstrates that the non-local model behaves ``macroscopically'' (follows the YL law) on the scale larger than $\varepsilon$ and ``microscopically'' (deviates from the YL law in a way consistent with molecular dynamics simulations of droplets), otherwise.
Under static conditions, eqs. \ref{Eq-Mom-Brackbill} and \ref{eq:Nonlocal_force} can be solved analytically for a circular interface separating two fluids in two dimensions.
The pressure as a function of the distance $r$ from the droplet center can be found as \cite{Howard2019}:
\begin{equation}
{P(r)-P(\infty)
= -4 \pi( s_{aa} -s_{ab}) \left(\varepsilon^4 G(r, \varepsilon) - A\varepsilon_0^4 G(r, \varepsilon_0)\right)}
\label{eq:Pexact}
\end{equation}
where $G(r, \varepsilon) $ is given by eq. \ref{eqn:G}:
\begin{align}
G(r, \varepsilon) &= \begin{cases} e^{-\frac{a^2}{2\varepsilon^2}} \sum_{l = 0}^\infty \frac{1}{l!(l+1)!} \left( \frac{a^2}{2\varepsilon^2} \right)^{l+1} \left[\Gamma\left(l+1, \frac{a^2}{2\varepsilon^2}\right) -\frac{1}{2}\Gamma\left(l+1, \frac{r^2}{2\varepsilon^2}\right) \right] & r < a \\
e^{-\frac{a^2}{2\varepsilon^2}} \sum_{l = 0}^\infty \frac{1}{l!(l+1)!} \left( \frac{a^2}{2\varepsilon^2} \right)^{l+1} \left[\frac{1}{2}\Gamma\left(l+1, \frac{r^2}{2\varepsilon^2}\right) \right] & r \geq a. \end{cases} \label{eqn:G}
\end{align}
\begin{figure}
\centering
\sidesubfloat[]{
\includegraphics[width=0.35\linewidth]{Figures/Figure1a.eps}} \sidesubfloat[]{
\includegraphics[width=0.35\linewidth]{Figures/Figure1b.eps}}
\caption{Comparison of numerical integration and the analytical solution with $a/\varepsilon = 6$. (a) Pressure profile from eq. \ref{eq:Pexact}. (b) Pressure jump at the center of a circular droplet calculated from the analytic solution in eq. \ref{eq:Pexact}. The dashed line shows the linear value from the YL equation, eq. \ref{Young-Laplace}. The pressure jump agrees with the YL equation when the value of $\varepsilon$ is much smaller than the radius $a$, $\varepsilon \ll a$. As the curvature increases relative to $\varepsilon$ the pressure jump decreases relative to the one predicted by the YL law. The triangles represent values for which the non-local model gives non-physical values.}
\label{fig:analytic_profiles}
\end{figure}
The pressure in eq. \ref{eq:Pexact} as a function of $r$ is plotted in Fig. \ref{fig:analytic_profiles}$A$ for a droplet with the radius $a$. The pressure profile qualitatively agrees with behavior observed in molecular dynamics simulations \cite{Park2001, Masuda2011, Nakamura2011, Malek2018}. The pressure is zero far from the droplet, then at the interface the pressure drops and becomes negative, as seen in MD simulations. At the center of the droplet the pressure reaches a constant value. The pressure jump from inside to outside the interface is given by the YL law (eq. \ref{eq:YLlaw}).
Fig. \ref{fig:analytic_profiles}$B$ demonstrates that the pressure difference $P_{\varepsilon, in} - P_{\varepsilon, out}$ as a function of curvature
agrees with the YL law (eq. \ref{Young-Laplace}) for droplet radius $a > 3.5 \varepsilon$ ($P_{\varepsilon, in}$ and $P_{\varepsilon, out}$ are pressures inside and outside of the droplet at the distance greater than $3.5\varepsilon$ from the interface). For smaller droplets with $a< 3.5\varepsilon$, the pressure jump begins to deviate from the YL law.
For ``tiny'' droplets with $a< 0.3\varepsilon$, the pressure difference begins to decrease (the triangle symbols in Fig. \ref{fig:analytic_profiles}$B$). This indicates the limit of incompressible treatment of small (nano) droplets.
\section{Numerical implementation of the non-local model}
\label{sec:numerical}
Eq. \ref{Eq-Mom} is discretized with a prediction-correction scheme based on the finite volume method presented in \cite{Tryggvason2012}. A temporary velocity is first calculated without the pressure,
\begin{equation}
\frac{\mathbf{u}^*-\mathbf{u}^n}{\Delta t} = \nabla_h \cdot \mathbf{u}^n\mathbf{u}^n + \mathbf{g}
+\frac{1}{\rho^n} \mathbf{F}^n
+ \frac{1}{\rho^n}\nabla_h^2\mathbf{u}^n
\end{equation}
where $\mathbf{u}^n$ is the velocity field at time step $n$, $\mathbf{u}^*$ is the temporary velocity, $\mathbf{g}$ represents external forces such as gravity, and $\rho^n$ is the density field. The subscript $h$ denotes discrete finite volume operators. The pressure $p^n$ is then calculated so that $\mathbf{u}^{n+1}$ satisfies the incompressibility condition, $\nabla_h \cdot \mathbf{u}^{n+1} =0$, which gives
\begin{equation}
\nabla_h \cdot \left( \frac{1}{\rho^n} \nabla_h p^n \right) = \frac{1}{\Delta t} \nabla_h \cdot \mathbf{u}^*
\end{equation}
for the pressure and
\begin{equation}
\frac{\mathbf{u}^{n+1}-\mathbf{u}^*}{\Delta t} = \frac{-\nabla_h p^n}{\rho^n}
\end{equation}
for the updated velocity at time step $n+1$. We use a staggered mesh, with the velocity calculated on the mesh edges and pressure, density, viscosity, and the color function updated on the cell centers.
The pressure $p^n$ is solved using a successive over relaxation (SOR) scheme \cite{Tryggvason2012}. Following \cite{Tryggvason2012}, we set the time step to $\Delta t = \frac{\Delta x^2\min(\rho_\alpha, \rho_\beta)}{4.25\max(\mu_\alpha, \mu_\beta)}$.
We use a linear interpolation within elements and the color function at time step $n$, $\phi^n$, to discretize eq. \ref{eq:Nonlocal_force}:
\begin{align}\label{Cont-SPH-Mom-Sum-phi}
\mathbf{F}^n(\mathbf{x}) &= \int_{\Omega_\alpha^n\cup\Omega_\beta^n } s(\mathbf{x}, \mathbf{y}) f_\varepsilon(|\mathbf{x}-\mathbf{y}|) \frac{{\mathbf x}-{\mathbf y}}{|{\mathbf x}-{\mathbf y}|} d {\mathbf y} \nonumber \\
&= \sum_j f_\varepsilon(|{\mathbf x} - {\mathbf y}|) \frac{{\mathbf x}-{\mathbf y}}{|{\mathbf x}-{\mathbf y}|} V_j \left(s_{\alpha \alpha}[\phi^n({\mathbf x})\phi^n({\mathbf y}) +(1-\phi^n({\mathbf x}) )(1-\phi^n({\mathbf y}) )] + s_{\alpha \beta}[\phi^n({\mathbf x})(1-\phi^n({\mathbf y}))+ (1-\phi^n({\mathbf x}) )\phi^n({\mathbf y}) ]
\right) ,
\end{align}
where $V_j$ is the size of element $j$. In the numerical simulations, $\varepsilon/\varepsilon_0 = 2$,
$A = (\varepsilon/\varepsilon_0)^3$, and $s_{\alpha \alpha } = {s}_{ \beta \beta} = 10^5 {s}_{ \alpha \beta} $. Unless otherwise noted we take
$\varepsilon = 4\Delta x /3.5$ where $\Delta x$ is the grid resolution and $V_j = \Delta x^2$.
When advecting the color function, $\phi$, it is important to use an accurate method that is conservative and also preserves the sharpness of the front. As noted in \cite{Sethian2003}, implementation of high order upwind schemes is difficult on a staggered mesh. We use the method presented in \cite{Olsson2005} and updated in \cite{Olsson2007} for a conservative advection scheme of the color function. To update the color function, we first solve
\begin{equation}
\phi_t + \textbf{u}\cdot \nabla \phi = 0 \label{eq:advection}
\end{equation}
using the conservative Total Variation Diminishing (TVD) method with Superbee limiter from \cite{Olsson2005} with a second-order Runge-Kutta scheme in time.
As noted in \cite{Olsson2005}, with eq. \ref{eq:advection} alone the interface width and profile will not remain constant over the course of a simulation. To remedy this, we solve the compression-diffusion equation from \cite{Olsson2005} after eq. \ref{eq:advection}. The compression-diffusion equation is given by
\begin{equation}
\phi_\tau + \nabla \cdot (\phi(1-\phi) \mathbf{n}) = \kappa \Delta \phi \label{eq:CDEquation}
\end{equation}
where $\mathbf{n}$ is the normal of the interface and $\kappa = \frac{1}{2}(\Delta x)^{1-d}$. $d$ is a parameter chosen so that $d \ll 1$ \cite{Olsson2005}. In the paper we take $d = 0.$ The time parameter is denoted by $\tau$ to distinguish it from the simulation time step, and is chosen as $\Delta \tau = \frac{1}{2}(\Delta x)^{1+d}$. Eq. \ref{eq:CDEquation} is solved until the solution reaches steady state, defined by $\int | \phi^{n+1}-\phi^n| \leq \text{TOL}\cdot\Delta \tau $ for a given tolerance $\text{TOL}$, where $\phi^n$ denotes the solution after $n$ iterations. To update $\phi^{n+1}$ we use a second-order Runge-Kutta scheme. The intermediate compression step provides an artificial compression normal to the front interface to maintain a sharp front, with constant thickness proportional to $\kappa$, preventing diffusion as the front is advected. The added viscosity term, $\kappa \Delta \phi $, prevents discontinuities at the interface. The compression flux $(\phi(1-\phi) \mathbf{n}) $ acts in regions where $0 < \phi < 1$ and provides compression normal to the interface, sharpening the interface profile.
Because the area each phase is defined as the area inside and outside of the $\phi = 0.5$ contour, conservation of $\phi$ does not necessarily imply conservation of each phase. However, our results in sec. \ref{sec:rising_bubble} indicate that the area loss is minimal. Once the color function is calculated, the density and viscosity are found directly by
\begin{equation}
\mu^{n+1} =\mu_1 + (\mu_2-\mu_1) \phi^{n+1},
\end{equation}
\begin{equation}
\rho^{n+1} =\rho_1 + (\rho_2-\rho_1) \phi^{n+1}.
\end{equation}
The numerical algorithm described in this section is implemented in a parallel C++ code with OpenMP. In this paper, simulations with the non-local methods are compared with simulations calculated with the CLS method. To provide a direct comparison, the simulations using the CLS method are completed with the same code, with the surface tension force in eq. \ref{Cont-SPH-Mom-Sum-phi} replaced by eq. \ref{eq:YLlaw}.
\section{Numerical results}
In this section, we validate the numerical implementation of the non-local model against analytical solutions, solutions obtained with the CLS \cite{Osher1988, Sussman1998, Olsson2005,Harten1977, Olsson2007} finite-volume discretization of the Navier-Stokes equations subject to the YL boundary condition, and published numerical simulations. In the following, we refer to CLS simulations of the Navier-Stokes equations subject to the YL boundary condition as a local model. Comparisons with the results of the local model will be used to illustrate advantages of the proposed here non-local model.
\subsection{Static pressure in a droplet}
We first test the code against analytical non-local and YL solutions for a static droplet with radius $a$.
In Fig. \ref{fig:Pressure}$A$, the pressure across the center of the droplet is plotted for varying resolution $\Delta x$. As the resolution across the domain increases, $P_{\varepsilon, in} - P_{\varepsilon, out}$ converges to the value expected by the YL law, with an error of less than $1\%$ for $a / \Delta x = 64$. In Fig. \ref{fig:Pressure}$B$, the resolution remains constant while the parameter $\varepsilon$ is varied relative to $a$. Here, $P_{\varepsilon, in} - P_{\varepsilon, out}$ is in excellent agreement with the YL law when $a/\varepsilon > 3.5$.
\begin{figure}
\centering
\sidesubfloat[]{ \includegraphics[width=0.35\columnwidth]{Figures/Figure2a.eps}}
\sidesubfloat[]{
\includegraphics[width=0.35\columnwidth]{Figures/Figure2b.eps}}
\caption{(a) Pressure across the center of a circular droplet with varying resolution. The droplet radius $a$ is fixed as $a = 0.0625L_x$, where $L_x = L_y = 4.0$ are the dimensions of the domain. The parameter $\varepsilon$ in the force shape function is fixed at $\varepsilon = 0.2a$. (b) Pressure across the center of a circular droplet with varying radius of support for the force shape function. The droplet radius $a$ is fixed as $a = 0.125L_x$, where $L_x = L_y = 2.0$ are the dimensions of the domain, with resolution $\Delta x = \Delta y = 1/200$. In both cases the fluid parameters are set to $\mu_f = \mu_d = 0.15$, $\rho_f = 1.0$, $\rho_d = 2.0$, and $\sigma_0 = 5.0$. }
\label{fig:Pressure}
\end{figure}
\subsection{Rising bubble}
\label{sec:rising_bubble}
\begin{table}
\centering
\caption{Parameters for benchmark test cases 1 and 2 from \cite{Hysing2009}.}
\begin{tabular}{ccccccc}
\hline
Test case & $\rho_1$ & $\rho_2$ & $\mu_1$ & $\mu_2$ & $g$ & $\sigma$ \\
\hline
1 & 1000 & 100 & 10 & 1 & 0.98 & 24.5 \\
2 & 1000 & 1 & 10 & 0.1 & 0.98 & 1.96 \\
\hline
\end{tabular}
\label{tab:Hysingparams}
\end{table}
A set of quantitative benchmarks for a rising bubble was proposed in \cite{Hysing2009} for validation and comparison between numerical methods for interfacial flow. \cite{Hysing2009} considered a bubble rising due to buoyancy effects in a two-dimensional domain with $L_x = 1$, $L_y =2$, no-slip boundary conditions at the top and bottom walls, and free slip conditions on the vertical walls. The droplet has an initial radius $a = 0.25$ and is centered at $(x_c, y_c) = (0.5, 0.5)$. Values of the viscosity, density, surface tension, and gravitational acceleration are given in table \ref{tab:Hysingparams} for the two test cases, denoted Case 1 and Case 2. In \cite{Hysing2009}, three numerical discretizations of local models were compared, two Eulerian Level Set finite-element codes (TP2D, denoted by Group 1, and FreeLIFE, denoted by Group 2) and an arbitrary Lagrangian-Eulerian moving grid method (MooNMD, denoted by Group 3.) The results of \cite{Hysing2009} have been compared with other numerical local models by many subsequent authors, including a conservative Level Set method with front sharpening \cite{Strubelj2009}, the volume of fluid method implemented in OpenFOAM\textsuperscript{\textregistered} \cite{Klostermann2013}, a diffuse interface model \cite{Aland2012}, and a finite element based Level Set method \cite{Zahedi2012}. In \cite{Strubelj2009} and \cite{Klostermann2013}, the results from \cite{Hysing2009} were matched to within $2\%$ and $4\%$, respectively, while in \cite{Zahedi2012}, it was found that higher resolution in their model is needed to achieve the accuracy in \cite{Hysing2009}.
We run the non-local model for Case 1 and Case 2 until the dimensionless time $t=3$ with the parameters given in table \ref{tab:Hysingparams} and $\varepsilon = \frac{8}{3.5}\Delta x$, where $\Delta x = 1/80$, $1/160$, or $1/320$ is the grid spacing. The characteristic length scale is $L = 2a$, where $a$ is the bubble radius, and time scale $L/\sqrt{2ga}$ where $g$ is the gravitational velocity.
In \cite{Hysing2009}, data are provided for the bubble area $A(t)$, the center of mass $\mathbf{x}_c$
\begin{equation}
\mathbf{x}_c = (x_c(t), y_c(t)) = \frac{\int_{\Omega_2(t)} \mathbf{x} \; d\mathbf{x}}{\int_{\Omega_2(t)} 1 \; d\mathbf{x}},
\label{eq:cm}
\end{equation}
where $\Omega_2(t)$ is the domain occupied by the bubble, the rise velocity $\mathbf{v}_c$
\begin{equation}
\mathbf{v}_c = (u_c(t), {v}_c(t)) = \frac{\int_{\Omega_2(t)} \mathbf{v}(\mathbf{x},t) \; d\mathbf{x}}{\int_{\Omega_2(t)} 1 \; d\mathbf{x}},
\label{eq:vc}
\end{equation}
where $\mathbf{v} = (u, v)$ is the fluid velocity, and the circularity $\hat{c}(t)$
\begin{equation}
\hat c(t) = \frac{2 \pi a_{circle}(t)}{P_b (t)}
\label{eq:hatc}
\end{equation}
where $a_{circle}(t)=\sqrt{A(t)/\pi}$ and $P_b (t)$ is the bubble perimeter.
A comparison of the front locations given by \cite{Hysing2009} and from the non-local model at time $t = 3.0$ for Case 1 and Case 2 is given in Figs. \ref{fig:benchmark1_front}a and \ref{fig:benchmark2_front}a. For Case 1 there is excellent agreement between the non-local model and the benchmark data from \cite{Hysing2009}, both in terms of the final bubble location and shape. Case 2 is computationally more challenging due to the topological changes that occur as the bubble breaks up to form satellite droplets. The results from \cite{Hysing2009} disagree on the point of breakup and the final bubble shape, as shown in Fig. \ref{fig:benchmark1_front}$c$.
Tables \ref{tab:Hysingerror} and \ref{tab:Hysingerror2} give values for Case 1 and Case 2 of $y_c(t=3)$, the maximum rise velocity $v_{c,max}$ and the time $t_{v_{c, max}}$ at which the maximum occurs, and the minimum circularity $\hat c_{min}$ and the time $t_{\hat c_{min}}$ at which the minimum occurs for the range of values presented by the three groups in \cite{Hysing2009} and the non-local model. Our results show that the non-local model converges to the benchmark values as its resolution increases. For Case I is an excellent agreement between the non-local models and the results from \cite{Hysing2009}, with a relative error for $\Delta x = 1/320$ of $0.02\%$ for $y_c(t = 3.0)$, $0.04\%$ for ${v}_{c, max}$, $0.96\%$ for $t_{{v}_{c, max}}$, $0.488\%$ for $\hat c_{min}$, and $1.19\%$ for $t_{\hat c_{min}}$. In each case the error is less than or on the same order of magnitude as the errors found by \cite{Strubelj2009} and \cite{Klostermann2013}. The inset of Fig. \ref{fig:benchmark1_front}$d$ shows that the final center of mass in the vertical direction $y_c$ for the non-local model with $\Delta x= 1/320$ is bounded by the benchmark values from \cite{Hysing2009}. For Case 2 the final bubble shape after breakup is inconclusive, however Table \ref{tab:Hysingerror2} shows good agreement between the non-local model and \cite{Hysing2009}. Due to computational limitations, we are unable to run higher resolution simulations of Case 2, which may improve the agreement between \cite{Hysing2009} and the non-local model.
While the numerical method used for advecting the color function $\phi$ is conservative, the bubble area is calculated as the contour corresponding to $\phi = 0.5$, and the area inside this contour is not necessarily conserved. Nevertheless, we see good conservation of the bubble area to within $0.226\% $ for Case 1 with $\Delta x = 1/ 80$, $0.077\%$ for $\Delta x = 1/ 160$, and $0.027\%$ for $\Delta x = 1/ 320$. The change in area for Case 2 is $0.283\%$ for $\Delta x = 1/ 80$ and $0.116\%$ for $\Delta x = 1/ 160$. The change in the bubble area is plotted in Figs. \ref{fig:benchmark1_front}$b$ and \ref{fig:benchmark2_front}$b$ for Case 1 and Case 2, respectively.
Figs. \ref{fig:benchmark1_front}$e$ and \ref{fig:benchmark2_front}$e$ show that the initial circularity is slightly under the expected value of one for a circular droplet at the beginning of the non-local simulations. This occurs because the droplet originally is fit to the square mesh. Over the first several time steps the compression algorithm for the advection of the color function rounds the edges of the droplet increasing the circularity to one.
\begin{table}
\centering
\caption{Comparison between benchmark values from the local models in \cite{Hysing2009} and the non-local model with $\Delta x = 1/80, 1/160,$ and $1/320$. The definitions for the center of mass $y_c$, rise velocity ${v}_{c}$, and circularity $\hat c$ are given in eqs. \ref{eq:cm}--\ref{eq:hatc}. }
\begin{tabular}{l c c c c }
\hline
& $\Delta x = \frac{1}{80}$ & $\Delta x = \frac{1}{160}$ & $\Delta x = \frac{1}{320}$ & \cite{Hysing2009} \\
\hline
$y_c(t = 3.0)$ & 1.0863 & 1.0824 & 1.0807 & 1.081 $\pm$ 0.001 \\
$v_{c, max}$ & 0.2438 & 0.2426 & 0.2418 & 0.2419 $\pm$ 0.0002 \\
$t_{v_{c, max}}$ & 0.9199 & 0.9030 & 0.9174 & 0.9263 $\pm$ 0.005 \\
$\hat c_{min}$ & 0.8961 & 0.8967 & 0.8968 & 0.9012 $\pm$ 0.0001 \\
$t_{\hat c_{min}}$ & 1.9206 & 1.9314 &1.9125 & 1.89 $\pm$ 0.015 \\
\hline
\end{tabular}
\label{tab:Hysingerror}
\end{table}
\begin{table}
\centering
\caption{Comparison between values from \cite{Hysing2009} and the non-local model with $\Delta x = 1/80$ and $1/160$. The definitions for the center of mass $y_c$ and the rise velocity ${v}_{c}$ are given in eqs. \ref{eq:cm} and \ref{eq:vc}. }
\begin{tabular}{c c c c }
\hline
& $\Delta x = \frac{1}{80}$ & $\Delta x = \frac{1}{160}$ & \cite{Hysing2009} \\
\hline
$y_c(t = 3.0)$ & 1.102 & 1.119 & 1.134 $\pm$ 0.009 \\
$v_{c,max}$ & 0.250 & 0.250 & 0.252 $\pm$ 0.002 \\
$t_{v_{c, max}}$ & 0.674 & 0.750 & 0.731$\pm$ 0.003 \\
\hline
\end{tabular}
\label{tab:Hysingerror2}
\end{table}
\begin{figure}
\centering
\begin{minipage}{.3\textwidth}
\sidesubfloat[]{ \includegraphics[width=\linewidth]{Figures/Figure4a.eps}}
\end{minipage}
\begin{minipage}{0.65\textwidth}
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure5a.eps}}
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure5b.eps}}\\
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure5c.eps}}
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure5d.eps}}
\end{minipage}
\caption{ Case 1. (a) Location of the $\phi = 0.5$ contour at $t = 3.0$ for the non-local model with grid spacing $\Delta x = 1/320$ compared with the front location from \cite{Hysing2009} Group 1 for cell size $h = 1/320$. (b-e) Comparison for Case 1 between the results of the non-local model with grid spacing $\Delta x = 1/80, 1/160$, and $1/320$ and the results from \cite{Hysing2009} Group 1 (cell size $h = 1/320$), Group 2 (element mesh size $h = 1/160$), and Group 3 (900 degrees of freedom on the interface). (b) Change in bubble area $A(t) - A(t = 0)$, (c) rise velocity $v_c$ given by eq. \ref{eq:vc}, (d) center of mass $y_c$ given by eq. \ref{eq:cm}, and (e) bubble circularity $\hat c$ given by eq. \ref{eq:hatc}. The parameters for Case 1 and 2 are given in table \ref{tab:Hysingparams}. (Color online.)
}
\label{fig:benchmark1_front}
\end{figure}
\begin{figure}
\centering
\begin{minipage}{.3\textwidth}
\sidesubfloat[]{ \includegraphics[width=\linewidth]{Figures/Figure4b.eps}}
\end{minipage}
\begin{minipage}{0.65\textwidth}
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure6a.eps}}
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure6b.eps}}\\
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure6c.eps}}
\sidesubfloat[]{ \includegraphics[width=0.45\columnwidth]{Figures/Figure6d.eps}}
\end{minipage}
\caption{Case 2. (a) Location of the $\phi = 0.5$ contour at $t = 3.0$ for the non-local model with grid spacing $\Delta x = 1/160$ compared with the front location from \cite{Hysing2009} for Group 1 (cell size $h = 1/320$), Group 2 (element mesh size $h = 1/160$), and Group 3 (900 degrees of freedom on the interface). (b-e) Comparison for Case 2 between the results of the non-local model with grid spacing $\Delta x = 1/80$ and $1/160$ and the results from \cite{Hysing2009} Group 1 (cell size $h = 1/320$), Group 2 (element mesh size $h = 1/160$), and Group 3 (900 degrees of freedom on the interface). (b) Change in bubble area $A(t) - A(t = 0)$, (c) rise velocity $v_c$ given by eq. \ref{eq:vc}, (d) center of mass $y_c$ given by eq. \ref{eq:cm}, and (e) bubble circularity $\hat c$ given by eq. \ref{eq:hatc}. The parameters for Case 1 and 2 are given in table \ref{tab:Hysingparams}. (Color online.)
}
\label{fig:benchmark2_front}
\end{figure}
\subsection{Circular bubble in shear flow}
\label{sec:shear}
A ``macroscopic'' spherical bubble in a linear shear flow $\mathbf{u} = (\dot \gamma y, 0)$ will form a steady elongated ellipsoidal shape. Denoting the major and minor axes of the ellipse as $l$ and $b$, the deformation ratio $D = \frac{l - b}{l + b}$ of an initially spherical bubble with radius $a$ was found by \cite{Taylor1932, Taylor1934} using the local model as
\begin{equation}
D = \text{Ca} \frac{(19 \lambda + 16)}{(16 \lambda + 16)}
\end{equation}
where $\lambda = \mu_d/\mu_f$ is the viscosity ratio between the drop and the surrounding fluid and $\text{Ca} = \mu_f a \dot \gamma / \sigma_0$ is the capillary number, the ratio of the magnitude of the viscous forces to the capillary forces. This formula has been validated for periodic suspensions of two-dimensional macroscopic droplets in a channel \cite{Zhou1993}, and corrections have been made for macroscopic bubbles in confinement \cite{Ioannou2016}. The three-dimensional results of \cite{Taylor1934} have been shown to agree well with two-dimensional simulations for small and moderate capillary number $\text{Ca} < 1$ \cite{Zhou1993, Guido1998}. The results of \cite{Taylor1932, Taylor1934} have been extended to second order \cite{Guido2011, Chaffey1967} to describe the angle the major axis of the deformed elliptical bubble makes with the horizontal axis, called the orientation angle and denoted by $\theta$, to give
\begin{equation}
\theta = \frac{\pi}{4} - \text{Ca} \frac{(19 \lambda + 16)(2 \lambda + 3)}{80( \lambda + 1)}.
\end{equation}
We simulate a macroscale bubble ($a/\varepsilon = 5$) in shear flow with the non-local and local (CLS) models for $\text{Ca} = 0.24$ and $\lambda = 1.4$ to provide a comparison to the experiments of \cite{Guido1998}. Results for the bubble deformation, $D$, and the location of the fronts are given in Fig. \ref{fig:Shear_D}. At times $0 \leq t \leq 4.0$, the non-local and local models agree well with the analytical results from \cite{Taylor1932, Taylor1934} and the experimental results \cite{Guido1998}. The orientation angle $\theta$ agrees as well with the second-order analytical results from \cite{Guido2011}. The relative error between the exact deformation, $D_{exact} = 0.26625$, and the non-local model with $a/\varepsilon = 5.0$ is $0.74\%$, and the relative error for the local model is $0.39\%$. The error between the analytical value for the orientation angle $\theta_{exact}$ and the non-local model is $4.13\%$, compared with $7.15\%$ for the CLS method.
Next, we model a ``microscopic'' bubble with the initial radius $a/\varepsilon = 2.5$, $\text{Ca} = 0.24$, and $\lambda = 1.4$, the same $\text{Ca}$ and $\lambda$ values as in the simulations of the macroscopic bubble. We observe that the deformation of the microscopic bubble is significantly larger than that of the macroscopic bubble and the resulting bubble has a pronounced sigmoid shape, see Fig. \ref{fig:Shear_D}$C$. In this case, the curvature of the front is significantly lower at the rounded corners of the droplet, which in turn decreases the force due to the surface tension at those points. The droplet then continues to elongate, coming into equilibrium with a higher deformation than for macroscale droplets. Mesoscale Dissipative Particle Dynamics (DPD) simulations of a spherical microscale droplet in shear flow also predicted a larger droplet deformation for capillary number $\text{Ca} > 0.2$ than that predicted by the Taylor theory \cite{Chen2004, Pan2014}.
The deformation values found in these DPD studies correspond well with the deformation in the non-local model for microscale bubble. Note that the local model, the behavior depends only on $\text{Ca}$ and $\lambda$ and not on the size of the bubble. {We examined the droplet circulation for each case and found very similar results for all simulations in Fig. \ref{fig:Shear_D}.}
\begin{figure}
\centering
\sidesubfloat[]{
\includegraphics[width=0.4\linewidth]{Figures/Figure7a.eps}}
\sidesubfloat[]{
\includegraphics[width=0.4\linewidth]{Figures/Figure7b.eps} } \\
\sidesubfloat[]{
\includegraphics[width=0.45\linewidth]{Figures/Figure7c.eps}}
\caption{(a) Bubble deformation for a droplet with $\text{Ca} = 0.24$ and $\lambda = 1.4$. Results are compared to the steady state results from \cite{Taylor1932} and experimental results from \cite{Guido1998} (open and closed symbols correspond to different digitization settings \cite{Guido1998}). {(b) Bubble orientation angle, $\theta$, compared with the analytical solution in \cite{Guido2011}. } (c) Location of the $\phi = 0.5$ contour for the non-local and CLS methods. The non-dimensional strain $\gamma$ is defined as $\gamma = \dot \gamma t$. }
\label{fig:Shear_D}
\end{figure}
\subsection{Droplet collisions in shear flow}
In this section we use the non-local model and the local CLS model to simulate the collision of two equal size droplets in shear flow.
This system has been studied previously, both experimentally and, more recently, numerically. Results have shown that the droplet behavior falls into one of three regimes: at low capillary number the droplets coalesce to form one large droplet, and at higher capillary number the droplets slide past each other \cite{Guido1998a, Chen2009, Shardt2013}. In some cases, particularly with small droplets, a third regime is possible at moderate capillary number where the droplets temporarily coalesce before breaking apart \cite{Shardt2013}. These three regimes are illustrated in Fig. \ref{fig:two_shear_sample_traj}, simulated with the non-local model for $a = 0.15$ at three capillary numbers.
At the lowest capillary number, the droplets coalesce, forming one large droplet. At long times, this large coalesced droplet will find a final shape as described for a single droplet in shear flow in sec. \ref{sec:shear}. As the capillary number increases there is a transitional regime where the droplets coalesce and then separate, known as temporary bridging. Finally, at high enough capillary number the droplets slide past each other without coalescence.
In \cite{Shardt2013}, all three regimes were simulated using a three-dimensional free-energy lattice Boltzmann method (LBM). The LBM offers an advantage over many front capturing models, including the Volume of Fluid (VoF) method and CLS method, where coalescence of droplets will depend on the grid resolution. As the grid is further refined, the film between the droplets will be better resolved, which increases the time needed for the film to fully drain and delays coalescence. \cite{Shardt2013} shows the LBM does offer grid independent results, but the behavior depends on the thickness of the diffuse droplet interface relative to the droplet size.
Here, we choose parameters to match the results in \cite{Shardt2013}, including the Reynolds number, $Re = 1$. In comparison, the Reynolds number in the experiments in \cite{Chen2009} is $Re < 10^{-7}.$ Further studies are needed to provide a detailed understanding of how the coalescence behavior may depend on the Reynolds number. In our simulations, the droplets have initial separation in both the $x$ and $y$ directions given by $| x_1-x_2| / a = 2.52$ and $| y_1-y_2| / a = 0.78$, where $(x_1, y_1)$ and $(x_2, y_2)$ are the centers of the two droplets. The relative channel width is $H/a = 5.1282$. The droplets have the same density and viscosity as the surrounding fluid, with $\rho = 6\mu$. We find that the transitions between the three regimes occur at capillary numbers of the same order of magnitude as presented by \cite{Shardt2013}, although we do not expect exact agreement due to the differences between two-dimensional and three-dimensional simulations.
\begin{figure}
\subfloat[][$\text{Ca} = 0.15$: Coalescence]{
\includegraphics[width=0.25\linewidth]{Figures/Figure8a1.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure8a2.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure8a3.eps} }\\
\subfloat[][$\text{Ca} = 0.25$: Temporary bridge] {
\includegraphics[width=0.25\linewidth]{Figures/Figure8b1.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure8b2.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure8b3.eps}}\\
\subfloat[][$\text{Ca} = 0.35$: Sliding]{
\includegraphics[width=0.25\linewidth]{Figures/Figure8c1.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure8c2.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure8c3.eps}}
\caption{Example front locations illustrating the three typical trajectories of two droplets in shear flow. (a) Coalescence occurs for $\text{Ca} = 0.15$. (b) A temporary bridge forms for $\text{Ca} = 0.25$. (c) No coalescence occurs for $\text{Ca} = 0.35$, and the droplets slide past each other. All simulations have resolution $\Delta x = 1/80$, $a = 0.15$, and $\varepsilon = 0.025$, which means $a/\varepsilon = 6$ and $\varepsilon/\Delta x = 2$. (Movie S1). (Color online.) }
\label{fig:two_shear_sample_traj}
\end{figure}
\begin{figure}
\subfloat[][CLS method, $\kappa = \Delta x/2$]{
\includegraphics[width=0.17\linewidth]{Figures/Figure9a1.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9a2.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9a3.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9a4.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9a5.eps}
\includegraphics[width=0.12\linewidth]{Figures2/Figure9clegend.eps}}\\
\subfloat[][{CLS method, $\kappa = 0.0125$}]{
\includegraphics[width=0.17\linewidth]{Figures2/Figure9c1.eps}
\includegraphics[width=0.17\linewidth]{Figures2/Figure9c2.eps}
\includegraphics[width=0.17\linewidth]{Figures2/Figure9c3.eps}
\includegraphics[width=0.17\linewidth]{Figures2/Figure9c4.eps}
\includegraphics[width=0.17\linewidth]{Figures2/Figure9c5.eps}
\includegraphics[width=0.12\linewidth]{Figures2/Figure9clegend.eps}}\\
\subfloat[][Non-local method]{
\includegraphics[width=0.17\linewidth]{Figures/Figure9b1.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9b2.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9b3.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9b4.eps}
\includegraphics[width=0.17\linewidth]{Figures/Figure9b5.eps}
\includegraphics[width=0.12\linewidth]{Figures2/Figure9clegend.eps}}
\caption{Collision trajectories for two droplets with resolution $\Delta x = 1/40, 1/80, 1/160$, and $1/240$ in shear flow with $\text{Ca} = 0.1$ calculated using the (a) CLS method with smoothing length $\kappa = \Delta x/2$, {(b) CLS method with fixed smoothing length $\kappa = 0.0125$,} and (c) non-local method with $\varepsilon = 0.025$. The times between Figures are not uniform and are chosen to be representative of the droplet behavior. The CLS method shows significantly different behavior, with simulations with finer resolution coalescing at later times, and not at all when $\Delta x = 1/240$. In contrast, the CLS simulations converge with respect to the grid size. Small satellite droplets can form at the intersection of the two droplets for the non-local model. (Movies S2 and S3).}
\label{fig:two_shear_res10}
\end{figure}
A key advantage of the non-local model for this problem is that droplet coalescence is controlled by the parameter $\varepsilon$, not the grid resolution. In \cite{Shardt2013}, the critical capillary number for the transition from coalescing to sliding droplets was found to decrease with the droplet radius, and follow a power-law with exponent that increases with increasing P\'{e}clet number. In our CLS simulations, the droplets coalesce at different times depending on the resolution and do not coalesce at all at the highest resolution considered with interface smoothing length $\kappa = \Delta x/2$, as shown in Fig. \ref{fig:two_shear_res10}$A$ at $\text{Ca} = 0.1$ and Fig. \ref{fig:two_shear_res15}$A$ at $\text{Ca} = 0.15$. This is because in the local model, the droplets will coalesce when they are within one grid point and there are no forces interacting between the droplets to assist coalesce when the droplets are separated by one more than one grid point. {When the smoothing length is kept fixed, $\kappa = 0.0125$, the droplets continue to coalesce at slightly different times, although the final behavior is the same, as shown in Fig. \ref{fig:two_shear_res10}$B$ at $\text{Ca} = 0.1$.}
In contrast, when the resolution is sufficient ($\Delta x \leq 1/80$), the behavior of the non-local method is similar regardless of resolution as long as $\varepsilon$ is kept constant with respect to the droplet radius as the mesh is refined. This is demonstrated in Fig. \ref{fig:two_shear_res10}$C$ at $\text{Ca} = 0.1$ and Fig. \ref{fig:two_shear_res15}$B$ at $\text{Ca} = 0.15$. At $\text{Ca} = 0.1$, small satellite droplets form as two droplets coalesce. Similar phenomenon can be seen in \cite{Shardt2013}.
\begin{figure}
\subfloat[][CLS method]{
\includegraphics[width=0.25\linewidth]{Figures/Figure10a1.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure10a2.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure10a3.eps}
\includegraphics[width=0.2\linewidth]{Figures/Figure10alegend.eps}}\\ \subfloat[][Non-local method]{
\includegraphics[width=0.25\linewidth]{Figures/Figure10b1.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure10b2.eps}
\includegraphics[width=0.25\linewidth]{Figures/Figure10b3.eps}
\includegraphics[width=0.2\linewidth]{Figures/Figure10blegend.eps}}
\caption{Collision trajectories for two droplets with resolution $\Delta x = 1/40, 1/80$, and $1/160$ with $\text{Ca} = 0.15$ calculated using the (a) CLS method and (b) non-local method with $\varepsilon = 0.025$.}
\label{fig:two_shear_res15}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.16\linewidth]{Figures/Figure11a.eps}
\includegraphics[width=0.16\linewidth]{Figures/Figure11b.eps}
\includegraphics[width=0.16\linewidth]{Figures/Figure11c.eps}
\includegraphics[width=0.16\linewidth]{Figures/Figure11d.eps}
\includegraphics[width=0.16\linewidth]{Figures/Figure11e.eps}
\includegraphics[width=0.12\linewidth]{Figures/Figure11legend.eps}
\caption{Collision trajectories for two droplets in shear flow with $\text{Ca} = 0.1$ calculated using the non-local method with varying values of $\varepsilon$. The times between figures is not uniform and are chosen to be representative of the droplet behavior. The values $a/\varepsilon = 20, 15, 12, 6$, and $2.5$ correspond to $\varepsilon = 0.0075, 0.01, 0.0125, 0.025$, and $0.06,$ respectively. The resolution is $\Delta x = 1/160$, so $\varepsilon/ \Delta x = 1.2, 1.6, 2.0, 4.0$, and $9.6$. (Movie S4).}
\label{fig:two_shear_eps}
\end{figure}
\begin{figure}
\centering
\sidesubfloat[]{
\includegraphics[width=0.35\linewidth]{Figures/Figure12a.eps}}
\sidesubfloat[]{ \includegraphics[width=0.35\linewidth]{Figures/Figure12b.eps}}
\caption{Front location and color function for two droplets with (a) the non-local method with $\varepsilon = 0.015$ ($a/\varepsilon = 10$), and (b) the non-local method with $\varepsilon = 0.025$ ($a/\varepsilon = 6$). At the same time, the droplets are closer together in the simulation with a higher value of $\varepsilon$. In both (a) and (b) the resolution is $\Delta x = 1/160$ and $a = 0.15$. (Color online.)}
\label{fig:two_shear_cf}
\end{figure}
The effect of $a/\varepsilon$ (the relative droplet size) is studied in Fig. \ref{fig:two_shear_eps}, which shows the droplet trajectories for five values of $\varepsilon$ at resolution $\Delta x = 1/160$ and droplet radius $a = 0.15$. The resulting droplets range from micro-droplets ($a/\varepsilon = 2.5$) to mesoscopic droplets ($a/\varepsilon = 6, 10$, and $12$) and macroscopic droplets ($a/\varepsilon = 15$, and $20$).
We find that the relatively smaller droplets (i.e., droplets with smaller $a/\varepsilon$) coalesce sooner. The droplets with $a/\varepsilon \geq 15$ do not coalesce during the simulation and instead slide past each other. The behavior of macroscopic bubbles does not depend on the $a/\varepsilon$ as long as $a/\varepsilon\geq15$.
The micro-scale droplets with $a/\varepsilon = 2.5$ display distinct characteristics, although the overall behavior is similar to that of the mesoscopic droplets. The fronts between these droplets flatten significantly at $\dot \gamma t = 1.2$, to a much larger degree than is seen in the other droplets. This results in many satellite droplets forming, which then migrate to the edge of the coalesced droplet over time. A zoomed-in view of the front locations and the corresponding color functions for macroscopic droplets can be seen in Fig. \ref{fig:two_shear_cf} for $a/\varepsilon = 10$ and $6$, showing that the front flattening occurs to a lesser degree than the micro-scale droplet in Fig. \ref{fig:two_shear_eps} (see Fig. \ref{fig:two_shear_cf}$B$.) The effect of increasing $\varepsilon$ relative to the droplet radius, for the same resolution and time, is to cause the two droplets to be significantly closer together leading to earlier coalescence.
An example of the coalescence process with the non-local simulation is shown in detail in Fig. \ref{fig:cf_coalescence}, which plots the droplet fronts as well as the color function. In this example $ \varepsilon/\Delta x = 2.4$, and $a/\varepsilon = 10$. When the droplets are sufficiently close an asymmetry forms in the front which becomes the initial location of coalescence. As the simulation continues, this neck widens until the droplets form a single large droplet.
\begin{figure}
\centering
\includegraphics[width=0.19\linewidth]{Figures/Figure13a.eps}
\includegraphics[width=0.19\linewidth]{Figures/Figure13b.eps}
\includegraphics[width=0.19\linewidth]{Figures/Figure13c.eps}
\includegraphics[width=0.19\linewidth]{Figures/Figure13d.eps}
\includegraphics[width=0.19\linewidth]{Figures/Figure13e.eps}
\caption{Examples of the color function and front locations for two coalescing droplets with $\varepsilon = 0.015$, $\Delta x = 1/160$, and $a = 0.15$, giving $a/\varepsilon = 10$ and $\varepsilon/\Delta x = 2.4$. (Color online.)}
\label{fig:cf_coalescence}
\end{figure}
\section*{Conclusion}
In this paper we propose a non-local PDE model for multiphase flow that replaces the Young-Laplace law and is valid at both nano and macroscopic scales.
Our model allows for simulations of multiscale systems with curvature radii ranging from nano to macro (micron and larger) scales, without the need to couple MD simulations with continuum NS models. Nanoscale multiphase flows, including nanobubbles, are important in biomedical applications, water and waste treatment, and membrane and surface defouling and cleaning. These applications typically include suspensions of many densely packed nanobubbles over macroscopic length scales, so the ability to use a continuum NS model will allow for efficient simulation of such systems in a way not possible with MD.
The non-local model limits numerical error from interpolating a sharp interface across a grid. Calculating the surface tension by integrating over a local neighborhood at each point on the surface reduces the spurious, or parasitic, currents that occur in other methods. It also handles merging interfaces without the behavior depending on the grid resolution. Instead, the dynamics are controlled by the relative curvature radius $a/\varepsilon$, where $\varepsilon$ is a parameter in the non-local surface tension force. The parameter $\varepsilon$ defines the relative width of the interface. Physically, the interface width is of the order of the support of the molecular forces. Therefore, when modeling nanodroplets, $\varepsilon$ should be on the order of the support of molecular forces, i.e., on the order of several $r{A}$. The parameter $\varepsilon$ also define the resolution of numerical simulations as the grid size should be smaller than $\varepsilon$. Therefore, using $\varepsilon$ on the order of several {\r{a}ngstr\"{o}ms} for modeling macroscopic droplets is computationally infeasible. Our results show that interfaces behave macroscopically when the radius of curvature larger than $3.5\varepsilon$ and the interface merging dynamics becomes independent of $\varepsilon$ for $a/\varepsilon >15$. Thus, it is sufficient to set $\varepsilon = a/3.5$ for modeling individual macroscopic droplets and $\varepsilon = a/15$ for modeling multiple interacting droplets.
When compared with existing benchmarks for macroscale droplets, the proposed non-local method matches or exceeds the accuracy of local numerical methods for multiphase flow. However, the non-local method can also capture mesoscale features that cannot be resolved by local methods, as demonstrated by a droplet in shear flow. { Calculating the surface tension with the non-local model does require taking an integral at each point, which does make the surface tension calculation more computationally expensive, however each integral is independent and therefore lends itself well to massive parallelization. While not considered here, the non-local model can be easily implemented in existing codes by changing only the surface tension calculation. Additionally, the number of points included in each integral can be truncated to those within a distance of 3.5$\varepsilon/\Delta x$ from the target point, reducing the additional amount of work per time step to be $\mathcal{O}\left(N \varepsilon^2/\Delta x^2\right)$. For a fixed mesh, the force shape function $f_\varepsilon(|{\mathbf x}-{\mathbf y}|)$ can be precomputed before the simulation and used each time step.} Our future work will include full three-dimensional simulations. Because the force due to surface tension only needs to be calculated near the interface, fast Level Set methods such as the one proposed by \cite{Adalsteinsson1995} could be adapted to the non-local model to reduce the computational time. In addition, simulations of multiscale systems are well fitted for adaptive mesh refinement, particularly when considering the interface between merging bubbles.
\section{Acknowledgements}
This work was supported by the U.S. Department of Energy (DOE) Office of Science, Office of Advanced Scientific Computing Research as part of the New Dimension Reduction Methods and Scalable Algorithms for Nonlinear Phenomena project. Pacific Northwest National Laboratory is operated by Battelle for the DOE under Contract DE-AC05-76RL01830.
The authors wish to thank Professor S. Guido for the data provided in Fig. \ref{fig:Shear_D}.
\section{}
\label{}
\section{Introduction}
\file{elsarticle.cls} is a thoroughly re-written document class
for formatting \LaTeX{} submissions to Elsevier journals.
The class uses the environments and commands defined in \LaTeX{} kernel
without any change in the signature so that clashes with other
contributed \LaTeX{} packages such as \file{hyperref.sty},
\file{preview-latex.sty}, etc., will be minimal.
\file{elsarticle.cls} is primarily built upon the default
\file{article.cls}. This class depends on the following packages
for its proper functioning:
\begin{enumerate}
\item \file{natbib.sty} for citation processing;
\item \file{geometry.sty} for margin settings;
\item \file{fleqn.clo} for left aligned equations;
\item \file{graphicx.sty} for graphics inclusion;
\item \file{txfonts.sty} optional font package, if the document is to
be formatted with Times and compatible math fonts;
\item \file{hyperref.sty} optional packages if hyperlinking is
required in the document;
\item \file{endfloat.sty} optional packages if floats to be placed at
end of the PDF.
\end{enumerate}
All the above packages (except some optional packages) are part of any
standard \LaTeX{} installation. Therefore, the users need not be
bothered about downloading any extra packages. Furthermore, users are
free to make use of \textsc{ams} math packages such as
\file{amsmath.sty}, \file{amsthm.sty}, \file{amssymb.sty},
\file{amsfonts.sty}, etc., if they want to. All these packages work in
tandem with \file{elsarticle.cls} without any problems.
\section{Major Differences}
Following are the major differences between \file{elsarticle.cls}
and its predecessor package, \file{elsart.cls}:
\begin{enumerate}[\textbullet]
\item \file{elsarticle.cls} is built upon \file{article.cls}
while \file{elsart.cls} is not. \file{elsart.cls} redefines
many of the commands in the \LaTeX{} classes/kernel, which can
possibly cause surprising clashes with other contributed
\LaTeX{} packages;
\item provides preprint document formatting by default, and
optionally formats the document as per the final
style of models $1+$, $3+$ and $5+$ of Elsevier journals;
\item some easier ways for formatting \verb+list+ and
\verb+theorem+ environments are provided while people can still
use \file{amsthm.sty} package;
\item \file{natbib.sty} is the main citation processing package
which can comprehensively handle all kinds of citations and
works perfectly with \file{hyperref.sty} in combination with
\file{hypernat.sty};
\item long title pages are processed correctly in preprint and
final formats.
\end{enumerate}
\section{Installation}
The package is available at author resources page at Elsevier
(\url{http://www.elsevier.com/locate/latex}).
It can also be found in any of the nodes of the Comprehensive
\TeX{} Archive Network (\textsc{ctan}), one of the primary nodes
being
\url{http://tug.ctan.org/tex-archive/macros/latex/contrib/elsarticle/}.
Please download the \file{elsarticle.dtx} which is a composite
class with documentation and \file{elsarticle.ins} which is the
\LaTeX{} installer file. When we compile the
\file{elsarticle.ins} with \LaTeX{} it provides the class file,
\file{elsarticle.cls} by
stripping off all the documentation from the \verb+*.dtx+ file.
The class may be moved or copied to a place, usually,
\verb+$TEXMF/tex/latex/elsevier/+,
or a folder which will be read
by \LaTeX{} during document compilation. The \TeX{} file
database needs updation after moving/copying class file. Usually,
we use commands like \verb+mktexlsr+ or \verb+texhash+ depending
upon the distribution and operating system.
\section{Usage}\label{sec:usage}
The class should be loaded with the command:
\begin{vquote}
\documentclass[<options>]{elsarticle}
\end{vquote}
\noindent where the \verb+options+ can be the following:
\begin{description}
\item [{\tt\color{verbcolor} preprint}] default option which format the
document for submission to Elsevier journals.
\item [{\tt\color{verbcolor} review}] similar to the \verb+preprint+
option, but increases the baselineskip to facilitate easier review
process.
\item [{\tt\color{verbcolor} 1p}] formats the article to the look and
feel of the final format of model 1+ journals. This is always single
column style.
\item [{\tt\color{verbcolor} 3p}] formats the article to the look and
feel of the final format of model 3+ journals. If the journal is a two
column model, use \verb+twocolumn+ option in combination.
\item [{\tt\color{verbcolor} 5p}] formats for model 5+ journals. This
is always of two column style.
\item [{\tt\color{verbcolor} authoryear}] author-year citation style of
\file{natbib.sty}. If you want to add extra options of
\file{natbib.sty}, you may use the options as comma delimited strings
as arguments to \verb+\biboptions+ command. An example would be:
\end{description}
\begin{vquote}
\biboptions{longnamesfirst,angle,semicolon}
\end{vquote}
\begin{description}
\item [{\tt\color{verbcolor} number}] numbered citation style. Extra options
can be loaded with\linebreak \verb+\biboptions+ command.
\item [{\tt\color{verbcolor} sort\&compress}] sorts and compresses the
numbered citations. For example, citation [1,2,3] will become [1--3].
\item [{\tt\color{verbcolor} longtitle}] if front matter is unusually long, use
this option to split the title page across pages with the correct
placement of title and author footnotes in the first page.
\item [{\tt\color{verbcolor} times}] loads \file{txfonts.sty}, if
available in the system to use Times and compatible math fonts.
\item [{\tt\color{verbcolor} reversenotenum}] Use alphabets as
author--affiliation linking labels and use numbers for author
footnotes. By default, numbers will be used as author--affiliation
linking labels and alphabets for author footnotes.
\item [{\tt\color{verbcolor} lefttitle}] To move title and
author/affiliation block to flushleft. \verb+centertitle+ is the
default option which produces center alignment.
\item [{\tt\color{verbcolor} endfloat}] To place all floats at the end
of the document.
\item [{\tt\color{verbcolor} nonatbib}] To unload natbib.sty.
\item [{\tt\color{verbcolor} doubleblind}] To hide author name,
affiliation, email address etc. for double blind refereeing purpose.
\item[] All options of \file{article.cls} can be used with this
document class.
\item[] The default options loaded are \verb+a4paper+, \verb+10pt+,
\verb+oneside+, \verb+onecolumn+ and \verb+preprint+.
\end{description}
\section{Frontmatter}
There are two types of frontmatter coding:
\begin{enumerate}[(1)]
\item each author is
connected to an affiliation with a footnote marker; hence all
authors are grouped together and affiliations follow;
\pagebreak
\item authors of same affiliations are grouped together and the
relevant affiliation follows this group.
\end{enumerate}
An example of coding the first type is provided below.
\begin{vquote}
\title{This is a specimen title\tnoteref{t1,t2}}
\tnotetext[t1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[t2]{The second title footnote which is a longer
text matter to fill through the whole text width and
overflow into another line in the footnotes area of the
first page.}
\end{vquote}
\begin{vquote}
\author[1]{Jos Migchielsen\corref{cor1}%
\fnref{fn1}}
\ead{<EMAIL>}
\author[2]{CV Radhakrishnan\fnref{fn2}}
\ead{<EMAIL>}
\author[3]{CV Rajagopal\fnref{fn1,fn3}}
\ead[url]{www.stmdocs.in}
\end{vquote}
\begin{vquote}
\cortext[cor1]{Corresponding author}
\fntext[fn1]{This is the first author footnote.}
\fntext[fn2]{Another author footnote, this is a very long
footnote and it should be a really long footnote. But this
footnote is not yet sufficiently long enough to make two
lines of footnote text.}
\fntext[fn3]{Yet another author footnote.}
\address[1]{Elsevier B.V., Radarweg 29, 1043 NX Amsterdam,
The Netherlands}
\address[2]{Sayahna Foundations, JWRA 34, Jagathy,
Trivandrum 695014, India}
\address[3]{STM Document Engineering Pvt Ltd., Mepukada,
Malayinkil, Trivandrum 695571, India}
\end{vquote}
The output of the above \TeX{} source is given in Clips~\ref{clip1} and
\ref{clip2}. The header portion or title area is given in
Clip~\ref{clip1} and the footer area is given in Clip~\ref{clip2}.
\deforange{blue!70}
\src{Header of the title page.}
\includeclip{1}{130 612 477 707}{1psingleauthorgroup.pdf
\deforange{orange}
\deforange{blue!70}
\src{Footer of the title page.}
\includeclip{1}{93 135 499 255}{1pseperateaug.pdf
\deforange{orange}
Most of the commands such as \verb+\title+, \verb+\author+,
\verb+\address+ are self explanatory. Various components are
linked to each other by a label--reference mechanism; for
instance, title footnote is linked to the title with a footnote
mark generated by referring to the \verb+\label+ string of
the \verb=\tnotetext=. We have used similar commands
such as \verb=\tnoteref= (to link title note to title);
\verb=\corref= (to link corresponding author text to
corresponding author); \verb=\fnref= (to link footnote text to
the relevant author names). \TeX{} needs two compilations to
resolve the footnote marks in the preamble part.
Given below are the syntax of various note marks and note texts.
\begin{vquote}
\tnoteref{<label(s)>}
\corref{<label(s)>}
\fnref{<label(s)>}
\tnotetext[<label>]{<title note text>}
\cortext[<label>]{<corresponding author note text>}
\fntext[<label>]{<author footnote text>}
\end{vquote}
\noindent where \verb=<label(s)>= can be either one or more comma
delimited label strings. The optional arguments to the
\verb=\author= command holds the ref label(s) of the address(es)
to which the author is affiliated while each \verb=\address=
command can have an optional argument of a label. In the same
manner, \verb=\tnotetext=, \verb=\fntext=, \verb=\cortext= will
have optional arguments as their respective labels and note text
as their mandatory argument.
The following example code provides the markup of the second type
of author-affiliation.
\begin{vquote}
\author{Jos Migchielsen\corref{cor1}%
\fnref{fn1}}
\ead{<EMAIL>}
\address{Elsevier B.V., Radarweg 29, 1043 NX Amsterdam,
The Netherlands}
\author{CV Radhakrishnan\fnref{fn2}}
\ead{<EMAIL>}
\address{Sayahna Foundations, JWRA 34, Jagathy,
Trivandrum 695014, India}
\author{CV Rajagopal\fnref{fn1,fn3}}
\ead[url]{www.stmdocs.in}
\address{STM Document Engineering Pvt Ltd., Mepukada,
Malayinkil, Trivandrum 695571, India}
\end{vquote}
\vspace*{-.5pc}
\begin{vquote}
\cortext[cor1]{Corresponding author}
\fntext[fn1]{This is the first author footnote.}
\fntext[fn2]{Another author footnote, this is a very long
footnote and it should be a really long footnote. But this
footnote is not yet sufficiently long enough to make two lines
of footnote text.}
\end{vquote}
The output of the above \TeX{} source is given in Clip~\ref{clip3}.
\deforange{blue!70}
\src{Header of the title page..}
\includeclip{1}{119 563 468 709}{1pseperateaug.pdf
\deforange{orange}
\pagebreak
Clip~\ref{clip4} shows the output after giving \verb+doubleblind+ class option.
\deforange{blue!70}
\src{Double blind article}
\includeclip{1}{124 567 477 670}{elstest-1pdoubleblind.pdf
\deforange{orange}
\vspace*{-.5pc}
The frontmatter part has further environments such as abstracts and
keywords. These can be marked up in the following manner:
\begin{vquote}
\begin{abstract}
In this work we demonstrate the formation of a new type of
polariton on the interface between a ....
\end{abstract}
\end{vquote}
\vspace*{-.5pc}
\begin{vquote}
\begin{keyword}
quadruple exiton \sep polariton \sep WGM
\end{keyword}
\end{vquote}
\noindent Each keyword shall be separated by a \verb+\sep+ command.
\textsc{msc} classifications shall be provided in
the keyword environment with the commands
\verb+\MSC+. \verb+\MSC+ accepts an optional
argument to accommodate future revisions.
eg., \verb=\MSC[2008]=. The default is 2000.\looseness=-1
\subsection{New page}
Sometimes you may need to give a page-break and start a new page after
title, author or abstract. Following commands can be used for this
purpose.
\begin{vquote}
\newpageafter{title}
\newpageafter{author}
\newpageafter{abstract}
\end{vquote}
\begin{itemize}
\leftskip-2pc
\item [] {\tt\color{verbcolor} \verb+\newpageafter{title}+} typeset the title alone on one page.
\item [] {\tt\color{verbcolor} \verb+\newpageafter{author}+} typeset the title
and author details on one page.
\item [] {\tt\color{verbcolor} \verb+\newpageafter{abstract}+}
typeset the title,
author details and abstract \& keywords one one page.
\end{itemize}
\section{Floats}
{Figures} may be included using the command, \verb+\includegraphics+ in
combination with or without its several options to further control
graphic. \verb+\includegraphics+ is provided by \file{graphic[s,x].sty}
which is part of any standard \LaTeX{} distribution.
\file{graphicx.sty} is loaded by default. \LaTeX{} accepts figures in
the postscript format while pdf\LaTeX{} accepts \file{*.pdf},
\file{*.mps} (metapost), \file{*.jpg} and \file{*.png} formats.
pdf\LaTeX{} does not accept graphic files in the postscript format.
The \verb+table+ environment is handy for marking up tabular
material. If users want to use \file{multirow.sty},
\file{array.sty}, etc., to fine control/enhance the tables, they
are welcome to load any package of their choice and
\file{elsarticle.cls} will work in combination with all loaded
packages.
\section[Theorem and ...]{Theorem and theorem like environments}
\file{elsarticle.cls} provides a few shortcuts to format theorems and
theorem-like environments with ease. In all commands the options that
are used with the \verb+\newtheorem+ command will work exactly in the same
manner. \file{elsarticle.cls} provides three commands to format theorem or
theorem-like environments:
\begin{vquote}
\newtheorem{thm}{Theorem}
\newtheorem{lem}[thm]{Lemma}
\newdefinition{rmk}{Remark}
\newproof{pf}{Proof}
\newproof{pot}{Proof of Theorem \ref{thm2}}
\end{vquote}
The \verb+\newtheorem+ command formats a
theorem in \LaTeX's default style with italicized font, bold font
for theorem heading and theorem number at the right hand side of the
theorem heading. It also optionally accepts an argument which
will be printed as an extra heading in parentheses.
\begin{vquote}
\begin{thm}
For system (8), consensus can be achieved with
$\|T_{\omega z}$
...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{thm}
\end{vquote}
Clip~\ref{clip5} will show you how some text enclosed between the
above code\goodbreak \noindent looks like:
\vspace*{6pt}
\deforange{blue!70}
\src{{\ttfamily\color{verbcolor}\expandafter\@gobble\string\\ newtheorem}}
\includeclip{2}{1 1 453 120}{jfigs.pdf}
\deforange{orange}
The \verb+\newdefinition+ command is the same in
all respects as its\linebreak \verb+\newtheorem+ counterpart except that
the font shape is roman instead of italic. Both
\verb+\newdefinition+ and \verb+\newtheorem+ commands
automatically define counters for the environments defined.
\vspace*{6pt}
\deforange{blue!70}
\src{{\ttfamily\color{verbcolor}\expandafter\@gobble\string\\ newdefinition}}
\includeclip{1}{1 1 453 105}{jfigs.pdf}
\deforange{orange}
The \verb+\newproof+ command defines proof environments with
upright font shape. No counters are defined.
\vspace*{6pt}
\deforange{blue!70}
\src{{\ttfamily\color{verbcolor}\expandafter\@gobble\string\\ newproof}}
\includeclip{3}{1 1 453 65}{jfigs.pdf}
\deforange{orange}
Users can also make use of \verb+amsthm.sty+ which will override
all the default definitions described above.
\section[Enumerated ...]{Enumerated and Itemized Lists}
\file{elsarticle.cls} provides an extended list processing macros
which makes the usage a bit more user friendly than the default
\LaTeX{} list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list counter
type and its attributes.
\begin{vquote}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.', so that the item counter will be suffixed
by a period.
\item You can use `a)' for alphabetical counter and '(i)' for
roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\end{vquote}
\deforange{blue!70}
\src{List -- Enumerate}
\includeclip{4}{1 1 453 185}{jfigs.pdf}
\deforange{orange}
Further, the enhanced list environment allows one to prefix a
string like `step' to all the item numbers.
\begin{vquote}
\begin{enumerate}[Step 1.]
\item This is the first step of the example list.
\item Obviously this is the second step.
\item The final step to wind up this example.
\end{enumerate}
\end{vquote}
\deforange{blue!70}
\src{List -- enhanced}
\includeclip{5}{1 1 313 83}{jfigs.pdf}
\deforange{orange}
\section{Cross-references}
In electronic publications, articles may be internally
hyperlinked. Hyperlinks are generated from proper
cross-references in the article. For example, the words
\textcolor{black!80}{Fig.~1} will never be more than simple text,
whereas the proper cross-reference \verb+\ref{tiger}+ may be
turned into a hyperlink to the figure itself:
\textcolor{blue}{Fig.~1}. In the same way,
the words \textcolor{blue}{Ref.~[1]} will fail to turn into a
hyperlink; the proper cross-reference is \verb+\cite{Knuth96}+.
Cross-referencing is possible in \LaTeX{} for sections,
subsections, formulae, figures, tables, and literature
references.
\section[Mathematical ...]{Mathematical symbols and formulae}
Many physical/mathematical sciences authors require more
mathematical symbols than the few that are provided in standard
\LaTeX. A useful package for additional symbols is the
\file{amssymb} package, developed by the American Mathematical
Society. This package includes such oft-used symbols as
$\lesssim$ (\verb+\lesssim+), $\gtrsim$ (\verb+\gtrsim+) or
$\hbar$ (\verb+\hbar+). Note that your \TeX{}
system should have the \file{msam} and \file{msbm} fonts installed. If
you need only a few symbols, such as $\Box$ (\verb+\Box+), you might try the
package \file{latexsym}.
Another point which would require authors' attention is the
breaking up of long equations. When you use
\file{elsarticle.cls} for formatting your submissions in the
\verb+preprint+ mode, the document is formatted in single column
style with a text width of 384pt or 5.3in. When this document is
formatted for final print and if the journal happens to be a double column
journal, the text width will be reduced to 224pt at for 3+
double column and 5+ journals respectively. All the nifty
fine-tuning in equation breaking done by the author goes to waste in
such cases. Therefore, authors are requested to check this
problem by typesetting their submissions in final format as well
just to see if their equations are broken at appropriate places,
by changing appropriate options in the document class loading
command, which is explained in section~\ref{sec:usage},
\nameref{sec:usage}. This allows authors to fix any equation breaking
problem before submission for publication.
\file{elsarticle.cls} supports formatting the author submission
in different types of final format. This is further discussed in
section \ref{sec:final}, \nameref{sec:final}.
\subsection*{Displayed equations and double column journals}
Many Elsevier journals print their text in two columns. Since
the preprint layout uses a larger line width than such columns,
the formulae are too wide for the line width in print. Here is an
example of an equation (see equation 6) which is perfect in a
single column preprint format:
\bigskip
\setlength\Sep{6pt}
\src{See equation (6)}
\deforange{blue!70}
\includeclip{4}{105 500 500 700}{1psingleauthorgroup.pdf}
\deforange{orange}
\noindent When this document is typeset for publication in a
model 3+ journal with double columns, the equation will overlap
the second column text matter if the equation is not broken at
the appropriate location.
\vspace*{6pt}
\deforange{blue!70}
\src{See equation (6) overprints into second column}
\includeclip{3}{59 421 532 635}{elstest-3pd.pdf}
\deforange{orange}
\vspace*{6pt}
\noindent The typesetter will try to break the equation which
need not necessarily be to the liking of the author or as it
happens, typesetter's break point may be semantically incorrect.
Therefore, authors may check their submissions for the incidence
of such long equations and break the equations at the correct
places so that the final typeset copy will be as they wish.
\section{Bibliography}
Three bibliographic style files (\verb+*.bst+) are provided ---
\file{elsarticle-num.bst}, \file{elsarticle-num-names.bst} and
\file{elsarticle-harv.bst} --- the first one can be used for the
numbered scheme, second one for numbered with new options of
\file{natbib.sty}. The third one is for the author year
scheme.
In \LaTeX{} literature, references are listed in the
\verb+thebibliography+ environment. Each reference is a
\verb+\bibitem+ and each \verb+\bibitem+ is identified by a label,
by which it can be cited in the text:
\verb+\bibitem[Elson et al.(1996)]{ESG96}+ is cited as
\verb+\citet{ESG96}+.
\noindent In connection with cross-referencing and
possible future hyperlinking it is not a good idea to collect
more that one literature item in one \verb+\bibitem+. The
so-called Harvard or author-year style of referencing is enabled
by the \LaTeX{} package \file{natbib}. With this package the
literature can be cited as follows:
\begin{enumerate}[\textbullet]
\item Parenthetical: \verb+\citep{WB96}+ produces (Wettig \& Brown, 1996).
\item Textual: \verb+\citet{ESG96}+ produces Elson et al. (1996).
\item An affix and part of a reference:
\verb+\citep[e.g.][Ch. 2]{Gea97}+ produces (e.g. Governato et
al., 1997, Ch. 2).
\end{enumerate}
In the numbered scheme of citation, \verb+\cite{<label>}+ is used,
since \verb+\citep+ or \verb+\citet+ has no relevance in the numbered
scheme. \file{natbib} package is loaded by \file{elsarticle} with
\verb+numbers+ as default option. You can change this to author-year
or harvard scheme by adding option \verb+authoryear+ in the class
loading command. If you want to use more options of the \file{natbib}
package, you can do so with the \verb+\biboptions+ command, which is
described in the section \ref{sec:usage}, \nameref{sec:usage}. For
details of various options of the \file{natbib} package, please take a
look at the \file{natbib} documentation, which is part of any standard
\LaTeX{} installation.
In addition to the above standard \verb+.bst+ files, there are 10
journal-specific \verb+.bst+ files also available.
Instruction for using these \verb+.bst+ files can be found at
\href{http://support.stmdocs.in/wiki/index.php?title=Model-wise_bibliographic_style_files}
{http://support.stmdocs.in}
\section{Graphical abstract and highlights}
A template for adding graphical abstract and highlights are available
now. This will appear as the first two pages of the PDF before the
article content begins.
\pagebreak
Please refer below to see how to code them.
\begin{vquote}
....
....
\end{abstract}
\begin{graphicalabstract}
\end{graphicalabstract}
\begin{highlights}
\item Research highlight 1
\item Research highlight 2
\end{highlights}
\begin{keyword}
....
....
\end{vquote}
\section{Final print}\label{sec:final}
The authors can format their submission to the page size and margins
of their preferred journal. \file{elsarticle} provides four
class options for the same. But it does not mean that using these
options you can emulate the exact page layout of the final print copy.
\lmrgn=3em
\begin{description}
\item [\texttt{1p}:] $1+$ journals with a text area of
384pt $\times$ 562pt or 13.5cm $\times$ 19.75cm or 5.3in $\times$
7.78in, single column style only.
\item [\texttt{3p}:] $3+$ journals with a text area of 468pt
$\times$ 622pt or 16.45cm $\times$ 21.9cm or 6.5in $\times$
8.6in, single column style.
\item [\texttt{twocolumn}:] should be used along with 3p option if the
journal is $3+$ with the same text area as above, but double column
style.
\item [\texttt{5p}:] $5+$ with text area of 522pt $\times$
682pt or 18.35cm $\times$ 24cm or 7.22in $\times$ 9.45in,
double column style only.
\end{description}
Following pages have the clippings of different parts of
the title page of different journal models typeset in final
format.
Model $1+$ and $3+$ will have the same look and
feel in the typeset copy when presented in this document. That is
also the case with the double column $3+$ and $5+$ journal article
pages. The only difference will be wider text width of
higher models. Therefore we will look at the
different portions of a typical single column journal page and
that of a double column article in the final format.
\begin{center}
\hypertarget{bsc}{}
\hyperlink{sc}{
{\bf [Specimen single column article -- Click here]}
}
\hypertarget{bsc}{}
\hyperlink{dc}{
{\bf [Specimen double column article -- Click here]}
}
\end{center}
\src{}\hypertarget{sc}{}
\deforange{blue!70}
\hyperlink{bsc}{\includeclip{1}{88 120 514 724}{elstest-1p.pdf}}
\deforange{orange}
\src{}\hypertarget{dc}{}
\deforange{blue!70}
\hyperlink{bsc}{\includeclip{1}{27 61 562 758}{elstest-5p.pdf}}
\deforange{orange}
\end{document}
\section{}
\label{}
\section{}
\label{}
|
\section{Introduction}
The isotropic background radiation that fills the Universe and extends over more than 16 orders of magnitude in frequency --from radio waves all the way up to high-energy gamma rays \cite{Hill:2018trh}-- carries information on the emission mechanisms of different astrophysical and cosmological sources over the history of the Universe, and can possibly shed further light on the nature of the elusive dark matter (DM) that constitutes the largest fraction of its mass.
We consider the hypothesis that the bulk of the DM is made up of {(non-rotating)} black holes of primordial origin (PBHs), formed from the collapse of overdense Hubble patches prior to the big-bang nucleosynthesis epoch \cite{ZelNov,Hawking:1971ei}. We focus on the current PBH mass window for DM ranging from approximately $10^{17}$~g to $10^{19}$~g, revisiting the Hawking radiation constraints on their abundance from extragalactic gamma-ray data \cite{Carr:2009jm} and showing how future gamma- and X-ray observations\footnote{The frontier between gamma- and X-rays is not sharply defined. A reasonable distinction can be made setting it at $\sim 100\, \rm{keV}$.} with an increased sensitivity have the potential for discovering a population of PBHs comprising the totality of the DM.
This mass range is indeed particulary relevant for DM, given that it has been found that previously claimed femtolensing bounds \cite{Barnacka:2012bm} were marred by an inadequate treatment of the involved optics, leaving much of that window open \cite{Katz:2018zrn}. In addition, the $\sim 10\%$ limit on the abundance of PBHs at $10^{19}$~g -- $10^{20}$~g from the observed distribution of white dwarfs \cite{Graham:2015apa}, as well as from the disruption of neutron stars in the PBH {mass} range from $10^{19}$~g to $10^{23}$~g \cite{Capela:2013yf} has been challenged \cite{Montero-Camacho:2019jte}, opening the possibility that PBHs of mass below
$\sim 5\times 10^{22}$ g --with higher masses being constrained by microlensing \cite{Niikura:2017zjd}-- could explain all the DM.
The lower end of the current PBH mass window for DM, at $\sim 10^{17}$ g, comes instead from Hawking evaporation limits. As we already mentioned, in this work we focus specifically on extragalactic gamma-ray bounds from evaporation \cite{Carr:2009jm}. However, there are other phenomena, related to Hawking radiation, which have been used to constrain this low mass region: the Voyager measurements of $e^\pm$ \cite{Boudaud:2018hqb}, the 511 keV positron-electron annihiliation line from INTEGRAL \cite{DeRocco:2019fjq,Laha:2019ssq,Dasgupta:2019cae}, the non-detection of a neutrino flux from PBH emission at Superkamiokande \cite{Dasgupta:2019cae}, distortions on the CMB anisotropies \cite{Poulin:2016anj,Stocker:2018avm}, and the Galactic emission of gamma/X-rays \cite{Carr:2016hva} (see also \cite{Laha:2020ivk}). The advantage of the bounds coming from the possible PBH extragalatic emission is that they are free from Galactic propagation uncertainties. As we will see, making use of their full potential requires an adequate characterization of the emission from other astrophysical sources, mostly active galactic nuclei (AGNs) and blazars.
Hawking radiation \cite{Hawking:1974sw,Hawking:1974rv,Page:1976wx, MacGibbon:1990zk, MacGibbon:1991vc} is an approximately thermal particle emission expected to be emitted by black holes, with temperature $T=(8\pi k_B)^{-1} c^2\, m_P^2/M\simeq6\times10^{-8} M_\odot/M$ K, being $M$ the mass of the black hole, $M_\odot=3\times 10^{33}$ g the mass of the Sun and $m_P=\sqrt{{\hbar\,c}/{G}}$ the Planck Mass. In Ref.\ \cite{Carr:2009jm}, a conservative (but nonetheless stringent) upper bound on the cosmological PBH abundance was set for $M \lesssim 10^{17}$ g by comparing the predicted Hawking gamma-ray emission with the isotropic gamma-ray background in the approximate energy range 0.1 MeV --- 10 GeV that was measured by EGRET \cite{Strong:2004ry}, Fermi-LAT \cite{2010PhRvL.104j1101A} and COMPTEL \cite{Weidenspointner:2000aq}. This bound has recently been updated in \cite{Arbey:2019vqx} (with the data from the same experiments), finding a good agreement with \cite{Carr:2009jm}.
The main focus of our analysis is on the isotropic gamma- and X-ray background in the 10 keV -- MeV domain. Using a power-law modeling of such background --motivated by the assumption that a population of unresolved extra-Galactic sources (mainly AGNs and blazars) represent the main contribution to it-- we place an upper limit on the abundance of PBHs (as a function of their mass) by considering an array of datasets acquired by a variety of missions has been proposed over the latest decades. We also estimate the expected improvement in the bounds from a putative future experiment. We do so by assuming that such a (more sensitive) experiment will resolve a significantly larger number of individual AGNs and blazars and will therefore provide a lower isotropic unresolved background for energies above $\sim 200$ keV. We show that, under this assumption, a significantly better upper limit on the PBH abundance than the current one may be placed in the future. This result motivates the investment in future gamma- and X-ray experiments in this range, as well as in further theoretical studies geared towards a more precise modeling of astrophysical sources.
\section{Hawking radiation from PBHs}
\label{sec:Hawking}
The photon emission from a population of PBHs of mass $M$ accounting for a fraction $f=\Omega_{PBH}/\Omega_{DM}$ of the total DM density in the Universe is
\begin{align}
\Phi_M=\frac{{\rm d} N}{{\rm d} E\,{\rm d} t} = f\frac{c\,\rho}{4\pi\,M} \int { {\rm d}z \frac{e^{-\tau(z)} }{ H(z)} \, \Psi_M[(1+z)E]} \,,
\label{eq:x-ray}
\end{align}
where $\rho=2.17 \times 10^{-30} {\rm g}/{\rm cm}^3$ is the current DM density of the Universe \cite{Aghanim:2018eyx}, and $H(z)$ is the Hubble rate of expansion as a function of redshift. The function $\Psi_M[E]$ denotes the differential flux emitted by a single PBH, as a function of the energy $E$, per unit of energy and time. For PBHs of masses above $10^{16}$ g is well approximated by the primary\footnote{The spectrum of BHs may feature a secondary emission component, depending on their mass, which is due to the interactions among the primary emitted particles, see e.g.\ \cite{Carr:2009jm}.} Hawking emission:
\begin{align}
\Psi_M[E]=(2\pi\hbar)^{-1}\Gamma_s/(\exp(E/k_B T)-1)\,,
\end{align}
where the so-called grey factor $\Gamma_s$ is a function of $M$ and $E$. In the high-energy limit $E\gg k_B T$, the grey factor approximately satisfies $\Gamma_s
\propto (M/m_P)^2 (E/ m_P\, c^2)^2$; whereas for $E\ll k_B T$, $\Gamma_s
\propto (M/m_P)^4 (E/ m_P\, c^2)^4$ \cite{MacGibbon:1990zk}. These expressions are insufficient to render adequately the peak height and position of $\Psi_M[E]$, which is best computed numerically. To do so we use the public code BlackHawk \cite{Arbey:2019mbc}, which also allows to include the (subdominant) secondary emission. We find that the differential flux for BHs of mass between $10^{16}$g and $10^{20}$g can be approximated by
\begin{align}
\Psi_M[E]\simeq \frac{2.5\times 10^{21}\,{\rm GeV}^{-1}{\rm s}^{-1}}{(M_{18}\,E/E_0)^{-2.7}+(M_{18}\,E/E_0)^{6.7}}\,,
\end{align}
where $E_0=6.54\times10^{-5}$ GeV and $M_{18}\equiv M/10^{18}\rm g$. This approximation is accurate to better than $\sim 1\%$ around the emission's peak (until $\Psi_M[E]$ decreases an order of magnitude), which is enough for our purposes. Nevertheless, we obtain the bounds on the PBH abundance from the instantaneous spectra given by BlackHawk.
The factor $(1+z)$ inside $\Psi_M[(1+z)E]$ accounts for the Doppler shift from the time of emission to the time of arrival to the detector. The optical depth $\tau(z)$ describes the attenuation due to the propagation of the signal over the relevant cosmological redshifts. Unlike for hard gamma rays, this is negligible for soft gamma-rays and $X$-rays. The integrand in \eq{eq:x-ray} decreases very rapidly with $z$ and accurate results are obtained integrating up to $z\sim \mathcal{O}(100)$.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Xraydata_w_PBH_draft_v2.pdf}
\caption{Cosmic X-ray background spectrum, as measured by various experiments. Overimposed are the Ueda+14 model (blue dashed line), the fit to a double power-law (black dashed line) of Eq.\ \ref{eq:x-raym}, and the corrections to the latter due to two hypothetical monochromatic PBH distributions with different masses $M$ and cosmological abundances $f=\Omega_{PBH}/\Omega_{DM}$.}
\label{fig:xray_plot}
\end{figure}
\section{The X-ray and gamma-ray AGN background}
\label{sec:XrayBackground}
There has been a considerable effort dedicated to interpreting the measurements of the X-ray and gamma-ray background from keV energies all the way up to $\sim 100$~GeV in terms of a superposition of a large number of unresolved extra-Galactic sources. In particular, the data in the range $\sim 5$--200 keV observed by \textit{Swift}/BAT \cite{Swift_paper}, MAXI \cite{MAXI_paper}, ASCA \cite{ASCA_paper}, XMM-Newton \cite{XMM_paper}, Chandra \cite{Chandra_paper} and ROSAT \cite{ROSAT_paper} are well reproduced by a population synthesis model of active galactic nuclei (AGNs) developed by Ueda et al.\ in \cite{Ueda:2014tma} (see the blue dotted line in Figure \ref{fig:xray_plot}).
AGNs are powered by gas accretion onto a supermassive black hole and are very efficient X-ray emitters. The model of \cite{Ueda:2014tma} is based on the extrapolation of the luminosity functions of AGNs in different redshift ranges inferred by a sample of 4039 AGNs in soft (up to 2 keV) and/or hard X-ray bands ($>$2 keV). The objects in the sample include both Compton-thin and Compton-thick AGNs (with the latter being heavily obscured by dust). As can be seen if Figure \ref{fig:xray_plot}, this AGN modeling fails to describe adequately the SMM data.
Indeed, for energies above $\sim 50-100$ keV the contribution from blazars is expected to become progressively more important. These objects correspond to the
AGNs that are detected at a small angle between the accretion disk axis and the observer line of sight \cite{Urry:1995mg}; together with star-forming galaxies and radio galaxies, they are thought to dominate the GeV-TeV gamma-ray isotropic background measured by Fermi-LAT {(not shown in Figure \ref{fig:xray_plot}).}
Although the details of the intermediate MeV -- GeV domain are still not clearly understood (see in particular the discussion in \cite{DeAngelis:2017gra}), the previous considerations lead us to employ the working assumption that a combination of different classes of extra-Galactic emitters explain the X-ray and gamma-ray unresolved diffuse background in a wide energy range and that the emission from these populations of sources can be modeled as a superposition of featureless power-laws.
Therefore, in the approximate energy range going from $20$~keV to $3000$~keV, which corresponds to the region where the Hawking emission from BHs in the mass range $10^{16}$~g -- $10^{19}$~g can contribute importantly to the Universe's diffuse spectrum, we model the astrophysical background as a double power-law fit to the data from the SMM \cite{Watanabe1997}, Nagoya balloon \cite{Fukada1975}, HEAO--1 and HEAO-A4 \cite{Gruber1999,Kinzer1997} experiments. Concretely, we use the following proxy for the combined AGN and blazar emission:
\begin{align}
\Phi_{\rm AGN}=\frac{{\rm d} N}{{\rm d} E\,{\rm d} t} \,= \frac{A}{\left(E/E_b\right)^{n_1}+\left(E/E_b\right)^{n_2}}\,.
\label{eq:x-raym}
\end{align}
For instance, assuming zero contribution to the data from PBH evaporation, the best fit is: $E_b = 35.6966$~keV, $A=0.0642\, {\rm keV^{-1}s^{-1}cm^{-2}sr^{-1} }$, $n_1=1.4199$ and $n_2=2.8956$;
see the black dashed line in Figure \ref{fig:xray_plot}.
\section{Constraints and prospects on $f$}
\label{sec:constraints}
We present now our results on the current upper limits on the PBH abundance $f= \Omega_{\rm PBH}/\Omega_{\rm DM}$, in the mass window of interest, and estimate the prospects for a future MeV mission. {We recall that we assume a population of non-rotating PBHs. Angular momentum makes black holes evaporate faster, making the bounds on their abundance stronger \cite{Arbey:2019vqx}. For PBHs formed during radiation domination --which is the most common scenario-- the assumption of negligible angular momentum is the most reasonable one. Some amount of angular momentum may be expected if the PBHs form during a phase of early matter domination \cite{Harada:2017fjm}. A moderate level of rotation would not change our results significantly. Only nearly extremal black holes (which is unlikely in common PBH formation models) would lead to a significant difference. In this sense, our bounds and forecast should be regarded as conservative.}\newline
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{bound_aggressive_refit_v3.pdf}
\caption{95\% c.l.\ upper limits on the PBH abundance versus the PBH mass assuming a monochromatic mass population. The black dot-dashed line is the result of the present work neglecting the background contribution from AGNs. The light blue-dotted line is the analogous result of \cite{Carr:2009jm}, see also \cite{Arbey:2019mbc}. The blue continuous line is the bound obtained assuming a power-law modeling of the AGN and blazar contribution to the observed spectrum. The red dashed curve indicates the sensitivity achievable with an experiment capable of reducing the astrophysical background by a factor of 10}
\label{fig:mainResult}
\end{figure}
\noindent $\star$ {\bf Current bound: Conservative approach.}
We start by deriving the present bound on $f$ under the most conservative approach, i.e.\ without assuming any astrophysical modeling of the data, as done in \cite{Carr:2009jm} (see also \cite{Arbey:2019vqx}). Under the assumption that the data are Gaussian distributed, we compute the estimator
\begin{align} \label{est}
\hat\chi^2 =\sum(D - E^2\Phi_M)^2/{\Delta^2}
\end{align}
over the energy bins for which the PBH emission $\Phi_M$ is larger than the X-ray data $D$ shown in Figure \ref{fig:xray_plot} and whose errors are denoted $\Delta$ in \eq{est}. The allowed PBH fraction $f$ for a given PBH mass $M$ at $95\%$ c.l.\ corresponds to values of $\hat\chi^2$ smaller than the threshold
\begin{align}
\hat\chi^2 \leq \chi^2_{0.05} (N-1)\,,
\end{align}
where $N$ --which depends on $f$-- is the number of bins in which the model overshoots the data. The threshold and the estimator are computed as functions of $f$ by adopting $N-1$ as the number of degrees of freedom for the $\chi^2$ distribution. The bound is shown in Figure \ref{fig:mainResult} as a black dot-dashed line. Our result is in good agreement with those of \cite{Carr:2009jm} (see blue dotted line) and \cite{Arbey:2019vqx} in the PBH mass region ($M \geq 10^{16}$g) that we consider, even though in the (relevant) region between 1 MeV and 20 MeV those works used COMPTEL data whereas we use SMM, which has a much smaller error and a broader range. \newline
\noindent $\star$ {\bf Sensitivity reach with the current data.}
We assume now a double power-law fit of the AGN background akin to that of Eq.\ \eq{eq:x-raym}
and consider the following estimator, evaluated {\it over all energy bins} (of the data we have used to obtain the fit):
\begin{align}
\hat\chi^2={\sum}{\left( D - E^2 \Phi_M- E^2 \Phi_{\rm AGN} \right)^2}/{\Delta^2}\,.
\end{align}
The upper limit on the PBH abundance is set at the value of $f$ above which $\hat\chi^2$ worsens beyond the threshold for the 95$\%$ c.l.\ with respect to the minimum: i.e. the threshold is set as
\begin{align}
\hat\chi^2 - \hat\chi^2_{min} \leq \chi^2_{0.05} (1) \simeq 3.84\,.
\end{align}
Importantly, for each mass bin we obtain the best-fitting combination of AGN and blazar plus PBH emission by letting free the PBH fraction $f$ as well as the four parameters ($n_1$, $n_2$, $E_b$ and $A$) of Eq.\ \eq{eq:x-raym}, and then vary the parameter $f$ (recomputing the AGN and blazar contribution at each step) until the threshold above is reached.\newline
In Figure \ref{fig:mainResult} we show the results of both approaches ({\it conservative} and {\it sensitivity reach}) for 95\% c.l.\ upper limits on $f$. Regarding the conservative approach (black solid line), our constraints are in good agreement with those from \cite{Carr:2009jm} (blue dotted line), while the bound assuming the power-law AGN model (blue solid line) is a factor $\sim10$ stronger for the same $M$ and reaches larger PBH masses (up to about $7\times10^{17}$g). \newline
\noindent $\star$ {\bf Future prospects.}
In the forthcoming {years,} both the hard X-ray and the MeV gamma-ray sky will be probed with increasing accuracy by several planned space observatories. In the MeV range, a variety of space missions have been proposed over the last years, for instance: e-ASTROGAM \cite{DeAngelis:2017gra}, AMEGO \cite{McEnery:2019tcm}, AdEPT \cite{Hunter:2013wla}, COSI \cite{Tomsick:2019wvo}, SMILE \cite{Tanimori:2015wma}.
In the hard X-ray domain, the instrument onboard the ASTRO-H mission \cite{Matsumoto:2018mba} will measure X-rays up to $80$ keV with unprecedented accuracy.
All these instruments will detect a larger number of point sources, both AGNs and blazars. They are thus expected to characterize a lower isotropic extragalactic background due to unresolved point sources. As a consequence, a hypothetical PBH signal will be easier to detect. It is therefore important to provide a quantitative estimate of the potential of possible future experiments in either setting a stronger upper limit on the PBH abundance in the mass range under investigation, or identifying a PBH signal with sufficient significance.
A careful assessment on how the expected increase in point-source sensitivity, especially in the MeV domain, results in a lower isotropic background due to sub-threshold sources is beyond the scope of the present work\footnote{Notice that such an analysis can be done only on a case-by-case basis, from the detailed specifications of the experiments (which in several cases are still being decided)}. Instead, we present a simple but useful calculation based on straightforward extrapolations of observed luminosity functions at low redshift. If we assume that most of the isotropic background in the MeV domain is due to blazars, and we assume a single-power law luminosity function of the form \cite{2009ApJ...699..603A}
\begin{equation} \label{spl}
N(S) = N_0 (S/S_0)^{-\alpha}\,,
\end{equation}
where $S$ is the source flux and $\alpha$ is a positive number, the isotropic background at redshift zero due to unresolved sources is simply given by the integral:
\begin{equation}
\mathcal{I}(S_a,S_b)=\int_{S_0}^{S_{th}} S\,\frac{dN}{dS}\,dS\,,
\end{equation}
where $S_0$ is the minimum luminosity and $S_{th}$ is the detection threshold of the experiment. If we now consider two experiments with different thresholds $S_1$ and $S_2$, the ratio between the unresolved flux in the two cases is
\begin{equation}
r = \frac{\mathcal{I}(S_2,S_0)}{\mathcal{I}(S_1,S_0)}\,.
\end{equation}
In the limit of $S_0 \ll S_1,S_2$ and for a exponent $0 < \alpha < 1$ (like the ones found in \cite{2009ApJ...699..603A}), we get:
\begin{equation}
r \simeq \left(\frac{S_2}{S_1}\right)^{1-\alpha}\,.
\end{equation}
Let us now consider a future experiment such as AMEGO, which is able to provide a factor of $\sim 10$ increase in point-source sensitivity with respect to COMPTEL at 1 MeV \cite{McEnery:2019tcm}. Let us assume a single power-law luminosity function for a population of blazars, as in \eq{spl}, following \cite{2009ApJ...699..603A}. For obtaining an order-of-magnitude estimate, we can ignore a possible (but mild) redshift dependence of the exponent $\alpha$. The above formalism implies that a background reduction by a factor of $10$ is possible with a luminosity function exponent $\alpha \simeq 0.5$, which is commensurate with the values quoted in \cite{2009ApJ...699..603A}.
Given these considerations, we consider here two different scenarios, characterized by {factors} of $10$ and $100$ reduction in the diffuse, unresolved background, accompanied by a reduction by the same factor of its uncertainty. We generate mock data according to this prescription, and adopt the same procedure we applied to the current data in the {\it sensitivity reach} approach described previously. The results are shown in Figure \ref{fig:mainResult}. The red dashed line indicates a potentially significant improvement of the current upper limits for a factor 10 reduction, together with a notable extension towards larger masses.
\section{Discussion and Conclusions}
In this paper we have studied the upper limits on the abundance of primordial black holes from X-ray and gamma-ray Hawking radiation, as a function of mass and assuming a monochromatic distribution in the $10^{16}$ g -- $10^{19}$~g mass range. This approximate mass window is particularly interesting {not only for the DM problem but also} from the model-building point of view. The known examples of the conceptually simplest mechanism capable of producing PBHs from single-field inflation --based on an approximate inflection point in the potential \cite{Ivanov:1994pa}, see e.g.\ \cite{Garcia-Bellido:2017mdw,Kannike:2017bxn,Ballesteros:2017fsr,Cicoli:2018asa,Dalianis:2018frf,Ballesteros:2019hus,Ballesteros:2020qam}-- tend to do so in this range, once a reasonable fit to the CMB data at cosmological scales is imposed, see e.g.\ the related discussion in \cite{Ballesteros:2017fsr}. As discussed in the introduction, this mass range is within the currently most promising region for the existence of a significant contribution of PBHs to the DM.
We have first computed the upper limit with a conservative approach, by requiring that the expected signal from PBHs does not overshoot the diffuse {(extragalactic)} background. The bound we have obtained is mostly driven by SMM data and agrees {well} with the results presented in \cite{Carr:2009jm,Arbey:2019vqx}, which instead used COMPTEL data in the relevant energy range. This bound is competitive with the recently derived upper limits based on the Voyager $e^\pm$ data \cite{Boudaud:2018hqb} and extends to larger masses.
Then, we have considered a less conservative approach, which {takes} into account in a simple way the expected effect of known classes of unresolved astrophysical sources --AGNs and blazars in particular-- that are thought to provide the dominant contribution to the X-ray and gamma-ray diffuse isotropic background. Under the well-motivated (and common in the literature, see \cite{Ajello:2008xb}, \cite{Ueda:2014tma}) assumption that the current data are mainly reproduced by such sources and that they can be characterized by a double power-law, we derive a stronger upper limit on the PBH abundance based on a simultaneous fit of current data to the emission of these sources and PBHs in the MeV domain. This (sensitivity reach) bound is about an order of magnitude stronger than the conservative upper limit described above.
We have also considered the potential of a future, more sensitive, experiment in the MeV domain that could resolve a larger number of individual sources, therefore providing a lower and more accurate estimate of the diffuse unresolved background. The prospects, supported by the extrapolation of currently measured AGN and blazar luminosity functions, are very promising: We remark in particular that PBH masses as large as $10^{18}$ g are within reach under the assumption of a background reduction by a factor of $10$.
The exploration of the {diffuse photon emission in the low-energy gamma-ray (or upper X-ray range)} is therefore a promising avenue towards a possible future detection of a signal associated to a population of PBHs that may constitute a significant part --perhaps even all-- {of} the DM in the Universe.
In order to obtain further progress in this {promising PBH mass region for DM ($\gtrsim 10^{17}\rm{g}$),} a more sensitive experiment is needed {both in the MeV domain and (particularly) in the keV one,} and a more detailed understanding of the population of astrophysical sources that contribute to the bulk of the diffuse background is essential. Above 1 MeV, such modeling should potentially include not only the contributions from sources (AGNs and blazars) that are known to be dominant in the sub- and sup-Mev bands, but also possible currently subdominant sources.
{For a fixed value of the PBH abundance $f$, the emission $\Phi_M$ grows if the mass $M$ of the PBHs is {\it decreased} --we recall that we assumed a monochromatic distribution-- and, simultaneously, the location of the peak of the PBH moves towards higher energies. This can be understood qualitatively (disregarding the secondary emission) by applying the scaling discussed in Section \ref{sec:Hawking}. As the measured diffuse background has a decreasing overall tendency from $\sim 30\, \rm{keV}$ to $\sim 10^3\, \rm{MeV}$ (see \cite{Hill:2018trh}), the bounds on the PBH abundance become stronger than those shown in figure \ref{fig:mainResult} for smaller masses (below $10 ^{16}\,\rm{g}$). This low mass region is thus of no relevance for the DM problem and we have not considered it in this work. It is interesting though that the emission of a bright but sub-dominant population of PBHs of mass $\mathcal{O}$($10^{15}$ g - $10^{16}$ g) would peak in the 1-10 MeV region. In this energy range, the Compton and Fermi X-ray data present a noticeable change in slope with respect to that of SMM (see e.g.\ figure 7 of ref. \cite{Hill:2018trh}). This energy region of the spectrum of the Universe is still not fully characterized from a theoretical point of view. These features make it a tantalizing target for the search of potentially exotic emitters (and subdominant PBHs in particular). Indeed, the bounds on the PBH abundance from gamma-rays in this region \cite{Carr:2009jm} necessarily use the approach that we have termed {\it conservative} (i.e.\ no astrophysical background is assumed).
Returning now to PBHs as a dark matter candidate, i.e.\ for masses $\sim 10^{18}$~g, the fact that $\Phi_M$ decreases as $M$ increases (and the measured diffuse background increases at the corresponding, lower, energies) a more accurate measurement and modeling of the soft X-ray diffuse background may play, as we have argued, a crucial role towards a potential discovery or, at least, to strengthen current bounds substantially. To conclude, we} remark that there exist other channels for PBH detection in {the mass window relevant for DM}. In particular, the energy injection associated to Hawking emission around or just after recombination can be probed with CMB data \cite{Poulin:2016anj}. Also, if the effect occurs during the reionization epoch it can be potentially detectable in the 21 cm absorption line of neutral Hydrogen \cite{Clark:2018ghm}. Moreover, the positron emission by PBHs of mass around $10^{17}$~g has been recently used to set a bound on $f$ using the keV line, which can improve over the gamma-ray bound, depending on the assumed Galactic density profile \cite{DeRocco:2019fjq}. All these channels are complementary and their further exploration could help to identify or rule out a significant population of PBHs in this promising mass window.
\vspace{0.5cm}
\mysections{Acknowledgments}
We thank M.A.\ S\'anchez-Conde for collaboration in the first stages of this project, M.\ Ajello for providing files with the X-ray data, P.\ Serpico and M.\ Taoso for discussions and useful comments and suggestions on a draft version of this work, J.R.\ Espinosa and A.\ Urbano for discussions and A.\ Arbey and J.\ Auffinger for correspondence about BlackHawk. We thank M.\ Boudaud for insightful comments which helped to improve this paper.
The work of GB is funded by a {\it Contrato de Atracci\'on de Talento (Modalidad 1) de la Comunidad de Madrid} (Spain), with number 2017-T1/TIC-5520, by {\it MINECO} (Spain) under contract FPA2016-78022-P and {\it MCIU} (Spain) through contract PGC2018-096646-A-I00.
DG has received financial support through the Postdoctoral Junior Leader Fellowship Programme from la Caixa Banking Foundation (grant n.\ LCF/BQ/LI18/11630014).
JCB is supported by an {\it Atracci\'on de Talento de la Comunidad de Madrid} (contract no.\ 2016-T1/TIC-1542).
DG and JCB are additionally supported by {\it MINECO} (Spain) through the grant PGC2018-095161-B-I00 and Red Consolider MultiDark FPA2017-90566-REDC.
All authors are supported by the IFT UAM-CSIC Centro de Excelencia Severo Ochoa SEV-2016-0597 grant.
\bibliographystyle{hunsrtm}
|
\section{Introduction}
In this paper, we address the following problem:
\smallskip
\noindent{\bf Problem 1.}
Let $M$ be a smooth real hypersurface in a complex manifold $X$.
Find necessary and sufficient conditions on $M$
to be CR-diffeomorphic to a real-analytic CR manifold.
\smallskip
If $M$ is CR-diffeomorphic to a real-analytic CR manifold,
we shall call it
{\em analytically regularizable}.
Problem 1 is of interest because of
the case of real-analytic CR manifolds
being much better studied
with more results and tools available
such as complexification and Segre varieties
(see e.g.\ \cite{ber}).
On the other hand,
the problem
seems to be widely open
even for strictly pseudoconvex hypersurfaces,
where it is non-trivial,
i.e.\ there exist smooth non-analytic hypersurfaces
that are analytically regularizable and there exist those that are not,
see the end of Section 2 for respective examples.
(The latter phenomenon is in contrast with the case of
hypersurfaces of mixed Levi form signature,
where any CR-diffeomorphism to a real-analytic hypersurface
extends holomorphically to both sides,
hence any analytically regularizable hypersurface must be
already real-analytic in its ambient complex manifold $X$.)
The goal of this paper is
to provide
a nontrivial {\em necessary and sufficient} condition (Condition E below)
giving a solution to Problem 1
in the case
when $M$ is {\em strictly pseudoconvex}.
Our condition is formulated
in terms of holomorphic extension of certain
functions invariantly associated to $M$ (these functions can be viewed as Fefferman type determinants \cite{feffer}, as shown in Section 2).
Thus our results provide a {\em two-way bridge} for Problem 1
with the much better studied
questions of holomorphic extension of smooth functions from real submanifolds in complex manifolds.
We proceed by giving a high level non-technical
formulation of our result,
while more precise details can be found
in Section 2.
Given a smooth real hypersurface $M$
in a complex manifold $X$ of dimension $n+1$,
consider smooth local defining equations
of the kind $\rho(z, \bar z)=0$ for $M$ with $d\rho\ne 0$.
Then $\rho$ can be formally complexified at each point,
i.e.\ there exist formal power series $\rho(z,\bar w)$
giving for $w=z$ the Taylor series of $\rho$.
This allows to invariantly define formal
Segre varieties $$Q_p=\{z: \rho(z, \bar p)=0\}$$ for $p\in M$,
as well as their $k$-jets $j^k_pQ_p$ for every $k\ge1$.
In particular (see \cite{websterrefl}),
if $M$ is strictly pseudoconvex,
then $j^kQ_p$, $p\in M$,
defines a canonical embedding of $M$
into the space $J^{k,n}(X)$ of all $k$-jets of complex-analytic
hypersurfaces for every $k\ge 1$,
given by $$p\mapsto j^kQ_p,$$ and the image of $M$ appears to be totally real.
Furthermore, for any $k,l\ge 1$,
we obtain canonical smooth maps
$s^{k,l}$
between respective images of those embeddings of $M$,
sending $j^kQ_p$ to $j^lQ_p$ for every $p$.
To formulate our main result,
denote by $M_J\subset J^{1,n}$
the image of the embedding $p\mapsto j^1Q_p$ of $M$ as above,
and consider the map
$$s^{1,2}\colon M_J\to J^{2,n}$$
sending $j^1_pQ_p$ to $j^2_pQ_p$
for every $p\in M$.
Let $\3M\subset J^{1,n}$ be
the (smooth) real hypersurface
consisting of all $1$-jets with base points in $M$.
\begin{theorem}\Label{main0}
Let $M$ be a smooth strictly pseudoconvex hypersurface
in a complex manifold $X$.
Then $M$ is analytically regularizable
(i.e.\ CR-diffeomorphic to a real-analytic CR manifold)
if and only if the map $s^{1,2}$
admits a local holomorphic extension
to the pseudoconvex side of
$\3M$ in $J^{1,n}$ valued in $J^{2,n}$
that is smooth up to $\3M$.
\end{theorem}
Note that since $M_J$ is generic in $J^{1,n}$,
if such holomorphic extension of $s^{1,2}$ exists,
it is necessarily unique.
Also, the CR-diffeomorphism to a real-analytic CR manifold
is unique if exists,
up to a real-analytic CR-diffeomorphism,
as follows directly from the reflection principle for real-analytic strictly pseudoconvex hypersurfaces.
We conclude by mentioning that,
while analyticity problems have been
studied for other geometric structures, e.g. Riemannian structures
(see e.g.\ \cite{kazdan}),
no similar nontrivial necessary and sufficient conditions
seem to be known.
\medskip
\begin{center}
\bf Acknowledments
\end{center}
\medskip
We would like to thank Laszlo Lempert for his helpful remarks
following a talk given by the first author on the subject.
The first author was supported by the GACR (Czech Grant Agency) and the FWF (Austrian Science Fund) during the preparation of this paper.
\section{Condition E for smoothly embedded real hypersurfaces}
We now describe in details the holomorphic extension condition
in \autoref{main0}
that we shall call
Condition E.
We shall give below both an invariant and a coordinate-based formulations of it.
For the basic concepts in CR-geometry (such as Segre varieties and formal submanifolds) we refer to \cite{ber}, and for {\em jet bundles} and related concepts to \cite{cs}.
Let
$$
\pi \colon J^{1,n}
\to \CC{n+1}
$$
be the bundle
of $1$-jets of complex hypersurfaces of $\CC{n+1}$,
which is a projective holomorphic bundle over $\CC{n+1}$ with the fiber dimension $n$,
and
$M\subset\CC{n+1},\,n\geq 1$,
be a smooth strictly pseudoconvex real hypersurface.
Then the complex tangent bundle $\tc{}M$ induces
the natural embedding
$$
\varphi\colon M\to J^{1,n}, \quad x \mapsto \bigl(x,[\tc{x}M]\bigr).
$$
The image
$$
\varphi(M)=:M_J \subset J^{1,n}
$$
is a smooth $(2n+1)$-dimensional real submanifold in the $(2n+1)$-dimensional complex manifold $J^{1,n}$. Webster in \cite{websterrefl} observed that
$M_J\subset J^{1,n}$ is {\em totally real} whenever $M$ is Levi-nondegenerate.
Next,
associated with $M$ is the smooth pseudoconvex real hypersurface
$$
\widehat M:=\pi^{-1}(M) \subset J^{1,n}.
$$
The manifold $M_J$ is a smooth submanifold in $\widehat M$.
Note that $\widehat M$ itself is locally CR-equivalent
to $M\times \CC{n}$ (and thus is {\em holomorphically degenerate}, see \cite{ber}).
In what follows we denote by $U^+$ the pseudoconvex side of $M$ and by $$\widehat U^+:=\pi^{-1}(U^+)$$ that of $\widehat M$ ($\widehat U^+$ is locally biholomorphic to $U^+\times\CC{}$ at a point in $\widehat M$).
We next fix a point $p\in M$.
Since $M$ is smooth, we may consider at each point $q\in M$
near $p$, its formal complexfication at $q$ as a formal complex hypersurface in $\CC{n+1}\times\overline{\CC{n+1}}$ obtained by complexifying the formal Taylor series of its defining function at $q$. In this way, the formal Segre variety $S_q$ of $M$ at $q$ is well defined. Then the $2$-jets
\begin{equation}\Label{2jets}
j^2_qS_q,\,\,q\in M
\end{equation} of such formal Segre varieties induce a smooth embedding
of $M$ (and hence $M_J\subset J^{1,n}$) into the bundle
$$J^{2,n}=J^{2,n}(\CC{n+1})$$
of $2$-jets of complex hypersurfaces in $\CC{n+1}$. The space $J^{2,n}$ is canonically a fiber bundle
$$
\pi^2_1 \colon J^{2,n}\to J^{1,n}.
$$
The above $2$-jet embedding defines a canonical section of $\pi^2_1$,
$$s\colon M_J\to J^{2,n}.$$
Now our analyticity condition for a smooth strictly pseudoconvex submanifold looks as follows.
\begin{definition}\Label{coorfree} We say that $M$ {\em satisfies Condition E at $p$}, if for some choice of a neighborhood $U$ of $p$, the section $s$ extends as a smooth section of $\pi^2_1$ over the pseudoconvex side $\widehat U^+ \cup \widehat M$, which is furthermore holomorphic in $\widehat U^+$.
\end{definition}
We next give an (equivalent to the above) coodinate formulation of Condition E. If $M\subset\CC{n+1}$ is a smooth hypersurface with the defining equation
\begin{equation}\Label{rhoeq}
\rho(Z,\bar Z)=0,\quad Z=(z,w)=(z_1,...,z_n,w)\in\CC{n+1},
\end{equation}
$p\in M$ the distinguished point and $\rho_w(p,\bar p)\neq 0$, then its formal Segre variety at a point $q=(\tilde q,q_{n+1})\in M$ nearby $p$ is a graph of a function $w(z)$ (considered as a formal power series in $(z-\tilde q)$). Then the $2$-jets \eqref{2jets} amount to either the scalar function
$\Phi$ defined pointwise as $w''(z)$ for $z=\tilde q$ (case $n=1$), or to the symmetric matrix function $\Phi=(\Phi_{ij}),\,\,i,j=1,...,n$, defined pointwise as the collection of $w_{z_iz_j}$ for $z=\tilde q$ (case $n>1$). It is possible to verify that, in turn, for $n=1$ we have
\begin{equation}\Label{Phi1}
\Phi=\frac{1}{(\rho_w)^3}
\begin{vmatrix} \rho & \rho_z & \rho_w\\ \rho_z & \rho_{zz}
& \rho_{zw} \\ \rho_w & \rho_{zw} & \rho_{ww} \end{vmatrix},
\end{equation}
and for $n>1$ we have
\begin{equation}\Label{Phi}
\Phi_{ij}=\frac{1}{(\rho_w)^3}
\begin{vmatrix} \rho & \rho_{z_j} & \rho_w\\ \rho_{z_i} & \rho_{z_iz_j} & \rho_{z_iw} \\ \rho_w & \rho_{z_jw} & \rho_{ww}, \end{vmatrix} \quad i,j=1,...,n.
\end{equation}
(To obtain \eqref{Phi1},\eqref{Phi}, one has to differentiate twice the identity \eqref{rhoeq} assuming $w$ to be a function of $z$). Both the scalar function \eqref{Phi1} and the matrix valued function \eqref{Phi} can be considered as either smooth functions on the strictly pseudoconvex hypersurface $M$ or as that on the totally real manifold $M_J$ introduced above.
We shall remark that the invariant determinants in \eqref{Phi1},\eqref{Phi} were used also by Ebenfelt and the second author \cite{ez}
as well as by Ebenfelt, Duong and the second author
\cite{edz} for characterizing the Cartan tensor of a Levi-degenerate hypersurface in terms of its defining function. They can be seen as certain generalizations of the determinants used by Fefferman in \cite{feffer} for studying asymptotics of the Bergman metric in a smoothly bounded strictly pseudoconvex domain at a boundary point.
In terms of the $\Phi$-function, Condition E reads as follows.
\begin{definition}\Label{coor} We say that $M$ {\em satisfies Condition E at $p$}, if for some choice of a neighborhood $U$ of $p$, the function $\Phi$ defined on $M_J$ by either \eqref{Phi1} or \eqref{Phi} extends to the pseudoconvex side $\widehat U^+ \cup \widehat M$ holomorphically and smoothly up to the boundary.
\end{definition}
It is obvious that \autoref{coor} is equivalent to \autoref{coorfree}.
We give now the more precise local version
of our main result, from which the global result in \autoref{main0}
follows directly in view of the uniqueness
of the extension and the real-analytic CR-structure.
\begin{theorem}\Label{main}
A smooth strictly pseudoconvex real hypersurface $M\subset\CC{n+1},\,n\geq 1$, is locally analytically regularizable
(i.e.\ CR-equivalent near a point $p\in M$ to a real-analytic hypersurface $\2M\subset\CC{n+1}$) if and only it satisfies Condition E at $p$.
\end{theorem}
\begin{remark}\Label{Ck}
In fact, the above Condition E for smooth hypersurfaces can be naturally extended to hypersurfaces of merely class $C^2$. Following then the details of the proof of \autoref{main} below, it can be seen that, {\em for a $C^k$ smooth strictly pseudoconvex hypersurface $M\subset\CC{n}$ with $k\geq 2$, Condition E is necessary and sufficient for the existence of a local $C^{k-1}$ smooth CR-diffeomorhism of $M$ onto a real-analytic hypersurface $\tilde M\subset\CC{n+1}$.}
\end{remark}
We finish this section by providing two simple examples of
smooth non-analytic strictly pseudoconvex hypersurfaces in $\CC{2}$,
one of which admits and the other does not a CR-diffeomorphism onto a real-analytic hypersurface,
i.e.\ one is analytically regularizable while the other isn't.
\begin{example}
Let $f(z,w)$ be a holomorphic function in the unit ball $\mathbb B^2\subset\CC{2}$ which is smooth up to $\partial\mathbb B^2=S^3$ but does not extend holomorphically across $S^3$, e.g. one can take
a branch of $e^{(w-1)^{-1/3}}$.
Then, for $\epsilon$ sufficiently small, the map
(small perturbation of identity)
$$
(z,w)\mapsto (z + \epsilon f(z,w),w)
$$
defines a CR-diffeomorphism from $S^3$ onto a smooth
but not analytic strictly pseudoconvex hypersurface $M\subset\CC{2}$
that is obviously analytically regularizable.
\end{example}
\begin{example}
The real hypersurface $M\subset\CC{2}$
(a flat perturbation at the origin of the standard hyperquadric)
given by
$$
\ensuremath{\mbox{\rm Im}\,} w=|z^2|+e^{-1/|z|^2}
$$
is smooth and strictly pseudoconvex near the origin,
but is not CR-diffeomorphic to a real-analytic hypersurface
$\tilde M\subset {\4C}^2$, not even locally at $0$.
Indeed, since $M$ is {\em formally spherical} at $0$
(i.e.\ spherical up to infinite order),
the existence of a CR-diffeomorphism $H$ onto $\tilde M$ with, say, $H(0)=0$ would mean that $\tilde M$ is also formally
and hence biholomorphically spherical at $0$.
(The formal expansion of $H$ at $0$ yields a formal transformation of the hyperquadric $\ensuremath{\mbox{\rm Im}\,} w=|z|^2$ onto $\tilde M$
which is locally biholomorphic in view of the Chern-Moser theory \cite{chern}.)
Hence $M$ must itself be spherical
in a neighborhood of $0$.
On the other hand,
one can see by computing the Chern-Moser's curvature
(e.g. following \cite{lobumbilic} or using
the determinant expression \cite{ez}), that $M$ is not spherical in any neighborhood of $0$.
Thus $M$ is not analytically regularizable
in any neighborhood of $0$.
\end{example}
\section{Associated differential equations and the necessity of Condition E}
In this section, we show that the necessity of Condition E follows rather easily from the construction of {\em holomorphic differential equations associated with a real-analytic hypersurface}. On the other hand, the sufficiency of Condition E is already quite nontrivial. It is addressed in Section 4.
\subsection{The method of associated differential equations} It was observed by Cartan \cite{cartan} and Segre \cite{segre} (see also Webster \cite{webster}) that the geometry of a real hypersurface in $\CC{2}$ parallels that of a second order ODE
\begin{equation}\Label{wzz}
w\rq{}\rq{}=\Phi(z,w,w\rq{}).
\end{equation}
More generally, the geometry of a real hypersurface in $\CC{n+1},\,n\geq 1$, parallels that of a complete second order system of PDEs
\begin{equation}\Label{wzkzl}
w_{z_kz_l}=\Phi_{kl}(z_1,...,z_n,w,w_{z_1},...,w_{z_n}),\quad \Phi_{kl}=\Phi_{lk},\quad k,l=1,...,n.
\end{equation}
Moreover, {\em in the real-analytic case} this parallel becomes algorithmic by using the Segre family of a real hypersurface. With any real-analytic Levi-nondegenerate hypersurface $M\subset\CC{n+1},\,n\geq 1$ one can uniquely associate a holomorphic ODE \eqref{wzz} ($n=1$) or a holomorphic PDE system \eqref{wzkzl} ($n\geq 2$). The Segre family of $M$ plays a role of a mediator between the hypersurface and the associated differential equations. A more recent exposition of this method was given in the work \cite{sukhov1,sukhov2} of Sukhov. For recent work on associated differential equations in the degenerate setting, see e.g. the papers
\cite{divergence, nonminimalODE, nonanalytic} of the first author with Lamel and Shafikov.
The associated differential equation procedure is particularly clear in the case of a Levi-nondegenerate hypersurface in $\CC{2}$. In this case the Segre family is a 2-parameter anti-holomorphic family of holomorphic curves. It then follows from standard ODE theory that there exists an unique ODE \eqref{wzz}, for which the Segre varieties are precisely the graphs of solutions. This ODE is called \it the associated ODE. \rm
In the general case, both right hand sides in \eqref{wzz},\eqref{wzkzl} appear as functions determining the $2$-jet of a Segre variety as an analytic function of the $1$-jet. More explicitly, we denote the coordinates in $\CC{n+1}$ by
$$
(z,w)=(z_1, \ldots, z_n,w).
$$
Let then fix $M\subset\CC{n+1}$ to be a smooth real-analytic
hypersurface, passing through the origin, and choose a small neighborhood $U$
of the origin. In this case
we associate a complete second order system of holomorphic PDEs to $M$,
which is uniquely determined by the condition that the differential equations are satisfied by all the
graphing functions $h(z,\zeta) = w(z)$ of the
Segre family $\{Q_\zeta\}_{\zeta\in U}$ of $M$ in a
neighbourhood of the origin.
To be more explicit we consider the
so-called {\em complex defining
equation } (see, e.g., \cite{ber})\,
$w=\rho(z,\bar z,\bar w)$ \, of $M$ near the origin, which one
obtains by substituting $u=\frac{1}{2}(w+\bar
w),\,v=\frac{1}{2i}(w-\bar w)$ into the real defining equation and
applying the holomorphic implicit function theorem.
The Segre
variety $Q_p$ of a point
$$x=(a,b)\in U,\,a\in\CC{n},\,b\in\CC{}$$
is now given
as the graph
\begin{equation} \Label{segredf}w (z)=\rho(z,\bar a,\bar b). \end{equation}
Differentiating \eqref{segredf} we obtain
\begin{equation}\Label{segreder}
w_{z_j}=\rho_{z_j}(z,\bar a,\bar b),
\quad
j=1,\ldots,n.
\end{equation}
Considering \eqref{segredf} and \eqref{segreder} as a holomorphic
system of equations with the unknowns $\bar a,\bar b$,
in view the Levi-nondegeneracy of $M$,
an
application of the implicit function theorem yields holomorphic functions
$A_1,...,A_n, B$ such that
\eqref{segredf} and \eqref{segreder} are solved by
$$
\bar a_j=A_j(z,w,w'),\quad
\bar b=B(z,w,w'),
$$
where we write
$$
w' = (w_{z_1}, \ldots, w_{z_n}).
$$
The implicit function theorem applies here because the
Jacobian of the system coincides with the Levi determinant of $M$
for $(z,w)\in M$ (\cite{ber}). Differentiating \eqref{segredf} twice
and substituting the above solution for $\bar a,\bar b$ finally
yields
\begin{equation}\Label{segreder2}
w_{z_kz_l}=\rho_{z_kz_l}(z,A(z,w,w'),
B(z,w,w'))=:\Phi_{kl}(z,w,w'),
\quad
k,l=1, \ldots, n,
\end{equation}
or, more invariantly,
\begin{equation}\Label{segreder2'}
j^2_{(z,w)} Q_x = \Phi(x, j^1_{(z,w)} Q_x).
\end{equation}
Now \eqref{segreder2} is the desired complete system of holomorphic second order PDEs
denoted by $\mathcal E = \mathcal{E}(M)$.
\begin{definition}\Label{PDEdef}
We call $\mathcal E = \mathcal{E}(M)$ \it the system of PDEs
associated with $M$. \rm
We also regard the collection $\{\Phi_{ij}\}_{i,j=1}^n$
as {\em the PDE system defining the CR structure} of a Levi-nondegenerate hypersurface $M$.
\end{definition}
\subsection{The necessity of Condition E}
We now explain the necessity of Condition E for the existence of a smooth CR-diffeomorphism $F$ of $(M,0)$ onto a real-analytic germ $(\tilde M,0)$. Indeed, given such a CR-diffeomorphism $F$ of $(M,0)$ onto $(\tilde M,0)$, we may consider the section $\tilde \Phi$, as in \eqref{segreder2'}, associated with $(\tilde M,0)$. Clearly, Condition E is satisfied by $(\tilde M,0)$ since $\tilde \Phi$, considered as a function on the $1$-jet bundle $J^{1,n}$, already gives the holomorphic extension required in Condition E. Further, we note that the CR-diffeomorphism $F$ extends holomorphically to the pseudoconvex side $U^+$. The latter extension lifts naturally to a fiber-preserving map $\widehat F$ of the pseudoconvex neighborhood $\widehat U^+$ of $\widehat M$ into $J^{1,n}$ which is smooth up to $\widehat M_J$ ($\widehat F$ is the {\em $1$-jet prolongation} of the extension of $F$, see e.g. \cite{cs}). Now, since $F$ transforms formal complexifications of $M,\tilde M$ respectively onto each other, we conclude that the $2$-jet prolongation of the extension of $F^{-1}$ transforms $\tilde \Phi$ into the desired holomorphic extension $\Phi$, as required.
\qed
\section{The sufficiency of Condition E}
In this section, we consider a smooth strictly pseudoconvex hypersurface $M\subset\CC{n+1},\,n\geq 1$, defined near the point $0\in M$ and satisfying condition E. We shall prove that $M$ is CR-diffeomorphic (locally near $0$) to a real-analytic hypersurface $\tilde M\subset\CC{n+1}$.
\subsection{Segre foliation in the space of $1$-jets}
We start by recalling that the affine subset
$$E\simeq\CC{n+1}\times\CC{n}$$ of the
bundle $J^{1,n}\to \CC{n+1}$
of $1$-jets of complex hypersurfaces
is endowed with the canonical
(up to a scalar function multiple)
$1$-form $$\omega_0:=dw-\sum_{1}^n\xi_jdz_j.$$ Here $(z_1,...,z_n,w)=(z,w)$ denote the coordinates in $\CC{n+1}=\CC{n}\times\CC{}$, and $\xi=(\xi_1,...,\xi_n)$ are the respective ``jet''-variables corresponding to the derivatives $w_{z_1},...,w_{z_n}$ respectively. The restriction to $E$ of the canonical projection
$\pi:\,J^{1,n}\mapsto \CC{n+1}$ then becomes
\begin{equation}\Label{PI}
\pi:\,\,(z,w,\xi)\mapsto (z,w)
\end{equation}
and $E$ consists precisely
of the $1$-jets of hypersurfaces
that project submersively onto
$\CC{n}\times\{0\} \subset \CC{n+1}$.
The main use of the canonical form $\omega_0$ here is the following. The (complex) tangent bundle $TS$ of a complex hypersurface $S\subset\CC{n+1}$ given as a graph of a function $w=w(z)$ allows to naturally lift $S$ to a complex $n$-dimensional submanifold of $E$. Then an $n$-dimensional submanifold $\tilde S\subset E$ of the kind $$w=w(z),\,\xi=\xi(z)$$ is a lifting in the above sense of a complex hypersurface $S\subset\CC{n+1}$ if and only if $\omega_0|_{\tilde S}=0$.
Further, we observe that, in the case of a {\em real-analytic} hypersurface $M\subset\CC{n+1}$,
the associated system \eqref{wzkzl} amounts to an integrable holomorphic
$n$-distribution in $E$ given by the condition:
\begin{equation}\Label{distrib}
\omega=(\omega_0,\omega_1,...,\omega_n)=0,
\end{equation}
where
\begin{equation}\Label{omegak}
\omega_k:=d\xi_k-\sum_1^n\Phi_{kl}dz_l,\,\,k=1,...,n
\end{equation}
(in the sense that the leaves of the foliation $\mathcal F$ determined by \eqref{distrib} are precisely the lifts to $E$ of graphs of solutions for \eqref{wzkzl}).
In accordance with the latter observation, let us denote the smooth symmetric matrix function \eqref{Phi} (or respectively \eqref{Phi1}) associated with a smooth hypersurface $M$ satisfying Condition E by $\Phi$, and consider the associated complex valued differential $1$-forms $\omega_1,...,\omega_n$, defined by \eqref{omegak}.
In view of Condition E, the function $\Phi$ and hence all the $1$-forms $\omega_0,\omega_1,...,\omega_n$ extend holomorphically to the pseudoconvex side $\widehat U^+$ of $\widehat M$, defining there a holomorphic $n$-distribution $D$. Alternatively, $D$ is spanned by the $n$ holomorphic vector fields
\begin{equation}\Label{Lj}
L_j:=\frac{\partial}{\partial z_j}+\xi_j\frac{\partial}{\partial w}+\sum_1^n \Phi_{sj}\frac{\partial}{\partial \xi_s}.
\end{equation}
We then have
\begin{proposition}\Label{integrable}
The distribution $D$ in $\widehat U^+$ is integrable.
\end{proposition}
\begin{proof}
Integrability of the distribution $D$ amounts to the conditions
\begin{equation}\Label{integrability}
L_j\Phi_{kl}-L_k\Phi_{jl}=0,\quad j,k,l=1,...,n,
\end{equation}
where $L_j$ are as in \eqref{Lj}.
(In terms of the system \eqref{wzkzl}, conditions \eqref{integrability} mean simply the symmetry of third order derivatives of $w$ in all indices). In view of the Condition E, the left hand side in \eqref{integrability} extends smoothly to the real hypersurface $\widehat M\subset E$. We claim that the latter extension vanishes on the totally real manifold $M_J$. Indeed, the fact of vanishing of the left hand side in \eqref{integrability} when restricted on $M_J$ is nothing but the symmetry of the third order jet of formal Segre varieties of $M$ in all indices, which proves the claim. Since $M_J$ is totally real of dimension $2n+1$, this implies that the left hand side in \eqref{integrability} vanishes identically in $\widehat U^+\cup \widehat M$, as required.
\end{proof}
\autoref{integrable} implies the existence of an $n$-dimensional holomorphic foliation $\mathcal F$ in $\widehat U^+$ generated by $D$. We specify that by {\em leaves} of the foliation $\mathcal F$ we mean maximal connected components of integral submanifolds of $\mathcal F$.
\begin{definition} In what follows we call $\mathcal F$
{\em the Segre type foliation in $\widehat U^+$.}
\end{definition}
\subsection{Changing the complex structure on the pseudoconcave side of $M$}
In this section, we show that the pseudoconcave side $U^-$ of $M$ can be interpreted as the space of leaves for the Segre foliation $\mathcal F$ constructed above, and this endows $U^-$ with a {\em different} (integrable) complex structure, which is smooth up to $M$ and which induces on $M$ a CR-structure coincident with the initially given CR-structure on $M$ induced from $\CC{n+1}$.
We construct the desired complex structures
in multiple steps discussed in detail below.
\medskip
{\noindent \bf Step I.} We recall that the distribution $D$ above can be, according to Condition E, smoothly extended to $\widehat U^+\cup\widehat M$ as a function valued in the complex Grassmannian $\mbox{Gr}(n,E)$. (Note though that this extension is {\em not} everywhere tangent to $\widehat M$!). Furthermore, we note that $M_J$ is precisely the locus of points in $\widehat M$, for which the value of the extension of $D$ {\em is tangent} (and hence complex tangent) to $\widehat M$. This follows directly from the construction of $M_J,\widehat M$.
\medskip
{\noindent \bf Step II.}
Our next goal is to show that the space of leaves of the Segre foliation
$\mathcal F$ is a smooth manifold in its natural (quotient) topology, which can be furthermore extended to a smooth manifold with boundary $M_J$.
We will make use of the following
\begin{proposition}\Label{extension}
Let $U\subset\RR{m}$ be a neighborhood of the origin and $M\subset U$ a smooth strictly convex hypersurface through the origin and, furtheremore, one has $T_0M=\{x_m=0\}$ and the
second fundamental form of $M$ at $0$ equals to
\begin{equation}\Label{ksum}
dx_1^2+...+dx_{k}^2
\end{equation} for some $1\leq k <m$. Let $U^+$ be the convex side of $M$, $U^-$ the concave side of $M$, and $D$ a smooth $k$-dimensional integrable distribution in $U^+$. Assume that
\smallskip
(i) $D$ extends to $M$ smoothly (as a function valued in $\mbox{Gr}(k,\RR{m})$);
\smallskip
(ii) The $k$-plane $D_0$ at $0$ is spanned by $\frac{\partial}{\partial x_1},...,\frac{\partial}{\partial x_k}$ (in particular, $D_0\subset T_0 M$) and, moreover, the $\frac{\partial}{\partial x_{k+1}},...,\frac{\partial}{\partial x_{m}}$ components of
some collection of vector fields spanning $D$ have zero linear parts
at the origin.
\smallskip
\noindent Then, after possibly changing the neighborhood $U$, the distribution $D$ extends to a smooth integrable distribution in $U$. Furthermore, if $D$ is given by
smooth up to $M$, (pointwise) linearly independent and commuting vector fields $X_1,..,X_k$ in $U^+$, then these vector fields can be extended smoothly to $U$ in such a way that the extensions still commute. The foliation $\mathcal F$ in $U^+$ generated by $D$ extends
therefore to a smooth $k$-dimensional foliation $\mathcal F'$ in $U$, in the sense that
each leaf of $\mathcal F$ is an open subset of a unique leaf of $\mathcal F'$ and each intersection of a leaf of $\mathcal F'$ with $U^+$ is connected.
\end{proposition}
\begin{proof}
We prove the proposition by induction.
For $k=1$, we choose a vector field $X$ generating $\mathcal F$. We split $x=(x_1,\tilde x)$. In view of (ii), we may assume
\begin{equation}\Label{conv}
X=p\frac{\partial}{\partial x_1}+q\frac{\partial}{\partial \tilde x},\quad p(0)=1,\,\,q(0)=0,\,\,\frac{\partial q}{\partial x_1}=0.
\end{equation}
Consider a smooth extension $X'$ of $X$ to $U$ (possible by (i)). Then \eqref{conv} immediately implies that the orbit of $X'$ at $0$ has the form
$$x_1=t,\quad \tilde x=O(t^3), \quad t\in(-\epsilon,\epsilon).$$
Now if, for example, $m=2$ (so that $\tilde x=x_2$), we consider the defining equation $x_2=\psi(x_1)$ of $M$ as well as the definining equation $x_2=\phi_0(x_1)$ of the orbit. Then \eqref{ksum} implies that $(\psi-\phi)''>0$ on $(-\epsilon,\epsilon)$ for small elough $\epsilon$, that is why the part of the orbit lying inside $U^+=\{x_2>\psi\}$ is the set of negative values of a convex function, i.e. the intersection of the orbit with $U^+$ is {\em connected}. By continuity, after shrinking possibly $U$, the same argument applies for all points nearby $0$, and this proves the proposition for $k=1$ and $m=2$. The case $k=1$ and $m>2$ can be reduced to the previous one by considering the intersection with a $2$-dimensional surface containing the orbit and the $x_m$ coordinate axis. We leave the details to the reader.
We now proceed with the induction step.
We choose $k$ linearly independent smooth vector fields $X_1,...,X_{k}$ spanning the distribution $D$ in $U^+$ and extending smoothly to $M$. Following a proof of the Frobenius theorem (e.g. \cite{book}), it is not difficult to show that we can choose $X_1,...,X_k$ to be furthermore {\em commuting}. Indeed, we have:
$$X_i=\sum_{j=1}^m \alpha_{ij}\frac{\partial}{\partial x_j}$$
for smooth up to $M$ functions $\alpha_{ij}$ in $U^+$
($(x_1,...,x_m)$ are the coordinates in $\RR{m}$). Since $X_i$ are linearly independent, we can assume without loss of generality that the matrix
$$\left(\alpha_{ij}\right)_{i,j=1}^k$$
is invertible in $U^+\cup M$, and consider the inverse smooth matrix $(\beta_{ij})$. Then it is straightforward to check that the vector fields
$$\sum_{j=1}^k\beta_{ij}X_j,\quad 1\leq i\leq k,$$ obviously spanning the same distribution in $U^+$ and smooth up to $M$, in fact in addition {\em commute}.
The latter allows us to consider the {\em integrable} distribution generated by the commuting vector fields $X_1,...,X_{k-1}$. Applying then the induction assumption, we obtain smooth {\em commuting} extensions $X_1',..,X_{k-1}'$ of $X_1,..,X_{k-1}$ and hence a $(k-1)$-dimensional integrable distribution in $U$. The foliation $\mathcal X$ given by this distribution has the property that its leaves can have only connected intersections with $U^+$. After that, let us perform (after possibly shrinking $U$) a smooth in $U$ diffeomorphism,
tangent to the identity at $0$ and
transforming the vector fields $X_1',...,X_{k-1}'$ to $\frac{\partial}{\partial x_1},...,\frac{\partial}{\partial x_{k-1}}$ respectively (the latter is possible since the vector fields commute). Now all the leaves of $\mathcal X$ become parallel to the $(x_1,...,x_{k-1})$-plane. We keep the same notation for $U,U^\pm$ in the new coordinates.
Then the vector field $X_k$ (defined so far in the closure of $U^+$)
have the form
\begin{equation}\Label{YY}
X_k=a_1(x_k,...,x_m)\frac{\partial}{\partial x_1}+\cdots+a_m(x_k,...,x_m)\frac{\partial}{\partial x_m}
\end{equation}
where
we use the commutativity $[X_j,X_k]=0,\,j=1,...,k-1$).
Consider the orthogonal projection $\Omega$ of the closure of $U^+$ onto the $(x_k,...,x_m)$-plane. Then the leaves of $\mathcal X$ intersecting the closure of $U^+$ project onto a single point in $\Omega$. As follows from \eqref{YY}, the vector field $X_k$ is {\em constant} on each leaf of $\mathcal X$, which allows to extend $X_k$ constantly along each leaf intersecting the closure of $U^+$. In view of the above, this gives a smooth function on $\Omega$, which we first extend smoothly to a neighborhood of the origin in the $(x_k,...,x_m)$-plane, and then again constantly along each leaf of $\mathcal X$.
Since the intersection of each leaf of $\mathcal X$ with the closure of $U^+$ is connected, the extension obtained is well defined.
We thus are able to extend $X_k$ smoothly to a full neighborhood of the origin still being constant on each leaf of $\mathcal X$.
In view of the latter property, the extended vector field $X_k'$ also
satisfies $[X_j',X_k']=0,\,1\leq j\leq k-1$.
In summary, we obtain an integrable distribution in $U$ spanned by $X_1',...,X_k'$. For the respective foliation $\mathcal F'$, each leaf of $\mathcal F$ is clearly contained in that of $\mathcal F'$. It remains to show that intersections of leaves of $\mathcal F'$ with $U^+$ are connected. This however can be seen from (ii) by an argument identical to the one in the $1$-dimensional case.
The proof is complete.
\end{proof}
We now apply \autoref{extension} to the situation of the
strictly pseudoconvex hypersurface $\widehat M\subset E$, its neighborhood $\widehat U$ and the distribution $D$ (considered as a real distribution). For that, we perform a biholomorphic (in fact polynomial) coordinate change mapping the origin and the tangent plane $\ensuremath{\mbox{\rm Im}\,} w=0$ onto themselves, and removing holomorphic quadratic terms from the formal Taylor expansion of $M$ in the origin
such that property \eqref{ksum} holds for $\widehat M$.
In fact, in such coordinates $M$ becomes approximated by a quadric to order $2$ at the origin:
\begin{equation}\Label{approx}
\ensuremath{\mbox{\rm Im}\,} w=Q(z,\bar z)+O(3),
\end{equation}
where $Q$ is a positive definite Hermitian form (we assume it to be simply sum of squares) and $O(3)$ stand for terms of degree $3$ and higher. In case $M$ is already this quadric itself, $D$ is spanned by the (real parts of) the vector fields
$$L_j:=\frac{\partial}{\partial z_j}+\xi_j\frac{\partial}{\partial w},$$
so that (ii) is satisfied. However, terms of order $3$ and higher in \eqref{approx} do not effect (ii), and
this finally shows that $\widehat M$ satisfies all the conditions of \autoref{extension}.
We end up with a smooth extension of the distribution $D$ to a full neighborhood of the origin in $E$ in such a way that it is still integrable and defines a foliation $\mathcal F'$ extending $\mathcal F$, in the sense that each leaf of $\mathcal F$ is an open subset of some leaf of $\mathcal F'$ and, furthermore, leaves of $\mathcal F'$ have only connected intersections with $\widehat U^+$ (so that each leaf of $\mathcal F$ is contained in exactly one leaf of $\mathcal F'$).
\bigskip
{\noindent \bf Step III.}
Based of the outcome of Step II, we are finally able
to endow the space of leaves of $\mathcal F$
with the structure of a smooth $(2n+2)$-manifold with boundary $M_J$ in the natural (quotient) topology.
Indeed, first note that the tangent plane at $0$ to the leaf of $\mathcal F'$ through $0$ is $w=0,\,\xi_j=0$ (as follows from the definition and the initial normalization of $M$). Hence, the $(2n+2)$-plane
\begin{equation}\Label{z0}
\{z=0\}
\end{equation}
has the property that all the leaves of (the ambient foliation) $\mathcal F'$ intersect it transversally at single points
(after possibly shrinking $\widehat U$).
Thus the space of leaves of $\mathcal F'$ can be identified with a domain in \eqref{z0} (viewed as $\RR{2n+2}$). Accordingly, the space of leaves of $\mathcal F$ is an open connected subset $G$ of the latter domain (since it is given by the condition of having nonempty intersection with the domain $\widehat U^+$).
This gives a structure of a smooth manifold on the space of leaves of $\mathcal F$ (in its natural quotient topology)!
It remains to show that the space of leaves of $\mathcal F$ has, furthermore, a structure of a smooth $(2n+2)$-manifold with a boundary.
Indeed, as follows from the discussion in Steps I and II, any leaf of $\mathcal F'$ intersecting the closure of $\widehat U^+$ either intersects $\widehat M$ transversally and consequently intersects the open part $\widehat U^+$, or intersects the boundary $\widehat M$ only at a point in $M_J$, or is contained in $\widehat U^+$ (the last possibility in fact does not occur, but we do not need this fact). In this way, {\em $M_J$ can be identified with the subset of leaves of $\mathcal F'$ intersecting the boundary $\widehat M$ but not the open part $\widehat U^+$}. Thus, we are looking for the set of leaves of $\mathcal F'$ which are tangent at some point to $\widehat M$. The latter set is a smooth $(2n+1)$-submanifold in $\RR{2n+2}$ (which is in fact the boundary of the above open set $G$ of leaves of $\mathcal F'$ intersecting $\widehat U^+$). Indeed, perform a local diffeomorphism (near the origin in $E\sim\RR{4n+2}$) with the identity linear part in such a way that in the new local coordinates the leaves become "horizontal" $2n$-planes (that is, they are given by $$x_j=c_j,\,j=2n+1,...,4n+2,$$ where $c_j$ are constant), and the hypersurface $\widehat M$ becomes
$$
x_{4n+2}=\varphi(x_1, \ldots ,x_{4n+1})
$$
for a smooth function $\varphi$. Note that, in particular, $\varphi$ has nondegenerate at the origin Hessian in $x_1,...,x_{2n}$ (as follows, for example, from \eqref{approx}). Now the condition that a leaf is tangent to $\widehat M$ at some point is:
\begin{equation}\Label{tang}
c_{4n+2}=\varphi(x_1,...,x_{2n},c_{2n+1},...,c_{4n+1}),\quad \varphi_{x_j}(x_1,...,x_{2n},c_{2n+1},...,c_{4n+1})=0,\,\, j=1,..,2n.
\end{equation}
Solving the last $2n$ equations of \eqref{tang} in $x_1,...,x_{2n}$ by the implicit function theorem and substituting the result into the first equation of \eqref{tang}, we obtain a smooth hypersurface
$$c_{4n+2}=\psi(c_{2n+1}, \ldots ,c_{4n+1})$$
in $\RR{2n+2}$ (endowed with the coordinates $c_{2n+1},...,c_{4n+1}$) for an appropriate smooth function $\psi$ with $\psi(0)=d\psi(0)=0$, as desired.
\bigskip
We thus have proved the following
\begin{proposition}\Label{manifold}
The leaf space of the foliation $\mathcal F$ is a smooth $(2n+2)$-manifold in its natural (quotient) topology. Furthermore, it can be regarded as a smooth $(2n+2)$-manifold with boundary $M_J$. The (germ at the origin of) the upper half space
$$\bar H=\{(x_{2n+1},...,x_{4n+2}):\,\,x_{4n+2}\geq 0\}$$ serves as a coordinate chart for it.
\end{proposition}
\medskip
{\noindent \bf Step IV.}
We denote the leave space from \autoref{manifold} by $\mathcal U$, and the respective manifold with boundary by $\overline{\mathcal U}$. As was mentioned above, $\mathcal U$ represents, in a certain sense, the pseudoconcave side $U^-$ endowed however with a {\em different} (still integrable) complex structure.
We
next note that the leave space $\mathcal U$ is, on the other hand, the quotient topological space $\widehat U^+/\mathcal F$, and this means (if read together with \autoref{manifold}) that $\mathcal U$ has also a structure of a {\em complex} manifold with boundary. We emphasize at this point that, exclusively for the purpose of obtaining the right complex structure on $\widehat U^+/\mathcal F$,
\smallskip
{\em we shall change the complex structure on
$\mathcal U$ to its conjugate structure.}
\smallskip
This is related to the antiholomorphic dependence of Segre varieties on their parameters, for a real-analytic hypersurface. Thus we obtain a smooth integrable complex structure on $\mathcal U$, which extends further to a smooth CR-structure on $\partial\mathcal U$. The boundary $\partial\mathcal U$ is naturally diffeomorphic to $M_J$ and hence to $M$. We will show later that the two CR-structures on $M$ (the one coming from the quotient space and the one induced from the embedding into $\CC{n+1}$) in fact agree with each other together with all higher order derivatives.
\medskip
{\noindent \bf Step V.}
We now have to take into consideration the ``one sided Segre varieties of $M$'', that is, images of leaves of the foliation $\mathcal F$ under the projection map $\pi$, as in \eqref{PI}. This gives us the family
$$\mathcal S^+:=\Bigl\{\pi(T)\Bigr\}_{T\in \mathcal F}.$$
Note that, as the extension construction in Step II above shows, all the leaves in $\mathcal F$
(after
possibly
shrinking
the basic neighborhood $\widehat U$) are open subsets of graphs of smooth functions of the kind $w=w(z),\,\xi=\xi(z)$ with
connected intersections with $\widehat U^+$. In this way, all elements of $\mathcal S^+$ are $n$-dimensional complex submanifolds in $U^+$, and we conclude that
{\em $\mathcal S^+$ is an $(n+1)$-dimensional anti-holomorphic family
of
pairwise transverse
complex $n$-dimensional submanifolds in $U^+$ of the form $w=w(z)$}.
(Transversality here means that no two elements of $\mathcal S^+$ are tangent at a point $p\in U^+$).
The anti-holomorphic parametrization of $\mathcal S^+$ here comes from the integrable complex structure on the space of leaves $\mathcal U$.
We note that $\mathcal U$ itself is endowed with a natural $(n+1)$-dimensional anti-holomorphic family of $n$-dimensional complex submanifolds as follows. We fix a point $p\in U^+$ and consider the set $S_p$ of all the manifolds from $\mathcal S^+$ passing through $p$ as a subset of $\mathcal U$. Following the geometric interpretation in the real-analytic case, we call $S_p$ {\em the Segre variety of $p$}. The structure of $S_p$ becomes particularly clear when considering the foliation $\mathcal F$: then the set of all elements of $\mathcal S^+$ passing through $p$ lifts to the set of all fibers in $\mathcal F$ intersecting the fiber $\pi^{-1}(p)$ of the bundle $E$. In this way, we easily see, from the construction of the manifold $\mathcal U$, that each $S_p$ can be identified via the $1$-jet map
with the fiber $\pi^{-1}(p)$ and thus is a complex $n$-dimensional submanifold in $\mathcal U$ (with respect to the above described complex structure on $\mathcal U$), as required. We denote the resulting family of submanifolds in $\mathcal U$ by $\mathcal S^-$ (it becomes an anti-holomorphic family parameterized by $U^+$).
For completeness of the picture, we also call, for each $p\in\mathcal U$, the respective leaf $T\in \mathcal S^+$ its {\em Segre variety} and denote the latter one by $S_p$. We then obtain the following familiar symmetry property:
$$p\in S_q \Leftrightarrow q\in S_p, \quad p\in U^+,\,\,q\in \mathcal U.$$
\medskip
{\noindent \bf Step VI.} We recall that the boundary manifold $\partial\mathcal U$ is naturally diffeomorphic to the initial CR-manfold $M$, as follows from the construction of $\overline{\mathcal U}$. We further extend (locally near $0$) the latter diffeomorphism smoothly to a diffeomorphism between the manidold with boundary $\overline{\mathcal U}$ and the pseudoconcave side $U^-$ of $M$ (which is possible since both are
manifolds
of equal dimension
with boundary).
We end up with a smooth manifold $U$ decomposed as a union of two manifolds with boundary:
$$U=(U^-\cup M)\cup (U^+\cup M),$$
where both $U^-$ and $U^+$ are endowed with their individual complex structures (for $U^-$ this is the integrable structure induced from $\mathcal U$ and for $U^+$ this is the standard complex structure induced from $\CC{n+1}$). Moreover, both structures admit a smooth extension to the boundary and induce boundary CR-structures on $M$. Our goal is now to show that the two structures (considered for the moment as $(2n+2)\times (2n+2)$ matrices) agree on $M$ (together with all derivatives). In particular, they define a smooth structure in a full neighborhood of the origin, and the two induced CR-structures on $M$ coincide.
For doing so, let us fix $p\in M$ and the respective point $\tilde p\in\partial\mathcal U$. We observe the following:
all
data required for computing the boundary value at $\tilde p$ of the complex structure on $\mathcal U$ comes from the $2$-jet of $M$ at $p$ (as follows from our construction). Similarly, for computing the $k$-jet at $p$ of the limit
of the complex
structure we just need to know the $(k+2)$-jet of $M$ at $p$. Since $M$ can be approximated to any order by a real-analytic hypersurface, we conclude that it suffies to show that for a {\em real-analytic} hypersurface $M$ the two above structures coincidence and define a real-analytic (in particular smooth) structure in a neighborhood of $p$.
If now $M$ is real-analytic, then, as a well know fact (e.g. \cite{ber}), after choosing an appropriate neighborhood $U$ of $p$, we have the property that a Segre variety of a point $q\in U$ intersects the pseudoconvex side $U^+$ iff $q$ lies in the pseudoconcave side $U^-$ of $M$. Thus, if $M=\{\rho(Z,\bar Z)=0\}$ near $p$, then the above manifold $\mathcal U$ consists of Segre varieties
$\{\rho(Z,\bar q)=0\}$
with $q\in U^-$, and thus can be identified with $U^-$
with the standard complex structure on it,
while its boundary can be identified with $M=\{\rho(q,\bar q)=0\}$ with the standard CR-structure on it.
This immediately yields the desired property.
\medskip
\subsection{End of proof of the main result}
In this section, we complete the proof of the main result.
\smallskip
Recall that, as an outcome of Step VI, we obtain a smooth manifold $U$ endowed with an {\em integrable} smooth complex structure $J$ (the integrability follows from that on both $U^-$ and $U^+$ and thus, by continuity, at points in $M$ as well). By the Newlander-Nirenberg theorem \cite{nirenberg}, there is a smooth diffeomorphism $\chi$ of $(U,J)$ (preserving the origin), mapping $U$ onto a neighborhood of the origin
and transforming the above complex structure $J$ on $U$ into the standard complex structure $J_{st}$ in $\CC{n+1}$ (that is, $\chi$ is a $(J,J_{st})$-biholomorphism). The resulting smooth strictly pseudoconvex image of $M$ we still denote by $M$, and the pseudoconvex and the pseudoconcave sides of it respectively we still denote by $U^\pm$.
We shall consider now the above families $\mathcal S^\pm$ on $U^\pm$ respectively, after applying the diffeomorphism $\chi$.
We recall that elements of both families $\mathcal S^\pm$ are $J$-invariant and have the transversality property, hence they become
{\em
families of holomorphic curves} on $U^\pm$ respectively. This allows us to consider, in the same fashion as in Section 2, their lifting to the space of $1$-jets $J^{1,n}$, and this results in two foliations defined in some domains
$\widehat U^\pm\subset E$ with $\pi(\widehat U^\pm)=U^\pm$ respectively (here $E$ is again the affine subset of the bundle of $1$-jets of complex hypersurfaces). In particular, we may consider the respective holomorphic direction fields defined in the same domains. As follows from the above, these two directions fields extend smoothly to $M_J$.
Importantly, we can {\em not} conclude at this step that the domain $\widehat U^-$ has the form $U^-\times\Omega$, where $\Omega$ is an open neighborhood of the origin in $\CC{n}$. This is because $1$-jets of elements of $\mathcal S^-$ passing through a point $p^-\in U^-$ do not cover a full neighborhood of the origin of a uniform size. To see the latter, we note that the
"pencil" of Segre varieties passing through $p^-\in U^-$ is the union of
"pencils"
of Segre varieties from $\mathcal S^+$ at points belonging to the Segre variety $S_{p^-}\in\mathcal S^+$. Unlike the situation in the real-analytic case, $S_{p-}$ is defined so far {\em only} as a variety in $U^+$,
which is not a full neighborhood of $0$.
(In the real-analytic case, such "one-sided" Segre varieties extend analytically across $M$ and become varieties in a uniform neighborhood of $0$).
That is why possible jets of Segre varieties through $p^-$ in our construction form a "smaller" set in the space of $1$-jets (compared to the real-analytic case) and thus do not give a uniform neighborhood of the origin.
(However, the domain $\widehat U^+$ does have the desired form $U^+\times\Omega$ for an open neighborhood $\Omega$ of the origin).
To overcome the latter difficulty, we use Webster's considerations from \cite{websterrefl} to conclude that, after possibly
shrinking the neighborhood $U$:
\medskip
\noindent (i) There exists a wedge $W^+\subset\widehat U^+$ with the totally real edge $M_J$;
\smallskip
\noindent (ii) The {\em reflection map} $\tau$ defined as
\begin{equation}\Label{tau}
\tau(z, T_z S_\zeta)=(\zeta,T_\zeta S_z), \quad z\in\widehat U^\pm,\,\, \zeta\in S_z
\end{equation}
is anti-holomorphic, well defined in $\widehat U^\pm$, and smooth up to $M_J$;
\smallskip
\noindent (iii) The image $\tau(W^+)$ contains a wedge $W^-\subset\widehat U^-$, and $\tau$ satisfies
$$\tau\circ\tau=\mbox{Id}$$
(thus $\tau$ is an anti-holomorphic involution);
\smallskip
\smallskip
\noindent (iv) $\tau|_{M_J}=\mbox{Id}$.
\medskip
To prove (i)-(iii) we note that, even though \cite{websterrefl} deals with the real-analytic case, the proof of (i)-(iii) in \cite{websterrefl} is based solely on the approximation of $M$ by a quadric to order $3$ and the implicit function theorem subsequently, that is why it can be essentially word-by-word repeated in our situation. We leave the details here to the reader. Statement (iv) follows directly from the construction in Section 4.2 (alternatively, in can be viewed from (i)-(iii) read together).
We now use the edge-of-the-wedge theorem (e.g. \cite{rosay}) to conclude that $\tau$ extends anti-holomorphically to a full neighborhood of the origin in $E$ (still being an involution, by uniqueness). For simplicity, we denote the latter neighborhood by $\widehat U$.
Let us consider finally (locally near the origin) the fixed point set $M'\subset U$ of $\tau$. In view of the fact that $\tau$ is an anti-holomorphic involution, $M'$ is a {\em real-analytic} totally real submanifold in $\widehat U$ of dimension $2n+1$. In view of (iv), we conclude that $M'=M_J$, i.e. both $M_J$ and $M$ are {real-analytic}. The proof
of
\autoref{main}
is complete.
\qed
|
\section{Introduction}
Nonlinear optics spans a broad class of phenomena that involve light-induced variation of optical properties of a system. Interestingly, the nonlinearity in optical systems often originates from an inherently quantum mechanical process, but the description of the resulting output can be either classical or quantum, depending on the system under consideration and intensity of the light fields. For instance, nonlinear crystals implementing frequency mixing and stimulated scattering~\cite{boyd2003nonlinear} are described using classical descriptions rooted in average nonlinear material susceptibilities. On the other hand, a new class of effects and non-classical states emerge on quantizing the light fields, which are dealt with in the framework of quantum optics~\cite{carmichael2009statistical}. In order to distinguish between these two descriptions, it is important to understand the transition from quantum to classical dynamics, especially given the burgeoning presence of nonlinear systems in applications triggered by quantum information processing~\cite{nielsen2010quantum, RevModPhys.77.513}.
An example of a nonlinear phenomenon that is of importance in both quantum and classical optics is the Kerr effect, which changes the optical properties in proportion to the intensity of the incident field. Under a coherent drive and single photon loss, the Hamiltonian of the $k$th mode of a physical system subject to the Kerr effect is given by ($\hbar = 1$)
\begin{align}
H_k = \omega_k \, a_k\dag a_k - K_k \, ( a_k\dag a_k )^2 + \varepsilon_k(t)\,a_k\dag + \varepsilon_k^*(t)\,a_k, \label{eq:hamiltonian}
\end{align}
and the dynamics of the mode, under the assumption of Markovian decay, is described by the master equation
\begin{align}
\dot{\rho} = -i [ H_k,\rho] + \kappa_k \Big[a_k \rho a_k\dag - \frac{1}{2} (a_k\dag a_k \rho + \rho a_k\dag a_k) \Big]. \label{eq:mastereq}
\end{align}
Here, $\omega_k$ is the bare resonance frequency mode, $K_k$ is the Kerr coefficient responsible for shifting the mode frequency, $\varepsilon_k(t) = \varepsilon_k e^{-i\omega_d t}$ is the amplitude of a coherent drive with angular frequency $\omega_d$ and $\kappa_k$ is the single-photon decay rate of the mode. We note that the Hamiltonian above assumes that all nonlinear modes are independent; in principle, there can be additional cross-Kerr coupling between different $k$ modes of the system. This driven-dissipative Kerr resonator is ubiquitous in physical systems spanning fiber optics~\cite{agarwal2006fiber}, superconducting quantum circuits with Josephson junctions (JJs)~\cite{Yurke:87,castellanos2008amplification,PhysRevB.86.024503,PhysRevA.86.013814,leib2012networks,Eichler2014,PhysRevApplied.5.024002}, optomechanical systems~\cite{Bose1997,Ludwig2012} and atomic ensembles~\cite{Gupta2007}.
In particular, JJ-based superconducting circuits naturally realize strong Kerr nonlinearities at the single photon level, which have been extensively utilized both for fundamental quantum optical studies \cite{bozyigit2011antibunching, kirchmair2013observation} and quantum information-inspired applications~\cite{PhysRevA.75.032329, corcoles2015demonstration, kelly2015state}. This breadth of applications is enabled by the highly flexible nature of the nonlinearity realized in superconducting circuits, which can be tuned either in-situ through the flux-tunable SQUID-based designs \cite{Bell2012} or ex-situ through appropriate selection of junction parameters; the latter is usually accomplished by designing junctions with different ratios of Josephson energy and charging energy or, alternatively, by using arrays of junctions~\cite{Eichler2014, PhysRevLett.109.137002, PhysRevB.92.104508}.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{kerr_regimes_v6.pdf}
\caption{Different regimes of Kerr resonator as a function of relative anharmonicity $K_k/\omega_{k}$ which is inversely related to the number of levels in the potential well, and the quality factor $\omega_k / \kappa_k$. The lower left corner corresponds to parametric and bifurcation amplifiers while the upper right corner corresponds to the transmon regime. Each of these regimes is accompanied by a sketch of the average steady-state population of the oscillator, $|\alpha_{k}|^{2}$, in response to a drive amplitude $\varepsilon_{k}$ detuned from resonance by $\Delta_k = \omega_{k} -\omega_{d}$ [see Appendix \ref{app:classical}]. The central region is the intermediate ``mesoscopic" regime that is explored theoretically and experimentally in this work.} \label{fig:kerrregimes}
\end{figure}
Figure~\ref{fig:kerrregimes} maps the different nonlinearity regimes readily accessible with JJ circuits. For nonlinearities much smaller than the mode linewidth ($K_k\ll \kappa_k$), the dynamics can be entirely captured by that of a driven classical Duffing oscillator~\cite{dykman2012fluctuating}. This is the domain of the Josephson parametric amplifiers (JPA) \cite{PhysRevLett.109.050507, PhysRevLett.109.050506, lin13singleshot} and Josephson bifurcation amplifiers (JBA) \cite{mallet2009single, vijay2009invited} that are routinely employed for quantum-limited measurements \cite{Murch2013, Sun2014}. Of particular interest in this regime is the bistable behavior occurring when a lossy Kerr resonator is strongly driven. In this bistable regime, the Kerr resonator may switch between two stable states corresponding, respectively, to small and large photon populations of the driven resonator~\cite{muppalla2017bi}. For weak nonlinearity, this switching behavior is well-described by a semi-classical quantum activation theory \cite{vijay2009invited, dykman1980fluctuations, dykman2012fluctuating}. On the other hand, for large nonlinearity ($K_k\gg \kappa_k$), the system enters the photon blockade regime in which the nonlinearity-induced frequency shift limits the number of photons in the oscillator under a drive of fixed frequency. The nonlinear oscillator then becomes an effective two- or `few'-level system; this is the so-called transmon regime \cite{PhysRevA.76.042319}. In contrast to the semi-classical picture, oscillators in this regime do not exhibit any bistability or hysteresis and a full quantum description is necessary to describe their dynamics.
A thorough understanding of fluctuation-induced switching rates in the cross-over region is important for understanding and optimizing Kerr oscillators, both from fundamental physics and application points of view. For instance, nonlinear oscillators exhibit scale invariant behavior near bistability; also, the time needed to switch from one metastable state to the other limits the qubit measurement time in JBAs \cite{mallet2009single, vijay2009invited}. However, the crossover of driven-dissipative dynamics from (semi)-classical to the quantum regime remains poorly understood. In this work, we perform a detailed investigation of this intermediate mesoscopic regime between semi-classical bistability and the transmon regimes for driven Kerr oscillators. The workhorse of our studies is a Kerr oscillator in the strongly nonlinear mesoscopic regime, based on a superinductance formed from an array of Josephson junctions~\cite{VladThesis}. The eigenmodes of the array form highly nonlinear oscillator modes, where the nonlinear shift per photon is larger than the natural oscillator linewidths. Specifically, we use the time scale, here labelled $\tau$, associated with decay into the steady state of the Kerr resonator as a benchmark to delineate semi-classical and quantum dynamics. We report the results of an experiment with a nonlinear resonator realized with a superconducting quantum circuit. The results are simulated using a stochastic master equation and we find signatures of oscillator relaxation time $\tau$ much longer than the intrinsic decay time $1/\kappa$, which motivates further theoretical investigations. We go beyond a linear treatment and present both numerical and analytical calculations of switching rates in this system, considering the situation where the oscillation amplitude is locked in one of the two metastable dynamical states in the presence of a strong drive. Our theoretical studies indicate a breakdown of usual semi-classical treatments that describe oscillator decay primarily as quantum activation in a metapotential or, alternatively, by thermal activation. Instead, in the mesoscopic regime, a quantum treatment is essential to capture the relaxation timsescale of the oscillator. We characterize this transition from a semi-classical to quantum description by introducing a crossover parameter, $\xi = T_{\gamma}/T_{\kappa}$, as a ratio of a temperature associated with tunneling-induced escape $T_{\gamma}$ and effective temperature associated with fluctuations seen by the oscillator $T_{\kappa}$. When $\xi>1$, quantum effects introduce a new decay channel and fluctuation-induced activation is inadequate to describe the switching dynamics in this regime.
This paper is organized as follows: Section \ref{sec:exp} introduces our experimental implementation of a Kerr resonator in the mesoscopic regime realized by an array of Josephson junctions \cite{PhysRevLett.109.137002}. In Sec. \ref{sec:theory}, we present the theoretical model describing the ground state properties and Kerr coupling between the distributed modes of the array. By doing a first principle quantization of the array, we confirm that the Kerr coefficients for its distributed modes lie in the regime of interest, $K_k/\kappa_k \approx 2-5$. We then describe how stochastic master equation simulations accurately predict the experimentally observed nonlinear resonance lineshape for an eigenmode of the array in the presence of an external drive. In Secs.~\ref{sec:numr} and \ref{sec:theo}, we perform detailed numerical and analytical investigations of fluctuation-induced switching rates in order to understand the nonlinear effects observed in the experiment. Sec.~\ref{sec:concl} concludes the paper with a discussion of the main results and provides an outlook for future theoretical and experimental studies. Additional numerical and analytical results are included in Appendicies~\ref{app:array} and \ref{app:classical}.
\section{Experiment with Josephson junction arrays}
\label{sec:exp}
Figure \ref{fig:experimental_setup} depicts the experimental system realizing a mesoscopic Kerr oscillator, which consists of an array of 80 Josephson junctions capacitively coupled to a transmission line in a hanger geometry~\cite{PhysRevLett.109.137002}. The sample is mounted inside a copper box in a dilution fridge at a temperature of 15 mK. Both coherent driving and measurement of the resonator are performed through the transmission line. The transmission signal at the output port~2 is amplified by a HEMT amplifier before being demodulated and recorded.
\par
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{experimental_setup_v3.pdf}
\caption{(a) Schematic of the experimental set up comprising an array of Josephson junctions coupled to a transmisson line. The capacitance $C_s$ is the coupling capacitance between the array and the tranmission line and $C_g$, $C_e$ represent the capacitances to ground at the two ends of the array. (b) Linear lumped element circuit model for the array. Each junction is represented as an LC-circuit with capacitance $C_J$, inductance $L_J$ and an extra parasitic capacitance to ground $C_0$. (c) Rotated SEM image of the junction array.} \label{fig:experimental_setup}
\end{center}
\end{figure}
The resonator has a fundamental frequency of $\omega_0/2\pi = 4.357$~GHz, with an internal quality factor of 37,000 and external quality factor of 5,000. Larger internal quality factors were observed for other samples \cite{PhysRevLett.109.137002} but, in all cases, the quality factor was dominated by the external coupling. The experimental setup is constrained by circulators and a low-pass filter to measure signals with frequencies ${<}\,12$ GHz. As a result, only the fundamental mode of the Kerr resonator can be directly probed. Leveraging the nonlinearity-induced mode-mode coupling, it is however possible to probe higher frequency modes. Indeed, a continuous drive applied to the array at the frequencies of its higher modes leads to a shift of the fundamental mode frequency. This shift arises due to the cross-Kerr coupling represented by the interaction term $\sum_{k\neq l} -K_{kl} a_k\dag a_k a_l\dag a_l$ where $K_{kl}$ is the cross-Kerr coupling between array modes $k$ and $l$ while $a_k^{(\dagger)}$ destroys (creates) an excitation in mode $k$. This mode-mode interaction leads to a dispersive readout mechanism for the modes that are above the experimental frequency cutoff (see Appendix~\ref{app:array}). This technique was already used in Ref.~\cite{PhysRevLett.109.137002} to map the full dispersion relation of the distributed modes of the array. In this way, the first mode was identified to be at $\omega_1/2\pi = 11.9$~GHz.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{figure_experiment_data.pdf}
\caption{Frequency shift of the fundamental mode frequency as a function of pump drive detuning and for different pump drive powers. The dots represent the experimental data while the lines represent the numerical simulations. The legend indicates the power, $P$, of the drive applied both in the experiment and in the simulations, normalized to $0$ dB for the weakest drive display here. For the numerics we have used power proportional to the experimentally applied power, parametrized such that $\varepsilon_r \equiv \varepsilon_1/\varepsilon_{0dB}$ with $\varepsilon_{0dB}/(2\pi) = 1.83$~MHz. The solid lines are a full stochastic two-mode simulation, while the dashed lines are the frequency shifts calculated from an one-mode steady-state assumption using Eqs.~\eqref{eq:mastereq} and~\eqref{eq:Dss}. For the simulation we have used $K_1/(2\pi) = 5.7$~MHz, $\kappa_1/(2\pi) = 2.9$~MHz and $\kappa_0/(2\pi) = 1.0$~MHz. For both experiments and simulations, the frequency shift is obtained by sweeping the probe detuning $\Delta_0$.} \label{fig:data}
\end{figure}
To probe the photon number in this first mode as a function of the drive detuning, a continuous pump tone is applied close to $\omega_1$. Simultaneously, the transmission spectrum of the fundamental mode is measured by sweeping the frequency of a weak continuous probe tone. The frequency at which the probe is maximally reflected indicates the resonance of the fundamental mode. Figure~\ref{fig:data} shows the experimentally measured frequency shift (dots) obtained using this method for different pump powers and as a function of the detuning of the pump drive. The driven-dissipative Kerr resonator always shows a single resonance peak. Similarly, we note that the measured frequency shifts show abrupt changes between a low and a high value, with no values in between the two. The lack of intermediate values for the frequency shift indicates that no switching occurs between two bistable states during the time of the experiment set by the the measurement time-scale, $t_m = 1 \, \mu$s and averaged 1000 times. We present a detailed theoretical description of the experimentally observed resonance lineshapes in the next section.
\section{Theoretical model}
\label{sec:theory}
Since only the fundamental mode is directly measurable experimentally, the Kerr coefficients are inferred from a theoretical model of the distributed junction array. To this end, we write the Lagrangian for the linear part of the array in terms of the system capacitances and Josephson inductance ($L_{J}$), as shown in Fig.~\ref{fig:experimental_setup}(c)~\cite{devoret1995quantum}. Following a quantization of these modes, we perturbatively include the nonlinear contributions from the Josephson junctions (see Appendix~\ref{app:array} for details)~\cite{PhysRevA.86.013814, PhysRevB.92.104508}. From room-temperature resistance measurements, we infer an inductance per junction of $L_J = 1.9$~nH and an effective array plasma frequency (highest mode frequency) of 18.2~GHz. The effective array plasma frequency is not equal to the single junction plasma frequency $\omega_p = 1 / \sqrt{C_J L_J}$, as would be the case if the junctions were purely linear elements. We moreover infer that the system capacitances illustrated in to have values $C_0 = 0.066$ fF, $C_J = 26.54$ fF, $C_g = 10.4$ fF, $C_s = 3$ fF and $C_e = 10.84$ fF close to the design parameters found by simulations~\cite{PhysRevLett.109.137002}. With these parameters, we estimate the self-Kerr coefficient of the fundamental mode to be $K_0/2\pi = 0.5$ MHz, and that for the first mode to be $K_1/2\pi = 5.7$ MHz. The latter is almost a factor of two bigger than the corresponding linewidth $\kappa_1/2\pi = 2.9$ MHz, and hence lies in the intermediate (or mesoscopic) regime identified in Fig. \ref{fig:kerrregimes}.
Reintroducing the junction nonlinearity in the analysis, we find that the mode frequencies get shifted by self-Kerr and cross-Kerr couplings. The $k$th mode's frequency, $\omega_k$, is shifted such that the new effective frequency becomes
\begin{align}
\omega_k \rightarrow \omega_k - \sum_l K_{kl},
\end{align}
where $K_{kl}$ is the cross-Kerr contribution between modes $k$ and $l$ (see Appendix~\ref{app:array}). Although each $K_{kl}$ is a factor of $10^2$--$10^3$ smaller than $\omega_k$ (on the order of a few MHz), the cumulative effect of all the modes results in significant shifts of the eigenfrequencies of the array by up to a GHz compared to the bare mode frequencies. Moreover, as a result of the cross-Kerr couplings, when driving the Kerr-resonator, the average photon number, $\exv{a_1\dag a_1}$, increases leading to a measurable frequency shift of the fundamental mode.
As a first, simple, approach to reproducing the data of Fig.~\ref{fig:data}, the steady-state photon number, $\exv{a_1\dag a_1}_{s} = \text{Tr}(\rho_{s} a_1\dag a_1)$, is numerically computed using the single-mode master equation Eq.~\eqref{eq:mastereq}. From this, the frequency shift of the fundamental mode due to photon population in the first mode is then estimated to be
\begin{align}
\Delta_{0} = - K_{01} \exv{a_1\dag a_1}_{s}, \label{eq:Dss}
\end{align}
where $K_{01} = 4\sqrt{K_0 K_1}$ denotes the cross-Kerr coupling between the two modes~\cite{PhysRevA.86.013814}. The result of this calculation corresponds to the dashed lines in Fig.~\ref{fig:data}. Evidently, this steady-state result does not match the experimental data (dots). Indeed only some features of the experiment are captured by this single-mode steady-state calculation. For instance, the detunings are approximately correct. On the other hand, the shape of the frequency shift as a function of the detuning is very poorly reproduced. In particular, the shoulders to the left of the maxima of the shift are not observed experimentally.
In order to theoretically reproduce the data, we now turn to an approach that resembles the experimental situation more closely. First, rather than computing the steady-state photon population, we compute the full time-dependent response of the transmitted power along the transmission line coupled to the array. As in the experiment, this response is integrated over a finite measurement time $t_m$ and averaged over 1000 realizations. Any dynamics that occur on a time-scale much slower than $t_m$, such as slow switching between two bistable states, are therefore neither resolved in the measurement nor in the simulations. Second, while the cross-Kerr interaction was used in computing the expected frequency shift in Eq.~\eqref{eq:Dss}, we now add this interaction directly to the system Hamiltonian. This allows for capturing the measurement-backaction in the form of measurement-induced dephasing of the first mode by the probe tone. The addition of measurement-induced dephasing increases the effective linewidth of the nonlinear mode, thereby leading to a reduction in the number of photons in the mode and, consequently, to a smaller observed frequency shift.
In order to obtain a description as close to the actual experiment, we numerically solve the full two-mode Hamiltonian including both the mesoscopic Kerr mode ($a_{1}$) and probe mode ($a_{0}$) with the stochastic master equation,
\begin{equation}\label{eq:stochasticmaster}
\begin{split}
d{\rho_Z} &= -i dt\, [H_1 + H_0 + H_c,\rho_Z]\\
&\quad + dt\, \kappa_1 \big[a_1 \rho_Z a_1\dag - \frac{1}{2}(a_1\dag a \rho_Z + \rho_Z a_1\dag a_1) \big] \\
&\quad + dt\, \kappa_0 \big[a_0 \rho_Z a_0\dag - \frac{1}{2}(a_0\dag a_0 \rho_Z + \rho_Z a_0\dag a_0) \big] \\
&\quad + \sqrt{\kappa_0} \Big[ dZ a_0 \rho_Z + dZ^* \rho_Z a_0\dag\\
&\quad - \text{Tr}(dZ a_0 \rho_Z + dZ^* \rho_Z a_0\dag ) \, \rho_Z \Big]
\end{split}
\end{equation}
valid for heterodyne detection of the fundamental mode $a_0$~\cite{wiseman2009quantum}. In this expression, $H_k$ is Hamiltonian given by Eq.~\eqref{eq:hamiltonian}, while
\begin{equation}
H_c = -K_{01} \, a_1\dag a_1\, a_0\dag a_0^{\phantom{\dagger}} \label{eq:crosskerrH}
\end{equation}
is the relevant cross-Kerr interaction. In Eq.~\eqref{eq:stochasticmaster}, the density matrix $\rho_Z$ is conditioned on the result of a heterodyne measurement where $dZ = dW_a + idW_b$, with $dW_i$ a stochastic Wiener processes with $\exv{dW_i} = 0$ and $\exv{dW_i^2} = dt$\footnote{The subscripts of $dW_i$ indicates solely that these are different independent Wiener processes.}.
The last term of Eq.~\eqref{eq:stochasticmaster} is associated to the homodyne measurement record which is used to update our description of the state of the system $\rho_Z$.
From the numerical integration of Eq.~\eqref{eq:stochasticmaster}, the transmitted power can then be calculated as
\begin{align}
J_T(t) =&\, \frac{1}{2}\Big\{\sqrt{\kappa_{0}}\,\text{Tr} [\rho_Z (a_0 + a_0\dag)] + dW_a/dt \Big\}^2 \nonumber \\&+ \frac{1}{2}\Big\{i\sqrt{\kappa_{0}}\,\text{Tr} [\rho_Z (a_0 - a_0\dag)] + dW_b/dt \Big\}^2, \label{eq:signalT}
\end{align}
which, averaged over time $t_m$, yields the measured signal. The value $\Delta_0$ corresponding to the maximum transmission signal, $J_T$, gives the frequency shift plotted in Fig.~\ref{fig:data}. The simulation were performed using standard numerical integration for the deterministic part of the equation and after each integration step, $\delta t$, a random number is drawn from a Gaussian distribution with variance $\delta t$ for each stochastic term $dW_i$ such that the stochastic part of $d\rho_Z$ can be readily calculated.
In Fig.~\ref{fig:data}, the numerical simulations using the parameters of the experiment are presented as solid lines, which are in excellent agreement with the experimental data. The difference between the steady-state values used in Eq.~\eqref{eq:Dss} and the stochastic master equation simulations is mainly due to the facts that the latter (i) includes the full two-mode interaction and, as a result, the measurement backaction, and (ii) explicitly takes into account the finite measurement time $t_{m}$. Therefore, dynamics on time-scales much longer than the characteristic measurement time, which are not probed by the experiment, are also not observed in the stochastic master equation simulation. The most predominant feature observed from the stochastic simulations, and not the simpler single-mode steady-state calculations, is the lack of intermediate values in the abrupt transition from low photon number to high photon number state when changing the drive detuning, i.e., the ``shoulders'' in the steady-state photon number. For strong Kerr nonlinearities the energy levels of a Kerr oscillator are {$\omega_k - K_k n$} and can be well-resolved. The ``shoulders'' at large detunings are, thus, signatures of switching events at corresponding photon-number-selective transitions. These intermediate values are not resolved in the experiment due to the finite measurement time and this fact indicates a time-scale of potential switching dynamics much longer than the measurement time $t_m$ and comparably or even longer than the total experimental time of $t_m$ times the number of averages. To fully capture the dynamics of the experiment, a two-mode model is needed, however, the contributions from the two-mode interaction do not significantly contribute to the switching time-scale. Rather, the effects of the cross-Kerr interaction is to account for the measurement-induced backaction. This backaction affects the detuning and drive strength for which the increased time-scales appear. Thus, the physics of the long time-scales can be understood from only a single mode model. In the next two sections, we confirm this by performing a detailed investigation of the switching rates for a single mode Kerr resonator with parameters in the mesoscopic regime, and explicitly comparing the predictions from semi-classical and quantum treatments.
\section{Numerical treatment of the switching rates}
\label{sec:numr}
Since both the experiment and stochastic master equation simulations show a qualitatively different behavior than the single-mode steady-state master equation, a more complete quantum description is necessary to quantitatively understand the switching dynamics of Kerr resonator in large-nonlinearity regimes. In particular, as evident from Fig.~\ref{fig:data}, the lack of intermediate photon number values in the bistable regime indicates that there is no dynamical switching between the two bistable states within the time-scale of the experiment. In this section, we perform numerical simulations of the steady-state photon number and switching rates of the Kerr oscillator introduced in the previous section. Specifically, in order to identify the appropriate description for the nonlinear oscillator in mesoscopic regime, we compare the results obtained from semi-classical simulations with that from quantum master equation simulations.
\begin{figure}[t]
\includegraphics[width=\linewidth]{timescale_regions.pdf}
\caption{Representative curves showing numerical calculations of steady-state photon number as a function of the drive detuning for $K=\kappa$ and $\varepsilon=6\kappa$ (as the red curves in Fig.~\ref{fig:semi_master} but with smaller nonlinearity, i.e., closer to the semi-classical regime). The solid black line shows the results calculated using the master equation, while the dotted lines show the results ontained using a semi-classical equation of motion (see Eq.~\eqref{eq:semi} and Appendix~\ref{app:classical}). In regions I and III, the resonator relaxes to its steady state on an average time scale $T_\kappa = \kappa^{-1}$. Region II corresponds to the switching region where the oscillator dynamics slow down considerably.} \label{fig:regions}
\end{figure}
We begin with a generic description of the steady-state response of a nonlinear oscillator. It is convenient to delineate the time-averaged response into three regions, as sketched in Fig.~\ref{fig:regions} for an oscillator with $K = \kappa$ (this corresponds to the mesoscopic regime depicted as the blue region in Fig.~\ref{fig:kerrregimes}). The thick full line is a representative curve obtained using a master equation simulation, while the narrow dashed line is obtained from a semi-classical analysis with the same parameters. In region I, the resonator relaxes to a state with a low photon number on a time scale of $T_\kappa = 1/\kappa$. There may be two stable classical solutions in this region, as indicated by the dashed line in Fig.~\ref{fig:regions}, but the fluctuations associated with the low photon-number state are not sufficient to bring the system to the high photon number state. For the lower branch solution of this region, the semi-classical and the full quantum treatment give identical results for both the relaxation time into the steady-state and for the steady-state photon number. Indeed, in this regime, only few photons are present and, as a consequence, the nonlinear effects play only a minor role. Region III is conceptually comparable to region I, except that the system relaxes into a high photon number state.
In region II, the dynamics are such that the photon number initially latches to a low photon-number state, but after some time it jumps to, and continues to fluctuate around, a high photon-number state \cite{Siddiqi2005, PhysRevApplied.1.054005}. This switching between the high and low photon number states continues and we refer to the time-scale of this dynamics as the switching time, $\tau$. We henceforth focus on region II, and study switching times as a function of the ratio $K/\kappa_{1}$. In particular we are interested in quantifying the difference between a weakly nonlinearly Kerr resonator, which we expect to behave classically, and a highly nonlinear Kerr resonator where quantum fluctuations are expected to play a larger role.
We first consider a semi-classical description which we obtain from the Heisenberg-Langevin equation satisfied by the first mode,
\begin{align}\label{eq:hleq}
\dot{a}_{1} = - i \big[a_{1}\, , \, H_{1} \big] - \frac{\kappa_{1}}{2} a_{1} + \sqrt{\kappa_{1}} a_{1}^{\rm in}(t),
\end{align}
where $H_1$ is given by Eq.~\eqref{eq:hamiltonian}. Here the input field, $a_{1}^{\rm in}(t)$, satisfying the commutation relation $[a_{k}^{\rm in}(t), a_{k'}^{\rm in \dagger}(t') ] = \delta_{kk'}\delta(t - t')$, accounts only for the quantum fluctuations induced by the environment. The driven-dissipative Kerr resonator is most conveniently treated in a frame rotating at the drive frequency, $\omega_d$, such that we make the replacements $\varepsilon_{1}(t) \rightarrow \varepsilon_{1}$ and $\omega_1 \rightarrow \Delta_1 = \omega_1 - \omega_d$ in the Hamiltonian.
We first attempt to solve the time-evolution using a semi-classical trajectory approach \cite{PhysRevA.73.063801},
which replaces the system field operators with complex numbers, $\langle a \rangle \rightarrow \alpha$. This replacements reduces Eq.~\eqref{eq:hleq} to an equation of motion for the phase-space variable $\alpha$ (see also Appendix~\ref{app:classical} for a deterministic classical treatment). However, when replacing the input field with a stochastic variable to account for the quantum noise associated with $a_{1}^{\rm in}(t)$, this approach becomes only approximate in the presence of nonlinear mixing terms in the Hamiltonian, since it does not account for up/down-converted quantum noise due to mode mixing. The resulting semi-classical stochastic equation of motion can then be written as,
\begin{align}
\dot{\alpha} = -i \Delta \alpha +2i\, K |\alpha|^2 \alpha - i \varepsilon - \frac{\kappa}{2} \alpha + \sqrt{\kappa} \, \zeta(t), \label{eq:semi}
\end{align}
corresponding to an equation for the coherent state amplitude of the system, and where $\zeta(t)$ is stochastic Wiener process that models the quantum vacuum noise \footnote{Here, $\zeta(t) = (dW_a(t)/dt + i\, dW_b(t)/dt)/\sqrt{2}$ represents a stochastic Wiener process with $\langle dW_i \rangle = 0$ and $\langle dW_i^2 \rangle = dt$ that, on average, corresponds to the input field being the quantum vacuum state.}. Here, we have suppressed the subscript 1 both for brevity and to underscore the generality of this treatment.
A particular realisation of $\zeta(t)$ is referred to as a semi-classical trajectory. In the bistable regime of the oscillator, a trajectory initially latches to a low photon-number state, and after some time jumps to a high photon-number state \cite{Siddiqi2005, PhysRevApplied.1.054005}. As the switching continues with a waiting time between the successive switches approximately Poisson distributed, we use an exponential function to fit the average over many trajectories to extract the time scale to reach steady-state.
To understand the quantum effects that arise in the switching dynamics of the Kerr resonator, we next perform numerical master simulations, Eq.~\eqref{eq:mastereq}, of the driven system with the Hamiltonian, Eq.~\eqref{eq:hamiltonian}, and analyze the relaxation towards the steady-state. From these simulations, we observe an oscillatory behavior of the photon number that relaxes on a time scale $T_\kappa$ followed by an exponential relaxation towards the steady state. The relaxation time scale, $\tau$, can therefore be readily extracted from the exponent of exponential decay obtained using master equation simulations.
\begin{figure}[t]
\includegraphics[width=\linewidth]{figure_kerr_semi_master_v2.pdf}
\caption{Comparison of steady-state photon number and the relaxation time scale $\tau$ (in units of $1/\kappa$) as a function of the drive detuning $\Delta/\kappa$, obtained from quantum master equation simulations (a) and the semi-classical simulations averaged over 200 semi-classical trajectories (b). The red lines represent steady-state photon numbers (right axis), while the blue lines are the time scales obtained from an exponential fit to the time evolution of the photon number (left axis). The solid lines are simulation results for $K=2\kappa$, while the dashed lines show the results for $K=0.2\kappa$. The blue lines indicating the relaxation times $\tau$ associated with the switching dynamics are only plotted in the bistable regime. In all cases, the resonator is initialized in the vacuum state and a drive of amplitude $\varepsilon = 6\kappa$ is used.}
\label{fig:semi_master}
\end{figure}
In Fig.~\ref{fig:semi_master}, we compare the results obtained from both simulations described above, for a strongly nonlinear Kerr oscillator with $K=2\kappa$ (solid lines) and for a weakly nonlinear Kerr oscillator with $K=0.2\kappa$ (dashed lines). The results of the master equation simulation are shown in the main panel, while those of the semi-classical trajectories are shown in the inset. In both cases, the relaxation time is shown in blue (left axis) and the average photon number in red (right axis). First, it is clear from both the classical and quantum numerical treatments that the switching time (blue lines) can significantly exceed the linear decay time $T_{\kappa}=1/\kappa$ near the bistability. Further, for weak nonlinearity $K=0.2\kappa$, both the quantum and the semi-classical approaches show a sharp rise in $\tau$ close to the detuning where the steady state maximum photon number reaches its maximum (compare dashed lines between the main figure and the inset). This behavior is a generic feature of weakly nonlinear systems. Semi-classically, it can be understood as being caused by the switching rates between two stable states becoming equal.
For larger nonlinearity, $K=2\kappa$, the discrepancy between the full quantum simulations and semi-classical numerics is much more pronounced (compare solid lines between the main figure and the inset). Specifically, a much wider distribution of time scales is predicted by the quantum treatment in the strong nonlinearity regime, which the semi-classical approach completely fails to capture. This observation is not surprising and is reinforced by the fact that the steady-state in this regime shows a negative $Q$-parameter, \mbox{$Q = (\exv{(\Delta a_1\dag a_1)^2} - \exv{a_1\dag a_1})/\exv{a_1\dag a_1} \approx {-}0.4$}, which indicates sub-Poissonian statistics. Similarly, the steady-state Wigner function displays negative values (not shown), a clear sign of the non-classical nature of the state of the system.
Numerical simulations with varying $K/\kappa$ further confirm that the average behavior observed in Fig.~\ref{fig:semi_master} gradually transitions from a situation where the semi-classical treatment matches well with full quantum simulation for small $K$, towards a larger discrepancy when $K$ is increased (not shown).
The numerical results presented here can now be compared to the results of Fig.~\ref{fig:data}. There, the steady-state behavior showed intermediate values in the bistable regime, while the experimental data was only correctly reproduced using a two-mode stochastic quantum analysis that take into account the finite measurement time. This is consistent with the emergence of a very slow time-scale, which can be significantly longer than the linear decay rate $\kappa$ of the oscillator; since dynamics on very long time-scales were not probed by the experiment, it invalidates a steady-state analysis of the data. Moreover, for sufficiently large nonlinearities ($K > \kappa$) as relevant to the experiment with JJ arrays, this slow time-scale cannot be attributed only to semi-classical switching dynamics.
In the next section, we present analytical studies highlighting the differences between the semi-classical picture and the full quantum calculations of the switching rate in the large nonlinearity regime ($K/\kappa \geq 1$), while comparing them against the numerical results obtained in this section.
\section{Analytical treatment of the switching rates}
\label{sec:theo}
\subsection{Quantum calculation}
\label{sec:calcquant}
In the last section, we saw that the increase in the switching time in this regime is not captured by a semi-classical treatment. Therefore, we now consider an analytical description that takes into account the full nonlinear quantum dynamics.
To this end, we consider the Liouvillian describing the time evolution of the driven-dissipative single-mode Kerr-oscillator in the mesoscopic regime [Eqs.~\eqref{eq:hamiltonian} and~\eqref{eq:mastereq}]. In a Hilbert-space of dimension $N$ ($N$ being the number of Fock states included in the calculation), the Liouvillian has $N^2$ eigenvalues. The time-evolved density matrix can be expressed in terms of these eigenvalues as
\begin{equation}
\rho(t) = \sum_\lambda c_\lambda e^{\lambda t} \rho_\lambda
\end{equation}
where $\rho_\lambda$ are the eigenstates of the Liouvillian, $\mathcal{L}[\rho_\lambda] = \lambda \, \rho_\lambda$. In the presence of damping, all the eigenvalues have negative real parts, with the exception of the steady state $\rho_{s}$ for which $\lambda =0$. This also represents the only physical state of the system that survives at long times. To preserve the trace at all times, including in steady state, $\rho_{s}$ always appears in this decomposition with a constant prefactor of $c_{s} = 1$. Since the time evolution is trace-preserving this implies that $\text{Tr}(\rho_{\lambda\neq0}) = 0$ and that $\rho_{\lambda\neq0}$ are not valid density matrices with direct physical significance \cite{peres1995quantum}.
Here, we are interested in the time-dependent fraction of population in the steady state and therefore we re-express the expansion as a linear combination of orthogonal unit-trace density matrices. To accomplish this, we consider the eigenvalue, $\chi \neq 0$ of the Liouvillian, $\mathcal{L}[\cdot]$, defined as the eigenvalue with the smallest real part~\cite{PhysRevA.35.1729}. We observe that $\langle \rho_{s}, \rho_\chi \rangle \neq 0$, with \mbox{$\langle A, B \rangle = \text{Tr}(A\dag B)$} respresenting the matrix inner product. To ensure orthogonality, we use a Gram-Schmidt construction to define
\begin{align}
\tilde{\rho}_\chi = \rho_\chi{-}\frac{\langle \rho_{s}, \rho_\chi\rangle }{\langle \rho_{s}, \rho_{s}\rangle} \rho_{s},
\end{align}
with $\langle \tilde{\rho}_\chi, \rho_{s} \rangle = 0$. The normalized density matrix $\tilde{\rho} = \tilde{\rho}_\chi / \text{Tr}(\tilde{\rho}_\chi)$ has unit trace, but it is not ensured to be positive definite. Nevertheless, we can express the density matrix $\rho(t)$ as
\begin{align}
\rho(t) = \beta_0(t) \rho_{s} + \beta_1(t) \tilde{\rho} + \sigma(t), \label{eq:rhosschi}
\end{align}
where $\sigma (t)$ only ensures the positive definiteness of the full density matrix, while the factors $\beta_{i}$ denote the decay rates.
In order to mimic the exponential relaxation observed in the numerical analysis of Sec. \ref{sec:numr}, we surmise that $\beta_0(t) = (1 - \tilde{\beta}_0 e^{-\lambda_e t})$ with $\lambda_e$ represents the ``escape rate", thereby explicitly assuming Markovian switching dynamics. To estimate the rate at which the density matrix approaches the steady state in the long-time limit, we can calculate the time scale $\lambda_{e}$ by inserting Eq.~\eqref{eq:rhosschi} into the master equation, Eq.~\eqref{eq:mastereq}, and taking the inner product with $\rho_{s}$ to obtain,
\begin{align}
\dot{\beta}_0(t) = \langle \,\rho_{s}, \mathcal{L}[\rho(t)]\, \rangle.
\end{align}
The quantum-induced switching rate can now be expressed as $\lambda_e = \dot{\beta}_0(t)/(1-\beta_0(t))$. Since $\sigma(t)$ only plays a role in preserving positive definiteness, the dominant time scale is not affected by it and we can omit $\sigma(t)$ in the analysis of the time scale without introducing unphysical behavior. Without $\sigma(t)$, the expression for $\lambda_{e}$ only accounts for the dynamics induced by the steady state and the smallest eigenvalue of $\mathcal{L}[\rho(t)]$, the two dominant contributions in the long time limit. While $\lambda_e$ is explicitly time-dependent, in the regime of interest it is approximately constant and we can therefore neglect the time dependence and simply evaluate for $t=0$ where $\rho(t=0)$ is the initial vacuum state. This leads to the following expression for $\lambda_e$:
\begin{align}
\lambda_e = {\langle \,\rho_{s}, \mathcal{L}[\rho(t=0)]\, \rangle} \Bigg({1 - \frac{\langle \rho_{s}, \rho(t=0) \rangle}{\langle \rho_{s}, \rho_s \rangle}}\Bigg)^{-1}. \label{eq:lambda}
\end{align}
The solid lines in Fig.~\ref{fig:lambda} (a) and (b) show the results of the analytical calculation of the switching times $\lambda_e^{-1}$ for $K/\kappa = 2$ and $K/\kappa = 0.2$, respectively. The dashed lines represent the time-scales extracted from the numerical master equation simulation also shown in Fig.~\ref{fig:semi_master}. Though Eq.~\eqref{eq:lambda} predicts a larger value of $\tau = \lambda_e^{-1}$ than that observed numerically, the analytical calculations qualitatively match the numerical results. We can understand the larger value for $\tau$ obtained analytically from the fact that we neglected $\sigma(t)$. Indeed, in general $\sigma(t)$ has contributions from all eigenvalues of the Liouvillian which includes contributions with a larger negative real part than $\chi$. Therefore, the dynamics associated with $\sigma(t)$ must be strictly faster than $\tau$. In particular, the Liouvillian spectrum may show a two-fold degeneracy in the real part of the eigenvalues, which may speed up the dynamics by up to a factor of two. The prolonged time-scales observed in both experiment and in numerical simulation can, therefore, directly be qualified using a simple analytical quantum calculation.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{figure_lambda_v2.pdf}
\caption{Comparison for switching times between oscillators with strong Kerr nonlinearity, $K=2\kappa$, (a) and (c), and weak Kerr nonlinearity $K = 0.2\kappa$, (b) and (d). The panels (a) and (b) present the analytical calculation of the relaxation time scale showed in solid lines, while the dashed lines are the time scales extracted from numerical simulations (also shown in Fig.~\ref{fig:semi_master}). In (c) and (d) the switching time-scales obtained from escape rates in a metapotential are displayed (Eq.~\eqref{eq:lambda_e}]). The dashed lines are the time scales extracted from the semi-classical numerical simulations. Inset: The inset in the lower-left panel displays an example of the metapotential $U(x)$ for the parameters marked by the blue dot.
The ticks on the x-axis are at 0 and $x_0$ with ticks on the y-axis showing the corresponding values of the energy in the metapotential.
For all simulations and calculations, driver power $\varepsilon = 6\kappa$ was used.} \label{fig:lambda}
\end{figure}
\subsection{Semi-classical calculation: Quantum activation}
\label{sec:calcclassical}
The quantum calculation based on the Liouvillian qualitatively reproduces the dynamics of the full master equation. However, for weak nonlinearity, we expect a semi-classical treatment to be sufficient. Here, we compare the switching rates predicted by the numerical simulations with analytical results for a fluctuation-induced escape from a metapotential \cite{dykman2012fluctuating}. A metapotential is an effective potential that corresponds to the same equation of motion as the semi-classical model.
In this treatment, the fluctuations associated with the noise $\zeta(t)$ in Eq.~\eqref{eq:semi} are transformed into an effective temperature and the switching rate $\tau$ can then be evaluated as a thermal escape rate from a local minimum of the metapotential \cite{dykman1980fluctuations, vijay2008josephson}. The increased time-scale in region II in this picture corresponds to a higher effective barrier in the metapotential.
To estimate the semi-classical escape rate, we begin by first considering the semi-classical equation, Eq.~\eqref{eq:semi}, with $\zeta(t) = 0$ (See Appendix~\ref{app:classical}), and later reinstate the effect of the noise. The general approach is to rephrase the complex equation into a real equation for a generalized position variable, $x$, that changes slowly in time. To this end, we solve Eq.~\eqref{eq:semi} with $\zeta(t) = 0$ and denote the low-amplitude solution as $\alpha_0$ and the unstable solution as $\alpha_u$. Next, we define a rotation angle \mbox{$\varphi = \text{tan}^{-1}[ -\text{Im} (\alpha_u - \alpha_0) / \text{Re} (\alpha_u - \alpha_0)]$} such that the quantity $x_0 = e^{i\varphi} (\alpha_u - \alpha_0)$ is a real number and the axis on the line from 0 to $x_0$ constitutes our position variable $x$.
In order to determine the metapotential $U(x)$, we make a substitution $\alpha(t)=\alpha_0+e^{-i\varphi}z(t)$ and rewrite the equation of motion, Eq.~\eqref{eq:semi}, in the form $\dot{z}=F(z)$, where
\begin{align}
F(z) =& (-i\Delta - \kappa/2)(z + e^{i\varphi}\alpha_0) - ie^{i\varphi} \varepsilon \nonumber\\&- 2iK (z + e^{-i\varphi}\alpha_0^*)(z + e^{i\varphi}\alpha_0)^2. \label{eq:dUx}
\end{align}
represents the effective force on the particle. The complex variable $z$ can be represented in terms of two real variables, playing the role of the coordinate and momentum, $z=x+ip$.
Since the two states $\alpha_0$ and $\alpha_u$ are steady states, it follows that $F(0) = F(x_0) = 0$.
Further, since the imaginary part accounts for the dynamics of the momentum $p$, we can write the derivative of the metapotential $U(x)$ as
\begin{align}
\text{Im}[\,F(x)\,] = \dot{p} = - \frac{d U(x)}{d x}.
\end{align}
Integrating this equation results in a one-dimensional metapotential,
\begin{align}
U(x) = - \int \text{Im}[\,F(x)\,] \, dx,
\end{align}
with a minimum at 0 and maximum at $x_0$ as illustrated in the inset of Fig.~\ref{fig:lambda}~(c). Following Kramer's escape law~\cite{kramers1940brownian, RevModPhys.62.251}, the escape rate over the barrier can be written as~\cite{dykman1980fluctuations, vijay2008josephson}
\begin{align}
\lambda_e = \gamma_0 \exp \Big( - \frac{\Delta U}{\kappa} \,\Big), \label{eq:lambda_e}
\end{align}
with \mbox{$\Delta U = U(x_0) - U(0)$} denoting the activation energy. Note that effective temperature is set by the mode linewidth, $\kappa$, since it enters as the prefactor for $\zeta(t)$ in Eq.~\eqref{eq:semi}. The rate $\gamma_0$ is the attempt frequency and it is extracted from the quadratic term of the potential, which can be expressed as $\frac{\gamma_0^2}{2\kappa}x^2$. When increasing the drive amplitude, $\epsilon$, the barrier height is decreased leading to a faster escape rate~\cite{Dykman2007}.
In Figs.~\ref{fig:lambda} (c) and (d), we show $\lambda_e$ obtained using this quantum-activation approach (solid lines) and compare it with the switching rates extracted from the semi-classical simulations (dashed lines, also shown in the inset of Fig.~\ref{fig:semi_master}). We observe that, for small detunings, the simulations match the escape rate calculation of Eq.~\eqref{eq:lambda_e} quite well. However, a large discrepancy is observed for larger detunings when the switching time dramatically increases near the bifurcation point, where the rates between the two stable states equilibrates as already explained in the context of the numerical simulations. In the escape rate calculation, we are however only calculating a one-way rate and, thus, we do not capture the same increase in switching time. More interestingly, and as should be expected, we observe that while the escape rate calculation captures the behavior of the semi-classical simulations well, it does not capture the quantum corrections relevant in the strongly nonlinear regime ($K/\kappa = 2$), as observed from its deviations from predictions of quantum calculation and master equation simulations [c.f. Figs \ref{fig:lambda} (a) and (c)].
To summarize this section, we have considered the time scales relevant for the relaxation of a Kerr-resonator towards the steady state. Going back to the regions introduced in Fig.~\ref{fig:regions}, we noticed that, most significantly, the time scale for relaxation towards the steady state becomes very large in region II. This is consistent with experimental data presented in Sec.~\ref{sec:exp}, where we found a steady-state description to be inadequate due to the finite time-scale set by the measurement rate (see also the experiments presented in \cite{muppalla2017bi}). We also found that while a semi-classical calculation adequately predicts the dynamics of a weakly nonlinear system, it fails to capture the structure of the time scales for relaxation, $\tau$, seen for large nonlinearities. The time scale for relaxation can be accurately predicted by a simplified but fully quantum model. On the other hand, an escape rate calculation matches the semi-classical trajectory simulations only partly, and completely fails to describe the dynamics for large nonlinearities.
\subsection{Crossover parameter}
\label{sec:crossovertemp}
The observed breakdown of the semi-classical theory can be understood in a phenomenological manner by introducing a damping-dependent crossover temperature $T_{\gamma}$~\cite{ingold1995},
\begin{eqnarray}
T_{\gamma} = \frac{\hbar \gamma_{0}}{2\pi k_{B}} \left(\sqrt{\frac{\kappa^{2}}{4 \gamma_{0}^2}+ 1}-\frac{\kappa}{2 \gamma_{0}}\right),
\end{eqnarray}
where $\gamma_{0}$ is the attempt frequency in the metapotential introduced in the semi-classical calculation. Using this definition, we can define a crossover parameter $\xi$ as
\begin{eqnarray}
\xi \equiv \frac{T_{\gamma}}{T_{\kappa}},
\end{eqnarray}
where ${T_{\kappa} = \hbar \kappa/k_{B}}$ denotes the effective temperature of the quantum fluctuations coupled to the oscillator. The semi-classical to quantum crossover boundary is set by $\xi =1$~\cite{ingold1995}.
\begin{figure}[t!]
\includegraphics[width=0.98\columnwidth]{figure_CrossoverTempv4.pdf}
\caption{Crossover parameter $\xi$ as a function of reduced detuning $\Delta/\kappa$ for a nonlinear oscillator. For $\Delta/\kappa < \sqrt{3}/2$, $\gamma_{0}$ depends entirely on the detuning of the resonator (dashed curve), while for $\Delta/\kappa > \sqrt{3}/2$, $\gamma_{0}$ is largely dominated by the photon amplitude in the resonator (solid curve). This change is captured by the rotation of the real axis of the one-dimensional metapotential denoted by angle $\varphi$. The crossover point is consistent with the region where breakdown of the semi-classical theory was observed previously.} \label{fig:crosstemp}
\end{figure}
Figure \ref{fig:crosstemp} shows $\xi$ as a function of detuning of the oscillator. For $\xi < 1$ or $T_{\kappa} > T_{\gamma}$, the decay dynamics can be described largely using the semi-classical activation treatment. However, as the detuning of the oscillator increases, and ${T_{\gamma} > T_{\kappa}}$, quantum tunneling effects can become essential to describe the nonlinear decay dynamics. We find that the attempt frequency $\gamma_{0}$, and consequently $\xi$, does not depend on the strength of the nonlinearity explicitly. This makes $\xi$ a universal quantity for a dissipative decay of a metastable state; nonetheless, strong nonlinearity $K/\kappa > 1$ is essential for the oscillator to bifurcate at large enough detunings and enter the quantum regime as shown in Fig.~\ref{fig:lambda}. This intuition is consistent with significant deviations of both numerical simulations and Liouvillian-based analytical estimates from quantum activation results at large nonlinearity. Nonlinear oscillators with large $K/\kappa$ thus access a qualitatively different regime from that described by (thermal or quantum) activation-based models, in which rate of escape via tunneling is exponentially small \cite{Dykman2007}.
\section{Discussion and Outlook}
\label{sec:concl}
In conclusion we have investigated the switching dynamics of a Kerr resonator in a combined experimental, numerical and analytical framework. We used experimental microwave spectroscopy data of a distributed mode of a Josephson junction array with $K/\kappa \approx 2$ and observed that the relaxation into the quantum steady state is not resolved in the experiment due to switching-times much longer than the probe time.
We also simulate the experiment using a two-mode stochastic master equation and find that the numerical simulations match very well with the experimental data, thus confirming our interpretation. A recent experiment \cite{muppalla2017bi} has directly measured the very slow switching dynamics in a weakly nonlinear Kerr resonator.
To further analyze the slow switching rates observed in the experiment we performed both semi-classical trajectory simulations and quantum master-equation simulations. We find that, for a range of parameters, the time scale to relax into the steady state is increased significantly beyond the natural decay time of the resonator, especially for strong nonlinearity ($K/\kappa \geqslant 1$) as compared to that for weak nonlinearity ($K/\kappa \ll 1$). We find that a semi-classical trajectory method is able to describe this slowdown only in the weak nonlinearity regime, while a full quantum master equation treatment is essential to calculate the switching rates for strongly nonlinear oscillators. To analytically estimate the time scales to reach the steady state, we compared a simplified quantum model and a semi-classical metapotential model. In the semi-classical metapotential treatment, the fluctuation-induced switching between bistable states is modelled as an activation over barrier in a metapotential, with the fluctuation-intensity determined by thermal or quantum noise. We find significant deviations from this model, especially when switching rates are small; moreover, these deviations are especially pronounced for large $K /\kappa$ and persist even far from bifurcation. In contrast, we obtain good qualitative agreement between quantum calculations and full master equation simulation in this regime. This is not entirely surprising since thermal or quantum activation models necessarily assume weak nonlinearity or weak driving conditions for the oscillator \cite{Dykman2007}. Our results indicate that switching dynamics in mesoscopic oscillators i.e. when $K/\kappa \gtrsim 1$ may be dominated by some other mechanism, such as dynamical quantum tunneling.
The conclusions drawn in this work regarding the time scales associated with switching dynamics in nonlinear Kerr-resonators are highly relevant for the characterization of state-of-the-art applications of Josephson devices especially in the strong nonlinearity regimes where quantum effects are more pronounced. In particular, we describe how the interplay of switching rates and the repetition rate of the experiment is essential to explain the experimentally measured nonlinear response of the superinductance presented in Fig. 3. We expect our results to guide the design of applications that similarly aim to observe non-linear spectroscopic signatures in Kerr resonators. In addition, our results can directly be used for assessing the performance of a bifurcation readout scheme for superconducting qubits~\cite{mallet2009single, vijay2009invited}. In such a scheme, the required measurement time has to be a few multiples of the characteristic switching time. Thus, Eqs.~\eqref{eq:lambda} and \eqref{eq:lambda_e} can be directly used for finding the required measurement time and, consequently, the expected qubit readout fidelity. These concerns become increasingly important as devices based on arrays of Josephson junctions, similar to the application in this work, continue to play an important role in experiments and theory proposals including various readout schemes~\cite{castellanos2008amplification, PhysRevA.90.062333, PhysRevLett.113.110502}, quantum controllers \cite{PhysRevA.93.012346, PhysRevX.5.041020, PhysRevApplied.4.034002} and even qubit architectures \cite{Bell2012,PhysRevLett.113.247001,Dempster2014}.
Furthermore, our calculations for the Kerr coefficients for the higher distributed modes of the Josephson array presented here (see Appendix~\ref{app:array}) indicate that these modes should be in the `mesoscopic' regime (defined as $K \gtrsim \kappa$) investigated here. This regime is also optimal for direct observation of dissipative quantum tunneling~\cite{Serban2007}, which is usually obscured by the activation-dominated switching observed in the JBA regime~\cite{Siddiqi2005}. Moreover, since the switching rate does not follow activation dependence on noise intensity, this suggests that mesoscopic Kerr oscillators can be useful platforms to test dynamics resulting from non-Gaussian noise. More generally, our study provides a framework to explore multi-photon quantum effects \cite{Kryuchkyan2012}, quantum noise properties, and quantum-to-classical transitions in strongly nonlinear systems.
\section*{Acknowledgements}
The authors acknowledge useful and insightful discussions with E.~Doucet, S.~Boutin and G.~Kirchmair. CKA acknowledges fruitful discussions with K. Mølmer, the hospitality of Universit\'{e} de Sherbrooke and financial support from the Villum Foundation and the Danish Ministry of Higher Education and Science. The authors acknowledge support by the U.S. Army Research Office under the grant number W911NF-18-1-0212, by the U.S. Department of Energy under grant number DE-SC0019515, by NSERC, and in part by funding from the Canada First Research Excellence Fund.
\begin{appendix}
\section{Kerr-resonators with arrays of Josephson junctions}
\label{app:array}
We consider a Kerr resonator, consisting of a linear 1D array of $N$ Josephson junctions (JJ) forming a nonlinear inductance, with first and last junctions capacitively shunted to ground. We will, at first, assume that each of these junctions are sufficiently linear such that we can neglect the nonlinearity of each junctions. Each junction is described by its effective inductance $L_J$ and capacitance $C_J$. Furthermore, we include a parasitic capacitance to ground for each junction, $C_0$. This gives us the (linearized) Lagrangian for the array:
\begin{align}
\mathcal{L}_{array} &= \sum_{n=1}^{N} \frac{C_J}{2} (\dot{\phi}_n - \dot{\phi}_{n+1})^2 + \frac{C_0}{2} \dot{\phi}_n^2 \nonumber\\&\qquad\quad- \frac{1}{2L_{J}} (\phi_n - \phi_{n+1})^2.
\label{eq:Larray}
\end{align}
The terms in Lagrangian corresponding to the end capacitances can be written as,
\begin{align}
\mathcal{L}_{end} = \frac{C_s}{2} \dot{\phi}_1^2 + \frac{C_g}{2} \dot{\phi}_1^2 + \frac{C_e}{2} \dot{\phi}_{N+1}^2,
\end{align}
where $C_s$ is the capacitance to the transmission line which controls the external quality factor of the array resonances, $C_g$ is the shunt capacitance on the first junction, and $C_e$ is the shunt capacitance on the last junction. Including the terms due to shunt capacitances, the full Lagrangian can then be witten as
\begin{align}
\mathcal{L} &= \mathcal{L}_{array} + \mathcal{L}_{end}\nonumber\\
&= \dot{\vec{\phi}}^{\,T} \frac{\mathbb{C}}{2} \dot{\vec{\phi}} - \vec{\phi}^{\,T} \frac{\mathbb{L}}{2} \vec{\phi}, \label{eq:CLmat}
\end{align}
where we have introduced the symmetric matrices $\mathbb{C}$ and $\mathbb{L}$ along with the vector $\phi = \big\{ \phi_1, \phi_2, \ldots, \phi_{N+1} \big\}^T$
Standing waves across the array now constitute a set of normal modes for the array. Formally, we find these standing mode by diagonalizing the matrix ${\Omega^2 = \mathbb{C}^{-1}\mathbb{L}^{-1}}$, such that the Euler-Lagrange equations for these modes decouple. The eigenvalues of the matrix, $\Omega^2$ are the squares of the normal-mode frequencies of the Kerr resonator. We can similarly define the effective capacitances and inductances for these modes as
\begin{align}
C_k = \vec{v}_k^{\,T} \mathbb{C} \vec{v}_k, && L_k^{-1} = \vec{v}_k^{\,T} \mathbb{L}\vec{v}_k,
\end{align}
where $\vec{v}_k$ are the corresponding eigenvectors of $\Omega^2$.
The eigenfrequencies are now, by construction, given by $\omega_k = 1/\sqrt{L_k C_k}$. Keep in mind that the actual modes are described by the physical phase variable $\phi(t) = \sum_k \phi_k(t) \vec{v}_k$, where $\phi_k(t)$ is a function oscillating with $\omega_k$; thus, the physical amplitude of the phase is in the $\phi_k(t)$ variables and no physical quantify depends on the normalization of $\vec{v}$.
\subsection{Quantization of the modes}
Having found the normal modes we can apply a canonical quantization scheme to develop a quantum model for the JJ array~\cite{devoret1995quantum}. This is done by first rewriting the Lagrangian into diagonal form
\begin{align}
\mathcal{L} = \sum_k \frac{C_k}{2} \dot{\phi}_k(t)^2 - \frac{1}{2L_k} \phi_k(t)^2.
\end{align}
This Lagrangian now yields the conjugate variables
\begin{align}
q_k = C_k \dot{\phi}_k^2,
\end{align}
such that we can introduce the quantum operators $\hat{\phi}_k$ and $\hat{q}_k$ satisfying the commutation relation $[\hat{\phi}_k, \hat{q}_k] = i\hbar$. The Hamiltonian for the system is then readily obtained as
\begin{align}
H = \sum_k \frac{\hat{q}_k^2}{2C_k} + \frac{\hat{\phi}_k^2}{2L_k},
\end{align}
which can be recast into the form
\begin{align}
H = \sum_k \omega_k a_k\dag a_k^{\phantom{\dagger}}.
\end{align}
by introducing the ladder operators $(a_k\dag, a_k)$ as
\begin{subequations}
\begin{align}
\hat{\phi}_k &= \sqrt{\frac{\hbar\omega_k L_k}{2}} (a_k\dag + a_k),\\
\hat{q}_k &= i\sqrt{\frac{\hbar\omega_k C_k}{2}} (a_k\dag - a_k).
\end{align}
\label{eq:phi}
\end{subequations}
The real physical phase variable is now described by the quantum variables, $\hat{\phi} = \sum_k \hat{\phi}_k \vec{\phi}_k$ and we confirm that for a given wave-function of the system the phase-variable is independent of the normalization of $\vec{\phi}_k$.
\subsection{Reintroduction of the nonlinearity}
\begin{figure}[t!]
\centering
\includegraphics[width=0.9\columnwidth]{figure_kerrkappa.pdf}
\caption{Calculated self-Kerr shifts per photon $K_{ll}$ for different modes of the Josephson junction array described in Sec. \ref{sec:exp}.} \label{fig:Kerrshifts}
\end{figure}
The potential energy quadratic in phase variables, as used in Eq. (\ref{eq:Larray}) implicitly assumes small phase excursions $\phi \ll \varphi_0$ (with $\varphi_0 = \hbar/(2e)$) for which Josephson junction potential $-E_J \cos(\phi / \varphi_0)$ is approximated as $\cos x \approx 1 - x^2$.
To study the effect of nonlinear contributions of the junctions in the array, we include the next term in the expansion that gives rise to a nonlinear term in the potential, $\mathcal{U}_{nl}$, seen by the phase $\phi$
\begin{align}
\mathcal{U}_{nl} = -\frac{1}{24L_J \varphi_0^2} \sum_{n=0}^N \big( \phi_n - \phi_{n+1} \big)^4.
\end{align}
Introducing the variable
\begin{align}
\Delta \phi_k (n) = {v}_k [n] - {v}_k [n+1],
\end{align}
where ${v}_k [n]$ denotes the $n$th entry in the vector $\vec{v}_k$, we can rewrite the potential in terms of the quantum mode operators as,
\begin{align}
\mathcal{U}_{nl} = -\sum_{k_1,k_2,k_3,k_4} \frac{ \prod_{j=1}^4 \hat{\phi}_{k_j}}{24L_J \varphi_0^2} \sum_{n=1}^N \prod_{j=1}^4 \Delta \phi_{k_j}(n). \label{eq:unl}
\end{align}
Generally, such a term will lead to Kerr terms as well as beam-splitter term. The beam-splitter terms are, however, only important between modes with small frequency difference, or if the system is strongly pumped by some external field of appropriate frequency that makes these terms resonant. Since we are primarily interested in the few photon regime of the lowest modes, which are all well-separated in frequency, we only consider the self-Kerr and the cross-Kerr terms. These can be expressed with the Hamiltonian
\begin{align}
H_{nl} = -\sum_{kl} K_{kl} a_k\dag a_k^{\phantom{\dagger}} a_l\dag a_l^{\phantom{\dagger}}, \label{eq:Hmkerr}
\end{align}
where the Kerr-coefficients are found by inserting Eq.~\eqref{eq:phi} into Eq.~\eqref{eq:unl} and rearranging the terms such that we obtain Eq.~\eqref{eq:Hmkerr} as
\begin{align}
K_{kl} = \frac{2-\delta_{kl}}{4L_J \varphi_0^2} \frac{\hbar \omega_k L_k}{2} \frac{\hbar \omega_l L_l}{2} \sum_{n=1}^N \Delta \phi_k (n)^2 \Delta \phi_l (n)^2, \label{eq:K}
\end{align}
with $\delta_{kl}$ being the Kronecker-delta. Figure \ref{fig:Kerrshifts} shows the Kerr shifts for the first eight array modes, calculated using the parameters for the array presented in Sec. \ref{sec:exp}. We note that the presence of end capacitances loads the array and decreases the mode frequencies significantly \cite{PhysRevLett.109.137002}. As evident from Eq.~(\ref{eq:K}), the mode frequencies and concomitant Kerr shifts per photon can be much larger for an unloaded Josephson array.
As a last important detail, we should mention that the nonlinear coupling between the modes also drags the eigenfrequencies down such that the real mode frequencies become
\begin{align}
\omega_k' = \omega_k - \sum_l K_{kl}.
\end{align}
This extra contribution is very important as this can shift the frequency of each mode by as much as 2 GHz for parameters similar to the experiment of Sec.~\ref{sec:exp}. The Hamiltonian for the array is, therefore, expressed as
\begin{align}
H = \sum_k \omega_k' a_k\dag a_k^{\phantom{\dagger}} - \sum_{kl} K_{kl} a_k\dag a_k^{\phantom{\dagger}} a_l\dag a_l^{\phantom{\dagger}}.
\end{align}
\section{Classical solution for Kerr oscillator}
\label{app:classical}
We adopt a simple classical description of the driven-dissipative Kerr resonator by replacing the quantum operator $a$ with the classical complex variable $\alpha$ and ignore vacuum fluctuations. This yields the equation of motion (for convenience we change the phase of the drive)
\begin{align}
\dot{\alpha} = -i\Delta \, \alpha + 2iK \, |\alpha|^2 \alpha - \frac{\kappa}{2} \, \alpha + \epsilon. \label{eq:simple_alpha}
\end{align}
In steady state, this leads to the following expression for $n = |\alpha|^2$,
\begin{align}
\varepsilon^2 = \Delta^2 n - 4\Delta K n^2 + 4K^2 n^3 + \frac{\kappa^2}{4} n, \label{eq:clas_ne}
\end{align}
which can be used to find the drive power needed for a given target photon number. We may take the inverse to obtain $n(\varepsilon)$, however, we are not ensured that this function will be single-valued. As a matter of fact, there will always be a set of parameters, $\Delta$ and $K$, for a given $\kappa$ where $n(\varepsilon)$ is multivalued and we can write the highest and lowest $n$'s in this bistable regime as
\begin{align}
n_{c,\pm} = -\frac{\Delta}{3 K} \Bigg[ 1 \mp \sqrt{1 - \frac{3}{4}\left(1 + \frac{\kappa^2}{4\Delta^{2}}\right)} \;\Bigg]. \label{eq:ncpm}
\end{align}
This expression for $n_{c,\pm}$ can now be inserted into Eq.~\eqref{eq:clas_ne} to get the critical drive power to be in the bistable regime. An alternative approach would be to numerically propagate the dynamical equation, Eq.~\eqref{eq:simple_alpha} until the steady state is reached. In contrast to the multi-valued solution of Eq.~\eqref{eq:clas_ne}, this would only give a single value, but if the parameters are chosen to be in bistable regime the steady state will depend on the initial condition. Experimentally this appears as hysteresis in the system, where, say, a continuous adiabatic change of $\Delta$ will yield different results for increasing or decreasing $\Delta$.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{figure_classical_regime.pdf}
\caption{The classically bistable regime with $K = \kappa/20$, $\Delta = 2.5\kappa$ and $\epsilon = 4\kappa$. In (a) we see a master equation simulation for an initial coherent state of amplitude $\alpha(0) = \sqrt{30}$ and with a different phase, $\alpha(0) = -\sqrt{30}$. Furthermore we have a simulation starting from vacuum and a simulation from vacuum for two different signs of the detuning. We do the same simulation in (b), but using the classical equation of Eq.~\eqref{eq:simple_alpha}. Panels on the right show the photon number distributions obtained for the final state, where $P(n)$ is the probability to be in the $n$-Fock state, for two different initial states: (c)~$\alpha(0) = -\sqrt{30}$ and (d)~$\alpha(0) =0$ (vacuum).} \label{fig:classical}
\end{figure}
It is worth noting that when $\kappa \gg K$, the dissipation in the resonator dominates the dynamics, destroying all the quantum coherent effects. In this respect, for a Kerr-resonator with $K/\kappa \ll 1$ a classical description of dynamics should suffice. To investigate this further, we compared the results of master equation and classical field equation simulations, shown in Fig.~\ref{fig:classical}. The first thing that we observe in Fig.~\ref{fig:classical} is that for an initial coherent state, $\alpha(0) = \sqrt{30}$, which is very close to the steady state, we have initially a small amount of oscillations in both the quantum [Fig.~\ref{fig:classical} (a)] and classical [Fig.~\ref{fig:classical} (b)] simulations before both converge to the same steady state value for the photon number. Interestingly, the initial phase of the oscillations is very different between the classical and quantum simulations.
The second observation is that if we change the initial phase of the coherent state to $\alpha(0) = - \sqrt{30}$, the difference between the quantum and the classical simulation becomes pronounced. The quantum simulation shows a small drop in photon number followed by a convergence towards a steady state value different from the one with opposite phase. The classical simulation shows oscillations, since we are far away from the steady state, but eventually the photon number converges to the same steady-state value that we found for the simulation with opposite sign on $\alpha(0)$. Now one might be tempted to interpret this behavior as the mean photon number being very different in the classical case and in the quantum case. Nonetheless, on taking a closer look at the steady state photon number distribution obtained from the quantum simulation [Fig.~\ref{fig:classical}(c)-(d)], we find that this is not the case. We see that the origin of the intermediate photon number comes from a dual-peaked photon number distribution; this can be interpreted as the steady state being a mixed state between the two peaks, each very close to a Poissonian distribution. The classical simulation is only single-valued, so it always selects only one of the solutions, depending on the size of the initial coherent state. We observe a similar behavior, when we start in the vacuum state. Here, there is always a very few number of photons, so the nonlinearity plays a smaller role. The quantum and the classical simulations are very similar, but, due to the quantum fluctuations, there is a small probability for the oscillator to end up in the high photon number state even when it starts in vacuum [see Fig.~\ref{fig:classical} (d)].
Finally, we see that if we change the sign of the detuning, $\Delta$, the quantum and classical simulation are identical. This is because the sign change moves us away from the bistable regime. So in both cases, the dynamics is deterministic and, since we are always at a low photon number, the nonlinearity again plays a small role.
\end{appendix}
|
\section{Introduction}
\label{sec:intro}
The geometry of the gas in the Central Molecular Zone (CMZ -- defined here as the region at a radial distance $R\lesssim 200\,{\rm pc}$ from the Galactic centre, or equivalently $ | l | \lesssim 1.5 \ensuremath{^\circ}$) has been intensively studied in recent years \citep[e.g.][]{Molinari+2011,Kruijssen+2015,Henshaw+16a,Sormani+2018}. In contrast, the region immediately surrounding the CMZ has received relatively little attention. However, it is well known that the CMZ is not an isolated system, but instead is strongly interacting with its surroundings. For example, the Galactic bar continuously drives a gas inflow into the CMZ which strongly affects its dynamics and may even drive the observed turbulence of the CMZ \citep{SormaniBarnes2019}.
Among the most enigmatic features in the region surrounding the CMZ is a discrete population of extremely broad-lined ($\Delta v > 100 \,{\rm km\, s^{-1}}$) compact clouds that are very prominent in molecular line datacubes (e.g.\ CO) in the region $| l |\leq 10 \ensuremath{^\circ}$ \citep[][]{Liszt2006,Liszt2008}. These features dominate the kinematics of molecular gas just outside the CMZ. The prototypical example is Bania Clump 2 \citep{StarkBania1986}. Despite their enormous velocity dispersion, these puzzling features are confined to a narrow longitude range. Similar features are not found anywhere else in the Galaxy. In this paper, we will refer to these features as Extended Velocity Features (EVFs) on account of their large velocity dispersions. We give a brief summary of the observational properties of the EVFs in Sect. \ref{sec:observations}.
Several possible interpretations of the EVFs have been put forward in the literature:
\begin{enumerate}[(a)]
\item They are gaseous structures extended in space that happen to coincidentally lie parallel to the line of sight \citep[e.g.][]{StarkBania1986,Boyce+89,Lee++1999,Baba++2010}. Such interpretations suffer from the `fingers of god' effect, i.e. they assume that we are at a special location in the universe in which these structures happen to point toward us.
\item Some of them have been interpreted as the footprints of giant magnetic loops caused by the \cite{Parker1966} instability near the Galactic centre \citep{Fukui+2006,Fujishita+2009,Machida+2009,Torii+2010,Suzuki+2015,Riquelme+2018}.
\item Some of them have been interpreted as evidence for the presence of intermediate-mass black holes (IMBH) \citep{Oka+2016,Oka+2017,Takekawa+2019,Takekawa+2019b}.
\item They are lumps which are just about to cross the dust lanes of the Milky Way bar (\citealt{Fux1999}, see also \citealt{Liszt2006,Liszt2008}).
\end{enumerate}
In this paper we show that features similar to the observed ones arise naturally in simulations of gas flow in a barred potential. We then use the insight gained from the simulations to interpret the observations. The paper is structured as follows. In Sect. \ref{sec:observations} we briefly review the observations and the key properties that characterise the EVFs. In Sect. \ref{sec:setup} we describe the numerical setup of our simulations. In Sects. \ref{sec:results} and \ref{sec:discussion} we discuss our results and interpret the observations. Finally in Sect. \ref{sec:conclusion} we sum up.
\section{Observations}
\label{sec:observations}
Here we briefly review the observational data. A more detailed analysis can be found for example in \citet{Liszt2006,Liszt2008} and \citet{Oka+2012} for CO, \citet{BoyceCohen1994} for OH, \citet{Longmore+2017} for NH$_3$ and \citet{McClureGriffiths+2012} for HI.
Fig. \ref{fig:data} shows molecular line emission from the inner Galaxy. The three most prominent EVFs are highlighted: these are the $l=5.4\ensuremath{^\circ}$, the $l=3.2\ensuremath{^\circ}$ (also known as Bania Clump 2) and the $l=1.3\ensuremath{^\circ}$ features. Other, less prominent EVFs can be found throughout the inner regions of our Galaxy (see references above).
Also highlighted are the dust lane features L1 to L4. These are not EVFs, but are often linked to them in $(l,b,v)$ space (see Property vi below). The "dust lane" terminology is used here for historical reasons despite these features being primarily (but not exclusively) detected in gas. The terminology originally comes from observations of external barred galaxies such as NGC 1300 or NGC 5383 in which one can see "the presence of two dust lanes leaving the nucleus one on each side of the bar and extending into the spiral arms" \citep{Sandage1961}. After it was realised that the MW is a barred galaxy, the features L1 and L4 were identified as the dust lanes of the MW bar \citep{Fux1999}, and the ``dust lane'' terminology was maintained despite the fact that they were initially observed in HI and CO emission, and not from dust emission/extinction. Subsequent work has identified the L1 and L4 features also from the dust \citep{Marshall+2008}. Beyond the two main dust lane features L1 and L4, \cite{Liszt2008} determined the presence of the two additional secondary dust lane features L2 and L3 using CO emission. As we will see later in the paper, the presence of multiple dust lanes also occurs in our simulations.
The key properties that characterise the EVFs are:
\begin{enumerate}
\item They are extremely broad-lined, with velocity dispersions of up to $200 \,{\rm km\, s^{-1}}$ when observed at low resolution.
\item They are compact, so they are very localised in the $(l,b)$ plane (the typical extensions of the largest EVFs are $\Delta l, \Delta b \sim 0.5\ensuremath{^\circ}$, )
\item They are usually more extended in latitude than in longitude. So they are typically elongated perpendicularly to the Galactic plane.
\item They are predominantly found in the $(v>0,l>0)$ and $(v<0,l<0)$ quadrants of the $(l,v)$ plane, although a few of them are found in the other two quadrants as well.
\item They never extend beyond the Terminal Velocity Curve (TVC)\footnote{The TVC at $l>0$ ($l<0$) is defined as the maximum (minimum) value of line-of-sight velocity at which the bulk of the emission from the Galactic disc is found, i.e. it is the curve that defines the envelope of the latitude-integrated $(l,v)$ diagram \citep[see for example][Chapter 9]{BM}.} at their value of longitude.
\item Some of them are clearly connected to some dustlane-like features associated with the Galactic bar (see for example how the $l=5.4\ensuremath{^\circ}$ EVF connects L1 to L3 or how the $l=3.2\ensuremath{^\circ}$ EVF is connected to L2, see also \citealt{Liszt2008}).
\item Some of them (e.g. Bania's Clump 2) show sharp HI emission profiles on one side \citep{McClureGriffiths+2012}.
\item When observed at very high resolution, they typically break-up into multiple kinematic sub-components with strong velocity gradients (see for example \citealt{Liszt2006} which resolved the internal velocity structure of several EVFs and fig. 30 in \citealt{Longmore+2017} which shows the complicated velocity structure of Bania Clump 2 in NH$_3$).
\end{enumerate}
Successful theoretical models should be able to reproduce the above properties.
\begin{figure*}
\includegraphics[width=0.8\textwidth]{images/data.pdf}
\caption{Molecular emission from the inner Galaxy. Some of the most prominent EVFs ($l=1.3\ensuremath{^\circ}$, $l=3.2\ensuremath{^\circ}$ a.k.a. Bania Clump 2 and $l=5.4\ensuremath{^\circ}$) and the dustlane-like features identified by \citet{Liszt2008} (L1 to L4) are indicated. The grey background shows the $^{12}$CO $J=1\to0$ data from \citet{Bitran+1997} (in the main panels) and \citet{Oka+1998} (in the zoom-in panels). The $l=5.4\ensuremath{^\circ}$, $l=3.2\ensuremath{^\circ}$ and L1 to L4 features are highlighted in the CO data. The magma colour scale in the centre shows HCN from the data of \citet{Jones+2012}. The HCN data only covers the region $ -0.7 < l < 1.8 \ensuremath{^\circ}$, $-0.3<b< 0.2\ensuremath{^\circ}$ and $-300 <v< 300 \,{\rm km\, s^{-1}}$. The $l=1.3\ensuremath{^\circ}$ feature is visible in the HCN data and is indicated with an arrow.}
\label{fig:data}
\end{figure*}
\section{Numerical setup} \label{sec:setup}
Our numerical setup is the same as that of \cite{Sormani+2018} except for a few differences. Therefore we only provide here a brief recap and state the differences from these previous simulations, and refer the reader to section 3 of \cite{Sormani+2018} and references therein for a more detailed description.
\subsection{Hydrodynamic code}
The simulations are three-dimensional and the gas is assumed to flow in a multi-component external barred potential $\Phi(\mathbf{x},t)$ which is constructed to fit the properties of the Milky Way (see next section and Appendix \ref{sec:potential}). The gas self-gravity and magnetic fields are neglected.
We use the moving-mesh code {\sc Arepo} \citep{Springel2010}, modified to treat the chemical evolution of the interstellar gas. The code solves the equations of fluid dynamics:
\begin{align}
& \frac{\partial \rho}{\partial t} + \nabla \cdot (\rho \mathbf{v}) = 0, \\
& \frac{ \partial (\rho \mathbf{v})}{\partial t} + \nabla \cdot \left( \rho \mathbf{v} \otimes \mathbf{v} + P \bold I \right) = - \rho \nabla \Phi, \\
& \frac{\partial(\rho e)}{\partial t} + \nabla \cdot \left[(\rho e + P) \mathbf{v} \right] = \dot{Q} + \rho \frac{\partial \Phi}{\partial t}, \label{eq:energy}
\end{align}
where $\rho$ is the gas density, $\mathbf{v}$ is the velocity, $P$ is the thermal pressure, $\bold I$ is the identity matrix, $e~=~e_{\rm therm} + \Phi + {\mathbf{v}^2}/{2} $ is the energy per unit mass, $e_{\rm therm}$ is the thermal energy per unit mass. We adopt the equation of state of an ideal gas, $P = (\gamma -1) \rho e_{\rm therm}$, where $\gamma=5/3$ is the adiabatic index.
We account for the chemical evolution of the gas using an updated version of the NL97 chemical network from \citet{gc12}, which itself was based on the work of \citet{gm07a,gm07b} and \citet{nl97}. With this network, we solve for the non-equilibrium abundances of H, H$_{2}$, H$^{+}$, C$^{+}$, O, CO and free electrons. An extensive description of the network is given in Section~3.4 of \citet{Sormani+2018} and in the interests of brevity we do not repeat it here.
The term $\dot{Q}$ in Equation~\ref{eq:energy} contains the contributions of the radiative and chemical processes that can change the internal energy of the system ($\dot{Q}=0$ for an adiabatic gas). It includes (i) a cooling function which depends on the instantaneous chemical composition of the gas \citep{glo10,gc12}; (ii) the heat absorbed or released in the most important chemical processes that occur in the interstellar medium, which are tracked in real time by the chemical network; (iii) external heating sources that represent the average Interstellar Radiation Field (ISRF) and cosmic ray ionisation rate. The strength of the ISRF is set to the standard value $G_0$ measured in the Solar neighbourhood \citep{draine78} diminished by a local attenuation factor which depends on the amount of gas present within 30~pc of each computational cell. This attenuation factor is introduced to account for the effects of dust extinction and H$_{2}$ self-shielding and is calculated using the {\sc Treecol} algorithm described in \cite{clark12}. The cosmic ray ionisation rate is fixed to $\zeta_{\rm H} = 3 \times 10^{-17} \: {\rm s^{-1}}$ \citep{gl78}. These values correspond to the `low' simulation of \cite{Sormani+2018}. We have shown in that paper the strength of the ISRF mainly controls the amount of molecular gas but makes little difference to the dynamics. Indeed, even if the ISRF field is a factor of a 1000 higher than in the Solar Neighbourhood \citep{Clark+2013}, the sound speed of the molecular gas comes nowhere close to the values of $c_{\rm s}=5\mhyphen10\,{\rm km\, s^{-1}}$ which would be needed to significantly affect the dynamics of the gas \citep{SBM2015a}. Hence the results of the present paper are not affected by the strength of the assumed ISRF.
\subsection{Differences between \citet{Sormani+2018} and the present paper}
The main difference between the simulations in \cite{Sormani+2018} and the one used in the present paper is that we modified the gravitational potential of the bar so that the size of the nuclear ring that naturally forms in the simulation matches the observed size of the CMZ (it was a factor of $\sim 2$ too large in the previous simulations). In general, the size of this ring is controlled by (i) the parameters of the gravitational potential, mainly the bar strength and the bar pattern speed \citep[e.g.][and references therein]{Sormani+2018b} (ii) the effective sound speed of the gas (see for example Fig.~1 of \citealt{SBM2015a}). Since the sound speed of the gas is fixed by our treatment of the heating \& cooling of the ISM and the pattern speed of the gas is independently constrained to be $\Omega_{\rm p} = 40 \,{\rm km\, s^{-1}} \,{\rm kpc}^{-1}$ \citep[e.g.][]{SBM2015c,Portail+2017,Sanders+2019}, we have increased the strength of the bar (compatibly with with known observational constraints) to achieve the desired result of a smaller ring. The gravitational potential and the resulting rotation curve are described in detail in Appendix \ref{sec:potential}.
The second difference is that we increased the resolution. The resolution in the simulation is determined by the condition that cells approximately have the same mass (so that denser gas has a higher spatial resolution). The system of mass refinement present in {\sc Arepo} ensures that this condition is satisfied by splitting cells whose mass becomes greater than twice this target mass and merging cells whose mass is too low. Here we use a target resolution of $25 \, \rm M_\odot$ per cell, while in \cite{Sormani+2018} we used a target resolution of $100 \, \rm M_\odot$.
The last difference is in the initial density profile of the gas. In \cite{Sormani+2018} the initial density distribution was approximately uniform inside a cylindrical slab of radius $10~\,{\rm kpc}$ and half-height $1~\,{\rm kpc}$, with the addition of some small random noise. Here instead we initialise the density according to the following axisymmetric density distribution:
\begin{equation}
\rho(R,z) = \frac{\Sigma_0}{4 z_{\rm d}} \exp\left(- \frac{R_{\rm m}}{R} - \frac{R}{R_{\rm d}}\right) \sech\left(\frac{z}{2 z_{\rm d}}\right)^2\, ,
\end{equation}
where $(R,\phi,z)$ denote standard cylindrical coordinates, $z_{\rm d} = 85 \,{\rm pc}$, $R_{\rm d} = 7 \,{\rm kpc}$, $R_{\rm m} = 1.5 \,{\rm kpc}$, $\Sigma_0 = 50 {\rm M_\odot} \,{\rm pc}^{-2}$ and we also have cut our disc so that $\rho=0$ for $R\geq 5\,{\rm kpc}$. This profile better matches the observed radial distribution of gas in the Galaxy \citep{KalberlaDedes2008,HeyerDame2015}. The initial density distribution is very smooth and we do not include any random noise. Despite this smoothness of the initial conditions, the gas flow in the bar region ends up being unsteady and turbulent because of the processes described in sect. 4 of \cite{Sormani+2018}.
\section{Results}
\label{sec:results}
\begin{figure*}
\includegraphics[width=1.0\textwidth]{images/patches.pdf}
\caption{The snapshot of our simulation at $t=178 \,{\rm Myr}$. \emph{Top row}: surface density of gas in the $(x,y)$ plane. \emph{Bottom row:} corresponding projections in the $(l,v)$ plane in the optically thin approximation and assuming that the angle between the Sun-GC line and the major axis of the bar is $\phi=20\ensuremath{^\circ}$ \citep{BlandHawthornGerhard2016}. The left and middle column show HI and CO respectively as calculated by the chemical network included in the simulation. The right column shows a colour coded map on top of the CO emission, allowing one to identify corresponding structures in the $(x,y)$ and $(l,v)$ views. A movie showing a 3D visualisation of the snapshot shown in this figure is available at \url{http://www.ita.uni-heidelberg.de/~mattia/videos/EVF/flyby.mp4}.}
\label{fig:patches}
\end{figure*}
\begin{figure*}
\includegraphics[width=1.0\textwidth]{images/patches_map.pdf}
\caption{Features in the $(x,y)$ plane and their projection to the $(l,v)$ plane for the simulation snapshot at $t=178 \,{\rm Myr}$. The top panels are zoom-ins of the bottom panels. Arrows in the left panels show the velocity field in the rotating frame of the bar. Labels mark some of the interesting features. The feature V1 resembles the EVF observed at $l=5.4\ensuremath{^\circ}$ in Fig. \ref{fig:data}. The feature V1 originates as the material on the `overshooting' feature O1 (which has passed very close to the CMZ and brushed it) crashes onto the dust lane feature D1.}
\label{fig:patchesmap}
\end{figure*}
Figure \ref{fig:patches} shows a snapshot of our simulation at $t=178\,{\rm Myr}$. The top row shows the HI and CO surface density in the $(x,y)$ plane, while the bottom row shows the corresponding projections in the $(l,v)$ plane. To produce these projections, we bin each {\sc Arepo} cell as a point in the $(l,v)$ plane with a weight proportional to the mass of the component of interest (HI or CO, as appropriate) and inversely proportional to the square of its distance from the Sun, as discussed in more detail in Section 3.6 of \citet{Sormani+2018}.
These projections assume that the gas is optically thin to HI and CO line emission, but accounting more accurately for line opacities would only change the strength of the emission and not its distribution in the $(l,v)$ plane. Figure \ref{fig:patchesmap} shows the correspondence between top down and projection plots in more detail, with labels that identify some of the interesting features.
Several features that resemble the observed EVFs can be identified in the various $(l,v)$ projections. A particularly prominent one is the red feature at $l\simeq 3 \ensuremath{^\circ}$ labelled V1 in Fig. \ref{fig:patchesmap}. This feature has an extreme velocity dispersion ($\Delta v \sim 200\,{\rm km\, s^{-1}}$) but is very localised in real $(x,y)$ space (it all originates from the small red patch visible in the top-left panel of Fig. \ref{fig:patchesmap}). This is precisely the main property that characterises the observed EVFs (see Sect. \ref{sec:observations}). The V1 feature connects the dust lane features D1 and D2 (see labels in Fig. \ref{fig:patchesmap}). This is remarkably similar to the what is observed for the EVF at $l\simeq 5.4 \ensuremath{^\circ}$ in Fig. \ref{fig:data}, which connects the main observed dust lane L1 to the secondary dust lane L3.
Inspection of the velocity fields in Figs. \ref{fig:patchesmap} and \ref{fig:losvelocity} reveals the origin of the feature V1. It originates as gas on the feature O1 crashes into the dust lane feature D1. The feature O1 is gas that has fallen along the dust lane on the opposite side, touched \& brushed the CMZ, and then continued its course until it crashed into the middle of feature D1.\footnote{Using high-sensitivity CO data \cite{MizunoFukui2004} have identified what might be the observational counterpart of the overshooting feature O1 (see crosses in their Fig. 3). This feature seems to connect to the $l=5.4\ensuremath{^\circ}$ feature in the three-dimensional $(l,b,v)$ space in a manner very similar to how the O1 feature connects to the V1 feature in our simulations. This however needs to be confirmed by future observations.} When the feature O1 comes in contact with feature D1, the two have enormously different velocities. The signature of this collision in the $(l,v)$ plane is the extreme velocity dispersion that characterises the feature V1.
Figure \ref{fig:patchesmap} also shows the presence of several further features with high velocity dispersion at negative longitudes. These are coloured green. These features originate with a similar mechanism as the feature V1 discussed above. They look more crowded in the $(l,v)$ plane partly on account of projection effects (they are on the far side of the Galaxy). The production of the EVFs is a stochastic process in the simulation on account of the unsteady and turbulent flow that develops due to the processes described in section 4 of \cite{Sormani+2018}.
A second type of broad-lined features that have a somewhat different origin than the ones described above also appear during the course of the simulation. Figure \ref{fig:patchesmapCMZ} shows an example of this second type of EVF. It is labelled V2 in the figure. This second type of feature originates as material that has been falling along the dust lanes crashes into the CMZ. The dense material in the CMZ typically has velocities much lower than those of the dust lanes, so when they collide they produce very broad-lined features like V2 in the figure. This feature has much in common with the observed EVF at $l=1.3\ensuremath{^\circ}$ (see Fig. \ref{fig:data}).
\begin{figure}
\includegraphics[width=0.5\textwidth]{images/losvelocity.png}
\caption{Line-of-sight velocity in the $(x,y)$ plane. The larger circle highlights where the feature V1 shown in Fig. \ref{fig:patchesmap} originates. In this region, material with very different line-of-sight velocities collides, producing the large velocity dispersion observed in the $(l,v)$ plane. The smaller circle highlights a region at the outer edges of the CMZ, where the dust lane brushes the CMZ. This behaviour also brings into contact material with very different velocities and can give rise to EVFs.}
\label{fig:losvelocity}
\end{figure}
\begin{figure*}
\includegraphics[width=1.0\textwidth]{images/patches_map_CMZ.pdf}
\caption{Features in the $(x,y)$ plane and their projection to the $(l,v)$ plane in the central regions for the simulation snapshot at $t=191 \,{\rm Myr}$. The feature V2 illustrates the second type of EVF. This is created as incoming material from the dust lanes crashes into the CMZ.}
\label{fig:patchesmapCMZ}
\end{figure*}
\begin{figure}
\includegraphics[width=0.45\textwidth]{images/zoomV1.png}
\caption{Zoom-in that shows the 3D CO Position-Position-Velocity structure of the feature V1 in Fig. \ref{fig:patchesmap}. A movie showing the feature from different orientations is available at \url{http://www.ita.uni-heidelberg.de/~mattia/videos/EVF/zoomV1.mp4}.}
\label{fig:zoomV1}
\end{figure}
\begin{figure}
\includegraphics[width=0.45\textwidth]{images/zoomV2.png}
\caption{Zoom-in that shows the 3D CO Position-Position-Velocity structure of the feature V2 in Fig. \ref{fig:patchesmapCMZ}. A movie showing the feature from different orientations is available at \url{http://www.ita.uni-heidelberg.de/~mattia/videos/EVF/zoomV2.mp4}.}
\label{fig:zoomV2}
\end{figure}
\section{Discussion}
\label{sec:discussion}
The results in the previous section suggest that at least some (perhaps most) of the EVFs found in the observations originate from collisions. These typically involve gas falling along the dust lanes that crashes with material with very different line-of-sight velocities. Our simulations show that this occurs naturally when gas flows in a barred potential and cannot be avoided: our initial conditions are prepared ensuring that the gas is as calm as possible (they are very smooth, symmetric and do not include any random noise), yet such collisions develop spontaneously. This happens even in the absence of any form of stellar feedback.\footnote{Indeed, the gas flow in a barred potential is inevitably unsteady and turbulent \citep[][]{Sormani+2018}. This is well illustrated for example in the top-left panel in Fig. \ref{fig:patches}, which shows that inside the bar region the flow is structured and unsteady, in striking contrast with the flow just outside the bar region which is extremely smooth and steady.}
The extended velocity features occur frequently in the simulations although perhaps at any given time there are somewhat fewer of them in a synthetic $(l,v)$ diagram than in its observational counterpart. This is probably a consequence of the fact that we have tried to keep the gas flow as smooth as possible, while in the real Galaxy more collisions should be expected on account of the facts that the initial conditions are most likely not smooth and that additional processes contribute to produce more unsteadiness and turbulence (stellar feedback, perturbations from satellite galaxies that punch through the MW disc, etc). Thus our simulations provide a lower limit on the number of EVF-producing collisions that might be expected in the real Galaxy.
Our interpretation naturally explains most of the key observational properties listed in Section~\ref{sec:observations}. Property (i) is satisfied because this is the property by which we select features in the simulation to compare to the observed EVFs. Property (ii) is satisfied because the collision sites have limited extension in real $(x,y)$ space, so the features are localised in the $(l,b)$ plane. Property (iv) is satisfied because collisions in the simulations happen preferentially in the two quadrants $(l>0,v>0)$ and $(l<0,v<0)$, although not exclusively (see for example the green material in Fig. \ref{fig:patchesmap}). Property (v) is satisfied because colliding clouds are part of the general large-scale flow and so their velocities are always within the limits defined by the TVC. Property (vi) is satisfied because in our interpretation some of the features are naturally connected with the dust lanes.
With our existing simulation, we are not able to verify whether features formed in this way satisfy Property (iii). One of the unrealistic properties of our simulation is that the gas layer is too thin compared to observations (typical thickness of molecular gas in the simulations is only $H \sim 10 \,{\rm pc}$), probably due to the lack of stellar feedback \citep[see the discussion in section 5.5.1 of][]{Sormani+2018}. The thinness of the simulations can also be appreciated from the movies linked in the Supplementary Information section below. Hence, on scales much larger than $\sim10\,{\rm pc}$ the gas is always more elongated in longitude than in latitude in our simulations, contrary to Property (iii). However, we might argue that both Property (iii) and (vii) may be expected for more realistic (and therefore more vertically `puffed up') clouds within the context of our interpretation. When two clouds collide at high speed, we expect them to be compressed in the direction of motion (in this case, the $l$ direction). This might explain Property (iii). Similarly, one might expect that a collision produces a strong compression shock on one side, visible as a sharp edge (Property vii).
The masses of the features in the simulations are comparable to the masses of the observed EVFs. For example. the mass of the feature V1 in the simulation is $\simeq 2.5\times 10^{6} \, \rm M_\odot$ while the mass of the observed $l=5.4\ensuremath{^\circ}$ feature has been estimated by \cite{Liszt2006} as $\simeq 5 \times 10^{6} \, \rm M_\odot$. This is a good agreement given that (a) the processes that produce the EVFs and therefore their masses are stochastic and (b) the masses measured from the observations are very uncertain due to the uncertainty in the CO-to-H$_{2}$ conversion factor ($X_{\rm CO}$). Indeed, standard assumptions made to calibrate $X_{\rm CO}$ such as virial equilibrium \citep[e.g.][]{Bolatto+2013} are most likely not valid for the features considered here which are in a highly dynamical environment.
As noted in Sect. \ref{sec:observations} (Property viii), observed EVFs typically have a very complicated internal Position-Position-Velocity (PPV) structure and break-up into several sub-components with strong velocity gradients when observed at very high resolution. What is the small scale structure of the EVFs obtained in the simulations? To investigate this we show in Figs. \ref{fig:zoomV1} and \ref{fig:zoomV2} the CO PPV maps of the features V1 and V2 studied above. Movies that show the same features from different orientations are also available at the link provided in the Supplementary Information section below. These show that V1 and V2 are indeed connected structures in 3D physical Position-Position-Position space, and not coincidental amalgamations of unconnected components. The simulated EVFs posses a certain degree of internal structure (particularly the kinematical structure of V2 appears to be significantly more complex than of V1), but the real observed EVFs display a much higher degree of complexity (compare Figs. \ref{fig:zoomV1} and \ref{fig:zoomV2} with figs. 6,7,8,9 of \citealt{Liszt2006} and fig. 30 of \citealt{Longmore+2017}). This is not unexpected given that our simulations start out very smooth and lack any kind of stellar feedback, self-gravity and/or initial noise that could generate substructure, so that on small scales clouds tend be much smoother than their counterparts in the real Galaxy. It is however interesting to note that the simulated EVFs \emph{do} have some substructure due to the unsteady gas flow caused by the bar, in contrast to the gas outside the bar region which is extremely smooth.\footnote{The smoothness outside the bar region can be appreciated for example in the visualisation at the following link: \url{http://www.ita.uni-heidelberg.de/~mattia/videos/EVF/flyby.mp4}.}\footnote{The observations, as seen for example in \cite{Longmore+2017}, also seem to indicate a possible connection between the statistics of the small-scale velocity structure and the type of EVF. Since our simulations are currently unable to reproduce the small-scale complexity of EVFs, we refrain from specifying the expected statistics of the different small-scale velocity structures. However, this is worth further study.} Another aspect that is evident from Figs. \ref{fig:zoomV1} and \ref{fig:zoomV2} is the small vertical extent (i.e. in the $z$ direction) of our simulations discussed above. Despite these caveats, the comparison shows that the simulated EVFs may be identified with the bulk gas of the observed EVFs.
Finally, we note the following. In the previous section we have identified two mechanisms that produce collisions (and therefore EVFs) in our simulations. The first is overshooting material which collides with the dust lanes on the opposite side, exemplified by feature V1 in Fig. \ref{fig:patchesmap}. The second is material on the dust lanes which collides with CMZ material, exemplified by feature V2 in Fig. \ref{fig:patchesmapCMZ}. However we cannot exclude that in a more turbulent, realistic environment further mechanisms that generate collisions are possible. For example, multiple dust lanes are generally expected to be very close in real space although they have very different line-of-sight velocities. A relatively small perturbation to the velocity field (induced for example by an external perturbation such as accretion from the circumgalactic medium or stellar feedback) may cause them to touch. This would lead to a transfer of material between the two (the faster dust lane is decelerated, while the slower one is accelerated), which in the $(l,v)$ diagram would show up as an EVF. The key point is that velocity dispersions of the order of $\sim 100 \mhyphen 200 \,{\rm km\, s^{-1}}$ (comparable to the velocity of the Sun around the Galactic centre) suggest that collisions between large-scale Galactic flows are involved. The presence of a bar creates the perfect environment to make such collisions likely.
\subsection{Comparison with previous work}
Compared to the other interpretations (a) to (d) listed in the introduction, we note the following. Unlike interpretation (a), according to which EVFs are extended structures that coincidentally line parallel to the line-of-sight, our interpretation does not suffer from the `finger of god' effect. The patches of gas producing the EVFs in our simulations are always localised in $(x,y)$ space and in general do not correspond to structures which are elongated along the line-of-sight. For example, we have verified that our features remain `extended' in the $(l,v)$ plane even if observed at different angles $\phi$ between the major axis of the bar and the Sun-GC line.
According to interpretation (b) magnetic instabilities alone (without a bar potential) are responsible for creating the EVFs. However, the synthetic $(l,v)$ diagrams produced from simulations of this mechanism performed to date \citep{Machida+2009,Suzuki+2015,Kakiuchi+2018} do not seem to be able to convincingly reproduce the morphology of the EVFs in the $(l,v)$ plane (Properties i and ii in Sect. \ref{sec:observations}). Moreover, in this interpretation the connection with the dust lanes of the MW bar (Property vi in Sect. \ref{sec:observations}) remains unexplained. Nevertheless, it is possible that magnetic fields, when added on top of the bar potential, play a role in shaping the properties and morphologies of the EVFs.
Interpretation (c) assumes that EVFs are created by gravitational kicks around IMBHs. According to this interpretation, the large velocity dispersion seen in an EVF should depend on the impact parameter of the incoming gas cloud relative to the IMBH and on the mass of the IMBH, and should have no relation to the TVC and/or to the dust lanes features of the MW. Hence in this interpretation it is unclear why the EVFs never extend beyond the TVC at their longitudes (Property v) and why they seem to be associated with the dust lanes of the MW (Property vi). This interpretation also posits an ad-hoc assumption, namely the presence of IMBHs, which is unnecessary since it can be avoided in our interpretation. Finally, we note that in the case of the CO-0.40-0.22 cloud, an EVF that has been claimed to be the signature of an IMBH close to the Galactic centre \citep{Oka+2017}, constraints on the radio spectrum and a detection of a mid-infrared point source both disfavour the presence of an IMBH \citep{Ravi+2018}.
The interpretation (d) of \cite{Fux1999} is essentially the same that we have given in this paper, but in an embryonic state. The simulations of \cite{Fux1999} did not possess the necessary resolution to actually see the EVFs in the synthetic $(l,v)$ diagrams. Fux speculated about the implications of his simulations and imagined what he would have seen if he had higher resolution. We have refined the \cite{Fux1999} interpretation by correcting some parts (e.g.\ the clumps are not really `crossing' the dust lane and exiting on the other side as Fux suggested, but instead are joining and merging with the dust lane and then flowing together towards the central regions) and filling in some details (e.g.\ the origin of some of the clumps hitting the dust lane are clumps from the dust lanes on the other side that have overshot).
\begin{figure}
\includegraphics[width=0.4\textwidth]{images/sketch.png}
\caption{Sketch of how the geometry of the gas surrounding the CMZ might look like according to our interpretation. Coloured straight lines represent the various dust lanes of the MW. The purple circle represents the CMZ. The two yellow clouds on the near side dust lanes represent the $l=5.4\ensuremath{^\circ}$ and $l=3.2\ensuremath{^\circ}$ (aka Bania Clump 2) EVFs in Fig. \ref{fig:data} respectively. The yellow cloud on at the intersection between dust lanes and CMZ represents the $l=1.3\ensuremath{^\circ}$ EVF in Fig. \ref{fig:data}.}
\label{fig:sketch}
\end{figure}
\subsection{Implications for the observations} \label{sec:implications}
In Section~\ref{sec:results}, we showed that we can distinguish two basic types of EVFs in the simulation. The first is produced by material on the dust lanes that collides with overshooting material from the other side. The second is produced by material on the dust lanes which collides with CMZ material.
In the observations some features can be associated quite clearly with one or the other of these two possibilities. For example, the features at $l=5.4\ensuremath{^\circ}$ and $l=3.2\ensuremath{^\circ}$ in Fig. \ref{fig:data} are most likely of the first type (because of the way they are associated with the dust lane features L1-L3 and because their longitudes place them outside the CMZ), while the feature at $l=1.3\ensuremath{^\circ}$ is most likely of the second type (because it connects with dense CMZ gas, see Fig \ref{fig:data}). Figure \ref{fig:sketch} shows a sketch of how the geometry of the gas surrounding the CMZ might look like according to our interpretation. For other features in the observations the situation is more ambiguous, and one needs to study this on a case by case basis, checking for example the connection between them and other features such as dust lane features and using high resolution data, which is outside the scope of the present paper.
The HI projection in Fig. \ref{fig:patches} displays many features that cannot be seen in the CO projection. Thus we expect that several features which are invisible in high-density tracers may be detected in low-density observational tracers such as HI (or the CO $J=1\to0$ line observed with high enough sensitivity). The features identified in low-density tracers can be used to connect the features seen in higher-density tracers such as NH$_3$ or HCN. This will be necessary to get a complete picture of the 3D geometry and gas flows in and around the CMZ.
Finally, it is worth noting that if our interpretation of the $l=5.4\ensuremath{^\circ}$ and $l=3.2\ensuremath{^\circ}$ EVFs is correct, we are directly witnessing collisions at a relative speed of $\Delta v\sim 200\,{\rm km\, s^{-1}}$. This is a perfect laboratory for studying what happens when two molecular clouds with masses in excess of $M=10^6 \, \rm M_\odot$ collide with each other with extreme velocities. We expect to find a rich chemistry and the presence of shock tracers associated with these features in the observations. If the interpretation of the $l=1.3\ensuremath{^\circ}$ EVF is correct, we are directly witnessing material that is accreting onto the CMZ. Studying this feature in more detail can, therefore, give insight on the process of accretion as it is happening and on the physical and chemical condition of the accreted gas.
\section{Summary} \label{sec:conclusion}
Surrounding the Galactic centre there exist an enigmatic population of compact molecular clouds with extreme velocity dispersions. These Extended Velocity Features (EVFs) dominate the kinematics of gas just outside the Central Molecular Zone. We have used hydrodynamical simulations of gas flow in a barred potential to interpret these clouds. We have found that similar features occur naturally in these simulations. They originate from collisions between material that is falling along the dust lanes of the bar and material with substantially different line-of-sight velocities. We have distinguished between two types:
\begin{enumerate}
\item EVFs like the feature V1 in Fig. \ref{fig:patchesmap} which originate from the collision between material on the dust lanes and material that has `overshot' from the dust lane on the opposite side;
\item EVFs like the feature V2 in Fig. \ref{fig:patchesmapCMZ} which originate from the collision between material on the dust lanes and material belonging to the CMZ.
\end{enumerate}
Examples of both types of features can be identified in the observations. The sketch in Fig. \ref{fig:sketch} shows our proposed interpretation of the most unambiguous features. Other features can be identified in the data, but the interpretation is more ambiguous and will require more careful analysis with higher resolution observations.
If our interpretation is correct, we are witnessing clouds colliding at relative velocities of $\Delta v \sim 200\,{\rm km\, s^{-1}}$ (e.g. the $l=5.4\ensuremath{^\circ}$ and $l=3.2\ensuremath{^\circ}$ clouds). This provides an excellent laboratory to study extreme cloud collisions. We are also directly witnessing gas being accreted onto the CMZ (e.g. the $l=1.3\ensuremath{^\circ}$ cloud). This provides a unique opportunity to study how gas is accreted and the physical and chemical properties of the accreted gas.
\section*{Supplementary information}
Movies showing the time evolution of the simulations, a 3D visualisation of the snapshot shown in Fig. \ref{fig:patchesmap} and the three dimensional PPV structure of the features V1 and V2 can be found at the following link: \url{http://www.ita.uni-heidelberg.de/~mattia/download.html}.
\section*{Acknowledgements}
MCS thanks Tom Dame and Harvey Liszt for insightful comments and discussions. We thank Jean-charles Lambert for developing {\sc GLNEMO2}, a freely-distributed interactive visualization 3D software for N-body snapshots which is publicly available at the following link: \url{https://projets.lam.fr/projects/glnemo2}. MCS, RGT, SCOG, and RSK acknowledge support from the Deutsche Forschungsgemeinschaft via the Collaborative Research Centre (SFB 881) ``The Milky Way System'' (subprojects B1, B2, and B8) and the Priority Program SPP 1573 ``Physics of the Interstellar Medium'' (grant numbers KL 1358/18.1, KL 1358/19.2, and GL 668/2-1). RSK furthermore thanks the European Research Council for funding in the ERC Advanced Grant STARLIGHT (project number 339177). ATB would like to acknowledge the funding provided from the European Union's Horizon 2020 research and innovation programme (grant agreement No 726384). CDB and HPC gratefully acknowledges support for this work from the National Science Foundation under Grant No. (1816715). PCC acknowledges support from the Science and Technology Facilities Council (under grant ST/N00706/1) and StarFormMapper, a project that has received funding from the European Union's Horizon 2020 Research and Innovation Programme, under grant agreement no. 687528. HPH thanks the LSSTC Data Science Fellowship Program, which is funded by LSSTC, NSF Cybertraining Grant No. 1829740, the Brinson Foundation, and the Moore Foundation; his participation in the program has benefited this work. RJS gratefully acknowledges an STFC Ernest Rutherford fellowship (grant ST/N00485X/1) and HPC from the Durham DiRAC supercomputing facility (grants ST/P002293/1, ST/R002371/1, ST/S002502/1, and ST/R000832/1). The authors acknowledge support by the state of Baden-W\"urttemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1134-1 FUGG.
\def\aap{A\&A}\def\aj{AJ}\def\apj{ApJ}\def\mnras{MNRAS}\def\araa{ARA\&A}\def\aapr{Astronomy \&
Astrophysics Review}\defApJS}\def\apjl{ApJ}\def\pasj{PASJ}\def\nat{Nature}\def\prd{Phys. Rev. D{ApJS}\def\apjl{ApJ}\def\pasj{PASJ}\def\nat{Nature}\def\prd{Phys. Rev. D}
\defSpace Sci. Rev.}\def\pasp{PASP}\def\aaps{A\&AS{Space Sci. Rev.}\def\pasp{PASP}\def\aaps{A\&AS}
\bibliographystyle{mn2e}
|
\section{Introduction}\label{sec:Introduction}
This paper addresses block-coordinate (BC) proximal gradient methods for problems of the form
\begin{equation}\label{eq:P}
\minimize_{\bm x=(x_1,\dots,x_N)\in\R^{\sum_in_i}}
\Phi(\bm x)
{}\coloneqq{}
F(\bm x)
{}+{}
G(\bm x),
\quad\text{where}\quad
\textstyle
F(\bm x)\coloneqq\tfrac1N\sum_{i=1}^N f_i(x_i),
\end{equation}
in the following setting.
\begin{ass}[problem setting]\label{ass:basic
In problem \eqref{eq:P} the following hold:
\begin{enumeratass}
\item\label{ass:f
function \(f_i\) is \(L_{f_i}\)-smooth (Lipschitz differentiable with modulus \(L_{f_i}\)), \(i\in[N]\);
\item\label{ass:g
function \(G\) is proper and lower semicontinuous (lsc);
\item\label{ass:phi
a solution exists: \(\argmin\Phi\neq\emptyset\).
\end{enumeratass}
\end{ass}
Unlike typical cases analyzed in the literature where \(G\) is separable \cite{tseng2001convergence,tseng2009coordinate,nesterov2012efficiency,beck2013convergence,bolte2014proximal,richtarik2014iteration,lin2015accelerated,chouzenoux2016block,hong2017iteration,xu2017globally}, we here consider the complementary case where it is only the smooth term \(F\) that is assumed to be separable. The main challenge in analyzing convergence of BC schemes for \eqref{eq:P} especially in the nonconvex setting is the fact that even in expectation the cost does not necessarily decrease along the trajectories.
Instead, we demonstrate that the forward-backward envelope (FBE) \cite{patrinos2013proximal,themelis2018forward} is a suitable Lyapunov function for such problems.
Several BC-type algorithms that allow for a nonseparable nonsmooth term have been considered in the literature, however, all in convex settings.
In \cite{tseng2008block,tseng2010coordinate} a class of convex composite problems is studied that involves a linear constraint as the nonsmooth nonseparable term.
A BC algorithm with a Gauss-Southwell-type rule is proposed and the convergence is established using the cost as Lyapunov function by exploiting linearity of the constraint to ensure feasibility. A refined analysis in \cite{necoara2013random,necoara2014random} extends this to a random coordinate selection strategy.
Another approach in the convex case is to consider randomized BC updates applied to general averaged operators. Although this approach can allow for fully nonseparable problems, usually separable nonsmooth functions are considered in the literature.
The convergence analysis of such methods relies on establishing quasi-Fej\'er monotonicity \cite{iutzeler2013asynchronous,combettes2015stochastic,pesquet2015class,bianchi2016coordinate,peng2016arock,latafat2019new}.
In a primal-dual setting in \cite{fercoq2019coordinate} a combination of Bregman and Euclidean distance is employed as Lyapunov function.
In \cite{hanzely2018sega} a BC algorithm is proposed for strongly convex algorithms that involves coordinate updates for the gradient followed by a full proximal step, and the distance from the (unique) solution is used as Lyapunov function.
The analysis and the Lyapunov functions in all of the above mentioned works rely heavily on convexity and are not suitable for nonconvex settings.
Thanks to the nonconvexity and nonseparability of \(G\), many machine learning problems can be formulated as in \eqref{eq:P}, a primary example being constrained and/or regularized finite sum problems \cite{bertsekas2011incremental,shalevshwartz2013stochastic,defazio2014finito,defazio2014saga,mairal2015incremental,reddi2016proximal,reddi2016stochastic,schmidt2017minimizing}
\begin{equation}\label{eq:FSP}
\textstyle
\minimize_{x\in\R^n}
\varphi(x)
{}\coloneqq{}
\tfrac1N\sum_{i=1}^N f_i(x)
{}+{}
g(x),
\end{equation}
where \(\func{f_i}{\R^n}{\R}\) are smooth functions and \(\func{g}{\R^n}{\Rinf}\) is possibly nonsmooth, and everything here can be nonconvex.
In fact, one way to cast \eqref{eq:FSP} into the form of problem \eqref{eq:P} is by setting
\begin{equation}\label{eq:FINITOG}
\textstyle
G(\bm x)
{}\coloneqq{}
\tfrac1N\sum_{i=1}^Ng(x_i)
{}+{}
\indicator_C(\bm x),
\end{equation}
where
\(
C
{}\coloneqq{}
\set{\bm x\in\R^{nN}}[x_1=x_2=\dots=x_N]
\)
is the consensus set, and \(\indicator_C\) is the indicator function of set \(C\), namely
\(
\indicator_C(\bm x)=0
\)
for \(\bm x\in C\) and \(\infty\) otherwise.
Since the nonsmooth term \(g\) is allowed to be nonconvex, formulation \eqref{eq:FSP} can account for nonconvex constraints such as rank constraints or zero norm balls, and nonconvex regularizers such as \(\ell^p\) with \(p\in[0,1)\), \cite{hou2012complexity}.
Another prominent example in distributed applications is the \emph{``sharing''} problem \cite{boyd2011distributed}:
\begin{equation}\label{eq:SP}
\minimize_{\bm x\in\R^{nN}}\Phi(\bm x)
{}\coloneqq{}
\textstyle
\tfrac1N\sum_{i=1}^Nf_i(x_i)
{}+{}
g\Bigl(\sum_{i=1}^Nx_i\Bigr)
.
\end{equation}
where \(\func{f_i}{\R^n}{\R}\) are smooth functions and \(\func{g}{\R^n}{\Rinf}\) is nonsmooth, and all are possibly nonconvex. The sharing problem is cast as in \eqref{eq:P} by setting \(G\coloneqq g \circ A\), where \(A\coloneqq[\I_n~\dots~\I_n]\in\R^{n\times nN}\) (\(I_r\) denotes the \(r\times r\) identity matrix).
\subsection{The main block-coordinate algorithm}\label{sec:BC}
While gradient evaluations are the building blocks of smooth minimization, a fundamental tool to deal with a nonsmooth lsc term \(\func{\psi}{\R^r}{\Rinf}\) is its \DEF{\(V\)-proximal mapping}
\begin{equation}\label{eq:prox}
\prox_\psi^V(x)
{}\coloneqq{}
\argmin_{w\in\R^r}\set{
\psi(w)
{}+{}
\tfrac12\|w-x\|^2_V
},
\end{equation}
where \(V\) is a symmetric and positive definite matrix and \(\|{}\cdot{}\|_V\) indicates the norm induced by the scalar product \((x,y)\mapsto\innprod{x}{Vy}\).
It is common to take \(V=t^{-1}\I_r\) as a multiple of the \(r\times r\) identity matrix \(\I_r\), in which case the notation \(\prox_{t\psi}\) is typically used and \(t\) is referred to as a stepsize.
While this operator enjoys nice regularity properties when \(g\) is convex, such as (single valuedness and) Lipschitz continuity, for nonconvex \(g\) it may fail to be a well-defined function and rather has to be intended as a point-to-set mapping \(\ffunc{\prox_\psi^V}{\R^r}{\R^r}\).
Nevertheless, the value function associated to the minimization problem in the definition \eqref{eq:prox}, namely the \emph{Moreau envelope}
\begin{equation}\label{eq:Moreau}
\psi^V(x)
{}\coloneqq{}
\min_{w\in\R^r}\set{
\psi(w)
{}+{}
\tfrac12\|w-x\|^2_V
},
\end{equation}
is a well-defined real-valued function, in fact locally Lipschitz continuous, that lower bounds \(\psi\) and shares with \(\psi\) infima and minimizers.
The proximal mapping is available in closed form for many useful functions, many of which are widely used regularizers in machine learning; for instance, the proximal mapping of the \(\ell^0\) and \(\ell^1\) regularizers amount to hard and soft thresholding operators.
In many applications the cost to be minimized is structured as the sum of a smooth term \(h\) and a proximable (\ie with easily computable proximal mapping) term \(\psi\).
In these cases, the \emph{proximal gradient method} \cite{fukushima1981generalized,attouch2013convergence} constitutes a cornerstone iterative method that interleaves gradient descent steps on the smooth function and proximal operations on the nonsmooth function, resulting in iterations of the form
\(
x^+
{}\in{}
\prox_{\gamma\psi}(x-\gamma\nabla h(x))
\)
for some suitable stepsize $\gamma$.
Our proposed scheme to address problem \eqref{eq:P} is a BC variant of the proximal gradient method, in the sense that only some coordinates are updated according to the proximal gradient rule, while the others are left unchanged.
This concept is synopsized in \Cref{alg:BC}, which constitutes the general algorithm addressed in this paper.
\begin{algorithm}
\caption{General forward-backward block-coordinate scheme}
\label{alg:BC}
\begin{algorithmic}[1]
\Require
\(\bm x^0\in\R^{\sum_in_i}\),~
\(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\),
{\small \(i\in[N]\)}
\Statex
\(
\Gamma=\blockdiag(\gamma_1\I_{n_1},\dots,\gamma_N\I_{n_N})
\),~
\(k=0\)
\item[{\sc Repeat} until convergence]
\State
\(
\bm z^k
{}\in{}
\prox_G^{\Gamma^{-1}}\bigl(
\bm x^k-\Gamma\nabla F(\bm x^k)
\bigr)
\)
\State\label{state:BC:sampling
select a set of indices \(I^{k+1}\subseteq[N]\)
\State
update~~
\(x_i^{k+1}= z_i^k\)
~for \(i\in I^{k+1}\)
~~and~~
\(x_i^{k+1}= x_i^k\)
~for \(i\notin I^{k+1}\),~
\(k\gets k+1\)
\item[{\sc Return}]
\(\bm z^k\)
\end{algorithmic}
\end{algorithm}
Although seemingly wasteful, in many cases one can efficiently compute individual blocks without the need of full operations.
In fact BC \Cref{alg:BC} bridges the gap between a BC framework and a class of incremental methods where a global computation typically involving the full gradient is carried out incrementally via performing computations only for a subset of coordinates.
Two such broad applications, problems \eqref{eq:FSP} and \eqref{eq:SP}, are discussed in the dedicated \Cref{sec:Finito,sec:Sharing}, where among other things we will show that \Cref{alg:BC} leads to the well known Finito/MISO algorithm \cite{defazio2014finito,mairal2015incremental}.
\subsection{Contribution}
\begin{enumerate}
leftmargin=0pt,
labelwidth=7pt,
itemindent=\labelwidth+\labelsep,
label=\rlap{{\bf\arabic*)}}\hspace*{\labelwidth},
]
\item
To the best of our knowledge this is the first analysis of BC schemes with a nonseparable nonsmooth term and in the fully nonconvex setting.
While the original cost \(\Phi\) cannot serve as a Lyapunov function, we show that the forward-backward envelope (FBE) \cite{patrinos2013proximal,themelis2018forward} decreases surely, not only in expectation (\Cref{thm:sure}).
\item
This allows for a quite general convergence analysis for different sampling criteria.
This paper in particular covers randomized strategies (\Cref{sec:random}) where at each iteration one or more coordinates are sampled with possibly time-varying probabilities, as well as essentially cyclic (and in particular cyclic and shuffled) strategies in case the nonsmooth term is convex (\Cref{sec:cyclic}).
\item
We exploit the Kurdyka-\L ojasiewicz (KL) property to show global (as opposed to subsequential) and linear convergence when the sampling is essentially cyclic and the nonsmooth function is convex, without imposing convexity requirements on the smooth functions (\Cref{thm:cyclic:global}).
\ifaccel
\item
When \(G\) is convex and \(F\) is twice continuously differentiable, the FBE is continuously differentiable.
If, additionally, \(F\) is (strongly) convex and quadratic, then the FBE is (strongly) convex and has Lipschitz-continuous gradient.
Owing to these favorable properties, we propose a new BC Nesterov-type acceleration algorithm for minimizing the sum of a block-separable convex quadratic plus a nonsmooth convex function, whose analysis directly follows from existing work on smooth BC minimization \cite{allen2016even}.
\fi
\item
As immediate byproducts of our analysis we obtain
{\bf (a)} an incremental algorithm for the sharing problem \cite{boyd2011distributed} that to the best of our knowledge is novel (\Cref{sec:Sharing}), and
{\bf (b)} the Finito/MISO algorithm \cite{defazio2014finito,mairal2015incremental} leading to a much simpler and more general analysis than available in the literature with new convergence results both for randomized sampling strategies in the fully nonconvex setting and for essentially cyclic samplings when the nonsmooth term is convex (\Cref{sec:Finito}).
\end{enumerate}
\subsection{Organization}
The rest of the paper is organized as follows.
The core of the paper lies in the convergence analysis of \Cref{alg:BC} detailed in \Cref{sec:convergence}: \Cref{sec:FBE} introduces the FBE, fundamental tool of our methodology and lists some of its properties whose proofs are detailed in the dedicated \Cref{sec:proofs:FBE}, followed by other ancillary results documented in \Cref{sec:auxiliary}.
The algorithmic analysis begins in \Cref{sec:sure} with a collection of facts that hold independently of the chosen sampling strategy, and later specializes to randomized and essentially cyclic samplings in the dedicated \Cref{sec:random,sec:cyclic}.
\Cref{sec:Finito,sec:Sharing} discuss two particular instances of the investigated algorithmic framework, namely (a generalization of) the Finito/MISO algorithm for finite sum minimization and an incremental scheme for the sharing problem, both for fully nonconvex and nonsmooth formulations.
Convergence results are immediately inferred from those of the more general BC \Cref{alg:BC}.
\Cref{sec:Conclusions} concludes the paper.
\section{Convergence analysis}\label{sec:convergence}
We begin by observing that \Cref{ass:basic} is enough to guarantee the well definedness of the forward-backward operator in \Cref{alg:BC}, which for notational convenience will be henceforth denoted as \(\T(\bm x)\).
Namely, \(\ffunc{\T}{\R^{\sum_in_i}}{\R^{\sum_in_i}}\) is the point-to-set mapping
\begin{align*}
\T(\bm x)
{}\coloneqq{} &
\prox_G^{\Gamma^{-1}}\left(\Fw{\bm x}\right)
\\
\numberthis\label{eq:T}
{}={} &
\argmin_{\bm w\in\R^{\sum_in_i}}\set{
F(\bm x)+\innprod{\nabla F(\bm x)}{\bm w-\bm x}
{}+{}
G(\bm w)
{}+{}
\tfrac12\|\bm w-\bm x\|_{\Gamma^{-1}}^2
}.
\end{align*}
\begin{lem}\label{thm:osc
Suppose that \Cref{ass:basic} holds, and let \(\Gamma\coloneqq\blockdiag(\gamma_1\I_{n_1},\dots,\gamma_N\I_{n_N})\) with \(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\), \(i\in[N]\).
Then \(\prox_G^{\Gamma^{-1}}\) and \(\T\) are locally bounded, outer semicontinuous (osc), nonempty- and compact-valued mappings.
\begin{proof}
See \Cref{proof:thm:osc}.
\end{proof}
\end{lem}
\subsection{The forward-backward envelope}\label{sec:FBE}
The fundamental challenge in the analysis of \eqref{eq:P} is the fact that, without separability of \(G\), descent on the cost function cannot be established even in expectation.
Instead, we show that the \emph{forward-backward envelope} (FBE) \cite{patrinos2013proximal,themelis2018forward} can be used as Lyapunov function.
This subsection formally introduces the FBE, here generalized to account for a matrix-valued stepsize parameter \(\Gamma\), and lists some of its basic properties needed for the convergence analysis of \Cref{alg:BC}.
Although easy adaptations of the similar results in \cite{patrinos2013proximal,themelis2018forward,themelis2019acceleration}, for the sake of self-inclusiveness the proofs are detailed in the dedicated \Cref{sec:proofs:FBE}.
\begin{subequations}
\begin{defin}[forward-backward envelope]\label{def:FBE
In problem \eqref{eq:P}, let \(f_i\) be differentiable functions, \(i\in[N]\), and for \(\gamma_1,\dots,\gamma_N>0\) let
\(
\Gamma=\blockdiag(\gamma_1\I_{n_1},\dots,\gamma_N\I_{n_N})
\).
The forward-backward envelope (FBE) associated to \eqref{eq:P} with stepsize \(\Gamma\) is the function
\(
\func{\FBE}{\R^{\sum_in_i}}{[-\infty,\infty)}
\)
defined as
\begin{equation}
\label{eq:FBE}
\FBE(\bm x)
{}\coloneqq{}
\inf_{\bm w\in\R^{\sum_in_i}}\set{
F(\bm x)+\innprod{\nabla F(\bm x)}{\bm w-\bm x}
{}+{}
G(\bm w)
{}+{}
\tfrac12\|\bm w-\bm x\|_{\Gamma^{-1}}^2
}.
\end{equation}
\end{defin}
\Cref{def:FBE} highlights an important symmetry between the Moreau envelope and the FBE: similarly to the relation between the Moreau envelope \eqref{eq:Moreau} and the proximal mapping \eqref{eq:prox}, the FBE \eqref{eq:FBE} is the value function associated with the proximal gradient mapping \eqref{eq:T}.
By replacing any minimizer \(\bm z\in\T(\bm x)\) in the right-hand side of \eqref{eq:FBE} one obtains yet another interesting interpretation of the FBE in terms of the \(\Gamma^{-1}\)-augmented Lagrangian associated to \eqref{eq:P}
\begin{align}
\nonumber
\LL(\bm x,\bm z,\bm y)
{}\coloneqq{} &
F(\bm x)+G(\bm z)+\innprod{\bm y}{\bm x-\bm z}
{}+{}
\tfrac12\|\bm x-\bm z\|_{\Gamma^{-1}}^2,
\shortintertext{namely,}
\label{eq:FBEz}
\FBE(\bm x)
{}={} &
F(\bm x)+\innprod{\nabla F(\bm x)}{\bm z-\bm x}
{}+{}
G(\bm z)
{}+{}
\tfrac12\|\bm z-\bm x\|_{\Gamma^{-1}}^2
\\
{}={} &
\LL(\bm x,\bm z,-\nabla F(\bm x)).
\shortintertext
Lastly, by rearranging the terms it can easily be seen that
}
\label{eq:FBEMoreau}
\FBE(\bm x)
{}={} &
F(\bm x)
{}-{}
\tfrac12\|\nabla F(\bm x)\|_\Gamma^2
{}+{}
G^{\Gamma^{-1}}(\Fw{\bm x}),
\end{align}
hence in particular that the FBE inherits regularity properties of \(G^{\Gamma^{-1}}\) and \(\nabla F\), some of which are summarized in the next result.
\end{subequations}
\begin{lem}[FBE: fundamental inequalities]\label{thm:FBEineq
Suppose that \Cref{ass:basic} is satisfied and let \(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\), \(i\in[N]\).
Then, the FBE \(\FBE\) is a (real-valued and) locally Lipschitz-continuous function.
Moreover, the following hold for any \(\bm x\in\R^{\sum_in_i}\):
\begin{enumerate}
\item\label{thm:leq}
\(\FBE(\bm x)\leq\Phi(\bm x)\).
\item\label{thm:geq}
\(
\tfrac12\|\bm z-\bm x\|^2_{\Gamma^{-1}-\Lambda_F}
{}\leq{}
\FBE(\bm x)-\Phi(\bm z)
{}\leq{}
\tfrac12\|\bm z-\bm x\|^2_{\Gamma^{-1}+\Lambda_F}
\)
for any \(\bm z\in\T(\bm x)\), where
\(
\Lambda_F
{}\coloneqq{}
\tfrac1N
\blockdiag\bigl(L_{f_1}\I_{n_1},\dots, L_{f_n}\I_{n_N}\bigr)
\).
\item\label{thm:strconcost}
If in addition each $f_i$ is $\mu_{f_i}$-strongly convex and $G$ is convex, then for every \(\bm x\in\R^{\sum_in_i}\)
\[
\tfrac12\|\bm z-\bm x^\star\|_{\mu_F}^2
{}\leq{}
\FBE(\bm x)-\min\Phi
\]
where \(\bm x^\star\coloneqq\argmin\Phi\),
\(
\mu_F
{}\coloneqq{}
\frac1N\blockdiag\bigl(\mu_{f_1}\I_{n_1},\dots,\mu_{f_N}\I_{n_N}\bigr)
\),
and \(\bm z=\T(\bm x)\).
\end{enumerate}
\begin{proof}
See \Cref{proof:thm:FBEineq}.
\end{proof}
\end{lem}
Another key property that the FBE shares with the Moreau envelope is that minimizing the extended-real valued function \(\Phi\) is equivalent to minimizing the continuous function \(\FBE\).
Moreover, the former is level bounded iff so is the latter.
This fact will be particularly useful for the analysis of \Cref{alg:BC}, as it will be shown in \Cref{thm:sure} that the FBE (surely) decreases along its iterates.
As a consequence, despite the fact that the same does not hold for \(\Phi\) (in fact, iterates may even be infeasible), coercivity of \(\Phi\) is enough to guarantee boundedness of \(\seq{\bm x^k}\) and \(\seq{\bm z^k}\).
\begin{lem}[FBE: minimization equivalence]\label{thm:FBEmin
Suppose that \Cref{ass:basic} is satisfied and that \(\gamma_i\in(0,\nicefrac{N}{L_i})\), \(i\in[N]\).
Then the following hold:
\begin{enumerate}
\item\label{thm:min}
\(\min\FBE=\min\Phi\);
\item\label{thm:argmin}
\(\argmin\FBE=\argmin\Phi\);
\item\label{thm:LB}
\(\FBE\) is level bounded iff so is \(\Phi\).
\end{enumerate}
\begin{proof}
See \Cref{proof:thm:FBEmin}.
\end{proof}
\end{lem}
We remark that the kinship of \(\FBE\) and \(\Phi\) extends also to local minimality; the interested reader is referred to \cite[Th. 3.6]{themelis2018proximal} for details.
\subsection{A sure descent lemma}\label{sec:sure}
We now proceed to the theoretical analysis of \Cref{alg:BC}.
Clearly, some assumptions on the index selection criterion are needed in order to establish reasonable convergence results, for little can be guaranteed if, for instance, one of the indices is never selected.
Nevertheless, for the sake of a general analysis it is instrumental to first investigate which properties hold independently of such criteria.
After listing some of these facts in \Cref{thm:sure}, in \Cref{sec:random,sec:cyclic} we will specialize the results to randomized and (essentially) cyclic sampling strategies.
\begin{lem}[sure descent]\label{thm:sure
Suppose that \Cref{ass:basic} is satisfied.
Then, the following hold for the iterates generated by \Cref{alg:BC}:
\begin{enumerate}
\item\label{thm:Igeq}
\(
\FBE(\bm x^{k+1})
{}\leq{}
\FBE(\bm x^k)
{}-{}
\sum_{i\in I^{k+1}}\tfrac{\xi_i}{2\gamma_i}\|z_i^k-x_i^k\|^2
\),
where \(\xi_i\coloneqq\frac{N-\gamma_iL_{f_i}}{N}\), \(i\in[N]\), are strictly positive;
\item\label{thm:decrease
\(\seq{\FBE(\bm x^k)}\) monotonically decreases to a finite value \(\Phi_\star\geq\min\Phi\);
\item\label{thm:omega
\(\FBE\) is constant (and equals \(\Phi_\star\) as above) on the set of accumulation points of \(\seq{\bm x^k}\);
\item\label{thm:xdiff
the sequence \(\seq{\|\bm x^{k+1}-\bm x^k\|^2}\) has finite sum (and in particular vanishes);
\item\label{thm:bounded
if \(\Phi\) is coercive, then \(\seq{\bm x^k}\) and \(\seq{\bm z^k}\) are bounded.
\end{enumerate}
\begin{proof}
\begin{proofitemize}
\item\ref{thm:Igeq}~
To ease notation, let
\(
\Lambda_F
{}\coloneqq{} \tfrac1N
\blockdiag\bigl(L_{f_1}\I_{n_1},\dots, L_{f_n}\I_{n_N}\bigr)
\)
and for \(\bm w\in\R^{\sum_in_i}\) let \(w_I\in\R^{\sum_{i\in I}n_i}\) denote the slice \((w_i)_{i\in I}\), and let \(\Lambda_{F_I},\Gamma_I\in\R^{\sum_{i\in I}n_i\times\sum_{i\in I}n_i}\) be defined accordingly.
Start by observing that, since \(\bm z^{k+1}\in\prox_G^{\Gamma^{-1}}(\Fw{\bm x^{k+1}})\), from the proximal inequality on $G$ it follows that
\begin{align*}
G(\bm z^{k+1})-G(\bm z^k)
{}\leq{} &
\tfrac12\|\bm z^k-\bm x^{k+1}+\Gamma\nabla F(\bm x^{k+1})\|_{\Gamma^{-1}}^2
{}-{}
\tfrac12\|\bm z^{k+1}-\bm x^{k+1}+\Gamma\nabla F(\bm x^{k+1})\|_{\Gamma^{-1}}^2
\\
={} &
\numberthis\label{eq:proxIneqFBS}
\tfrac12\|\bm z^k-\bm x^{k+1}\|_{\Gamma^{-1}}^2
{}-{}
\tfrac12\|\bm z^{k+1}-\bm x^{k+1}\|_{\Gamma^{-1}}^2
{}+{}
\innprod{\nabla F(\bm x^{k+1})}{\bm z^k-\bm z^{k+1}}.
\end{align*}
We have
\ifarxiv\else
\bgroup\mathtight[0.5
\fi
\begin{align*}
\FBE(\bm x^{k+1})-\FBE(\bm x^k)
{}={} &
{\red
F(\bm x^{k+1})
}
{}+{}
\innprod{\nabla F(\bm x^{k+1})}{\bm z^{k+1}-\bm x^{k+1}}
{\blue
{}+{}
G(\bm z^{k+1})
}
{}+{}
\tfrac12\|\bm z^{k+1}-\bm x^{k+1}\|_{\Gamma^{-1}}^2
\\
&
{}-{}
\left(
{\red
F(\bm x^k)+\innprod{\nabla F(\bm x^k)}{\bm z^k-\bm x^k}
}
{\blue
{}+{}
G(\bm z^k)
}
{}+{}
\tfrac12\|\bm z^k-\bm x^k\|_{\Gamma^{-1}}^2
\right)
\shortintertext
{\red
apply the upper bound in \eqref{eq:Lip} with \(\bm w=\bm x^{k+1}\)
and
{\blue
the proximal inequality \eqref{eq:proxIneqFBS}
}
{}\leq{} &
{\red
\innprod{\nabla F(\bm x^k)}{\bm x^{k+1}-\bm z^k}
{}+{}
\tfrac12\|\bm x^{k+1}-\bm x^k\|_{\Lambda_F}^2
}
{}+{}
\innprod{\nabla F(\bm x^{k+1})}{{\blue\bm z^k}-\bm x^{k+1}}
\\
&
{}-{}
\tfrac12\|\bm z^k-\bm x^k\|_{\Gamma^{-1}}^2
{\blue
{}+{}
\tfrac12\|\bm z^k-\bm x^{k+1}\|_{\Gamma^{-1}}^2
}.
\end{align*}
\ifarxiv\else
\egroup
\fi
To conclude, notice that the \(\ell\)-th block of \(\nabla F(\bm x^k)-\nabla F(\bm x^{k+1})\) is zero for \(\ell\notin I\), and that the \(\ell\)-th block of \(\bm x^{k+1}-\bm z^k\) is zero if \(\ell\in I\).
Hence, the scalar product vanishes.
For similar reasons, one has
\(
\|
\bm z^k-\bm x^{k+1}
\|^2_{\Gamma^{-1}}
{}-{}
\|\bm z^k-\bm x^k\|_{\Gamma^{-1}}^2
{}={}
{}-{}
\|z_I^k-x_I^k\|_{\Gamma_I^{-1}}^2
\)
and
\(
\|\bm x^{k+1}-\bm x^k\|_{\Lambda_F}^2
{}={}
\|z_I^k-x_I^k\|_{\Lambda_{F_I}}^2
\),
yielding the claimed expression.
\item\ref{thm:decrease}~
Monotonic decrease of \(\seq{\FBE(\bm x^k)}\) is a direct consequence of assert \ref{thm:Igeq}.
This ensures that the sequence converges to some value \(\Phi_\star\), bounded below by \(\min\Phi\) in light of \Cref{thm:min}.
\item\ref{thm:omega}~
Directly follows from assert \ref{thm:decrease} together with the continuity of \(\FBE\), see \Cref{thm:FBEineq}.
\item\ref{thm:xdiff}~
Denoting
\(
\xi_{\rm min}
{}\coloneqq{}
\min_{i\in[N]}\set{
\xi_i
}
\)
which is a strictly positive constant, it follows from assert \ref{thm:Igeq} that for each \(k\in\N\) it holds that
\begin{align*}
\FBE(\bm x^{k+1})-\FBE(\bm x^k)
{}\leq{} &
{}-{}
\sum_{\mathclap{i\in I^{k+1}}}{
\tfrac{\xi_i}{2\gamma_i}\|z_i^k-x_i^k\|^2
}
\\
{}\leq{} &
{}-{}
\tfrac{\xi_{\rm min}}{2}
\sum_{i\in I^{k+1}}{
\gamma_i^{-1}\|z_i^k-x_i^k\|^2
}
\\
\numberthis\label{eq:SDx}
{}={} &
{}-{}
\tfrac{\xi_{\rm min}}{2}
\|\bm x^{k+1}-\bm x^k\|_{\Gamma^{-1}}^2.
\end{align*}
By summing for \(k\in\N\) and using the positive definiteness of \(\Gamma^{-1}\) together with the fact that \(\min\FBE=\min\Phi>\infty\) as ensured by \Cref{thm:min} and \Cref{ass:phi}, we obtain that
\(
\sum_{k\in\N}\|\bm x^{k+1}-\bm x^k\|^2
{}<{}
\infty
\).
\item\ref{thm:bounded}~
It follows from assert \ref{thm:decrease} that the entire sequence \(\seq{\bm x^k}\) is contained in the sublevel set \(\set{\bm w}[\FBE(\bm w)\leq\FBE(\bm x^0)]\), which is bounded provided that \(\Phi\) is coercive as shown in \Cref{thm:LB}.
In turn, boundedness of \(\seq{\bm z^k}\) then follows from local boundedness of \(\T\), cf. \Cref{thm:osc}.
\qedhere
\end{proofitemize}
\end{proof}
\end{lem}
\subsection{Randomized sampling}\label{sec:random}
In this section we provide convergence results for \Cref{alg:BC} where the index selection criterion complies with the following requirement.
\begin{ass}[randomized sampling requirements]\label{ass:random
There exist \(p_1,\dots,p_N>0\) such that, at any iteration and independently of the past, each $i\in[N]$ is sampled with probability at least $p_i$.
\end{ass}
Our notion of randomization is general enough to allow for time-varying probabilities and mini-batch selections.
The role of parameters \(p_i\) in \Cref{ass:random} is to prevent that an index is sampled with arbitrarily small probability.
In more rigorous terms,
\(
\P*{i\in I^{k+1}}
{}\geq{}
p_i
\)
shall hold for all \(i\in[N]\), where \(\P{}\) represents the probability conditional to the knowledge at iteration \(k\).
Notice that we do not require the \(p_i\)'s to sum up to one, as multiple index selections are allowed, similar to the setting of \cite{bianchi2016coordinate,latafat2019new} in the convex case.
Due to the possible nonconvexity of problem \eqref{eq:P}, unless additional assumptions are made not much can be said about convergence of the iterates to a unique point.
Nevertheless, the following result shows that any accumulation point \(\bm x^\star\) of sequences $\seq{\bm x^k}$ and $\seq{\bm z^k}$ generated by \Cref{alg:BC} is a stationary point, in the sense that it satisfies the necessary condition for minimality
\(
0\in\hat\partial\Phi(\bm x^\star)
\),
where \(\hat\partial\) denotes the (regular) nonconvex subdifferential, see \cite[Th. 10.1]{rockafellar2011variational}.
\begin{thm}[randomized sampling: subsequential convergence]\label{thm:random:subseq
Suppose that \Cref{ass:basic,ass:random} are satisfied.
Then, the following hold almost surely for the iterates generated by \Cref{alg:BC}:
\begin{enumerate}
\item\label{thm:res
the sequence \(\seq{\|\bm x^k-\bm z^k\|^2}\) has finite sum (and in particular vanishes);
\item\label{thm:decreasez
the sequence \(\seq{\Phi(\bm z^k)}\) converges to \(\Phi_\star\) as in \Cref{thm:decrease};
\item\label{thm:cluster
\(\seq{\bm x^k}\) and \(\seq{\bm z^k}\) have same cluster points, all stationary and on which \(\Phi\) and \(\FBE\) equal \(\Phi_\star\)
\end{enumerate}
\begin{proof}
In what follows, \(\E{}\) denotes the expectation conditional to the knowledge at iteration \(k\).
\begin{proofitemize}
\item\ref{thm:res}~
Let
\(
\xi_i\coloneqq\frac{N-\gamma_iL_{f_i}}{N}>0
\),
\(i\in[N]\), be as in \Cref{thm:Igeq}.
We have
\begin{align*}
\E{\FBE(\bm x^{k+1})}
{}\overrel*[\leq]{\ref{thm:Igeq}}{} &
\E{
\FBE(\bm x^k)
{}-{}
\sum_{i\in I^{k+1}}{
\tfrac{\xi_i}{2\gamma_i}\|z_i^k-x_i^k\|^2
}
}
\\
{}={} &
\FBE(\bm x^k)
{}-{}
\sum_{I\in\Omega}{
\P{\mathcal I^{k+1}=I}
\sum_{i\in I}{
\tfrac{\xi_i}{2\gamma_i}\|z_i^k-x_i^k\|^2
}
}
\\
{}={} &
\FBE(\bm x^k)
{}-{}
\sum_{i=1}^N{
\sum_{I\in\Omega,I\ni i}{
\P{\mathcal I^{k+1}=I}
\tfrac{\xi_i}{2\gamma_i}\|z_i^k-x_i^k\|^2
}
}
\\
\numberthis\label{eq:EFBE+}
{}\leq{} &
\FBE(\bm x^k)
{}-{}
\sum_{i=1}^N{
\tfrac{p_i\xi_i}{2\gamma_i}\|z_i^k-x_i^k\|^2
},
\end{align*}
where $\Omega\subseteq 2^{[N]}$ is the sample space ($2^{[N]}$ denotes the power set of $[N]$).
Therefore,
\begin{equation}\label{eq:ExSD}
\E{\FBE(\bm x^{k+1})}
{}\leq{}
\FBE(\bm x^k)
{}-{}
\tfrac\sigma2
\|\bm x^k-\bm z^k\|_{\Gamma^{-1}}^2
\quad\text{where }
\sigma
{}\coloneqq{}
\min_{i=1\dots N}{
p_i\xi_i
}
{}>{}
0.
\end{equation}
The claim follows from the Robbins-Siegmund supermartingale theorem, see \eg \cite{robbins1985convergence} or \cite[Prop. 2]{bertsekas2011incremental}.
\item\ref{thm:decreasez}~
Observe that
\(
\FBE(\bm x^k)-\|\bm z^k-\bm x^k\|^2_{\Gamma^{-1}+\Lambda_F}
{}\leq{}
\Phi(\bm z^k)
{}\leq{}
\FBE(\bm x^k)-\|\bm z^k-\bm x^k\|^2_{\Gamma^{-1}-\Lambda_F}
\)
holds (surely) for \(k\in\N\) in light of \Cref{thm:geq}.
The claim then follows by invoking \Cref{thm:decrease} and assert \ref{thm:res}.
\item\ref{thm:cluster}~
In the rest of the proof, for conciseness the ``almost sure'' nature of the results will be implied without mention.
It follows from assert \ref{thm:res} that a subsequence \(\seq{\bm x^k}[k\in K]\) converges to some point \(\bm x^\star\) iff so does the subsequence \(\seq{\bm z^k}[k\in K]\).
Since \(\T(\bm x^k)\ni\bm z^k\) and both \(\bm x^k\) and \(\bm z^k\) converge to \(\bm x^\star\) as \(K\ni k\to\infty\), the inclusion \(0\in\hat\partial\Phi(\bm x^\star)\) follows from \Cref{thm:critical}.
Since the full sequences \(\seq{\FBE(\bm x^k)}\) and \(\seq{\Phi(\bm z^k)}\) converge to the same value \(\Phi_\star\) (cf. \Cref{thm:decrease} and assert \ref{thm:decreasez}), due to continuity of \(\FBE\) (\Cref{thm:FBEineq}) it holds that \(\FBE(\bm x^\star)=\Phi_\star\), and in turn the bounds in \Cref{thm:geq} together with assert \ref{thm:res} ensure that \(\Phi(\bm x^\star)=\Phi_\star\) too.
\qedhere
\end{proofitemize}
\end{proof}
\end{thm}
When \(G\) is convex and \(F\) is strongly convex (that is, each of the functions \(f_i\) is strongly convex), the FBE decreases \(Q\)-linearly in expectation along the iterates generated by the randomized BC-\Cref{alg:BC}.
\begin{thm}[randomized sampling: linear convergence under strong convexity]\label{thm:random:linear
Additionally to \Cref{ass:basic,ass:random}, suppose that \(G\) is convex and that each \(f_i\) is \(\mu_{f_i}\)-strongly convex.
Then, for all \(k\) the following hold for the iterates generated by \Cref{alg:BC}:
\begin{subequations}\label{subeq:random:linear}
\begin{align}
\label{eq:random:Qlinear}
\E{\FBE(\bm x^{k+1})-\min\Phi}
{}\leq{} &
(1-c)
\bigl(\FBE(\bm x^k)-\min\Phi\bigr)
\\
\E[]{\Phi(\bm z^k)-\min\Phi}
{}\leq{} &
\bigl(\Phi(\bm x^0)-\min\Phi\bigr)(1-c)^k
\\
\tfrac12\E[]{\|\bm z^k-\bm x^\star\|^2_{\mu_F}}
{}\leq{} &
\bigl(\Phi(\bm x^0)-\min\Phi\bigr)(1-c)^k
\end{align}
\end{subequations}
where \(\bm x^\star\coloneqq\argmin\Phi\),
\(
\mu_F
{}\coloneqq{} \tfrac1N
\blockdiag\bigl(\mu_{f_1}\I_{n_1},\dots\mu_{f_n}\I_{n_N}\bigr)
\),
and denoting \(\xi_i=\frac{N-\gamma_iL_{f_i}}{N}\), \(i\in[N]\),
\begin{equation}\label{eq:cwc}
c
{}={}
\min_{i\in[N]}{
\set{\tfrac{\xi_ip_i}{\gamma_i}}
{}\bigg/{}
\max_{i\in[N]}\set{\tfrac{N-\gamma_i\mu_{f_i}}{\gamma_i^2\mu_{f_i}}}
}.
\end{equation}
Moreover, by setting the stepsizes \(\gamma_i\) and minimum sampling probabilities \(p_i\) as
\begin{equation}\label{eq:gammaLinear}
\gamma_i
{}={}
\tfrac{N}{\mu_{f_i}}
\left(1-\sqrt{1-1/\kappa_{i}}\right)
\quad\text{and}\quad
p_i
{}={}
\frac{
\left(\sqrt{\kappa_i}+\sqrt{\kappa_i-1}\right)^2
}{
\sum_{j=1}^N\left(\sqrt{\kappa_j}+\sqrt{\kappa_j-1}\right)^2
}
\end{equation}
with
\(
\kappa_i\coloneqq\frac{L_{f_i}}{\mu_{f_i}}
\),
\(i\in[N]\), then the constant \(c\) in \eqref{subeq:random:linear} can be tightened to
\begin{equation}\label{eq:cbc}
c
{}={}
\tfrac{1}{
\sum_{i=1}^N\left(
\sqrt{\kappa_i}+\sqrt{\kappa_i-1}
\right)^2
}.
\end{equation}
\begin{proof}
Since \(\bm z^k\) is a minimizer in \eqref{eq:FBE}, the necessary stationarity condition reads
\(
\Gamma^{-1}(\bm x^k-\bm z^k)-\nabla F(\bm x^k)
{}\in{}
\partial G(\bm z^k)
\).
Convexity of \(G\) then implies
\[
G(\bm x^\star)
{}\geq{}
G(\bm z^k)
{}+{}
\innprod{\Gamma^{-1}(\bm x^k-\bm z^k)-\nabla F(\bm x^k)}{\bm x^\star-\bm z^k},
\]
whereas from strong convexity of \(F\) we have
\[
F(\bm x^\star)
{}\geq{}
F(\bm x^k)
{}+{}
\innprod{\nabla F(\bm x^k)}{\bm x^\star-\bm x^k}
{}+{}
\tfrac12\|\bm x^k-\bm x^\star\|^2_{\mu_F}.
\]
By combining these inequalities into \eqref{eq:FBEz}, and denoting \(\Phi_\star\coloneqq\min\Phi=\min\FBE\) (cf. \Cref{thm:min}), we have
\begin{align*}
\FBE(\bm x^k)-\Phi_\star
{}\leq{} &
\tfrac12\|\bm z^k-\bm x^k\|^2_{\Gamma^{-1}}
{}-{}
\tfrac12\|\bm x^\star-\bm x^k\|^2_{\mu_F}
{}+{}
\innprod{\Gamma^{-1}(\bm z^k-\bm x^k)}{\bm x^\star-\bm z^k}
\\
{}={} &
\tfrac12\|\bm z^k-\bm x^k\|_{\Gamma^{-1}-\mu_F}^2
{}+{}
\innprod{(\Gamma^{-1}-\mu_F)(\bm z^k-\bm x^k)}{\bm x^\star-\bm z^k}
{}-{}
\tfrac12\|\bm x^\star-\bm z^k\|_{\mu_F}^2.
\end{align*}
Next, by using the inequality
\(
\innprod{\bm a}{\bm b}
{}\leq{}
\tfrac12\|\bm a\|_{\mu_F}^2
{}+{}
\tfrac12\|\bm b\|^2_{\mu_F^{-1}}
\)
to cancel out the last term, we obtain
\begin{align*}
\FBE(\bm x^k)-\Phi_\star
{}\leq{} &
\tfrac12\|\bm z^k-\bm x^k\|_{\Gamma^{-1}-\mu_F}^2
{}+{}
\tfrac12\|(\Gamma^{-1}-\mu_F)(\bm x^k-\bm z^k)\|_{\mu_F^{-1}}^2
\\
{}={} &
\tfrac12\|\bm z^k-\bm x^k\|_{\Gamma^{-2}\mu_F^{-1}(\I-\Gamma\mu_F)}^2,
\numberthis\label{eq:QUB}
\end{align*}
where the last identity uses the fact that the matrices are diagonal.
Combined with \eqref{eq:EFBE+} the claimed \(Q\)-linear convergence \eqref{eq:random:Qlinear} with factor \(c\) as in \eqref{eq:cwc} is obtained.
The $R$-linear rates in terms of the cost function and distance from the solution are obtained by repeated application of \eqref{eq:random:Qlinear} after taking (unconditional) expectation from both sides and using \Cref{thm:FBEineq}.
To obtain the tighter estimate \eqref{eq:cbc}, observe that \eqref{eq:EFBE+} with the choice
\[
\textstyle
p_i
{}\coloneqq{}
\tfrac{1}{\gamma_i\mu_{f_i}}
\tfrac{N-\gamma_i\mu_{f_i}}{N-\gamma_iL_{f_i}}
\left(
\sum_j{
\tfrac{1}{\gamma_j\mu_{f_j}}
\tfrac{N-\gamma_j\mu_{f_j}}{N-\gamma_jL_{f_j}}
}
\right)^{-1},
\]
which equals the one in \eqref{eq:gammaLinear} with $\gamma_i$ as prescribed, yields
\begin{align*}
\E{\FBE(\bm x^{k+1})-\Phi_\star}
{}\leq{} &
\FBE(\bm x^k)-\Phi_\star
{}-{}
\left(
\textstyle
2N\sum_j{
\tfrac{1}{\gamma_j\mu_j}
\tfrac{N-\gamma_j\mu_j}{N-\gamma_jL_j}
}
\right)^{-1}
\sum_{i=1}^N{
\tfrac{N-\gamma_i\mu_{f_i}}{\gamma_i^2\mu_{f_i}}
\|z_i^k-x_i^k\|^2
}
\\
{}={} &
\textstyle
\FBE(\bm x^k)-\Phi_\star
{}-{}
\left(
2N\sum_j{
\tfrac{1}{\gamma_j\mu_j}
\tfrac{N-\gamma_j\mu_j}{N-\gamma_jL_j}
}
\right)^{-1}
\|\bm z^k-\bm x^k\|_{\Gamma^{-1}\mu_F^{-1}(\Gamma^{-1}-\mu_F)}^2.
\end{align*}
The assert now follows by combining this with \eqref{eq:QUB} and replacing the values of \(\gamma_i\) as proposed in \eqref{eq:gammaLinear}.
\end{proof}
\end{thm}
Notice that as \(\kappa_i\)'s approach \(1\) the linear rate tends to \(1-\nicefrac1N\).
\subsection{Cyclic, shuffled and essentially cyclic samplings}\label{sec:cyclic}
In this section we analyze the convergence of the BC-\Cref{alg:BC} when a cyclic, shuffled cyclic or (more generally) an essentially cyclic sampling \cite{tseng1987relaxation,tseng2001convergence,hong2017iteration,chow2017cyclic,xu2017globally} is used.
As formalized in the following standing assumption, an additional convexity requirement for the nonsmooth term \(G\) is needed.
\begin{ass}[essentially cyclic sampling requirements]\label{ass:cyclic
In problem \eqref{eq:P}, function \(G\) is convex.
Moreover, there exists $T\geq 1$ such that in \Cref{alg:BC} each index is selected at least once within any interval of $T$ iterations.
\end{ass}
Note that having \(T<N\) is possible because of our general sampling strategy where sets of indices can be sampled within the same iteration.
For instance, \(T=1\) corresponds to \(I^{k+1}=[N]\) for all \(k\), in which case \Cref{alg:BC} would reduce to a (full) proximal gradient scheme.
Two notable special cases of single index selection rules are the cyclic and shuffled cyclic sampling strategies.
\begin{itemize}
leftmargin=*,
label={},
itemindent=0cm,
labelsep=0pt,
partopsep=0pt,
parsep=0pt,
listparindent=0pt,
topsep=0pt,
]
\item{\sc Shuffled cyclic sampling:} corresponds to setting
\begin{equation}\label{eq:ShufCyclicRule}
I^{k+1}=\set{\pi_{\lfloor\nicefrac kN\rfloor}\bigl(\mod(k,N)+1\bigr)}\quad \text{for all}\quad k\in\N,
\end{equation}
where $\pi_0,\pi_1,\dots$ are permutations of the set of indices $[N]$ (chosen randomly or deterministically).
\item{\sc Cyclic sampling:} corresponds to the case \eqref{eq:ShufCyclicRule} with $\pi_{\lfloor\nicefrac kN\rfloor}=\id$, \ie,
\begin{equation}\label{eq:cyclicRule}
I^{k+1}=\set{\mod(k,N)+1}\quad \text{for all}\quad k\in\N.
\end{equation}
\end{itemize}
We remark that in practice it has been observed that an effective sampling technique is to use random shuffling after each cycle \cite[\S2]{bertsekas2015convex}.
Consistently with the deterministic nature of the essentially cyclic sampling, all results of the previous section hold surely, as opposed to almost surely.
\begin{thm}[essentially cyclic sampling: subsequential convergence]\label{thm:cyclic:subseq
Suppose that \Cref{ass:basic,ass:cyclic} are satisfied.
Then, all the asserts of \Cref{thm:random:subseq} hold surely.
\begin{proof}
We first establish an important descent inequality for \(\FBE\) after every $T$ iterations, cf. \eqref{eq:Essential_cyclic_descent}.
Convexity of \(G\), entailing $\prox_{G}^{\Gamma^{-1}}$ being Lipschitz continuous (cf. \Cref{thm:FNE}), allows the employment of techniques similar to those in \cite[Lemma 3.3]{beck2013convergence}.
Since all indices are updated at least once every \(T\) iterations, one has that
\begin{equation}\label{eq:ki}
\ki
{}\coloneqq{}
\min\set{t\in[T]}[
\text{\(i\) is sampled at iteration \(T\nu+t-1\)}
]
\end{equation}
is well defined for each index \(i\in[N]\) and \(\nu\in\N\).
Since \(i\) is sampled at iteration \(T\nu+\ki-1\) and \(x_i^{T\nu}=x_i^{T\nu+1}=\dots=x_i^{T\nu+\ki-1}\) by definition of \(\ki\), it holds that
\begin{align*}
x_i^{T\nu+\ki}
{}={} &
x_i^{T\nu+\ki-1}
{}+{}
\trans{U_i}\,\left(
\T(\bm x^{T\nu+\ki-1})
{}-{}
\bm x^{T\nu+\ki-1}
\right)
\\
\numberthis\label{eq:equiiter_ki}
{}={} &
x_i^{T\nu}
{}+{}
\trans{U_i}\,\left(
\T(\bm x^{T\nu+\ki-1})
{}-{}
\bm x^{T\nu+\ki-1}
\right),
\end{align*}
\ifaccel\else
where $U_i\in \R^{(\sum_jn_j)\times n_i}$ denotes the $i$-th block column of the identity matrix so that for a vector $v\in \R^{n_i}$
\begin{equation}\label{eq:U}
U_iv
{}={}
\trans{(0,\dots, 0, \!\overbracket{\,v\,}^{\mathclap{i\text{-th}}}\!, 0, \dots, 0)}.
\end{equation}
\fi
For all $t\in[T]$ the following holds
\begin{align*}
\FBE(\bm x^{T(\nu+1)})
{}-{}
\FBE(\bm x^{T\nu})
{}={} &
\sum_{\tau=1}^T\left(
\FBE(\bm x^{T\nu+\tau})
{}-{}
\FBE(\bm x^{T\nu+\tau-1})
\right)
\\
{}\leq{} &
\FBE(\bm x^{T\nu+t})
{}-{}
\FBE(\bm x^{T\nu+t-1})
\\
{}\leq{} &
-\tfrac{\xi_{\rm min}}{2}
\|\bm x^{T\nu+t}-\bm x^{T\nu+t-1}\|_{\Gamma^{-1}}^2,
\numberthis\label{eq:descent_esscyc}
\end{align*}
where $\xi_i\coloneqq \tfrac{N-\gamma_{i}L_{f_{i}}}{N}$ as in \Cref{thm:Igeq}, $\xi_{\rm min}\coloneqq \min_{i\in[N]}\set{\xi_i}$, and the two inequalities follow from \Cref{thm:Igeq}.
Moreover, using triangular inequality for $i\in[N]$ yields
\begin{align*}
\|\bm x^{T\nu+\ki-1}-\bm x^{T\nu}\|_{\Gamma^{-1}}
{}\leq{} &
\sum_{\tau=1}^{\ki-1}\|\bm x^{T\nu+\tau}-\bm x^{T\nu+\tau-1}\|_{\Gamma^{-1}}
\\
\numberthis\label{eq:new25}
{}\leq{} &
\tfrac{T}{\sqrt{\xi_{\rm min}\nicefrac{}{2}}}
\left(
\FBE(\bm x^{T\nu})
{}-{}
\FBE(\bm x^{T(\nu+1)})
\right)^{\nicefrac12},
\end{align*}
where the second inequality follows from \eqref{eq:descent_esscyc} together with the fact that \(\ki\leq T\).
For all \(i\in[N]\), from the triangular inequality and the \(L_{\bf T}\)-Lipschitz continuity of \(\T\) (\Cref{thm:TLip}) we have
\begin{align*}
\gamma_i^{-\nicefrac12}
\|\trans{U_i}\,(\bm x^{T\nu}-\T(\bm x^{T\nu}))\|
{}\leq{} &
\gamma_i^{-\nicefrac12}
\|\trans{U_i}\,\bigl(\bm x^{T\nu}-\T(\bm x^{T\nu+\ki-1})\bigr)\|
\\
&
{}+{}
\gamma_i^{-\nicefrac12}
\|\trans{U_i}\,\bigl(\T(\bm x^{T\nu+\ki-1})-\T(\bm x^{T\nu})\bigr)\|
\\
{}\leq{} &
\gamma_i^{-\nicefrac12}
\|x_i^{T\nu+\ki-1}-x_i^{T\nu+\ki}\|
\\
&
{}+{}
\|\T(\bm x^{T\nu+\ki-1})-\T(\bm x^{T\nu})\|_{\Gamma^{-1}}
\\
{}\leq{} &
\|\bm x^{T\nu+\ki-1}-\bm x^{T\nu+\ki}\|_{\Gamma^{-1}}
{}+{}
L_{\bf T}\|\bm x^{T\nu+\ki-1}-\bm x^{T\nu}\|_{\Gamma^{-1}}
\\
\numberthis\label{eq:sqrtbound}
{}\overrel[\leq]{\eqref{eq:descent_esscyc},~\eqref{eq:new25}}{} &
\tfrac{1+TL_{\bf T}}{\sqrt{\xi_{\rm min}/2}}
\left(
\FBE(\bm x^{T\nu})
{}-{}
\FBE(\bm x^{T(\nu+1)})
\right)^{\nicefrac12}.
\end{align*}
By squaring and summing over \(i\in[N]\), we obtain
\begin{equation}\label{eq:Essential_cyclic_descent}
\FBE(\bm x^{T(\nu+1)})-\FBE(\bm x^{T\nu})
{}\leq{}
-\tfrac{\xi_{\rm min}}{2N(1+TL_{\bf T})^2}
\|\bm z^{T\nu}-\bm x^{T\nu}\|^{2}_{\Gamma^{-1}}.
\end{equation}
By telescoping the inequality and using the fact that \(\min\FBE=\min\Phi\)
\ifarxiv
by
\else
shown in
\fi
\Cref{thm:min}, we obtain that
\(
\seq{\|\bm z^{T\nu}-\bm x^{T\nu}\|^2_{\Gamma^{-1}}}[\nu\in\N]
\)
has finite sum, and in particular vanishes.
Clearly, by suitably shifting, for every \(t\in[T]\) the same can be said for the sequence
\(
\seq{\|\bm z^{T\nu+t}-\bm x^{T\nu+t}\|^2_{\Gamma^{-1}}}[\nu\in\N]
\).
The whole sequence
\(
\seq{\|\bm z^k-\bm x^k\|^2}
\)
is thus summable, and we may now infer the claim as done in the proof of \Cref{thm:random:subseq}.
\end{proof}
\end{thm}
In the next theorem explicit linear convergence rates are derived under the additional strong convexity assumption for the smooth functions.
The cyclic and shuffled cyclic cases are treated separately, as tighter bounds can be obtained by leveraging the fact that within cycles of \(N\) iterations every index is updated exactly once.
\begin{thm}[essentially cyclic sampling: linear convergence under strong convexity]\label{thm:cyclic:linear
Additionally to \Cref{ass:basic,ass:cyclic}, suppose that each function \(f_i\) is \(\mu_{f_i}\)-strongly convex.
Then, denoting
\(
\delta
{}\coloneqq{}
\min_{i\in[N]}\set{
\tfrac{\gamma_i\mu_{f_i}}{N}
}
\)
and
\(
\Delta
{}\coloneqq{}
\max_{i\in[N]}\set{
\tfrac{\gamma_iL_{f_i}}{N}
}
\),
for all \(\nu\in\N\) the following hold for the iterates generated by \Cref{alg:BC}:
\begin{subequations}\label{subeq:cyclic:linear}
\begin{align}
\label{eq:cyclic:Qlinear}
\FBE(\bm x^{T(\nu+1)})-\min\Phi
{}\leq{} &
(1-c)
\bigl(\FBE(\bm x^{T\nu})-\min\Phi\bigr)
\\ \label{eq:cyclic:Rlinear1}
\Phi(\bm z^{T\nu})-\min\Phi
{}\leq{} &
\bigl(\Phi(\bm x^0)-\min\Phi\bigr)(1-c)^\nu
\\ \label{eq:cyclic:Rlinear2}
\tfrac12\|\bm z^{T\nu}-\bm x^\star\|^2_{\mu_F}
{}\leq{} &
\bigl(\Phi(\bm x^0)-\min\Phi\bigr)(1-c)^\nu
\end{align}
\end{subequations}
where \(\bm x^\star\coloneqq\argmin\Phi\),
\(
\mu_F
{}\coloneqq{} \tfrac1N
\blockdiag\bigl(\mu_{f_1}\I_{n_1},\dots\mu_{f_n}\I_{n_N}\bigr)
\),
and
\begin{equation}\label{eq:cyclic:cwc}
c
{}={}
\frac{
\delta(1-\Delta)
}{
N\bigl(1+T(1-\delta)\bigr)^2
(1-\delta)
}.
\end{equation}
In the case of shuffled cyclic \eqref{eq:ShufCyclicRule} or cyclic \eqref{eq:cyclicRule} sampling, the inequalities can be tightened by
replacing \(T\) with \(N\) and with
\begin{equation}\label{eq:linearShuffledCyclic}
c
{}={}
\frac{\delta(1-\Delta)}{N\left(2-\delta\right)^{2}\left(1-\delta\right)}.
\end{equation}
\begin{comment}
the following tighter bound holds
\begin{equation}\label{eq:linearShuffledCyclic}
{\FBE(\bm x^{N(\nu+1)})-\min\Phi}
{}\leq{}
\left(1-c\right)
\left(\FBE(\bm x^{N\nu})-\min\Phi\right),
\quad \text{where} \quad
c
{}={}
\frac{\delta(1-\Delta)}{N\left(2-\delta\right)^{2}\left(1-\delta\right)}.
\end{equation}
\end{comment}
\begin{proof}
\begin{proofitemize}
\item\emph{The general essentially cyclic case.}~
Since \(\T\) is \(L_{\bf T}\)-Lipschitz continuous with \(L_{\bf T}=1-\delta\) as shown in \Cref{thm:contractive}, inequality \eqref{eq:Essential_cyclic_descent} becomes
\[
\FBE(\bm x^{T(\nu+1)})
{}-{}
\FBE(\bm x^{T\nu})
{}\leq{}
-\tfrac{1-\Delta}{2N(1+T(1-\delta))^2}
\|\bm z^{T\nu}-\bm x^{T\nu}\|^2_{\Gamma^{-1}}.
\]
Moreover, it follows from \eqref{eq:QUB} that
\begin{equation}\label{eq:strongLB}
\FBE(\bm x^{T\nu})-\Phi_\star
{}\leq{}
\tfrac12
(\delta^{-1}-1)
\|\bm z^{T\nu}-\bm x^{T\nu}\|_{\Gamma^{-1}}^2.
\end{equation}
By combining the two inequalities the claimed \(Q\)-linear convergence \eqref{eq:cyclic:Qlinear} with factor \(c\) as in \eqref{eq:cyclic:cwc} is obtained.
In turn, the $R$-linear rates \eqref{eq:cyclic:Rlinear1} and \eqref{eq:cyclic:Rlinear2} follow from \Cref{thm:FBEineq}.
\item\emph{The shuffled cyclic case.}~
Let us now suppose that the sampling strategy follows a shuffled rule as in \eqref{eq:ShufCyclicRule} with permutations \(\pi_0,\pi_1,\dots\) (hence in the cyclic case $\pi_\nu=\id$ for all $\nu\in\N$).
Let $U_i$ be as in \eqref{eq:U} and $\xi_{\rm min}$ as in the proof of \Cref{thm:cyclic:subseq}. Observe that \(\ki=\pi_\nu^{-1}(i)\leq N\) for \(\ki\) as defined in \eqref{eq:ki}.
For all $t\in[N]$
\begin{align*}
\FBE(\bm x^{N(\nu+1)}) -\FBE(\bm x^{N\nu})
{}\leq{} &
\FBE(\bm x^{N\nu+t-1}) -\FBE(\bm x^{N\nu})
\\
{}\leq{} &
-\tfrac{\xi_{\rm min}}{2}\sum_{\tau=1}^{t-1} \|\bm x^{N\nu+\tau}-\bm x^{N\nu+\tau-1}\|^2_{\Gamma^{-1}}
\\
\numberthis\label{eq:tighterNoTri}
{}={} &
-\tfrac{\xi_{\rm min}}{2}\|\bm x^{N\nu+t-1}-\bm x^{N\nu}\|^2_{\Gamma^{-1}},
\end{align*}
where the equality follows from the fact that at every iteration a different coordinate is updated (and that $\Gamma$ is diagonal), and the inequalities from \Cref{thm:Igeq}. Similarly, \eqref{eq:descent_esscyc} holds with $T$ replaced by \(N\) (despite the fact that \(T\) is not necessarily \(N\), but is rather bounded as \(T\leq 2N-1\)).
By using \eqref{eq:tighterNoTri} in place of \eqref{eq:new25}, inequality \eqref{eq:sqrtbound} is tightened as follows
\[
\gamma_i^{-\nicefrac12}
\|\trans{U_i}(\bm x^{N\nu}-\T(\bm x^{N\nu}))\|
{}\leq{}
\tfrac{1+L_{\bf T}}{\sqrt{\xi_{\rm min}/2}}
\left(
\FBE(\bm x^{N\nu})
{}-{}
\FBE(\bm x^{N(\nu+1)})
\right)^{\nicefrac12}.
\]
By squaring and summing for \(i\in[N]\) we obtain
\begin{equation}\label{eq:cyclic_descent}
\FBE(\bm x^{N(\nu+1)})-\FBE(\bm x^{N\nu})
{}\leq{}
-\tfrac{\xi_{\rm min}}{2N(1+L_{\bf T})^2}
\|\bm z^{N\nu}-\bm x^{N\nu}\|^2_{\Gamma^{-1}}
{}={}
-\tfrac{1-\Delta}{2N(1+L_{\bf T})^2}
\|\bm z^{N\nu}-\bm x^{N\nu}\|^2_{\Gamma^{-1}},
\end{equation}
where \(L_{\bf T}=1-\delta\) as discussed above.
By combining this and \eqref{eq:strongLB} (with \(T\) replaced by \(N\)) the improved coefficient \eqref{eq:linearShuffledCyclic} is obtained.
\qedhere
\end{proofitemize}
\end{proof}
\end{thm}
Note that if one sets $\gamma_i = \alpha N/L_{f_i}$ for some $\alpha\in(0,1)$, then $\delta = \alpha\min_{i\in[N]} \set{\nicefrac{\mu_{f_i}}{L_{f_i}}}$ and $\Delta=\alpha$.
With this selection, as the condition number approaches $1$ the rate in \eqref{eq:linearShuffledCyclic} tends to $1-\frac{\alpha}{N\left(2-\alpha\right)^{2}}$.
\begin{comment}
{\color{red}Similarly to the argument in the randomized case, the $R$-linear rate
\[
\|\bm x^{N\nu}-\bm x^\star\|^2_M
{}\leq{}
4(1-c)^k
\left(\FBE(\bm x^0)-\min\Phi\right)
\]
for the (shuffled) cyclic case with $M$ as in \Cref{thm:strconcost} is obtained.}
\end{comment}
\subsection{Global and linear convergence with KL inequality}
The convergence analyses of the randomized and essentially cyclic cases both rely on a descent property on the FBE that quantifies the progress in the minization of \(\FBE\) in terms of the squared forward-backward residual \(\|\bm x-\bm z\|^2\).
A subtle but important difference, however, is that the inequality \eqref{eq:ExSD} in the former case involves a conditional expectation, whereas \eqref{eq:Essential_cyclic_descent} in the latter does not.
The \emph{sure} descent property occurring for essentially cyclic sampling strategies is the key for establishing global (as opposed to subsequential) convergence based on the Kurdyka-\L ojasiewicz (KL) property \cite{lojasiewicz1963propriete,lojasiewicz1993geometrie,kurdyka1998gradients}.
A similar result is achieved in \cite{xu2017globally}, which however considers the complementary case to problem \eqref{eq:P} where the nonsmooth function \(G\) is assumed to be separable, and thus the cost function itself can serve as Lyapunov function.
\begin{defin}[KL property with exponent \(\theta\)]\label{def:KL
A proper lsc function \(\func{h}{\R^n}{\Rinf}\) is said to have the \DEF{Kurdyka-{\L}ojasiewicz} (KL) property with exponent \(\theta\in(0,1)\) at \(\bar w\in\dom h\) if there exist \(\varepsilon,\eta,\varrho>0\) such that
\[
\psi'(h(w)-h(\bar w))\dist(0,\partial h(w))\geq 1
\]
holds for all \(w\) such that \(\|w-\bar w\|<\varepsilon\) and \(h(\bar w)<h(w)<h(\bar w)+\eta\), where \(\psi(s)\coloneqq\varrho s^{1-\theta}\).
We say that \(h\) satisfies the KL property with exponent \(\theta\) (without mention of \(\bar w\)) if it satisfies the KL property with exponent \(\theta\) at any \(\bar w\in\dom\partial h\).
\end{defin}
Semialgebraic functions comprise a wide class of functions that enjoy this property \cite{bolte2007clarke,bolte2007lojasiewicz}, which has been extensively exploited to provide convergence rates of optimization algorithms \cite{attouch2009convergence,attouch2010proximal,attouch2013convergence,bolte2014proximal,frankel2015splitting,ochs2014ipiano,li2016douglas,xu2013block}.
Based on this, in the next result we provide sufficient conditions ensuring global and \(R\)-linear convergence of \Cref{alg:BC} with essentially cyclic sampling.
\begin{thm}[essentially cyclic sampling: global and linear convergence]\label{thm:cyclic:global
Additionally to \Cref{ass:basic,ass:cyclic}, suppose that \(\Phi\) has the KL property with exponent \(\theta\in(0,1)\) (as is the case when \(f_i\) and \(G\) are semialgebraic), and is coercive.
Then, any sequences \(\seq{\bm x^k}\) and \(\seq{\bm z^k}\) generated by \Cref{alg:BC} converge to (the same) stationary point \(\bm x^\star\).
Moreover, if \(\theta\leq\nicefrac12\) then \(\seq{\|\bm z^k-\bm x^k\|}\), \(\seq{\bm x^k}\) and \(\seq{\bm z^k}\) converge at $R$-linear rate.
\begin{proof}
Let \(\seq{\bm x^k}\) and \(\seq{\bm z^k}\) be sequences generated by \Cref{alg:BC} with essentially cyclic sampling, and let \(\Phi_\star\) be the limit of the sequence \(\seq{\FBE(\bm x^k)}\) as in \Cref{thm:decrease}.
To avoid trivialities, we may assume that \(\FBE(\bm x^k)\gneqq\Phi_\star\) for all \(k\), for otherwise the sequence \(\seq{\bm x^k}\) is asymptotically constant, and thus so is $\seq{\bm z^k}$.
Let \(\Omega\) be the set of accumulation points of \(\seq{\bm x^k}\), which is compact and such that \(\FBE\equiv\Phi_\star\) on \(\Omega\), as ensured by \Cref{thm:cyclic:subseq}.
It follows from \Cref{thm:loja} and \cite[Lem. 1(ii)]{attouch2009convergence} that \(\FBE\) enjoys a \emph{uniform} KL property on \(\Omega\); in particular,
\(
\psi'(\FBE(\bm x^k)-\Phi_\star)\dist(0,\partial\FBE(\bm x^k))
{}\geq{}
1
\)
holds for all \(k\) large enough such that \(\bm x^k\) is sufficiently close to \(\Omega\) and \(\FBE(\bm x^k)\) is sufficiently close to \(\Phi_\star\), where \(\psi(s)=\varrho s^{1-\theta'}\) for some \(\varrho>0\) and \(\theta'=\max\set{\theta,\nicefrac12}\).
Combined with \Cref{thm:subdiffdist}, for all \(k\) large enough we thus have
\begin{equation}\label{eq:KL}
\psi'(\FBE(\bm x^k)-\Phi_\star)
{}\geq{}
\frac{c}{\|\bm x^k-\bm z^k\|_{\Gamma^{-1}}},
\end{equation}
where
\(
c
{}\coloneqq{}
\frac{
N\min_i\set{\sqrt{\gamma_i}}
}{
N+\max_i\set{\gamma_iL_{f_i}}
}
{}>{}
0
\).
Let
\(
\Delta_k\coloneqq\psi(\FBE(\bm x^k)-\Phi_\star)
\).
By combining \eqref{eq:KL} and \eqref{eq:Essential_cyclic_descent} we have that there exists a constant \(c'>0\) such that
\begin{equation}\label{eq:KLinequality}
\Delta_{(\nu+1)T}
{}-{}
\Delta_{\nu T}
{}\leq{}
\psi'(\FBE(\bm x^{\nu T})-\Phi_\star)
\left(\FBE(\bm x^{(\nu+1)T})-\FBE(\bm x^{\nu T})\right)
{}\leq{}
-c'
\|\bm x^{\nu T}-\bm z^{\nu T}\|_{\Gamma^{-1}}
\end{equation}
holds for all $\nu\in\N$ large enough (the first inequality uses concavity of \(\psi\)).
By summing over \(\nu\) (sure) summability of the sequence \(\seq{\|\bm x^{\nu T}-\bm z^{\nu T}\|}[\nu\in\N]\) is obtained.
By suitably shifting, for every \(t\in[T]\) the same can be said for the sequence
\(
\seq{\|\bm z^{T\nu+t}-\bm x^{T\nu+t}\|}[\nu\in\N]
\),
and since \(T\) is finite we conclude that the whole sequence
\(
\seq{\|\bm z^k-\bm x^k\|}
\)
is summable.
Since \(\|\bm x^{k+1}-\bm x^k\|\leq\|\bm z^k-\bm x^k\|\) we conclude that \(\seq{\bm x^k}\) has finite length and is thus convergent (to a single point), and consequently so is \(\seq{\bm z^k}\).
\begin{comment}
Consider the \(\Gamma^{-1}\)-augmented Lagrangian defined in \eqref{eq:AugLagrangian} and let
\(
\mathcal{L}_k
{}\coloneqq{}
\LL(\bm x^k,\bm z^k,-\nabla F(\bm x^k))
\)
and similarly
\(
\partial\mathcal{L}_k
{}\coloneqq{}
\partial\LL(\bm x^k,\bm z^k,-\nabla F(\bm x^k))
\).
Note that
\(
\FBE(\bm x^k)
{}={}
\mathcal{L}_k
\);
to avoid trivialities, we may thus assume that \(\mathcal{L}_k\gneqq\Phi_\star\) for all \(k\), for otherwise the sequence \(\seq{\bm x^k}\) is asymptotically constant, cf. \eqref{eq:SDx}, and thus so is $(\bm x^k,\bm z^k,-\nabla F(\bm x^k))$.
Let \(\Omega\) be the set of accumulation points of \(\seq{\bm x^k}\), which is compact and such that \(\FBE\equiv\Phi_\star\) on \(\Omega\) for some \(\Phi_\star\in\R\), as ensured by \Cref{thm:cluster}.
Then, since \(\|\bm x^k-\bm z^k\|\to0\) and \(\nabla F\) is continuous,
\(
\Omega'
{}\coloneqq{}
\set{\bigl(\bm x,\bm x,-\nabla F(\bm x)\bigr)}[\bm x\in\Omega]
\)
is the set of cluster points of $(\bm x^k,\bm z^k,-\nabla F(\bm x^k))$, which is also
compact and on which \(\LL\) is constantly equal to \(\Phi_\star\).
Since \(F\) and \(G\) are semialgebraic, known properties of semialgebraic functions (see \eg \cite[\S8.3.1]{ioffe2017variational}) ensure that \(\LL\) is semialgebraic, and as such it possesses the KL property on \(\Omega'\), see \cite[Thm. 3 and Lem. 6]{bolte2014proximal}: there exists a continuous increasing concave function
\(
\func{\psi}{[0,\varepsilon)}{[0,\infty)}
\)
(for some \(\varepsilon>0\)) which is differentiable on \((0,\varepsilon)\) and with \(\psi(0)=0\), such that
\(
\psi'(\mathcal{L}_k-\Phi_\star)\dist(0,\partial\mathcal{L}_k)
{}\geq{}
1
\)
for all \(k\) large enough such that $(\bm x^k,\bm z^k,-\nabla F(\bm x^k))$ is sufficiently close to \(\Omega'\) and \(\mathcal{L}_k\) is sufficiently close to \(\Phi_\star\).
The optimality condition for \(\bm z\in\T(\bm x)\) reads
\begin{equation}\label{eq:proxsubgrad}
\Gamma^{-1}(\bm x-\bm z)-\nabla F(\bm x)
{}\in{}
\partial G(\bm z),
\end{equation}
hence
\(
\partial\mathcal{L}_k
{}\ni{}
\bigl(
\Gamma^{-1}(\bm x^k-\bm z^k),~
0,~
\bm x^k-\bm z^k
\bigr)
\),
which implies that
\[
\textstyle
\dist(0,\partial\mathcal{L}_k)
{}\leq{}
\sqrt{\gamma_{\rm min}^{-1}+\gamma_{\rm max}}\,
\|\bm x^k-\bm z^k\|_{\Gamma^{-1}},
\]
where \(\gamma_{\rm min}\coloneqq\min_{i\in[N]}\set{\gamma_i}\) and \(\gamma_{\rm max}\coloneqq\max_{i\in[N]}\set{\gamma_i}\).
Combined with the KL inequality, we obtain
\begin{equation}\label{eq:KL}
\textstyle
\psi'(\mathcal L_k-\Phi_\star)
{}\geq{}
\frac{1}{
\sqrt{\gamma_{\rm min}^{-1}+\gamma_{\rm max}}\,
\|\bm x^k-\bm z^k\|_{\Gamma^{-1}}
}.
\end{equation}
Denote
\(
\Delta_k\coloneqq\psi(\mathcal{L}_k-\Phi_\star)
\), \(
\sigma'
{}={}
\tfrac{\xi_{\rm min}}{N(1+TL_{\bf T})^2}
\)
and let \(\xi_{\rm min}\) and \(L_{\bf T}\) be as in the proof of \Cref{thm:cyclic:subseq}.
We have for all $\nu\in\N$
\begin{equation}\label{eq:KLinequality}
\Delta_{(\nu+1)T}
{}-{}
\Delta_{\nu T}
{}\leq{}
\psi'(\mathcal{L}_{\nu T}-\Phi_\star)
\left(\mathcal{L}_{(\nu+1)T}-\mathcal{L}_{\nu T}\right)
{}\overrel*[\leq]{\eqref{eq:Essential_cyclic_descent},\,\eqref{eq:KL}}{}
-\tfrac{\sigma'}{2\sqrt{\gamma_{\rm min}^{-1}+\gamma_{\rm max}}}
\|\bm x^{\nu T}-\bm z^{\nu T}\|_{\Gamma^{-1}},
\end{equation}
where the first inequality uses concavity of \(\psi\).
By summing over \(\nu\in\N\) (sure) summability of the sequence \(\seq{\|\bm x^{\nu T}-\bm z^{\nu T}\|}[\nu\in\N]\) is obtained.
By suitably shifting, for every \(t\in[T]\) the same can be said for the sequence
\(
\seq{\|\bm z^{T\nu+t}-\bm x^{T\nu+t}\|}[\nu\in\N]
\),
and since \(T\) is finite we conclude that the whole sequence
\(
\seq{\|\bm z^k-\bm x^k\|}
\)
is summable.
Since \(\|\bm x^{k+1}-\bm x^k\|\leq\|\bm z^k-\bm x^k\|\) we conclude that \(\seq{\bm x^k}\) has finite length and is thus convergent (to a single point), and consequently so is \(\seq{\bm z^k}\).
Suppose now that \(\theta\leq\nicefrac12\), so that \(\psi(s)=\varrho\sqrt s\).
Then,
\[
\|\bm x^{\nu T}-\bm z^{\nu T}\|_{\Gamma^{-1}}
{}\overrel[\geq]{\eqref{eq:KL}}{}
\tfrac{2c}{\varrho}
\sqrt{\FBE(\bm x^{\nu T})-\Phi_\star}
{}={}
\tfrac{2c}{\varrho^2}
\psi(\FBE(\bm x^{\nu T})-\Phi_\star)
{}={}
\tfrac{2c}{\varrho^2}
\Delta_{\nu T}.
\]
Combined with \eqref{eq:KLinequality} it follows that $\seq{\Delta_{\nu T}}[\nu\in\N]$ conveges $Q$-linearly.
By rearranging \eqref{eq:KLinequality} as
\[
c'\|\bm x^{\nu T}-\bm z^{\nu T}\|_{\Gamma^{-1}}
{}\leq{}
\Delta_{\nu T}
{}-{}
\Delta_{(\nu+1)T}
{}\leq{}
\Delta_{\nu T},
\]
\(R\)-linear convergence of \(\seq{\|\bm x^{\nu T}-\bm z^{\nu T}\|}[\nu\in\N]\) follows.
By suitably shifting, for every \(t\in[T]\) the same can be said for the sequence
\(
\seq{\|\bm z^{T\nu+t}-\bm x^{T\nu+t}\|}[\nu\in\N]
\),
and since \(T\) is finite we conclude that the whole sequence
\(
\seq{\|\bm z^k-\bm x^k\|}
\)
converges $R$-linearly. On the other hand, since \(\|\bm x^{k+1}-\bm x^k\|\leq\|\bm z^k-\bm x^k\|\), also \(\seq{\|\bm x^{k+1}-\bm x^k\|}\) converges \(R\)-linearly, hence so does \(\seq{\bm x^k}\). By combining the two, we conclude that also $\seq{\bm z^k}$ converges $R$-linearly.
\begin{comment}
As in the proof of \Cref{thm:cyclic:global} for \(k\) large enough inequality \eqref{eq:KL} holds, that is,
\[
\textstyle
\rho(1-\theta)(\mathcal L_{\nu T}-\Phi_\star)^{-\theta}
{}\geq{}
\frac{1}{
\sqrt{\gamma_{\rm min}^{-1}+\gamma_{\rm max}}\,
\|\bm x^k-\bm z^k\|_{\Gamma^{-1}}
}
\]
owing to the fact that \(\psi(s)=\rho s^{1-\theta}\).
Therefore,
\begin{align*}
\Delta_{\nu T}
{}={}
\rho(\mathcal L_{\nu T}-\Phi_\star)^{1-\theta}
{}\leq{} &
\rho
\left(
\rho(1-\theta)\sqrt{\gamma_{{\rm min}}^{-1}+\gamma_{{\rm max}}}\,\|{\bm x}^{\nu T}-{\bm z}^{\nu T}\|_{\Gamma^{-1}}
\right)^{\frac{1-\theta}{\theta}}
\\
{}\leq{} &
\rho^2(1-\theta)\sqrt{\gamma_{{\rm min}}^{-1}+\gamma_{{\rm max}}}\,\|{\bm x}^{\nu T}-{\bm z}^{\nu T}\|_{\Gamma^{-1}},
\end{align*}
where in the second inequality we used the fact $(1-\theta)/\theta\geq 1$ and that the base of the exponent is smaller than one for \(k\) large enough since $\|{\bm x}^{\nu T}-{\bm z}^{\nu T}\|_{\Gamma^{-1}}$ converges to zero.
Combined with \eqref{eq:KLinequality} it follows that $\seq{\Delta_{\nu T}}[\nu\in\N]$ conveges $Q$-linearly.
By rearranging \eqref{eq:KLinequality} as
\[
\tfrac{\sigma'}{2\sqrt{\gamma_{\rm min}^{-1}+\gamma_{\rm max}}}
\|\bm x^{\nu T}-\bm z^{\nu T}\|_{\Gamma^{-1}}
{}\leq{}
\Delta_{\nu T}
{}-{}
\Delta_{(\nu+1)T}
{}\leq{}
\Delta_{\nu T}
\]
\(R\)-linear convergence of \(\seq{\|\bm x^k-\bm z^k\|}\) follows.
By suitably shifting, for every \(t\in[T]\) the same can be said for the sequence
\(
\seq{\|\bm z^{T\nu+t}-\bm x^{T\nu+t}\|}[\nu\in\N]
\),
and since \(T\) is finite we conclude that the whole sequence
\(
\seq{\|\bm z^k-\bm x^k\|}
\)
converges $R$-linearly. On the other hand since $\|\bm x^{k+1}-\bm x^k\|\leq \|\bm z^k-\bm x^k\|$, $\seq{\|\bm x^{k+1}-\bm x^k\|}$ also converges $R$-linearly, hence so does $\seq{\bm x^k}$.
\end{comment}
\end{proof}
\end{thm}
\section{Nonconvex finite sum problems: the Finito/MISO algorithm}\label{sec:Finito}
As mentioned in \Cref{sec:Introduction}, if \(G\) is of the form \eqref{eq:FINITOG} then problem \eqref{eq:P} reduces to the finite sum minimization presented in \eqref{eq:FSP}.
Most importantly, the proximal mapping of the original nonsmooth function \(G\) can be easily expressed in terms of that of the small function \(g\) in the reduced finite sum reformulation, as shown in the next lemma.
\begin{lem}
Given \(\gamma_i>0\), \(i\in[N]\), let
\(
\Gamma
{}\coloneqq{}
\blockdiag(\gamma_1I_n,\dots,\gamma_NI_n)
\)
and
\(
\hat\gamma
{}\coloneqq{}
\bigl(\sum_{i=1}^N\gamma_i^{-1}\bigr)^{-1}
\).
Then, for \(G\) as in \eqref{eq:FINITOG} and any \(\bm u\in\R^{Nn}\)
\[
\textstyle
\prox_G^{\Gamma^{-1}}(\bm u)
{}={}
\set{(\hat v,\dots,\hat v)}[
\hat v
{}\in{}
\prox_{\hat\gamma g}(\hat u)
]
\quad\text{where}\quad
\hat u
{}\coloneqq{}
\hat\gamma
\sum_{i=1}^N\gamma_i^{-1}u_i.
\]
\begin{proof}
Observe first that for every \(w\in\R^n\) one has
\begin{align*}
\textstyle
\sum_i\gamma_i^{-1}\|w-u_i\|^2
{}={} &
\textstyle
\sum_i\gamma_i^{-1}\|\hat u-u_i\|^2
{}+{}
\sum_i\gamma_i^{-1}\|w-\hat u\|^2
{}+{}
\smashoverbrace{
\textstyle
2\sum_i\gamma_i^{-1}\innprod{\hat u-u_i}{w-\hat u}
}{
=0
}
\\
\numberthis\label{eq:mean}
{}={} &
\textstyle
\sum_i\gamma_i^{-1}\|\hat u-u_i\|^2
{}+{}
\hat\gamma^{-1}\|w-\hat u\|^2.
\end{align*}
Next, observe that since \(\dom G\subseteq C\) (the consensus set),
\begin{align*}
\prox_G^{\Gamma^{-1}}(\bm u)
{}={} &
\argmin_{\bm w\in\R^{Nn}}\set{\textstyle
G(\bm w)+\sum_{i=1}^N\tfrac{1}{2\gamma_i}\|w_i-u_i\|^2
}
\\
{}={} &
\argmin_{\bm w\in\R^{Nn}}\set{\textstyle
G(\bm w)+\sum_{i=1}^N\tfrac{1}{2\gamma_i}\|w_i-u_i\|^2
}[
w_1=\dots=w_N
]
\\
{}={} &
\argmin_{(w,\dots,w)}\set{\textstyle
g(w)+\sum_{i=1}^N\tfrac{1}{2\gamma_i}\|w-u_i\|^2
}
\\
{}\overrel*{\eqref{eq:mean}}{} &
\argmin_{(w,\dots,w)}\set{\textstyle
g(w)
{}+{}
\tfrac{1}{2\hat\gamma}\|w-\hat u\|^2
}
{}={}
\set{(\hat v,\dots,\hat v)}[
\hat v\in\prox_{\hat\gamma g}(\hat u)
]
\end{align*}
as claimed.
\end{proof}
\end{lem}
If all stepsizes are set to a same value \(\gamma\), so that \(\Gamma=\gamma\I_{Nn}\), then the forward-backward step reduces to
\begin{align*}
\bm z
{}\in{}
\prox_G^{\Gamma^{-1}}(\bm x-\Gamma\nabla F(\bm x))
\quad\Leftrightarrow\quad &
\bm z=(\bar z,\dots,\bar z),
\\
\numberthis\label{eq:FBFinito}
&
\bar z
{}\in{}
\prox_{\gamma g\nicefrac{}N}\left(
\textstyle
\tfrac1N\sum_{j=1}^N\bigl(
x_j-\tfrac\gamma N\nabla f_j(x_j)
\bigr)
\right).
\end{align*}
The argument of \(\prox_{\gamma g\nicefrac{}{N}}\) is the (unweighted) average of the forward operator.
By applying \Cref{alg:BC} with \eqref{eq:FBFinito}, Finito/MISO \cite{defazio2014finito,mairal2015incremental} is recovered.
Differently from the existing convergence analyses, ours covers fully nonconvex and nonsmooth problems, more general sampling strategies and the possibility to select different stepsizes \(\gamma_i\) for each block, which can have a significant impact on the performance compared to the case where all stepsizes are equal.
Moreover, to the best of our knowledge this is the first work that shows global convergence and linear rates even when the smooth functions are nonconvex.
The resulting scheme is presented in \Cref{alg:Finito}.
We remark that the consensus formulation to recover Finito/MISO (although from a different umbrella algorithm) was also observed in \cite{davis2016smart} in the convex case.
Moreover, the Finito/MISO algorithm with cyclic sampling is also studied in \cite{mokhtari2018surpassing} when \(g\equiv0\) and $f_i$ are strongly convex functions; consistently with \Cref{ass:cyclic}, our analysis covers the more general essentially cyclic sampling even in the presence of a nonsmooth convex term $g$ and allowing the smooth functions $f_i$ to be nonconvex.
Randomized Finito/MISO with $g\equiv 0$ is also studied in the recent work \cite{qian2019miso}; although their analysis is limited to a single stepsize, in the convex case it is allowed to be larger than our worst-case stepsize \(\min_i\gamma_i\).
\begin{algorithm}
\caption{Nonconvex proximal Finito/MISO for problem \eqref{eq:FSP
}
\label{alg:Finito}
\begin{algorithmic}[1]
\item[{\sc Require}]
\(
x^{\rm init}\in\R^n
\),~
\(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\),
{\small \(i\in[N]\)}
\Statex
\(
\hat\gamma
{}\coloneqq{}
\bigl(\sum_{i=1}^N\gamma_i^{-1}\bigr)^{-1}
\),~~
\(
s_i
{}={}
x^{\rm init}-\frac{\gamma_i}{N}\nabla f_i(x^{\rm init})
\)~
\(i\in[N]\),~~
\(
\hat s
{}={}
{\hat\gamma}\sum_{i=1}^N\gamma_i^{-1}s_i
\)
\item[{\sc Repeat} until convergence]
\State
select a set of indices \(I\subseteq[N]\)
\State
\(
z
{}\in{}
\prox_{\hat\gamma g}(\hat s)
\)
\For{ \(i\in I\) }
\State
\(
v
{}\gets{}
z-\frac{\gamma_i}{N}\nabla f_i(z)
\)
\State
update~~
\(
\hat s
{}\gets{}
\hat s+\tfrac{\hat\gamma}{\gamma_i}(v-s_i)
\)
~~and~~
\(
s_i\gets v
\
\EndFor
\item[{\sc Return} $z$ ]
\end{algorithmic}
\end{algorithm}
The convergence results from \Cref{sec:convergence} are immediately translated to this setting by noting that the bold variable ${\bm z}^k$ corresponds to $(z^k,\dots,z^k)$.
Therefore, $\Phi({\bm z^k})= \varphi(z^k)$ where $\varphi$ is the cost function for the finite sum problem.
\begin{cor}[subsequential convergence of \Cref{alg:Finito}]\label{thm:Finito:convergence
In the finite sum problem \eqref{eq:FSP} suppose that \(\argmin\varphi\) is nonempty, \(g\) is proper and lsc, and each \(f_i\) is \(L_{f_i}\)-Lipschitz differentiable, \(i\in[N]\).
Then, the following hold almost surely (resp. surely) for the sequence $\seq{z^k}$ generated by \Cref{alg:Finito} with randomized sampling strategy as in \Cref{ass:random} (resp. with any essentially cyclic sampling strategy and $g$ convex as required in \Cref{ass:cyclic}):
\begin{enumerate}
\item
the sequence \(\seq{\varphi(z^k)}\) converges to a finite value \(\varphi_\star\leq\varphi(x^{\rm init})\);
\item
all cluster points of the sequence \(\seq{z^k}\) are stationary and on which \(\varphi\) equals \(\varphi_\star\).
\end{enumerate}
If, additionally, \(\varphi\) is coercive, then the following also hold:
\begin{enumerate}[resume]
\item
\(\seq{z^k}\) is bounded (in fact, this holds surely for arbitrary sampling criteria).
\end{enumerate}
\end{cor}
\begin{cor}[linear convergence of \Cref{alg:Finito} under strong convexity
Additionally to the assumptions of \Cref{thm:Finito:convergence}, suppose that \(g\) is convex and that each \(f_i\) is \(\mu_{f_i}\)-strongly convex.
The following hold for the iterates generated by \Cref{alg:Finito}:
\begin{itemize}[leftmargin=*,label={},itemindent=-0.5cm,labelsep=0pt,partopsep=0pt,parsep=0pt,listparindent=0pt,topsep=0pt]
\item{\sc Randomized sampling:}
under \Cref{ass:random},
\begin{align*}
\E[]{\varphi(z^k)-\min\varphi}
{}\leq{} &
(\varphi(x^{\rm init})-\min\varphi)
(1-c)^k
\\
\tfrac12\E[]{\|z^k-x^\star\|^2}
{}\leq{} &
\frac{
N(\varphi(x^{\rm init})-\min\varphi)
}{
\sum_i\mu_{f_i}
}
(1-c)^k
\end{align*}
holds for all \(k\in\N\), where \(c\) is as in \eqref{eq:cwc} and \(x^\star\coloneqq\argmin\varphi\).
If the stepsizes \(\gamma_i\) and the sampling probabilities \(p_i\) are set as in \Cref{thm:random:linear}, then the tighter constant \(c\) as in \eqref{eq:cbc} is obtained.
\item{\sc Shuffled cyclic or cyclic sampling:}
under either sampling strategy \eqref{eq:ShufCyclicRule} or \eqref{eq:cyclicRule},
\begin{align*}
\varphi(z^{\nu N})-\min\varphi
{}\leq{} &
(\varphi(x^{\rm init})-\min\varphi)
(1-c)^\nu
\\
\tfrac12\E[]{\|z^{\nu N}-x^\star\|^2}
{}\leq{} &
\frac{
N(\varphi(x^{\rm init})-\min\varphi)
}{
\sum_i\mu_{f_i}
}
(1-c)^\nu
\end{align*}
holds surely for all \(\nu\in\N\), where \(c\) is as in \eqref{eq:linearShuffledCyclic}.
\end{itemize}
\end{cor}
The next result follows from \Cref{thm:cyclic:global} once the needed properties of \(\Phi\) as in the umbrella formulation \eqref{eq:P} are shown to hold.
\begin{cor}[global convergence of \Cref{alg:Finito}]\label{thm:Finito:global
In the finite sum problem \eqref{eq:FSP}, suppose that \(\varphi\) has the KL property with exponent \(\theta\in(0,1)\) (as is the case when \(f_i\) and \(g\) are semialgebraic) and coercive, \(g\) is proper convex and lsc, and each \(f_i\) is \(L_{f_i}\)-Lipschitz differentiable, \(i\in[N]\).
Then the sequence \(\seq{z^k}\) generated by \Cref{alg:Finito} with any essentially cyclic sampling strategy as in \Cref{ass:cyclic} converges surely to a stationary point for \(\varphi\).
Moreover, if \(\theta\leq\nicefrac12\) then it converges at \(R\)-linear rate.
\begin{proof}
Function \(\Phi=F+G\) be as in \eqref{eq:FINITOG} clearly is coercive and satisfies \Cref{ass:basic}.
In order to invoke \Cref{thm:cyclic:global} is suffices to show that there exists a constant \(c>0\) such that
\begin{equation}\label{eq:Dist}
\dist(0,\partial\Phi(\bm x))
{}\geq{}
c\dist(0,\partial\varphi(x))
\quad
\text{for all \(x\in\R^n\) and \(\bm x=(x,\dots,x)\),}
\end{equation}
as this will ensure that \(\Phi\) enjoys the KL property at \(\bm x^\star=(x^\star,\dots,x^\star)\) with same desingularizing function (up to a positive scaling).
Notice that for \(x\in\R^n\) and \(\bm x=(x,\dots,x)\), one has
\(
\bm v\in\partial G(\bm x)
\)
iff
\(
\frac1N\sum_{i=1}^Nv_i
{}\in{}
\partial g(x)
\).
Since
\(
\partial\Phi(\bm x)
{}={}
\tfrac1N\mathop\times_{i=1}^N\nabla f_i(x_i)+\partial G(\bm x)
\)
and
\(
\partial\varphi(x)
{}={}
\tfrac1N\sum_{i=1}^N\nabla f_i(x)
{}+{}
\partial g(x)
\),
see \cite[Ex. 8.8(c) and Prop. 10.5]{rockafellar2011variational}, for \(x\in\R^n\) and denoting \(\bm x=(x,\dots,x)\) we have
\begin{align*}
\dist(0,\partial\varphi(x))
{}\leq{} &
\inf_{\bm v\in\partial G(\bm x)}{
\left\|\textstyle
\tfrac1N\sum_{i=1}^N\nabla f_i(x)
{}+{}
\tfrac1N\sum_{i=1}^Nv_i
\right\|
}
\\
{}\leq{} &
\tfrac1N\inf_{\bm v\in\partial G(\bm x)}{
\textstyle
\sum_{i=1}^N\|\nabla f_i(x)+v_i\|
}
{}={}
\tfrac1N\inf_{\bm u\in\partial\Phi(\bm x)}{
\newnorm{\bm u}
},
\end{align*}
where \(\newnorm{{}\cdot{}}\) is the norm in \(\R^{Nn}\) given by
\(
\newnorm{\bm w}=\sum_{i=1}^N\|w_i\|
\).
Inequality \eqref{eq:Dist} then follows by observing that
\(
\inf_{\bm u\in\partial\Phi(\bm x)}{
\newnorm{\bm u}
}
\)
is the distance of \(0\) from \(\partial\Phi(\bm x)\) in the norm \(\newnorm{{}\cdot{}}\), hence that \(\newnorm{{}\cdot{}}\leq c'\|{}\cdot{}\|\) for some \(c'>0\).
\end{proof}
\end{cor}
\section{Nonconvex sharing problem}\label{sec:Sharing}
In this section we consider the sharing problem \eqref{eq:SP}.
As discussed in \Cref{sec:Introduction}, \eqref{eq:SP} fits into the problem framework \eqref{eq:P} by simply letting \(G\coloneqq g \circ A\), where
\(A\coloneqq[\I_n~\dots~\I_n]\in\R^{n\times nN}\).
By arguing as in \cite[Th. 6.15]{beck2017first} it can be shown that, when \(A\) has full row rank, the proximal mapping of $G=g\circ A$ is given by
\begin{equation}\label{eq:sharingprox}
\prox_G^{\Gamma^{-1}}(\bm u)
{}={}
\bm u+\Gamma\trans A(A\Gamma\trans A\,)^{-1}\left(\prox^{(A\Gamma\trans A\,)^{-1}}_g\left(A\bm u\right)-A\bm u\right).
\end{equation}
Since \(A\Gamma\trans A=\sum_{i=1}^N\gamma_i\) for the sharing problem \eqref{eq:SP},
\begin{align*}
\bm v
{}\in{}
\prox_G^{\Gamma^{-1}}(\bm u)
~~\Leftrightarrow~~ &
\bm v
{}={}
(u_1+\gamma_1w,\dots,u_N+\gamma_Nw)
\\
&
\textstyle
w
{}\in{}
\tilde\gamma^{-1}\left(\prox_{\tilde{\gamma}g}(\tilde u)-\tilde u\right),
~~
\tilde\gamma\coloneqq\sum_{i=1}^N\gamma_i,
~~
\tilde u\coloneqq \sum_{i=1}^Nu_i.
\end{align*}
Consequently general BC \Cref{alg:BC} when applied to the sharing problem \eqref{eq:SP} reduces to \Cref{alg:Sharing}.
\begin{algorithm}
\caption{Block-coordinate method for nonconvex sharing problem \eqref{eq:SP}
\label{alg:Sharing
\begin{algorithmic}[1]
\item[{\sc Require}]
\(
x_i^{\rm init}\in\R^{n}
\),~
\(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\),
{\small \(i\in [N]\)}
\Statex
\(
\tilde\gamma
{}\coloneqq{}
\sum_{i=1}^N\gamma_i
\),~~
\(
s_i
{}={}
x_i^{\rm init}-\frac{\gamma_i}{N}\nabla f_i(x_i^{\rm init})
\)~
\(i\in [N]\),~~\(
\tilde s
{}={}
\sum_{i=1}^N s_i
\
\item[{\sc Repeat} until convergence]
\State
select a set of indices \(I\subseteq[N]\)
\State $w \gets \tilde{\gamma}^{-1}(\prox_{\tilde{\gamma}g}(\tilde s)-\tilde s)$
\For{ \(i\in I\) }
\State
\(
v_i {}\gets{} s_i + \gamma_i w - \tfrac{\gamma_i}{N} \nabla f_i(s_i + \gamma_i w )
\)
\State
update~~
\(
\tilde s
{}\gets{}
\tilde s+(v_i-s_i)
\)
~~and~~
\(
s_i \gets v_i
\
\EndFor
\item[{\sc Return}]
$\bm z=(s_1 + \gamma_1 w ,\dots,s_N + \gamma_N w)$ with $w\in\tilde{\gamma}^{-1}(\prox_{\tilde{\gamma}g}(\tilde s)-\tilde s)$
\end{algorithmic}
\end{algorithm}
\begin{rem}[generalized sharing constraint]
Another notable instance of $G=g\circ A$ well suited for the BC framework of \Cref{alg:BC} is when \(g=\indicator_{\set0}\) and \(A=[A_1~\dots~A_N]\), \(A_i\in\R^{n\times n_i}\) such that $A$ is full rank.
This models the generalized sharing problem
\[
\minimize_{\bm x\in\R^{\sum_in_i}}{\textstyle
\tfrac1N\sum_{i=1}^Nf_i(x_i)
}
\quad\stt{}\textstyle
\sum_{i=1}^NA_ix_i=0.
\]
In this case \eqref{eq:sharingprox} simplifies to
\[
\left(\prox_G^{\Gamma^{-1}}(\bm u)\right)_i
{}={}
u_i-\gamma_i\trans{A_i}\mathcal A^{-1}\sum_{i=1}^NA_iu_i,
\]
where $\mathcal A\coloneqq A\Gamma\trans A$ can be factored offline and \(\sum_{i=1}^NA_ix_i\) can be updated in an incremental fashion in the same spirit of \Cref{alg:Sharing}.
\end{rem}
The convergence results for \Cref{alg:Sharing} summarized below fall as special cases of those in \Cref{sec:convergence}.
\begin{cor}[convergence of \Cref{alg:Sharing}]\label{thm:sharing:convergence
In the sharing problem \eqref{eq:SP}, suppose that \(\argmin\Phi\) is nonempty, \(g\) is proper and lsc, and each \(f_i\) is \(L_{f_i}\)-Lipschitz differentiable, \(i\in[N]\). Consider the sequences $\seq{w^k}$ and $\seq{\bm s^k}$ generated by \Cref{alg:Sharing} and let $\seq{\bm z^k}=\seq{s_1^k + \gamma_1 w^k ,\dots,s_N^k + \gamma_N w^k}$.
Then, the following hold almost surely (resp. surely) with randomized sampling strategy as in \Cref{ass:random} (resp. with any essentially cyclic sampling strategy and $g$ convex as required in \Cref{ass:cyclic}):
\begin{enumerate}
\item
the sequence \(\seq{\Phi(\bm z^k)}\) converges to a finite value \(\Phi_\star\leq\Phi(\bm x^{\rm init})\);
\item
all cluster points of the sequence \(\seq{\bm z^k}\) are stationary and on which \(\Phi\) equals \(\Phi_\star\).
\end{enumerate}
If, additionally, \(\Phi\) is coercive, then the following also hold:
\begin{enumerate}[resume]
\item
\(\seq{\bm z^k}\) is bounded (in fact, this holds surely for arbitrary sampling criteria).
\end{enumerate}
\end{cor}
\begin{cor}[linear convergence of \Cref{alg:Sharing} under strong convexity]\label{cor:RLinSharing
Additionally to the assumptions of \Cref{thm:sharing:convergence}, suppose that \(g\) is convex and that each \(f_i\) is \(\mu_{f_i}\)-strongly convex.
The following hold:
\begin{itemize}[leftmargin=*,label={},itemindent=-0.5cm,labelsep=0pt,partopsep=0pt,parsep=0pt,listparindent=0pt,topsep=0pt]
\item{\sc Randomized sampling:}
under \Cref{ass:random},
\begin{align*}
\E[]{\Phi(\bm z^k)-\min\Phi}
{}\leq{} &
\bigl(\Phi(\bm x^{\rm init})-\min\Phi\bigr)(1-c)^k
\\
\tfrac12\E[]{\|\bm z^k-\bm x^\star\|^2_{\mu_F}}
{}\leq{} &
\bigl(\Phi(\bm x^{\rm init})-\min\Phi\bigr)(1-c)^k
\end{align*}
holds for all \(k\in\N\), where \(\bm x^\star\coloneqq\argmin\Phi\),
\(
\mu_F
{}\coloneqq{} \tfrac1N
\blockdiag\bigl(\mu_{f_1}\I_{n_1},\dots\mu_{f_n}\I_{n_N}\bigr)
\),
and \(c\) is as in \eqref{eq:cwc}.
If the stepsizes \(\gamma_i\) and the sampling probabilities \(p_i\) are set as in \Cref{thm:random:linear}, then the tighter constant \(c\) as in \eqref{eq:cbc} is obtained.
\item{\sc Shuffled cyclic or cyclic sampling:}
under either sampling strategy \eqref{eq:ShufCyclicRule} or \eqref{eq:cyclicRule},
\begin{align*}
\Phi(\bm z^{N\nu})-\min\Phi
{}\leq{} &
\bigl(\Phi(\bm x^{\rm init})-\min\Phi\bigr)(1-c)^\nu
\\
\tfrac12\|\bm z^{N\nu}-\bm x^\star\|^2_{\mu_F}
{}\leq{} &
\bigl(\Phi(\bm x^{\rm init})-\min\Phi\bigr)(1-c)^\nu
\end{align*}
holds surely for all \(\nu\in\N\), where \(c\) is as in \eqref{eq:linearShuffledCyclic}.
\end{itemize}
\end{cor}
We conclude with an immediate consequence of \Cref{thm:cyclic:global} that shows that (strong) convexity is in fact not necessary for global or linear convergence to hold.
\begin{cor}[global and linear convergence of \Cref{alg:Sharing}]\label{thm:Sharing:global
In problem \eqref{eq:SP}, suppose that \(\Phi\) has the KL property with exponent \(\theta\in(0,1)\) (as is the case when \(g\) and \(f_i\) are semialgebraic) and is coercive, \(g\) is proper convex lsc, and each \(f_i\) is \(L_{f_i}\)-Lipschitz differentiable, \(i\in[N]\).
Then the sequence $\seq{\bm z^k}$ as defined in \Cref{thm:sharing:convergence} with any essentially cyclic sampling strategy as in \Cref{ass:cyclic} converges surely to a stationary point for \(\Phi\).
Moreover, if \(\theta\leq\nicefrac12\) it converges with \(R\)-linear rate.
\end{cor}
\ifaccel
\section{Accelerated block-coordinate proximal gradient}
The work \cite{allen2016even} introduced a coordinate descent method for smooth convex minimization, in which each coordinate is randomly sampled according to an ad hoc probability distribution that provably leads to a remarkable speed up with respect to uniform sampling strategies.
The unified analysis of BC-algorithms and the analytical tool introduced in this paper, the forward backward envelope function, allow the extention of this approach to nonsmooth convex minimization of the form \eqref{eq:P},
where functions \(f_i\) are convex quadratic and \(G\) is convex but possibly nonsmooth:
\begin{ass}[requirements for the fast BC-\Cref{alg:Fast}]\label{ass:Fast
In problem \eqref{eq:P}, \(\func{G}{\R^{\sum_in_i}}{\Rinf}\) is proper convex and lsc, and
\(f_i(x_i)\coloneqq\tfrac12\trans{x_i}H_ix_i+\trans{q_i}x_i\) is convex quadratic, with \(L_{f_i}\coloneqq\lambda_{\rm max}(H_i)\) and \(\mu_{f_i}\coloneqq\lambda_{\rm min}(H_i)\geq0\),
\(i\in[N]\).
\end{ass}
Let $U_i\in \R^{{\sum_{i=1}^N n_i}\times n_i}$ denote the $i$-th block column of the identity matrix so that for a vector $v\in \R^{n_i}$
\begin{equation}\label{eq:U}
U_iv= (0,\dots, 0, \!\overbracket{\,v\,}^{\mathclap{i\text{-th}}}\!, 0, \dots, 0).
\end{equation}
The accelerated BC scheme based on \cite{allen2016even} (for both strongly convex and convex cases) is given in \Cref{alg:Fast}.
Similarly to the approach of \cite{patrinos2014douglas} where an accelerated Douglas-Rachford algorithm is proposed, in order to derive \Cref{alg:Fast} we consider the scaled problem \( \minimize_{\tilde{\bm x}} \FBEC(\tilde{\bm x})\) where $\FBEC \coloneqq \FBE\circ Q^{-1/2}$, and $Q$ is the symmetric positive definite matrix
\begin{equation} \label{eq:QQ}
Q
{}\coloneqq{}
\blockdiag(Q_1,\dots,Q_N)\succ0
\quad \text{with } Q_i
{}\coloneqq{}
\gamma_i^{-1}\I-\tfrac{1}{N}H_i\in\R^{n_i\times n_i},~i\in[N].
\end{equation}
As detailed in \Cref{thm:convex}, whenever \Cref{ass:Fast} is satisfied $\FBEC$ is a convex Lipschitz-differentiable function, and its gradient is given by $\nabla \FBEC(\tilde{\bm x}) = Q^{1/2}(\bm x-\prox_G^{\Gamma^{-1}}(\bm x-\Gamma\nabla F(\bm x)))$ where $\bm x=Q^{-1/2}\tilde{\bm x}$.
Note that, based on \Cref{thm:convex}, \(\FBEC\) is \(1\)-smooth along the \(i\)-th block (in the notation of \cite{allen2016even}, \(L_i=1\), $S_\alpha=N$, and \(p_i=\nicefrac1N\)).
Hence the parameters of the algorithm simplify substantially resulting in uniform sampling.
Moreover, when functions \(f_i\) are \(\mu_{f_i}\)-strongly convex, by \Cref{thm:convex} \(\FBEC\) is $\sigma$-strongly convex with $\sigma= \frac{1}{N} \min_{i\in [N]} \{\gamma_i\mu_{f_i}\}$.
\Cref{alg:Fast} is obtained by applying the fast BC to this problem and scaling the variables by $Q^{-1/2}$.
Specifically, the update rule as in \cite{allen2016even} reads
\[
\begin{cases}[ r @{{}={}} l ]
\tilde{\bm x}^+ & \tau\tilde{\bm w}+(1-\tau)\tilde{\bm y}
\\
\tilde{\bm y}^+ & \tilde{\bm x}-U_i\trans{U_i}\nabla\FBEC(\tilde{\bm x}^+)
{}={}
\tilde{\bm x}-U_iQ_i^{\nicefrac12}(x_i^+-z_i^+)
\\
\tilde{\bm w}^+ & \tfrac{1}{1+\eta\sigma}(\tilde{\bm w}+\eta\sigma\tilde{\bm x}^+-N\eta U_i\trans{U_i}\nabla\FBEC(\tilde{\bm x}^+))
{}={}
\tfrac{1}{1+\eta\sigma}(\tilde{\bm w}+\eta\sigma\tilde{\bm x}^+-N\eta U_iQ_i^{\nicefrac12}(x_i^+-z_i^+)),
\end{cases}
\]
where \(\bm z^+=\prox_G^{\Gamma^{-1}}(\bm x^+-\Gamma\nabla F(\bm x^+))\).
Since \(Q^{-\nicefrac12}U_iQ_i^{\nicefrac12}=U_i\), premultiplying by \(Q^{-\nicefrac12}\) yields
\[
\begin{cases}[ r @{{}={}} l ]
\bm x^+ & \tau\bm z+(1-\tau)\bm y
\\
\bm z^+ & \prox_G^{\Gamma^{-1}}(\bm x^+-\Gamma\nabla F(\bm x^+))
\\
\bm y^+ & \bm x+U_i(z_i^+-x_i^+)
\\
\bm w^+ & \tfrac{1}{1+\eta\sigma}(\bm w+\eta\sigma\bm x^++N\eta U_i(z_i^+-x_i^+)).
\end{cases}
\]
For computational efficiency, vectors $\Gamma\nabla F(\bm x^k)$ and $\Gamma\nabla F(\bm w^k)$ are stored in variables $\bm r^k$ and $\bm v^k$ and updated recursively using the fact that gradients are affine, in such a way that each iteration requires only the evaluation of the sampled gradient (see \Cref{state:d}).
For similar reasons, in \Cref{alg:Fast} the iterates start with the $\bm y$-update rather than the $\bm x$-update as in \cite{allen2016even}.
Moreover, in the same spirit of \Cref{alg:BC} this accelerated variant can be implemented efficiently whenever the individual blocks of \(\bm z^+\) can be computed efficiently, similarly to the cases discussed in \Cref{sec:Finito,sec:Sharing}.
\begin{algorithm}
\caption{Accelerated block-coordinate proximal gradient for problem \eqref{eq:P} under \Cref{ass:Fast}
\label{alg:Fast
\begin{algorithmic}[1]
\item[{\sc Require}]
\(\bm x^0\in\R^{\sum_in_i}\),~
\(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\),~$i\in[N]$,
\(
\sigma
{}\coloneqq{}
\frac{1}{N} \min_{i\in [N]} \{\gamma_i\mu_{f_i}\}
\)
\State\label{state:sigma_beta
{\bf if} \(\sigma=0\),~~{\bf then}~
\(\eta = \nicefrac{1}{N^2}\)
~~{\bf otherwise}~
set
\(
\tau
{}={}
\frac{2}{1+\sqrt{1+\nicefrac{4N^2}{\sigma}}}
\),
\(
\eta
{}={}
\frac{1}{\tau N^2}
\)~
{\bf end if}
\State
\(
\bm w^0
{}={}
\bm x^0
\),~
\(
(\bm v^0,\bm r^0)
{}={}
(\Gamma\nabla F(\bm x^0),\Gamma\nabla F(\bm x^0))
\),~
\(
\bm z^{0}
{}={}
\prox_G^{\Gamma^{-1}}\bigl(\bm x^{0}-\bm r^{0}\bigr)
\)
\def\myVar#1
\fillwidthof[l]{\bm x^{k+1}}{#1
\For{ \(k=0,1,\dots\) }
\State
sample \(i\in[N]\) uniformly
\State\label{state:d
\(
\myVar{\bm y^{k+1}}
{}\gets{}
\bm x^{k}+U_{i}\bigl(z_{i}^{k}-x_{i}^{k}\bigr)
\),\qua
\(
d
{}\gets{}
\tfrac{\gamma_i}{N}\nabla f_i(z_i^{k})-r_i^{k}
\)
\State
\(
\myVar{\bm v^{k+1}}
{}={}
\frac{1}{1+\eta\sigma}\Bigl(\bm v^{k}+\eta\sigma\bm r^{k}+N{\eta} U_id\Bigr)
\),\qua
\(
\bm w^{k+1}
{}={}
\frac{1}{1+\eta\sigma}\Bigl(\bm w^{k}+\eta\sigma\bm x^{k}+{N\eta}U_{i}\bigl(z_{i}^{k}-x_{i}^{k}\bigr)\Bigr)
\)
\State{\bf if}~~\(\sigma=0\),~~{\bf then}~
\(
\eta
{}\gets{}
\frac{k+3}{2N^2}
\),~
\(
\tau
{}\gets{}
\frac{2}{k+3}
\)
~~{\bf end if
\State\label{state:FBEgrad
\(
\myVar{\bm x^{k+1}}
{}={}
\tau \bm w^{k+1}+(1-\tau)\bm y^{k+1}
\),\qua
\(
\bm r^{k+1}
{}={}
\tau \bm v^{k+1}+(1-\tau)(\bm r^{k}+U_i d)
\
\State
\(
\myVar{\bm z^{k+1}}
{}={}
\prox_G^{\Gamma^{-1}}\bigl(\bm x^{k+1}-\bm r^{k+1}\bigr)
\)
\EndFor{}
\end{algorithmic}
\end{algorithm}
The convergence rate results follow directly from those of \cite{allen2016even} with parameters \(L_i=1\) and $S_\alpha=N$ as described above.
\begin{thm}[convergence rates of \Cref{alg:Fast}]
Suppose that \Cref{ass:basic,ass:Fast} are satisfied.
Then, the iterates generated by \Cref{alg:Fast} satisfy
\[
\mathbb{E}\left[{\FBE(\bm y^k)-\min \Phi}\right]
{}\leq{}
\frac{2N^2\|\bm x^0 - \bm x^\star\|^2_Q}{(k+1)^2},
\]
where $Q$ is as in \eqref{eq:QQ}.
Moreover, in the strongly convex case ($\sigma= \frac{1}{N} \min_{i\in [N]} \{\gamma_i\mu_{f_i}\}>0$)
\[
\mathbb{E}\left[{\FBE(\bm y^k)-\min \Phi}\right]
{}\leq{}
O(1)
(1-c)^k\left(
\FBE(\bm x^0)-\min \Phi
\right)
\quad\text{where}\quad
\textstyle
c
{}={}
\left(
\frac12
{}+{}
\sqrt{
\frac14
{}+{}
\frac{N^2}{\sigma}
}
\right)^{-1}.
\]
\end{thm}
Note that in the strongly convex case it follows from \Cref{thm:strconcost} that the distance from the solution decreases \(R\)-linearly as
\[
\E[]{ \|\bm y^{k}-\bm x^\star\|^2_M}
{}\leq{}
O(1)\left(1-c\right)^k
\left(\FBE(\bm x^0)-\min\Phi\right),
\]
where $M$ is as in \Cref{thm:strconcost}.
\fi
\section{Conclusions}\label{sec:Conclusions}
We presented a general block-coordinate forward-backward algorithm for minimizing the sum of a separable smooth and a nonseparable nonsmooth function, both allowed to be nonconvex.
The framework is general enough to encompass regularized finite sum minimization and sharing problems, and leads to (a generalization of) the Finito/MISO algorithm \cite{defazio2014finito,mairal2015incremental} with new convergence results and with another novel incremental-type algorithm.
The forward-backward envelope is shown to be a particularly suitable Lyapunov function for establishing convergence: additionally to enjoying favorable continuity properties, \emph{sure} descent (as opposed to in expectation) occurs along the iterates.
Possible future developments include extending the framework to account for a nonseparable smooth term, for instance by ``quantifying the strength of coupling'' between blocks of variables as in \cite[\S7.5]{bertsekas1989parallel}.
\ifarxiv
\clearpage
\fi
\begin{appendix}
\section{The key tool: the forward-backward envelope}\label{sec:appendix}
This appendix contains some proofs and auxiliary results omitted in the main body.
We begin by observing that, since \(F\) and \(-F\) are 1-smooth in the metric induced by
\(
\Lambda_F\coloneqq\tfrac1N\blockdiag(L_{f_1}\I_{n_1},\dots,L_{f_N}\I_{n_N})
\),
one has
\begin{equation}\label{eq:Lip}
F(\bm x)+\innprod{\nabla F(\bm x)}{\bm w-\bm x}
{}-{}
\tfrac12\|\bm w-\bm x\|_{\Lambda_F}^2
{}\leq{}
F(\bm w)
{}\leq{}
F(\bm x)+\innprod{\nabla F(\bm x)}{\bm w-\bm x}
{}+{}
\tfrac12\|\bm w-\bm x\|_{\Lambda_F}^2
\end{equation}
for all \(\bm x,\bm w\in\R^{\sum_in_i}\), see \cite[Prop. A.24]{bertsekas2016nonlinear}.
Let us denote
\[
\M(\bm w,\bm x)
{}\coloneqq{}
F(\bm x)+\innprod{\nabla F(\bm x)}{\bm w-\bm x}
{}+{}
G(\bm w)
{}+{}
\tfrac12\|\bm w-\bm x\|_{\Gamma^{-1}}^2
\]
the quantity being minimized (with respect to \(\bm w\)) in the definition \eqref{eq:FBE} of the FBE.
It follows from \eqref{eq:Lip} that
\begin{equation}\label{eq:bounds}
\Phi(\bm w)
{}+{}
\tfrac12\|\bm w-\bm x\|^2_{\Gamma^{-1}-\Lambda_F}
{}\leq{}
\M(\bm w,\bm x)
{}\leq{}
\Phi(\bm w)
{}+{}
\tfrac12\|\bm w-\bm x\|^2_{\Gamma^{-1}+\Lambda_F}
\end{equation}
holds for all \(\bm x,\bm w\in\R^{\sum_in_i}\).
In particular, \(\M\) is a \emph{majorizing model} for \(\Phi\), in the sense that \(\M(\bm x,\bm x)=\Phi(\bm x)\) and \(\M(\bm w,\bm x)\geq\Phi(\bm w)\) for all \(\bm x,\bm w\in\R^{\sum_in_i}\).
In fact, as explained in \Cref{sec:FBE}, while a \(\Gamma\)-forward-backward step \(\bm z\in\T(\bm x)\) amounts to evaluating a minimizer of \(\M({}\cdot{},\bm x)\), the FBE is defined instead as the minimization value, namely \(\FBE(\bm x)=\M(\bm z,\bm x)\) where \(\bm z\) is any element of \(\T(\bm x)\).
\subsection{Proofs of \texorpdfstring{\Cref{sec:FBE}}{\S\ref*{sec:FBE}}}\label{sec:proofs:FBE}
\begin{appendixproof}{thm:osc}
For \(\bm x^\star\in\argmin\Phi\) it follows from \eqref{eq:Lip} that
\[
\min\Phi
{}\leq{}
F(\bm x)
{}+{}
G(\bm x)
{}\leq{}
G(\bm x)
{}+{}
F(\bm x^\star)
{}+{}
\innprod{\nabla F(\bm x^\star)}{\bm x-\bm x^\star}
{}+{}
\tfrac12\|\bm x^\star-\bm x\|_{\Lambda_F}^2.
\]
Therefore, \(G\) is lower bounded by a quadratic function with quadratic term \(-\tfrac12\|{}\cdot{}\|_{\Lambda_F}^2\), and thus is prox-bounded in the sense of \cite[Def. 1.23]{rockafellar2011variational}.
The claim then follows from \cite[Th. 1.25 and Ex. 5.23(b)]{rockafellar2011variational} and the continuity of the forward mapping \(\Fw{}\).
\end{appendixproof}
\begin{appendixproof}{thm:FBEineq
Local Lipschitz continuity
\ifarxiv
of the FBE
\fi
follows from \eqref{eq:FBEMoreau} in light of \Cref{thm:osc} and \cite[Ex. 10.32]{rockafellar2011variational}.
\begin{proofitemize}
\item\ref{thm:leq}~
Follows by replacing \(\bm w=\bm x\) in \eqref{eq:FBE}.
\item\ref{thm:geq}~
Directly follows from \eqref{eq:bounds} and the identity \(\FBE(\bm x)=\M(\bm z,\bm x)\) for \(\bm z\in\T(\bm x)\).
\item\ref{thm:strconcost}~
By strong convexity, denoting \(\Phi_\star\coloneqq\min\Phi\), we have
\[
\Phi_\star
{}\leq{}
\Phi(\bm z)-\tfrac12\|\bm z-\bm x^\star\|_{\mu_F}^2
{}\leq{}
\FBE(\bm x)
{}-{}
\tfrac12\|\bm z-\bm x^\star\|_{\mu_F}^2
\]
where the second inequality follows from \Cref{thm:geq}.
\qedhere
\end{proofitemize}
\end{appendixproof}
\begin{appendixproof}{thm:FBEmin}
\begin{proofitemize}
\item\ref{thm:min} and \ref{thm:argmin}~
It follows from \Cref{thm:leq} that \(\inf\FBE\leq\min\Phi\).
Conversely, let \(\seq{\bm x^k}\) be such that \(\FBE(\bm x^k)\to\inf\FBE\) as \(k\to\infty\), and for each \(k\) let \(\bm z^k\in\T(\bm x^k)\).
It then follows from \Cref{thm:leq,thm:geq} that
\[
\inf\FBE
{}\leq{}
\min\Phi
{}\leq{}
\liminf_{k\to\infty}\Phi(\bm z^k)
{}\leq{}
\liminf_{k\to\infty}\FBE(\bm x^k)
{}={}
\inf\FBE,
\]
hence \(\min\Phi=\inf\FBE\).
Suppose now that \(\bm x\in\argmin\Phi\) (which exists by \Cref{ass:basic}); then it follows from \Cref{thm:geq} that \(\T(\bm x)=\set{\bm x}\) (for otherwise another element would belong to a lower level set of \(\Phi\)).
Combining with \Cref{thm:leq} with \(\bm z=\bm x\) we then have
\[
\min\Phi
{}={}
\Phi(\bm z)
{}\leq{}
\FBE(\bm x)
{}\leq{}
\Phi(\bm x)
{}={}
\min\Phi.
\]
Since \(\min\Phi=\inf\FBE\), we conclude that \(\bm x\in\argmin\FBE\), and that in particular \(\inf\FBE=\min\FBE\).
Conversely, suppose \(\bm x\in\argmin\FBE\) and let \(\bm z\in\T(\bm x)\).
By combining \Cref{thm:leq,thm:geq} we have that \(\bm z=\bm x\), that is, that \(\T(\bm x)=\set{\bm x}\).
It then follows from \Cref{thm:geq} and assert \ref{thm:min} that
\[
\Phi(\bm x)
{}={}
\Phi(\bm z)
{}\leq{}
\FBE(\bm x)
{}={}
\min\FBE
{}={}
\min\Phi,
\]
hence \(\bm x\in\argmin\Phi\).
\item\ref{thm:LB}~
Due to \Cref{thm:leq}, if \(\FBE\) is level bounded clearly so is \(\Phi\).
Conversely, suppose that \(\FBE\) is not level bounded.
Then, there exist \(\alpha\in\R\) and \(\seq{\bm x^k}\subseteq\lev_{\leq\alpha}\FBE\) such that \(\|\bm x^k\|\to\infty\) as \(k\to\infty\).
Let \(\lambda=\min_i\set{\gamma_i^{-1}-L_{f_i}N^{-1}}>0\), and for each \(k\in\N\) let \(\bm z^k\in\T(\bm x^k)\).
It then follows from \Cref{thm:geq} that
\[
\min\Phi
{}\leq{}
\Phi(\bm z^k)
{}\leq{}
\FBE(\bm x^k)
{}-{}
\tfrac\lambda2\|\bm x^k-\bm z^k\|^2
{}\leq{}
\alpha
{}-{}
\tfrac\lambda2\|\bm x^k-\bm z^k\|^2,
\]
hence \(\seq{\bm z^k}\subseteq\lev_{\leq\alpha}\Phi\) and
\(
\|\bm x^k-\bm z^k\|^2
{}\leq{}
\tfrac2\lambda(\alpha-\min\Phi)
\).
Consequently, also the sequence \(\seq{\bm z^k}\subseteq\lev_{\leq\alpha}\Phi\) is unbounded, proving that \(\Phi\) is not level bounded.
\qedhere
\end{proofitemize}
\end{appendixproof}
\subsection{Further results}\label{sec:auxiliary}
This section contains a list of auxiliary results invoked in the main proofs of \Cref{sec:convergence}.
\begin{lem}\label{thm:critical
Suppose that \Cref{ass:basic} holds, and let two sequences \(\seq{\bm u^k}\) and \(\seq{\bm v^k}\) satisfy \(\bm v^k\in\T(\bm u^k)\) for all \(k\) and be such that both converge to a point \(\bm u^\star\) as \(k\to\infty\).
Then, \(\bm u^\star\in\T(\bm u^\star)\), and in particular \(0\in\hat\partial\Phi(\bm u^\star)\).
\begin{proof}
Since \(\nabla F\) is continuous, it holds that \(\Fw{\bm u^k}\to\Fw{\bm u^\star}\) as \(k\to\infty\).
From outer semicontinuity of \(\prox_G^{\Gamma^{-1}}\) \cite[Ex. 5.23(b)]{rockafellar2011variational} it then follows that
\[
\bm u^\star
{}={}
\lim_{k\to\infty}
\bm v^k
{}\in{}
\limsup_{k\to\infty}
\prox_G^{\Gamma^{-1}}(\Fw{\bm u^k})
{}\subseteq{}
\prox_G^{\Gamma^{-1}}(\Fw{\bm u^\star})
{}={}
\T(\bm u^\star),
\]
where the limit superior is meant in the Painlevé-Kuratowski sense, cf. \cite[Def. 4.1]{rockafellar2011variational}.
The optimality conditions defining \(\prox_G^{\Gamma^{-1}}\) \cite[Th. 10.1]{rockafellar2011variational} then read
\begin{align*}
0
{}\in{} &
\hat\partial\left(
G+\tfrac12\|{}\cdot{}-(\Fw{\bm u^\star})\|_{\Gamma^{-1}}^2
\right)(\bm u^\star)
{}={}
\hat\partial G(\bm u^\star)
{}+{}
\Gamma^{-1}\left(
\bm u^\star - (\Fw{\bm u^\star})
\right)
\\
{}={} &
\hat\partial G(\bm u^\star)
{}+{}
\nabla F(\bm u^\star)
{}={}
\hat\partial\Phi(\bm u^\star),
\end{align*}
where the first and last equalities follow from \cite[Ex. 8.8(c)]{rockafellar2011variational}.
\end{proof}
\end{lem}
\begin{lem
Suppose that \Cref{ass:basic} holds and that function \(G\) is convex.
Then, the following hold:
\begin{enumerate}
\item\label{thm:FNE}
\(\prox_G^{\Gamma^{-1}}\) is (single-valued and) firmly nonexpansive (FNE) in the metric $\|{}\cdot{}\|_{\Gamma^{-1}}$; namely,
\[
\|
\prox_G^{\Gamma^{-1}}(\bm u)
{}-{}
\prox_G^{\Gamma^{-1}}(\bm v)
\|_{\Gamma^{-1}}^2
{}\leq{}
\innprod{
\prox_G^{\Gamma^{-1}}(\bm u)
{}-{}
\prox_G^{\Gamma^{-1}}(\bm v)
}{
\Gamma^{-1}(\bm u-\bm v)
}
{}\leq{}
\|
\bm u
{}-{}
\bm v
\|_{\Gamma^{-1}}^2
\quad\forall\bm u,\bm v;
\]
\item\label{thm:MoreauGrad}
the Moreau envelope \(G^{\Gamma^{-1}}\) is differentiable with \(\nabla G^{\Gamma^{-1}}=\Gamma^{-1}(\id-\prox_G^{\Gamma^{-1}})\);
\item\label{thm:subdiffdist}
for every \(\bm x\in\R^{\sum_in_i}\) it holds that
\(
\dist(0,\partial\FBE(\bm x))
{}\leq{}
\tfrac{
N+\max_i\set{\gamma_iL_{f_i}}
}{
N\min_i\set{\sqrt{\gamma_i}}
}
\|\bm x-\T(\bm x)\|_{\Gamma^{-1}}
\);
\item\label{thm:TLip}\label{thm:contractive}
\(\T\) is \(L_{\bf T}\)-Lipschitz continuous in the metric $\|{}\cdot{}\|_{\Gamma^{-1}}$ for some \(L_{\bf T}\geq0\); if in addition \(f_i\) is \(\mu_{f_i}\)-strongly convex, \(i\in[N]\), then \(L_{\bf T}\leq 1-\delta\) for \(\delta=\frac1N\min_{i\in[N]}\set{\gamma_i\mu_{f_i}}\).
\end{enumerate}
\begin{proof}
\begin{proofitemize}
\item\ref{thm:FNE} and \ref{thm:MoreauGrad}~
See \cite[Prop.s 12.28 and 12.30]{bauschke2017convex}.
\item\ref{thm:subdiffdist}
Let \(D\subseteq\R^{\sum_in_i}\) be the set of points at which \(\nabla F\) is differentiable.
From the chain rule of differentiation applied to the expression \eqref{eq:FBEMoreau} and using assert \ref{thm:MoreauGrad}, we have that \(\FBE\) is differentiable on \(D\) with gradient
\[
\nabla\FBE(\bm x)
{}={}
\bigl[
\I-\Gamma\nabla^2F(\bm x)
\bigr]
\Gamma^{-1}
\bigl[
\bm x-\T(\bm x)
\bigr]
\quad
\forall\bm x\in D.
\]
Since \(D\) is dense in \(\R^{\sum_in_i}\) owing to Lipschitz continuity of \(\nabla F\), we may invoke \cite[Th. 9.61]{rockafellar2011variational} to infer that \(\partial\FBE(\bm x)\) is nonempty for every \(\bm x\in\R^{\sum_in_i}\) and
\[
\partial\FBE(\bm x)
{}\supseteq{}
\partial_B\FBE(\bm x)
{}={}
\bigl[
\I-\Gamma\partial_B\nabla F(\bm x)
\bigr]
\Gamma^{-1}
\bigl[
\bm x-\T(\bm x)
\bigr]
{}={}
\bigl[
\Gamma^{-1}-\partial_B\nabla F(\bm x)
\bigr]
\bigl[
\bm x-\T(\bm x)
\bigr],
\]
where \(\partial_B\) denotes the (set-valued) Bouligand differential \cite[\S7.1]{facchinei2003finite}.
The claim now follows by observing that
\(
\partial_B\nabla F(\bm x)
{}={}
\tfrac1N\blockdiag(\partial_B\nabla f_1(x_1),\dots,\partial_B\nabla f_N(x_N))
\)
and that each element of \(\partial_B\nabla f_i(x_i)\) has norm bounded by \(L_{f_i}\).
\item\ref{thm:TLip}~
Lipschitz continuity follows from assert \ref{thm:FNE} together with the fact that Lipschitz continuity is preserved by composition.
Suppose now that \(f_i\) is \(\mu_{f_i}\)-strongly convex, \(i\in[N]\).
By \cite[Thm 2.1.12]{nesterov2013introductory} for all $x_i,y_i\in\R^{n_i}$
\begin{equation}\label{eq:smoothStrcvx}
\langle\nabla f_i(x_i)-\nabla f_i(y_i),x_i-y_i\rangle\geq\tfrac{\mu_{f_i}L_{f_i}}{\mu_{f_i}+L_{f_i}}\|x_i-y_i\|^2+\tfrac1{\mu_{f_i}+L_{f_i}}\|\nabla f_i(x_i)-\nabla f_i(y_i)\|^2.
\end{equation}
For the forward operator we have
\begin{align*}
&
\|
(\id-\tfrac{\gamma_i}{N}\nabla f_i)(x_i)
{}-{}
(\id-\tfrac{\gamma_i}{N}\nabla f_i)(y_i)
\|^2
\\
{}={} &
\|x_i-y_i\|^2
{}+{}
\tfrac{\gamma_i^2}{N^2}
\|\nabla f_i(x_i)-\nabla f_i(y_i)\|^2
{}-{}
\tfrac{2\gamma_i}{N}
\innprod{x_i-y_i}{\nabla f_i(x_i)-\nabla f_i(y_i)}
\\
\overrel[\leq]{\eqref{eq:smoothStrcvx}}{} &
\Bigl(
1-\tfrac{\gamma_i^2\mu_{f_i}L_{f_i}}{N^2}
\Bigr)
\|x_i-y_i\|^2
{}-{}
\tfrac{\gamma_i}{N}
\Bigl(
2-\tfrac{\gamma_i}{N}(\mu_{f_i}+L_{f_i})
\Bigr)
\innprod{\nabla f_i(x_i)-\nabla f_i(y_i)}{x_i-y_i}
\\
{}\leq{} &
\left(1-\tfrac{\gamma_i^2\mu_{f_i}L_{f_i}}{N^2}\right)
\|x_i-y_i\|^2
{}-{}
\tfrac{\gamma_i\mu_{f_i}}{N}
\left(2-\tfrac{\gamma_i}{N}(\mu_{f_i}+L_{f_i})\right)
\|x_i-y_i\|^2
\\
{}={} &
\left(1-\tfrac{\gamma_i\mu_{f_i}}{N}\right)^2
\|x_i-y_i\|^2,
\end{align*}
where strong convexity and the fact that $\gamma_i<\nicefrac{N}{L_{f_i}}\leq\nicefrac{2N}{(\mu_{f_i}+L_{f_i})}$ was used in the second inequality.
Multiplying by $\gamma_i^{-1}$ and summing over $i$ shows that \(\id-\Gamma\nabla F\) is \((1-\delta)\)-contractive in the metric \(\|{}\cdot{}\|_{\Gamma^{-1}}\), and so is \(\T=\prox_G^{\Gamma^{-1}}\circ(\Fw{})\) as it follows from assert \ref{thm:FNE}
\qedhere
\end{proofitemize}
\end{proof}
\end{lem}
The next result recaps an important property that the FBE inherits from the cost function \(\Phi\) that is instrumental for establishing global convergence and asymptotic linear rates for the BC-\Cref{alg:BC}.
The result falls as special case of \cite[Th. 5.2]{yu2019deducing} after observing that
\[
\FBE(\bm x)
{}={}
\inf_{\bm w}\set{
\Phi(\bm w)
{}+{}
D_H(\bm w,\bm x)
},
\]
where
\(
D_H(\bm w,\bm x)
{}={}
H(\bm w)-H(\bm x)-\innprod{\nabla H(\bm x)}{\bm w-\bm x}
\)
is the Bregman distance with kernel \(H=\tfrac12\|{}\cdot{}\|_{\Gamma^{-1}}^2-F\).
\begin{lem}[{\cite[Th. 5.2]{yu2019deducing}}]\label{thm:loja
Suppose that \Cref{ass:basic} holds and for \(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\), \(i\in[N]\), let \(\Gamma=\blockdiag(\gamma_1\I_{n_1},\dots,\gamma_N\I_{n_N})\).
If \(\Phi\) has the KL property with exponent \(\theta\in(0,1)\) (as is the case when \(f_i\) and \(G\) are semialgebraic), then so does \(\FBE\) with exponent
\(
\max\set{\nicefrac12,\theta}
\).
\end{lem}
\ifaccel
\begin{lem}[FBE: convexity and block-smoothness]\label{thm:convex
Suppose that \Cref{ass:basic,ass:Fast} are satisfied, and consider the notation introduced therein.
Let \(\gamma_i\in(0,\nicefrac{N}{L_{f_i}})\) be fixed.
Define
\(
Q_i
{}\coloneqq{}
\gamma_i^{-1}\I-\tfrac{1}{N}H_i\in\R^{n_i\times n_i}
\),
\(
Q
{}\coloneqq{}
\blockdiag(Q_1,\dots,Q_N)
\),
and
\(
H
{}\coloneqq{}
\tfrac1N\blockdiag(H_1,\dots,H_N)
\).
Then, $\FBEC \coloneqq\FBE \circ Q^{-1/2}$ is convex and smooth with $\nabla\FBEC (\tilde{\bm x}) = Q^{1/2}(\bm x-\T(\bm x))$ where $\bm x=Q^{-1/2}\tilde{\bm x}$.
In fact, for any \(\tilde{\bm x},\tilde{\bm x}'\in\R^{\sum_in_i}\) it holds that
\begin{equation}\label{eq:FBEComposedsmooth}
0
{}\leq{}
\innprod{\nabla\FBEC(\tilde{\bm x}')-\nabla\FBEC(\tilde{\bm x})}{\tilde{\bm x}'-\tilde{\bm x}}
{}\leq{}
\|\tilde{\bm x}'-\tilde{\bm x}\|^2.
\end{equation}
In particular, function $\FBEC$ is $1$-smooth along each block $i\in[N]$. If, additionally, all functions \(f_i\) are strongly convex, then
$\FBEC$ is \(\sigma\)-strongly convex with $\sigma\coloneqq \tfrac{1}{N}\min_{i\in [N]}\left\{\gamma_i\mu_{f_i}\right\}$.
\begin{proof}
Since $\gamma_i<N/L_{f_i}$, $Q$ is positive definite.
We begin by showing that for any \(\bm x,\bm x'\in\R^{\sum_in_i}\) it holds that
\begin{equation}\label{eq:FBEsmooth}
0\leq \|\bm x'-\bm x\|^2_{Q}- \|Q(\bm x'-\bm x)\|^2_{\Gamma}
{}\leq{}
\innprod{\nabla\FBE(\bm x')-\nabla\FBE(\bm x)}{\bm x'-\bm x}
{}\leq{}
\|\bm x'-\bm x\|^2_Q.
\end{equation}
It follows from \Cref{thm:MoreauGrad}, the chain rule of differentiation applied to \eqref{eq:FBEMoreau}, and the twice continuous differentiability of \(F\) that \(\FBE\) is continuously differentiable with
\(
\nabla\FBE(\bm x)
{}={}
Q(\bm x-\bm z)
\).
For \(\bm z^x\coloneqq\T(\bm x)\) and \(\bm z^{x'}\coloneqq\T(\bm {x'})\) it holds that
\begin{equation}\label{eq:innprodGrad}
\innprod{\nabla\FBE(\bm {x'})-\nabla\FBE(\bm x)}{\bm {x'}-\bm x}
{}={}
\innprod{
Q(\bm {x'}-\bm z^{x'}-\bm x+\bm z^x)
}{
\bm {x'}-\bm x
}
{}={}
\|\bm {x'}-\bm x\|^2_Q
{}-{}
\innprod{
\bm z^{x'}-\bm z^x
}{
Q(\bm {x'}-\bm x)
}.
\end{equation}
In order to bound the last scalar product, observe that
\[
0
{}\leq{}
\innprod{
\Gamma^{-1}(\bm z^{x'}-\bm z^x)
}{
(\bm {x'}-\Gamma\nabla F(\bm {x'}))
{}-{}
(\bm x-\Gamma\nabla F(\bm x))
}
{}\leq{}
\bigl\|
(\bm {x'}-\Gamma\nabla F(\bm {x'}))
{}-{}
(\bm x-\Gamma\nabla F(\bm x))
\bigr\|_{\Gamma^{-1}}^2,
\]
as it follows from \Cref{thm:FNE}.
Since \(\id-\Gamma\nabla F=\Gamma Q{}\cdot{} - \Gamma\bm q\) (with $\bm q\coloneqq(\tfrac1Nq_1,\dots,\tfrac1Nq_N)$), the above inequality simplifies to
\[
0
{}\leq{}
\innprod{
\bm z^{x'}-\bm z^x
}{
Q(\bm {x'}-\bm x)
}
{}\leq{}
\|\Gamma Q(\bm {x'}-\bm x)\|_{\Gamma^{-1}}^2,
\]
which combined with \eqref{eq:innprodGrad} results in the claimed \eqref{eq:FBEsmooth}.
If additionally \(\mu_{f_i}>0\) for all \(i\), then \(\FBE\) is \(1\)-strongly convex in the metric \(\|{}\cdot{}\|^2_{Q - Q\Gamma Q}\) (by observing that \(Q-Q\Gamma Q\succ 0\)). The result in \eqref{eq:FBEComposedsmooth} follows by using \eqref{eq:FBEsmooth} with the change of variables $\bm x=Q^{-1/2}\tilde{\bm x}$, $\bm x'=Q^{-1/2}\tilde{\bm x}'$ and noting that $\nabla \FBEC(\tilde{\bm x}) = Q^{-1/2}\nabla \FBE (\bm x)$.
Since $\Gamma$ is block-wise a multiple of identity it commutes with any block-diagonal matrix. Therefore, when $f_i$ are strongly convex, using the lower bound in \eqref{eq:FBEsmooth} and the above change of variable we obtain that $\FBEC$ is strongly convex in the metric \(\|{}\cdot{}\|^2_{\I - \Gamma Q}\). The result follows by noting that $\I - \Gamma Q= \Gamma H$.
\end{proof}
\end{lem}
\fi
\end{appendix}
\ifarxiv
\bibliographystyle{plain}
\else
\phantomsection
\addcontentsline{toc}{section}{References}
\bibliographystyle{spmpsci}
\fi
|
\section{Introduction}
Time series forecasting is a critical analysis tool in retail~\cite{larson01,johnson15}, finance, and commerce, and has been extensively researched across communities for multiple decades~\cite{box64,box68,hyndman08,quinlan92,Seeger2016,FlunkertSG17,Araujo:2018} (See \cite{Faloutsos2018} for a recent survey).
However, with the remarkable success of deep learning in other classically hard tasks like speech recognition, the time series forecasting problem also requires revisiting.
We consider typical forecasting settings where we are given multiple series along different points in time. Each series at time $t$ is characterized by a real-valued output $y^i_t$ and a vector of input features $\vx^i_t$. Our goal is to predict the $y$ for $K$ future points of time for which we are given inputs $\vx^i_{T+1}, \ldots, \vx^i_{T+K}$.
Classical methods of time series analysis like ARIMA, exponential smoothing, and other state space models~\cite{hyndman08}, train separate parameters $\vtheta^i$ for each time series $i$ and can be considered as local models.
In contrast recent deep learning methods~\cite{FlunkertSG17,wen2017multi,Mukherjee2018} train shared parameters $\vtheta$ for predicting each $y$ as a function of a summary of each time series' previous observation.
Typically, a Recurrent Neural Networks (RNNs) is used for representing such sequential history as fixed dimensional vectors. For sequences representing time series, in addition to general purpose gated RNNs like LSTMs, recent proposals also include un-gated RNNs like Statistical Recurrent Units (SRUs~\cite{OlivaPS17}) and Fourier Recurrent Units (FRU~\cite{ZhangLSD18}). During training, data from multiple time series teaches the RNN to capture enough context from the history relevant for making predictions at
each of the future $K$ points in time.
Such models have been found to surpass purely local models like ARIMA and conventional globally trained models like boosted regression trees~\cite{quinlan92}.
In general however such global model may be sub-optimal when the various series are heterogeneous, or when not all time series are observed during training. The only local influence in these models is the RNN state
that evolves with time separately for each time series $i$.
We take a different view and treat the per-series state as {\em local parameters} that map a future input along with its context to a prediction. We use a clever trick to update the local parameters in closed form, and in a streaming manner much like normal RNN parameters. We call our unit a Adaptive Recurrent Unit (ARU) whose finite-sized state
can be used to compute
per-series local parameters in closed form.
Recently, other methods have also been proposed to augment a global time series model with local parameters\cite{Rangapuram2018,Goel2017}.
However, their method of local adaptation requires the global model to be retrained for each update to a time series.
More generally, localization of a globally trained model can be treated as a domain adaptation problem~\cite{bendavid06Analysis,BlitzerMP06} for which many methods have been proposed including model fine-tuning, meta-learning~\cite{Finn2017,ravi17,mishra2018a}, and memory-augmented networks~\cite{SantoroBBWL16,Shankar2018}. These methods
either require gradient-based iterative updates, or are memory-intensive, and/or rely on techniques like self-attention that is
quadratic in $T$ for large time-series.
ARU does not entail iterative training of local parameters, its adaptation is light-weight and streaming, and only requires a time-series specific state of constant size.
This makes it particularly efficient for long time-series since its storage and compute requirement is constant in the sequence length.
The core principle of ARU is to use RNN-like updates to incrementally maintain per-series sufficient statistics required for fitting a local conditional Gaussian distribution in closed form. Our contribution is in exploiting this classical trick in the context of modern deep learning models, and designing network architectures that provide the right local-global tradeoff. Surprisingly, we show that this strategy of local adaptation is more robust across different types of time-series than recently proposed more ambitious methods that tax the deep network to learn to generate local parameters from RNN states.
We compare our ARU-based adaptation to two state-of-art adaptation methods on five time-series datasets under various settings. We show that ARU is effective in reducing error by 10 to 20\% compared to the baseline, and is generally better than existing adaptation methods. This reduction in error comes at very little overheads of running time, whereas existing state-of-the-art adaptation methods are up to a factor of four slower.
\newcommand{{\mathcal N}}{{\mathcal N}}
\renewcommand{\vv}{{\vx}}
\newcommand{{\vek{s}}}{{\vek{s}}}
\section{Our Model}
We start with a review of state-of-the-art global models and classical local models for time series forecasting.
\subsubsection*{Review: Global Models}
Typical deep learning based global models for multi-horizon time series forecasting~\cite{FlunkertSG17,wen2017multi} deploy the encoder-decoder architecture.
First an input layer maps the input features $\vx^i_t$ to a real vector.
This could include embedding lookups for categorical attributes and any rescaling for continuous attributes. Next the transformed input along with the previous
output
$[\vv^i_t, y_{t-1}^i]$ is fed to one or more encoder RNN layers.
The RNN could be either a generic LSTM or time-series specific units like SRU or FRU that capture
context useful for future predictions.
The output of the encoder is its final state ${\vek{g}}^i_T$ at the end of $T$ steps. This can be treated as a summary of the known $y$ values that is relevant as a {\em context} for future predictions.
The decoder initialized with ${\vek{g}}^i_T$ is responsible for making the predictions on the $K$ future time horizons as a function of respective inputs $\vx^i_{t}: t=T+1\ldots,T+K$. The decoder could be auto-regressive,
where
the previous predicted $y$ is fed as input to be next step, or independently make each of the $K$ predictions. We found the independent model to provide higher accuracy for long-term forecasts than the auto-regressive model fed with noisy previous predictions. This is corroborated in \cite{wen2017multi}.
The decoder
takes each transformed future input $\vv^i_{t}: t=T+1\ldots,T+K$ concatenated with ${\vek{g}}^i_T$ and generates an output vector $\vh^i_t$ using an optional RNN and one or more feed-forward layers. Lastly, a Gaussian distribution is imposed on the output by using a linear layer to map $\vh^i_t$ into a mean and variance. The final equations driving the prediction at future times are:
\begin{eqnarray}
\label{eq:globalmu}
& &{\vek{g}}^i_T = RNN([y^i_{t-1},\vx^i_t]:t=1\ldots T|\theta_{\text{enc}}) \\
& & \vh^i_t = FF([{\vek{g}}^i_T, \vx^i_t]:t=T+1\ldots T+K | \theta_{\text{dec}}) \\
& &\mu^i_t =\theta_\mu[\vh^i_t,1],~~~\sigma^i_t = \log(1+\exp(\theta_\sigma[\vh^i_t,1])) \\
& &\Pr(y^i_t | \vx^i_t, (\vx^i_1,y^i_1),\ldots, (\vx^i_T,y^i_T)) = {\mathcal N}(\mu^i_t, \sigma^i_t)
\end{eqnarray}
where $\theta_{\text{enc}}, \theta_{\text{dec}}, \theta_\mu, \theta_\sigma$ are all parts of the global parameters $\vtheta$.
\begin{figure}
\centering
\input{hybrid_diag}
\caption{Diagram of the global model with encoder size of $3$ and decoder size of $3$. All the nodes/states with the same label are copies.}
\label{Fig:global_model}
\end{figure}
During training we are given data comprising of
input-output
pairs over several time series. $D=\{(\vx^i_1,y^i_1),\ldots,(\vx^i_T,y^i_T): i = 1 \ldots, N\}$. We simulate multiple encoder-decoder windows from this history using a sliding window of stride $\ell$ and compute the data likelihood as follows:
$$\max_\vtheta \sum_{i=1}^N \sum_{T'=2:\ell}^{T-K}\sum_{t=T'+1}^{T'+K}\log {\mathcal N}(y^i_t ; \mu^i_t, \sigma^i_t | \vtheta)$$
where ${\mathcal N}$ denotes the Gaussian density function. The parameter $\vtheta$ includes all network parameters spanning the input-layer, the encoder RNN, the decoder RNN, and the last output layer that generates $\mu^i_t, \sigma^i_t$. Figure \ref{Fig:global_model} shows the global model.
The global model is driven by parameters trained across multiple
time-series, and the only local influence in the final equation that
outputs
$\Pr(y^i_t|\vx^i_t, \lbrace(\vx^i_t,y^i_t)\rbrace_{t=1}^{T})$
is the RNN state ${\vek{g}}^i_T$ that serves as
context. The RNN parameters are trained end-to-end to find the most
relevant context.
\subsubsection*{Review: Local Models}
A local model would train separate parameters $\theta^i$ for each $i$.
However, unless the length of each time-series is very large, training
parameters of a complex network like a multi-layer deep network would
not work. Hence, local models have traditionally been simple models
such as linear state space models that use linear parameters to
transition from one state to the next, and make state-ful predictions. Let $\vs_t^i$ denote the local state at time $t$ of series $i$. Parameters $\theta^i_{\text{tr}}$ control transition from one state to the other via affine transforms, and parameter $\theta^i_\mu$ controls the linear transform to generate the mean output from the local state. The variance is a fixed learned parameter, and that makes the local model as:
\begin{eqnarray}
\label{eq:local}
{\vek{s}}^i_t & = \theta^i_{\text{tr}} [{\vek{s}}^i_{t-1}~ \vx^i_t ~1] \\
\mu^i_t & =\theta^i_\mu[{\vek{s}}^i_t~1],~~~\sigma^i_t=\theta^i_\sigma
\end{eqnarray}
where the parameters $\theta^i_{\text{tr}}$ and $\theta^i_\mu,
\theta^i_\sigma$ are learned locally for each time-series using its known values up to $T$ that is $(\vx^i_1,y^i_1),\ldots, (\vx^i_T,y^i_T)$. Assuming $T$ is large, local
models can capture peculiarities of individual time-series in ways
that cannot be approximated by local RNN state of global models.
\subsection{The ARU RNN}
Our model not only attempts to capture the best of the local and global models,
but does so in a manner that allows the local parameters to be
computed recurrently in a streaming manner. We call our unit Adaptive
Recurrent Unit (ARU) that views the per-series state as local
predictive parameters that are fitted on the fly.
Additionally, ARU
is designed to capture local predictive parameters without expensive
parameter retraining.
We use the shared global network to learn the complex transformation of the history and the input into a final output vector $\vh^i_t$. However, instead of the global last linear layer (Equation~3 in the global model), we use a local model. We perform the local fit in a clever way to allow easy end-to-end training of the global network parameters without inner loop of local updates as in existing methods like \cite{Finn2017}.
The main idea behind the ARU is to use a RNN to provide the best possible least square fit locally for each time series based on observed $\vx_t,\vh_t$ values.
The motivation behind this choice is that the optimal local parameters can be obtained in closed form using sufficient statistics that can be maintained in a streaming manner. In contrast local parameters that depend on iterative gradient-based updates, require multiple data passes and are difficult to embed within a larger global network that is trained via gradient descent.
Although the ARU unit itself only models linear interaction between
its inputs and outputs, it is embedded in a larger neural network that
provides non-linear transformation of both the input to the ARU and
its output.
We describe the ARU-RNN in this section
and then in Section~\ref{sec:aru_in_global_model} describe how we fit
it in the global model.
\newcommand{\vek{sxx}^{i}}{\vek{sxx}^{i}}
\newcommand{\vek{sxy}^{i}}{\vek{sxy}^{i}}
\newcommand{\vek{ss}^{i}}{\vek{ss}^{i}}
\newcommand{\vek{sn}^{i}}{\vek{sn}^{i}}
\subsubsection{The ARU Update Equations}
Let $\vh^i_t$ be the output vector at time $t$ from the global model.
The ARU RNN has two modes: adapt and predict. In the adapt mode it is fed the input $\vh^i_t$ along with the true label $y^i_t$ at that time and produces an updated state $\vs^i_t$. In the predict mode denoted $ARU(\vs^i_{t-1}, \vh_t^i)$ it outputs the local prediction based on state $\vs^i_{t-1}$ on the input $\vh_t^i$.
The ARU state $\vs_t$ at each time $t$ keeps four types of sufficient statistics $\vs_t=[\vek{sxx}^{i}_t,\vek{sxy}^{i}_t,\vek{sn}^{i}_t,\vek{ss}^{i}_t]$ to provide a local least square fit between the $\vh$ and $y$s. A vector $\valpha$ of $J$ aging factors maintains this statistics for varying amounts of aging of the old data. Initially, all states are zero and at each time-step it is updated as follows:
ARU\_Update($\vek{s}^i_{t-1}, \vh^i_t, y^i_t):$
\begin{align*}
\vek{sxx}^{i}_t &= \valpha\,\vek{sxx}^{i}_{t-1} + [\vh_t^i~1]^{\text{T}}[\vh_t^i~1] \\
\vek{sxy}^{i}_t &= \valpha\,\vek{sxy}^{i}_{t-1} + [\vh_t^i~1]^{\text{T}}(\vy_t^i) \\
\vek{sn}^{i}_t &= \valpha\,\vek{sn}^{i}_{t-1} + 1 \\
\vek{ss}^{i}_t &= \valpha\,\vek{ss}^{i}_{t-1} + (\vy_t^i-ARU(\vs^i_{t-1}, \vh_t^i))^2 \\
\text{Return} &~\vs^i_t = [\vek{sxx}^{i}_t,\vek{sxy}^{i}_t,\vek{sn}^{i}_t,\vek{ss}^{i}_t]
\end{align*}
From these update equations, it is easy to interpret what each of the ARU state components represent.
The state $\vek{sxx}^{i}$ represents age-weighted sum of pairwise feature product of the input vector, $\vek{sxy}^{i}$ represents age-weighted sum of product of input features and output $y$, $\vek{sn}^{i}_t$ represents the age adjusted count so far, and $\vek{ss}^{i}_t$ represents the accumulated fitting noise that will serve as the variance from the local predictions. In this equation $ARU(\vs^i_{t-1}, \vh_t^i)$ denotes the mean local prediction as explained below.
The ARU state is updated in a streaming manner every time a true $y^i_t$ is known.
\sloppy We next provide the equations of ARU in the predict mode. $ARU(\vs^i_{t-1}, \vh_t^i)$ represents the predicted values from the ARU at state $\vs^i_{t-1}$ on the input $\vh_t^i$.
The predicted output using only ARU states can be computed by exploiting the closed form of the least squares solution. First, the local parameter is calculated in closed form using the sufficient statistics. The $\lambda I$ term regularizes the local parameters. The local parameter is used to compute the mean and variance of the local prediction.
ARU\_Predict($\vek{s}^i_t, \vh^i_t$):
\begin{alignat}{2}
\label{Eqn: ARU_Predict}
\vtheta^i_{t,\mu} & = (\vek{sxx}^{i}_t + \lambda I)^{-1}\vek{sxy}^{i}_t , \quad\quad & \theta^i_{t,\sigma} & = \vek{ss}^{i}_t / \vek{sn}^{i}_t \\
\vm_t^i & = \vtheta^i_{t,\mu} [\vh_t^i 1], \quad\quad & \va_t^i & = \theta^i_{t,\sigma}
\end{alignat}
where $\vm_t^i$ denotes the mean local prediction and $\va_t^i$ denotes the local variance.
The ARU fits local parameters, which can be used to compute the mean
prediction (and variance) much like in the local model in Equation ~\ref{eq:local}. However, unlike in local models, we do not output these directly as predictions. Instead, we exploit the availability of multiple time series to further combine these local predictions with global parameters as follows.
\subsection{ARU in Global Model}
\label{sec:aru_in_global_model}
We concatenate the local mean and variance predictions with the final vector $\vh^i_t$ and use a two-layer feed forward network to transform them as follows.
\begin{equation}
\label{eq:aru_in_globalmu}
\mu^i_t=\theta_\mu (FF2 [\vh^i_t,\vm^i_t,1],~~~~~~~~\sigma^i_t = \log(1+\exp(\theta_\sigma(FF2[\vh^i_t,\va^i_t,1]))
\end{equation}
Figure \ref{Fig:aru_in_global_model} shows how ARU is embedded in the global model for one decoder step. We only show mean computation in the figure. The variance computation can also be handled similarly. The output $\vm_4$ obtained from ARU is combined with hidden layer and passed further to a two layer feed-forward network. The $FF2$ network further evaluates the importance of the global output $\vh^i_t$ and local ARU output $\vm^i_t$ and makes the final predictions.
\begin{figure}
\centering
\input{aru_diag.tex}
\caption{ARU cell combined with the decoder of the global model. Showing only the mean computation here.}
\label{Fig:aru_in_global_model}
\end{figure}
\section{Related Work}
Before we present our empirical evaluation we discuss related work that could have served as alternatives to our method of adaptation.
\paragraph{Generate Local Parameters from RNN State}
Recently \cite{Goel2017} and \cite{Rangapuram2018} propose to use the power of deep learning to directly compute local parameters $\vtheta^i$ from the local state
of the RNN.
As a state-of-art representative of this class of methods, we discuss the DeepState method of \cite{Rangapuram2018}. In DeepState the RNN state ${\vek{g}}^i_t$ computed from global parameters (Eq~\ref{eq:globalmu}) is passed through feed-forward networks to directly output the local parameters $\theta^i_{\text{tr}}$ and $\theta^i_\mu,
\theta^i_\sigma$, that is,
\begin{eqnarray}
[\theta^i_{\text{tr}}, \theta^i_\mu,
\theta^i_\sigma] = FF({\vek{g}}^i_t; \theta_\text{meta})
\end{eqnarray}
where $\theta_\text{meta}$ are additional global parameters. The
computed
local parameters are used to generate the mean and variance of the output $y$s using a linear state-space model (e.g. Equation~\ref{eq:local}). The global parameters are trained via joint likelihood on all training examples. Unlike our approach, these methods cannot adapt to new time series since local $y$-s are only transferred via the joint likelihood based training. In our experiment section we show that these models do not train well unless the length of each time-series is long.
\paragraph{Dedicated Local parameters jointly trained with Global} An approach often used for localization of global models is to dedicate per-locale parameters that play specific role in the global model and are trained jointly end-to-end. For example~\cite{Rei15} uses per-document parameter along with global parameters to develop better embeddings. In our case, we would train time-series specific parameters $\vtheta^i$ along with the rest of the global parameters. This method is only applicable when all time-series are known during training time. A downside of this approach is that it cannot be easily evolved when new values are observed for a time series. Also, each time series has to be long enough for the local parameters to train well.
\paragraph{Fine-tune Global Parameters for each series}
Another approach is to treat localization as a problem of domain adaptation, for which a huge literature exists~\cite{bendavid06Analysis,BlitzerMP06}. A well-known solution is to fine-tune the global parameters on labeled data of each time-series using gradient descent on the loss over the limited labeled data for each series~\cite{ravi17}.
This method would require storage of separate local parameters for each series. Also, parameter fine-tuning could lead to unpredictable local performance. This has led to an explosion of meta-learning methods that exploit multiple local datasets during training so as to learn the adaptation process~\cite{Finn2017,mishra2018a}.
Another method of adaptation is using memory~\cite{Shankar2018, rae2018fast}
that combines parameter fine-tuning with memory-based recall.
Of these a recent state-of-art approach is the SNAIL model that uses a simple method of learned deep self-attention on local data for adaptation. As a state-of-art representative of learned adaptation methods, we will compare with this method in our experiments. Our approach is closed-form self-attention that exploits the special regression form of our prediction function. Self-attention based models for time series forecasting are also explored in \cite{QinSCCJC17}.
For theoretical insights on the learning efficiency of global vs local models see \cite{Kuznetsov2019}.
\paragraph{Special RNNs for Time Series}
We designed ARU so that it can be embedded in a global network much like any other RNN. Recently, \cite{OlivaPS17} and \cite{ZhangLSD18} also propose special RNNs for time-series. Oliva et al. ~\cite{OlivaPS17} proposed an un-gated alternative to LSTMs, called the Statistical Recurrent Unit (SRU). SRUs maintain moving averages of summary statistics in their hidden states and generalize exponential moving average like statistics. However, they treat the SRU state like 'context' to be used for predictions in place of LSTM states. Fourier Recurrent Units FRUs~\cite{ZhangLSD18} is another such un-gated time-series unit. In contrast, we convert the ARU state to local parameters by exploiting the closed-form fit from the the least square sufficient statistics. \nocite{Li:2010}
\section{Experiments}
We now empirically compare our ARU-based streaming adaptation method to state-of-the-art deep learning based global models as baselines, and to two recent approaches of adapting them to local time-series. In addition to five real-life public time-series datasets, we also present controlled experiments on synthetic datasets to gain insights on different methods of localizing deep global models.
Our code and experiments can be found at \footnote{\url{https://github.com/pratham16/ARU.git}}.
This section is organized as follows: Section~\ref{sec:data} describes the datasets, Section~\ref{sec:methods} describes the five methods we compare with, Section~\ref{sec:setup} describes detailed setup of the experiments, Section~\ref{sec:synth} presents experiments on the synthetic dataset to qualitatively understand the difference between DeepState and ARU, and Section~\ref{sec:real} presents anecdotes, accuracy, and running times on the real datasets.
\begin{table}[]
\centering
\begin{tabular}{l|r|r|r|r|r}
Dataset & N & T & K & Enc.Len & \#Features \\ \hline
Rossman & 1115 & 1600 & 16 & 16 & 39 \\
Walmart & 3331 & 143 & 8 & 8 & 16 \\
Electricity & 370 & 44000 & 24 & 168 & 5 \\
Traffic & 963 & 2100 & 24 & 168 & 3 \\
Parts & 2246 & 52 & 8 & 8 & 1 \\
\end{tabular}
\caption{Statistics of the datasets used in our experiments. N is the number of time series, T is the number of values in a series, K is the forecast horizon we used in our experiments, Enc.Len is the length of the encoder in each prediction instance.}
\label{tab:data}
\end{table}
\subsection{Datasets}
\label{sec:data}
We perform our experiments on five publicly available time-series datasets. The first two datasets we collected from Kaggle contains rich $\vx$ features at each point and are more aligned with typical retail forecasting setups where the deep models have gained most traction recently. The remaining three contain only derived time features and are included for comparison with the existing literature. A summary of the datasets appears in Table~\ref{tab:data}.
\myparagraph{Rossman}
The Rossmann dataset\footnote{\url{https://www.kaggle.com/c/rossmann-store-sales/data}} contains daily sales history
of 1115 Rossmann stores. The data is collected between 1st Jan. 2013 to 31st July 2015, i.e., roughly 900 values per series. Each time series is associated with various co-variate features.
These
include predetermined store-specific features such as store type and distance to the nearest competing store and time-varying features such as whether the store is running any promotion, and external features such as weather conditions, is current day a school holiday or a state holiday, etc.
\myparagraph{Walmart}
The Walmart\footnote{\url{https://www.kaggle.com/c/walmart-recruiting-store-sales-forecasting/data}} dataset maintains weekly sales in each department of 45 walmart stores from Feb 2010 to Oct 2010. Here, (store, department) tuple uniquely identifies each time-series, resulting in 3331 time-series. Each time-series contains roughly 143 values.
This dataset also provides promotional features and external features such as temperature, unemployment, consumer price index, major holidays in a given week etc.
\myparagraph{Electricity}
The electricity\footnote{https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014} dataset contains hourly energy consumption of 370 houses (in kW) from 1st Jan. 2011 to 31st Dec. 2015, a total of 44,000 values per time-series.
\myparagraph{Traffic}
The Traffic\footnote{https://archive.ics.uci.edu/ml/datasets/PEMS-SF} dataset contains hourly occupancy rates of 963 car lanes in San Francisco bay area. The occupancy rates are in the range $[0, 1]$. The data is collected between 1st Jan. 2008 to 30th March 2009, i.e., 10560 values per series.
\myparagraph{Parts}
This dataset \cite{chapados2014effective} contains monthly sales information of 2246 car parts over 52 months. This is a small dataset compared to others in our collection.
\paragraph{Features}
For each dataset, we extract time-related features in addition to any features available in the data (e.g. in Rossman and Walmart).
For Walmart, we use month of the year and week of the year features. For Rossman, we use day of the month and month of the year features. Since Parts is a monthly dataset, we use month of the year feature for Parts. Since Electricity and Traffic are hourly datasets, we use hour of the day and day of the week features. In addition to these, we use month of the year feature for electricity dataset.
We embed categorical features into real vectors and rescale the real valued attributes between 0 and 1 across the time-series.
\begin{table}[]
\centering
\begin{tabular}{|l|r|c|}
\hline
Set & \# RNN units & Hidden layer sizes \\ \hline
{\tt Small} & 8 & 8,6,6 \\
{\tt Medium} & 16 & 16,15,10 \\
{\tt Large} & 50 & 32,20,15 \\ \hline
\end{tabular}
\caption{Hyperparameter sets}
\label{tab:hyperparameters}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=0.4\textwidth]{synthetic_plots/SyntheticVaryTWithTsId_large_w.eps}
\includegraphics[width=0.4\textwidth]{synthetic_plots/SyntheticVaryTWithTsId_small_w.eps}
\includegraphics[width=0.4\textwidth]{synthetic_plots/SyntheticVaryTWithoutTsId_large_w.eps}
\includegraphics[width=0.4\textwidth]{synthetic_plots/SyntheticVaryTWithoutTsId_small_w.eps}
\caption{Performance of ARU and DeepState on synthetic data. The columns correspond to different values of $\gamma$ ($\gamma=20$ first column, $\gamma=1$ second column). The first row are with time-series ids and second row are without time-series ids.}
\label{fig:synth}
\end{figure*}
\subsection{Methods}
\label{sec:methods}
We compare ARU against two state-of-the-art global deep learning models as baselines and two most recent
local adaptation methods.
We drop comparison with classical local-only models such as ARIMA because the DeepState and DeepAR methods that we compare with have been shown to conclusively surpass them.
\myparagraph{Baseline}
As a baseline we use the globally trained model where the decoder makes independent predictions for each
of the future $K$ points in time.
Note that the encoder used to compute the per-series state ${\vek{g}}^i_T$ continues to be auto-regressive.
Our experiments and \cite{wen2017multi} have found the independent decoder to yield
better performance
than an auto-regressive decoder. This baseline is part of our code base and uses exactly the same setup as ARU.
\myparagraph{DeepAR}
As another global model, we use
DeepAR, a production quality deep forecasting model that is built-in Amazon's SageMaker machine learning service \footnote{\url{https://docs.aws.amazon.com/sagemaker/latest/dg/deepar.html}}. As described in \cite{FlunkertSG17}, DeepAR is also an encoder-decoder model but their decoder is auto-regressive. Since the source code is not public, further details of specific features, scaling, etc that they deploy is unknown. To the best of our ability we tried to match the features and experiment setup with our global baseline but we cannot be sure without access to the source code.
\myparagraph{SNAIL} The SNAIL method proposed in \cite{mishra2018a} is a local adaptation model that uses interleaving of dilated causal convolution and self-attention layers to capture long range dependencies and for localization respectively. We use two convolution layers with dilation rates 2 and 4 respectively, and two self-attention layers.
\myparagraph{DeepState} This model uses an RNN to predict the parameters of a local state-space model \cite{Rangapuram2018} and is currently the
best known local adaptation model for time-series forecasting.
The source code is not publicly available, so we only compare with their published numbers.
\subsection{Experiment setup}
\label{sec:setup}
We use a single RNN layer in the encoder. The decoder has three ReLU layers on a concatenation of the encoder state ${\vek{g}}^i_T$ and input features $\vx^i_t$. A skip connection from $\vx^i_{T+k}$ to the second ReLU layer is added.
The batch size for all experiments is set to 64. Adam optimizer is used with learning rate 0.0001. ARU regularization $\lambda$ and aging vector $\valpha$ are chosen based on validation loss.
In all our experiments, $\valpha \in \lbrace 1.0, 0.99, 0.95, 0.9 \rbrace^J$. In all experiments, best performing value of $\valpha$ turns out to be $[1.0]$, except for Traffic dataset in Table \ref{tab:deepstate}, where best performing $\valpha = [1.0, 0.95, 0.9]$. For number of RNN units and number of hidden units, we use three sets of hyper-parameter settings for baseline and ARU -- {\tt Small},
{\tt Medium},
and {\tt Large} as given in Table \ref{tab:hyperparameters}. We use the {\tt Small} configuration for Parts since it contains time-series of length 52. For Rossman and Walmart, we use the
{\tt Medium}
configuration. For Electricity and Traffic datasets, we use the {\tt Large} configuration in Table \ref{tab:deepstate} and the
{\tt Medium}
configuration in the rest of the experiments.
We rescale $y$ values of each time-series with its average $y$ value as follows:
$$ {\overline{y}}^i_t = 1 + \frac{1}{T} \sum_{t'=1}^T y^i_{t'} $$
This rescaling is also used by \cite{FlunkertSG17}. On the Walmart dataset since each time-series is small, we perform above rescaling on a batch instead of the entire time-series i.e.\ $ {\overline{y}}^i_t = 1 + \frac{1}{E} \sum_{t'=1}^{E} y^i_{t'} $ where $E$ is encoder length.
We also experimented with SRU \cite{OlivaPS17}, which is an un-gated alternative to LSTMs. Since we obtained similar results with SRUs, we omit those numbers from our experiments.
In addition to RMSE, we report Normalized Deviation (ND), which is used in \cite{FlunkertSG17, yu2016temporal} as an evaluation metric. It is defined as:
$$ \text{ND}(y^i_t, \hat{y}^i_t) = \bigg( \sum_{i=1}^N \sum_{t=T+1}^{T+K} |y^i_t - \hat{y}^i_t| \bigg) \bigg/ \bigg( \sum_{i=1}^N \sum_{t=T+1}^{T+K} |y^i_t| \bigg) $$
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{example_plots/rossman_7.eps}
\includegraphics[width=0.45\textwidth]{example_plots/rossman_8.eps}
\includegraphics[width=0.45\textwidth]{example_plots/rossman_2.eps}
\includegraphics[width=0.45\textwidth]{example_plots/rossman_3.eps}
\caption{Examples of four time series from the Rossman dataset showing how well our ARU is able to track the local patterns of each series compared to the global baseline.}
\label{Fig:example_rossman}
\end{figure*}
\subsection{Qualitative Comparison on Synthetic Data}
\label{sec:synth}
In order to get an insight into
the working of ARU vis-a-vis the most recent DeepState method of local adaptation, we first present a qualitative comparison using a very simple synthetic dataset.
We generate 10 synthetic time-series of hourly data. The $y$ values for these time-series are obtained as a linear combination of $x$ features and local parameters $\vtheta_\mu^i$, which are uniformly sampled in the range $[-\gamma,\gamma]$. We use only hour of day and day of week as $\vx$ features and generate time-series of different lengths
by sampling $y \sim \vtheta_\mu^i [\vx_t~1] + \epsilon$ where $\epsilon \sim \mathcal{N}(0, 1)$ is a Gaussian noise with unit variance.
We compute RMSE for predictions on last 24 hour forecasts in each time-series. This non-auto-regressive model might look too simplistic but it helps us bring out the essence of different approaches.
Also, after conditioning on the encoder-state, we found non-auto-regressive decoder to be superior to the auto-regressive decoder even on real-life datasets.
In this simple case, the state-space model would need to learn to generate $\vtheta_\mu^i$ from its RNN state ${\vek{g}}^i_t$ in order to match the generative process exactly. In contrast, the ARU would compute
$\vtheta_\mu^i$ from the data whose sufficient statistics is summarized in the ARU state. The global network just needs to learn to provide the right input to the ARU cell and also correctly integrate its output with the global output.
We used the {\tt Medium} setting (Table~\ref{tab:hyperparameters}) for the global network without any special customization for the synthetic experiments. Thus, the global network needs to learn the correct meta-learning strategy for both methods of adaptation.
When id $i$ is part of $\vx_t$, in theory, a powerful enough network can memorize each set of the 10 local parameter in its feed-forward layers, making them both identical. We therefore compare these models under two settings: one where each time-series $i$ feeds its unique id $i$ to the network, and second where it does not.
We compare the two methods against increasing length of time-series in Figure~\ref{fig:synth} for two different values of $\gamma$ (20 and 1) both with and without time-id. When $\gamma$ is large (first column) the local parameters differ a lot from one time series to the next and the DeepState method of computing them from the RNN state is not at all effective. With increasing sequence length, errors of both methods reduce but ARU converges much faster to error values less then 0.1. When local time-series are more similar to each other ($\gamma=1$), DeepState starts matching ARU only after seeing many samples per time-series ($\gamma=1$, with time-series id). Also, parameter generation methods like DeepState also crucially exploit the time-series specific identifiers to generate local parameters (compare red plots across the two rows). This implies that they memorize local parameters of the finite number of training time-series, and cannot perform adaptation of a series not seen during training. Also, they cannot exploit streaming availability of labeled data after training.
\begin{table}[ht]
\centering
\begin{tabular}{|l|l|r|r|} \hline
Dataset & Method & Normalized & RMSE \\
& & Deviation (ND) & \\ \hline
Rossman & Baseline & 0.0987 & 1020.7 \\
& DeepAR & 0.2273 & 2197.4 \\
& SNAIL & 0.0963 & 993.4 \\
& ARU & {\bf 0.0851} & {\bf 908.6} \\ \hline
Walmart & Baseline & 0.1187 & 3585.8 \\
& DeepAR & 0.2215 & 5894.9 \\
& SNAIL & 0.1082 & 3465.0 \\
& ARU & {\bf 0.1058} & {\bf 3413.3} \\ \hline
Electricity & Baseline & 0.1264 & 377.3 \\
& DeepAR & 0.1838 & 530.9 \\
& SNAIL & 0.1307 & 363.2 \\
& ARU & {\bf 0.1260} & {\bf 344.6} \\ \hline
Traffic & Baseline & 0.1991 & 0.0374 \\
& DeepAR & {\bf 0.1894} & 0.0367 \\
& SNAIL & 0.2187 & 0.0380 \\
& ARU & 0.1964 & 0.0372 \\ \hline
Parts & Baseline & 1.5763 & 1.300 \\
& DeepAR & {\bf 1.4440} & {\bf 1.268} \\
& SNAIL & 1.4910 & 1.276 \\
& ARU & 1.5409 & 1.285 \\ \hline
\end{tabular}
\caption{Performance of different methods on the datasets in the fixed mode.}
\label{tab:overall}
\end{table}
\subsection{Comparisons on Real Data}
\label{sec:real}
We start by showing some anecdotes, then compare different methods on prediction error in two different settings, and finally compare them on running time.
\subsubsection{Anecdotes} In Figure \ref{Fig:example_rossman}, we show how ARU enables a global model to adapt to the
characteristics
of each time series from the Rossman dataset. The vertical line denotes the prediction horizon. The ARU model is able to match the peaks of the data much better than the global model. The location and pattern of peaks in the actual data varies a lot across the four time-series and a single global model is unlikely to be able to capture them. The ARU by intervening at the last layer, with local sufficient statistics, is able to learn the relation between the inputs and peaks
more accurately.
For example, look at the two peaks after the green line in the first plot. For the time-series in the second row, first column we see how the global model's peak prediction is offset by one
step
whereas ARU aligns perfectly with the true $y$s. In the last time-series (second row, second column), we see how nicely the gradual decline is matched by ARU, whereas the global baseline is needlessly jittery.
\subsubsection{Quantitative comparison on prediction error}
Next we move to a quantitative comparison. We first compare in a fixed horizon setting where we predict for a fixed horizon
$T+1,\ldots,T+K$
that is known during training ($T$ and $K$ values summarized in Table~\ref{tab:data}).
In Table~\ref{tab:overall} we compare different methods for a fixed horizon on the five datasets on four methods.
The DeepState method is compared separately since
we could not get
access to the DeepState code base.
We observe that on feature-rich datasets like Rossman and Walmart ARU is significantly better than all existing methods.
Another property of ARU is that even when it does not improve beyond the baseline it is not much worse. This is because we meta-learn how to combine ARU's local predictions with the global predictions.
We next compare different methods
in
a streaming setting
where we provide the true $E$ (encoder length) forecasts after the training time period $T$, and ask for prediction for the next $K$ time periods.
We do not retrain model parameters to simulate a true online deployment setting. For the electricity and traffic datasets, we use the rolling-window forecast setting used in \cite{Rangapuram2018} and \cite{yu2016temporal}. Hourly values of 7 days are input and we make prediction for the next day. Then the forecast window is shifted by one day and we repeat this for the next 7 days. Since the frequency of the Walmart dataset is weekly, we make rolling-window forecasts for 16 weeks, with window size 8. For the Rossman data, forecast window size is 16 days, and we make forecasts for 32 days.
\begin{table}[ht]
\centering
\begin{tabular}{|l|l|r|r|} \hline
Dataset & Method & Normalized & RMSE \\
& & Deviation (ND) & \\ \hline
Rossman & Baseline & 0.094 & 983.2 \\
& DeepAR & 0.245 & 2389.0 \\
& SNAIL & 0.093 & 972.6 \\
& ARU & {\bf 0.089} & {\bf 934.9} \\ \hline
Walmart
& Baseline & 0.137 & 4548.6 \\
& DeepAR & 0.233 & 6704.0 \\
& SNAIL & 0.144 & 4690.5 \\
& ARU & {\bf 0.114} & {\bf 3938.6} \\ \hline
Electricity
& Baseline & 0.136 & 416.7 \\
& DeepAR & 0.172 & 544.0 \\
& SNAIL & 0.135 & 400.6 \\
& ARU & {\bf 0.127} & {\bf 396.4} \\ \hline
Traffic
& Baseline & 0.170 & 0.0224 \\
& DeepAR & {\bf 0.145} & {\bf 0.0216} \\
& SNAIL & 0.165 & 0.0227 \\
& ARU & 0.161 & 0.0220 \\
\hline
\end{tabular}
\caption{Performance of different methods in the streaming mode.}
\label{tab:stream}
\end{table}
In Table~\ref{tab:stream}, we summarize the results on four datasets. The car parts dataset was dropped since each time-series is too small for streaming. We observe that even in this mode, ARU provides significant gains over all existing methods. A close second is the SNAIL method of adaptation but we will see later that SNAIL incurs significant runtime overheads.
\begin{table}[]
\centering
\begin{tabular}{|l|l|l|l|l|} \hline
Dataset & Setting & DeepAR & DeepState & ARU \\
& Enc Len-$K$ & & & \\
\hline
Parts & 12-12 & 1.245 & 1.470 & 1.318 \\
Electricity & 336-168 & 0.199 & 0.087 & 0.116 \\
Traffic & 336-168 & 0.148 & 0.168 & 0.159 \\
Tourism-monthly & 48-24 & 0.107 & 0.138 & 0.096 \\ \hline
\end{tabular}
\caption{ND loss for datasets on which DeepState reported results. The setting denotes the respective encoder and decoder lengths (prediction horizon).}
\label{tab:deepstate}
\end{table}
\subsubsection{Comparison with DeepState}
Since the DeepState code is not publicly available, we compare ARU with DeepState on exactly those dataset and settings for which we were able to approximately match the publicly available DeepAR numbers in the DeepState paper. Here we also include a sixth dataset Tourism-monthly since it was used in
the DeepState paper.
This is a small dataset of 366 time-series containing monthly tourism demand for approximately 15 years \cite{athanasopoulos2011tourism}. In this dataset, the start timestamp as well as length of each time-series are different. Since this is a monthly dataset, we use month of the year feature.
We observe that ARU is better than DeepState in 3 out of 4 datasets. Both DeepAR and DeepState models use a time-series id, whereas we do not.
Perhaps these datasets do not show too much local variation, making it possible for large deep models to memorize local patterns in its parameters as we showed in our synthetic experiments.
\subsubsection{ARU integration}
Another way in which we differ from DeepState is that instead of directly
yielding
the local predictions, we integrate them with global predictions using more learned global parameters. This allows us to safeguard against the high variance in local predictions and to meta-learn the best fall-back strategy. In Table~\ref{tab:ARU-Direct} we compare ARU with ARU-direct whose local predictions are output directly. We observe that ARU-Direct is much worse than ARU and also worse than the Baseline.
\begin{table}[]
\centering
\begin{tabular}{|l|l|r|r|} \hline
Dataset & Method & Normalized & RMSE \\
& & Deviation (ND) & \\ \hline
Rossman & Baseline & 0.0987 & 1020.7 \\
& ARU & 0.0851 & 908.6 \\
& ARU-Direct & 0.1358 & 1835.68 \\ \hline
Walmart & Baseline & 0.1187 & 3585.8 \\
& ARU & 0.1058 & 3413.3 \\
& ARU-Direct & 0.1271 & 4041.88 \\\hline
Electricity & Baseline & 0.1264 & 377.3 \\
& ARU & 0.1260 & 344.6 \\
& ARU-Direct & 0.1389 & 398.11 \\ \hline
Traffic & Baseline & 0.1991 & 0.0374 \\
& ARU & 0.1964 & 0.0372 \\
& ARU-Direct & 0.2195 & 0.0379 \\ \hline
Parts & Baseline & 1.5763 & 1.300 \\
& ARU & 1.5409 & 1.285 \\
& ARU-Direct & 1.6555 & 1.5332 \\ \hline
\end{tabular}
\caption{Comparison of performance of ARU-Direct method with Baseline and ARU}
\label{tab:ARU-Direct}
\end{table}
\begin{table}[]
\setlength\tabcolsep{10.0pt}
\centering
\begin{tabular}{|l|l|l|l|} \hline
Dataset & Baseline & ARU & SNAIL \\ \hline
Electricity & 1.0458 & 1.1788 & 3.0754 \\
Traffic & 2.0740 & 2.3383 & 5.7673 \\
Walmart & 0.7034 & 0.9434 & 1.2693 \\
Rossman & 0.4379 & 0.6717 & 2.2837 \\ \hline
\end{tabular}
\caption{Inference time (in seconds)}
\label{tab:running_time}
\end{table}
\subsubsection{Running time comparison}
In Table \ref{tab:running_time}, we compare running times of different models of inference on the test data. The running time of SNAIL is larger, up to a factor of four higher than the baseline, due to its costly self-attention mechanism. In contrast, the running time of ARU is at most
1.5 times more than the baseline.
Also, since ARU stores a fixed sized state per time-series, its storage overhead is also very small.
These experiments demonstrate that ARU is an effective light-weight method of streaming adaptation that can be easily plugged into any existing global model.
\section{Conclusion}
In this paper we presented ARU, an adaptive recurrent unit that provides streaming local adaptation of globally
trained
deep time-series models. The core principle in ARU is simple --- exploit the kernel trick to obtain a closed-form optimal solution to linear least square regression loss. While this trick is classically known, our contribution is in recognizing the imminently practical use of this trick in modern deep network settings and designing a method of effectively integrating global-local patterns. Unlike existing self-attention or memory-augmented models with linearly increasing memory requirements, ARU allocates only constant sized state per time-series. Unlike methods that learn to generate local parameters, ARU can adapt to streaming data. Experiments on five real-life and synthetic datasets establish that ARU is an effective light-weight method of streaming adaptation of global deep forecasting models.
In future, we plan to apply the ARU method of adaptation to linear layers in other parts of a deep network.
\noindent
\paragraph{Acknowledgements: }We thank Flipkart for sponsoring the project, Srayanta Mukherjee and P Kompalli for getting us interested in the problem, and M Prashanth for performing initial experiments.
\bibliographystyle{ACM-Reference-Format}
|
\section{Arnold-Euler geodesics}
\label{sec0}
\setcounter{equation}{0}
{\tiny
\tableofcontents
}
\*\*
\defArnold-Euler geodesics{Introduction}
\section{Arnold-Euler geodesics}
\label{sec1}
Many works investigate the Fluctuation Relation, (FR): but its
interpretation is often very far from the original one proposed in
\cite{GC995} and in subsequent works, see for instance \cite{Ga013b}.
Here I present the original point of view referring also a few
of its interpretations, consequences and related conjectures.
In this review only the foundations of the theory are discussed, as I
cannot present the many (pertinent) developments that followed,
starting with the early ones, \cite{JP998b,Ku998,EPR999,LS999,Ma999}.
The FR arises from a simple theorem on dynamical systems, the {\it
Fluctuation Theorem}, (FT). The FT applies, under further suitable
assumptions, to {\it Anosov systems}: which can be considered as playing a
role analogous to that played for non chaotic systems by the harmonic
oscillators, although sometimes they are considered an abstract
mathematical notion.\footnote{\small{\it e.g.\ } ``Whether or not speculations concerning
such hypothetical Anosov systems are an aid or a hindrance to
understanding seems to be an aesthetic question'', \cite[p.221]{Ho999}.
\label{H1}}
In natural observations initial data are generated by a well defined
procedure, that is sometimes called a {\it protocol}, but are always
affected by unavoidable errors, no matter how carefully one fixes the
protocol.
Therefore initial data are generated with a probability distribution:
defined by the protocol and {\it unknown}. Yet it is subject to the
fundamental assumption that it is a probability distribution on the ``phase
space'' $M$ (a smooth Riemannian manifold here) which {\it admits a
density} with respect to the volume of $M$. The probability of $x\in M$
being in a open set $dx$ around $x$ has the form $\r(x)dx$ where $\r$ is
some continuous function (or slightly more general). This is an assumption
which should not be overlooked: it {\it cannot be proved} but it is always
assumed (at least tacitly) and is, therefore, a {\it law of nature}, with
the far reaching consequence that it leads to the determination of the
probability distributions of the stationary states in equilibrium as well
as in nonequilibrium systems: see below.
The connection with Physics is established via the hypothesis (called {\it
Chaotic Hypothesis}, (CH)) stating that ``all'' systems exhibiting
chaotic motions can be treated for many purposes as {\it Anosov systems}.
Informally, in such systems an observer co-moving, in phase space $M$, with
a point $x$ sees it as a ``saddle point'' (mathematically a ``hyperbolic fixed
point''), while it wanders invading an attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$
in $M$, The notion of Anosov maps and flows,
\cite{AA966,Sm967},\cite[Ch.4]{GBG004}, is briefly recalled in the
footnote.%
\footnote{\small If $M$ is a smooth ({\it i.e.\ } $\infty$-differentiable) bounded
manifold and $S$ is an invertible smooth ({\it i.e.\ } $\infty$-differentiable
together with the inverse $S^{-1}$) map on $M$, the system $(M,S)$ is an
Anosov map if
%
\\
(a) at every point $x\in M$ there are two complementary tangent planes
$T_s(x)$ and $T_u(x)$, transverse in $x$, which depend continuosly on
$x$, are covariant in the sense that the Jacobian ${\partial} S(x)^{\pm1}$ acts
on the plane tangent to the attracting set so that ${\partial} S(x)^{\pm1}
T_\g(x)=T_\g(S(x)^{\pm1}), \g=u,s$,
%
\\
(b) furthermore there are
$C>0,\l<1$ such that $|{\partial} S^n(x) v|< C \l^n |v|, n>0$ if $v\in T_s(x)$
and $|{\partial} S^{-n}(x) v|< C \l^n |v|, n>0$ if $v\in T_u(x)$,
%
\\
(c) there is a point whose orbit is dense in $M$.
%
\\
%
The definition of Anosov flow is similar: the covariant mutually
transversal planes are now three: $T_u(x),T_s(x)$ on which expansion and
contraction take place under action of the flow $S_t$ as in the map case
and a third $1$-dimensional tangent plane parallel to $f(x)$, if $\dot
x=f(x)$ is the differential equation defining the flow $S_t$; this is a
tangent vector supposed not zero, $|f(x)|>0$, and which, of course,
neither expands nor contracts under the ${\partial} S_t(x)$. \\
The stable and unstable planes of Anosov systems (maps or flows) can be
{\it integrated} ({\it i.e.\ } there is a smooth surface everywhere tangent to them)
to define stable and unstable manifolds $W_u(x),W_s(x)$ which are dense on
$M$.\label{hyperbolic}}
Suppose that evolution ($\infty$-smooth, for simplicity) of a mechanical
system is defined by a map $S$ on a phase space $M$ and is attracted by an
$\infty$-smooth surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\subset M$ on which $S$ is an Anosov map
$S{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\,{\kern-1.truept\leftarrow\kern-5.truept\to\kern-1.truept}\,{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, as considered here in most cases (for simplicity). It
transforms an initial datum $x$ into the new datum $Sx$ in a single time
step; then the main property of the dynamical system $(M,S)$ is that the
evolution is chaotic and a phase space point, with exceptions forming a set
of zero volume, moves accumulating at the attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$
densely.\footnote{\small This means that for all times $t_0$ the closure of
the trajectory of $\{S^tx\}_{t>t_0}$ is ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$.} Such $S$ will be
called a {\it map with an Anosov attractor}.\footnote{\small Evolutions
defined by maps arise typically when studying evolutions in continuous
time through observations triggered by ``timing events'', {\it i.e.\ }
observations that are made every time a specified event takes place: for
instance every time that the evolving trajectory crosses a given surface
in phase space. These are referred as observations performed on a {\it
Poincar\'e's section}.%
\label{Poincare' section}}
The key property of maps with an Anosov attractor is that the fraction of
time asymptotically spent in any open (or just measurable and with $>0$
volume) phase space region $dx$, defines a stationary probability
distribution $\m(dx)=\m(Sdx)$ which is independent of the initial $x$,
again except for data $x$ forming a set of $0$ phase space volume,
\cite{Si977,Si994}. The probability $\m$ is concentrated on the attracting
surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, {\it i.e.\ } $\m({\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F})=1$; the data outside the attracting surface
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ evolve while exponentially attracted by ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$,
\cite{Si968a,Si968b}.
Likewise if the evolution is instead described by a flow $x\to S_tx$,
generated by a ($\infty$-smooth) differential equation on
$M$, $\dot x=f(x)$, which on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\subset M$ is an Anosov flow and ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is
an attractive smooth surface, then any initial point $x\in M$, with
exceptions forming a set of zero volume, moves accumulating densely at
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and spending a well defined fraction of time in every measurable
region of $M$ with positive volume: thus uniquely defining a stationary
probability distribution $\m(dx)=\m(S_t dx)$, concentrated on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, which
is independent of the initial $x$, again except for data forming a set of
$0$ volume in phase space $M$, \cite{Si977,Si994,BR975}: such flow $S_t$
will be called a {\it flow with an Anosov attractor}.
\0Formally: \*\noindent{\it Theorem: Suppose that a smooth dynamical system is
defined on a manifold $M$ containing a globally attracting smooth surface
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ on which the motion is an Anosov map or flow. Then all initial data
$x$, aside from a set of zero volume, evolve visiting open sets $D$ with
a time frequency $\m(D)$: the probability distribution $\m$ is
independent on the protocol generating $x$. The probability $\m$ is
ergodic and mixing at exponential rate for all smooth observables.} \*
Phase space $M$ can be a smooth manifold or, more generally, an open set in
$R^n$ for some $n$, which is a domain of attraction of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\subset M$. For
systems with a global Anosov attractor (flows or maps) the associated
unique stationary probability distribution $\m(dx)$ that defines the
statistical properties of the evolution, ({\it i.e.\ } frequencies of visit to
regions of $M$), is called the {\it SRB distribution},
\cite{RT971,Ru995,Ru989}.
The just mentioned theorems on maps or flows become relevant for systems
evolving chaotically in the cases in which the following hypothesis holds
{\it i.e.}, as its name suggests (as intended in \cite{GC995,BG997,BGG997}, see
also the warning in \cite[endnote 18]{GC995}), always when motions are
{\it empirically chaotic}: \*
\noindent{\it Chaotic hypothesis (CH):} {\it A chaotic evolution takes place on a
phase space $M$ being attracted by a bounded smooth attracting surface
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\subset M$ and on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ the map $S$ (or the flow $S_t$) is an Anosov
map (or flow).} \*
The SRB distribution $\m$ has support on the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$: so
smoothness of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is a strong assumption.%
\footnote{\small Which contrasts the picture of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ as a fractal set. In
systems of $\sim10^{20}$ molecules with a fractal attractor of dimension
$6\cdot10^{19}+3.141$ this means that it 'behaves' as a smooth surface of
dimension $6\cdot 10^{19}$; or in a Navier-Stokes fluid (an
$\infty$-dimensional system) at large Reynolds number $R$ an attracting
set of dimension $R^{\frac93}+.33$ 'behaves' as a smooth surface of
dimension integral part of $R^{\frac93}$.} Of course there is the
possibility that ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is a surface of dimension lower than that of $M$, as
specified in the CH.\footnote{\small Often, if the dynamical system depends
on a parameter $\e$, the chaotic motion might occupy, asymptotically, an
attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}_\e$ with a dimension dependent on $\e$ and equal to
that of the full phase space only for a small (if any) interval of
variability of $\e$.}
The CH is a general and heuristic (more restrictive) interpretation of
original ideas on turbulence phenomena, \cite{Ru980}, and has been
introduced in \cite{GC995,GC995b} to interpret simulations on evolutions
with attracting set coinciding with the full phase space, and extended to
the more general case in which the attracting surface is lower dimensional,
\cite{BG997,BGG997}).
Hereafter the CH will
be supposed to hold {\it for all dynamical systems considered}, unless
stated otherwise.
The above formulation in which the Anosov system is realized on a
attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\ne M$, rather than on the full phase space, already
hinted in \cite{GC995}, has become relevant as soon as attempts were
undertaken to apply CH to systems for which the strict inclusion
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\subset M$ was manifest, \cite{BG997,BGG997}.
Besides smoothness of the attracting set the further strong assumption of CH
is that the evolution on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is hyperbolic in the sense of Anosov. Both
aspects of the CH can be simultaneously weakened by supposing that the
motion has an attractor which satisfies the ``Axiom A'', \cite{BR975};
however such generality will not be envisaged here.
It will be convenient to distinguish between {\it attractor} and {\it
attracting set} ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$: the latter is an invariant set approached
asymptotically by all points $x$ in its {\it basin} of attraction (which is
an open set around ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$): $\lim_{t\to\infty} distance (S_tx,{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F})=0$; while
an attractor $A\subset {\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is an invariant subset dense on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$
which has full SRB measure ({\it i.e.\ } $\m(A)=1$) and minimal Hausdorff dimension,
often smaller than the dimension of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ which, in turn, could be {\it much
smaller} than the dimension of $M$.
The SRB distributions have strong ergodic properties (see Sec.\ref{sec14}
for some details) and in particular the average value over time of a smooth
observable $O(x)$ is reached exponentially fast: however, more
generally, it is possible that in $M$ there are several attracting sets
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}_i$, each with its own SRB distribution, just as in equilibrium
statistical mechanics there are cases in which the Gibbs state is not
unique and the extremal ones correspond to different phases. It will appear
that the analogy is a deep one, see Sec.\ref{sec7}.
Anosov systems are chaotic systems whose properties can be studied in {\it
great detail}: certainly they correspond to an idealization of chaos; but
it should be kept in mind that Statistical Mechanics arose from the
idealization, far more surprising, that microscopic motion could be
regarded as periodic, \cite{Bo866,Cl871,Bo868,Ma879}, see also
\cite[Sec.6\&7]{Ga016}.
\defArnold-Euler geodesics{Stationary Distributions (SRB)}
\section{Arnold-Euler geodesics}
\label{sec2}
Let time evolution on $M$ be a map $S_r$, (or a flow $S_{r,t}$), that
may depend on a parameter $r$ (or on more but imagine, to simplify, that
only $r$ will be varied). Then as $r$ changes the stationary SRB distribution
$\m_r(dx)$, for the system $(M,S_r)$, changes and the {\it collection} of
such distributions will be called ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{mc}$ and its elements will be
thought of as {\it ensembles of stationary states}.
Volumes of regions in phase space $M$ {\it change}. in general, when
transformed by the discrete evolution map $S$ or by the flow $S_t$
generated by a differential equation $\dot x=f(x)$; and the rate of change
per unit volume can be measured from the Jacobian matrix $J(x)_{ij}={\partial}_i
S(x)_j$ for maps (here ${\partial}_i{\buildrel def\over=} {\partial}_{x_i}$) or by the matrix
$J_{ij}={\partial}_i f_j(x)$ in the case of a flow. And the {\it phase space
contraction rate} is defined by:
\begin{equation}}\def\ee{\end{equation} \kern-3mm
\eqalign{\s(x)=&-\log |\det J(x)|,\qquad{\rm discrete\ evolution} \cr
\s(x)=&-{\rm div} f(x)\equiv -{\rm Tr} J(x), \ {\rm continuous\
evol.}\cr}\Eq{e2.1}\ee
Changing variables or the metric on phase space implies a change of $\s(x)$
into $\s(x)+ u(S_rx)-u(x)$ for a suitable function $u(x)$:\footnote{\small{\it e.g.\ } a
change in variables $y=w(x)$ leads to $u(x)=-\log |\det
{\partial}_{x_j}{w_i}|$.} which implies that the time average
$\lim_{T\to\infty} \frac1T\sum_{j=0}^T \s(S_r^jx)$, if existing, does not
depend on the metric used on $M$. Likewise and with the same implication on
the averages, in the continuous evolution systems, $\s(x)$ changes into
$\s(x)+\dot u(x)$ (where $\dot u(x){\buildrel def\over=}\sum_j {\partial}_{x_j} U(x) \,f(x)_j$
for a suitable function $U(x)$).\footnote{\small{\it e.g.\ } changing $x$ into $y=w(x)$
leads to $U(x)= -\log |{\partial}_{x_j}w_j|$.}
The time average $\s_+$ of $\s$ {\it i.e.\ } of $n\to\s(S^nx)$ or $t\to\s(S_tx)$
coincides, except for a set of $x$'s with $0$ volume, with the average
$\s_+=\int \s(x)\m(dx)$ with respect to the SRB distribution $\m$.
Remark that $\s_+$ is a quantity which has the dimension of an inverse time
in the case of continuous systems while it is dimensionless for maps, (as
time is an integer in the case of maps). It will play an important role in
the following, particularly when $\s_+\ne0$, as it sets a time
scale\footnote{\small Since the CH-evolutions that we consider proceed towards a
bounded attracting set it is $\s_+\ge0$, \cite{Ru996}.} that will be
called the {\it dissipation time scale}.
\defArnold-Euler geodesics{Symmetries, (time reversal)}
\section{Arnold-Euler geodesics}
\label{sec3}
\setcounter{equation}{0}
Trying to evince information from the statistical properties of the stationary
distributions of a time evolution, discrete or continuous, it is
important to take into account symmetries of the underlying equations of
motion, \cite{DGM984}.
There are not many such symmetries and a key role is played by the
fundamental symmetries, like translation and rotation invariance or time
reversal, often enjoyed by the molecular constituents of the systems of
interest and perhaps reflected by their macroscopic properties.
Of particular interest will be systems in which the evolution is $S$,
discrete in time, or is a continuous time evolution $S_t$, which satisfies a
``{\it time reversal\,}'' symmetry, {\it i.e.\ } such that there is an
$\infty$-smooth map $x\to Ix$ on the phase space $M$, {\it independent or
smoothly dependent} on any parameter that might affect the dynamics, with
the property
\begin{equation}}\def\ee{\end{equation}\eqalign{
&(a)\ I S^{-1}=S I,\qquad I^2=1, \qquad {\rm discrete \ evol.}\cr
&(a')\ IS_{-t}=S_t I,\qquad I^2=1, \qquad {\rm cont.\ evol.}\cr
&(b)\ I \ {\rm is\ isometric}\cr
}\Eq{e3.1}\ee
For isolated particle systems $I$ is just the reversal of all velocities
and it is a basic law of nature in Newtonian physics.
A typical situation is described in the following section presenting a
simple, but quite general, model of a nonequilibrium system.
The model will also illustrate the notion of {\it phase space contraction}
and its relation with the thermodynamic notion of {\it entropy
generation}. It will appear that although there is a relation between
entropy creation rate and phase space contraction, still the two notions
are quite different. {\it Nevertheless} their difference can be expressed
as a variation of a suitable phase space observable evaluated at successive
map iterations or, in the cases of flows, as a {\it time derivative} of a
suitable observable: therefore it has no influence, or a controlled one,
on the {\it average phase space contraction}, \cite[Ch2.5]{Ga013b}.
\defArnold-Euler geodesics{Example: reversible dissipation}
\section{Arnold-Euler geodesics}
\label{sec4}
\setcounter{equation}{0}
The system consists in $N\equiv N_0$ particles in a container $\CC_0$ and
of $ N_a$ particles in $n$ containers $\CC_a$ which play the role of {\it
thermostats}: their positions will be denoted $\V X_a,\,a=0,1,\ldots,n$,
and $\V X{\buildrel def\over=}(\V X_0,\V X_1,\ldots,\V X_n)$. Interactions will be
described by a potential energy
\begin{equation}}\def\ee{\end{equation}
W(\V X)=\sum_{a=0}^{n} U_a(\V X_a) +\sum_{a=1}^n W_a(\V X_0,\V X_a)
\Eq{e4.1}\ee
{\it i.e.} particles in different thermostats only interact indirectly, via
the system. All masses will be $m=1$, for simplicity.
\eqfig{110}{80}{}{fig2}{}
\noindent{\small Fig.1:\label{Fig1}\it The reservoirs occupy finite regions outside
$\CC_0$, {\it e.g.\ } sectors $\CC_a\subset R^3$, $a=1,2\ldots$. Their particles are
constrained to have a {\it total} kinetic energy $K_a$ constant, by
suitable forces $\V F_a$, so that the reservoirs ``temperatures'' $T_a$,
are well defined, by $K_a=\sum_{j=1}^{N_a} \frac12\, (\dot{\V
X}_{a,j})^2{\buildrel def\over=} \frac32 N_a k_B T_a{\buildrel def\over=} \frac32 N_a\b_a^{-1}$. The
set-up, classical and quantum, is introduced in \cite{FV963}.}
\vskip2mm
Particles in $\CC_0$ may also be subject to external, possibly
non conservative, forces $\V F(\V X_0,{\V E})$ depending on a few strength
parameters ${\V E}=(E_1,E_2,\ldots)$. It is convenient to imagine that
the forces due to the confining potentials determining the geometrical shape
of the region $\CC_0$ are included in $\V F$, so that one of the
parameters is the volume $V=|\CC_0|$. See Fig.1.
Following Sec.\ref{sec1} the statistical properties of the stationary
states of the system should be described, assuming the CH, by the SRB
distributions $\m_{\V E}$ on phase space.
The equations of motion are:
\begin{equation}}\def\ee{\end{equation}\eqalign{
\ddot{\V X}_{0i}=&-{\partial}_i U_0(\V X_0)-\sum_{a}
{\partial}_i W_a(\V X_0,\V X_i)+\V F_i\cr
\ddot{\V X}_{ai}=&
-{\partial}_i U_a(\V X_a)-
{\partial}_i W_a(\V X_0,\V X_i)-\a_a \dot{\V X}_
\cr}\Eq{e4.2}\ee
where the last term $-\a_a \dot{\V X}_a$ is a {\it phenomenological} force
that implies that thermostats particles keep constant total kinetic energies
$K_a=\frac32 N_a k_B T_a$: $\a_a$ is therefore (as checked by
direct computation of the time derivative of $K_a=\frac12\, \dot{\V
X}_{a,j}^2$ defined by
\begin{equation}}\def\ee{\end{equation}\a_a \,{\buildrel def\over=}\,\frac{L_a-\dot U_a} {3N_a k_B
T_a}\Eq{e4.3}
\ee
where $L_a=-\partial_{\V X_a} W_a(\V X_0,\V X_a)\cdot \dot{\V X}_a$ is
the work done per unit time by the forces that the particles in
$\CC_0$ exert on the particles in $\CC_a, a>0$; here $k_B$ denotes
Boltzmann's constant.
The exact form of the forces that have to be added in order to insure the
kinetic energies constancy should not really matter, within wide
limits. But this is a property that is not obvious and which is much
debated.\footnote{\small The above thermostatting forces choice
can be seen to coincide with the ones obtained via Gauss' {\it
least effort} principle for ideal anholonomic constraints applied to
the constraints $K_a=const$, see
\cite[Ch.2]{Ga013b}: this is a criterion that
has been adopted in several simulations,
\cite[Sec.5.2,p103]{EM990}. {\it Independently} of Gauss' principle it is
immediate to check that if $\a_a$ is defined by Eq.\equ{e4.3} then the
kinetic energies $K_a$ are, strictly, constants of motion.}
The work $L_a$ in Eq.\equ{e4.3} will be interpreted as {\it heat} $\dot
Q_a$ ceded, per unit time, by particles in $\CC_0$ to the \hbox{$a$-th}
thermostat. The
{\it entropy production rate} due to heat exchanges between the system and
the thermostats can, therefore, be naturally defined by
\begin{equation}}\def\ee{\end{equation}\s^0(\dot{\V X},\V X){\buildrel def\over=}\sum_{a=1}^{N_a} \frac{\dot Q_a}{k_B
T_a}\Eq{e4.4}\ee
because the ``temperature'' of $\CC_a$ remains constant, and at
stationarity the thermostats can be regarded in thermal equilibrium.
It should be stressed that here {\it no entropy notion} is introduced for
the stationary state: only {\it variation} of the thermostats entropy is
considered and it should not be regarded as a new quantity because, in the
stationary states, the thermostats should be considered in equilibrium at a
fixed temperature.
A question is whether there is any relation between $\s^0$ and the phase
space contraction $\s$ of Eq.\equ{e2.1} for the equations of motion in
Eq.\equ{e4.2} ({\it i.e.\ } minus the divergence of the equations
Eq.\equ{e4.2}).
The latter, in the recent literature, has been identified
with the entropy production: and in the present case can be immediately
computed by the appropriate differentiation of Eq.\equ{e4.2},\equ{e4.3} and
is (neglecting $O(\min_{a>0} N_a^{-1})$)
\begin{equation}}\def\ee{\end{equation}\textstyle\s(\dot{\V X},\V X)={\mathop\sum\limits_{a>0}}
{\textstyle\frac{3N_a-1}{3 N_a}} \frac{\dot Q_a-\dot U_a}{k_B T_a} =
{\mathop\sum\limits_{a>0}} \frac{\dot Q_a}{k_B T_a}-\dot U
\Eq{e4.5}\ee
where $U=\sum_{a>0} \frac{3N_a-1}{3 N_a} \frac{U_a}{k_B T_a}$. Hence in
this example, physically interesting, in which the thermostats are
``external'' to the system volume (unlike to what happens in several
examples in which they act inside the volume of the system), the phase
space contraction is not the entropy production rate, \cite[Ch.2]{Ga013b}. {\it
However it differs from the entropy production rate by a total time
derivative}.
Consequence: entropy creation rate $\s^0$ and phase space contraction
$\s$ differ, {\it but their time averages coincide}.
This is relevant because the definition Eq.(\equ{e4.4}) {\it has meaning
independently of the equations of motions} and can, therefore, be
suitable for experimental tests, \cite[Ch.5-2.Ch.3,Ch.4]{Ga013b}.
It should be stressed that the numbers $N_a$ of particles in the reservoirs,
$a>0$, enters through $\frac{3 N_a-1}{3N_a}$, hence is is essentially
independent on the thermostat sizes (provided large).
Finally I mention that the identification, up to a total time derivative,
of phase space contraction with entropy production rate can be shown, as
discussed in Sec.\ref{sec7}\,,\,\ref{sec15} below, to cover the entropy
production rate in systems whose evolution can be approximated by
macroscopic continua equations, like fluids described by Navier-Stokes
equations. In the sense that, again, phase space contraction of the
interacting particles systems that underlay the macroscopic equations is
related, in the stationary states, to the entropy production rate
independently defined in classical nonequilibrium Thermodynamics,
\cite{DGM984}: and, at most, it differs from it by a total time derivative,
\cite[Ch.4]{Ga013b}.
\defArnold-Euler geodesics{Hamiltonian dissipation?}
\section{Arnold-Euler geodesics}
\label{sec5}
\setcounter{equation}{0}
No entropy production is possible in a stationary state of a
Hamiltonian system ({\it i.e.\ } an isolated system). At least not if it is
finite. However things are different when the system is in contact
with infinite systems.
As an example consider a Hamiltonian version of the model of
Sec.\ref{sec4}, Fig.1 above. If the constraints on the kinetic energy in
the containers $\CC_a, a=1,\ldots,n$, are removed and the containers are
extended to infinity, interesting stationary states can be obtained from
initial configurations which, in each $\CC_a$, are naturally chosen from
the {\it canonical equilibrium ensemble} with density $\r_a$ and
temperature $T_a$ and from any distribution\footnote{\small With density on
the phase space $\CC_0\times R^{3N_0}$.} for the particles in
$\CC_0$. The initial distribution will be called $\m_0$ and, although not
invariant in time, it may evolve towards an invariant one, $\m$ (as
reasonable as this looks, however, a mathematical proof of this is {\it far
from known}). The existence of $\m$ will be assumed in this example.
Since the containers are infinite the stationary state that will be reached
can be expected to keep an average kinetic energy per particle remaining
$\frac32 k_B T_a$, identically equal to the initial value. For rather
general models of microscopic interaction between particles, it can be
shown, see for instance \cite{GP010b},
that the time evolution of the particles in $\CC_a$ that are far from the
boundary of $\CC_0$, are little affected by the interactions with the
particles in $\CC_0$ and the average kinetic energy per particle in each
$\CC_a$ will be an exact constant of motion, equal to the initial
$\frac32k_B T_a$, for all finite times (although it is still possible that
in the limit of infinite time this might change).\footnote{\small The
physical picture is that the energy generated by work performed by the
active forces on the particles in $\CC_0$ and by the interactions between
the particles in $\CC_0$ and those in the thermostats is ceded to the
thermostats creating in them heat currents $J_a$ which
decrease as the inverse of the square distance to $\CC_0$: so the
thermostats remain asymptotically, as the distance from $\CC_0$ tends to
$\infty$, in equilibrium.}
Consider as initial distribution $\m_0(dx)$ which is a product of
independent canonical distributions in each container $\CC_a,a\ge0$, with
given densities and temperatures $\r_a,T_a\equiv \frac1{k_B
\b_a}$.\footnote{\small But the choice of $T_0$ has no particular physical
meaning and the distribution in $\CC_0$ could be replaced by ``any''
distribution, with some density on the $X_0,\dot X_0$ variables.}
Although {\it now purely Hamiltonian} the system is infinite
and the phase space volume {\it measured by the evolving distribution}
$\m_t(dx)=\m_0(S_{-t} dx)$, changes per unit time by $\s(x)$ with:
\begin{equation}}\def\ee{\end{equation} \s(x)=\frac{d}{dt} \log \frac{\m_0(S_{-t}dx)}{\m_0(dx)}\equiv
\sum_{a=1}^n \b_a \dot Q_a+\b_0 \dot Q_0\Eq{e5.1}\ee
where $\s(x)$ is computed from the equations of motion as:
\begin{equation}}\def\ee{\end{equation} \kern-3mm\eqalign{
\dot Q_a=& L_a=-{\partial}_{\V X_a} W_a(\V X_0,\V X_a)\cdot\dot {\V X}_a,
\qquad a\ge1\cr
\dot Q_0=&\dot K_0+\dot U_0\cr}\Eq{e5.2}\ee
and $\b_0 \dot Q_0=\b_0\,(\dot K_0+\dot
U_0)=\b_0\,\V F\cdot \dot{\V X}_0-\sum_{j>0}(\dot U_{0j}-\dot Q_j))$
is a time derivative so that it {\it does not contribute} to the time average
$\s_+=\lim _{T\to\infty}\frac1T\int \s(S_t x) dt$, \cite[Ch.4]{Ga013b}.
\*
\noindent{\it Remark:} If $\m_0$ is defined, as proposed above, as a product of
independent canonical distributions it might be surprising that the system
in $\CC_0$ plays a special role: could one write the same formulae with
$\CC_1$ playing the role of $\CC_0$ and find that $\dot Q_1$ is a total
derivative of $U_1+K_1$?
\\ However $U_1+K_1$ is infinite unlike $U_0+K_0$, so nothing can be
concluded about its time derivative. If the regions $\CC_j$ were finite
then the system would evolve and all $\dot U_a+\dot K_a$ would
uninterestingly average to $0$: so $\CC_0$ plays a special role and is the
only container for which $U_0+K_0$ and its time derivative are meaningful.
\*
The expression Eq.\equ{e5.1}, and the irrelevance of the contribution from
$\dot Q_0$ to the average of $\s$ (see Eq.\equ{e5.2}) is the key to the
interpretation of fluctuations on $\s$ in systems modeled by particles:
even in cases (essentially in all experimental settings) in which the
evolution is not describable in terms of equations of motion, in the sense
that the equations of motion are not analytically known.
In such cases the average of $\s(x)$ (equal to that of
$\s^0(x)=\sum_{a\ge1}\b_a \dot Q_a$) {\it is still accessible} via measurement
of the heats exchanged and the temperatures of the reservoirs with which
heat is exchanged. Of course it is a delicate and difficult task to measure
{\it all} such quantities, {\it i.e.\ } the full entropy production.
It is remarkable, and it will be discussed in Sec.\ref{sec6}, that quite
generally, under the Chaotic Hypothesis, that it becomes possible to obtain
a general, universal, property of the (rare) fluctuations of the entropy
production.
\defArnold-Euler geodesics{Fluctuation Relation (FR)}
\section{Arnold-Euler geodesics}
\label{sec6}
\setcounter{equation}{0}
The {\it Fluctuation Relation} deals with the entropy production
$\s^0(x)=\sum_a\frac{\dot Q_a}{k_B T_a}$, see \equ{e5.1},\equ{e5.2}, or
more generally, with the phase space volume contraction rate
$\s(x)=-\sum_i{\partial}_{x_i} f_i(x)$ and it applies to {\it finite} systems
whose evolution takes place on a phase space $M$, via an equation $\dot
x=f(x)$, and is {\it time reversal symmetric} in the sense of
Sec.\ref{sec3} and hyperbolic on $M$ ({\it i.e.\ } it is a Anosov
system).\footnote{\small A positive $\s(x)$ means that the volume {\it
contracts} near $x$: hence in stationary states the average $\s_+$ of
$\s$ must be $\ge0$, \cite{Ru996}.} It equally deals with evolutions
which are time reversible Anosov maps: attention will be mostly
concentrated on the continuous time case, to avoid repetitions.
The $\s(x)$ is defined in terms of the metric on $M$ and via appropriate
covariant derivatives ${\partial}_{x_i}$: but it is simpler to imagine that $M$
is a Euclidean space with coordinates measured in prefixed units, so that
${\partial}_{x_i}$ are the usual partial derivatives.
In general $\s(x)$ {\it depends} on the metric and changing metric on $M$ the
expression of $\s(x)$ changes; however the variation can be, in general,
expressed by the time derivative of a suitable observable so that the long
time averages $\s_+$ of $\s(x)$ {\it do not depend on the metric}, like most
physically interesting observables ({\it e.g.\ } the Lyapunov exponents). See
comments following Eq.\equ{e2.1} and Eq.\equ{e5.2}
Given an attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ on which the evolution is an Anosov system,
{\it i.e.\ } motion on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is a smooth continuous hyperbolic
flow\footref{hyperbolic} $x\to S_tx$, or a smooth discrete hyperbolic
map\footref{hyperbolic} $x\to Sx$; let $\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(x)$ be the {\it surface
contraction rate} on the surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and consider the
quantity
\vglue-6mm
\begin{equation}}\def\ee{\end{equation} \eqalign{
p=&\frac1\t \int_0^\t \frac{\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(S_t)}{\s_+} dt \qquad {\rm continuous\
time}, \ \t>0\cr
p=&\frac1\t\sum_{k=0}^\t \frac{\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(S^kx)}{\s_+}\qquad{\rm discrete\ time},
\ \t \ {\rm integer}
\cr}\Eq{e6.1}\ee
where $\s_+$ is the infinite time average of $\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(x)$, which is
$x$-independent, aside exceptional $x$'s in a $0$-volume set, and
coincides with the average of $\s$ with respect to the SRB distribution
$\m_{srb}$ on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, {\it i.e.\ } the probability distribution generated by the motion of
any point $x$ chosen randomly with some density with respect to the volume,
\ref{sec1}.
Then consider the stationary $\m_{srb}$-probability of the above variable $p$,
Eq.\equ{e6.1}, and define its {\it large deviation rate} as a function $\z(p)$
such that:
\begin{equation}}\def\ee{\end{equation} P_{srb,\t}(D)= e^{ \t\,\max_{p\in D}\z(p) +o(\t)}\Eq{e6.2}\ee
for all domains $D$ ({\it i.e.\ } for all regions $D$ which are closures of their
interiors).
If the dynamical system is an Anosov system (or, more generally, if is satisfies
the ``axiom A'') then
\*
\noindent{(a)} the rate $\z(p)$ exists and is defined in the interior of an
interval $[p',p'']$ containing $p=1$,
\\(b) it is analytic in $p$ if $p'<p''$ while it is $-\infty$ for
$p\not\in [p',p'']$
\\(c) $\s_+=0$ if and only if the SRB distribution $\m$ admits a density
over the attracting surface, \cite{Ru996}.
\* A simple universal result
for reversible Anosov systems, is the following {\it Fluctuation Theorem},
\cite{GC995},\cite{Si977,Ga995b}.
\*
\0FT: {\it If the evolution is time reversal symmetric then the rate
function\footnote{\small The rate $\z(p)$ is defined so that the
probability of finding $\frac1\t\int_0^\t \frac{\s(S_tx)}{\s_+}dt dt\in
[p,p+\d p]$ is $\exp \t \max_{[p,p+\d p]}\z(p)$ for $p\in (-p^*,p^*)$
where $p^*\ge1$; it exists and is analytic if $\s(x)$ is the phase
space contraction of an Anosov evolution.} verifies the symmetry
property:
\begin{equation}}\def\ee{\end{equation} \z(-p)=\z(p)-p \s_+\Eq{e6.3}\ee
%
for $p\in [-p^*,p^*]$ with $p^*\ge1$.
}
\*
An immediate consequence is that if the Chaotic Hypothesis is considered
valid, time reversibility holds and the attracting set can be supposed to
coincide with the full phase space (so that ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}=M$ and $\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\equiv\s$),
then the entropy generation rate or more generally the phase space
contraction rate are expected to satisfy the large deviation property which
is, in this case, called {\it Fluctuation Relation}, FR, and it is {\it
informally} written as
\begin{equation}}\def\ee{\end{equation} \log\frac{P_{srb}(p)}{P_{srb}(-p)}=\t\,1\, p\, \s_+\, +\, o(\t)
\Eq{e6.4}\ee
and more precisely formulated, as existence of a ``large deviation'' rate
$\z(p)$ satisfying Eq.\equ{e6.3}: where the $1$ is inserted for later
reference.
While the formal probability density for the events $\pm p$, {\it i.e.\ }
$P_{srb}(\pm p)$, is a difficult quantity strongly dependent on the
dynamical system, the interest of the FR is that Eq.\equ{e6.3},\equ{e6.4}
are, under the above assumptions, an {\it exact symmetry} of $\z(p)$ and,
at least in some cases, FR {\it deals with a quantity ($\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(x)$) which
has physical meaning} (entropy generation rate) and mathematical meaning
(phase space contraction rate): therefore a check of Eq.\equ{e6.4} can
become a test of the chaotic hypothesis.
The fluctuations relation is, for time reversible evolutions, a symmetry of
the SRB distributions. {\it However} it requires that:
\*
\noindent(i) the motion on the attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ has the Anosov property,
\\
(ii) and {\it at the same time} it is reversible; hence, in the frequent
cases in which ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is not the full phase space but just a smooth surface
in it, it should be $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}={\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, quite {\it unlikely} if $I$ is the usual time
reversal symmetry ({\it i.e.\ } velocities reversal),\\
(iii) furthermore, if (i) and (ii) hold, the FR concerns the fluctuations
of the surface area of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, and {\it not of the full volume}: which is
very hard to access, as ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ itself. \*
\0and the three conditions strongly limit a literal
applicability of FR and lead to the analysis of further properties of the
considered evolutions, see Sec.\ref{sec9}\,,\,\ref{sec17}.
Nevertheless it can be applied to systems that are only mildly out of
equilibrium. If the system, remaining time reversal symmetric, is set out
of equilibrium by the action of small forces and is in contact with
thermostats with small differences of the respective temperatures, call
$\e$ a parameter measuring the size of the forces and of the temperature
differences. Then, if for $\e=0$ the system has the Anosov property on the
full phase space, it will continue to have such property also for small
$\e\ne0$, because Anosov systems are structurally stable,
\cite{AA966}. Hence the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ will remain identical to the
full phase space and time reversal will be a symmetry of the motions on the
attracting set and the FR assumptions will remain verified. This was the
case of the systems to which the FR has been applied, \cite{GC995}, and
tested, \cite{BGG997}, to explain the fluctuations of the phase space
contraction observed in the simulation in \cite{ECM993}.
In attempting to test or use the FR in systems which are not very small it
is not reasonable to hope that the smallness of the above $\e$ does not
depend on the system size, although important cases (lattices of coupled
Anosov maps) are known in which $\e$ can be taken independent of the size,
\cite{BK995,GBG004}: therefore it might be thought that FR becomes
irrelevant in most interesting cases, \footref{H1}. Clearly more properties
are needed to deal with the systems that are not small perturbations of
Anosov systems. In Sec.\ref{sec9} the {\it applicability far beyond the
latter cases} will be discussed. \*
\noindent{\it Remark:}\label{cil} Often the Eq.\equ{e4.5} raises the question
``how can it be relevant'' as the Boltzmann's constant in the denominator
is likely to give a huge value to the inverse time scale $\media{\s}$ which
determines the time scale over which the FR yields predictions? For
instance imagining to put $1 \,cm^3$ of steel (with faces of $1\,cm^2$) in
contact between two reservoirs at temperatures $T=300{}^oK$ and $T+\d
T=310{}^oK$ the average of the entropy production rate, ${\dot Q}\frac{\d
T}{k_B T^2}$, can be expressed via the steel thermal conductivity $\ch$
as $\ch (\frac{\d T}T)^2 \frac{\D}{k_B}$: and the result is $\sim
10^{18}\,sec^{-1}$, see also \cite[p.4]{STX005}. If FR could be applied
{\it literally} there would be no way to see a heat flow from cold to warm
during $10^{-6} sec$ before ``trying'' to see it, say once every second,
for at least $\sim 10^{18}/{10^6}$ times ({\it i.e.\ } $\sim 10^3$ billion
days). See Sec.\ref{sec9} for a possible answer to the problem.
\footnote{\small The problem is
considered, by some colleagues, a ``disaster'' for FR, making it
physically irrelevant.} \*
The FR bears {\it formal} similarity with identities arising in the
evolution of equilibrium states, or more generally with the evolution of
initial distributions on phase space {\it which are symmetric under time
reversal but not stationary}. The deep difference between the latter
identities and the above FR is briefly commented in Sec.\ref{sec19}) below.
Unfortunately the name ``fluctuation relation'' has been often used in all
cases, causing great confusion to loom on the subject.
\*
\defArnold-Euler geodesics{Nonequilibrium ensembles. Ensembles equivalence}
\section{Arnold-Euler geodesics}
\label{sec7}
\setcounter{equation}{0}
In general given an evolution equation on a phase space $M$ depending on
one or more parameters, denoted $\V E=\{\n,E,\ldots\}$, the SRB stationary
states, {\it i.e.\ } the distributions that are generated by all points of $M$,
excepting a subset of $M$ with $0$ volume, form a collection ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$ of
probability distributions $\m_{\V E}$ parameterized by the given parameters
and each of which can be called an ``ensemble''.
Even in the cases, considered in this section, in which there is only one
parameter $\n$ and CH holds, there might be several distinct attracting
surfaces and therefore more than a single SRB distribution $\m_\n$: if so,
further parameters will have to be added to distinguish the various
possibilities, {\it just} as done in equilibrium statistical mechanics in
presence of phase transitions to distinguish the different pure phases,
\cite{Ru969,LR969,Ga000}.
A key question is whether the same system can be described by {\it
different} equations of motion. There are several instances in which this
is possible: for instance a fluid motion can be equally well described by,
say, a Navier-Stokes equation or by a
(far more complex) collection of molecules, in contact with a thermostat and
at given density, at least if attention is given to
observations depending on large scale properties and performed over long
time scales, \cite{Ma867-b}.
Even for Navier-Stokes (NS) fluids there might be several different
equations, simpler than the ultimate molecular models, that can describe
the class of phenomena considered relevant in given physical situations.
For instance it has been convincingly argued that macroscopic transport
coefficients can be obtained by replacing the equations of motion of
molecules by simple(r) models, suitable for simulations, obeying modified
equations of motion which can even be non-Newtonian: in the context of
molecular simulations this has been originated in the early '80s,
\cite{Ho999,EM990}. A first example of equations alternative to the NS
equations to describe a developed turbulence flow is found in \cite{SJ993}.
It is natural to consider, together with the collection of SRB
distributions $\m_\n\in{\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$ for a given equation depending on a parameter
$\n$, the collection $\m'_E\in{\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$ of SRB distributions corresponding to a
different equation parameterized by a new parameter denoted $E$, which on
physical or just heuristic grounds describes equivalently the same class of
phenomena, {\it i.e.\ } predicts the same properties for large classes of
observables.
\*
\noindent{\it Remarks:} (i) The equivalence should mean that it is possible to
establish a correspondence between the ensembles ({\it i.e.\ } the distributions) in
${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$ and the ones in ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$ so that for each $\n$ there is a
corresponding $E(\n)$ and the average of ``most'' observables in the
$\m_\n\in {\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$ and $\m'_{E(\n)}$ in ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$ should
coincide (or be close).
\\
(ii) This is quite analogous to the description of equilibrium states in
Statistical Mechanics (SM): the canonical distribution $\m^V_\b$ of $N=\r V$
molecules of a gas in a container of volume $V$ depends, at fixed density
$\r$, on a parameter (inverse temperature) $\b=(k_B T)^{-1}$ and the
microcanonical distribution $\m^{'V}_E$ depends on a parameter (total energy)
$E$. If $E$ and $\b$ are so related that
\vglue-6mm
\begin{equation}}\def\ee{\end{equation}\m^{'V}_E(\sum_{i=1}^N\frac1{2m} p_i^2)=\frac32 N \b^{-1}\Eq{e7.1}\ee
then the average of ``many'' observables $O$, $\m^{'V}_E(O)$, is equal or
close to $\m_\b(O)$. \\
(iii) Actually, in SM, in the limit as
$V\to\infty,\ \r=N/V\ {fixed}$, for any {\it local observable}. {\it i.e.\ }
depending only on the configuration of the molecules located in a finite
region, the canonical and microcanonical averages are not only close but
{\it strictly equal}, at least in absence of long range forces or of phase
transitions, \cite{Ru969,Ga000}. \\
(iv) And, still considering (SM), in presence of phase transitions at $\b$
it will be necessary to label the distributions in ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$ by further
parameters $a$: in this case the distributions in the ensemble ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$ will
also have to be distinguished by an equal number of parameters $a'$ and a
correspondence between $a$ and $a'$ can be established so that, under the
condition Eq.\equ{e7.1}, it is still $\m^{'V}_{a',E}(O)=\m^V_{a,\b}(O)$, for
local observables, in the
limit $V\to\infty$. \*
A first example is obtained by considering a system described by equations
on $x\in R^n$ which are obtained as follows\*
\noindent(a) let $\dot x= G(x)$ be a time
reversible equation for the time reversal $Ix=-x$ ({\it i.e.\ } $G(x)=G(-x)$) \\
(b) add a reversible forcing $f(x)$, with $If=fI$ ({\it i.e.\ } $f(x)=f(-x)$)
\\
(c) and a
dissipative term, $-\n Lx$, with $L$ linear and positive $(Lx\cdot
x>0$, if $x\ne0$) depending on a parameter $\n$, whose effect is to
balance, in the average, the ``energy'' injected by the forcing
\*
The complete equation has therefore the form
\begin{equation}}\def\ee{\end{equation} \dot x=G(x)+f(x)-\n L x \Eq{e7.2}\ee
At fixed $f$ and for each ``friction'' $\n>0$ small enough, the evolution
will be supposed to satisfy the CH and to lead to a unique
stationary (``SRB'') distribution $\m_\n$.
The collection of the SRB distributions $\m_\n$, as $\n$ varies, will be
denoted ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$, and each of them defines a {\it nonequilibrium ensemble}.
Next consider a {\it different} equation in which the friction coefficient
$\n$ in Eq.\equ{e7.2} is replaced by a multiplier $\a(x)$ so defined that a
selected observable $\O$ is an exact constant of motion. For instance the
cases $\O(x)= x^2$ or $\O(x)=(x\cdot Lx)$ lead to new equations of motion
$\dot x=G(x)+f(x)-\a(x) L x$ with, respectively:
\begin{equation}}\def\ee{\end{equation} \eqalign{
\a(x)=&\frac{x\cdot G(x)+ x\cdot f(x)} {x\cdot Lx}\qquad{\rm or}\cr
\a(x)=&\frac{Lx\cdot G(x)+ Lx\cdot f(x)} {L x\cdot Lx}\cr}
\Eq{e7.3} \ee
%
Then the stationary states for the new equations form a collection of
stationary states ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$ with elements parameterized by the value $E$ of
the constant of motion $\O$ introduced by the multiplier $\a$.
Quite generally the motion generated by the new equations is eventually
restricted to a bounded region, because of the action of the friction
and of conservation laws possibly valid for the time reversible system in
absence of forcing.
Therefore for $\n$ small it can be expected that in the stationary states
$\a(x)$ fluctuates leading to a {\it homogenization phenomenon}, {\it i.e.\ } to the
property that in the stationary state for the new equation
\begin{equation}}\def\ee{\end{equation} \dot x=G(x)+f(x)-\a(x) L x\Eq{e7.4}\ee
large classes of observables have the same averages in the distribution
$\m^c_\n$ and in the distribution $\m'_{E}$ belonging to the new ensemble
${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$, of stationary distributions for Eq.\equ{e7.4}, {\it provided} $\n$
and $E$ are kept related by $\n=\m'_{E}(\a)$:
\begin{equation}}\def\ee{\end{equation} \lim_{\n\to0} \m^c_\n(O)= \lim_{\n\to0} \m'_{E}(O),\Eq{e7.5}\ee
or, equivalently, if $E=\m_\n(\O)$. More formally: \*
{\it If motions following Eq.\equ{e7.2} eventually develop on a ball $M$
generating a family ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^c$ of stationary distributions parameterized by
$\n$ and if the motions following the Eq.\equ{e7.4} are also eventually
confined in a ball $M'$, generating a family ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}'$, then Eq.\equ{e7.5}
holds for arbitrarily fixed observables $O$, provided the correspondence
between $\m_\n^c$ and $\m'_{E(\n)}$ is such that $\n=\m'_{E(\n)}(\a)$ or,
equivalently, $E(\n)=\m_\n(\O)$.} \*
As in the case of ensembles equivalence in equilibrium not all ensembles
are equivalent, not even in the thermodynamic limit, therefore the
observables $\O$ defining the ensemble have to be selected on a case by
case basis.
The above statement has been tested in a few cases: involving strongly
truncated NS equations, \cite{GRS004},\footnote{\small Doubts have been
raised in \cite{RM007}: which might be related to the use of a rather
large value of $\n$ in a strongly truncated NS equation in 2D: it is
hoped that the latter results will be tested again at smaller $\n$ (in
spite of computational difficulties).} Lorenz96 equations, \cite{GL014},
shell model for turbulence, \cite{BCDGL018}.
The conjecture will be analyzed in some detail, and considerably
strengthened, in Sec.\ref{sec17} for the stationary states of the
incompressible NS equation with periodic boundary conditions. But it is
convenient to discuss first in which sense the FR can be made relevant for
systems irreversibly evolving in presence of strong friction, and to
exhibit a few more applications of the FR to classical and new problems.
In particular a {\it key problem} is whether the FR can be of any utility
if the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is a surface of dimension lower than that of
$M$ and, although the evolution equations remain reversible, it is
$I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\ne{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, {\it i.e.\ } reversibility does not hold as a symmetry for motions on
the attracting set (as, instead, required for the validity of the FR), see
comments (i-iii) in Sec.\ref{sec6}\,.
\defArnold-Euler geodesics{Strong Dissipation: attracting set size. Lyapunov pairs.}
\section{Arnold-Euler geodesics}
\label{sec8}
\setcounter{equation}{0}
As forcing and dissipation increase the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ may become a
small subset of phase space: and, if the CH holds, it
becomes a smooth surface of dimension {\it lower} than the full dimension
of phase space.
In this case although the motion on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is a Anosov system
it may appear, at first, that it does not even make sense to ask whether
a FR holds because:
\* \noindent(1) it is not possible to think that it could express properties of
the volume contraction on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$: the main difficulty is that time
reversal symmetry, essential for the FR, is lost since, even if the
equations of motion are time reversible, the action of the time reversal $I$
will likely map the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ into a ``repelling'' set $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$
disjoint from ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. The motion on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ remains hyperbolic, as assumed by
CH, but it no longer has the desired symmetry in time. \\
(2) Furthermore the FR deals with the volume contraction on the full phase
space but the hyperbolic character (assumed by the CH) of the motion on the
attracting set could establish, {\it if} for some reason a new time
reversal $\wt I$ were spawned as a symmetry on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, a property of the
contraction of surface elements in ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$; however their analysis would
require a, highly unlikely, detailed understanding the geometry of the
attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. \*
The latter two, seemingly insurmountable, difficulties are however
intertwined and tend, in several cases, to ``compensate''. We begin with
a simple case.
A remarkable property was discovered for a Hamiltonian
evolution with $n$ degrees of freedom for $x=(\V p,\V q)$ with $H(\V p,\V
q)=\frac12\V p^2+V(\V q)$ and {\it subject also} to a friction force $-\n\V
p$. Namely, under very general conditions on the potential $V$ (typically
just boundedness of the surfaces $H=const$), the Lyapunov exponents
$\l_0\ge\l_1\ge \l_{d/2-1}\ge\ldots\l_{2n-1}$ of the motion are such that,
\cite{Dr988},
\begin{equation}}\def\ee{\end{equation} \frac12(\l_j+\l_{2n-1-j})=-\frac\n2\qquad j=0,\ldots,n-1\Eq{e8.1}\ee
In other words the symplectic symmetry of the Hamiltonian systems (which in
absence of friction implies $\l_j+\l_{2n-1-j}\equiv0$) leaves, in presence
of friction, Eq.\equ{e8.1} as a ``remnant'', at least {\it if the friction
force has the simple form} $-\n\V p$. Furthermore Eq.\equ{e8.1} holds
identically for the eigenvalues $\l_j(x)$ of the Jacobian matrix $J(x)$ of
the flow at each point $x$.
{\it Remarkably} the relation Eq.\equ{e8.1} has been extended, \cite{DM996}, to
the time reversible cases in which the friction $-\n \V p$ is replaced by a
force $-\a(\V p,\V q) \V p$ with the multiplier $\a$ such that evolution
conserves the total kinetic energy $\frac12\V p^2$ exactly, as in some of
the simplest thermostat models, \cite{CELS993}, {\it i.e.\ } $\a=-\frac{\V
p\,\cdot\,{\partial}_{\V q} V(\V q)}{\V p^2}$.
In the latter systems Eq.\equ{e8.1} not only holds with $\n$ replaced by
the time average of $\a$ but it follows from the stronger property that the
evolution $t\to S_tx$ is such that, given $t_0>0$, the matrix $W={\partial}_i
(S_{t_0} x)_j$ has the property that the logarithms of the
eigenvalues of $(W^TW)^{\frac12}$
are $t_0$ times
$\l_{t_0,j}(x)>0,\,j=0,\ldots 2n$ (depending on $t_0$), which satisfy
$\frac12(\l_{t_0,j}(x)+\l_{t_0,2n-1-j}(x))=$ $-\frac\n2$ or respectively
$-\frac12\media{\a}$, where the average
is intended over $S_tx$ for $t\in[0,t_0]$.%
\footnote{\small The proof of the pairing symmetry in the above mentioned
cases is that the Jacobian matrix ${\partial}_i (S_t x)_j|_{t=0}$ is seen to be
the sum of the Jacobian for the Hamiltonian flow of $H(\V p,\V
q)-\frac\n2 pq$ plus the identity times $-\frac\n2$,
\cite{Dr988}. In the case of $\a$ a similar property holds replacing
$\n$ with $\a$, as is seen via a calculation. If $J(t)={\partial}_i (S_t x)_j$
then $J(t)PJ(t)^T=P e^{-n\n t}$ (or $P e^{-n\int_0^t\a(x(t))dt}$) where
$P=\pmatrix{0&{\bf 1}\cr{\bf-1}&0\cr}$ (${\bf 1}$ being the $n\times n$
identity) because $J(t){\buildrel def\over=} J_0e^{-\frac12 \n t}$ with $J_0$ a symplectic
matrix so that $ J_0PJ_0^T=P$ (proposition 24, Sec.3.12 in
\cite{Ga983}). Therefore let $v$ be an eigenvector $J_0^TJ_0 v=\l v$ then
the following chain of identities, using $P^2=-1$ shows that $\l^{-1}$ is
an eigenvalue, with eigenvector $Pv$:
$$\eqalign{&J_0^TJ_0 v=\l v \ \to\
P J_0^TJ_0 v=\l P v\ \to\
-J_0^{-1}P J_0 P P v=\l P v \cr
&\to
\ J_0^{-1}J_0^{-1T} Pv=\l Pv\ \to\ (J_0^T J)^{-1} Pv=\l Pv\cr}$$
%
impliyng pairing to $-\frac\n2$ (respectively to $-t^{-1}\int_0^t\frac12
\a(x(t))dt$) for the matrix $(J(t)^TJ(t))^{\frac12}$.
\cite{Dr988,DM996}.\label{OR}}
Eq.\equ{e8.1}, called {\it pairing symmetry}, is certainly very special,
\cite{Ho999}, but it suggests, \cite{BGG997}, that the dimension of the
attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is equal to {\it twice the number of non negative
Lyapunov} exponents: because it suggests that the pairs with two negative
exponents simply correspond to the phase space compression in the
directions that ``stick out'' of the surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$.
If so the latter directions certainly do not contribute to the contraction
of the surface of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$.\footnote{\small Failure to realize the difference
between Anosov motions on the attracting surface versus Anosov motions on
the entire phase space in chaotic systems satisfying the CH together with
time reversal symmetry is responsible for statements,
\cite[p.220]{Ho999}, like: ``{\it If there {\it were} such systems then
it could be proved that they would generate relatively simple
attractors, with equal numbers of positive and negative Lyapunov
exponents. Because the simple geometric argument of Section 7.8 shows
that nonequilibrium attractors are actually generated by {\it any}
stable time reversible, steady dynamics, the applicability of the
Anosov proofs is evidently rare to vanishing}'',
\cite{GC995},\cite{BG997},\cite{Ga998}.}
An arbitrary number of negative exponents\label{Hoo} can be added to any
spectrum by adding arbitrarily many dimensions whose coordinates contract
to $0$. It is only if there is a pairing symmetry that the negative pairs
can be {\it conjectured} to be unambiguously identified: and it can be
hoped that the same remains valid if the pairing is only approximate, which
is a property that is {\it often encountered}, see Sec.\ref{sec18} for
examples.
This idea has been discussed in the analysis of
a simulation dedicated to tests of the CH and FR in a system with
pairing symmetry, \cite[Sec.6]{BGG997}, and its relevance for strongly
dissipative systems like the Navier-Stokes flows has been proposed in
\cite[Sec.5]{Ga997b}, see Sec.\ref{sec18} below.
\*
\noindent{\it Remarks:}
(1) Accepting the above proposal, the dimension of the attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$
is determined when the system has a (possibly approximate) pairing
symmetry, and it is identified as twice the number of non negative Lyapunov
exponents.
\\
(2) It is worth stressing the general difference between the latter dimension,
that will be called {\it fluctuation dimension} of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ (or {\it
fd-dimension}), and the Kaplan-Yorke dimension of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$: the Kaplan-Yorke
dimension (or {\it ky-dimension}) is a measure of the fractal properties of
the SRB attractor contained in the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and it is not
larger than the fluctuation dimension; with which it coincides if the
SRB distribution has a density on phase space. In general the ky-dimension
is a fraction of the fd-dimension.
\\
(3) The above discussion, {\it heuristically} proposes how to determine the
dimension of the attracting surface under the CH when the pairing symmetry
holds.
\*
However it is unclear whether the pairing symmetry, exact or approximate, can be
of any help to address the second of the above difficulties, {\it i.e.\ } the lower
dimension of the attracting set and the accompanying breakdown of time
reversal symmetry for the motions confined to ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, the only ones of
statistical interest, and the consequent apparent irrelevance of the phase
space contraction $\s(x)$, to which {\it also} contribute the contracting
directions sticking out of the attracting surface.
The second of the two difficulties mentioned is addressed in the next section,
on the basis of the proposal in \cite{BG997,Ga997b}.
\defArnold-Euler geodesics{Dissipation. Time Reversal \& FR.}
\section{Arnold-Euler geodesics}
\label{sec9}
\setcounter{equation}{0}
Consider a time reversible evolution depending on a forcing parameter and,
still assuming the CH, suppose the forcing, hence the dissipation, to grow
so strong that the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ becomes a surface of dimension
smaller than that of phase space.
Then the time reversal symmetry $I$ is
{\it spontaneously broken} in the sense that it ceases to be a symmetry for
the motions that develop on the attracting set. It remains a symmetry for
the motions in phase space, but it has little relevance for the statistical
properties (with respect to the SRB distribution) of the motions because,
asymptotically, they are attracted to ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. The time reversal
image $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is quite generally a {\it repeller} and no motion
(except a set of data with $0$ volume) evolves towards it.
Therefore a natural question is whether the continuing existence of the
global time reversal symmetry $I$ can be accompanied by a map $\wt I$ of
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ to itself which is still a smooth isometry,
with $\wt I^2\equiv1$ and $S_t\wt I=\wt I S_{-t}$ in the flow case or $S\wt
I=\wt IS^{-1}$ in the case of maps.
The question has been analyzed in \cite{BG997} where a {\it geometric
property} has been identified which, when holding, shows that a ``local
time reversal symmetry'' $\wt I$, defined as a map of the attracting set
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ into itself, is {\it spawned out} of a global time reversal symmetry $I$,
as a parameter varies and changes the dimension of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ making it a
surface of dimension smaller than the dimension of the phase space $M$.
The latter property will seem at first sight quite special. However it is
enjoyed by a class of systems of interest in applications and at the same
time is a {\it structurally stable property} ({\it i.e.\ } it remains valid under
small perturbations of the dynamics). The property was named ``Axiom C''
because it is a modification of the ``Axiom B'' property introduced in
\cite{Sm967}.
To visualize the geometry of the Axiom C property consider the simpler case
of a time reversible map $S$ and imagine that the attracting
surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ becomes disjoint from its time reversal image $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, because
a parameter controlling the evolution is raised above a critical value, see
Fig.2.
Then the stable manifolds of the points in ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ are not entirely contained
in ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ but extend out of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and intersect $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ on manifolds which
are {\it unstable manifolds} for the points of $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. Likewise the
evolution $S^{-1}$ will have $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ as a attracting set out of which the
$S^{-1}$-stable manifolds of the points of $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ emerge and extend until
they intersect the surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ on its unstable manifolds.
So out of each point $x$ of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ emerge two manifolds intersecting ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$
respectively on the contracting and expanding manifolds at $x$ restricted to
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ for $S_t$ and {\it at the same time} the two manifolds intersect also
$I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and the intersections are the stable and unstable manifolds for
$S$ at some point $x'=Ix$ (linked by a $1$-dimensional curve).
The correspondence $x'=Px$, thus established between ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, {\it
commutes} with the time evolution, because the manifolds whose
intersection defines the correspondence $P:I {\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\,{\kern-1.truept\leftarrow\kern-5.truept\to\kern-1.truept}\,{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ are covariant
under the action of $S$: hence $PSx=S Px$ for all $x\in{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ or $x\in
I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. \*
\eqfig{220}{80}{
\ins{26}{9}{$x$}
\ins{83}{69}{$x'$}
\ins{145}{11}{$x$}
\ins{190}{15}{${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$}
\ins{190}{70}{$I\,{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$}
\ins{144}{63}{$\wt I\, x$}}{fig721}{}
\*
\noindent{\small Fig.2: \it Case of a map $S$. The first figure in Fig.2
illustrates a point $x\in {\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and its attracting manifold, and a local
part of its stable manifold that extends until $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ intersecting it in
the hatched line (stable manifold on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and unstable on
$I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$). Likewise the second figure describes a point $x'$ on $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ with
a local part of its stable manifold for $S^{-1}$ (extending to intersect
${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ on a unstable manifold, hatched). The third figure shows the
($1$-dimensional) intersection between the stable manifold of a point
$x\in {\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ and the unstable manifold of the point $x'\in I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$: in the
figure such intersection is a unidimensional curve that connects $x$ with
$x'$ (uniquely determined by $x$) {\it establishing the correspondence
$P:{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}\to I {\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ defining $P$}, with $x'=P x$. See caption to Fig.2.} \*
The picture requires a few assumptions of technical nature to avoid
occurrence of some more complex possibilities (for instance it is necessary
to {\it exclude} that the contracting manifold emerging from ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ wraps
around $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ rather than meeting it transversally): the mathematical
definition of the ``axiom C'' property can be found in \cite{BG997} and is
a modification of the notion of ``axiom B'', \cite{Sm967}.
A consequence is that the map $\wt I=P I$ maps ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ into itself (as well
as $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ into itself) and is a time reversal symmetry for the restriction
of $S$ to ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ (and to $I{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$).
The above analysis exhibits a structurally stable mechanism, \cite{BG997},
which, if holding, implies that although time reversal is {\it lost as a
symmetry} on an attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, it might be accompanied by a new
map $\wt I$ on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ which can be regarded as a {\it new} time reversal
symmetry for motions evolving on ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. Therefore it is interesting to see
whether a FR can also be established: heuristic ideas about such question,
with attention to a few possible applications will now be presented in the
rest of this section.
In presence of a pairing symmetry to the level $-\frac12\n$, Eq.\equ{e8.1},
suppose that the pairs of negative exponents describe the approach to the
attracting set and call $n_+$ the maximum number of {\it non negative}
Lyapunov exponents. Then the local exponents with labels $j=0,\ldots,n_+-1$
and the corresponding negative ones contribute
$\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(x)=\sum_{j=0}^{n_+}{\l_j(x)+\l_{2n-1-j}(x)}=n_+\n$ to the phase
space contraction and $n-n_+$ pairs of negative exponents should be
discarded in computing $\a_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$; so that the total average phase space
contraction on the attracting set $\s_{{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F},+}$ will be proportional to the
total average phase space contraction ({\it i.e.\ } average $\s_+$ of minus the
divergence of the equation of motion) $\s_{{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F},+}=n_+\n=\frac{n_+}{n} n\n=
\frac{n_+}{n} \s_+$.
Remark that the number $n-n_+$ is defined in terms of the Lyapunov exponents:
hence it does not depend on the point $x$. The conclusion is that in
systems with {\it time reversal and pairing symmetry} satisfying the CH and
axiom C, a fluctuation relation for the surface contraction of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$,
$\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(x)=\frac{n_+}n\s(x)$, holds.%
\footnote{\small It would seem that smoothness
of $\wt I$ should also be required because the axiom C implies only
H\"older continuity for $\wt I$, see \cite{Po010}. However a careful
examination of the FT proof shows that it is sufficient that the
restrictions to ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ of the stable and unstable manifolds of the points
$a\in{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ are smooth manifolds and this is implied by the assumed
smoothness of ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ itself (by CH) and by the smoothness of the global
manifolds.}
Set $\frac{n_+}n\equiv \frac{\NN_{attr}}{\NN}$ with $\NN_{attr}=$
dimension of the attracting surface and $\NN=$ dimension of the phase
space. Then the CH combined with Axiom C and a parity property will give,
for the probability of $\t^{-1}\int_0^\t \frac{\s(S_tx)}{\s_+} d t=
\t^{-1}\int_0^\t \frac{\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(S_tx)}{\s_{{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F},+}} d t$, the relation
\begin{equation}}\def\ee{\end{equation} \frac{P_{srb}(p)}{P_{srb}(-p)}=e^{\t \frac{\NN_{attr}}{\NN}\s_+p + o(\t)}
\Eq{e9.1}\ee
in the notation of Eq.\equ{e6.4}, \cite{BGG997}, because
$\s_{{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F},+}=\frac{\NN_{attr}}{\NN}\s_+$: {\it i.e.\ } the universal constant $1$ in
\equ{e6.4} is replaced by $\frac{\NN_{attr}}{\NN}$.
This covers the FR in systems verifying the pairing rule: {\it but, admittedly,
such systems are not really common in the applications}.
A much larger class of systems can be imagined if the Lyapunov exponents,
arranged as in Eq.\equ{e8.1}, satisfy $\frac12(\l_j+\l_{2n-1-j})=c_j$
with $c_j$ close to a constant ({\it i.e.\ } $c_j\sim C(\frac{j}{2n})$ with $C(\x)$ a
smooth function.
This property arises in a few important cases. For instance in simulations
of reversible models for the 2D incompressible Navier-Stokes equation in
periodic geometry, \cite{GRS004,Ga017a}.
Suppose that the local Lyapunov exponents are such that
$\frac12(\l_j(x)+\l_{2n-1-j}(x))=c_j(x)$ and let $\lis \l_j,\lis c_j$ be
the respective time averages. Then it
can be still imagined that corresponding pairs $\lis \l_j,\lis\l_{2n-1-j}$
of exponents of {\it opposite sign} are exponents concerning the motion on
the attracting set; particularly if the system is close to one for which
the pairing rule holds and $\lis c_j$ is close to a constant.
By an argument similar to the one presented in the pairing symmetrical
cases above, a large deviations relation might be obtained, for the total
contraction $\s(x)$, similar to Eq.\equ{e9.1} with $\frac{\NN_{attr}}{\NN}
\s$ replaced by $\s_{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}(x)=P\s(x)$ with $P{\buildrel def\over=} 1-\frac{\sum_{j\in L_-}
\lis\l_j}{\sum_{j\in L} \lis\l_j}$, where $L$ is the set of Lyapunov
exponents and $L_-$ the subset formed by the pairs of negative exponents.
Leading to a FR {\it with a controlled modification of the slope} in $p$,
at least if the pairing functions $c_j(x)$ can be found, see
Sec.\ref{sec18} for a non trivial example.
The above may apply to time reversible systems with Lyapunov spectrum
obeying a pairing rule at least approximately; and could be extended,
possibly, to irreversible ones if the latter fall under the equivalence
properties mentioned in Sec.\ref{sec7}.
\*
\noindent{\it Remarks} (i) In this respect it is worth coming back to the issue
mentioned in the remark concluding Sec.\ref{sec6}. Which pointed out that
when considering many particle systems, like for instance the steel cube
brought up as an example, the phase space contraction might have an average
too large, due to the size of the Boltzmann's constant. And consequently
the FR would fail to be of any relevance for the fluctuations statistics.
\\
(ii) However in such cases, as in any macroscopic system, the phase space
contraction is very large: but the FR should be applied to the contraction
of the surface of the attracting set. To do so the above axiom C and a
pairing property may be of help: for instance in presence of exact pairing
the FR holds with the $\s_+$ replaced by $\frac{\NN_{attr}}{\NN} \s_+$.
The dimension $\NN_{attr}$ in large systems will usually be $\ll \NN$:
and in order to apply the FR the average phase space contraction must
be corrected by the factor $\frac{\NN_{attr}}{\NN}$.
\\
(iii) Since $\NN$ is, in the case (i), a number of the order of a multiple
of Avogadro's number this is sufficient to turn $k_B T_a$ into $R_0=\NN
k_B$, with $R_0$ being of the order of the gas constant: and this converts
Eq.\equ{e4.5} to $\NN_{attr}\sum_a \frac{\dot Q_a}{R_0 T_a}$
and $R_0T_a$ is no longer very small. \\
(iv) Also $\NN_{attr}$, in the systems like the ones in the example in
Sec.\ref{sec4},\,\ref{sec5}, is typically {\it not of the order of
Avogadro's number}: macroscopic systems often can be described by
macroscopic equations and the number of positive exponents can be
identified with the number of positive exponents of the ``equivalent''
macroscopic equation; the latter very often has a rather small number of
positive exponents so $\NN_{attr}$ can be small: in conclusion the FR could
be applied to classes of system which admit microscopic and macroscopic
equivalent representations.
\\
(iv) A non trivial example is analyzed in
Sec.\ref{sec17}\,,\,\ref{sec18}.
\defArnold-Euler geodesics{Fluctuation Dissipation Theorem}
\section{Arnold-Euler geodesics}
\label{sec10}
\setcounter{equation}{0}
Suppose that a dynamical system equations
\*
\noindent(a) depend on parameters $\V E=(E_1,E_2,\ldots)$ and satisfy a $\V
E$-independent, smooth, time reversal symmetry $I$,\footnote{\small To fix
ideas think of a Hamiltonian system constrained to keep the total kinetic
energy constant, for instance via a Gaussian constraint, as considered in
many applications, \cite{EM990}: in absence of external forcing, and
assuming CH, the SRB distribution is quite generally explicitly known and
equivalent to the canonical distribution, \cite{EM990}.\label{example}}
\\
(b) for $\V E=\V 0$ the equations are supposed to satisfy the CH with
attracting set coinciding with the full phase space (as in the cases in the
footnote \footref{example}).\footnote{\small In other words at $\V E=\V0$ the
evolution is a Anosov system.}
\*
For $\V E\ne\V0$ the equations continue to be time reversal
symmetric with the same symmetry map $I$.\footnote{\small The symmetry
could also depend on $\V E$, becoming $I_{\V E}$, however further
assumptions would be needed, like differentiability, \cite{Po010b}.}
The dynamics is a Anosov system and it remains such at small $\V E$: the
attracting set coincides with the full phase space (by the structural
stability of CH) and the FT holds for the SRB
distributions.\footnote{\small Remark that this is an important case whose
occurrence has been considered ``{\it rare to evanescent}'' in
\cite[p.220]{Ho999}.}
It is therefore interesting to find whether the average phase space
contraction $\s_{\V E,+}$ is a function of $\V E$ with interpretation that
goes beyond its being a quantity associated with universal large
fluctuations of the dissipation. In particular it is interesting to find
an interpretation of the multiple derivatives
${\mbox{\boldmath$ \partial$}}^{\V n}_{\V E}\s_{\V E,+}|_{\V E=\V0}$.
The phase space contraction $\s_{\V E}(x)$, briefly $\s(x)$, will be supposed to have the
Taylor expansion:
\begin{equation}}\def\ee{\end{equation}
\s(x)=\sum_{i=1}^s E_i J^0_i(x)+ O(\V E^2)\Eq{e10.1}\ee
and, having assumed CH, the large deviation rate $\z(p)$ exists (model
dependent) and is analytic in $p$ for $p$ in the interval $(-p^*,p^*)$,
$p^*\ge1$, within which it can vary, \cite{Ga995b,Ga996a}.
On general grounds the function $\z(p)$ is the Laplace transform of
$\l(\b)=\lim_{\t\to\infty} \frac1\t\log\int e^{\b \t (p-1)} P_\t(d p)$
where $P_\t(d p)$ is the PDF of the variable $p=\frac1{\t \s_{\V
E,+}}\int_0^t\s(S_t x)dt$ in the SRB distribution. Once $\l(\b)$ is
"known" then $\z(p)$ is recovered via a Legendre transform; $\z(p)=
\max_\beta\big(\beta\langle\s\rangle_+(p-1)-\lambda(\beta)\big) $,
\cite{Ga995b,Ge998}.
By using the cumulant expansion for $\lambda(\b)$ we find that
$\lambda(\b)={1\over 2!}\b^2 C_2+{1\over3!}\b^3 C_3+\ldots$ where
the coefficients $C_j$ are
$\int_{-\infty}^\infty\langle\s(S_{t_1}\cdot)\s(S_{t_2}\cdot)\ldots
\s(S_{t_{j-1}}\cdot)\s(\cdot)\rangle^T_+\,dt_1\ldots$ if
$\langle\ldots\rangle^T_+$ denote the cumulants of the variables
$\s(x)$.
In our case the cumulants of order $j$ have size $O(G^j)$ with $G{\buildrel def\over=}|\V
E|$, by Eq.\equ{e10.1}, so that:
\begin{equation}}\def\ee{\end{equation}
{\zeta(p)={\langle\s\rangle_+^2\over2 C_2}(p-1)^2+ O((p-1)^3
G^3)}\Eq{e10.2}\ee
\noindent{}(remark that the first term in r.h.s. gives the central limit
theorem). Eq.\equ{e10.2}, together with the FR Eq.\equ{e6.3}, yields at
fixed $p$ the key relations:
\begin{equation}}\def\ee{\end{equation}\langle\s\rangle_+={1\over 2}C_2+ O(G^3)\Eq{e10.3}\ee
Define, \cite{Ga995b,Ge998}: $J_i(x)=\partial_{E_i}\sigma(x)$ = {\it
current}, $L_{ij}=\partial_{E_j}\langle J_i(x)\rangle_+|_{\V E=\V 0}$ =
{\it transport coefficients}; and study $L_{ij}$.
In the {\it r.h.s.}}\def\lhs{{\it l.h.s.} of the first of Eq.\equ{e10.3} discard $O(G^3)$: it becomes
quadratic in $\V E$ with coefficient $\frac12C_2$,
making use of the exponential decay of
SRB-correlations in Anosov systems, given by:
\begin{equation}}\def\ee{\end{equation} {1\over 2}\int_{-\infty}^\infty dt\, \big(\langle J^0_i(S_t\cdot)
J^0_j(\cdot)\rangle_+- \langle J^0_i\rangle_+\langle J^0_j\rangle_+
\big)\big|_{E=0}\Eq{e10.4}\ee
where convergence is implied by the strong mixing properties of the SRB
distribution due to the CH.
On the other hand the expansion of $\langle\sigma\rangle_+$ in the
\lhs of Eq.\equ{e10.3} to second order in $E$ gives:
\begin{equation}}\def\ee{\end{equation}\langle\sigma\rangle_+={1\over2}
\sum_{ij}\big(\partial_{E_i}\partial_{E_j}\langle\sigma\rangle_+
\big)\big|_{E=0} E_i E_j \Eq{e10.5}\ee
because the first order term vanishes, see Eq.\equ{e10.1}.
If $\m_+(dx)$ denotes the SRB distribution, the {\it r.h.s.}}\def\lhs{{\it l.h.s.} of Eq.\equ{e10.5} is the
sum of ${1\over2}E_iE_j$ times $\partial_{E_i}\partial_{E_j} \int \sigma(x)
\mu_+(dx)$ which equals the sum of the following three terms:\\
(i)
$\int\partial_{E_i} \partial_{E_j}\sigma(x) \mu_+(dx)$,\\
(ii)
$\int\partial_{E_i}\sigma(x)\partial_{E_j}\mu_+(dx)+ (i\leftrightarrow j)$
\\
(iii) $\int\sigma(x)\partial_{E_i}\partial_{E_j}\mu_+(dx))$, all
evaluated at $\V E=\V 0$.
The first addend is $0$ (by time reversal), the third
addend is also $0$ (as $\sigma=0$ at $\V E=\V 0$). Hence:
\begin{equation}}\def\ee{\end{equation}\partial_{E_i}\partial_{E_j} \langle\sigma\rangle_+|_{E=0}=
\Big(\partial_{E_j}\langle{J^0_i}\rangle_++
\partial_{E_i}\langle{J^0_j}\rangle_+\Big)|_{E=0}
\Eq{e10.6}\ee
\noindent{}and it is easy to check, again by using time reversal,
that:
\begin{equation}}\def\ee{\end{equation}
\partial_{E_j}\langle{J^0_i}\rangle_+|_{E=0}\,=
\,\partial_{E_j}\langle{J_i}\rangle_+|_{E=0}
= \,L_{ij}\Eq{e10.7}\ee
Thus equating r.h.s and l.h.s. of Eq.\equ{e10.3}, as expressed
respectively by Eq.\equ{e10.4} and Eq.\equ{e10.6} the matrix
${L_{ij}+L_{ji}\over2}$ is obtained, \cite{Ga996}.
At least if $i=j$ this is a ``Green-Kubo formula'', a relation sometimes
called "fluctuation dissipation theorem". It is however {\it very different}
from ``Onsager's reciprocity'' which would be $L_{ij}=L_{ji}$. The latter
will be discusses in the next section.
\defArnold-Euler geodesics{Onsager's Reciprocity}
\section{Arnold-Euler geodesics}
\label{sec11}
\setcounter{equation}{0}
A far reaching extension is necessary to obtain $L_{ij}=L_{ji}$ which will
lead to reciprocity, \cite{Ga996a}, and to further extensions,
\cite{GR997}.
The main remark is that FT theorem can be {\it extended} to give properties
of {\it joint} SRB distribution of $\sigma(x)$ and of the observable
$q(x)=E_j\partial_{E_j}\sigma$. Defining {\it dimensionless }
$j$-current $q=q_j(x)$ (at fixed $j$) as:
\begin{equation}}\def\ee{\end{equation}
{1\over\t}
\int_{-\t/2}^{\t/2}
E_j{\partial}_{E_j} \s(S_tx)dt{\buildrel def\over=}
q
\Eq{e11.1}\ee %
where the factor $E_j$ is there only to keep $\s$ and $E_j{\partial}_{E_j}\s$
with the same dimensions, the really essential property of $q_j(x)$ is
its odd {\it symmetry} under time reversal, as $\s(x)$.
Then if $P_\t(dp,dq)$ is the joint PDF of $p,q$ the {\it same}
proof of the FT in \cite{Si977,GC995,Ga995b} yields also the existence of a
rate function $\z(p,q)$ for $P_\t$ with the symmetry:
\begin{equation}}\def\ee{\end{equation} \z(p,q)=\z(-p,-q)\,+\, p\,\s_{\V E,+}, \qquad
{\rm for\ all} \ p,q \Eq{e11.2}\ee
for the {\it joint large fluctuations} of the variables $\s_{\V
E}(x)$, $E_j{\partial}_{E_j}\s(x)$.
The $\zeta(p,q)$ can be computed, in the same way as $\zeta(p)$ in
Sec.\ref{sec10}, by considering first the transform $\lambda(\b_1,\b_2)$:
\begin{equation}}\def\ee{\end{equation}
\lim_{\t\to\infty} \textstyle
{1\over\t}\log\kern-1pt\int e^{\t(\b_1\, (p-1)
\langle\s\rangle_++
\b_2\,(q-1)\langle E_j{\partial}_{E_j}\s\rangle_+)}P_\t(dp,dq)
\Eq{e11.3}\ee
and then the Legendre transform, abridging the SRB average
$\media{\cdot}_{\V E,+}$ with $\media{\cdot}_+$,
\begin{equation}}\def\ee{\end{equation}\eqalign{
\max_{\b_1,\b_2}\big(&
\b_1\, (p-1)\langle\s\rangle_++
\b_2\,(q-1)
\langle E_j{\partial}_{E_j}\s\rangle_+\cr
&-\lambda(\b_1,\b_2)\big)\,=\,\zeta(p,q)\cr}\Eq{e11.4}
\ee
The function $\lambda({\mbox{\boldmath$ \beta$}})$, ${\mbox{\boldmath$ \beta$}}=(\b_1,\b_2)$, is
evaluated by the cumulant expansion, as above, and one finds:
\begin{equation}}\def\ee{\end{equation}
\lambda({\mbox{\boldmath$ \beta$}})={1\over2}\,\big({\mbox{\boldmath$ \beta$}},
C\,{\mbox{\boldmath$ \beta$}})+O(E^3)\Eq{e11.5}\ee
where $C$ is the $2\times2$ matrix of the second order cumulants. The
coefficient $C_{11}$ is given by $C_2$ appearing in
Eq.\equ{e10.3},\equ{e10.4}; $C_{22}$ is given by the same expression with
$\s$ replaced by $E_j{\partial}_{E_j}\s$ while $C_{12}$ is the mixed cumulant:
\begin{equation}}\def\ee{\end{equation}
\int_{-\infty}^\infty
\big(\langle\s(S_t\cdot)\,E_j{\partial}_{E_j}\s(\cdot)\rangle_+
-\langle\s(S_t\cdot)\rangle_+\,\langle
E_j{\partial}_{E_j}\s(\cdot)\rangle_+\big)dt\Eq{e11.6}\ee
and convergence is again implied by the mixing properties of the SRB
distributions due to the CH.
Hence if $\V w=\pmatrix{(p-1)\langle\s\rangle_+\cr
(q-1)\langle E_j{\partial}_{E_j}\s\rangle_+\cr}$ we get:
\begin{equation}}\def\ee{\end{equation}
\zeta(p,q)={1\over2}\,\big(C^{-1}\V w,\V w\big)+O(E^3)
\Eq{e11.7}\ee
completely analogous to Eq.\equ{e10.2}. But the FT in Eq.\equ{e11.2}, implies
that $\zeta(p,q)-\zeta(-p,-q)$ is {\it $q$ independent}: this immediately
means:
\begin{equation}}\def\ee{\end{equation} -(C^{-1})_{22}
\langle E_j{\partial}_{E_j}\s\rangle_+-(C^{-1})_{21}
\langle \s\rangle_+=0+O(E^3)\Eq{e11.8} \ee
which, from $(C^{-1})_{22}=C_{11}/\det C$, becomes the analogue of
Eq.\equ{e10.3}:
\begin{equation}}\def\ee{\end{equation}
\langle E_j{\partial}_{E_j}\s\rangle_+={1\over2}C_{12}+O(E^3)
\Eq{e11.9}
\ee
\noindent{}Then, proceeding as in the derivation of Eq.\equ{e10.4} through
\Eq.\equ{e10.7} ({\it i.e.} expanding both sides of Eq.\equ{e11.9} {\it to
first order} in the $E_i$'s and using Eq.\equ{e11.8} we get that
${\partial}_{E_i}\langle{\partial}_{E_j}\s\rangle_+$ is given by the integral in
Eq.\equ{e10.4}. This means that $L_{ij}=L_{ji}$ and the general Green-Kubo
formulae follow together with Onsager's reciprocity.
Thus GK, and OR, are in the cases considered here, a consequence of FT and
of its extension, Eq.\equ{e11.2}, in the limit $E\to0$, when combined with
the expansion Eq.\equ{e10.2} for entropy fluctuations. Those theorems and
the fast decay of the $\s\-\s$ correlations, \cite{Si977}, are all natural
consequences of (CH) for reversible systems (which are the
starting point of our considerations). Reversibility is here assumed {\it
both in equilibrium and in non equilibrium}: this is a feature of
Gaussian thermostat models but by no means of all models; the $\V
E$-independence of the reversibility map is also essential but in most
reversible models it is just the velocity reversal map, which is
independent of $\V E$.
Of course the OR and GK only hold around equilibrium, {\it i.e.} they are
properties of $\V E$--derivatives evaluated at $\V E=0$; on the other hand
the expansion for $\lambda(\b)$ is a general consequence of the correlation
decay and the FT also holds for {\it non equilibrium} stationary states,
{\it i.e.} for $\V E\ne\V0$ small\footnote{\small {\it i.e.\ } as long as structural
stability maintains the system an Anosov system.}, and {\it can be
considered a generalization of the OR and GK}.
Evidence for the relation between $L_{ii}$, Green-Kubo formulae, and FT was
pointed out by P.Garrido in \cite{BGG997} in an effort to interpret results
of various numerical experiments and an apparent incompatibility between the
{\it a priori} known non Gaussian nature of the distribution $\pi_\t(p)$
and the "Gaussian looking" empirical distributions; the extension to the
reciprocity followed naturally (see also \cite{Ga996,Ga996a}). In
\cite{BGG997} the situation arising at really large fields, when the
attractor is strictly smaller than the whole phase space, is also
discussed (eventually leading to the analysis in Sec.\ref{sec9} above).
\defArnold-Euler geodesics{Fluctuation Patterns}
\section{Arnold-Euler geodesics}
\label{sec12}
\setcounter{equation}{0}
The derivation of Onsager's reciprocity for reversible Anosov systems
with a time reversal map $I$ smooth and parameters independent (usually
just a ``velocity reversal''), and therefore for systems verifying the Chaotic
Hypothesis, suggests that the fluctuation relations might be extended to
fluctuations of more general observables. At least for small perturbations
of Anosov systems and for smooth Axiom C systems, see caption to Fig.2.
Consider first the fluctuations of the phase space contraction $\s(x)$ and
those of a second observable $\f(x)$ with {\it definite parity} under time
reversal: so $\s(Ix)=-\s(x)$ and $\f(x)=-\f(Ix)$ (or $\f(x)=\f(Ix)$).
Consider a SRB distribution $\m_{srb}$ for the system: let
$\media{\s}_+>0,\media{\f}_+$ be the SRB time averages of $\s,\f$.
Call ``fluctuation pattern'' $\p$ a function on $[0,\t]$:
$t\to \p(t)=(s(t),f(t))$.
The evolution of a point $x$ in phase space such that
\begin{equation}}\def\ee{\end{equation} \eqalign{|&s(t)-\s(S_t x)|<\e,\cr
&|f(t)-\f(S_tx))|<\h\cr} \qquad {\rm for} \ t\in[0,\t]\
\Eq{e12.1}\ee
will be called a motion which {\it shadows the pattern $\p$ in the time
interval $[0,\t]$} and it will be
written
$x{\buildrel\t,\e,\h\over \sim}\p$.
The ``time reversal'' of the pattern $\p$ will be the pattern
$I\p=(-s(\t-t),-f(\t-t))$ (or if $\f(x)$ is even under time reversal
$I\p=(-s(\t-t),f(\t-t))$).
The SRB probability of a trajectory $x\to S_tx$ to follow a pattern $\p$
will be denoted
$P_\t(\{ x{\buildrel\t,\e,\h\over \sim}\p\})$; the argument at the basis of
the Fluctuation Theorem can be applied to study the ratio:
\begin{equation}}\def\ee{\end{equation} \frac1\t \log \frac{P_\t(\{ x{\buildrel\t,\e,\h\over \sim}\p\})}
{P_\t(\{ x{\buildrel\t,\e\over \sim}I\p\})}\Eq{e12.2}\ee
and for reversible Anosov systems leads, at first
surprisingly, immediately to the result:
\begin{equation}}\def\ee{\end{equation} \frac{P_\t(\{ x{\buildrel\t,\e,\h\over \sim}\p\})}
{P_\t(\{ x{\buildrel\t,\e,\h\over \sim}I\p\})}
=e^{\t\, \media{\s}_+\, p + o(\t)}
\Eq{e12.3}\ee
asymptotically as $\t\to\infty$ and to lowest order in the precision $\e,\h$.
More generally several observables can be considered $\s,\f_1,\f_2,\ldots$
and the notion of pattern can be accordingly extended; with the same result
that the ratio of the probability of a fluctuation pattern to that of the
time reversed pattern is $e^{\t\, \media{\s}_+\, p + o(\t)}$, to leading
order as $\t\to\infty$ and in the precision, {\it independent on the
specification} of the fluctuations of $\f_1,\f_2,\ldots$.
Also the $\f$ independence of Eq.\equ{e12.2} implies, given two fluctuation
patterns $\p$ for the observables $\s,\f$ and $\p'$ for the observables
$\s,\ps$, to leading order in $\e,\d,\t^{-1}$:
\begin{equation}}\def\ee{\end{equation} \frac{P_\t(\{ x{\buildrel\t,\e,\d\over \sim}\p\})}
{P_\t(\{ x{\buildrel\t,\e,\d\over \sim}\p'\})}
=\frac{P_\t(\{ x{\buildrel\t,\e,\d\over \sim}I\p\})}
{P_\t(\{ x{\buildrel\t,\e,\d\over \sim}I\p'\})}
\Eq{e12.4}\ee
The above relations show that {\it once the rare event of a sign
change of the entropy production } is realized then the time reversed
patterns have the {\it same relative probability} that they have when
the entropy production has the opposite sign.
In other words to see that time reversed patterns occur it is
``sufficient'' to just {\it change the sign of entropy production} (!):
``no further efforts'' are needed.
In Sec.\ref{sec11} the Eq.\equ{e11.2} has been shown to be essentially
equivalent to Onsager's reciprocity and it is a special case of the
general Eq.\equ{e12.3}: therefore the above Eq.\equ{e12.3} can be
considered an extension of Onsager's reciprocity to stationary
states of time reversible Anosov systems or more generally (if $\s_+$ is
intended as the average area contraction of the attracting surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$)
to systems verifying the CH and the ``axiom C'', see Sec.\ref{sec9}
and caption to Fig.2.
\defArnold-Euler geodesics{Irreversibility time scale}
\section{Arnold-Euler geodesics}
\label{sec13}
\setcounter{equation}{0}
The notion of ``reversible transformation'' between equilibrium states is
defined (often) to be an {\it infinitely slow} transformation through a
sequence of equilibrium states. The latter {\it oxymoron} is not really
satisfactory:\footnote{\small ``\`A la v\'erit\'e, les choses ne peuvent pas se
passer rigoureusement comme nons l'avons suppos\'e ...,
\cite[p.13-14]{Ca824}.} it would be instead desirable to have a
quantitative definition or, better, a way to associate with a
transformation, defined by a given ``protocol'' leading from an equilibrium
state to another, a time scale $\Th$ whose size indicates how long it takes
to realize that the process is irreversible.
Then a reversible transformation should be one with $\Th=\infty$ (to be
interpreted that irreversibility is {\it impossible} to detect). If
$\Th<\infty$ then it should be said that the evolution irreversible nature
is revealed after time $\Th$ which could be taken as the {\it
irreversibility time scale}, \cite[Ch.5-11]{Ga013b}.
So let $\m_0$ be the PDF of an equilibrium state and suppose that the
protocol of action on the system is enforced by a change on the parameters on
which the Hamiltonian depends: like the temperature of an external
thermostat, or the volume available to the molecules, in the case of a gas
enclosed in a container, or like the intensity of a volume force acting on
an incompressible fluid.
The protocol has a duration $\t$ and remains constant afterwards: during
the time $\t$ the system is no longer in equilibrium: the latter is reached
after the time $\t$ elapsed and the system remains isolated or in contact
with thermostats at the same temperature reaching the new equilibrium on a
characteristic time scale $\t'$.\footnote{\small Strictly speaking
equilibrium will be reached after infinite time; however it can be
considered reached for practical purposes after $\t'$, which has the
meaning of a time scale.}
In the following the general system introduced in Sec.\ref{sec4}, see
Fig.1, will be considered to fix ideas. The entropy production is given
by Eq.\equ{e4.5}. It is a quantity {\it with dimension of an
inverse time}, coinciding with the phase space total contraction rate.
In nonequilibrium situations the thermostats temperatures can be time
dependent and also the force $f$ as well as the volume of the container
$\CC_0$ can be time dependent. The thermostats temperatures are fixed
phenomenologically and the mechanism of variation of the stirring forces
and of the geometric variation of the container shape or volume are more
difficult to understand and to model physically.
For instance the variation of the force $f$ can be imagined due to the
varying speed of a paddle, rotating in the gas contained in $\CC_0$, which
in turn can be imagined to be controlled by a motor; but it is impossible
to take into account, without a {\it Daemon} helping, how to keep control
of the direction and intensity of the collisions on the paddle. Hence
assuming that the paddle has constant speed, or that it follows a given
protocol of variation, is a phenomenological assumption.
There are experimental setups in which a paddle, or varying forces, are
present and act on the particles in the container $\CC_0$. Or the external
thermostats temperatures and the volume of $\CC_0$ change following
prescribed paths, {\it e.g.\ } in the case of volume variations due to a moving
piston. Invariably the entropy production is measured via the amount of
work that the motor and the forces perform maintaining (or trying to
maintain) the external force constant, or constraining it to follow a
prefixed protocol, and via the heat ceded to the thermostats.
Here few cases will be considered in which the protocol contemplates only
variations of the external thermostats temperatures or of the volume of the
containers.
Given the general interpretation of the entropy production rate in terms of
phase space volume variation, the case of volume variation in the system of
Fig.1 (Sec.\ref{sec4}) can be treated {\it phenomenologically} by simply
adding to Eq.\equ{e4.5} the quantity $N\frac{\dot V}V$, which is the rate
of variation of the phase space volume $V^N$ allowed to the $N$ particles
in $\CC_0$.
Consider the system in Fig.1, Sec.\ref{sec4}, and express the
total phase space contraction per unit time, \equ{e4.5}, as
\kern-3mm
\begin{equation}}\def\ee{\end{equation}\s(x)\equiv\s_{tot}(\dot{\V X},\V X)= \sum_a \frac{\dot Q_a}{k_B
T_a}-N\frac{\dot V_t}{V_t}-\dot U\Eq{e13.1}\ee
\kern-3mm
Let $[0,\t]$ be the time during which a transformation protocol $\G:\,t\to
(T_a(t),V(t))\equiv P(t)$ acts on an initial {\it equilibrium state} with
SRB distribution $\m_0(dx)$ ({\it e.g.\ } a canonical ``Gibbs distribution''). Then
it is possible to define \*
\noindent(1) $\m_t(dx)=\m_0(S_{-t}dx)$, {\it i.e.\ } the distribution into which
$\m_0$ evolves in time $t$ under the flow generated in phase space by the
equations of motion (remark that $S_t$ is not a group in $t$ because the
evolution is now time dependent). \\
\noindent(2) the $SRB$ distribution $\m_{srb,t}$ corresponding to the stationary
distribution that corresponds to parameters $(T_a,V)$ fixed
(``frozen'') at their value at time $t$, $P(t)=(T_a(t),V(t))$.
\\
\noindent(3) the ``relative'' phase space contraction
\begin{equation}}\def\ee{\end{equation} r(t){\buildrel def\over=} (\lis \s_t -\lis\s_{srb,t})\Eq{e13.2}\ee
where $\lis\s_{srb,t}$ is the time average of the entropy production rate
in the SRB distribution corresponding to the control parameters $P=(T_a,V)$
frozen at time $t$, while $\lis \s_t$ is the average
phase space contraction in the
non stationary distribution $\m_t(dx)$ evolved from $\m_0$.
\footnote{\small The two terms in Eq.\equ{e13.1} have in general different
dependence on particles number $N$: as of $O(N)$ in the case of
volume variations or $O(N^{\frac23})$ if $P(t)$ only involves boundary
temperature variations (hence the heat exchange is a boundary effect).}
\* Assuming the chaotic hypothesis the approach to the SRB states will be
exponential: the state $\m_t$ would evolve under the ``frozen evolution''
exponentially fast, on some time scale\footnote{\small Supposed to be the
same for all $t$, for simplicity.} $\k^{-1}$ to $\m_{srb,t}$. Therefore
the integral (with a inverse time dimension):
\vskip-5mm\begin{equation}}\def\ee{\end{equation} \Th(\G)^{-1}=\int_0^\infty r(t)^2 dt\Eq{e13.3} \ee
will converge, provided the final values of the control parameters are
reached fast enough ({\it e.g.\ } in a finite time, as in actual protocols which
last a prefixed finite time).
The {\it time scale of irreversibility} of the protocol could be defined by
$\Th(\G)$: the larger $\Th$ is, the closer to a quasi static one the
transformation is, as suggested by the following remarks.
A physical definition of ``quasi static'' transformation protocol is a
transformation that is ``very slow'' during its duration time
$\t$. This can be translated mathematically into an evolution in which
$P(t){\buildrel def\over=} (T_a(t),V(t))$ evolves like, if not exactly, as
\kern-3mm
\begin{equation}}\def\ee{\end{equation} P(t)=P(0)+ (1-e^{-\e t})(P(\infty) -P(0))\Eq{e13.4}\ee
with $\e>0$ small.
An evolution $\G$ ``close to quasi static'', but simpler for computing
$\Th(\G)$, would proceed changing $P(0)$ into $P(\infty)=P(0)+\D$ by $\t/\d$
steps of size $\d$, each of which has a time duration $t_\d$ long enough so
that, at the $k$-th step, the evolving system closely settles onto its
stationary state $P(0)+k\d$.
The $t_\d$ can be defined\footnote{\small Remark that the variation of
$\lis\s_{(k+1)\d,+}-\lis\s_{k\d,+}$ is, in general, of order $\k\d$ as a
consequence of the differentiability, \cite{Ru997b}, of the SRB states
with respect to the parameters.} by $e^{-\k t_\d}\ll \k\d$ then by
Eq.\equ{e13.3}:
\begin{equation}}\def\ee{\end{equation} \Th(\G)^{-1}\simeq\, const\, \k^{-1}\,(\lis\s'\d)^2\,\log(\k\d)^{-1}
\Eq{e13.5}\ee
where $\lis\s'$ is an estimate of ${\partial}_t \lis\s_{srb,t}$. Therefore the
``slower'' is the protocol $\G$ ({\it i.e.\ } the larger the time scale $\d^{-1}$ is)
the closer to $\infty$ is the irresversibility scale $\Th(\G)$.
Another way of reading the above: {\it the closer the actual entropy
production $\lis\s_t$ is to the ``ideal'' $\lis\s_{srb,t}$ the longer is
the irreversibility time scale, {\it i.e.\ } the time beyond which the process
cannot be considered reversible.}
\*
\noindent{\it Remark:} particularly interesting are adiabatic processes in which
external forces vary remaining conservative:
\\
(a) an example is an adiabatic
expansion of a gas in a piston. The irreversibility time scale can be
evaluated from the piston velocity, see
\cite[Ch.5]{Ga013b}
\\
(b) a second example is a rarefied gas, with mass $m$ molecules in a {\it
fixed} adiabatic container, subject to a force of potential $mgz$. At
time $0$ the gas is in equilibrium at temperature $\b^{-1}$
and the process $G$ simply raises the acceleration $g$ to $g'>g$ at time
$0$ and then decreases it back to $g$ after a time $\t>0$ (or just stays
$g'$ forever).\footnote{\small This means that the initial potential energy
is $m g N h_0$, where $h_0$ is the height of the center of mass, and
varies at time $0$ to $m g' N h_0$.} Since $\s_t\equiv0$ (by Liouville's
theorem) and $\s_{srb,t}\equiv0$ it is $\Th=\infty$: {\it i.e.\ } the transformation
is {\it reversible} according to the above proposal of reversibility time
scale, as also suggested by Gibbs' entropy constancy in Hamiltonian
evolutions (even when the Hamiltonian is time dependent). Nevertheless the
cycle leads to an intermediate temperature {\it variation} $\d T$ (with
$\frac{\d T}T\simeq\frac{\d g}{g}$, up to finite volume corrections): an
apparent disagreement with the independence on the rapidity of the process,
see Appendix \ref{appC} for details. \*
\defArnold-Euler geodesics{Chaos. Structure of Anosov systems. Their digital codes.}
\section{Arnold-Euler geodesics}
\label{sec14}
\setcounter{equation}{0}
Since the early works on Statistical Mechanics the concept of {\it coarse
graining} played a major role in relating macroscopic and microscopic
descriptions of mechanical systems, \cite{DGM984}.
Anosov systems, through the chaotic hypothesis, offer {\it new}
perspectives. For simplicity here will be considered the case of a discrete
time evolution via a map $S$ on a phase space $M$, which could be a
Poincar\'e's section of either a macroscopic model of evolution (possibly
infinite dimensional, like Navier-Stokes equation) or a microscopic one
(like Newton's equations for $10^{19}$ molecules) or a phenomenological
model (like Lorenz96 or Lorenz63 models or GOY shell model).
What follows can be extended to the case of Anosov flows,
\cite{BR975,Ge998}, essentially by reduction of the problem to the Anosov
maps case by replacing the flow with a Poincar\'e's map between timed
events ({\it i.e.\ } by fixing a surface in phase space and studying the return map
to it). Extension to axiom A maps or flows is also possible,
\cite{Bo970a,Bo975,BR975}.
The discussion below is necessarily somewhat technical as it tries to
convey the reason why Anosov maps lead to stationary states which can be
identified with equilibria of {\it one dimensional spin chains with short range
interactions}: the extreme simplicity of Anosov maps will be manifest after
understanding the formalism. It will reward the necessary time, thus
providing strong support to the statement (see Sec.\ref{sec1}) that Anosov
maps play, for chaotic systems, a role parallel to that of the harmonic
oscillators for ordered dynamics.
A main feature of Anosov maps is that the stable and unstable manifolds of
each point $x$ are smooth manifolds which depend ``almost'' differentiably
on $x$ (they are H\"older continuous and the H\"older exponent can be taken
as close to $1$ as wished, paying the price of a larger H\"older
constant). The manifolds can be used to build ``cells'' in $M$ enclosed
within boundaries which are unions of subsets of stable or unstable
manifolds.
The key remark, \cite{Si968a,Si968b}, is that the phase space $M$ can be
paved with cells, ${\mathcal P}}\def\LL{{\mathcal L}=(P_1,\ldots,P_N)$, which are connected sets,
closures of their interiors, and are either pairwise disjoint or have only
common boundary points; furthermore are ``{\it covariant}'' if transformed
by the map $S$ in the following sense: \*
\noindent(1) the boundary ${\partial} P_j$ of $P_j$ has the form ${\partial}_u P_j\cup {\partial}_s
P_j$ with ${\partial}_u P_j$ consisting of surface elements which are unions of
portions of unstable manifolds and ${\partial}_s P_j$ consisting of surface
elements unions of portions stable manifolds: call ${\partial}_u {\mathcal P}}\def\LL{{\mathcal L}=\cup_j
{\partial}_u P_j$ and ${\partial}_s {\mathcal P}}\def\LL{{\mathcal L}=\cup_j {\partial}_s P_j$. \\
(2) the images $S^{\pm1} P_j$ of the cell $P_j$ will have boundary still
consisting of stable or unstable surface elements (because images of stable
or unstable manifolds are still stable or unstable manifolds) and,
furthermore, will have the {\it covariance property}, see Fig.3:
\begin{equation}}\def\ee{\end{equation} S {\partial}_s P_j \subset {\partial}_s{\mathcal P}}\def\LL{{\mathcal L},\quad
S^{-1} {\partial}_u P_j \subset {\partial}_u{\mathcal P}}\def\LL{{\mathcal L}\Eq{e14.1}\ee
This means that the $P_j$ are so deformed by $S$ (resp. $S^{-1}$) that no
new stable (resp. unstable) boundaries are created. Furthermore the points
$x$ in their evolution will never end up on any of the cells boundaries
with the exception of a set of zero volume ({\it i.e.\ } the set
$\cup_{i=-\infty}^\infty S^i{\partial}{\mathcal P}}\def\LL{{\mathcal L}$).\label{cup}
\\
(3) the ${\partial}_u {\mathcal P}}\def\LL{{\mathcal L},{\partial}_s{\mathcal P}}\def\LL{{\mathcal L}$ have $0$ volume
\*
\eqfig{150}{50}{}{fig0}{}
\noindent{\small Fig.3: \it Very symbolically, as $2$-dimensional squares, a few
elements of ${\mathcal P}}\def\LL{{\mathcal L}$ are shown as an array of squares. An element $P_i$
(shaded, left) of ${\mathcal P}}\def\LL{{\mathcal L}$ is transformed by $S$ into $S P_i$ (shaded,
right) in such a way that the part of the boundary that contracts ends up
exactly on a boundary of some element among $P_1,P_2,\ldots,P_n$. A
similar figure with horizontal and vertical lines exchanged would
illustate the action of $S^{-1}$.} \*
Obviously if ${\mathcal P}}\def\LL{{\mathcal L}$ is a pavement of $M$ with the above properties (1),(2)
then also the pavement whose elements are $P_{i,j}=S P_i\cap P_j$ has the
same properties: hence the hyperbolicity of the map yields that there
exist pavements with elements with diameter smaller than a prefixed
$\e>0$.
Such pavements ${\mathcal P}}\def\LL{{\mathcal L}=(P_0,\ldots,P_n)$ of $M$ are called {\it Markovian
partitions} if the maximum diameter of the $P_i$'s is so small that
the intersection between $S^{\pm} P_i$ and $P_j$ is a connected set:
as inthe figure.%
\footnote{\small Disconnected intersections may happen if the maximum
diameter of the $P_i$ can be dilated by the action of $S$ or $S^{-1}$ to
become larger than the diameter of $M$.}
The hyperbolicity of $S$ implies existence of
Markovian partitions and they can be constructed iteratively,
\cite{Si968b,Bo970a,FR981,GBG004}.
The elements of ${\mathcal P}}\def\LL{{\mathcal L}$ are called ``rectangles'' as they have boundaries
formed by portions of stable and unstable manifolds which in the case of
the simplest Anosov maps, {\it i.e.\ } algebraic hyperbolic maps of the
$2$-dimensional torus, are really quadrilaterals with opposite sides
parallel and equal: algrebraic means that the maps are defined by a
constant matrix with integer entries, no eigenvalue with
modulus $1$ and determinant $\pm1$ (like $\pmatrix{0&1\cr1&1\cr}$).
In dimension $2$, in general, they have the aspect of deformed rectangles
(as the manifolds constituting their boundaries are neither parallel nor
flat) with smooth boundaries. If the map $S$ is an algbraic map of the
$2$-torus ({\it i.e.\ } $S$ is a $2\times2$ matrix with integer entries)
they are rectangles, in the literal sense.
In $\ge 3$ dimensions the intersections between the stable manifolds and
the unstable manifolds meeting at the edges of the rectangle are not
smooth: in general a portion of unstable manifold of dimension $u>1$,
contained in ${\partial} P_i$, may have a boundary\footnote{\small {\it i.e.\ } the
intersection with the stable manifolds in ${\partial} P_i$.} which does not
contain a smooth surface of dimension $u-1$ ({\it e.g.\ } in dimension $3$ and if
the unstable manifold had dimension $2$ it does not contain a
differentiable arc, as one might naively imagine, \cite{Bo978}: {\it i.e.\ } the
edge is not a smooth line).
Likewise a portion of stable manifold, of dimension $s>1$, contained in
${\partial} P_i$, may have a boundary which does not contain a smooth surface of
dimension $s-1$. So the rectangles edges may be {\it quite
rugged}. Nevertheless the boundaries of the sets $P_i$ can be shown to
have {\it zero volume}.
\*
Given a point $x\in M$ its {\it history}
${\mbox{\boldmath$ \sigma$}}=(\s_k)_{k=-\infty}^\infty$, $\s_k=1,2,\ldots,n$, on a Markovian
partition ${\mathcal P}}\def\LL{{\mathcal L}$ is
defined by
\begin{equation}}\def\ee{\end{equation} S^k x \in P_{\s_k}, \quad \forall k\in(-\infty,\infty)\Eq{e14.2}\ee
{\it uniquely} with the exception of the set, with zero volume in $M$, of
the points $x\in\cup_{-\infty}^\infty S^k{\partial} {\mathcal P}}\def\LL{{\mathcal L}$, {\it i.e.\ } except the set of
points which in their evolution fall on the boundary of some of the
$P_i\in{\mathcal P}}\def\LL{{\mathcal L}$. The history is a {\it digital code} for the points of M and the
labels $k$ can, naturally, be called {\it ``times''}.
The history is very convenient as it transforms the evolution $x\to Sx$
into the simple ``translation'': if $x$ has history
${\mbox{\boldmath$ \sigma$}}=\{\s_i\}_{i=-\infty}^{\infty}$ and $x$ evolves into $Sx$ then its history
${\mbox{\boldmath$ \sigma$}}$ evolves into $\t {\mbox{\boldmath$ \sigma$}}=\{\s_{i+1}\}_{i=-\infty}^\infty$.
Define the $n\times n$ ``transitivity matrix'' $T_{\s,\s'}=1$ if there
is an interior point $x\in P_\s$ whose image $Sx$ is an interior point of
$P_{\s^\prime}$ and $T_{\s,\s^\prime}=0$ otherwise. Then only sequences
${\mbox{\boldmath$ \sigma$}}$ with $T_{\s_k,\s_{k+1}}\equiv1$, that will be called ``${\mathcal P}}\def\LL{{\mathcal L}$-compatible'',
can arise as histories of points.
Viceversa given any ${\mathcal P}}\def\LL{{\mathcal L}$-compatible history ${\mbox{\boldmath$ \sigma$}}$ there is at least one
$x\in M$ whose history is ${\mbox{\boldmath$ \sigma$}}$ and the correspondence $x\,{\kern-1.truept\leftarrow\kern-5.truept\to\kern-1.truept}\,{\mbox{\boldmath$ \sigma$}}$ is
one-to-one with the exception of points $x$ in the above mentioned zero
volume set $\cup_i S^i{\partial}{\mathcal P}}\def\LL{{\mathcal L}$. This geometric property follows from
hyperbolicity and the covariance of the boundaries of ${\mathcal P}}\def\LL{{\mathcal L}$, \cite{Ga013b}.
The history ${\mbox{\boldmath$ \sigma$}}$ determines the corresponding point $x$ ``exponentially
fast'', meaning that there is a constant $\k>0$ such that the
$\{\s_i\}_{i=-n}^n$ determines $x$ within $const\,e^{-\k n}$ (and $\k$ can
be taken any smaller than the minimum of the expansion rates for $S$ and
$S^{-1}$).
In Anosov maps there are points with a dense trajectory (see
\footref{hyperbolic} in Sec.\ref{sec1}): hence the compatibility matrix $T$
is ``transitive'', {\it i.e.\ } there is $K>0$ such that $T^K_{\s,\s^\prime}\ge1$
for all pairs $\s,\s^\prime$: this means that among compatible histories it
is possible that any symbol $\s$ is followed by any other symbol
$\s^\prime$ after at most $K$ steps.
The symbolic history can, therefore be used to code the distributions
$\m_0(dx)=\r(x)dx$, with density $\r(x)$ with respect to the volume element
$dx$ in $M$, into {\it stochastic processes}, {\it i.e.\ } into probability distributions
on the space of the compatible histories.\footnote{\small The code that
associates with $x\in M$ the history ${\mbox{\boldmath$ \sigma$}}$ of $x$ is closely related to
the ``structural stability'' of Anosov maps. Structural stability of
Anosov maps means that if an Anosov map $S$ is smoothly perturbed to
become a map $S_\e$, as a function of a parameter $\e$, with $S_0=S$
then, if $\e$ is small enough, also $S_\e$ is an Anosov map and $S_\e$ is
conjugated to $S$ in the sense that there is a H\"older continuous
homeomorphism $\Th$ of $M$ such that $S_\e=\Th\,S_0\,\Th^{-1}$. \\
An essential step to prove this property is to show that a Markovian
partition ${\mathcal P}}\def\LL{{\mathcal L}$ for $S$ can be deformed ``by continuity'' into a Markovian
partition ${\mathcal P}}\def\LL{{\mathcal L}_\e$ for $S_\e$ {\it with the same transition matrix}: so
that the conjugation is the map $x\,{\kern-1.truept\leftarrow\kern-5.truept\to\kern-1.truept}\, x'$ associating pairs with the same
histories: under $S_0$ on ${\mathcal P}}\def\LL{{\mathcal L}$ and under $S_\e$ on ${\mathcal P}}\def\LL{{\mathcal L}_\e$.%
\label{struct-stab}}
\defArnold-Euler geodesics{Volume as stochastic process.
SRB as Ising spin chain equilibrium}
\section{Arnold-Euler geodesics}
\label{sec15}
\setcounter{equation}{0}
The key to the theory of Anosov maps is the {\it representation of the volume
measure} as a probability distribution on the set of compatible sequences,
{\it i.e.\ } as a stochastic process, which in the above case has been proved to be
a ``Gibbs process'' with a short range potential, which {\it however}, in
general, {\it is not translation invariant}, \cite{Si968a,Si968b,Si972a},
see below. A connection with the Gibbs processes emerges naturally also
when attempting to interpret results of simulations,
\cite[Sec.3]{BGG997},\cite{LS999,Ma999}.
Given a Anosov map $S$, its phase space $M$ can be thought as the space of
states of a spin system on a $1$-dimensional lattice: evolution of $x\in M$
being just the shift of the history ${\mbox{\boldmath$ \sigma$}}$ on a Markovian partition
${\mathcal P}}\def\LL{{\mathcal L}=(P_1,P_2,\ldots,P_m)$, see Sec.\ref{sec14}, into which $x$ is
coded. Therefore points of $M$ are still digitally represented although {\it
the usual digital sequences} for their Cartesian coordinates) are abandoned.
\* \noindent{\it Remark:} Representation via histories on Markovian partitions in
not universal, like the the one via the digits of the cartesian
coordinates, but is specifically adapted to the particular dynamical system
$(M,S)$. \*
The normalized volume is then coded into a probability distribution
$\m_{vol}(d{\mbox{\boldmath$ \sigma$}})$ on the space $C({\bf Z})$ of compatible strings. In the
language of Statistical Mechanics, it would be an ``Ising model'', in which
the ${\mbox{\boldmath$ \sigma$}}$'s can be regarded as sequences of {\it spins},\footnote{\small
Here a {\it spin} is a variable that can assume a finite number of
values, {\it e.g.\ } $\s=\pm1$ or $\s=1,2,..,m$.} so that the time label $i$ of
$\s_i\in{\mbox{\boldmath$ \sigma$}}$ becomes the location of the spin on a ({\it one dimensional})
lattice.
The $\m_{vol}$ can be contructed via a function ${\mbox{\boldmath$ \Phi$}}$, called ``{\it
potential}'', defined for all integers $a\le b$
on the finite strings ${\mbox{\boldmath$ \sigma$}}=\{\s_a,\ldots,\s_b\}\in
C([a,b])$ that are compatible ({\it i.e.\ } that are restrictions to $[a,b]$ of
a string in $C({\bf Z})$. The ${\mbox{\boldmath$ \Phi$}}$ has the ``short range'' property, {\it i.e.\ }
${\mbox{\boldmath$ \Phi$}}_{[a,b]}({\mbox{\boldmath$ \sigma$}})$ tends to $0$ {\it exponentially} if $b-a\to\infty$ and {\it
uniformly} in $a$ as
\begin{equation}}\def\ee{\end{equation}||{\mbox{\boldmath$ \Phi$}}||=\sup_{a}\sum_{b\ge a}\sum_{{{\mbox{\boldmath$ \sigma$}}}\in C([a,b])} |{\mbox{\boldmath$ \Phi$}}_{[a,b]}({\mbox{\boldmath$ \sigma$}})|
e^{\k|b-a|}<\infty\Eq{e15.1}\ee
for some $\k>0$: at fixed time $a$ the potential ${\mbox{\boldmath$ \Phi$}}$ is {\it
exponentially localized} at time $a$.
The potential ${\mbox{\boldmath$ \Phi$}}$ will attribute to spin
configurations ${\mbox{\boldmath$ \sigma$}}\in C([-\t,\t])$, an ``energy'':
\begin{equation}}\def\ee{\end{equation} U({\mbox{\boldmath$ \sigma$}},\t)=\sum_{B\subset [-\t,\t]}{\mbox{\boldmath$ \Phi$}}_B({\mbox{\boldmath$ \sigma$}}_B)\Eq{e15.2}\ee
where ${\mbox{\boldmath$ \sigma$}}_B$ is the part of ${\mbox{\boldmath$ \sigma$}}$ with time labels in the interval $B$ and
the summation is over the intervals $B$ in $[-\t,\t]$.
The basic property concerns the set of $x$'s whose history ${\mbox{\boldmath$ \sigma$}}$ restricted
to $i\in\L=[-\ell,\ell]$ coincides with a given ${\mbox{\boldmath$ \sigma$}}_\L\in C(\L)$: by the
definition of history of $x$ this set is simply $P_{{\mbox{\boldmath$ \sigma$}}_\L}=\cap_{k\in\L}
P_{\s_k}$. Fixed ${\mbox{\boldmath$ \sigma$}}_\L\in C(\L)$ the normalized volume
$\m_{vol}(P_{{\mbox{\boldmath$ \sigma$}}_\L})$ is expressed in terms of the potential ${\mbox{\boldmath$ \Phi$}}$ as:
\begin{equation}}\def\ee{\end{equation} \m_{vol}(P_{{\mbox{\boldmath$ \sigma$}}_\L})=\lim_{\t\to\infty}
\frac{\sum_{{\mbox{\boldmath$ \sigma$}}\in C([-\t,\t])}^\L e^{U({\mbox{\boldmath$ \sigma$}},\t)}}
{\sum_{{\mbox{\boldmath$ \sigma$}}\in C([-\t,\t])} e^{U({\mbox{\boldmath$ \sigma$}},\t)}}\Eq{e15.3}\ee
where the superscript $\L$ restricts the sum in the numerator to the
configurations ${\mbox{\boldmath$ \sigma$}}$ coinciding with ${\mbox{\boldmath$ \sigma$}}_\L$ in the sites of $\L$.
The ${\mbox{\boldmath$ \Phi$}}$ is a suitable potential, expressible in terms of the
representation of the expansion and contraction rates at the point $x$ coded
into ${\mbox{\boldmath$ \sigma$}}$, \cite{Si968a,Si968b,Ga013b}.
Eq.\equ{e15.3} can be fairly easily checked in systems of dimension $2$
(particularly if $S$ is an algebraic map of the torus, see also
\cite{FR981}) because the description, see Sec.\ref{sec14}, of the
Markovian partition can be well visualized via geometric drawings, see
Fig.3 Sec.\ref{sec14}, but requires some effort in higher dimension,
\cite{Bo970a}.
Furthermore the potential ${\mbox{\boldmath$ \Phi$}}$ tends asymptotically to become ${\mbox{\boldmath$ \Phi$}}^+$ to
the right of the origin but it becomes asymptotically ${\mbox{\boldmath$ \Phi$}}^-$ to the left
and ${\mbox{\boldmath$ \Phi$}}^\pm$ are translationally invariant.\footnote{\small
If $B=[a,b]$ and $B+t=[a+t,b+t]$ then ${\mbox{\boldmath$ \Phi$}}^\pm_{B+t}({\mbox{\boldmath$ \sigma$}}_{B})\equiv
{\mbox{\boldmath$ \Phi$}}^\pm_{B}({\mbox{\boldmath$ \sigma$}}_{B})$ for all $t$.}
This means:
\begin{equation}}\def\ee{\end{equation} \sum_{B}^\th\sum_{{\mbox{\boldmath$ \sigma$}}_B}|{\mbox{\boldmath$ \Phi$}}_B({\mbox{\boldmath$ \sigma$}}_B)-{\mbox{\boldmath$ \Phi$}}_B^\pm ({\mbox{\boldmath$ \sigma$}}_B)| e^{\k
(|B|+d(0,B))}, \ \th=\pm
\Eq{e15.4}
\ee
where $B$ are intervals $[a,b]$ to the right
of the origin if $\th=+$ or to the left if $\th=-$ (respectively) and
$d(0,B)$ is the distance of $B$ to the origin.
Hence, if
${\mbox{\boldmath$ \sigma$}}=(\s_{-\t},\ldots,\s_{\t})=({\mbox{\boldmath$ \sigma$}}_-,{\mbox{\boldmath$ \sigma$}}_+)$, with
${\mbox{\boldmath$ \sigma$}}_-=\{\s_k\}_{k=-\t}^0$ and ${\mbox{\boldmath$ \sigma$}}_+=\{\s_k\}_{k=1}^\t$,
$U({\mbox{\boldmath$ \sigma$}})$ can be split as
\begin{equation}}\def\ee{\end{equation} U({\mbox{\boldmath$ \sigma$}},\t)= U_-({\mbox{\boldmath$ \sigma$}}_-,\t)+ U_+({\mbox{\boldmath$ \sigma$}}_+,\t) +{\mbox{\boldmath$ \Psi$}}({\mbox{\boldmath$ \sigma$}},\t)\Eq{e15.5}\ee
where $U_\pm({\mbox{\boldmath$ \sigma$}}^\pm,\t)=\sum_{B\subset [0,\pm\t]} {\mbox{\boldmath$ \Phi$}}^\pm_B({\mbox{\boldmath$ \sigma$}}_B)$ and
${\mbox{\boldmath$ \Psi$}}({\mbox{\boldmath$ \sigma$}},\t)$ can be expressed in terms of a potential
${\mbox{\boldmath$ \Psi$}}$ which satisfies a bound like Eq.\equ{e15.1} with ${\mbox{\boldmath$ \Psi$}}_B=0$
unless $B$ contains at least one of the three sites $\pm\t,0$: in other
words ${\mbox{\boldmath$ \Psi$}}$ is a {\it suitable interpolation} between ${\mbox{\boldmath$ \Phi$}}^\pm$ and ${\mbox{\boldmath$ \Phi$}}$.
The limit Eq.\equ{e15.3} exists as a consequence of the $1$-dimensionality
of the ${\mbox{\boldmath$ \sigma$}}$'s, of the short range of ${\mbox{\boldmath$ \Phi$}}^\pm,{\mbox{\boldmath$ \Phi$}}$ and of the absence of
phase transitions in stochastic processes with such potentials: the usual
SM analysis is presented only for the case of translation invariant
potentials, \cite[Sec.5.8]{Ga000}, but it works, essentially word-by-word,
also for non translation invariant potentials like the above ${\mbox{\boldmath$ \Phi$}}$'s. \*
The proof of the Eq.\equ{e15.3} is technical, \cite{Si968b},\cite{Si968a}:
in {\it heuristic} form can be found in Ch.3 of \cite{Ga013b} where it is
discussed together with several important corollaries which are summarized
in the following remarks, see also Ch.6 in \cite{GBG004}. \*
\noindent{\it Remarks}
(1) As a byproduct of the proof of Eq.\equ{e15.3} an interesting expression for
the phase space contraction emerges. Let $x$ be selected in $P_{\mbox{\boldmath$ \sigma$}}$ with
${\mbox{\boldmath$ \sigma$}}=({\mbox{\boldmath$ \sigma$}}^-,{\mbox{\boldmath$ \sigma$}}^+)$ (as above), then the logarithm of the total phase space
contraction in the interval $[-\t,0]$ at $y=S^{-\t}x$ can be expressed by
\begin{equation}}\def\ee{\end{equation} \eqalign{
\t \s_{t[-\t,0]}(y){\buildrel def\over=}&-\log |\det {\partial}_i S^{\t}_j(y)|\cr
=&-U_-({\mbox{\boldmath$ \sigma$}}^-,\t)+U_+({\mbox{\boldmath$ \sigma$}}^-,\t)\cr} \Eq{e15.6}\ee
{\it up to a correction} ${\mbox{\boldmath$ \Psi$}}'({\mbox{\boldmath$ \sigma$}},\t)$ with ${\mbox{\boldmath$ \Psi$}}'$ a potential with the
same properties as ${\mbox{\boldmath$ \Psi$}}$ in Eq.\equ{e15.5}, hence up to a $\t$-independent
constant.
\\
(2) Eq.\equ{e15.6} is a function which has average towards the future equal
to $\t \s_+$ with $\s_+$ being the SRB average of the single step phase
space contraction $-\log |\det {\partial}_i S_jx|$. Eq.\equ{e15.6} says that the
r.h.s. $-U_-({\mbox{\boldmath$ \sigma$}}^-,\t)+U_+({\mbox{\boldmath$ \sigma$}}^-,\t)$ and be used to replace $\sum_{i=-\t}^\t
\s(S^ix)$ up to a correction bounded by a $\t$ independent constant.
\\
(3) A second byproduct, see
Eq.3.8.5, Eq.3.11.2 in \cite{Ga013b},is
\begin{equation}}\def\ee{\end{equation} \frac{\m_{srb}(P_{{\mbox{\boldmath$ \sigma$}}})}{\m_{vol}(P_{{\mbox{\boldmath$ \sigma$}}})}=e^{-\t\s_{[-\t,0]}(S^{-\t}
x) +\ldots}\Eq{e15.7}\ee
where the dots indicate a correction which is bounded by a $\t$-independent
constant: this gives details about the singularity of the SRB distribution
with respect to the volume.
\*
Existence of ${\mbox{\boldmath$ \Phi$}}^+,{\mbox{\boldmath$ \Phi$}}^-$ is behind the theorem on
Anosov maps, \cite{Si968a,Si968b,Si972a}, stating that the SRB distribution
$\m_{srb}$ can be naturally represented as a PDF on the set of compatible
sequences associated with a Markovian pavement ${\mathcal P}}\def\LL{{\mathcal L}$ (any one, as there are
infinitely many of them to choose). From the general theory of the
one-dimensional Gibbs states, and from Eq.\equ{e15.3}, it can be read:
\*
\noindent(a) the SRB is given by
Eq.\equ{e15.3} with ${\mbox{\boldmath$ \Phi$}}={\mbox{\boldmath$ \Phi$}}^+$,\\
(b) the volume distribution has the form in
Eq.\equ{e15.3} with ${\mbox{\boldmath$ \Phi$}}$ in general $\ne{\mbox{\boldmath$ \Phi$}}^\pm$, and
\\
(c) the SRB distribution for the {\it backward evolution}, $S^{-1}$, is
given by Eq.\equ{e15.3} with ${\mbox{\boldmath$ \Phi$}}={\mbox{\boldmath$ \Phi$}}^-$: (a),(b),(c) together imply the
theorem of Sec.\ref{sec1}.
\\
(d) the phase space contraction $\s(x)$ is expressed in terms of the
symbolic history ${\mbox{\boldmath$ \sigma$}}$ of $x$ and of the potentials ${\mbox{\boldmath$ \Phi$}}^\pm$ via
Eq.\equ{e15.6}. This is the key to derive the FT.
\*
With the above ``Ising model interpretation'' of the phase space volume, the
short range nature of the potentials ${\mbox{\boldmath$ \Phi$}},{\mbox{\boldmath$ \Phi$}}^-,{\mbox{\boldmath$ \Phi$}}^+$ and the
$1$-dimensionality of the time ({\it i.e.\ } of the labels of the strings ${\mbox{\boldmath$ \sigma$}}$)
imply, from a SM viewpoint and as a {\it theorem}, that the
volume distribution is a simple stochastic process with very strong
ergodicity properties.
Therefore a randomly chosen point $x$ (except for a set of $x$ in a set
with zero volume) will have a well defined statistics, the SRB statistics,
such that $S^tx$ is coded into a string ${\mbox{\boldmath$ \sigma$}}_t=\{\s_{i+t}\}$ which, for
$t>0$ and large, is a {\it typical string} for the process with the ``future
potential'' ${\mbox{\boldmath$ \Phi$}}^+$, while for $t<0$ and large is a typical string for the
process with the ``past potential'' ${\mbox{\boldmath$ \Phi$}}^-$.
With probability $1$, with respect to the volume measure, or to any one
which has a density with respect to the volume, a point $x$ will generate a
well defined SRB statistics, {\it in general different} for the evolution
$S$ towards the future or for the evolution $S^{-1}$ towards the past. This
explains why in general the SRB distribution for $S$ and that for $S^{-1}$
are singular with respect to each other and to the volume.
The result can be suitably adapted to Anosov flows and also extended to
more general maps or flows, called Axiom A maps or flows,
\cite{BR975,Ge998}.
The structure of Anosov systems as a stochastic process with potential
${\mbox{\boldmath$ \Phi$}}$ is basic in the derivation of the fluctuation relation in
Sec.\ref{sec6}\,; it also indicates a urgent problem: namely that what said
so far might be simply insufficient to define a local phase entropy
production and to formulate a {\it local fluctuation relation} dealing with
some of the fluctuations taking place in a small region.
The problem is interesting as the fluctuations of the phase space
contraction, just because of its physical meaning, will be often
macroscopic quantities which, therefore, will be difficult to observe in
measurements.\footnote{\small Very large fluctuations can hint at
``violations'' of the second principle, \cite{ECM993}, hence cannot be
observed in large systems.}
Nevertheless there is some relation that can be established between the
latter problem and the structure of the just described global SRB
distributions, and it indicates that a fluctuation relation valid for
locally observed fluctuations ({\it i.e.\ } observed in small regions compared to
the system size) might be possible:\footnote{\small The importance of the
problem is made obvious by a few recent experimental works, {\it e.g.\ }
\cite{STX005,VCBCDGGP019}}: more details are deferred to Appendix
\ref{appB}.
\defArnold-Euler geodesics{Entropy\,? Stationarity \& Approach to it}
\section{Arnold-Euler geodesics}
\label{sec16}
\setcounter{equation}{0}
Boltzmann's $H$-theorem for {\it rarefied gases} led to the general
definition of equilibrium entropy as $S=k_B \log W$, as written by Planck,
where $W$ is the volume of phase space where the equilibrium distribution
is concentrated. In the $H$-theorem $S$ is the limit value of the more
general $H$-function, defined even for a nonequilibrium distribution of a
rarefied gas, which reaches its maximum $S$ on the equilibrium state.
Therefore $H$ can be regarded as an extension to nonequilibrium evolutions
(of rarefied gases {\it not in equilibrium}, but isolated and evolving
towards equilibrium) with the main feature that it is a ``Lyapunov
function'' varying with time and approaching (monotonically) a maximum
value, namely the equilibrium entropy.
Recently the Boltzmann's formula $S=k_B\log W$ has been extended to general
evolutions towards equilibrium, \cite{GGL004}, defining appropriately the
volume $W$ as the volume in phase space of the macrostate associated with
the initial microscopic state, determined by a local a coarse grained
empirical density and by the total energy (initial data consisting of
single (typical) phase space points and for a dense gas), and showing that
the new quantity appears to increase monotonically in time (towards an
equilibrium state)
This is different from a natural question arising here:
namely whether an entropy function can be associated with a
nonequilibrium stationary state, and if it even admits an extension to the
evolution towards stationary states which plays the role of a Lyapunov
function.
Going back to the origin of the ergodic hypothesis imagine the phase space
compatible with the constraints {\it as a discrete set of points} located
in the usual continuum phase space.
This is tempting as it would bring back the idea that a phase space point
wanders visiting successively all other points: it would explain the
existence of a unique stationary distribution, to be {\it therefore},
identified with the SRB distribution, which would be simply the
distribution giving {\it equal weight to all points}, whether in isolated
systems ({\it i.e.\ } Hamiltonian evolutions) or in systems out of equilibrium ({\it i.e.\ }
under the action of non conservative forces and thermostats): thus a
unification of the equilibrium and nonequilibrium phenomena would be
achieved.
To discuss the latter question consider a chaotic system {\it defined by a
map} on a manifold $M$ (and satisfying the CH).
Form a Markovian partition ${\mathcal P}}\def\LL{{\mathcal L}$ of the continuum phase space of a system
into finitely many ``cells'' $P_i$ and call $\m_{srb}(P_i)$ the SRB
probability of each set, {\it i.e.\ } the frequency of visit to $P_i$ from a
randomly chosen initial data. $\m_{srb}(P_i)$ is well defined although
singular, {\it {\it i.e.\ } not expressible} in general via an integral over $P_i$
of a density function: hence it is different from the volume $\m_{vol}(P_i)$ of
$P_i$. Then replace the continuum phase space by a {\it finite number} of
points $\NN_0$, with $\NN_0 \m_{srb}(P_i)$ of them in each $P_i\in{\mathcal P}}\def\LL{{\mathcal L}$.
The evolution should be a {\it one cycle permutation} of the phase space
points: in this way each cell $P_i$ is visited, in a very long time, with a
frequency which is, therefore, uniquely determined and is a
representation of the SRB distribution, at least for the computation of the
averages of observables whose variations in the cells $P_i$ can be
considered negligible. The time necessary might even be not be too long if
the cells $P_i$ are not too small and contain a order $1$ fraction of the
total number of (discretized) points.
But in the case of nonequilibrium the equations of motion are no longer
Hamiltonian, and are dissipative. This means that, in general, the
divergence $-\s$ of the equations of motion is not $0$ (as it is for the
isolated evolutions, {\it i.e.\ } in the Hamiltonian cases) and must have a non
negative average $\media\s\ge0$.\footnote{\small The average $\media\s$
cannot be $<0$, {\it i.e.\ } phase space cannot keep expanding forever if a
stationary state can be reached, \cite{Ru996}.} If $\media{\s}>0$ this
means that motion evolves towards an attracting set which has zero volume:
it can be imagined (by CH) dense on a smooth surface ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ of dimension
lower than that of phase space and initial data $x$ starting out of it
evolve in time with their distance to ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ tending to $0$ exponentially
fast.
A discretization of phase space should therefore be a discrete
representation of the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. Under the chaotic hypothesis
heuristic arguments can be developed to {\it estimate the number $\NN$ of
discrete points} necessary to give an accurate description of the motions
of data on the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$, \cite[Ch.3.11]{Ga013b}.
The points on which the dynamics develops can be obtained by covering phase
space with a {\it uniform lattice} with meshes $\d p,\d q$ in momentum and
position and representing the dynamics as a map on the discrete set of
$\NN_0$ points so obtained: then {\it select $\NN\le \NN_0$ of them which
are recurrent}; such points exist, having supposed that the motion can be
represented on a regular discrete lattice and the SRB is ergodic. Here one
should have in mind the {\it numerical simulations} of chaotic dynamical
systems: there the evolution is {\it literally} simulated as a map
(``code'') of a discrete set of points, digitally represented and regularly
spaced.
Then it is natural to try to define entropy of the SRB state the quantity
$S=k_B \log \NN$. A {\it heuristic estimate} of $\NN$, under the CH, has
been proposed in \cite[Ch3.11]{Ga013b} as sketched below.
First refine the Markovian partition ${\mathcal P}}\def\LL{{\mathcal L}=\{P_\s\}_{\s=1}^n$ into $\wh
{\mathcal P}}\def\LL{{\mathcal L}=\vee_{-\t}^{\t} S^k{\mathcal P}}\def\LL{{\mathcal L}$, {\it i.e.\ } define the partition whose elements have
the form $\cap_{i=-\t }^{\t } P_{\s_i}=P_{{\mbox{\boldmath$ \sigma$}}}$, choosing $\t$ so large
that the size of each element is so small that the {\it few observables}
of interest have a {\it constant value} in each $\wh P=P_{{\mbox{\boldmath$ \sigma$}}}$.
Therefore choose $\t$ so that $e^{-\l \t}\d=\d'$ where $\d$ is the maximal
linear dimension of the $P\in{\mathcal P}}\def\LL{{\mathcal L}$, $\d'$ is the maximal linear dimension of
$\wh P\in\wh {\mathcal P}}\def\LL{{\mathcal L}$ and $\l$ is the minimum Lyapunov exponent of $S$: thus
$\t$ depends on the precision
$\t=\l^{-1}\log\frac{\d}{\d'}$.\footnote{\small The CH implies that there
is no vanishing exponent for the map.}
Let $\NN$ be the number of the points on the attractor and $\NN_0$ be the
number of points in the regular lattice over which the dynamics is
discretized: then in a cell $P_{\mbox{\boldmath$ \sigma$}}$ the numbers of points will be
respectively $\NN\m_{srb}(P_{\mbox{\boldmath$ \sigma$}})$ and $\NN_0\m_{vol}((P_{\mbox{\boldmath$ \sigma$}}))$.
Therefore a simple estimate, \cite{Ga013b}, of the number points of the
uniform lattice that must be recurrent to guarantee a ``faithful'' discrete
representation of the dynamics over a time $\t$ is
\begin{equation}}\def\ee{\end{equation} \NN\le \NN_0 \min_{\mbox{\boldmath$ \sigma$}}
\frac{\m_{srb}(P_{\mbox{\boldmath$ \sigma$}})}{\m_{vol}(P_{\mbox{\boldmath$ \sigma$}})}\Eq{e16.1}\ee
where the minimum is over the histories ${\mbox{\boldmath$ \sigma$}}\in C([-\t,\t])$. The
Eq.\equ{e15.7} leads to: $\NN\le \NN_0 e^{-\media{\s}\t}$ with $\media{\s}$
equal to the SRB average of the phase space contraction $\s(x)$. Hence:
\begin{equation}}\def\ee{\end{equation} S=k_B\log\NN\le k_B(\log\NN_0
-\frac{\media{\s}}{\l}\log\frac{\d}{\d'})\Eq{e16.2}\ee
and changing the precision of the observations, {\it i.e.\ } changing the
observables determining $\d'$, $S$ changes by a quantity which depends on
the SRB distribution, if $\media{\s}>0$, via $\media{\s}$ and the smallest
non zero Lyapunov exponent, except when {\it $\media{\s}=0$},
\cite{Ru996}). This provides some evidence that $S$ is not defined just up
to an additive constant.
This is in sharp contrast with the equilibrium result (in which
$\media{\s}=0$) where changing the precision changes $\log\NN$ by a
constant {\it independent of the particular equilibrium state} studied.
And, although the derivation of the estimate is heuristic (and is an
inequality), it seems to indicate that entropy, as a {\it function of
state}, might not be definable for stationary states out of equilibrium,
\cite[Sec.3.10,3.11]{Ga013b}.
Nevertheless one of the main features of the {\it extension} of entropy, as
$S=k_B\log W$, to rarefied gases not in equilibrium but isolated and
evolving towards equilibrium, is that it is a ``Lyapunov function'' varying
with time and approaching (monotonically) a maximum value as a limit value,
namely the equilibrium entropy, \cite{GGL004}.
It is conceivable that {\it also} in the evolution to a stationary state it
could be possible to define a Lyapunov function with the same property of
evolving (possibly not monotonically) to a maximum which is reached at
stationarity, \cite[Ch3.11]{Ga013b}, as briefly discussed below.
Consider as an initial non stationary distribution {\it a delta function on
a single point in phase space}, for simplicity. Then the fraction
$P(\x,t)$ of times that the point $\x\in{\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is visited tends to
$\frac1\NN$, as prescribed by the SRB distribution in the above discrete
representation, where $\NN$ is the number of points in ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$. Therefore:
\begin{equation}}\def\ee{\end{equation} \kern-3mm\eqalign{&S(t)=k_B\sum_\x -\lis{P(\x,t)}\log\lis{P(\x,t)}\cr
&\tende{t\to+\infty} S_\infty=k_B \sum_\x -\frac1\NN \log \frac1\NN=k_B
\log\NN\cr}\Eq{e16.2}\ee
Hence $S_\infty$ is the maximum value that $S(t)$ can
reach: so that $S(t)$ can play the role of a Lyapunov function.
Although $S_{\infty}$ depends {\it non trivially} on the precision of the
discretisation used still, for all choices of the discretisation, the $S(t)$
will have the property of evolving to reach ({\it however not necessarily
monotonically}) the maximum value on the SRB distribution, {\it i.e.\ } on the
natural stationary state. Entropy might be not defined in general
stationary states, as a function of state, although in the approach to
stationarity it could be a Lyapunov function (not unique) extending the
equilibrium entropy function, \cite[Sec.3.12]{Ga013b}.
\defArnold-Euler geodesics{Viscous Fluids and Reversibility}
\section{Arnold-Euler geodesics}
\label{sec17}
\setcounter{equation}{0}
The analysis of the previous sections deal essentially with systems of
particles and leaves out the important class of stationary distributions
that arise in systems normally described via PDE's, but often can be also
described by properties of assemblies of microscopic particles, via
suitable scaling limits, \cite{Pr009}.
This suggests that it should be possible to apply the same ideas to
macroscopic systems, like fluids. Of course the theory of
chaos was developed precisely for such systems,
\cite{RT971,Ru989,Ru989b,Ru995}: however, if systems like fluids are
considered, the reversibility is usually lost in the macroscopic
descriptions.
Yet friction, responsible for the loss of reversibility, is a
phenomenological notion and it can be thought that the same systems could
admit equivalent descriptions via other equations, possibly even reversible.
A key might be the theory of ``ensembles'' for stationary non
equilibrium states, following the proposals considered in
Sec.\ref{sec7}\,-\,\ref{sec9}. An attempt in this direction is presented now
focusing attention on the incompressible Navier-Stokes fluids. A first
step is to propose, via the example of the NS equations, that the
stationary states of macroscopic systems that are scaling limits derived
from microscopic molecular evolutions can be described, in suitable
circumstances, by reversible equations, and equally well.
In the case of the NS equations the proposal goes back, in a related
context, to \cite{SJ993} and, in the form proposed below, to works
summarized in \cite{Ga013b}:
it appeared already in
\cite{GC995,BGG997,Ga996b,Ga997b,Ga000a,Ga000b,Ga001a,Ga006d}.
The classical NS equation in dimension $d=2,3$, for a velocity
${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(\xx)=\sum_{\kk\ne\V 0} e^{-i\kk\cdot\xx} {\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_\kk$, with periodic boundary
conditions in $[0,2\p]^d$, is
\begin{equation}}\def\ee{\end{equation}\dot {\T {\bf u}}}\def\Ff{{\T{\bf F}}+({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}\cdot{\mbox{\boldmath$ \partial$}}) {\T {\bf u}}}\def\Ff{{\T{\bf F}}=\n\Delta {\T {\bf u}}}\def\Ff{{\T{\bf F}} + {\bf f} - {\mbox{\boldmath$ \partial$}} p ~, \quad
{\mbox{\boldmath$ \partial$}} \cdot {\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f} = 0
\Eq{e17.1}\ee
where the external forcing ${\bf f}$ is supposed to be {\it concentrated on the
large scale} Fourier components, actually it will be supposed to have only
one Fourier's component $f_{\pm\kk_0}$ with $|f_{\pm\kk_0}|=\frac1{\sqrt2}$ and
$\kk_0=(2,-1)$, to fix ideas.%
\footnote{\small More generally the forcing can be supposed to have
$f_\kk\ne0$ only for $|\kk|<F$, with $F$ being a fixed cut-off. The cases
$\kk=(0,\pm1)$ and $\kk=(\pm1,0)$ are somewhat trivial, see
\cite{Ma986}.} The equation is not reversible for the time reversal map
$I{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}=-{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ and will be called INS.\footnote{\small Viscosity plays the
role of a model of thermostat: the fluid keeps a constant temperature in
spite of the viscosity: therefore viscosity is a model for the
undisclosed mechanism keeping the temperature constant.}
In the above dimensionless form the viscosity is written $\n=\frac1R$,
where $R$ is usually called ``Grashof's number''. The viscosity is a
phenomenological notion derived from reversible microscopic equations of
motion, \cite{Ma867-b}, and it is possible to think that the coefficient
$\frac1R$ could be replaced by a Lagrange multiplier designed to hold
constant a property characteristic of the flow.
The dissipation per unit time is $\n\DD({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})\equiv\frac1R\DD({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})$ with:
\begin{equation}}\def\ee{\end{equation} \DD({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})=\frac1{(2\p)^2}
\int ({\mbox{\boldmath$ \partial$}}{\T {\bf u}}}\def\Ff{{\T{\bf F}})^2d\xx=\sum_{\kk\ne\V0} \kk^2|u_\kk|^2\Eq{e17.2}\ee
which is called {\it enstrophy}, controls statistical properties
of the flow through its average. Therefore a first proposal is to replace
the viscosity $\frac1R$ with a multiplier such that
${\partial}_t\DD({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})\equiv0$. This leads immediately to
\begin{equation}}\def\ee{\end{equation}\eqalign{
&\dot {\T {\bf u}}}\def\Ff{{\T{\bf F}}+({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}\cdot{\mbox{\boldmath$ \partial$}}) {\T {\bf u}}}\def\Ff{{\T{\bf F}}=\a({\T {\bf u}}}\def\Ff{{\T{\bf F}})\Delta {\T {\bf u}}}\def\Ff{{\T{\bf F}} + \V f - {\T{\BDpr}} p ~, \quad
{\T{\BDpr}} \cdot {\T {\bf u}}}\def\Ff{{\T{\bf F}} = 0\cr
&\a({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})= \frac{\sum_\kk \kk^2\lis {{\bf f}}_\kk\cdot {\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_\kk}
{\sum_{\kk\ne\V0} \kk^4 |{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_\kk|^2}\cr
}\Eq{e17.3}\ee
in space dimension $d=2$. The equation is reversible for the time reversal
map $I{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}=-{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ and will be called RNS.%
\footnote{\small In dimension $d=3$: $\a=\a({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})$ has to be modified by
adding to the numerator of Eq.\equ{e17.3} be quantity $
\sum_{\kk_1,\kk_2} (\kk_1+\kk_2)^2$ $({\T {\bf u}}}\def\Ff{{\T{\bf F}}_{\raise 3mm
\hbox{$\kk_1$}}\cdot i\, {\T\kk}_{\raise 3mm \hbox{$\scriptstyle2$}})
(\lis{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_{\kk_1+\kk_2}\cdot{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_{\kk_2})$.}
Let $N$ be a cut-off and consider the evolutions for INS and RNS
above in $d=2$ for simplicity. Then the evolution equation for
${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_\kk=i\frac{\kk^\perp}{|\kk|} u_\kk$ are
\begin{equation}}\def\ee{\end{equation} \eqalign{
\dot u_\kk=&-\sum_{\kk_1+\kk_2=\kk}
\frac{(\kk_1^\perp\cdot\kk_2)(\kk_2^2-\kk_1^2)}
{2\,|\kk_1|\,|\kk_2|\,|\kk|}u_{\kk_1} u_{\kk_2}\cr
&-\b\, \kk^2u_\kk + f_\kk,\quad |\kk_2|,|\kk_2|,|\kk|\le N\cr}\Eq{e17.4}\ee
%
where $\b=\frac1R$ in the case of INS and $\b=\a(u)$ in the case of RNS and
in both cases $|\V u_{\V k}|=0$ for $|\V k|>N$ or $|\V k|=0$.
The size of the parameter $R$ controls the stability of the evolution: {\it for
simplicity} it will be supposed that if $R$ is large enough then all
initial data, with the exception of a set of zero volume, evolve towards a
unique attracting set and define a unique stationary distribution at least
if $N$ is not too small.
The stationary distributions of the two equations will be parameterized by
$R$ for INS and by $D$, the constant value of the enstrophy. And the
question will be whether there is a correspondence $R\,{\kern-1.truept\leftarrow\kern-5.truept\to\kern-1.truept}\, D$ which
associates distributions which are ``equivalent'', {\it i.e.\ } assign equal
averages to suitable classes of observables. \*
\noindent{\it Remarks:} (1) It is well known that at fixed $N$ it is not true that
there is a unique stationary distribution at given $R$ or $D$: at small
$R$ ({\it e.g.\ } $R<60$) by direct calculations by and accurate simulations this is
shown. In
\cite{BF980,FT979,FTZ984,FT985,FGN988,BF991} the phenomenon of
``hysteresis'', {\it i.e.\ } coexistence of several attracting sets, is discussed in
detail.
//
(2) One of the reasons behind the phenomenon (but by no means the only
one) is the ``gauge'' symmetry of the NS equations: if there is $\V a$ such
that $\kk\cdot\V a=0$ for all $\kk$ for which $f_\kk\ne0$, then if $u_\kk$
is a solution also $e^{i \kk\cdot\V a}u_\kk$ is a solution. Hence if
$f_\kk=c\,\d_{\kk,\pm\kk_0} e^{\pm i\g}, \, c\in R$ also $u_\kk
e^{i\th_0\kk\cdot\kk_0^\perp}$ is a solution: for each $\th_0$ an invariant
set of data is then defined. \\
(3) In the several cases some (or all) of the invariant sets may be stable and
several stationary states will coexist.\footnote{\small For instance if
$f_\kk$ is real and $\kk\cdot \V a=0$ if $f_\kk\ne0$, there is a solution
with $u_\kk$ real: and we have infinitely many invariant sets in which
$u_\kk$ has the form $v_\kk e^{\th_0\,\kk\cdot\V a}$, parameterized by
$\th_0$.}
\\
(4) Symmetry, or more generally existence of more than one attracting surface,
divides the stationary distributions into equivalence classes: and the
particular stationary distribution that is reached starting from a given
initial data $u$ may depend on $u$, see remark (iv) in Sec.\ref{sec6}. In
these instances equivalence may be difficult to check. Unless for all data
$u$, aside a set of zero volume, the stationary distribution is unique and
we are interested only in generic behavior in the space of velocity fields
$u$ {\it with complex components}.
\\
(5) The simplification of uniqueness of the attracting set (on which the CH
holds) that will be used below means that all invariant sets, except one,
become unstable at large $R$.
\\
(6) In the following we shall proceed under the above uniqueness
assumption. However this is not essential: if there are several possible
attracting sets then they will have to be distinguished by labels $\g$ and
the stationary distributions will be parameterized by the labels:
equivalence becomes in this case the existence for both equations of an
equal number of attracting sets and all parameters determining them can be
put in correspondence so that the corresponding stationary distributions
assign the same averages to the local observables. \*
Call $S^{irr,N}_t, S^{rev,N}_t$ the evolutions generated on the phase space
(of dimension $4N(N+1)$ if $d=2$) by the two equations. The SRB
distributions will be parameterized respectively by $R,N$ or by $D,N$ where
$D=\DD(u){\buildrel def\over=}\sum_{\kk}\kk^2 |u_\kk|^2$ is the (constant) enstrophy and
constitute elements of the ``ensembles'' ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{irr,N}$ and ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{rev,N}$
respectively, whose elements will be denoted $\m^{irr,N}_\n$ and
$\m^{rev,N}_D$, respectively.
The discussion in Sec.\ref{sec7}-\ref{sec9} suggests considering the two
collections of SRB distributions and estabilsh a correspondence $\sim$
between $\m^{irr,N}_R\in {\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{irr,N}$ and $\m^{rev,N}_D\in {\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{rev,N}$
by, see Eq.\equ{e17.2},
\begin{equation}}\def\ee{\end{equation} \eqalign{
&\m^{irr,N}_R \sim \m^{rev,N}_D\quad {\rm if} \quad
\m^{irr,N}_R(\DD)=D,\cr
}
\Eq{e17.5}\ee
\noindent{\it Conjecture:} {\it If $O(u)$ is a ``local'' observable, in the sense
that $O$ depends only on the components $u_\kk$ with $|\kk|<K$:
\begin{equation}}\def\ee{\end{equation} \lim_{N\to\infty} \m^{irr,N}_R(O)=\lim_{N\to\infty} \m^{rev,N}_D(O), \
\forall K\ {\rm prefixed}\Eq{e17.6}\ee
provided the equality in Eq.\equ{e17.5} holds as a relation between
$R\equiv\frac1\n$ and $D$.} \*
Multiplying both sides of Eq.\equ{e17.4} by $ u_{-\kk}$ yields that the
time derivative $\dot E$ of the energy $E=\frac12\sum_\kk |u_\kk|^2$ is given by
$-\frac1R \DD(u)+W(u)$ or $-\a(u)D+W(u)$ with $W(u)=\sum_\kk f_\kk
u_{-\kk}$: where $W$ is the work done per unit time by the
external force $\V f$.
Hence since $W$ is a local observable, as $f_\kk$ has been
supposed such, the average of $W$, which will be called $W^a$ for
$a={(rev,N)},{(irr,N)}$ respectively, has to be the same in equivalent
stationary states if $N\to\infty$, {\it i.e.\ }:
\begin{equation}}\def\ee{\end{equation} W^{irr,N}\equiv\n\m^{irr,N}_R({\cal }D), \quad
W^{rev,N}\equiv\m^{rev,N}_{D}(\a) En\Eq{e17.7}\ee
because the average of $\dot E$ has to vanish. Hence the equivalence condition
$\m^{irr,N}_R({\cal D})=D$ immediately implies:
\begin{equation}}\def\ee{\end{equation} R \m^{rev,N}_D(\a)\tende{N\to\infty} 1\Eq{e17.8}\ee
which becomes a key preliminary test of the conjecture when initial data are
randomly chosen and the evolution has a unique stationary state.
And the equivalence condition, if the conjecture
holds, receives the interpretation that the average work done by the
forcing and dissipated per unit time {\it is the same in the two
evolutions}.
In the cases in which there are several attracting sets, hence several SRB
distributions, the conjecture has to be modified (see remarks (iv) in
Sec.\ref{sec6} and (6) above) simply by saying that if $\g,\g'$ are labels
distinguishing the extremal distributions in ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{irr,N},{\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{rev,N}$,
with a given $R$ and the corresponding $D,$ then a correspondence between
$\g$ and $\g'$ is eventually, for $N$ large enough, possible so that
Eq.\equ{e17.8} holds.
If holding, the conjecture would establish a strong analogy between, on one
hand, the theory of the thermodynamic limit of the canonical and
microcanonical equilibrium ensembles and, on the other hand, the above
proposed equivalence of ensembles of SRB distributions for the INS and RNS
equations. The ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{irr,N}$ is analogous to the canonical ensemble with
$\n=R^{-1}$ corresponding to $k_BT$ and ${\mathcal E}}\def\DD{{\mathcal D}}\def\TT{{\mathcal T}^{rev,N}$ is analogous to the
microcanonical ensemble with $D$, the enstrophy, corresponding to the
energy. The observables $O$ play the role of the local observables and
their localization in momentum corresponds to the localization in space in
the thermodynamical equilibrium ensembles.
The above conjecture can be tested and some tests are being made in
simulations. It is also emerging that the conjecture could be strengthened
to cover also the Lyapunov spectra of equivalent elements of the two
nonequilibrium ensembles.
\defArnold-Euler geodesics{Simulations on 2D-NS}
\section{Arnold-Euler geodesics}
\label{sec18}
\setcounter{equation}{0}
Consider the two equations Eq.\equ{e17.4} and fix $R=2048$: the
conjecture stated in the previous section can be tested in simulations. The
cut-off will be set, in the tests that follow, at $960$ Fourier's modes {\it i.e.\ }
$|k_i|\le15$. The first test is to check the Eq.\equ{e17.8}: in all cases
below the evolution is empirically chaotic.
The figure, as well as the subsequent ones, is obtained after running the
{\it irreversible evolution} at $R=2048$, with $960$ modes for a long time
to obtain the average value $D$ for the enstrophy: this realizes the
equivalence condition Eq.\equ{e17.5}. Then the conjecture would predict
that in {\it reversible evolution}, run from an initial data with enstrophy
$D$, the average of $\a(\V u)$ should be $\frac1R$ . The first simulations
yields Fig.4.
\eqfig{200}{120}{}{fig-A01-64191711}{}
\noindent{\small Fig.4:\it Reversible evolution \alert{$NS_{rev}$}: running average
of the ``reversible friction'' $R \a(u)\equiv R\frac{2 Re(f_{-\kk_0}
u_{\kk_0})\kk_0^2}{\sum_\kk \kk^4|u_\kk|^2}$, superposed to the
conjectured value $1$ and to the fluctuating values $R \a(u)$:
R=2048, 960 modes, $\l_{max}$= max. Lyapunov exp. $\simeq
1.5$, integration step $h=2^{-17}$, $x$-axis time unit $4h$,
forcing $f_\kk=0$ except $f_{\pm(2,-1)}= {e^{\pm i\p/3}}/{\sqrt2}$; hence
time unit in abscissa corresponds to $2^{19}$ integration steps: data are
plotted by lines at such time intervals.
Superposed also to the running average of $R\a(u)$ in the
equivalent irreversible NS eq. The two running averages and the line 1
are not easy to distinguish on the scale of the drawing.} \*
A daring test, which goes beyond the conjecture, deals with the equivalence
of the exponents of the Jacobian matrix evaluated in the two equations
under the equivalence conditions: the result for the same truncation of the
equations ($960$ modes) are drawn, on the same frame which reports the
exponents for the Jacobian matrix over a time of the order of
$\l_{max}^{-1}$ where $\l_{max}$ is (an approximation of) the largest
Lyapunov exponent. Such exponents will be called ``local Lyapunov
exponents'',\footnote{\small They are defined in
terms of the diagonal elements of the QR decomposition of the matrix
${\partial}_u (S_\t u)$, linearizing the flow, with $\t$ fixed: in the picture
$\t=h 2^{10}$ was chosen, which is a small fraction of the time unit
fixed by the integration step (which is in the pictures
$h=2^{-17}$). Approximating the matrix ${\partial}_u (S_\t u)$ as $V=(1+h
J(u))^{\t h^{-1}}$, where $J(u)={\partial}_u\dot u$ is the Jacobian matrix of
the flow, the QR decomposition of $V$ gives a triangular matrix $R>0$ and
the logarithms of its eigenvalues divided by $\t$, denoted $\L_k(u)$,
$k=1,2,\ldots$, are sampled as $u(t)$ evolves in time every $4 h^{-1}$
steps and their averages are the definition adopted here of the local
Lyapunov exponents $\l_k$. A much more accurate definition would be
replacing $V$ with the time ordered product $\prod_{k=0}^{\t h^{-1}} (1+h
J(S_k u))$: but this greatly increases the computation time (by a factor
$2^7$ here). Alternatively one could consider the eigenvalues of the
symmetric part of $J(u)$, \cite{Ru982}: but this also requires a large
computation time.\label{QR0}} and are different from the Lyapunov
exponents whose evaluation would require substantially larger computation
time), \cite{Ru982,Li984}.
\eqfig{180}{120}{}{fig-L64191711}{}
\noindent{\small Fig.5:\it The {\it local}, over a time step $4h^{-1}$, Lyapunov
spectra for $960$ modes truncation: {\it reversible and irreversible
superposed}. The sum of the (local) exponents in Fig.4 is $<0$. } \*
The two spectra look quite identical: and the relative difference of
corresponding exponents
($|\l_i^{rev}-\l_i^{irr}|/{\max(|\l_i^{rev}|,|\l_i^{irr}|)}$ is perhaps
more informative:
\eqfig{180}{120}{}{fig-Ldiff191711}{}
\noindent{\small Fig.6: \it Relative difference betweeen (local) Lyapunov exponents
in the previous Fig.5; \alert{R=2048}, $960$ modes. The bar marks the $5\%$
discrepancy, and the lines are visual aids.
}
\*
It is remarkable that the (local) Lyapunov exponents may provide an example
of a pairing rule, see Sec.\ref{sec8}:
\eqfig{180}{120}{}{fig-LC64191711}{}
\noindent{\small Fig.7: \it The approximate pairing rule: graph of
$\frac12(\l_k+\l_{n-1-k})$, $n=960$, with the $\l_k$ the local exponents
in the previous Fig.5; \alert{R=2048}, $960$ modes.
}
\*
A pairing rule emerges from Fig.7. This remarkable fact possibly {\it
suggests} that the pairs consisting of two negative exponents are
associated with the attraction by the attracting set and the dimension of
the latter is therefore twice the number of exponents $>0$, while the
fractal dimension of the attractor is the KY dimension computed using only
the pairs of exponents of opposite sign. In the case of the previous
picture the following Fig.8 provides a detail with a clearer
pairing illustration:
\eqfig{180}{120}{}{fig-Ldet64191711}{}
\noindent{\small Fig.8: \it Detail of Fig.7 showing the pairs of opposite sign and the
ones of equal (negative) sign. The vertical line marks the $k\simeq452$
where the negative pairs begin to appear: hence suggest a dimension of
the attracting set $~904$ out of $960$. Graph of
$\frac12(\l_k+\l_{n-1-k})$, $n=960$, with the $\l_k$ the local exponents
in the previous Fig.5. } \*
The pairing appears exact, but is is not: as it could be seen by drawing
the pairing line on a larger scale. Still even on the scale of Fig.5 it is
not possible to distinguish the pairing line from an exactly horizontal
line.
A pairing property, quite manifest in Fig.5,6, was proposed in
\cite{Ga997b} as possible in NS fluids. It could be an approximate pairing
reflecting an exact one which should hold for the spectrum of the fluid
equations with ``Ekmann friction'' ({\it i.e.\ } with viscosity force $-\n {\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$
instead of $\n\D{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$).
For $\n=0$ the equation can be considered a Hamiltonian equation with
conjugate variables $({\mbox{\boldmath$ \delta$}},{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})$, called Arnold-Euler equation, where
${\mbox{\boldmath$ \delta$}}(x)$ is the displacement (with respect to an initial configuration of
fluid particles) of the ``fluid particle'' that reaches the point $x$ at
the instant in which the fluid velocity at $x$ is ${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(x)$. So ${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ is a
momentum variable while ${\mbox{\boldmath$ \delta$}}$ is a position variable and
$\frac{d}{dt}{\mbox{\boldmath$ \delta$}}(x)={\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(x)$ ({\it e.g.\ } see \cite{Ga997b}).
Formally the Ekmann equation, aside the infinite dimensionality, is covered
by the pairing theorem \cite{Dr988}: hence its Lyapunov spectrum should
have the exponents paired to $-\n$, see Appendix\,\ref{appD}\,. In
\cite{Ga997b} the idea has been proposed that the Lyapunov spectrum of the
$({\mbox{\boldmath$ \delta$}},{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})$-system and that of the ${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ are related by a symmetry which
would imply pairing also for the latter spectrum.
The NS equation viscosity is not proportional to ${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ and the argument in
\cite{Dr988} does not apply, not even formally. Still in \cite{Ga997b} the
possible pairing in the NS spectrum is discussed (called ``barometric
formula'') and for large cut-off is proposed to pair $\l_k,\l_{n-1-k}$ to a
suitable curve $c_k$ (which would be close to a constant in large intervals
of $k$).
A few more simulations have been performed to test the conjecture, all in
2D, because the 3D case is too demanding.
For a few further tests in systems with $48,224,960,3968$ modes ({\it i.e.\ }
increasing the cut-off $N$) and for $R$ up to $8192$, see \cite{Ga019}
where particular attention is dedicated to the approximate pairing, see
Sec.\ref{sec8}, of the Lyapunov exponents. Very few tests have been done
for $R$ small: but the conjecture should hold even in the laminar regimes;
{\it i.e.\ } when at given forcing the attractors can be coexisting stable periodic
motions.
Furthermore changing the forcing to allow a $\V f$ with more than a single
mode, but still keeping it acting only on the large
scale $\V k$'s and of size $||\V f||_2=1$, the average enstrophy can change
substantially but the results on the equivalence remain encouraging.
Also the precision, {\it i.e.\ } the integration time step $h$, can have strong
influence: even hysteresis may appear if $h$ is not small enough even
though it disappears for smaller $h$.
The results are still preliminary and hopefully will be continued not only
to check those so far obtained but also to study further tests and
refinements.
\*
\noindent{\it Remarks:}
(1) Since a real force $f_\kk$ transforms real data $u_\kk$ into real ones,
there will be an invariant distribution concentrated on real velocity
fields $u_\kk$: it may, as $N$ (or $R$ or both) grows, become unstable to
perturbations of $u$ which break the symmetry ({\it i.e.\ } reality of
$u_\kk$). Nevertheless such distribution may be unique among those which
are generated by a real initial $u_\kk$: hence, as mentioned in
Sec.\ref{sec17}, to check equivalence it
becomes necessary, in general, to identify other invariant conditions on
$u$ on which to base the selections of pairs of equivalent distributions
besides the corresponding $R$ and $E$; for instance compare only
distributions concentrated on real $u_\kk$'s.
\\
(2) The simplest checks of the equivalence concern ``gauge invariant''
observables: at least possible different stationary states related by the
symmetry ({\it i.e.\ } that can be transformed into each other by application of the
symmetry) will attribute the same averages to such observable. The
$R\media{\a}_{rev}=1$ is an example and the average work
$\media{W}_{rev}=\media{W}_{irr}$ or also
$\media{|u_\kk|^2}_{rev}=\media{|u_\kk|^2}_{irr}$ are examples.
\defArnold-Euler geodesics{Other relations. Comments.}
\section{Arnold-Euler geodesics}
\label{sec19}
\setcounter{equation}{0}
Several {\it universal} relations have been proposed in the recent
literature. I select below two among them.
\subsection{Transient fluctuation theorem}
Deals, \cite{ES994}, with reversible evolutions starting from random
initial data {\it chosen from an equilibrium distribution} of particles
(hence, in the nontrivial cases, not stationary), of {\it Boltzmann-Gibbs kind},
or more generally from a distribution {\it symmetric under time reversal}
(which, in most cases, is velocity reversal) and with density with
respect to the volume.
In this case the statement is that the probability density that a phase space
volume contracts by a factor $e^A$ compared to the probability that it
contracts by $e^{-A}$ in a time interval $\t$ is such that:
\vglue-6mm
\begin{equation}}\def\ee{\end{equation} \frac{P(A)}{P(-A)}=e^{A}, \quad \forall \t< \infty\Eq{e19.1}\ee
which is an immediate consequence of the definition ({\it i.e.\ } of the above few
lines preceding it): {\it no further assumption is necessary}.
Since Eq.\equ{e19.1} is sometimes compared to Eq.\equ{e6.3} (or
Eq.\equ{e6.4}) then, for the purpose of comparison, the $\s_+$ should be
defined as the average as $\t\to\infty$ of $\frac{A}\t$ and $p$ should be
set $p{\buildrel def\over=} \frac{A}{\t \s_+}$. In terms of $p,\s_+,\t$ Eq.\equ{e19.1}
becomes formally identical to Eq.\equ{e6.4}.
It is claimed that Eq.\equ{e19.1}, being valid for all $\t$, will imply
the fluctuation relation for the stationary state reached by the evolution
at infinite time and for the variable $\frac1{\t\s_+} A$. However the
stationary state in nonequilibrium cases is singular with respect to the
initial state and typical fluctuations observed in time $\t$ have the form
$p\s_+\t$ so, if $\s_+>0$ as it is in nonequilibrium cases, the quantity
$A$ in Eq.\equ{e19.1} has an unclear meaning when the system has reached a
stationary state and a time $\t=+\infty$ has already elapsed.%
\footnote{\small Simple examples of the meaning of Eq.\equ{e19.1} compared
to FR can be constructed: which exhibit systems, {\it as chaotic as
wished}, evolving towards a stationary state with average phase space
contraction $\s_+>0$ and which for every finite time satisfy
Eq.\equ{e19.1} but at infinite time do not satisfy the FR. An example of
such a map follows: let $S_0$ be a map on the unit
circle $\TT$ defined by the evolution at time $t=1$ (say) of
$\dot\f=-\sin\f$: it has $\f=\p$ as an unstable fixed point and $\f=0$ as
a stable fixed point (with Lyapunov exponents $\l_0=\pm1$,
respectively). Let $I$ be the reflection of the point $\f$ at the circle
center. Then the evolution is $I$-reversible and the distribution
$\m_0(d\f)=\frac{d\f}{2\p}$ is $I$-symmetric. Hence Eq.\equ{e19.1} holds
for all finite $\t$: at $\t=\infty$ the distribution of $p=\frac1\t
\sum_{k=0}^\t \cos(S_0^k\f)=\frac1\t A$ evolves to $\d(p)$ which does not
satisfy the FR for any $p>0$ although $\s_+=1$. The example can be easily
adapted to deal with a chaotic evolution: it is enough to consider the
dynamical system acting on pairs $(\f,x)$ which evolve in $(S_0\f,S_\X
x)$, where $\X$ is any Anosov map reversible under a map $J$. This is
reversible under the time reversal $(\f,x)\to (I\f,J x)$. Then
Eq.\equ{e19.1} holds but leads to a relation with slope $\s_+=\s(\X)+1$,
where $\s(\X)$ is the phase space contraction of the map $\X$), while FR
predicts the correct slope $\s(\X)$, because the example is a simple
example of a system with a smooth hyperbolic attracting set ({\it i.e.\ } the
pairs $(0,x)$), hence it satisfies the Chaotic Hypothesis: a case in
which the FR is a theorem. Likewise a flow example can be easily
constructed. The example is due to F.Bonetto.\label{TrFl}}
A proof of any relation between Eq.\equ{e19.1} and the FR discussed in the
present review, in any event,
has never been published, in spite of several announcements.
\subsection{The Jarzinsky relation}
The Eq.\equ{e19.1} is an identity but nevertheless it can be useful, as
shown by its applications in various domains and this might be the
explanation of the lack of interest on the FR and the Chaotic
hypothesis.
In this respect there are other relations which are exact and useful
identities with several interdisciplinary applications in nonequilibrium
phenomena.
An example is provided by an implementation of the simplest ``Monte Carlo
method'': here the general purpose of the Monte Carlo methods is intended
as the use of a controlled random number generator to produce random events
with a prescribed distribution.
For instance suppose that it is necessary to produce spin configurations on
a $N$-points lattice $\LL$ with a distribution proportional to
$e^{-\b U({\mbox{\boldmath$ \sigma$}})},\,\b>0$, with $U({\mbox{\boldmath$ \sigma$}})=\sum_R J_R \s_R$, where $J_R\ge0$ are
are given ``couplings'' for the spins $\s_r=\pm1$ with $r$ in a subset
$R=(r_1,\ldots,r_N)\subset \LL$. Suppose available a random number
generator $G_0$ able to generate a known distribution of ${\mbox{\boldmath$ \sigma$}}$, for
instance a Bernoulli shift $(\frac12,\frac12)$ distribution; then follow
the algorithm, also called a ``protocol'':
\*
\noindent(1) generate a spin configuration ${\mbox{\boldmath$ \sigma$}}$ for the Bernoulli distribution
using a deterministic random number generator $G_0$ (initialized beforehand
once and for all with a fixed number). This plays the role of selection of
initial data from a known initial state (here a sample Bernoulli path). And
compute the weight of the ${\mbox{\boldmath$ \sigma$}}$ in the Bernoulli shift (which in the case
under consideration would be $e^{-\b U_0({\mbox{\boldmath$ \sigma$}})}\equiv 1$, {\it i.e.\ } probability
$Z_0^{-1}=2^{-N}$). \\
(2) compute
$U({\mbox{\boldmath$ \sigma$}})$ and the ratio $e^{-\b (U({\mbox{\boldmath$ \sigma$}})-U_0({\mbox{\boldmath$ \sigma$}}))}$
\\
(3) attribute to ${\mbox{\boldmath$ \sigma$}}$ the weight $e^{W({\mbox{\boldmath$ \sigma$}})}\equiv e^{-\b (U({\mbox{\boldmath$ \sigma$}})-U_0({\mbox{\boldmath$ \sigma$}}))}$.
\*
Repeat the protocol many times: the statistic determined on the outputs
${\mbox{\boldmath$ \sigma$}}_1,{\mbox{\boldmath$ \sigma$}}_2,\ldots$ by assigning them the weights $e^{-\b
(U({\mbox{\boldmath$ \sigma$}}_i)-U_0({\mbox{\boldmath$ \sigma$}}_i))}$ acquires eventually the PDF $P({\mbox{\boldmath$ \sigma$}})=Z^{-1} \exp
-\b U({\mbox{\boldmath$ \sigma$}})$ and the relation
\begin{equation}}\def\ee{\end{equation} \media{e^W}=\frac{Z}{Z_0}\Eq{e19.2}\ee
holds with the average being
taken with respect to the initial distribution ({\it i.e.\ } the Bernoulli shift in
the present case).
The procedure can be used to generate the Gibbs distribution at temperature
$\b$ and Hamiltonian $H_1(x)$ from a Gibbs distribution with Hamiltonian
$H_0(x)$, \cite{Ja997}. Imagine to have at hand a system in equilibrium
with Hamiltonian $H_0(x)$ and a way (``protocol'') to force the evolution
of a configuration via equations of motion following a time dependent
Hamiltonian $H_t(x)$ which evolves from $H_0$ at time $0$ to $H_1$ at time
$1$: \*
\noindent(1) generate an initial state $x$ by picking a sample out
of the initial distribution, and
\\
(2) act, by changing the parameters of the Hamiltonian, so that $x$ evolves
with the time dependent Hamiltonian $H_t$ as $x\to S_{0,t}x$ and keep track
of the energy $W_t(x)=H(S_{0,t}x)-H_0(x)$\\
(3) weigh the output at time $t=1$ with $e^{-\b
W_1(x)}$: eventually the statistics of the weighted outputs will be the
distribution $Z^{-1}e^{-\b H_1(x)}$, as it is immediately checked using the
Liouville theorem $S_{0,t} dx=dx$ (where $dx=dpdq$ in canonical
coordinates).
\*
The above two protocols are realizations of a (naive) ``Monte Carlo''
method: the second can be particularly useful, aside numerical simulations,
even in applications to bio-systems where it has been possible to find a
way to measure $W_1$ at each run of the protocol.
The quantity $W_1$ has been identified, in several cases,
with the work performed on the system during one iteration of the protocol:
it has then been used particularly to measure the free energy variation
between two different equilibria at the same temperature: $\b\D F=-\log
\media{e^{-\b W_1}}$ (with the average being over the statistics of the
initial data), \cite{Ja997}.
Notice that access to $W_1$ is the only requirement necessary: the random
generator being the {\it initial equilibrium state} and the evolution $H_t$
only needs to be always the same, each time the protocol is run. Of course
it is necessary to be able to justify that the measurements of $W_1$ really
evaluate the work $W$ done on the system: in concrete cases it may be not easy
to be sure that all forces are taken into account, in particular the ones
that change $H_0$ to $H_t$. It may also be difficult to make sure that the
protocol used is always exactly the same.
\subsection{Ruelle-Lieb bounds}
There are remarkable {\it rigorous} bounds on the averages, with respect to
the stationary distributions, of the eigenvalues of the symmetric part of
the Jacobian matrix $J$ for the NS equations, symbolically given by
$J_{i,j}=\n\d_{ji}\D-\frac12({\partial}_i u_j+{\partial}_j u_i)$, acting on the
incompressible velocity fields.\footref{QR0} The averages over time of such
eigenvalues are a kind of ``local exponents''. The estimates give an upper
estimate $\wt N_J$ to the maximum number of exponents which add up, ordered
by decreasing size, to a non negative value, hence also an upper bound on
the number of non negative Lyapunov exponents.
The numbers $\wt N_J$ are, \cite{Ru982,Li984}, bounded in dimension
$2$ and $3$ (and more); and {\it in dimension $2$} the bounds can be expressed,
\cite[Eq(34)]{Li984}, in terms of the average
$D=\media{\sum_\kk \kk^2
|u_\kk|^2}_{irr}$ of the enstrophy $D(u)$:
\begin{equation}}\def\ee{\end{equation}\eqalign{
\wt N_J\le&A (2\p)^2 \sqrt{R^2D},\qquad A=0.55...
\cr}\Eq{e19.3}\ee
As seen in Sec.\ref{sec17} $\frac1R D=\media{W}_{irr}$ with $W$ the
power spent by the external force, see comment on Eq.\equ{e17.7}.
The $2$-dimensional estimates are in \cite[Eq/(43)]{Li984}: there are also
found similar estimates, in higher dimension, extending earlier ones in
\cite{Ru982}. In $\ge3$ dimensions the $\wt N_J$ are not bounded in terms
of $ \m^{a}_\cdot(\int dx({\mbox{\boldmath$ \partial$}}{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(x))^2),\ a =rev,irr$, but involve powers
of ${\mbox{\boldmath$ \partial$}}{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ higher than $2$.
The estimates apply to the irreversible NS
equations, truncated at arbitrary ultraviolet cut-off, and involve only the
eigenvalues of $\frac12(J+J^*)$ averaged in time: which can be evaluated in
simulations. Being rigorous they can be important in checks of the accuracy
of simulations. In the reversible equations $R^2 D$ should be replaced
by $\media{\a(u)^{-2} D(u) \ch(\a(u))}_{rev}$ where $\ch(z)=1$ if $z\ge0$
and $0$ otherwise.
\subsection{Wishes}
Several tests of the FR have been performed in the literature.
Unfortunately the FR is confused with the similar relation called above
the ``transient fluctuation theorem'' (possibly even omitting the
qualification of transient).
It is quite interesting that in most {\it simulations} the tests performed
really deal with the FR; hence it would be very interesting to mount
experiments to test the FR (in nonequilibrium situations).
Many experimental works, very delicate and difficult, that claim to have
tested the FR have, unfortunately, instead only tested the above transient
relation; and could be, {\it perhaps even easily}, devoted to a real test
of FR. Or the tests have been devoted to check the linearity in the
symmetry relation Eq.\equ{e6.3} for the fluctuations of a quantity identified
with $p$ but neither attempting to check the relation with the phase space
contraction nor examining the validity of the assumption that the attracting
set has dimension equal to the of phase space.
Beautiful laboratory experiments on nano materials,
proteins, granular materials, ... have been performed and are, very often,
remarkable and innovative from the technical view point but in all cases,
that I am aware of, at best they test the transient theorem.
The study of the FR has not yet attracted sufficient interest, with the
notable exception of the many numerical simulations.
The check of the fluctuation relation in stationary states of systems in
nonequilibrium is a test of the Chaotic Hypothesis, which is a {\it
physical assumption}, unlike the checks of the transient fluctuation
relation (which {\it per se} does not test any physical assumption, because
testing an identity or a theorem does not provide new information).
Also some experimental works limit the analysis to studying the PDF of the
work done or of the heat arising in the experiment but, in my view, not
always enough attention is dedicated to check that all forces acting are taken
into account. This leads, sometimes, to claim (more or less openly) that
the PDF of the work or heat generated at temperature $T$ in a process
cannot be normalized with $k_B T$, as in Eq.\equ{e4.3}: therefore, it is
concluded, the stationary FR is false.
The problem seems to be a certain {\it resilience} to invest time to follow
the ideas behind the chaotic hypothesis ({\it i.e.\ } Anosov systems) and the
general Axiom A attractors: an example of the attitude towards these ideas
is in the quoted statements in \cite{Ho999}; see also Appendix \ref{appA} below.
In spite of all the above remarks I still hope that the FR relation will be
tested in ``real'' experimental contexts.
\APPENDICE{1}
\defArnold-Euler geodesics{About certain comments on CH}
\section{Arnold-Euler geodesics}
\label{appA}
In the above sections quotes from \cite{Ho999} have been reproduced
without really commenting them. The reason is that the quotes were written
when the Chaotic Hypothesis had been just developed and many had not
yet had the time to really study the subject.
But the same comments have appeared in a second edition of the just quoted
book, \cite{HG012}, which I have seen only very recently (after completion
of the present text). Since the comments have had some resonance the next
few lines try to clarify some of the issues.
The Author of \cite{HG012} criticizes the use of the Anosov systems as
paradigm of chaotic motions. The full section from p.344 to p.347 discusses
the merits and demerits of Anosov systems. On p. 344 begins
\*
\noindent{\it ... has discussed the possibility that the useful
properties exhibited by certain oversimplified and quite rare dynamical
systems, termed "Anosov systems", have counterparts in the more usual
thermostatted systems studied with nonequilibrium simulation
methods. Anosov systems are oversimplifications, like square clouds or
spherical chickens...}
\*
This seems to refer to the proposal that the ``Axiom A'' systems should be
the right paradigm for generic chaotic systems, \cite{Ru989,Ru999}: a
proposal which however is not centered on Anosov systems. The Axiom A
systems are systems which have an the attracting set ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ on which motion
has strong chaotic properties (is essentially hyperbolic).
And the CH just proposes, in its final formulation, (1996), that for many
purposes the axiom A paradigm can be strengthened and simplified by
requiring {\it in addition} that ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ is an attracting surface, possibly
of dimension lower than that of phase space, on which the motion is an
Anosov system. Even in time reversible cases ${\mathcal A}}\def\CC{{\mathcal C}}\def\FF{{\mathcal F}$ {\it can be different
from its time reversal image}. This is explicitly stated with related
problems and examples in \cite{BGG997,BG997} and in several successive
publications.
The underlying idea being that it is not possible to distinguish, in a
system of physical interest, a fractal of Hausdorff dimension
$=10^6+3.1415...$ from a surface of exactly $10^6$ dimensions.
In summary the Chaotic Hypothesis only assumes that the dynamics under
consideration behaves (in some respects) like Anosov dynamics. This is
after all not too astonishing if the most relevant
degrees of freedom are chaotic like those of Anosov systems.
Most of the subsequent criticism in \cite{HG012} is anchored on keeping the
identification between the CH and the proposal that the whole dynamical
system is an Anosov system.
On p.346 the fluctuation theorem is called a ``retrospective result'' and
identified with the true Fluctuation Theorem, Sec.\ref{sec6} above,
claiming: \*
``{\it These same "results" were actually given earlier by Denis Evans and
several of his coworkers, for more general circumstances and through more
elementary arguments.}''
\*
\0but no reference is made here to the applicability of the ``earlier
retrospective result'' to stationary nonequilibria to which the Fluctuation
Theorem applies, see\footref{TrFl}.
Then on p. 347 the view is found that:
\*
``{\it Theoretical constructs such as ``measures'', should be viewed with a
healthy suspicion until algorithms for evaluating them are supplied. The
chaos inherent in {\it interesting} differential equations guarantees that
our only access to the "strange sets" which constitute attractors and
repellers will be representative time series from dynamical simulations. In
no way can we construct, or even conceive of constructing, a
Sinai-Ruelle-Bowen measure for an interesting system.}''
\*
\0However for most purposes {\it by the CH Hamiltonian systems should be
considered Anosov systems} (literally, except of course the integrable
ones). Hence the assumption that the attracting set is the full phase
space is not always unreasonable.
Furthermore it is useful to stress that there are easy examples of systems
satisfying {\it the CH, with equal or disjoint attracting and repelling
surfaces, time reversible, with as many degrees of freedom and negative
Lyapunov exponents as wished (unrelated to the number of positive ones)
and whose SRB measure is explicitly and completely constructed},
\cite[Sec.10.2]{GBG004}.
\APPENDICE{2}
\defArnold-Euler geodesics{Local Fluctuations. An example.}
\section{Arnold-Euler geodesics}
\label{appB}
\setcounter{equation}{0}
The phase space contraction in the evolution of a macroscopic system
is typically a macroscopic quantity: whether it is the amount of heat
ceded to the thermostats or the amount of work performed by the systems.
Therefore the average phase space contraction $\s_+$ which controls the
large fluctuations, Sec.\ref{sec6} and the occurrence of ``anomalous''
patterns, Sec.\ref{sec12}, cannot be really observed in measurements on
macroscopic systems.
Avoiding comments on the many experimental fluctuations observations which
claim to check the FR,\footnote{\small Sometimes claiming to have checked
it and sometimes claiming the opposite, while very often dealing with
unrelated transient phenomena.} the question asked here is whether a kind
of fluctuation relation could be defined, and constrain quantities
depending on events that can be observed in very small parts of the system.
In other words is it possible to give a meaning
to a {\it local fluctuation relation}? \cite[Ch4.9]{Ga013b}
The following relies on Sec.\ref{sec15}: it is inserted as it provides a
quite interesting example on how to make use of the symbolic dynamics
representation of the Anosov systems.
A simple example, in a system with time reversal symmetry, will be
discussed in which a {\it local entropy production rate} can be defined and
checked to satisfy a local version of FR. A general view on the matter can
be found in \cite{Ga998b,Ga997c}.
The analysis deals again with maps rather than flows.\footnote{\small This
time {\it the reason is not ``for simplicity''}, but because in the case
of {\it coupled flows}, even if the coupling has short range and is weak,
there seems to be no detailed and constructive general theory of the SRB
distributions, because no simple conditions are known that, via
perturbation techniques, yield hyperbolicity of the flow and allow
studying its properties. For a glimpse on the kind of
complications which arise when studying flows consult,
\cite{RW001,GGG006}. Instead, at least in the case of coupled maps,
the theory is quite well understood, \cite{PS991,BK995,BK997,GBG004}, as
in the example in Eq.\equ{eB.1} below, at small coupling $\e$.}
Consider a system with a translation invariant spatial structure, {\it e.g.\ } a
periodic chain, or a $d$-dimensional square lattice $[-L,L]^d$ with
periodic boundary, of $(2L)^d$ weakly interacting Anosov maps.
The phase space of the system is $M=\{\V x=(\ldots
x_1,x_0,x_1,x_2,\ldots)\}=\MM_0^{(2L)^d}$, where $x_i\in \MM_0$ are points in
a manifold $\MM_0$: to fix ideas we {\it take $\MM_0$ to be a torus}, on which
an Anosov map $\lis S_0$ acts; then define the ``coupled map'':
\begin{equation}}\def\ee{\end{equation} \lis S_\e(\V x)_i=\lis S_0 x_i+\e g(x_{i-1},x_i,x_{i+1}), \ i=...,0,1,\ldots
\Eq{eB.1}\ee
where $g$ is a smooth perturbation, {\it i.e.\ } a smooth periodic function on
$\MM_0^3$.
If $\e$ is small and the perturbation has short range it is proved in
\cite{PS991,BK995,BK997} that, defining the map $\lis S_\e$ as in
Eq.\equ{eB.1} with periodic boundary condition ({\it i.e.\ } identifying the site
$-L $ with $L $), the map $\lis S_\e$ remains, if $\e$ is
small enough, still Anosov. It is conjugated to $\lis S_0$, via a H\"older
continuous correspondence ${\mbox{\boldmath$ \Theta$}}_\e$, see\footref{struct-stab}, by
associating points $x$ and $x'$ with the same history under $\lis S_0$ and
$\lis S_\e$. Furthermore there is $\e_0>0$ such that the above holds for
$|\e|<\e_0$ {\it uniformly} in the system size $L$.
\footnote{\small
The SRB distribution for the evolution $\lis S_\e$ with $L=\infty$ could
also
be
defined via the SRB distribution $\m_{srb,\e}^L$ for the system $\lis S_\e$
consisting of the sites labeled from $-L $ and $L $ and then taking the
limit $\m_{srb}=\lim_{L\to\infty}\m_{srb,L}$, \cite{PS991}.
\\
This is possible because of the uniformity in $\e<\e_0$: below, however,
$L<\infty$ will be fixed, keeping in mind that the results will hold for
all $L$ if $|\e|<\e_0$.}
Here the purpose is to study whether a local version of the FR can hold at
least in an example derived from $\lis S_\e$: but the $\lis S_\e$ is, in
general, not reversible. A related reversible map $S^{rev}_\e$ can be
easily constructed on the ``doubled'' phase space $\MM=\MM_0\times\MM_0$ by
setting:
\begin{equation}}\def\ee{\end{equation} S_\e^{rev}(\V x,\V y)= (\lis S_\e(\V x),
(\lis S_\e)^{-1}(\V y))\Eq{eB.2}\ee
which is reversible for the time reversal map $I:(\V x,\V y)=(\V y,\V
x)$. In the rest of this section this system will be considered in more
detail.
A Markovian partition ${\mathcal P}}\def\LL{{\mathcal L}^L_0$ for $\lis S^L_0\times (\lis
S^L_0)^{-1}$ will be chosen to be the product of partitions ${\mathcal P}}\def\LL{{\mathcal L}_{-\frac
L2},\ldots,{\mathcal P}}\def\LL{{\mathcal L}_{\frac L2-1}$ for the single site maps $\lis S_0$ and
$\lis S_0^{-1}$; and ${\mathcal P}}\def\LL{{\mathcal L}^L_\e$ will be the partition ${\mbox{\boldmath$ \Theta$}}_\e{\mathcal P}}\def\LL{{\mathcal L}^L_0$
existing and defined by the structural stability map ${\mbox{\boldmath$ \Theta$}}_\e$, conjugating
$\lis S_0^{rev}$ to $\lis S_\e^{rev}$, \cite[Sec.10.2]{GBG004}.
Hence the history of a point $x$ will be a sequence of labels $\s_{i;j}$
with $i\in M$ and $j\in (-\infty,\infty)$: naturally
$i$ can be called a ``space label'' while $j$ a ``time label''. The
superscript $rev$ will be omitted in what follows, to simplify notations.
The analysis in Sec.\ref{sec15} applied to the Anosov map $S_\e$, will give
a representation of the volume distribution $\m_0$ and of the SRB
distribution $\m_{srb,\e}$ for $S_\e$ via, respectively, suitable
potentials ${\mbox{\boldmath$ \Phi$}}_\e,{\mbox{\boldmath$ \Phi$}}_{\e}^\pm$.
Let $[0,\t]$ be a time interval and $\L=[-\frac12L,\frac12L]^d=M$,
$|\L]=L^d$. Via the Jacobian matrix $J_\L(x)={\partial}_x (S_\e x)$ define the
{\it phase space contraction} and the {\it time averaged contraction} per
site as, respectively:
\begin{equation}}\def\ee{\end{equation} \eqalign{
\h_{\L,\e}(x)&=-\frac1{|\L|}\log |\det ({\partial}_x
(S_\e x))|\cr
\h_{\L,\e,\t}&=\frac1{|\L|}\lim_{\t\to\infty}
\frac1\t\sum_{j=0}^\t\h_{\L,\e}(S_{\e}^j x)\cr
}\Eq{eB.3}\ee
The limit of $\h_{L,\e,+}$ in Eq.\equ{eB.3} as $\t\to\infty$ exists with
probability $1$ with respect to the volume $\m_{vol}$, as well to the SRB
distribution $\m_{srb,\e}$, and is $x$-independent aside $x$'s in a set of
$0$ volume: because the statistical properties of the volume distribution
are those of the Gibbs distribution with potential ${\mbox{\boldmath$ \Phi$}}^+_\e$, hence enjoy
strong ergodicity properties, as any SRB distribution, with respect to time
translations.
\footnote{\small Likewise the space-time limit $\h_{\e,+}$ exists because
of the space-time ergodicity of the short range Gibbs processes
describing the volume as well as the SRB distributions.}
The phase space contraction $\sum_{t=0}^\t \h_{L,\e}(S_\e^t x)$ can be
expressed, see Sec.\ref{sec15}, via the potentials
${\mbox{\boldmath$ \Phi$}}^+_{\e},{\mbox{\boldmath$ \Phi$}}^-_{\e},{\mbox{\boldmath$ \Phi$}}_{\e}^\f$, where ${\mbox{\boldmath$ \Phi$}}^\f_{\e}$ is a potential
that describes the interpolation between ${\mbox{\boldmath$ \Phi$}}_{\e}^-$ to ${\mbox{\boldmath$ \Phi$}}_{\e}^+$ and
which is therefore ``localized'' (see comment to Eq.(\equ{e15.1})) in the
sense that ${\mbox{\boldmath$ \Phi$}}^\f_{\e}({\mbox{\boldmath$ \sigma$}}_I)\ne0$ only if $I$ contains the sites $0$ or
$L$ and $|{\mbox{\boldmath$ \Phi$}}_{\e,I}^\f({\mbox{\boldmath$ \sigma$}}_{I})|\le C e^{-\k |I|}$ for some $C,\k>0$).
Given the symbolic history ${\mbox{\boldmath$ \sigma$}}$ of $x$, the Eq.\equ{e15.6} can be
expressed as:
\begin{equation}}\def\ee{\end{equation}\eqalign{
\frac1{\t L^d}
\sum_{K\subset M\times [0,\t]}
({\mbox{\boldmath$ \Phi$}}^+_{\e,K}({\mbox{\boldmath$ \sigma$}}_K)-{\mbox{\boldmath$ \Phi$}}^-_{\e,K}({\mbox{\boldmath$ \sigma$}}_K))+\ldots\cr}
\Eq{eB.4}\ee
where $K=I\times [a,b]$ is a {\it parallelepiped} in $\L\times[0,\t]$,
and ${\mbox{\boldmath$ \Phi$}}^z_{\e,K}{\buildrel def\over=}\sum_{t\in[a,b]}{\mbox{\boldmath$ \Phi$}}^z_{I+t}$ for $z=\pm,\f$,
and the $\ldots$
indicate a correction $\sum_K {\mbox{\boldmath$ \Phi$}}^\f_{\e,K}({\mbox{\boldmath$ \sigma$}}_K)$,\footnote{\small
Relatively vanishing as $L^{-1}$ uniformly in $x$.}.
t
A natural mathematical definition of the ``local average phase space
contraction'' could be the $-\frac1{\t\L_0}\sum_{t=0}^\t \log J_{\L_0}(S^x)$
where $J_{\L_0}(x)=|\det ({\partial}_i (S_\e x)_{i'})|$. But this is a quantity
difficult to express in a useful way.
However it is also possible to propose a {\it different} definition of
local average phase space contraction based on the representation
Eq.\equ{eB.4} of the average of the logarithms of the full Jacobian. The
latter can be expressed as Eq.\equ{eB.4} up to a quantity uniformly bounded
in $L$: and the contribution to Eq.\equ{eB.4} from the parallelepipeds
$K$'s {\it entirely contained} in $\L_0\times[0,\t]$ is:
\begin{equation}}\def\ee{\end{equation} \h_{\L_0,\e,\t}^{loc}
{\buildrel def\over=}\frac1{\t L_0^d}
\sum_{K\subset\L_0\times[0,\t]}
\Big({\mbox{\boldmath$ \Phi$}}_{+,K}({\mbox{\boldmath$ \sigma$}})-
{\mbox{\boldmath$ \Phi$}}_{-,K}({\mbox{\boldmath$ \sigma$}})\Big)
\Eq{eB.5}\ee
see Eq.\equ{e15.1},\equ{e15.4}; the $ \h_{\L_0,\e,\t}^{loc}$ can be,
heuristically, called the ``local contraction rate''. It can be uniformly
bounded (in $\t,L$).
Given $\L_0$ let $\h^{loc}_{\L_0,\e,+}$ be the time average
$\h_{\L_0,\e,\t}^{loc}$ define:
\begin{equation}}\def\ee{\end{equation} p'=\frac1\t
\frac{\h_{\L,\e,\t}(x)}{\h_{\L,\e,+}},\quad p=\frac1\t
\frac{\h_{\L_0,\e,\t}^{loc}(x)}{\h_{\L_0,\e,+}^{loc}}\Eq{eB.6}\ee
and remark that $\h_{\L_0,\e,+}^{loc}=\h_{\L,\e,+}+o(L_0^{-1})$
(because of the SRB distribution representation of a Gibbs process).
It can also be shown that {\it to leading order} as $L_0,L,\t\to\infty$
the large deviation rates for $p',p$ in Eq.\equ{eB.6}
have the form $\t L^d\z_\infty(p'),\ \t L_0^d\z^0_\infty(p)$, with
$\z_\infty=\z^0_\infty$ because
$\z_\infty$ is obtained as a thermodynamic limit of a kind of partition
function: for a proof
see \cite[(5.14)]{Ga999b}.
Therefore by the FT applied to $S_\e$ it is
$L^d\z_\infty(p')-L^d\z_\infty(-p')= L^d p' \h_{\L,\e,+}$ and, since
$\h_{\L_0,\e,+}^{loc}= \h_{\L,\e,+}+o(L_0^d)$, the large deviations rate
for $p$ in Eq.\equ{eB.6} satisfies a FR of the form
\begin{equation}}\def\ee{\end{equation} \eqalign{
&L^d_0(\z_\infty(p)-\z_\infty(-p))=p \,L_0^d\h_{\L_0,\e,+}^{loc}\cr&=
p\, r\, (L^d\h_{\L,\e,+})\cr}\Eq{eB.7}\ee
with $r=\frac{L_0^d}{L^d}$ and $|p|\le p^*, p^*\ge1$, up to corrections of
$O(L_0^{d-1})$: which means that the global and local large fluctuations
rates are proportional and trivially related by a rescaling which equals
$r=(\frac{L_0}{L})^d$ up to a correction bounded $\k^{-1} L_0^{d-1}$ with
$\k$ bounding the range of the SRB potential, as in Eq.\equ{e15.1}.
The universal slope $1$ in the global FR is modified into
$r=(\frac{L_0}{L})^d$ in the local FR. The Eq.\equ{eB.7} {\it can be proved
for the system in} Eq.\equ{eB.2}.
However $p$ in Eq.\equ{eB.6} {\it is not related to a measurable quantity},
as it cannot be hoped to be able to measure directly the
local phase space contraction .defined as in Eq.\equ{eB.5}.
Still the phase space contraction is often related to the amount of heat
ceded or the work done on the surroundings by a system in a stationary
state, as exemplified in the case of Eq.\equ{e4.4}: hence it is tempting to
test, in cases in which the latter quantities are accessible to local
measurements, whether Eq.\equ{eB.7} holds. This is attempted in some
simulations, \cite{GL014}.
The interest of the above special example lies in the statements
independence on the total size of the systems: they also mean that the
fluctuation theorems {\it may lead to observable consequences} if one looks
at the far more probable microscopic f\/luctuations of the local entropy
production rate,\cite{PS991,BK995,BK997}. For more details see
\cite{Ga995b}.
\APPENDICE{3}
\defArnold-Euler geodesics{Reversible heating}
\section{Arnold-Euler geodesics}
\label{appC}
Imagine a rarefied gas enclosed in a cubic container of side $L$ described
by a canonical distribution at inverse temperature $\b^{-1}$. The potential
energy is $\sum_{i=1}^N mg z_i +\sum_{i,j} v(x_i-x_j){\buildrel def\over=} M\,g\,H +V$,
with $M$=total mass and $H$ the height of the center of mass. The
initial free energy if $F(\b,g)=-\b^{-1} \log \int e^{-\b(V+gP)}
d^{3N}pd^{3N}q$. The entropy can be computed via Gibbs' formula $S_0=-\int
\r(p,q)\log\r(p,q)d^{3N}pd^{3N}q$.
The gas is set out of equilibrium by changing the gravity $g$ to a new value
$g'$ for instance suddenly at time $t=0$ or following a given prescription
$t\to g(t), t\in [0,\t]$ with $g(\t)=g', \t<\infty$. Then it is let to
evolve.
Since the evolution is Hamiltonian (although not autonomous) $\r(p,q)$
evolves in $\r(p,q;t)$ and the latter tends, as $t\to\infty$, to a new
equilibrium state in the gravity potential $m g'z$; but $-\int
\r(p,q;t)\log\r(p,q;t)d^{3N}pd^{3N}q$ remains equal to $S_0$. Therefore at
the end of the evolution the new distribution $\r(p,q,\infty)$ will be an
equilibrium state of the system in the modified gravity field.
It will not be, however, any more a canonical Gibbs state at temperature
$\b^{-1}$ in a gravity field with acceleration $g'$; if the system is
ergodic on the energy surface then the final distribution reached at
infinite time after suddenly increasing the gravity $g$ to a new value $g'$
will be (integration over $p',q'$ only)''
\begin{equation}}\def\ee{\end{equation} \m^\infty(dpdq)=\int \m_\b(dp'dq') \m^{mc}_{E(p',q')}(dpdq)\Eq{eC.1}\ee
where $\m^{mc}_{E}(dpdq)$ is the microcanonical distribution with energy
$E$ and $E'(p',q')= K(p')+V(q')+Mg'H(q')$ is the sum of the kinetic energy,
internal potential energy and energy of the center of mass in a gravity
acceleration $g'$.
The distribution Eq.\equ{eC.1} will be equivalent to a canonical Gibbs
distribution (with temperature different from $\b^{-1}$) only in the
thermodynamic limit: in the finite system that we are considering it will
be different by corrections vanishing in the thermodynamic limit. Yet the
new state will be a stationary state close to a canonical (or any other
equivalent) equilibrium state,
To estimate, actually to define, the temperature $\b^{'-1}$ of the new
state imagine to identify the above $\m^\infty$ with a canonical
distribution $\m_{\b'}$, {\it i.e.\ } {\it neglect the finite volume
corrections}. Computing the Gibbs entropy $S_\infty$ of the new equilibrium
(reached after infinite time) and make use of the identity between the
Gibbs entropies of the initial and final states:
\begin{equation}}\def\ee{\end{equation}\kern-3mm\eqalign{
&F=-\frac1\b\log \kern-1mm\int e^{-\b(V+gP)} d^{3N}pd^{3N}q=-T \log Z_0\cr
&S=-{\partial}_T F=-\log Z_0-\b\media{V+ gP}\cr
&{\partial}_g S|_\b=\b^{2}(\media{P V}-\media{P}\media{V})
+g(\media{P^2}-\media{P}^2)\cr}\Eq{eC.2}\ee
with $T=\b^{-1}$, $P=MgH$ and ${\partial}_g S|_\b\ne0$ at $g=0$: thus the new
equilibrium {\it cannot} have the same entropy as the initial state if the
temperature remained the same unless $\b'\ne\b$ (because in general ${\partial}_g
S|_\b\ne0$, {\it e.g.\ } if $V\simeq0$ it is $\media{P^2}> \media{P}^2$). If the final
state has to become a canonical distribution at some temperature ({\it e.g.\ } the
above estimated $\b^{'-1}$) then the system will have to be attached to a
thermostat and some heat exchange will take place and the entire
transformation will be irreversible: in any event, if the system container
was really adiabatic and at any finite (or infinite) time the gravity
acceleration was dropped back to the initial value, then the system should
in the same time return to the initial canonical state. See also
\cite{SSHT002} for the analysis of equally interesting cases. \*
\APPENDICE{4}
\defArnold-Euler geodesics{Arnold-Euler geodesics}
\section{Arnold-Euler geodesics}
\label{appD}
The analysis of \cite{Dr988} applies to the geodesic flows: which are
Hamiltonian flows with Hamiltonian $H(\V p,\V q)= \frac12 g(\V q)^{-1}{\V
p}{\V p}$. Hence the Lyapunov exponents of $\dot {\V q}=g(\V q) \V p,
\dot{\V p}=-\frac12({\mbox{\boldmath$ \partial$}}_{\V q}g(\V q)^{-1}){\V p}\,{\V p}-\n\V p+\ff$
with $\V f$ independent of $\V p,\V q$ are paired to $-\frac12\n$, see
footnote\footref{OR}, with a general $H(\V p,\V q)$.
The Euler flow is a geodesic flow, \cite{Ar966b,TT010}, and the canonical
coordinates are $\V u,{\mbox{\boldmath$ \delta$}}$ where ${\mbox{\boldmath$ \delta$}}$ is the diffeomorphism bringing the
reference state $O$ of the fluid into the actual state $A$: if a force $\V
f$ is added which does not change the Hamiltonian nature of the motion ({\it e.g.\ }
if, as in \cite{Dr988}, $\V f(\V q)$ is locally conservative) and if,
furthermore, a viscosity force of the form $-\n\V u$ is also added, then
the above shows that pairing takes place (formally) to $-\frac\n2$.
A formal proof that the Euler equations for a fluid can be written as
Arnold's geodesic flow is summarized as follows (correcting also typos in
the somewhat obscure argument in the appendix theof \cite{Ga997b}). Let
${\mbox{\boldmath$ \xi$}}\to \xx={\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})$ be the diffeomorphism in $Sdiff(T)\subset diff(T)$
\footnote{\small Notation: $diff(T)$ denotes the set of
the general torus diffeomorphisms and $Sdiff(T)$ is the subset
of diffeomorphisms with Jacobian
$J({\mbox{\boldmath$ \delta$}})({\mbox{\boldmath$ \xi$}})\equiv \det(\frac{{\mbox{\boldmath$ \partial$}} {\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})}{\d{\mbox{\boldmath$ \partial$}}{\mbox{\boldmath$ \xi$}}})=1$ {\it e.g.\ }, if $d=3$,
$J({\mbox{\boldmath$ \delta$}})({\mbox{\boldmath$ \xi$}})={\mbox{\boldmath$ \partial$}}\d_1({\mbox{\boldmath$ \xi$}})\wedge{\mbox{\boldmath$ \partial$}}\d_2({\mbox{\boldmath$ \xi$}})\cdot{\mbox{\boldmath$ \partial$}}\d_3({\mbox{\boldmath$ \xi$}})\equiv1$.}
mapping the reference fluid configuration $O$ into the actual one $A$.
Here and below the differential operators are denoted ${\partial},\ldots$ but are
intended to be ${\partial}_{{\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})},\ldots$ when operating on functions of
${\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})$.
A map from $O$ to a configuration $A'\in diff(T)$ {\it infinitesimally
close} to one in $Sdiff(T)$ can be parameterized by ${\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$\zeta$}}$ with
${\mbox{\boldmath$ \delta$}}\in Sdiff(T)$ and ${\mbox{\boldmath$\zeta$}}$ an infinitesimal variation of the form ${\mbox{\boldmath$ \partial$}}
z({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))$ for some scalar $z$.
The coordinates ${\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$\zeta$}}$ can be checked to form a system of coordinates
``{\it orthogonal and well adapted}'', in the sense of definition 12 in
\cite{Ga983}, to the surface $J({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))=1$ in the space $diff(T)$ of
torus diffeomorphisms in the metric $g({\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$\zeta$}})$ attributing to an
infinitesimal variation $(\V w+\V z)$ of $({\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$\zeta$}})$ the square length:
\begin{equation}}\def\ee{\end{equation} \int (\V w({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))^2+{\mbox{\boldmath$ \partial$}} z({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))^2)d{\mbox{\boldmath$ \xi$}}\Eq{eD.1}\ee
for the {\it infinitesimal} variations $(\V w,{\mbox{\boldmath$ \partial$}} z)$ of $({\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$\zeta$}})$ with
$\V w$ a divergence free field and ${\mbox{\boldmath$ \partial$}} z$ a gradient field. Furthermore if
${\mbox{\boldmath$ \xi$}}\to{\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})$ is an incompressible configuration in $Sdiff(T)$ and
${\mbox{\boldmath$ \delta$}}'({\mbox{\boldmath$ \xi$}})={\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})+\V w({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))+{\mbox{\boldmath$ \partial$}} z({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))$ an {\it infinitesimally
close} one in $diff(T)$ then
\begin{equation}}\def\ee{\end{equation}\kern-3mm
\eqalign{&J({\mbox{\boldmath$ \delta$}}'({\mbox{\boldmath$ \xi$}}))-1=\D z({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})),\
\hbox{(``Liouville's theorem'')}\cr
&\int d{\mbox{\boldmath$ \xi$}}
(J({\mbox{\boldmath$ \delta$}}'({\mbox{\boldmath$ \xi$}}))-1)^2=\int d{\mbox{\boldmath$ \xi$}} (\D z({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})))^2\cr
&\kern9mm=\int d\xx (\D
z(\xx))^2\cr}
\Eq{eD.2}\ee
depending only on the ``violation of the constraint'' $J-1$ in the
above orthogonal and well adapted coordinates.
Therefore a general theorem on constrained motions, (for the constraint
$J({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))=1$) can be applied: because the {\it perfection criteria} for
the constraint are met as a consequence of Eq.\equ{eD.1}\,,\,\equ{eD.2},
see definition 13, Sec. 3.7, and proposition 13, Sec. 3.8 in
\cite{Ga983}. Hence the motions obey Euler-Lagrange equations for the
Lagrangian $\LL(\dot{\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$ \delta$}})=\int \frac12\dot{{\mbox{\boldmath$ \delta$}}}({\mbox{\boldmath$ \xi$}})^2 d{\mbox{\boldmath$ \xi$}}$ with the
{\it ideal holonomic constraint} $J({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))\equiv1$.
In other words the motions driven by the Lagrangian
$\LL_\L(\dot{\mbox{\boldmath$ \delta$}},{\mbox{\boldmath$ \delta$}})=\int d{\mbox{\boldmath$ \xi$}}\Big(\frac12\dot{{\mbox{\boldmath$ \delta$}}}({\mbox{\boldmath$ \xi$}})^2 + \L
(J({\mbox{\boldmath$ \delta$}}(\x))-1)^2\Big)$, in the space $diff(T)\supset Sdiff(T)$, are motions
which depend on the auxiliary parameter $\L$ and, if the initial data are
an incompressible configuration ${\mbox{\boldmath$ \delta$}}\in Sdiff(T)$, in the limit as
$\L\to+\infty$ converge, at any prefixed time, to motions driven by the
constrained Lagrangian $\LL$.
Explicitly, let ${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(\xx)=\dot {\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})$: since at time $\e>0$
${\mbox{\boldmath$ \delta$}},{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}$ become ${\mbox{\boldmath$ \delta$}}_\e({\mbox{\boldmath$ \xi$}})$ and ${\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}_\e(\xx)={\mbox{\boldmath$ \delta$}}_\e(\xx+\e {\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(\xx))$, it
follows that $\ddot{\T {\mbox{\boldmath$ \delta$}}}({\mbox{\boldmath$ \xi$}})={\partial}_t{\T {\bf u}}}\def\Ff{{\T{\bf F}}(\xx)+({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(\xx)\cdot{\mbox{\boldmath$ \partial$}})
{\T {\bf u}}}\def\Ff{{\T{\bf F}}(\xx)$ and the Euler-Lagrange equations for the constrained $\LL$
become the Euler equations for a perfect incompressible fluid,
obtained as a geodesic flow for the Hamiltonian $H({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f},{\mbox{\boldmath$ \delta$}})$ defined as
\begin{equation}}\def\ee{\end{equation} \int d{\mbox{\boldmath$ \xi$}} \big(\frac12
{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))^2+(\D^{-1}(({\mbox{\boldmath$ \partial$}}{\T {\bf u}}}\def\Ff{{\T{\bf F}})\cdot(\T{\partial} {\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})))({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))\big)
\Eq{eD.3}
\ee
with $p(\xx)=-\D^{-1}((\T{\mbox{\boldmath$ \partial$}}{\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f})\cdot({\mbox{\boldmath$ \partial$}}{\T {\bf u}}}\def\Ff{{\T{\bf F}}))(\xx)$ interpreted as
``pressure''.
If a force $\V f(\xx)=(-{\partial}_2\F(\xx),{\partial}_1\F(\xx))$, with $\F$ a
given scalar independent of the fluid configuration, acts on the fluid
then to Eq.\equ{eD.3} a term ${\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})\cdot
(-{\partial}_2\F({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})),{\partial}_1\F({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})))$ has to be added.
Recalling that $\xx={\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})$, the equations of motion are the pair:
\begin{equation}}\def\ee{\end{equation}\kern-4mm\eqalign{ &{\partial}_t {\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}})={\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}({\mbox{\boldmath$ \delta$}}({\mbox{\boldmath$ \xi$}}))\cr
&{\partial}_t{\T {\bf u}}}\def\Ff{{\T{\bf F}}(\xx)\kern-1mm+\kern-1mm({\V u}}\def\kk{{\V k}}\def\xx{{\V x}}\def\ff{{\V f}(\xx)\cdot{\mbox{\boldmath$ \partial$}})
{\T {\bf u}}}\def\Ff{{\T{\bf F}}(\xx)=-\n {\T {\bf u}}}\def\Ff{{\T{\bf F}}(\xx)-\T{\mbox{\boldmath$ \partial$}} p(\xx)+\T
\ff(\xx)\cr}\kern-2mm\Eq{eD.4}\ee
and by the just mentioned result in \cite{Dr988}, display a set of
Lyapunov exponents, local and global, paired to $-\n/2$.
In the Euler-Arnold case the exponents are naturally divided in two
classes: the exponents relative to the $\V u$ coordinates, 'fluid
exponents', which do not depend (explicitly) from the evolution of the
${\mbox{\boldmath$ \delta$}}$ coordinates and the others.
It would be natural to think that the
other exponents are simply a copy of the first: if so the fluid exponents,
alone, would be paired. However if true this cannot be a general property
of geodesic flows to which friction and forcing are added (as it fails in
the flow with Hamiltonian $H=\frac12 p^2$, {\it i.e.\ } for the equations $\dot q=p,
\dot p=-\n p+f$). I neither succeded in proving the just mentioned
double degeneracy of the Lyapunov spectrum for the Arnold-Euler equations
nor in convincing myself that it is a reasonable hypothesis.
\* The results just discussed {\it do not apply to the NS flow} because the
viscosity is $\n \D\V u$: however a pairing to a line which is not constant
but which depends on the scale in which the motion is studied, compatible
with the above simulations, was proposed in \cite{Ga997b}. \*
\noindent{\bf Acknowledgments:} I profited from discussions with L.Biferale,
M.Cencini, M.De Pietro, V.Lucarini in Sec.\ref{sec18}\,, and L.S.Young
in Sec.\ref{appB}\,, whom I thank warmly.
\vfill\eject
\bibliographystyle{unsrt}
|
\section{Introduction}\label{sec:introduction}
Over the past few years, the effectiveness of scene understanding for self-driving cars has substantially increased both for object detection and vehicle navigation \cite{Vaquero_2017_ECMR,Chen_2017_CVPR}.
Even though these improvements allowed for more advanced and sophisticated \ac{ADAS} and maneuvers, the current state of the art is far from the SAE full-automation level, especially in complex scenarios such as urban areas.
Most of these algorithms depend on very accurate localization estimates, which are often hard to obtain using common \acp{GNSS}, mainly for \ac{NLOS} and multipath issues.
Moreover, applications that require navigation in indoor areas, \emph{e.g., } valet parking in underground areas, necessarily require complementary approaches.
Different options have been investigated to solve the localization problem, including approaches based on both vision and \ac{LIDAR}; they share the exploitation of an a-priori knowledge of the environment in the localization process \cite{7759304,6942558,8206067}.
Localization approaches that utilize the same sensor for mapping and localization usually achieve good performances, as the map of the scene is matched to the same kind of data generated by the on-board sensor.
However, their application is hampered by the need for a preliminary mapping of the working area, which represents a relevant issue in terms of effort both for building the maps as well as for their maintenance.
On the one hand, some approaches try to perform the localization exploiting standard cartographic maps, such as OpenStreetMap or other topological maps, leveraging the road graph \cite{ParraAlonso2012} or high-level features such as lane, roundabouts, and intersections \cite{Flade2016, Raaijmakers2015, Ballardini2019}.
On the other hand, companies in the established market of maps and related services, like \emph{e.g., } HERE or TomTom, are nowadays already developing so-called \acp{HD map}, which are built using \ac{LIDAR} sensors \cite{heremaps}. This allows other players in the autonomous cars domain, to focus on the localization task.
\acp{HD map}, which are specifically designed to support self-driving vehicles, provide an accurate position of high-level features such as traffic signs, lane markings, etc. as well as a representation of the environment in terms of point clouds, with a density of points usually reaching 0.1m.
In the following, we denote as \ac{LIDAR}-maps the point clouds generated by processing data from \acp{LIDAR}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=.48\textwidth]{images/ITSC2019-Teaser_05.png}
\end{center}
\caption{A sketch of the proposed processing pipeline.
Starting from a rough camera pose estimate (\emph{e.g., } from a GNSS device), CMRNet compares an RGB image and a synthesized depth image projected from a LiDAR-map into a virtual image plane (red) to regress the 6DoF camera pose (in green).
Image best viewed in color.
}
\label{fig:figure_1_teaser}
\end{figure}
Standard approaches to exploit such maps localize the observer by matching point clouds gathered by the on-board sensor to the \ac{LIDAR}-map; solutions to this problem are known as point clouds registration algorithms.
Currently, these approaches are hampered by the huge cost of \ac{LIDAR} devices, the de-facto standard for accurate geometric reconstruction.
In contrast, we here propose a novel method for registering an image from an on-board monocular RGB camera to a \ac{LIDAR}-map of the area. This allows for the exploitation of the forthcoming market of \ac{LIDAR}-maps embedded into \acp{HD map} using only a cheap camera-based sensor suite on the vehicle.
In particular, we propose CMRNet, a \ac{CNN}-based approach that achieves camera localization with sub-meter accuracy, basing on a rough initial pose estimate.
The maps and images used for localization are not necessarily those used during the training of the network.
To the best of our knowledge, this is the first work to tackle the localization problem without a \emph{localized} \ac{CNN}, \emph{i.e., } a \ac{CNN} trained in the working area \cite{Kendall_2015_ICCV}.
CMRNet does not learn the map, instead, it learns to match images to the \ac{LIDAR}-map.
Extensive experimental evaluations performed on the KITTI datasets \cite{Geiger2013IJRR} show the feasibility of our approach.
The remainder of the paper is organized as follows: \Cref{sec:related-work} gives a short review of the most similar methods and the last achievements with DNN-based approaches.
In \Cref{sec:proposed-approach} we present the details of the proposed system.
In \Cref{sec:experimental} we show the effectiveness of the proposed approach, and \Cref{sec:conclusions,sec:future-works} present our conclusions and future work.
\section{Related work}\label{sec:related-work}
In the last years, visual localization has been a trending topic in the computer vision community.
Although many variations have been proposed, most of them are either based on images gathered from a camera sensor only or exploit some kind of 3-dimensional reconstruction of the environment.
\subsection{Camera-only approaches}
The first category of techniques deals with the 6-DoF estimate of the camera pose using a single image as input.
On the one hand, traditional methods face this problem by means of a two-phase procedure that consists of a coarse localization, performed using a place recognition algorithm, followed by a second refining step that allows for a final accurate localization \cite{Zamir_2010_ECCV, Sattler_2012_ECCV}.
On the other hand, the latest machine learning techniques, mainly based on deep learning approaches, face this task in a single step.
These models are usually trained using a set of images taken from different points of view of the working environment, in which the system performs the localization.
One of the most important approaches of this category, which inspired many subsequent works, is PoseNet \cite{Kendall_2015_ICCV}.
It consists in a \ac{CNN} trained for camera pose regression.
Starting from this work, additional improvements have been proposed by introducing new geometric loss functions \cite{Kendall_2017_CVPR}, by exploiting the uncertainty estimation of Bayesian \acp{CNN} \cite{kendall2016modelling}, by including a data augmentation scheme based on synthetic depth information \cite{Naseer_2017_IROS}, or using the relative pose between two observations in a \acp{CNN} pipeline \cite{Brahmbhatt_2018_CVPR}.
One of the many works that follow the idea presented in PoseNet is VLocNet++ \cite{8458420}.
Here the authors deal with the visual localization problem using a multi-learning task (MLT) approach.
Specifically, they proved that training a \ac{CNN} for different tasks at the same time yields better localization performances than single task learning.
As for today, the literature still sees \cite{8458420} as the best performing approach on the 7Scenes dataset \cite{Shotton_2013_CVPR}.
Clark \emph{et al.} \cite{Clark_2017_CVPR} developed a CNN that exploits a sequence of images in order to improve the quality of the localization in urban environments.
Brachmann \emph{et al.}{}, instead, integrated a differentiable version of RANSAC within a CNN-based approach in an end-to-end fashion \cite{Brachmann_2017_CVPR, Brachmann_2018_CVPR}.
Another camera-only localization is based on \textit{decision forests}, which consists of a set of decision trees used for classification or regression problems.
For instance, the approach proposed by Shotton \emph{et al.}{} \cite{Shotton_2013_CVPR} exploits RGBD images and regression forests to perform indoor camera localization.
The aforementioned techniques, thanks to the generalization capabilities of machine learning approaches, are more robust against challenging scene conditions like lighting variations, occlusions, and repetitive patterns, in comparison with methods based on hand-crafted descriptors, such as SIFT \cite{Lowe2004}, or SURF \cite{BAY2008346}.
However, all these methods cannot perform localization in environments that have not been exploited in the training phase, therefore these regression models need to be retrained for every new place.
\subsection{Camera and \ac{LIDAR}-map approaches}
The second category of localization techniques leverages existing maps, in order to solve the localization problem.
In particular, two classes of approaches have been presented in the literature: geometry-based and projection-based methods.
Caselitz \emph{et al.} \cite{7759304} proposed a geometry-based method that solves the visual localization problem by comparing a set of 3D points, the point cloud reconstructed from a sequence of images and the existing map.
Wolcott \emph{et al.} \cite{6942558}, instead, developed a projection-based method that uses meshes built from intensity data associated to the 3D points of the maps, projected into an image plane, to perform a comparison with the camera image using the \textit{Normalized Mutual Information} (NMI) measure.
Neubert \emph{et al.} \cite{8206067} proposed to use the similarity between depth images generated by synthetic views and the camera image as a score function for a particle filter, in order to localize the camera in indoor scenes.
The main advantage of these techniques is that they can be used in any environment for which a 3D map is available.
In this way, they avoid one of the major drawbacks of machine learning approaches for localization, \emph{i.e., } the necessity to train a new model for every specific environment.
Despite these remarkable properties, their localization capabilities are still not robust enough in the presence of occlusions, lighting variations, and repetitive scene structures.
The work presented in this paper has been inspired by Schneider \emph{et al.} \cite{Schneider_2017}, which used 3D scans from a \ac{LIDAR} and RGB images as the input of a novel CNN, RegNet.
Their goal was to provide a CNN-based method for calibrating the extrinsic parameters of a camera \emph{w.r.t.}~a~\ac{LIDAR} sensor.
Taking inspiration from that work, in this paper we propose a novel approach that has the advantages of both the categories described above.
Differently from the aforementioned literature contribution, which exploits the data gathered from a synchronized single activation of a 3D \ac{LIDAR} and a camera image, the inputs of our approach are a complete 3D \ac{LIDAR} map of the environment, together with a single image and a rough initial guess of the camera pose.
Eventually, the output consists of an accurate 6-DoF camera pose localization.
It is worth to notice that having a single \ac{LIDAR} scan taken at the same time as the image imply that the observed scene is exactly the same.
In our case, instead, the 3D map usually depicts a different configuration, \emph{i.e., } road users are not present, making the matching more challenging.
Our approach combines the generalization capabilities of \acp{CNN}, with the ability to be used in any environment for which a \ac{LIDAR}-map is available, without the need to re-train the network.
\section{Proposed Approach}\label{sec:proposed-approach}
In this work, we aim at localizing a camera from a single image in a 3D \ac{LIDAR}-map of an urban environment.
We exploit recent developments in deep neural networks for both pose regression \cite{Kendall_2015_ICCV} and feature matching \cite{Sun_2018_CVPR}.
The pipeline of our approach is depicted in Fig.~\ref{fig:figure_1_teaser} and can be summarized as follows.
First, we generate a synthesized depth image by projecting the map points into a virtual image plane, positioned at the initial guess of the camera pose.
This is done using the intrinsic parameters of the camera.
From now on, we will refer to this synthesized depth image as \ac{LIDAR}-image.
The \ac{LIDAR}-image, together with the RGB image from the camera, are fed into the proposed CMRNet, which regresses the rigid body transformation $H_{out}$ between the two different points of view.
From a technical perspective, applying $H_{out}$ to the initial pose $H_{init}$ allows us to obtain the 6-DoF camera localization.
In order to represent a rigid body transformation, we use a $(4,4)$ homogeneous matrix:
\begin{equation}
\bf{H} =
\begin{pmatrix}
\bf{R}_{(3,3)} & \bf{T}_{(3,1)} \\
\bf{0}_{(1,3)} & \bf{1}
\end{pmatrix}
\in SE(3)
\end{equation}
Here, $\bf{R}$ is a $(3,3)$ rotation matrix and \textbf{T} is a $(3, 1)$ translation vector, in cartesian coordinates.
The rotation matrix is composed of nine elements, but, as it represents a rotation in the space, it only has three degrees of freedom.
For this reason, the output of the network in terms of rotations is expressed using quaternions lying on the 3-sphere ($S^3$) manifold.
On the one hand, even though normalized quaternions have one redundant parameter, they have better properties than Euler angles, \emph{i.e., } gimbal lock avoidance and unique rotational representation (except that conjugate quaternions represent the same rotation).
Moreover, they are composed of fewer elements than a rotation matrix, thus being better suited for machine learning regression approaches.
The outputs of the network are then a translation vector $\textbf{T} \in \mathbb{R}^3$ and a rotation quaternion $\textbf{q} \in S^3$.
For simplicity, we will refer to the output of the network as $H_{out}$, implying that we convert $\textbf{T}$ and $\textbf{q}$ to the corresponding homogeneous transformation matrix, as necessary.
\subsection{LiDAR-Image Generation}\label{sec:lidargeneration}
In order to generate the LiDAR-image for a given initial pose $H_{init}$, we follow a two-step procedure.
\noindent
\textit{Map Projection.}
First, we project all the 3D points in the map into a virtual image plane placed at $H_{init}$, \emph{i.e., } compute the image coordinates $p$ of every 3D point $P$.
This mapping is shown in \Cref{eq:proj}, where $K$ is the camera projection matrix.
\begin{equation}
\label{eq:proj}
p^i = K \cdot H_{init} \cdot P^i
\end{equation}
The LiDAR-image is then computed using a z-buffer approach to determine the visibility of points along the same projection line.
Since \Cref{eq:proj} can be computationally expensive for large maps, we perform the projection only for a sub-region cropped around $H_{init}$, ignoring also points that lay behind the virtual image plane.
In \Cref{fig:occlusion_filtering_pre} is depicted an example of LiDAR-image.
\begin{figure}[htb]
\centering
\subfloat[Without Occlusion Filter]{%
\label{fig:occlusion_filtering_pre}\includegraphics[width=\linewidth]{images/pre_post_occlusion_2_top.png}}\hfill
\subfloat[With Occlusion Filter]{%
\label{fig:occlusion_filtering_post}\includegraphics[width=\linewidth]{images/pre_post_occlusion_2_bottom.png}}\hfill
\caption{Top: a LiDAR-image with the associated RGB overlay.
Please note how the points behind the building on the right, \emph{i.e., } lighter points on the fence, are projected into the LiDAR-image.
Bottom: an example of the occlusion filtering effect.
Color codes distance from close (blue) to far point (red). }\label{fig:occlusion_filtering}
\end{figure}
\noindent
\textit{Occlusion Filtering.}
The projection of a point cloud into an image plane can produce unrealistic depth images.
For instance, the projection of occluded points, \emph{e.g., } laying behind a wall, is still possible due to the sparsity nature of point clouds.
To avoid this problem, we adopt the point clouds occlusion estimation filter presented in \cite{pintus2011real};
an example of the effect of this approach is depicted in \Cref{fig:occlusion_filtering_post}.
For every point $P_i$, we can build a cone, about the projection line towards the camera, that does not intersect any other point.
If the cone has an aperture larger than a certain threshold \textit{Th}, the point $P_i$ is marked as visible.
From a technical perspective, for each pixel with a non-zero depth $p_j$ in the LiDAR-image, we compute the normalized vector $\Vec{v}$ from the relative 3D point $P_j$ to the pin-hole.
Then, for any 3D point $P_i$ whose projection lays in a neighborhood (of size \textit{KxK}) of $p_j$, we compute the vector ${\Vec{c} = \frac{P_i-P_j}{\|P_i-P_j\|}}$ and the angle between the two vectors ${\vartheta = \arccos(\Vec{v} \cdot \Vec{c})}$.
This angle is used to assess the visibility of $P_j$.
Occluded pixels are then set to zero in the LiDAR-image.
More detail is available in \cite{pintus2011real}
\subsection{Network Architecture}
PWC-Net \cite{Sun_2018_CVPR} was used as baseline, and we then made some changes to its architecture.
We chose this network because PWC-Net has been designed to predict the optical flow between a pair of images, \emph{i.e., } to find matches between them.
Starting from a rough camera localization estimate, our insight is to exploit the correlation layer of PWC-Net and its ability to match features from different points of view to regress the correct 6-DoF camera pose.
We applied the following changes to the original architecture.
\begin{itemize}
\item First, as our inputs are a depth and an RGB image (instead of two RGB images), we decoupled the feature pyramid extractors by removing the weights sharing.
\item Then, as we aim to perform pose regression, we removed the up-sampling layers, attaching the fully connected layers just after the first cost volume layer.
\end{itemize}
Regarding the regression part, we added one fully connected layer with 512 neurons before the first optical flow estimation layer (conv6\_4 in PWC-Net), followed by two branches for handling rotations and translations.
Each branch is composed of two stacked fully connected layers, the first with 256 while the second with 3 or 4 neurons, for translation and rotation respectively.
Given an input pair composed of a RGB image $\mathcal{I}$ and a LiDAR-image $\mathcal{D}$, we used the following loss function in \Cref{eq:loss}, where $\mathcal{L}_t(\mathcal{I}, \mathcal{D})$ is the translation loss and $\mathcal{L}_q(\mathcal{I}, \mathcal{D})$ is the rotation loss.
\begin{equation}
\label{eq:loss}
\mathcal{L}(\mathcal{I}, \mathcal{D}) = \mathcal{L}_t(\mathcal{I}, \mathcal{D}) + \mathcal{L}_q(\mathcal{I}, \mathcal{D})
\end{equation}
For the translation we used a $\mbox{smooth}_{L1}$ loss \cite{girshick2015fast}.
Regarding the rotation loss, since the Euclidean distance does not provide a significant measure to describe the difference between two orientations, we used the angular distance between quaternions, as defined below:
\begin{eqnarray}\label{eq:lq}
\mathcal{L}_q(\mathcal{I}, \mathcal{D}) & = & D_a(q*inv(\Tilde{q})) \\
\label{eq:atan2}
D_a(m) &=& atan2(\sqrt{b^2_m + c^2_m + d^2_m}, |a_m|)
\end{eqnarray}
Here, $q$ is the ground truth rotation, $\Tilde{q}$ represents the predicted normalized rotation, $inv$ is the inverse operation for quaternions, $\{a_m, b_m, c_m, d_m\}$ are the components of the quaternion $m$ and $*$ is the multiplicative operation of two quaternions.
In order to use Equation~\eqref{eq:atan2} as a loss function, we need to ensure that it is differentiable for every possible output of the network.
Recalling that $atan2(y,x)$ is not differentiable for ${y=0 \land x\leq0}$, and the fact that $m$ is a unit quaternion, we can easily verify that Equation~\eqref{eq:atan2} is differentiable in $S^3$.
\subsection{Iterative refinement}
\label{seq:iterative}
When the initial pose strongly deviates with respect to the camera frame, the map projection produces a LiDAR-image that shares just a few correspondences with the camera image.
In this case, the camera pose prediction task is hard, because the \ac{CNN} lacks the required information to compare the two points of view.
It is therefore quite likely that the predicted camera pose is not accurate enough.
Taking inspiration from \cite{Schneider_2017}, we propose an iterative refinement approach.
In particular, we trained different CNNs by considering descending error ranges for both the translation and rotation components of the initial pose.
Once a LiDAR-image is obtained for a given camera pose, both the camera and the LiDAR-image are processed, starting from the CNN that has been trained with the largest error range.
Then, a new projection of the map points is performed, and the process is repeated using a CNN trained with a reduced error range.
Repeating this operation \textit{n} times is possible to improve the accuracy of the final localization.
The improvement is achieved thanks to the increasing overlap between the scene observed from the camera and the scene projected in the \textit{$n^{th}$} LiDAR-image.
\subsection{Training details}
We implemented CMRNet using the PyTorch library \cite{paszke2017automatic}, and a slightly modified version of the official PWC-Net implementation.
Regarding the activation function, we used a leaky RELU (REctified Linear Unit) with a negative slope of $0.1$ as non-linearity.
Finally, CMRNet was trained from scratch for 300 epochs using the ADAM optimizer with default parameters, a batch size of $24$ and a learning rate of $1e^{-4}$ on a single NVidia GTX 1080ti.
\section{Experimental results}\label{sec:experimental}
This section describes the evaluation procedure we adopted to validate CMRNet, including the used dataset, the assessed system components, the iterative refinements and finally the generalization capabilities.
We wish to emphasize that, in order to assess the performance of CMRNet itself, in all the performed experiments each input was processed independently, \emph{i.e., } without any tracking or temporal integration strategy.
\subsection{Dataset}
We tested the localization accuracy of our method on the KITTI odometry dataset.
Specifically, we used the sequences from 03 to 09 for training (11697 frames) and the sequence 00 for validating (4541 frames).
Note that the validation set is spatially separated from the train set, except for a very small sub-sequence (approx 200 frames), thus it is fair to say that the network is tested in scenes never seen during the training phase.
Since the accuracy of the provided GPS-RTK ground truth is not sufficient for our task (the resulting map is not aligned nearby loop closures), we used a LiDAR-based SLAM system to obtain consistent trajectories.
The resulting poses are used to generate a down-sampled map with a resolution of 0.1m.
This choice is the result of our expectations on the format of HD-maps that will be soon available from map providers \cite{heremaps}.
Since the images from the KITTI dataset have different sizes (varying from 1224x370 to 1242x376), we padded all images to 1280x384, in order to match the \ac{CNN} architecture requirement, \emph{i.e., } width and height multiple of 64.
Note that we first projected the map points into the LiDAR-image and then we padded both RGB and LiDAR-image, in order not to modify the camera projection parameters.
To simulate a noisy initial pose estimate $H_{init}$, we applied, independently for each input, a random translation, and rotation to the ground truth camera pose.
In particular, for each component, we added a uniformly distributed noise in the range of [-2m, +2m] for the translation and [$-10^{\circ}$, $+10^{\circ}$] for the rotation.
Finally, we applied the following data augmentation scheme: first, we randomly changed the image brightness, contrast and saturation (all in the range [0.9, 1.1]).
Then we randomly mirrored the image horizontally, and last we applied a random image rotation in the range [$-5^\circ$, $+5^\circ$] along the optical axis.
The 3D point cloud was transformed accordingly.
Both data augmentation and the selection of $H_{init}$ take place at run-time, leading to different LiDAR-images for the same RGB image across epochs.
\subsection{System Components Evaluation}
We evaluated the performances of CMRNet by assessing the localization accuracy, varying different sub-components of the overall system.
Among them, the most significative are shown in \Cref{tab:risultatigrossi}, and derive from the following operational workflow.
First, we evaluated the best CNN to be used as backbone, comparing the performances of state-of-the-art approaches, namely PWC-Net, ResNet18 and RegNet \cite{Sun_2018_CVPR, he2016deep, Schneider_2017}.
According to the performed experiments, PWC-Net maintained a remarkable superiority with respect to RegNet and ResNet18 and therefore was chosen as a starting point for further evaluation.
Thereafter, we estimated the effects in modifying both inputs, \emph{i.e., } camera images and \ac{LIDAR}-images.
In particular, we added a random image mirroring and experimented different parameter values influencing the effect of the occlusion filtering presented in \Cref{sec:lidargeneration}, \emph{i.e., } size \textit{K} and threshold \textit{Th}.
At last, the effectiveness of the rotation loss proposed in \Cref{eq:atan2} was evaluated with respect to the commonly used $L_1$ loss. The proposed loss function achieved a relative decrease of rotation error of approx. 35\%.
The noise added to the poses in the validation set was kept fixed on all the experiments, allowing for a fair comparison of the performances.
\begin{table}[]
\scriptsize
\centering
\begin{threeparttable}
\caption{Parameter Estimation}
\label{tab:risultatigrossi}
\begin{tabular*}{\linewidth}{ccccc|cc}
\toprule
& \multicolumn{2}{c}{Occlusion} & & &\multicolumn{2}{c}{\textbf{Error}} \\ \cmidrule(lr){2-3} \cmidrule(lr){6-7}
Backbone & \textit{K} & \textit{Th} & Mirroring & Rot. Loss & Transl. & Rot. \\ \midrule
Regnet & - & - & \ding{55} & $D_a$& 0.64m & 1.67$^\circ $ \\
ResNet18 & - & - & \ding{55} & $D_a$& 0.60m & 1.59$^\circ $ \\
PWC-Net & 11 & 3.9999 & \ding{55} & $D_a$& 0.52m & 1.50$^\circ $ \\
PWC-Net & 13 & 3.9999 & \ding{55} & $D_a$& 0.51m & 1.43$^\circ $ \\
PWC-Net & 5 & 3.0 & \ding{55} & $D_a$& 0.47m & 1.45$^\circ $ \\
PWC-Net & 5 & 3.0 & \ding{51} & $D_a$& \textbf{0.46m} & $\textbf{1.36}^\circ$ \\
PWC-Net & 5 & 3.0 & \ding{51} & $L_1$& 0.46m & 2.07$^\circ$ \\ \bottomrule
\end{tabular*}
\begin{tablenotes}[para,flushleft]
\footnotesize
\item Median localization accuracy varying different sub-components of the overall system. \textit{K} and \textit{Th} correspond to the occlusion filter parameters as described in \Cref{sec:lidargeneration}.
\end{tablenotes}
\end{threeparttable}
\end{table}
\subsection{Iterative Refinement and Overall Assessment}
In order to improve the localization accuracy of our system, we tested the iterative approach explained in \Cref{seq:iterative}.
In particular, we trained three instances of CMRNet varying the maximum error ranges of the initial camera poses.
To assess the robustness of CMRNet, we repeated the localization process for 10 times using different initial noises. The averaged results are shown in \Cref{tab:iterative} together with the correspondent ranges used for training each network.
Moreover, in order to compare the localization performances with the state-of-the-art monocular localization in \ac{LIDAR} maps \cite{7759304}, we calculated mean and standard deviation for both rotation and translation components over 10 runs on the sequence 00 of the KITTI odometry dataset.
Our approach shows comparable values for the translation component ($0.33\pm0.22$m \emph{w.r.t.}~$0.30\pm0.11$m), with a lower rotation errors ($1.07\pm0.77^\circ$ \emph{w.r.t.}~$1.65\pm0.91^\circ$).
Nevertheless, it is worth to note that our approach still does not take advantage of any pose tracking procedure nor multi-frame analysis.
Some qualitative examples of the localization capabilities of CMRNet with the aforementioned iteration scheme are depicted in \Cref{fig:composition}.
In \Cref{fig:pdf} we illustrate the probability density functions (PDF) of the error, decomposed into the six components of the pose, for the three iterations of the aforementioned refinement.
It can be noted that the PDF of even the first network iteration approximates a Gaussian distribution and following iterations further decrease the variance of the distributions.
An analysis of the runtime performances using this configuration is shown in \Cref{tab:runtime}.
\begin{figure*}
\begin{center}
\includegraphics[width=1.0\textwidth]{images/composition_2.png}
\end{center}
\caption{Four examples of the localization results.
From left to right: Input LiDAR-image, CMRNet result after the third iteration, ground truth.
All \ac{LIDAR}-images are overlayed with the respective RGB image for visualization purpose.}
\label{fig:composition}
\end{figure*}
\begin{table}[]
\centering
\begin{threeparttable}
\caption{Iterative Pose Refinement}
\label{tab:iterative}
\begin{tabular*}{\linewidth}{lcc|cc}
\toprule
& \multicolumn{2}{c}{Initial Error Range}& \multicolumn{2}{c}{Localization Error} \\ \cmidrule(lr){2-3} \cmidrule(lr){4-5}
& Transl. [m] & Rot. [deg] & Transl. [m] & Rot. [deg] \\ \midrule
Iteration 1 & [-2, +2]& [$-10$, $+10$] & 0.51 & $1.39$ \\
Iteration 2 & [-1, +1]& [$-2$, $+2$] & 0.31 & $1.09$ \\
Iteration 3 & [-0.6, +0.6]& [$-2$, $+2$]& \textbf{0.27} & $\textbf{1.07}$ \\ \bottomrule
\end{tabular*}
\begin{tablenotes}[para,flushleft]
\footnotesize
\item Median localization error at each step of the iterative refinement averaged over 10 runs.
\end{tablenotes}
\end{threeparttable}
\end{table}
\begin{table}[t!]
\centering
\begin{threeparttable}
\caption{Runtime Performances}
\label{tab:runtime}
\begin{tabular*}{\linewidth}{cccc|c}
\toprule
& Z-Buffer & Occlusion Filter & CMRNet & \textbf{Total} \\ \midrule
Time {[}ms{]} & 8.6 & 1.4 & 4.6 & \textbf{14.7} ($\sim$68Hz) \\ \bottomrule
\end{tabular*}
\begin{tablenotes}[para,flushleft]
\footnotesize
\item In the table, an analysis of the time performances of the system steps for a single execution, \emph{i.e., } 44.1ms for the 3-stages iterative refinement.
All the code was developed in CUDA, achieving 68fps runtime performances on the KITTI dataset.
CPU-GPU transfer time was not here considerated.
\end{tablenotes}
\end{threeparttable}
\end{table}
\begin{figure*}[htb]
\centering
\subfloat[Longitudinal Errors]{%
\includegraphics[width=.32\textwidth]{images/x_error.png}}\hfill
\subfloat[Lateral Errors]{%
\includegraphics[width=.32\textwidth]{images/y_errors.png}}\hfill
\subfloat[Vertical Errors]{%
\includegraphics[width=.32\textwidth]{images/z_errors.png}}\hfill
\subfloat[Roll Errors]{%
\includegraphics[width=.32\textwidth]{images/roll_errors.png}}\hfill
\subfloat[Pitch Errors]{%
\includegraphics[width=.32\textwidth]{images/pitch_errors.png}}\hfill
\subfloat[Yaw Errors]{%
\includegraphics[width=.32\textwidth]{images/yaw_errors.png}}\hfill
\caption{Iterative refinement error distributions: a PDF has been fitted (using Gaussian kernel density estimation) on the network error outcome for each iteration step and each component.
The dashed red lines are the theoretic PDFs of the initial $H_{init}$ errors.}\label{fig:pdf}
\end{figure*}
\subsection{Generalization Capabilities}
In order to assess the generalization effectiveness of our approach, we evaluated its localization performance using a 3D \ac{LIDAR}-map generated on a different day with respect to the camera images, yet still of the same environment.
This allows us to have a completely different arrangement of parked cars and therefore to stress the localization capabilities.
Unfortunately, there is only a short overlap between the sequences of the odometry dataset (approx. 200 frames), consisting of a small stretch of roads in common between sequences "00" and "07".
Even though we cannot completely rely on the results of this limited set of frames, CMRNet achieved 0.57m and 0.9$^\circ$ median localization accuracy on this test.
Indeed, it is worth to notice that the network was trained with maps representing the same exact scene of the respective images, \emph{i.e., } with cars parked in the same parking spots, and thus cannot learn to ignore cluttering scene elements.
\section{Conclusions} \label{sec:conclusions}
In this work we have described CMRNet, a \ac{CNN} based approach for camera to LiDAR-Map registration, using the KITTI dataset for both learning and validation purposes.
The performances of the proposed approach allow multiple specialized CMRNet to be stacked as to improve the final camera localization, yet preserving realtime requirements.
The results have shown that our proposal is able to localize the camera with a median of less than $0.27$m and $1.07^\circ$.
Preliminary and not reported experiments on other datasets suggests there is room for improvement and the reason seems to be due to the limited vertical field-of-view available for the point clouds
Since our method does not learn the map but learn how to perform the registration, it is suitable for being used with large-scale HD-Maps.
\section{Future Works} \label{sec:future-works}
Even though our approach does not embed any information of specific maps, a dependency on the intrinsic camera calibration parameters still holds.
As part of the future works we plan to increase the generalization capabilities so to not directly depend from a specific camera calibration.
Finally, since the error distributions reveal a similarity with respect to Gaussian distributions, we expect to be able to benefit from standard filtering techniques aimed to probabilistically tackle the uncertainties over time.
\section*{Acknowledgments}\label{sec:ack}
The authors would like to thank Tim Caselitz for his contribution related to the ground truth SLAM-based trajectories for the KITTI sequences and Pietro Colombo for the help in the editing of the associated video.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Statistical testing based on randomized equal allocation is a widespread state-of-the-art approach in the design of experiments for around $ 100 $ years,
known today as the \emph{randomized controlled trial} in biostatistics, the \emph{between-group design} in social sciences, and the \emph{A/B testing} in
digital marketing. Already \citet{Thompson1933biometrika}, a biostatistician from Yale University, proposed a data-driven approach which would in
expectation lead to a higher reward from an experiment, using the following words:
\begin{quote}
``...there can be no objection to the use of data, however meagre, as a guide to action required before more can be collected ... Indeed, the fact that
such objection can never be eliminated entirely---no matter how great the number of observations---suggested the possible value of seeking other modes of
operation than that of taking a large number of observations before analysis or any attempt to direct our course...''
\end{quote}
\citet{Robbins1952}, a prominent mathematician and statistician, emphasized that this problem is of a much wider importance:
\begin{quote}
``In fact, the problem represents in a simplified way the general question of how we learn---or should learn---from past experience.''
\end{quote}
A formulation of the problem using the Bayesian decision-theoretic framework allows for Bayes-optimality. Practical application of this Bayesian approach
has however been long hindered by its computational complexity, since the optimal solution is known in analytical form only for infinite horizon
\citep{Kelly1981}. A variety of practical approximations and heuristics have been developed and studied across several disciplines in order to overcome
this issue, but their analysis failed to give exact results or bounds sufficiently close to Bayes-optimality for finite horizon problems, which are the
problems most relevant to many situations in practice.
\subsection{Paper Structure and Contributions}
In this paper we thus focus on the finite-horizon setting. We also restrict the discussion to two arms, which often naturally appears per se or as a
subproblem in some multi-armed generalizations (e.g. if new arms appear over time), and serves as a starting point for introducing additional problem
features. The consideration of binary responses is motivated by its widespread applicability and by being one of the most studied settings.
Our main objective is to give an account of modelling and solution approaches arising in different disciplines, in a unified framework and using a unified
terminology. The problem description, origins and terminology is given in \autoref{section:problem}. Our unified model is described in
\autoref{section:model}, cast in the Markov decision process framework, with subject responses modelled using the Bernoulli distribution, and the
corresponding Beta distribution for Bayesian updating. Different problem settings, assumptions and objectives are summarized in \autoref{section:settings}.
\autoref{section:approaches} gives an account of the history and state of the art of approaches from several disciplines. In \autoref{section:performance}
we evaluate these designs, together with a few newly proposed, accurately computationally (using a newly written package in Julia programming language by
the author) in order to compare their performance, showing that conclusions are different for moderate horizons (typical in practice) than for small
horizons (typical in academic literature reporting computational results). We further list and clarify a number of myths about this problem in
\autoref{section:myths}, e.g., we show that, computationally, much larger problems can be designed to Bayes-optimality than what is commonly believed.
\autoref{section:conclusion} concludes.
\section{Problem}
\label{section:problem}
We consider the problem with two \emph{arms} (or, \emph{interventions}), called $ C $ (mnemonically for ``control'' or ``comparator'' or standard of
``care'') and $ D $ (for ``discovery'' or ``development'').\footnote{In the existing literature, it is common to denote the two arms as $ 1 , 2 $ (but we
prefer to keep these names for actions defined below) or $ A , B $ (but we prefer to keep $ A $ for the action process and $ B $ for the Beta function).} $
T $ \emph{subjects} become available one by one, and each subject must be allocated to exactly one of the arms. Upon allocation of a subject to arm $ C $
($ D $), subject's \emph{response} is observed, which is binary (success/failure), where the success probability is $ \theta_{ C } $ ($ \theta_{ D } $) and
the failure probability is $ 1 - \theta_{ C } $ ($ 1 - \theta_{ D } $). The primary objective is to find a \emph{design}, i.e. a strategy composed of
\emph{randomized actions} of allocating the subjects to arms, which, in expectation, achieves the highest number of observed successes from the $ T $
subjects, assuming that the success probabilities are unknown. A formal model is given in \autoref{section:model}.
\subsection{Problem Origins}
The first statement of the two-armed bandit problem is in \citet{Thompson1933biometrika}, extended in \citet{Thompson1935ajm} to multiple arms, in a
Bayesian setting. Apparently unaware of Thompson's works, \citet{Robbins1952} formulated the two-armed bandit problem in a frequentist setting. Neither
\citet{Thompson1933biometrika, Thompson1935ajm} nor \citet{Robbins1952} used the terms ``arm'' or ``bandit''. The term \emph{two-armed bandit} problem
first appeared in \citet{BradtEtal1956}, referring to the setting with binary responses in which one knows the set $ \{ \theta_{ C } , \theta_{ D } \} $,
but does not know which arm is which. \citet{BradtEtal1956} also proposed a generalization of that problem, in which $ \{ \theta_{ C } , \theta_{ D } \} $
is unknown, which is the two-armed bandit problem as known today and as considered in this paper. \citet{Bellman1956} referred to the latter problem as the
\emph{two-machine problem}.
P. Whittle stated on several occasions that researchers were aware of this type of problem since the 1940s and considered it an important but very hard
open problem.\footnote{``...it was formulated during the war, and efforts to solve it so sapped the energies and minds of Allied analysts that the
suggestion was made that the problem be dropped over Germany, as the ultimate instrument of intellectual sabotage'' \citet{Whittle1979discussion};
``...propounded during the Second World War, and soon recognized as so difficult that it quickly became a classic, and a by-word for intransigence.''
\citet{Whittle1989foreword}; ``...had resisted analysis, however, to the point of being regarded by some as intrinsically insoluble.''
\citet{Whittle2002}.} That could be attributed to the absence of a suitable mathematical framework and theory, as the progress on the problem occurred
slowly alongside the emergence and the development of areas such as sequential analysis, Bayesian statistics, decision theory, dynamic programming,
stochastic processes, and concentration inequalities.
Indeed, early papers describing theoretical solutions on the bandit problem were often among the pioneers in these areas, introducing novel terminology and
notation, not all of which has been adopted more generally, and might thus be hard to read for today's researchers. Drawing on the early research in 1950s
and 1960s, three dominant ``schools'' have emerged:
\begin{itemize}
\item the \emph{Berry's school}: starting with \citet{Berry1972}, and rewritten and further developed in \citet{BerryFristedt1985}, focussing
predominantly on the finite-horizon setting;
\item the \emph{Gittins' school}: starting with \citet{GittinsJones1974}, and rewritten and further developed in \citet{Gittins1979, Gittins1989,
GittinsEtal2011}, focussing predominantly on the discounted infinite-horizon setting;
\item the \emph{Robbins' school}: starting with \citet{LaiRobbins1985aam}, and rewritten and further developed in \citet{Agrawal1995aap,
KatehakisRobbins1995pnas, BurnetasKatehakis1996}, focussing predominantly on the time-average infinite-horizon setting.
\end{itemize}
The author's suggestion is that this pioneering literature should be on the must-read list of researchers on bandit problems, regardless of their
discipline.
While the Robbins' school makes complete learning (i.e. identification of the better arm in infinite time almost surely) lexicographically more important
than the way of attaining it, both the Berry's and Gittins' schools replace the lexicographic ordering by resolving the trade-off between complete
\emph{learning and earning} (of rewards), with the relative weights implicitly given by the horizon and the discount factor, respectively. In
\autoref{section:approaches} we describe these ``schools'' and their relationships in more detail.
Of course, there are several other fascinating variants of the bandit problem, with different objective (e.g., risk-averse, adversarial, final-period-only,
etc.), different control (e.g., randomized, multi-mode, multi-resource, duelling, multi-player, etc.), and/or different dynamics (e.g., non-binary
responses, delayed responses, partial observability, arriving arms, covariates, correlation, restlessness, non-stationarity, non-Markovian, etc.); all
these are unfortunately beyond the scope of this paper.
\subsection{Applications}
The problem has been formulated, addressed or applied in a number of disciplines, each developing its own terminology, see \autoref{table:terminology}. In
this paper we use the terminology which we believe is a reasonable compromise and should not cause confusion for researches and practitioners from all the
disciplines. The author wishes to encourage researchers from all the disciplines to follow this terminology as closely as possible to facilitate for
researchers and practitioners from other disciplines to learn about their work.
According to \citet{Scott2010asmbi}: ``Multi-armed bandits have an important role to play in modern production systems that emphasize \emph{continuous
improvement}, where products remain in a perpetual state of feature testing even after they have been launched.'' The most commonly listed applications,
often requiring to adapt the generic multi-armed bandit problem to specific features, are as follows:
\begin{itemize}
\item \emph{Digital marketing}: In the digital world it is relatively easy to introduce and quick to get feedback on new document variants, and so
bandit problems have been proposed for social media advertising, personalized websites and user interfaces, email campaigns, influence
maximization, etc; see, e.g., \citet{LiberaliEtal2017handbook}. Bandit problems can also be used to address the problem of dynamic pricing with
demand uncertainty, which requires to solve a trade-off of learning (of the demand curve) and earning (the highest revenue); for a survey, see,
e.g., \citet{denBoer2015sorms}.
\item \emph{Clinical trials}: \citet{Thompson1933biometrika} pointed out that his bandit problem ``...would be important in cases where either the rate
of accumulation of data is slow or the individuals treated are valuable, or both.'' \citet{Gluss1962iac} further explained the motivation primarily
by \emph{rare diseases}. Following the focus on rare and/or life-threatening diseases, a few novel bandit-based designs have been developed and
proposed recently, and are being implemented in a growing number of trials, mainly in several types of cancer, where patients are stratified into
smaller groups using genetic biomarkers. Discussions about the advantages and disadvantages of bandit-based designs are ongoing, e.g.,
\citet{BerryEsserman2016nejm} argue that, in certain clinical trials, data-driven approaches make great sense ethically, statistically,
economically, scientifically, and logistically. For a survey on real adaptive trials, see e.g., \citet{BothwellEtal2018bmj}.
\item \emph{Search}: Bandit designs have been proposed for recommender systems in which new items and users appear frequently in order to assure
sufficient exploration, see, e.g., \citet{Aggarwal2016book}. Although digital search is typically considered by recommender systems, many
non-digital search applications exist, e.g. search for natural resources, search and rescue, surveillance and monitoring. Related to this category
are also the so-called best-arm identification problem and problems appearing in ranking and selection.
\end{itemize}
\begin{table}[tbp]
\setlength{\tabcolsep}{5pt}
\centering
\begin{tabular}{lcccc}
\toprule
Anecdotic & strategy & choice & pull & arms \\
\midrule
Operations \& Management & policy & allocation & resource & projects \\
Reinforcement learning & algorithm & decision & time step & actions \\
Biometrics \& Biostatistics & design & randomization & patient & treatments \\
Ranking \& selection & policy & spread over & measurement & alternatives \\
Economics & strategy & choice & resource & experiments \\
Computing \& Telecom. & scheduler & allocation & server & jobs \\
Marketing & policy & allocation & impression & advertisements \\
Transportation & driver & selection & vehicle & roads \\
\midrule
This paper & \emph{design} & \emph{randomized action} & \emph{subject} & \emph{interventions/arms} \\
\bottomrule
\end{tabular}
\caption{An illustration of typical terminology for bandit problems across disciplines}\label{table:terminology}
\end{table}
\section{Model}
\label{section:model}
In this section we formulate a general two-armed problem with binary responses as a Markov decision process, which provides sufficient generality to
accommodate all the solution approaches discussed in \autoref{section:approaches}.
\paragraph{Interventions.}
We consider arms (or, \emph{interventions}) labelled by $ k \in \setK := \{ C , D \} $. A subject must be allocated to exactly one intervention, and such
allocation yields a \emph{binary} response from that intervention: $ 0 $ (failure) or $ 1 $ (success). The response set is denoted by $ \setO := \{ 0 , 1
\} $. Subject responses are uncertain, i.e., modelled as Bernoulli-distributed with parameter $ 0 \le \theta_{ k } \le 1 $, the \emph{success probability},
independent across arms. The responses are \emph{immediate}, meaning that the response of an allocated subject is observed before the next decision needs
to be done.
\paragraph{Timing.}
Subjects arrive (i.e., are recruited) sequentially (i.e., one by one) at random moments in continuous time. Since we do not discount the future, we can
without loss of generality focus only on the moments of subjects' arrivals, which we call discrete \emph{time epochs} and see as regularly spaced. That is,
equivalently, we can consider that subjects arrive at time epochs $ t \in \setT := \{ 0 , 1 , 2 , \dots , T - 1 \} $, where $ T \le +\infty $ is the number
of subjects in the trial, i.e., the \emph{trial size}, or the \emph{time horizon}. To clarify, the $ ( t + 1 ) $-st subject arrives at time epoch $ t $.
Note that $ t = T $ is the time epoch denoting the end of the trial, when the response of the last subject is observed and no subject arrives.
\paragraph{States.}
At any moment in continuous time, the \emph{physical state} is represented by the numbers of observed successes and failures on each arm, the number of
allocated subjects without an observed response on each arm, the number of arrived subjects without being allocated, and the number of remaining subjects
to arrive. This is a vector with $ 8 $ elements summing up to $ T $ at any moment (all are non-negative integers). At time epochs, this can be simplified
without loss of generality to a vector with $ 5 $ elements, with the numbers of observed successes and failures on arm $ C $ denoted by $ s_{ C } $ and $
f_{ C } $, respectively, the numbers of observed successes and failures on arm $ D $ denoted by $ s_{ D } $ and $ f_{ D } $, respectively, and the number
of remaining subjects to be allocated (exactly one of which has arrived), $ n $. Since at time epochs $ s_{ C } + f_{ C } + s_{ D } + f_{ D } + n = T $, it
is sufficient to keep track of any four of these five numbers, leading to a state as vector with $ 4 $ elements, which we choose to be $ \vecx := ( s_{ C }
, f_{ C } , s_{ D } , f_{ D } ) $. Note that at time epoch $ t $, $ s_{ C } + f_{ C } + s_{ D } + f_{ D } = t $.
In addition to the physical state, there is an \emph{information state}, which at any moment in continuous time captures all the information that could
possibly affect the decisions. This may include real-world evidence and/or modelling assumptions. The real world evidence may be available before the start
and/or it can arrive anytime during the trial. The modelling assumptions typically refer to the parameters of the prior distributions (built on historical
data or expert opinions) for the success probability of each arm (whose weight may change over time, and can be either informative or non-informative), but
may also include other parameters such as the probability of dropouts, the probability of errors in recording the observations and/or the subject
allocations, the probability of mistakes in the statistical analysis and/or in the administration process, the timing of planned interim analyses, the
probability and/or timing of unplanned stopping of the trial due to safety concerns, the estimate of the size of the subject population after the end of
the trial, etc. For full generality, we consider the information state $ \veci $ (potentially dependent on the current physical state $ \vecx $ and/or
otherwise changing during the trial), and thus the state is $ ( \vecx , \veci ) $.
\paragraph{Actions.}
At every time epoch $ t \in \setT $ the design must prescribe how the arrived subject should be randomized (i.e., randomly allocated) to interventions.
While there are only two possible allocations, in every state we consider a possibly infinite \emph{action set} $ \setA_{ ( \vecx , \veci ) } $ of
randomized actions $ a $ identified by probabilities $ ( p_{ C }^{ a } , p_{ D }^{ a } ) $, meaning that the subject is allocated to intervention $ C $ ($
B $) with probability $ p_{ C }^{ a } $ ($ p_{ D }^{ a } $). Formally, $ \setA_{ ( \vecx , \veci ) } \subseteq \{ a : p_{ C }^{ a } \ge 0 , p_{ D }^{ a }
\ge 0 , p_{ C }^{ a } + p_{ D }^{ a } = 1 ) \}. $ Since from the theory of Markov decision processes it follows that an action which is a randomized
combination of other two actions is optimal only if all three are optimal, it is sufficient to consider only an action set of two \emph{pure randomized
actions}, which we call action $ 1 $ (identified by $ ( p_{ C }^{ 1 } , p_{ D }^{ 1 } ) $) and action $ 2 $ (identified by $ ( p_{ C }^{ 2 } , p_{ D }^{ 2
} ) $). For convenience in situations when both actions are optimal, we also consider an \emph{equally-weighted mixed randomized action}, which we call
action $ 3 $ (identified by $ ( p_{ C }^{ 3 } , p_{ D }^{ 3 } ) := ( ( p_{ C }^{ 1 } + p_{ C }^{ 2 } ) / 2 , ( p_{ D }^{ 1 } + p_{ D }^{ 2 } ) / 2 ) $),
which is a combination of the two pure randomized actions with equal weights. Formally, $ \setA_{ ( \vecx , \veci ) } = \{ 1 , 2 , 3 \} $, and without loss
of generality we assume $ p_{ C }^{ 1 } \ge p_{ C }^{ 2 } $. In some approaches discussed in \autoref{section:approaches}, there is no choice of actions,
meaning that the cardinality of $ \setA_{ ( \vecx , \veci ) } $ is one, which can be obtained by setting $ 1 \equiv 2 \equiv 3 $, effectively reducing the
Markov decision process to a Markov reward process.
In some approaches, the action set depends on the observations only via their sum $ t = s_{ C } + f_{ C } + s_{ D } + f_{ D } $, thus can be written as $
\setA ( t ) $. Finally, the simplest case is the one in which the action set is constant, which we write as $ \setA $.
\paragraph{Transition Probabilities.}
Denote by $ q_{ k , ( \vecx , \veci ) , o } $ the probability of observing response $ o \in \setO $ for the current subject if it is allocated to arm $ k
\in \setK $ in state $ ( \vecx , \veci ) $. We assume that $ \sum_{ o \in \setO } q_{ k , ( \vecx , \veci ) , o } = 1 $ for all $ k $, but this can be
relaxed in some models, e.g. if allowing for dropouts (i.e., missing responses).
If the information state $ \veci $ does not change during the trial, then the transition probabilities of moving from state $ ( \vecx , \veci ) $ to state
$ ( \vecx^{ \prime } , \veci ) $ under action $ a $ are
\begin{align*}
h_{ ( \vecx , \veci ) , ( \vecx^{ \prime } , \veci ) }^{ a } =
\begin{cases}
p_{ C }^{ a } q_{ C , ( \vecx , \veci ) , 1 } & \text{ if } \vecx^{ \prime } = \vecx + \vece_{ 1 } \\
p_{ C }^{ a } q_{ C , ( \vecx , \veci ) , 0 } & \text{ if } \vecx^{ \prime } = \vecx + \vece_{ 2 } \\
p_{ D }^{ a } q_{ D , ( \vecx , \veci ) , 1 } & \text{ if } \vecx^{ \prime } = \vecx + \vece_{ 3 } \\
p_{ D }^{ a } q_{ D , ( \vecx , \veci ) , 0 } & \text{ if } \vecx^{ \prime } = \vecx + \vece_{ 4 } \\
\end{cases}
\end{align*}
where $ \vece_{ j } $ is the standard basis vector.
If the information state changes during the trial, then these transition probabilities need to be amended to reflect its dynamics.
\paragraph{Expected One-Period Rewards.}
The expected one-period reward $ r_{ ( \vecx , \veci ) }^{ a } $ for all states $ ( \vecx , \veci ) $ and all actions $ a $ needs to be defined. If the
information state $ \veci $ does not change during the trial, then it is as follows: $ r_{ ( \vecx , \veci ) }^{ a } = 0 $ for all states such that $ s_{ C
} + f_{ C } + s_{ D } + f_{ D } \le T - 1 $ and $ r_{ ( \vecx , \veci ) }^{ a } = s_{ C } + s_{ D } $ for all states such that $ s_{ C } + f_{ C } + s_{ D
} + f_{ D } = T $ (i.e., the reward is the number of observed successes in all states in which the trial can eventually end).
The above definition of the reward is novel. The conventional one is to set the reward to the expected value of observing one success in a given state
under a given action, i.e., $ r_{ ( \vecx , \veci ) }^{ a } = p_{ C }^{ a } q_{ C , ( \vecx , \veci ) , 1 } + p_{ D }^{ a } q_{ D , ( \vecx , \veci ) , 1 }
$ at all time epochs $ t = 0 , 1 , \dots , T - 1 $ and $ r_{ ( \vecx , \veci ) }^{ a } = 0 $ at time epoch $ t = T $. In
\autoref{section:reward_equivalence} we prove the following theorem.
\begin{theorem}\label{theorem:1}
The two reward definitions give the same expected total reward for any fixed design.
\end{theorem}
\paragraph{State and Action Processes.}
The evolution of a Markov decision process is captured by the state process, which in full generality is 2-dimensional in order to keep the physical and
information states separately, $ ( \vecX ( \cdot ) , \vecI ( \cdot ) ) $, and the action process which depends on the state process, but can be briefly
written as $ A ( \cdot ) $, where $ A_{ ( \vecX ( t ) , \vecI ( t ) ) } \in \setA_{ ( \vecX ( t ) , \vecI ( t ) ) } $.
\section{Assumptions, Settings and Objectives}
\label{section:settings}
\subsection{Usage}
There are three principal types of usage of the model described above.
\paragraph{Evaluation by Simulation.}
Computer simulation is now a commonly used evaluation tool as it is relatively straightforward and the accuracy vs runtime trade-off can be addressed by
adjusting the number of simulation runs. But we believe that it has the \emph{law-of-the-hammer} syndrome of all simple universal tools: ``if the only tool
you have is a hammer, to treat everything as if it were a nail.''
\paragraph{Evaluation by Backward Recursion.}
In this paper, we give evidence that it is possible and preferable to use backward recursion instead of simulation for evaluation. This yields a perfectly
accurate evaluation (subject to computational accuracy of the chosen numerical type). We discuss its runtime in \autoref{section:myths}.
\paragraph{Optimization.}
When the action set is not singular in all states, there is room for choosing one of the actions for every state according to an objective of maximizing
some function. This does not necessarily need to be done by backward recursion; we describe several approaches in \autoref{section:approaches}. In this
case we assume that the success probabilities are unknown as otherwise it is trivial to optimize.
\subsection{Knowledge Assumptions}
While the success probabilities are assumed unknown for optimization, they may not necessarily be so for evaluation. There are two principal ways of
specifying the probabilities $ q_{ k , ( \vecx , \veci ) , o } $ of observing response $ o $, which depend on the knowledge assumption about the success
probabilities.
\paragraph{Known Success Probabilities.}
Evaluation of all the approaches described in \autoref{section:approaches} can be done by assuming that success probabilities $ \theta_{ k } $ are known,
and part of $ \vecI ( t ) $ for all $ t $. In that case the transition probabilities are independent of the physical state, so we can write $ q_{ k , \veci
, o } $. If the information state is $ \veci = ( \theta_{ C } , \theta_{ D } ) $ during the whole trial, then
\begin{align*}
q_{ k , \veci , o } =
\begin{cases}
\theta_{ k } & \text{ if } o = 1 \\
1 - \theta_{ k } & \text{ if } o = 0 \\
\end{cases}
\end{align*}
\paragraph{Unknown Success Probabilities.}
Approaches that allow for optimization assume that the success probabilities are unknown (otherwise the decision with the objective of maximising the
expected number of successes is trivial), and so require estimates of these, which can be obtained using Bayesian updating. Following the existing
literature, we use the Bayesian Beta-Bernoulli model for each arm $ k \in \setK $, in which $ \theta_{ k } $ is assumed to be a random variable drawn from
Beta distribution dependent on the state. At the initial time epoch $ t = 0 $, i.e., in physical state $ ( 0 , 0 , 0 , 0 ) $, each arm $ k $ is given a
prior Beta distribution with parameters $ \widetilde{ \successes }_{ k } ( 0 ) , \widetilde{ \failures }_{ k } ( 0 ) $. These parameters can be interpreted
as the numbers of pseudo-observations of successes and failures before the start of the trial. They are thus part of the information state $ \vecI ( t ) $
for all $ t $ and do not change over time. At every time epoch $ t $, in physical state $ \vecx = ( \successes_{ C } , \failures_{ C } , \successes_{ D } ,
\failures_{ D } ) $, each arm $ k $ has the posterior distribution given, because of conjugacy, by the Beta distribution with parameters $ \widetilde{
\successes }_{ k } , \widetilde{ \failures }_{ k } $, briefly $ \text{Beta} ( \widetilde{ \successes }_{ k } , \widetilde{ \failures }_{ k } ) $, where $
\widetilde{ \successes }_{ k } = \widetilde{ \successes }_{ k } ( 0 ) + \successes_{ k } $ and $ \widetilde{ \failures }_{ k } = \widetilde{ \failures }_{
k } ( 0 ) + \failures_{ k } $.
If the information state is $ \veci = ( \widetilde{ \successes }_{ C } ( 0 ) , \widetilde{ \failures }_{ C } ( 0 ) , \widetilde{ \successes }_{ D } ( 0 ) ,
\widetilde{ \failures }_{ D } ( 0 ) ) $ during the whole trial, then
\begin{align*}
q_{ k , ( \vecx , \veci ) , o } =
\begin{cases}
\frac{ \widetilde{ \successes }_{ k } }{ \widetilde{ \successes }_{ k } + \widetilde{ \failures }_{ k } } & \text{ if } o = 1 \\
\frac{ \widetilde{ \failures }_{ k } }{ \widetilde{ \successes }_{ k } + \widetilde{ \failures }_{ k } } & \text{ if } o = 0 \\
\end{cases}
\end{align*}
The conventional assumption is to take the uniform distribution as a prior distribution on each arm $ k $, i.e., $ ( \widetilde{ \successes }_{ k } ( 0 ) ,
\widetilde{ \failures }_{ k } ( 0 ) ) = ( 1 , 1 ) $; for a discussion on choosing a different prior Beta distribution, see \autoref{section:priors}.
\subsection{Performance Measures and Objectives}
In this paper we focus on the \emph{number of successes} as the principal performance measure (a.k.a. operating characteristic in clinical trials
literature). But instead of the number of successes, we report two equivalent measures: the \emph{proportion of successes} and the \emph{regret number of
successes}, as these provide complementary interpretation and insights. Yet another equivalent measure (not reported in this paper) is the \emph{fraction
of subjects allocated to the better arm}. Many other additive measures, e.g. monetary cost typical in health economics and health technology assessment,
can be defined analogously, but are not discussed in this paper.
A particular design $ \pi $ prescribes the action process $ A ( \cdot ) $. Let $ \Pi $ be the set of designs that are non-anticipating\footnote{A
non-anticipating design is a design which cannot see into the future; i.e., an action prescribed by the design at a given time epoch does not require the
knowledge of states which have not been observed yet.} and satisfy the above constraints on $ A ( \cdot ) $.
Let us denote by $ \Expectation^{ \pi }_{ t } [ \cdot ] $ the expectation under design $ \pi \in \Pi $ conditioned on information available at time epoch $
t \in \setT $. The mean number of successes is
\begin{align}
\totalNumber^{ \pi }_{ t } \left( \vecx , \veci \right) := \Expectation^{ \pi }_{ t } &\left[ \left. \sum_{ u = t }^{ T } r_{ ( \vecX ( u ) , \vecI ( u ) ) }^{ A_{ ( \vecX ( u ) , \vecI ( u ) ) } } \right| \left( \vecX ( t ) , \vecI ( t ) \right) = \left( \vecx , \veci \right) \right] .
\end{align}
and the mean proportion of successes is
\begin{align}
\totalProportion^{ \pi }_{ t } \left( \vecx , \veci \right) := \frac{ 1 }{ T - t } \totalNumber^{ \pi }_{ t } \left( \vecx , \veci \right) .
\end{align}
These two are measures of \emph{subject benefit}, while the former is on the absolute scale, the latter yields the average per-subject probability of
observed success, i.e., the subject benefit on the percentage scale.
We further define the mean regret number of successes,
\begin{align}
\totalRegret^{ \pi }_{ t } \left( \vecx , \veci \right) := \Expectation \left[ \left. \max\{ \theta_{ C } , \theta_{ D } \} \right| \vecI ( t ) = \veci \right] - \totalNumber^{ \pi }_{ t } \left( \vecx , \veci \right) ,
\end{align}
which is a measure of \emph{subject loss}. Note that all the three measures depend on parameter $ T $, although we have suppressed the explicit notation.
The objective is to find an optimal design $ \pi^{ * } $ that maximises the mean number of successes as evaluated at time epoch $ t = 0 $ when there are no
observations ($ \vecx = \veczero $), i.e.,
\begin{align}
\pi^{ * } := \argmax_{ \pi \in \Pi } \totalNumber^{ \pi }_{ 0 } \left( \veczero , \veci \right)
\end{align}
or, equivalently, maximizes the mean proportion of successes, or, equivalently, minimizes the mean regret number of successes.
Following the two knowledge assumptions above, we have two approaches to performance evaluation.
\paragraph{Known Success Probabilities.}
When the success probabilities are assumed to be known, we call the above measures the \emph{frequentist number of successes}, the \emph{frequentist
proportion of successes}, and the \emph{frequentist regret number of successes}, respectively. Due to symmetry, we can assume that $ \theta_{ C } \le
\theta_{ D } $ without loss of generality, and thus
\begin{align}
\totalRegret^{ \pi }_{ t } \left( \vecx , \veci \right) = \theta_{ D } - \totalNumber^{ \pi }_{ t } \left( \vecx , \veci \right) .
\end{align}
\paragraph{Unknown Success Probabilities.}
When the success probabilities are assumed to be unknown, we evaluate performance in terms of the quantities known in the literature as the Bayes return,
Bayes worth, or Bayes risk. In particular, we call the above measures the \emph{Bayes number of successes}, the \emph{Bayes proportion of successes}, and
the \emph{Bayes regret number of successes}, respectively.
For a problem with uniform distribution as a prior distribution on each arm, as considered in this paper, $ \Expectation \left[ \left. \max\{ \theta_{ C }
, \theta_{ D } \} \right| \vecI ( t ) = \veci \right] = 2 / 3 $ \citep{Berry1978jasa}, thus
\begin{align}
\totalRegret^{ \pi }_{ t } \left( \vecx , \veci \right) = \frac{ 2 }{ 3 } - \totalNumber^{ \pi }_{ t } \left( \vecx , \veci \right) ,
\end{align}
Note that our definition is different from the so-called Bayesian regret \citep[cf.][Section 34.6]{LattimoreSzepesvari2019book}, which is the average of
the frequentist regret with respect to the prior distribution, i.e. it is Bayesian only in the initial time epoch.
\section{Approaches}
\label{section:approaches}
In this section we describe a number of common approaches to the problem described in the previous section. We start with Equal Randomized Allocation,
which was originally developed for static setting, in which all the subjects arrive at the same time. All the remaining approaches are suitable only in
dynamic setting, i.e., are data-driven.
\subsection{Design of Experiments}
Statistical testing based on \emph{equal (i.e., \design{1:1}) randomized allocation} (a.k.a. random assignment) is a widespread state-of-the-art approach
in the design of experiments today, known as \emph{randomized controlled trial} in medicine, \emph{between-group design} in social sciences, and \emph{A/B
testing} in digital marketing. I theory, it allows the greatest reliability and validity of statistical estimate of the intervention effect (i.e., the
difference between the two success probabilities) under an initial equipoise assumption. Originally developed, advocated and popularized by the founders of
statistics such as C. S. Peirce in the fields of psychology and education in the late 19th century, and J. Neyman and R. A. Fisher in agriculture and other
fields in the early 20th century. In the middle of the 20th century, A. B. Hill popularized the method in the field of medicine and it is now the preferred
approach by regulatory agencies for assessing efficacy when deciding about marketing authorisation of new medicinal interventions in many countries. The
approach can be evaluated using routine (frequentist) statistical methods.
In this approach the transition probabilities are frequentist, and
\begin{itemize}
\item the information state $ \veci $ is ignored;
\item the action set is constant, $ | \setA | = 1 $ and $ p_{ C }^{ a } = p_{ D }^{ a } = 1 / 2 $;
\end{itemize}
If the success probabilities are known, then its performance can be evaluated directly: the proportion of successes with mean $ ( \theta_{ C } +
\theta_{ D } ) / 2 $ and standard deviation $ \sqrt{ ( \theta_{ C } + \theta_{ D } ) ( 2 - \theta_{ C } - \theta_{ D } ) / 4 T }$; the regret
number of successes with mean $ T \max \{ ( \theta_{ C } - \theta_{ D } ) / 2 , ( \theta_{ D } - \theta_{ C } ) / 2 \} $ and standard deviation
$ \sqrt{ T ( \theta_{ C } + \theta_{ D } ) ( 2 - \theta_{ C } - \theta_{ D } ) / 4 }$.
The approach is believed to be well understood. However, there are several potential concerns, e.g., (i) it was developed under the assumption of an
infinite population of subjects to which the best arm will be applied, but it is not clear how valid it is when the population is finite or when new,
potentially better arms become available in future; (ii) it was developed under the assumption of parallel allocation of subjects, but in dynamic (i.e.
sequential) setting, in which subjects arrive one by one, there is a risk of introducing accrual and/or allocation bias in the estimation of the success
probabilities if the person (or machine) delivering an intervention is aware of the characteristics and/or responses of the previous subjects; (iii) it was
developed under the assumption of the simple random sampling method, which is not true if the subjects need to provide a consent (which is a legal
requirement, for instance, in clinical trials); (iv) it was developed under the assumption of the use of random numbers, which not true for computer
generated randomization (pseudo-random numbers) provided by error-prone software and procedures of external companies.
\subsection{Bayesian Decision Theory}
Bayesian Decision Theory emerged together with game theory and mathematical programming in the middle of the 20th century, building on models and
techniques from Bayesian statistics, applied mathematics, and economics. A number of prominent researchers contributed to the development of the field,
including A. Wald, J. Wolfowitz, L. J. Savage, K. J. Arrow, D. Blackwell, H. Raiffa, R. Schlaifer, R. Bellman, etc.
The two-armed problem with binary responses was first formulated in this framework in \citet{Bellman1956}, with a known success probability of one arm,
over an infinite horizon $ T = + \infty $, and with geometric discounting of future rewards. He proposed to employ backward recursion based on the Bellman
equation from dynamic programming he had developed. \citet{Gluss1962iac} extended the model to two unknown arms using the same approach and discussed
the theoretical memory requirement relevant for online optimization assuming a finite-horizon truncation, after which the better arm is used forever.
\citet{Steck1964moc} programmed the backward recursion on one of the scientific computers of those times (Univac 1105) and listed the optimal allocations
with the truncation at $ T = 25 $ (and the discount factor $ 0.95 $).
\begin{figure}[tbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/time_action_lin_32-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/memory_online_action_lin_32-pure.pdf}
\end{subfigure}%
\caption{An illustration of computational complexity of online calculation of the deterministic Bayesian decision-theoretic design over a range of trial sizes.}\label{fig:1}
\end{figure}
For finding the best action for each state via this approach, the transition probabilities are Bayesian, and
\begin{itemize}
\item the conventional assumption is to take as information state $ \veci $ (that does not change over time) the prior distribution on each arm $ k $;
\item the action set is constant, $ | \setA | = 3 $;
\end{itemize}
Since \autoref{theorem:1} is true for any design, it is also true for the optimal design, which is derived using the Bellman equation, which is an
optimization equivalent of the Poisson equation.
\begin{theorem}
The two reward definitions give the same optimal design and to the same optimal expected total reward.
\end{theorem}
The conventional assumption is that the two pure actions are deterministic, i.e., $ p_{ C }^{ 1 } = p_{ D }^{ 2 } = 1 , p_{ C }^{ 2 } = p_{ D }^{ 1 } = 0
$, which we will refer to as the \emph{deterministic Bayesian decision-theoretic design}, despite the fact that randomization is allowed when both actions
are optimal.
A more general setting, in which the pure actions are allowed to be randomized instead of deterministic, was proposed in \citet{ChengBerry2007biometrika}
and further developed in \citet{WilliamsonEtal2017csda}. In fact, such a setting provides a continuum of designs, ranging from the deterministic Bayesian
decision-theoretic design to the equal randomized allocation design, recovered by setting $ p_{ C }^{ 1 } = p_{ C }^{ 2 } = p_{ D }^{ 1 } = p_{ D }^{ 2 } =
1 / 2 $.
\subsubsection{Optimal --- Dynamic Programming}
The Bayesian decision-theoretic model can be solved by (stochastic) dynamic programming (\design{DP}), which comprises of a calculation starting by
enumerating all the possible states in the final time period, continuing backwards in time while employing the Bellman equation in every state.
Unfortunately, complete structure of the optimal design is unknown, thus numerical computation is the only way of obtaining it.
\autoref{fig:1} illustrates the computational complexity of \emph{online calculation} (i.e., outputting the optimal action of the initial state) of this
design, using a state-of-the-art package \emph{BinaryBandit} in Julia programming language. In the number of elements (i.e., without multiplying by the
memory required for each element), the memory requirement presented in \autoref{fig:1}(right) is $ 2 $ times larger than that of \citet{Gluss1962iac}
because of his elimination of symmetric states, which can be done if the prior Beta distributions are the same for the two arms.
The computational complexity of \emph{offline calculation} (i.e., outputting the optimal actions of all possible states) of this design is very similar to
online calculation in terms of the runtime, but radically different in terms of memory requirement. Using the \emph{BinaryBandit} Julia package, a computer
with 32GB RAM is able to solve the problem of up to trial size $ T = 1,440 $, keeping all the optimal actions in RAM during the calculations. For practical
purposes, however, it may be possible to store parts of the solution on hard disk and thus relieve RAM memory to allow calculation for larger trial sizes.
A number of author's colleagues and PhD students who had programmed this design for their needshave kindly provided runtimes of their code implementations,
which are presented in \autoref{table:code_runtime} for comparison, indicating that the BinaryBandit package is two orders of magnitude faster than ad hoc
codes and is able to solve a few times larger problems.
\begin{table}[tbp]
\centering
\begin{tabular}{lrrrrrrr}
\toprule
Software & RAM & $ T = 60 $ & $ T = 120 $ & $ T = 180 $ & $ T = 240 $ & $ T = 300 $ & $ T^{ \max } $ \\
\midrule
Julia 0.6.2 \& ad hoc & 12 GB & 2sec & 22sec & 108sec & 331sec & 789sec & 420 \\
Julia 1.0.1 \& ad hoc & 12 GB & 1sec & 17sec & 82sec & 262sec & 643sec & 420 \\
R \& ad hoc & 16 GB & 1sec & 12sec & 59sec & 191sec & N/A & 240 \\
Julia 1.0.1 \& BB & 31 GB & 0.0036sec & 0.046sec & 0.23sec & 0.73sec & 1.6sec & 1440 \\
\midrule
R \& ad hoc & 5 GB & 1sec & 6sec & 26sec & 84sec & 209sec & 420 \\
Julia 1.0.1 \& BB & 31 GB & 0.0040sec & 0.056sec & 0.27sec & 0.91sec & 2.8sec & 4440 \\
\bottomrule
\end{tabular}
\caption{A comparison of runtime of the Bayesian decision-theoretic design for $ T = 60 : 60 : 300 $ and the largest horizon $ T^{ \max } $ (as a multiple of $ 60 $) which does not give an out-of-memory error using code implementations by author's colleagues and students. The top four are for offline calculation, the bottom two are for online calculation. BB refers to the use of the BinaryBandit package.}\label{table:code_runtime}
\end{table}
\subsubsection{Asymptotically Optimal --- Gittins and Whittle Index Rules}
\label{section:Gittins}
The structure of the optimal Bayesian decision-theoretic designs with deterministic pure actions is however known when considering the rewards over an
infinite horizon $ T \to + \infty $ and discounted with a geometrically-distributed discount factor $ 0 < \gamma < 1 $. \citet{GittinsJones1974} discovered
that optimal allocations can be characterized by an \emph{index rule}, which allows (Gittins) index values to be calculated for every arm separately, and
which at every time epoch allocates a subject to the arm with the highest Gittins index value (breaking the ties arbitrarily). See, e.g., \citet{Gittins1979,
Gittins1989, GittinsEtal2011} for general theory on the Gittins index.
Although theoretically appealing and useful in many other problems, in the setting of the Bayesian decision-theoretic design it does not provide an
ultimate solution, because the states are time-dependent and the horizon $ T $ is finite. Still, the Gittins index rule can be used as an approximation to
the optimal design, being asymptotically optimal as $ T \to +\infty $. It decreases the computational complexity of the offline calculation notably because
of a decreased size of the problem (one arm, i.e. two dimensions), but is still computed by dynamic programming, which is computationally demanding. In its
calculation, however, it is necessary either to use a discount factor and/or a horizon truncation, or to compute it for every horizon $ T $ considered by
adding the remaining number of subject allocations as a third dimension. In the former case, the resulting index rule is only an approximation and tends to
focus on learning less than the optimal design, see e.g. \citet{VillarEtal2015survey}. In the latter case, \citet[Section 6.2]{Nino2011} provides a
comparison of an approximate, so-called calibration, algorithm (using a grid of values to desired accuracy) and an exact algorithm for the calculation of
the Whittle index values, showing that already $ T \le 100 $ requires several GB of RAM, which suggests that the calibration algorithm with a grid of not
more than $ 4 $ significant digits is the only practical method for larger horizons.
Besides the requirement of infinite horizon, the theory of the Gittins index only applies when the pure actions are deterministic. When these are
randomized, the problem becomes so-called \emph{restless}, meaning that more than one arm can change its state in every period. Also, even when the pure
actions are deterministic, but the horizon is finite, the problem can be seen as restless, by adding the remaining number of subject allocations to the
state of each arm. \citet{Whittle1988} proposed to solve the restless problem also by an index rule, acknowledging that such a rule would not necessarily
be optimal, but conjecturing that it would admit a form of asymptotic optimality as both the number of arms and the number of allocated arms in each period
grow to infinity at a fixed proportion, which was eventually proved in \citet{WeberWeiss1990} under some technical assumptions. Whittle defined an index,
which reduces to the Gittins index in the non-restless setting, which became known as the \emph{Whittle index}. The above discussion suggests that the
Whittle index rule is conceptually more appropriate and more accurate than the Gittins index rule, and this was confirmed for the Bayesian
decision-theoretic design numerically, see, e.g., \citet{VillarEtal2015survey, Villar2018peis}. The computational complexity of the approximate Whittle
index values calculated by the calibration method for every horizon $ T $ considered (using a grid of values to desired accuracy) is similar to that of the
approximate Gittins index values using the same method.
Note that both the Gittins and Whittle index rule require to keep $ 3 $ elements in the one-arm state, so the reduction from the optimal two-armed problem
is only by one dimension. Moreover, in offline calculation, the index values that need to be stored are non-integer, thus require $ 64 $ bits per one-arm
state, while the offline calculation of the solution to the two-armed problem stores directly optimal actions, which require $ 2 $ bits per two-arm state.
Thus, the index rules (calculated to a few significant digits) are typically preferable to dynamic programming in calculation for horizons around $ T \ge
1000 $ and only if planned to be used for evaluation by simulation or for implementation in practice. However, \citet{Kaufmann2018aos} reports that she was
able to compute the Gittins index values only up to $ T = 1000 $.
Of course, advantages of index rules become important in problems with more than two arms, in which dynamic programming suffers from the \emph{curse of
dimensionality}. A discussion of such problems is however beyond the scope of this paper.
\subsubsection{Approximately Optimal --- Approximate Dynamic Programming}
Several general approaches have been proposed to deal with the curse of dimensionality of stochastic optimization problems, which are collectively known as
the approximate dynamic programming. There is a number of approximation techniques, but broadly focus on problem size reduction (e.g., the state space is
approximated by a grid for which optimal actions are computed, and interpolated on non-included states) and/or on simplification of function computations
(e.g. the value function is approximated by looking at decisions only a few periods ahead). See, e.g., \citet[Section 6.6]{PowellRyzhov2018book},
\citet{AhujaBirge2019report}.
We will describe one approach, which leads to a Bayesian design known as the \emph{knowledge gradient} (\design{BKG}). The fundamental idea to reduce the
amount of information and computation required for a decision in a given state at a given time epoch is to assume that this decision is the last one, and
in the next period we will identify the better arm which will whence be allocated to all the remaining subjects. \citet[Section 7.2]{FrazierEtal2008sjco}
showed that it is optimal for a search variant (i.e., maximizing the final-period expected reward) of the two-armed bandit problem with continuous
responses. A specific variant of this approach for our setting was presented in \citet[Section 4.7.1]{PowellRyzhov2012book} and further studied and
improved in \citet{EdwardsEtal2017peis}.
Let us denote by $ \mu_{ k } ( \widetilde{ \successes }_{ k } ( t ) , \widetilde{ \failures }_{ k } ( t ) ) $, or briefly $ \widetilde{ \mu }_{ k } $, the
\emph{belief}, i.e., the mean of the posterior Beta distribution of arm $ k $ calculated using posterior observations (i.e., both the observations and
prior pseudo-observations), i.e.,
\begin{align}
\widetilde{ \mu }_{ k } := \frac{ \widetilde{ \successes }_{ k } ( t ) }{ \widetilde{ \successes }_{ k } ( t ) + \widetilde{ \failures }_{ k } ( t ) } .
\end{align}
We will further denote the belief conditional on an additional (not necessarily binary) observation $ \delta $, respectively, by
\begin{align}
\widetilde{ \mu }_{ k }^{ +\delta } := \frac{ \widetilde{ \successes }_{ k } ( t ) + \delta }{ \widetilde{ \successes }_{ k } ( t ) + \widetilde{ \failures }_{ k } ( t ) + 1 } .
\end{align}
Under independent prior Beta distributions, the allocation at epoch $ t $ by this design is to the arm $ k $ with currently the largest value of
the following score
\begin{align}
\begin{cases}
\widetilde{ \mu }_{ k } + ( T - t - 1 ) \widetilde{ \mu }_{ \ell } & \text{ if } \widetilde{ \mu }_{ \ell } \ge \widetilde{ \mu }_{ k }^{ +1 } \\
\widetilde{ \mu }_{ k } + ( T - t - 1 ) \left[ ( 1 - \widetilde{ \mu }_{ k } ) \widetilde{ \mu }_{ \ell } + \widetilde{ \mu }_{ k } \widetilde{ \mu }_{ k }^{ +1 } \right] & \text{ if } \widetilde{ \mu }_{ k }^{ +1 } \ge \widetilde{ \mu }_{ \ell } \ge \widetilde{ \mu }_{ k }^{ +0 } \\
\widetilde{ \mu }_{ k } + ( T - t - 1 ) \widetilde{ \mu }_{ k }^{ +\widetilde{ \mu }_{ k } } & \text{ if } \widetilde{ \mu }_{ k }^{ +0 } \ge \widetilde{ \mu }_{ \ell }
\end{cases}
\end{align}
where $ \ell $ refers to the other arm, and ties are broken randomly. Although this score can be interpreted as capturing the allocation priority similarly
to the Gittins and Whittle indices, it depends on the state of the other arm, thus, strictly speaking, it is not an index.
\subsection{Na\"{\i}ve Designs}
While in the previous subsection we described designs which are horizon-dependent and forward-looking in the way of making allocation decisions, we now
turn our attention to horizon-independent designs and present a number of different na\"{\i}ve approaches in this subsection. Such designs are
computationally simple and easy to interpret, hence sometimes more appealing for practical use. It turns out that such deterministic designs are either
related to \emph{myopia} ($ T = 1 $) or to \emph{utopia} ($ T = +\infty $).
We will say that a design leads to \emph{complete learning}, if in the problem considered over an infinite horizon (i.e., $ T = +\infty $) it identifies
the better arm almost surely. It is easy to see that to achieve complete learning it is necessary to allocate each arm infinitely often.
\subsubsection{Myopic Index Rules
\citet{BradtEtal1956} considered the setting in which one knows the two-point set $ \{ \theta_{ C } , \theta_{ D } \} $, but does not know which arm is
which, and proved that if $ \theta_{ C } + \theta_{ D } = 1 $, then it is optimal, under any prior distributions, to deterministically allocate every
subject to the arm with currently largest Bayesian expected one-period reward. The also proved the same in the case $ \theta_{ C } + \theta_{ D } = 1 $
when the set $ \{ \theta_{ C } , \theta_{ D } \} $ is unknown. For the case $ \theta_{ C } + \theta_{ D } \neq 1 $ with known set $ \{ \theta_{ C } ,
\theta_{ D } \} $ they proved that such a design leads to complete learning and conjectured that it is optimal in terms of the Bayesian expected number of
successes. They also discussed that such a design is not optimal if both of these assumptions are dropped.
Note that under independent Beta prior distributions, the allocation at epoch $ t $ by this design is to the arm $ k $ with currently the largest mean
calculated using posterior observations (i.e., both the observations and prior pseudo-observations),
i.e.,
\begin{align}
\frac{ \widetilde{ \successes }_{ k } ( t ) }{ \widetilde{ \successes }_{ k } ( t ) + \widetilde{ \failures }_{ k } ( t ) }
\end{align}
where ties are broken randomly. We call this the \emph{Bayesian myopic} (\design{BM}) design, because it takes an action which is best for the next subject
only and ignores the future.
\citet{Feldman1962ams} also considered the setting with known two-point set $ \{ \theta_{ C } , \theta_{ D } \} $ and further generalized the optimality of
the Bayesian myopic design. \citet[Section 8]{Berry1972} realized that the setting with known two-point set $ \{ \theta_{ C } , \theta_{ D } \} $ (i.e.,
with strong between-arm dependence) is equivalent to the two-armed problem with independent arms with two-point prior distributions. \citet{Kelley1974as}
further generalized the setting with dependent arms and identified conditions for optimality of the Bayesian myopic design.
We also consider the \emph{frequentist myopic} (\design{FM}) design, which allocates each arm once in the first two periods, and then deterministically
allocates every subject to the arm with currently the largest mean calculated using the observations only (breaking the ties randomly), i.e.,
\begin{align}
\frac{ \successes_{ k } ( t ) }{ \successes_{ k } ( t ) + \failures_{ k } ( t ) }
\end{align}
Such a design was mentioned in \citet[Eq. (1.1)]{Bather1981jrssb} and called ``play the favourite''. \footnote{In the machine learning literature since the
2000s, this design is also known as ``follow the leader'' (as a simpler variant of ``follow the perturbed leader''). The term has however been used also
for other designs: \citet{SobelWeiss1972} are the first to use the term ``follow the leader'', but they mean a variant of ``stay-with-a-winner \&
switch-on-a-loser'', with a difference of randomizing (rather than switching) when a failure is observed and the number of failures on both arms is equal.
Many papers studying continuous bandit problems use the term ``follow the leader'' for (Gittins) index rule.}
Note that \design{BM} and \design{FM} are equivalent if and only if $ \widetilde{ \successes }_{ C } ( 0 ) = \widetilde{ \failures }_{ C } ( 0 ) =
\widetilde{ \successes }_{ D } ( 0 ) = \widetilde{ \failures }_{ D } ( 0 ) = 0 $.
\citet[Section 3]{Berry1978jasa} established that, in the setting with known two-point set $ \{ \theta_{ C } , \theta_{ D } \} $, in which the \design{BM}
design is optimal, at any epoch $ t $ at which the posterior probability that arm $ C $ is better than arm $ D $ is $ 1 / 2 $, it is optimal to allocate to
the arm with currently the greatest difference between the number of successes and the number of failures. Several tie-breaking rules can be used but they
only affect the performance marginally, so these are not discussed here in detail. We refer to this as the \emph{Bayesian greatest difference first}
(\design{BGDF}) design. The Bayesian performance of this design was illustrated numerically to be near optimal \citep{Berry1978jasa} and to outperform the
\design{BM} design \citep{Villar2018peis}. The mentioned condition applies when the one-arm prior distributions satisfy $ \widetilde{ s }_{ C } ( 0 ) =
\widetilde{ f }_{ C } ( 0 ) = \widetilde{ s }_{ D } ( 0 ) = \widetilde{ f }_{ D } ( 0 ) $, as in this paper, and note that in that case the frequentist
analogue, \design{FGDF}, yields an equivalent design.
\subsubsection{Utopic Index Rules}
\label{section:utopic}
In the frequentist setting, \citet{Robbins1952} introduced the deterministic ``stay-with-a-winner \& switch-on-a-loser'' design, in which the first
allocation is made randomly, and then the same arm is allocated whenever the observed response is a success, while the other arm is allocated whenever the
observed response is a failure. He realized that it leads to complete learning, but fails to achieve the maximum expected proportion of successes, because
it allocates the worse arm at a regular rate.
Interestingly, many of the deterministic designs actually have the ``stay-with-a-winner'' property. It is easy to see that the \design{BM}, \design{FM},
\design{BGDF} and \design{FGDF} all have it. In the Bayesian setting, \citet[Theorem 6.2]{Berry1972} proved the ``stay-with-a-winner'' property for the
deterministic \design{DP} design. Formally, the ``stay-with-a-winner'' property in this approach is: if arm $ k $ is optimal in period $ t \le T - 2 $, the
subject is allocated to it and its response is a success, then arm $ k $ is uniquely optimal in period $ t + 1 $. Contrary to the above, the
``switch-on-a-loser'' property is not, in general, satisfied by these designs. When one arm looks significantly better than the other one, the designs
would not switch after observing a failure on the former arm. See, e.g., \citet[p. 79]{BerryFristedt1985}.
\citet[Conjecture A, p. 892]{Berry1972} conjectured that for a large number of remaining allocations (i.e. at epochs $ t \ll T $), the only criterion for
Bayesian optimality in the \design{DP} design is the difference between the posterior number of failures on the two arms. (He did not specify what to do if
there is the same number of failures on the two arms.)
\citet{Kelly1981} studied the Gittins index rule in the infinite-horizon setting (in which it is optimal) as the discount factor approaches one (i.e. the
undiscounted setting) and established, under a technical condition on the prior distributions, that the Gittins index rule reduces to the following rule
which we call the \emph{frequenist least failures first} (\design{FLFF}) design: at every time epoch, allocate the subject to the arm with least observed
number of failures, breaking the ties in favor of any arm with greatest observed number of successes (breaking the double ties arbitrarily).
Although the similarity to Berry's conjecture is remarkable, \citet[Remark 4.15]{Kelly1981} stated that there seems to be no immediate link between these
two statements, and that Berry's conjecture may require a technical condition on the prior distributions. \citet[Remark 4.7]{Kelly1981} further elucidated
that \design{FLFF} is a slight variation of the ``stay-with-a-winner \& switch-on-a-loser'' design in \citet{Robbins1952}. Indeed, it is easy to see that
\design{FLFF} has the ``stay-with-a-winner'' property. After observing a failure, it switches if the other arm has lower number of failures, but it stays
or switches accordingly to the greater number of successes if the number of failures on both arms is the same.
Analogously, we call the \emph{Bayesian least failures first} (\design{BLFF}) design the one based on posterior numbers of successes and failures rather
than on the observed ones. If the prior distributions on the two arms are the same, then this design is equivalent to the \design{FLFF} design, but if they
are not, then the arm with lower number of posterior failures will be allocated without switching (at least) until the number of failures on the two arms
becomes the same.
We would like to highlight that the three radically different ``schools'' of the bandit problem, namely Robbins', Berry's and Gittins', all led to the
\design{FLFF} design with only slight variations. The ``stay-with-a-winner \& switch-on-a-loser'' design may be useful in some practical situations because
it only depends on the last observation, not on the numbers of successes and failures over the whole history.
\subsection{Reinforcement Learning --- UCB Index Rules}
\citet{LaiRobbins1985aam} laid out the theory of asymptotically optimal allocation and were the first to actually use the term \emph{``upper confidence
bound''} (UCB). They wrote it in quotes, as the quantity it refers to depends on time period $ t $, and is thus not the conventional upper bound of
confidence intervals, but can be interpreted as the upper confidence bound with significance level $ 1 / t $. \citet[Eq. (2.6)]{Lai1987as} introduced a UCB
index rule using the Kullback-Leiber divergence and \citet[Example 5.7]{Agrawal1995aap} developed UCB index values which are inflations of the observed
mean proportion of successes. The theory was further extended in \citet{BurnetasKatehakis1996} to multivariate and non-parametric distributions. The first
simple UCB index rules with finite-time theoretical guaranties were developed in \citet{AuerEtal2002}. See, e.g.,
\citet{BubeckCesa-Bianchi2012ftml,KaufmannGarivier2017esaim,LattimoreSzepesvari2019book} for accounts of subsequent developments.
The use of the observed mean proportion of successes in defining an index rule is attractive mainly because it is the maximum likelihood estimator of the
success probability in the static setting of the design of experiments. Index rules that use time-dependent inflations of the observed mean proportion of
successes were proposed and investigated even before \citet{LaiRobbins1985aam}, e.g., \citet{Bather1980aap, Bather1981jrssb, AbdelHamid1981thesis} in
frequentist setting and \citet{GittinsJones1974, Glazebrook1980jrssb, GittinsWang1992aos} in Bayesian setting.
Following \citet[Section 2]{BubeckCesa-Bianchi2012ftml}, we consider the popular \design{$ \alpha $UCB} design which allocates each arm once in the first
two periods, and then deterministically allocates every subject to the arm with currently the largest index (breaking ties randomly) of the form
\begin{align}
\frac{ \successes_{ k } ( t ) }{ \successes_{ k } ( t ) + \failures_{ k } ( t ) } + \sqrt{ \frac{ \alpha \cdot \ln( t + 1 ) }{ \successes_{ k } ( t ) + \failures_{ k } ( t ) } }
\end{align}
where $ \alpha > 0 $. The original design introduced in \citet{AuerEtal2002} used $ \alpha = 2 $. Theoretical upper bounds currently exist for $ \alpha > 1
$, but researchers have noticed empirically that lower values of $ \alpha $ typically lead to better performance and some used $ \alpha = 1 $, see, e.g.,
\citet{CsernaEtal2017uai}. In our numerical experiments (not reported here) we found that approximately the best performance is achieved with $ \alpha =
0.18 $.
Many other types of UCB designs have been developed, see e.g., \citet[Figure 2]{Kaufmann2018aos} for comparison of some of them. Besides designs based on
idea of UCB, there is a number of popular designs with randomized actions, e.g., epsilon-greedy, Boltzmann exploration, Thompson sampling, etc.
\subsection{Biostatistics}
Blinding of patients and personnel to the allocated treatment is an important desideratum in many types of clinical trials to mitigate a variety of biases
such as performance bias, detection bias and attrition bias \citep[Section 8.4]{Cochrane2011}. If allocation is deterministic, the patients and personnel
can, if they know the state of the trial, identify the allocated treatment with certainty or high probability. Thus, some amount of randomness in the
allocation decision is desirable. Moreover, the theory of design of experiments explains that randomization is important for mitigation of the selection
bias, for ensuring similarity in the treatment groups and for providing a basis for inference, which are essential for making valid conclusions at the end
of a trial \citep{RosenbergerEtal2019sim}. Thus, in clinical trials theory and practice, a lot of attention is paid to designs with randomized actions
\citep{RosenbergerLachin2015book}, although it should also be noted that not all types of clinical trials require or allow to implement blinding and
randomization. There are three major approaches to adaptive randomization: (i) \emph{Bayesian Response-Adaptive Randomization}, starting with
\citet{Thompson1933biometrika}, generalized in \citet{ThallWathen2007ejoc} (see also \citet[p. 156]{BerryEtal2011book}), and recently studied also in the
reinforcement learning literature (see, e.g., \citet{AgrawalGoyal2012colt}); (ii) \emph{Frequentist P\'{o}lya Urn Randomization}, which is a randomized
version of the ``stay-with-a-winner \& switch-on-a-loser'' design, starting with \citet{WeiDurham1978jasa}; and (iii) \emph{Randomized Index Rules} such as
those proposed by \citet{Bather1980aap, Bather1981jrssb, AbdelHamid1981thesis} in frequentist setting and by \citet{Glazebrook1980jrssb} in Bayesian
setting.
Note that the term ``bandit'' usually does not appear in the relevant biostatistics literature.
\subsection{Combination Designs}
Many researchers have suggested to use one design for the first $ N $ subjects and use another one for the remaining $ ( T - N ) $ subjects. This idea
appeared implicitly in early papers on sequential design of experiments, considering the \design{1:1} design initially (on a sample of size $ N $), and
then sticking to the arm with higher mean (concluded at a given significance level) forever, since that approach assumes $ T = +\infty $.
\citet{ChengEtal2003biometrika} found that when $ T $ is finite, for this combination design it is optimal for $ N $ to be of the order of $ \sqrt{ T } $,
and, in the particular case of prior Beta distribution with parameters $ ( 1 , 1 ) $ on each arm, the optimal $ N \to \sqrt{ 2 T } $ asymptotically as $ T
\to +\infty $ (and is slightly lower for finite horizons).
\citet{HoelEtal1972biometrika} considered a setting in which $ T $ is not fixed in advance, but the allocation is stopped if a predefined difference in the
number of successes is reached. They proposed to start with the \emph{arm alternating} design (i.e., a deterministic version of \design{1:1}), and then to
switch to the ``stay-with-a-winner \& switch-on-a-loser'' design if the estimate of $ \max \{ \theta_{ C } , \theta_{ D } \}$ is greater than $ 0.6 $ and
continue with arm alternating otherwise. This paper is just one example from the work in the area of \emph{ranking and selection}.
\citet[Section 4]{Zelen1969jasa} proposed to initially use the ``stay-with-a-winner \& switch-on-a-loser'' design, and then to stick to the arm with higher
mean. He found that this is better than the classic combination if $ \theta_{ C } + \theta_{ D } \ge 1 $ and the value of $ N \approx T / 3 $ yields a
nearly-optimal number of successes.
\citet[Remark 4.10]{Kelly1981} elucidated that the optimal Bayesian decision-theoretical design in the discounted setting can be regarded as having three
stages: first, \design{BLFF}, which he refers to as ``information gathering'', then it moves further and further away from that design, and finally it
always allocates to the same arm. \citet[Section 3.4]{Villar2018peis} provided some illustrative numerical evaluation of these stages for the Gittins and
Whittle index rules.
Motivated by the above and by the discussion in \autoref{section:utopic}, in this paper we consider three combination designs in which the designs used are
\design{BLFF} in the first stage, followed by \design{BM}, \design{$ \alpha $UCB} and \design{BMSF}, respectively, and one combination rule in which the
designs used are \design{1:1} in the first stage, followed by \design{BMSF}:
\begin{itemize}
\item \design{BLFF+BM} with $ N = \sqrt{ 4 T } $
\item \design{BLFF+$ \alpha $UCB} with $ N = \sqrt{ 4 T } $
\item \design{BLFF+BMSF} with $ N = \sqrt{ 9 T } $
\item \design{1:1+BMSF} with $ N = \sqrt{ 2 T } $
\end{itemize}
The lengths of the first stage using \design{BLFF} have been obtained heuristically based only on a small numerical study and so they are not necessarily
overall optimal, and surely not optimal for each particular scenario. To the best of our knowledge, these three designs have not been studied previously
and their theoretical analysis is an open problem. Note that using \design{BMSF} in the second stage represents the situation in which the arm with higher
mean (most successes is approximately equivalent to highest mean since \design{BLFF} keeps the number of failures balanced up to a difference of $ 1 $)
after the first stage is allocated to all the subjects in the second stage since only the allocated arm can collect additional successes. This is inspired
by but not fully equivalent to the design studied by \citet[Section 4]{Zelen1969jasa}, because \design{BLFF} is not fully equivalent to
``stay-with-a-winner \& switch-on-a-loser''.
Using \design{BM} in the second stage is similar to \design{BMSF}, but allows for additional learning during the second stage: the mean of the arm with
higher mean at the end of the first stage may eventually decrease below the mean of the other arm, at which moment the allocation switches to that arm
(there may be more than one such ``correction''). The combination design that uses \design{$ \alpha $UCB} in the second stage is included because it has
interesting performance, to be discussed in \autoref{section:performance}.
Design \design{1:1+BMSF} is inspired by \citet{ChengEtal2003biometrika}, although using \design{BMSF} in the second stage is not fully equivalent to
sticking to the arm with higher mean at the end of the first stage, because \design{1:1} may lead to unbalanced allocation due to sampling variability.
\section{Designs Performance}
\label{section:performance}
For some of the response-adaptive designs, theoretical values of or bounds on their performance exist. However, these are either asymptotic as $ T \to
+\infty $, or up to an additive and/or multiplicative constants, which are often large or unknown. For small ($ T \approx 1 : 10^{ 2 } $) and moderate ($ T
\approx 10^{ 2 } : 10^{ 4 } $) horizons, computational evaluation is the most appropriate to illustrate designs performance.
In this section we report computational experiments in which every design is evaluated by backward recursion, i.e. at full computational accuracy (of
Float64, which is of the order of $ 10^{ \log( T ) - 16 } $). For fairness, we only include deterministic designs and \design{1:1} as a benchmark (i.e. we
exclude the biostatistics ones).
To the best of our knowledge, such accurate evaluation has only been reported for small horizons in the existing literature. For moderate horizons, which
are often the most relevant in practice, for instance in clinical trials, designs are usually evaluated by simulation, which is remarkably less accurate.
We present the performance in terms of both the proportion of successes and (equivalently) the regret number of successes, as these provide complementary
insights. See \autoref{section:performance_continued} for small horizons, in which the performance of some of the designs is \emph{fundamentally
different}.
\subsection{Bayesian Performance --- Moderate Trial Sizes}
\autoref{fig:2} illustrates the Bayesian performance of the Bayesian decision-theoretic design for $ T = 120 : 120 : 4440 $. Although it is an interesting
way of summarizing the performance, it has three major drawbacks: (i) it is a simple average over all the possible pairs of parameters $ ( \theta_{ C } ,
\theta_{ D } ) $, while in practice one would be interested in a particular subset of the whole parameter space, possibly weighted in a particular way;
(ii) it presents the Bayesian value in which the observations happen according to the belief at every time epoch rather than being fixed over the whole
horizon, which blurs its interpretation because the beliefs are biased and which is less desirable since in practice one would be interested in the value
under the true (unknown) success probabilities; (iii) it depends on the prior distributions, so a choice needs to be made or it needs to be computed for a
number of different options.
Nevertheless, \autoref{fig:2} is instrumental in giving an idea about the order of magnitude of the objective, and in particular it is interesting to
observe that the regret number of successes is increasing and concave, taking value of around $ 5 $ for the trial size of $ 3000 $ and, by extrapolation,
it is likely to be below one per mille of the trial sizes beyond $ 10^{ 4 } $.
\begin{figure}[tbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/proportions_action_lin_32-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/regret_action_lin_32-pure.pdf}
\end{subfigure}%
\caption{An illustration of two (equivalent) Bayesian performance measures evaluated for the deterministic Bayesian decision-theoretic design over a range of moderate trial sizes.}\label{fig:2}
\end{figure}
\subsection{Frequentist Performance --- Moderate Trial Sizes}
In this subsection we evaluate and compare the frequentist performance of the above designs for for $ T = 120 : 60 : 1200 \cup 1200 : 300 : 2400 $. We do
so in four scenarios: $ ( \theta_{ C } , \theta_{ D } ) \in \{ ( 0.1 , 0.3 ) , ( 0.3 , 0.5 ) , ( 0.5 , 0.7 ) , ( 0.7 , 0.9 ) \} $. These scenarios are a
natural choice, and have been previously considered, e.g., $ ( 0.3 , 0.5 ) $ in \citet[Table 3]{Lai1987as}, \citet[Table 2]{BrezziLai2002jedc},
\citet[Table 5]{VillarEtal2015survey} and \citet[Table 6]{Villar2018peis}, $ ( 0.7 , 0.9 ) $ in \citet[Table 3]{Lai1987as}, and all of them in
\citet{HoelEtal1972biometrika}. Although conclusions from a number scenarios do not guarantee their validity in other scenarios, we have found these four
scenarios to be illustrative enough to provide negative conclusions for many designs.
We report both the proportion of successes, in \autoref{fig:PS_2400}, and the regret number of successes, in \autoref{fig:regret_2400}. Besides the mean,
for each measure we report also the standard deviation, which can be considered as a secondary criterion providing additional insights into the performance
of the designs.
\design{1:1} is the worst design in terms of the mean and often the best in terms of the SD but, surprisingly, it is not always the best, meaning that in
some scenarios there are designs that are better under both measures (all of these are combination designs, to be discussed below). \design{BLFF} notably
improves the mean, increasingly so in the scenarios with higher success probabilities, while only marginally deteriorates the SD. They both over-explore.
The curves of \design{BMSF}, \design{FM}, \design{BM}, \design{BGDF} and \design{BKG} look approximately linear in both the mean and SD of the regret
number of successes (approximately constant in both the mean and SD of the proportion of successes). The performance of \design{BGDF} and \design{BKG} is
quite bad in general and heavily scenario dependent. \design{BGDF}'s mean deteriorates with more extreme success probabilities, but its SD improves with
lower success probabilities. \design{BKG}'s mean and SD both improve with lower success probabilities, remarkably so in scenario $ ( 0.1 , 0.3 ) $, in
which the mean regret is the best of all designs and seems to be constant, while all the other designs' mean regret is increasing.
There are three designs that perform particularly badly in all four scenarios: \design{BMSF}, \design{FM} and \design{BM} are dominated by almost all the
other designs. High SD indicates that these three designs are extremely \emph{under-exploring}: too aggressively sticking to one of the arms, and
relatively often choosing the worse one. Among these three, \design{BMSF} is worse than the other two in both the mean and SD in all four scenarios. The
performance order of \design{FM} and \design{BM} depends on scenario; \design{BM}'s mean remarkably improves with lower success probabilities, while
\design{FM}'s mean improves with more extreme success probabilities. For these three designs it also holds that the better the mean, the better the SD, so
in every scenario the order is identical under both measures.
The curves for the mean of all the remaining designs (\design{DP}, \design{$ \alpha $UCB}, and the combination designs) look approximately concave, leading
to significant improvement in the mean. Quantitatively, all three versions of \design{$ \alpha $UCB} are almost identical across the four scenarios,
indicating that its performance depends on the difference $ \theta_{ C } - \theta_{ D } $ rather than on their respective values. Interestingly, the mean
regret number of successes of \design{2UCB} is more than $ 50\% $ higher than that of \design{1UCB}, which is in turn around $ 3 $ times higher than that
of \design{0.18UCB}, which is in turn still distinctively higher (between $ 20\% - 100\% $, depending on scenario) than that of \design{DP} which is the
best performing design except when dominated by \design{BKG} in scenario $ ( 0.1 , 0.3 ) $. The mean regret of \design{DP} in scenario $ ( 0.7 , 0.9 ) $
seems to be constant, while all the other designs' mean regret is increasing. The SD of \design{2UCB} and \design{1UCB} is practically undistinguishable,
while the SD of \design{0.18UCB} is notably higher in the scenarios with higher success probabilities, with the SD of \design{DP} being in between.
The combination designs bring some surprising results. First, \design{BLFF+2UCB} is practically identical to \design{2UCB} and \design{BLFF+1UCB} to
\design{1UCB} in all the scenarios and trial sizes, indicating that \design{$ \alpha $UCB} initially behaves essentially as \design{BLFF}. That is however
not true for \design{BLFF+0.18UCB}, which is similar but not identical to \design{0.18UCB}: in terms of the mean being better in scenarios with higher
success probabilities while being worse in those with lower success probabilities. The SD of \design{BLFF+0.18UCB} is always lower than that of
\design{0.18UCB}, and in particular it is the lowest of all designs except when dominated by \design{1:1} and \design{BLFF} in scenarios with low success
probabilities.
Somewhat surprisingly, \design{BLFF+BM} performs quite well, roughly similarly to \design{BLFF+0.18UCB} in terms of the mean, although its SD is sometimes
higher. \design{BLFF+BMSF} performs worse than \design{BLFF+BM} but still significantly better than \design{1UCB} in terms of the mean, but its SD is
notably higher. \design{1:1+BMSF} is in most instances significantly worse than \design{BLFF+BMSF} both in terms of the mean and SD.
We summarize the above discussion as follows:
\begin{itemize}
\item \design{DP} may be the preferred design in terms of the mean in scenarios with high and moderate success probabilities;
\item \design{BKG} may be the preferred design in terms of the mean in scenarios with low success probabilities;
\item \design{BLFF+0.18UCB} may be the overall preferred design in terms of both the mean (higher weight) and the SD (lower weight);
\item \design{BLFF} may be the overall preferred design in terms of both the mean (lower weight) and the SD (higher weight);
\item \design{1:1} may be the preferred design in terms of the SD in scenarios with low and moderate success probabilities;
\item \design{BLFF+0.18UCB} may be the preferred design in terms of the SD in scenarios with high success probabilities;
\end{itemize}
\begin{figure}[tbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_mean_70_90_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_SD_70_90_2400-pure.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_mean_50_70_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_SD_50_70_2400-pure.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_mean_30_50_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_SD_30_50_2400-pure.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_mean_10_30_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/PS_SD_10_30_2400-pure.pdf}
\end{subfigure}%
\caption{An illustration of performance (mean on the left, standard deviation on the right) in terms of the expected proportion of successes evaluated for
deterministic designs over a range of moderate trial sizes, for $ ( \theta_{ C } , \theta_{ D } ) = ( 0.7 , 0.9 ) $ in the first row, $ ( 0.5 , 0.7 ) $ in
the second row, $ ( 0.3 , 0.5 ) $ in the third row, $ ( 0.1 , 0.3 ) $ in the fourth row.}\label{fig:PS_2400}
\end{figure}
\begin{figure}[tbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_mean_70_90_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_SD_70_90_2400-pure.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_mean_50_70_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_SD_50_70_2400-pure.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_mean_30_50_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_SD_30_50_2400-pure.pdf}
\end{subfigure}%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_mean_10_30_2400-pure.pdf}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[trim=0pt 0pt 0pt 0pt, clip=true, width=\textwidth]{pic/NS_regret_SD_10_30_2400-pure.pdf}
\end{subfigure}%
\caption{An illustration of performance (mean on the left, standard deviation on the right) in terms of the regret number of successes evaluated for
deterministic designs over a range of moderate trial sizes, for $ ( \theta_{ C } , \theta_{ D } ) = ( 0.7 , 0.9 ) $ in the first row, $ ( 0.5 , 0.7 ) $ in
the second row, $ ( 0.3 , 0.5 ) $ in the third row, $ ( 0.1 , 0.3 ) $ in the fourth row.}\label{fig:regret_2400}
\end{figure}
\section{Myths}
\label{section:myths}
The hardness of the problem and the diversity of approaches mean that it is difficult for a single person to have complete information and full
understanding of the many details. This paper is, to the best of our knowledge, first attempt to survey key ideas from multiple disciplines. Author's
engagement with researchers and practitioners from a number of fields has given him an impression of common beliefs that may not be fully correct, listed
next.
\subsection{Myth \#1: The Bayesian Decision-Theoretic Design is Intractable}
This myth seems to be widespread across disciplines, e.g., ``The curse of dimensionality renders exact approaches impractical.''
\citep{AhujaBirge2019report}; ``Even in relatively benign setups the computation of the Bayesian optimal policy appears hopelessly intractable.''
\citep[Section 35.1]{LattimoreSzepesvari2019book}; ``Trials of a size up to 3500 patients would be feasible with today’s number \#1 supercomputer (with 1.3
PB of RAM).'' \citep{WilliamsonEtal2017csda}.
\autoref{table:DP_horizons} illustrates the evolution of what was reported as computationally tractable in the literature. Indeed, besides the two results
in the 1960s that pre-date the era of personal computers, the existing literature gives an impression that solving the problem beyond $ T = 100 $ is
impractical or unfeasible. A closer look however reveals that no improvement seem to have been achieved since \citet{Berry1978jasa} which appeared more
than 40 years ago, despite the theoretical progress in computer science and technological progress in personal computers. In this paper we use a
state-of-the-art package in Julia programming language written by the author, and show that on a standard laptop or desktop computer, the \design{DP}
design can be computed for offline use or evaluated in online fashion in a few minutes ($ T \approx 1,000 $), in a few hours ($ T \approx 2,000 $), or in a
few days ($ T \approx 4,000 $). 32GB of RAM allows storing (e.g. for offline use) of the whole design up to trial size around $ 1440 $; when its storing
is not needed (e.g. for Bayesian evaluation or for calculation of the initial action) it allows up to trial size around $ 4440 $. \footnote{The first
version of the package is planned to be released to public in mid-2019. The package will be described in a separate paper, whose abstract has been accepted
by the editors of the special issue on Bayesian Statistics of the Journal of Statistical Software, and will be submitted by its deadline 30 June 2019.}
\begin{table}[tbp]
\centering
\begin{tabular}{llll}
\toprule
Publication & $ T $ & $ T^{ \max } $ & SW, HW, RAM \\
\midrule
\citet{Steck1964moc} & $ 25 $ & N/A & N/A, UNIVAC 1105, 54 kB \\
\citet{Yakowitz1969book} & $ 5 $ & N/A & Fortran, N/A, N/A \\
\citet{Berry1978jasa} & $ 100 $ & N/A & Basic (?), Atari (?), N/A \\
\citet{GinebraClayton1999jspi} & $ 150 $ & $ 180 $ & N/A, N/A, N/A \\
\citet{HardwickEtal2006} & $ 100 $ & $ 200 $ & N/A, N/A, N/A \\
\citet{AhujaBirge2016ejor} & $ 96 $ & $ 240 $ & N/A, Mac 4GB \\
\citet{WilliamsonEtal2017csda} & $ 100 $ & $ 215 $ & R, PC, 16GB \\
\citet{Villar2018peis} & $ 100 $ & N/A & Matlab, PC, N/A \\
\citet{Kaufmann2018aos} & $ 70 $ & N/A & N/A, N/A, N/A \\
This paper & $ 4440 $ & $ 4440 $ & Julia 1.0.1 \& BB, PC, 32GB \\
\bottomrule
\end{tabular}
\caption{Horizons $ T $ with reported results and $ T^{ \max } $ reported as the largest computationally tractable on a standard computer by backward recursion in the literature on two-armed problem with binary responses.}\label{table:DP_horizons}
\end{table}
\subsection{Myth \#2: The Bayesian Decision-Theoretic Design is Optimal}
The \design{DP} is Bayes-optimal, but that does not necessarily mean that it is optimal in the non-Bayesian objective, i.e. that it achieves the largest
possible mean number of successes. \autoref{fig:regret_2400} and \autoref{fig:regret_240} clearly show that other designs perform better in some scenarios.
\subsection{Myth \#3: The Gittins Index Rule Leads to Incomplete Learning}
According to \citet{BrezziLai2000econometrika}: ``...we give in Section 3 a simple proof of the incompleteness of optimal learning from endogenous data in
the discounted multi-armed bandit problem... This generalizes \citet{Rothschild1974jet}'s result for Bernoulli two-armed bandits, and also the result of
\citet{BanksSundaram1992econometrika} who show that there is positive probability of incomplete learning in multi-armed bandits with general distributions
of rewards if the priors have finite support.'' This is however true only in the discounted setting; \citet{Kelly1981} proved that the structure of the
Gittins index rule in the undiscounted setting (see \autoref{section:utopic}) leads to complete learning.
\subsection{Myth \#4: The Gittins Index is Computationally Simpler than Dynamic Programming for the Two-Armed Problem}
As we explain in \autoref{section:Gittins}, there is a trade-off between accuracy and computational complexity of the Gittins index. For computing the
Gittins index, all the algorithms use a truncation of the horizon, and due to numerical instability most of the numerical algorithms use a discount factor
strictly lower than one although the appropriate one would be equal to one. So, there are three levels of approximation that the Gittins index requires,
and such an approximate Gittins index rule may in the end not even perform better than the Myopic index rule (see e.g. \citet{AhujaBirge2016ejor}).
\subsection{Myth \#5: The UCB Index Rule is Near-Optimal for the Finite-Horizon Problem}
\autoref{section:performance} illustrates that the \design{2UCB} and \design{1UCB} designs are significantly suboptimal, yielding five to tenfold mean
regret number of successes than the optimal design. Tuning the coefficient even beyond the intervals required by theoretical analysis can significantly
improve the performance, e.g. \design{0.18UCB} yields around twofold mean regret number of successes than the optimal design, which is still too large to
be considered near-optimal.
\subsection{Myth \#6: The Frequentist Mean Regret Number of Successes is Increasing}
\citet{LaiRobbins1985aam} presented a lower bound, valid under certain technical assumptions, indicating that the frequentist mean regret number of
successes increases proportionally to $ \log T $ as $ T \to \infty $. Our computational results indicate that this is not necessarily true for all designs
over finite horizons. For instance, the mean regret of \design{DP} in scenario $ \theta_{ C } = 0.7 , \theta_{ D } = 0.9 $ (see
\autoref{table:performance79}) at $ T = 240 $ is lower than that at $ T = 300 $, and it is non-increasing over $ T = 660 : 60 : 1200 $.
\section{Conclusion}
\label{section:conclusion}
The aim of this paper to provide a unified account of approaches from various disciplines to the two-armed bandit problem. We have proposed problem
terminology that we believe should not create conflicts with other existing terminology in most of the disciplines, to facilitate mutual learning. We have
proposed to use backward recursion instead of simulation for more accurate evaluation. We have created a computational comparison of designs from different
disciplines in a standardized set of scenarios. The computational results have showed that some of the simple ones (e.g. \design{BLFF+BM} and
\design{BLFF+BMSF}) perform surprisingly well in our scenarios and outperform may of the more sophisticated and more studied ones. We have also showed that
\design{DP} is tractable for much larger horizons than it is commonly believed. This suggests that there is a case for these to be used among benchmark
designs when developing new designs.
We have given an account of approaches to the problem with the objective of maximizing the mean number of successes, which is linear across arms and over
time. Most of the above designs, especially those horizon-ignorant, crucially depend on that property, and it is not clear how they could be modified if
the objective changed to another one. The only exception is the \design{DP} and its variants, which are quite flexible to accommodate other finite-horizon
objectives.
A significant area of research left out of this paper deals with practicalities of implementation of the designs, especially in the context of clinical
trials, in which (i) the objective is different, because it focusses much more on estimation of the success probabilities, and (ii) there are additional
constraints, e.g. requirement of a certain degree of randomization. A fair comparison of such designs would actually require a series of comparisons fixing
the randomization degree and including only those designs that satisfy it. Such work is extensive and is left for a separate paper.
|
\section{Introduction}\label{sec1}
Consider the Cauchy problem for the Kawahara equation
\begin{equation}\label{Kawahara}
\begin{cases}
u_t + uu_x + \alpha u_{xxx} + \beta u_{xxxxx} = 0,\\
u(0, x) = u_0(x),
\end{cases}
\end{equation}
where $u:\mathbb{R}^{1+1}\rightarrow\mathbb{R}$ and $\alpha$, $\beta$ are real constants with $\beta \neq 0$.
This fifth-order KdV type equation has been derived to model gravity-capillary waves on a shallow layer
and magneto-sound propagation in plasmas (\cite{Ka,KS}).
The well-posedness of the above Cauchy problem with initial data in Sobolev spaces $H^s$ has been studied by several authors
(see e.g. \cite{CDT,WCD,CLMW,CG}).
In particular, it was shown in \cite{CDT} that \eqref{Kawahara} has a global solution when $s=0$.
This was improved by Wang, Cui and Deng \cite{WCD} to $s>-1/2$
and then by Chen, Li, Miao and Wu \cite{CLMW} to $s>-7/4$.
More recently, Chen and Guo \cite{CG} obtained the global well-posedness for $s\geq-7/4$.
In this paper, we are concerned with the persistence of spatial analyticity for the solutions of \eqref{Kawahara},
given initial data in a class of analyticity functions.
While the well-posedness theory in Sobolev spaces is well developed, nothing is known about the spatial analyticity for the Kawahara equation.
From now on, we focus on the situation where we consider a real-analytic initial data
with uniform radius of analyticity $\sigma_0>0$,
so there is a holomorphic extension to a complex strip
$$S_{\sigma_0}=\{x+iy:x,y\in\mathbb{R},\,|y|<\sigma_0\}.$$
Now, it is natural to ask whether this property may be continued analytically to a complex strip $S_{\sigma(t)}$ for all later times $t$,
but with a possibly smaller and shrinking radius of analyticity $\sigma(t)>0$.
This type of problem was first introduced by Kato and Masuda \cite{KM},
and has recently received a lot of attention for the Korteweg-de Vries (KdV) equation.
It was shown by Bona and Gruji\'{c} \cite{BG} that
the radius $\sigma(t)$ for the KdV equation can decay no faster than $e^{-t^2}$ as $|t|\rightarrow\infty$.
This was improved greatly by Bona, Gruji\'{c} and Kalisch \cite{BGK} to a polynomial decay rate of $|t|^{-12}$.
Later, Selberg and da Silva \cite{SS} obtained a further refinement, $\sigma(t)\geq c|t|^{-4/3+\varepsilon}$,
where the $\varepsilon$ exponent was also removed by Tesfahun \cite{Te2}. The rate was again improved by Huang and Wang \cite{HW} up to $|t|^{-1/4}$.
See also a recent related result for the quartic generalised KdV equation by Selberg and Tesfahun \cite{ST2}.
In spite of these many works for KdV equations, there have been no results on this issue for the Kawahara equation \eqref{Kawahara}
which is a fifth-order KdV type equation.
Motivated by this, we aim here to obtain the spatial analyticity for the Kawahara equation.
The Gevrey space, denoted $G^{\sigma,s}(\mathbb{R})$, $\sigma\geq0$, $s\in\mathbb{R}$, is a suitable function space to
study analyticity of solution. In our case, it will be used with the norm
$$\|f\|_{G^{\sigma,s}}=\| e^{\sigma|D|} \langle D\rangle^s f\|_{L^2},$$
where $\langle D\rangle=1+|D|$ and $D$ denotes the derivative.
According to the Paley-Wiener theorem\footnote{The proof given for $s=0$ in \cite{K} applies also for $s\in\mathbb{R}$
with some obvious modifications.} (see e.g. \cite{K}, p. 209),
a function $f$ belongs to $G^{\sigma,s}$ with $\sigma>0$
if and only if it is the restriction to the real line of a function $F$ which is holomorphic in the strip
$S_\sigma=\{ x+iy:x,y\in\mathbb{R},\,|y| < \sigma \}$ and satisfies
$\sup_{|y| < \sigma} \| F(x+iy)\|_{H_x^s} <\infty$.
Therefore every function in $G^{\sigma,s}$ with $\sigma>0$ has an analytic extension to the strip $S_\sigma$.
Based on this property of the Gevrey space, which is the key to studying spatial analyticity of solution,
our result below gives a lower bound $|t|^{-1}$ on the radius of analyticity $\sigma(t)$ of the solution to \eqref{Kawahara}
as the time $t$ tends to infinity.
\begin{thm}\label{thm1}
Let $u$ be the global $C^\infty$ solution of \eqref{Kawahara} with $u_0 \in G^{\sigma_0,s}(\mathbb{R})$ for some $\sigma_0>0$ and $s\in\mathbb{R}$.
Then, for all $t\in\mathbb{R}$
$$u(t)\in G^{\sigma(t),s}(\mathbb{R})$$
with $\sigma(t)\geq c|t|^{-1}$ as $|t| \rightarrow \infty$.
Here, $c>0$ is a constant depending on $\|u_0\|_{G^{\sigma_0,s}(\mathbb{R})}$.
\end{thm}
It should be noted that the existence of the global $C^\infty$ solution in the theorem is always guaranteed.
Indeed, observe that ${G^{0,s}}$ coincides with the Sobolev space $H^s$ and the embeddings
\begin{equation}\label{emb}
G^{\sigma,s}\subset G^{\sigma^\prime,s^\prime}
\end{equation}
hold for all $0\leq\sigma'<\sigma$ and $s,s'\in\mathbb{R}$.
As a consequence of this embedding with $\sigma'=0$ and the existing global well-posedness theory in $H^{s'}(=G^{0,s'})$ for $s'\geq-7/4$,
the Cauchy problem \eqref{Kawahara} has a unique smooth solution for all time, given initial data $u_0\in G^{\sigma_0,s}$
for any $\sigma_0>0$ and $s\in\mathbb{R}$.
We close this section with further references on the spatial analyticity for other dispersive equations such as
Schr\"odinger equations \cite{BGK2,Te,AKS}, Klein-Gordon equations \cite{P} and Dirac-Klein-Gordon equations \cite{ST,S}.
The outline of this paper is as follows:
In Section \ref{sec2} we introduce some function spaces such as Bourgain and Gevrey-Bourgain spaces, and their basic properties
which will be used in later sections.
In Section \ref{sec3} we present a bilinear estimate (Lemma \ref{bilinear}) in Gevrey-Bourgain spaces.
By making use of a contraction argument involving this estimate,
we prove that in a short time interval $0\leq t \leq\delta$ with $\delta>0$ depending on the norm of the initial data,
the radius of analyticity remains strictly positive.
Next, we prove an approximate conservation law, although the conservation of $G^{\sigma_0,s}$-norm of the solution does not hold exactly,
in order to control the growth of the solution in the time interval $[0, \delta]$, measured in the data norm $G^{\sigma_0,s}$.
Section \ref{sec4} is concerned with the proofs of such a local result and the almost conservation law.
In Section \ref{sec5}, we finish the proof of Theorem \ref{thm1} by iterating the local result based on the conservation law.
The final section, Section \ref{sec6}, is devoted to the proof of Lemma \ref{bilinear}.
Throughout this paper, the letter $C$ stands for a positive constant which may be different
at each occurrence. We denote $A\lesssim B$ and $A\sim B$ to mean $A\leq CB$ and $B\lesssim A\lesssim B$, respectively.
We also use $A\ll B$ to mean $A\leq cB$ for some small constant $c>0$.
\section{Function spaces}\label{sec2}
In this section we introduce some function spaces and their basic properties
which will be used in later sections for the proof of Theorem \ref{thm1}.
For $s,b\in\mathbb{R}$, we use $X^{s,b}=X^{s,b}(\mathbb{R}^2)$ to denote the Bourgain space defined by the norm
$$
\|f\|_{X^{s,b}} =\| \langle\xi\rangle^s\langle\tau-p(\xi)\rangle^b\widehat{f}(\tau,\xi)\|_{L^2_{\tau,\xi}},
$$
where $p(\xi)=\alpha \xi^3 -\beta\xi^5$, $\langle\cdot\rangle=1+|\cdot|$ and $\widehat{f}$ denotes the space-time Fourier transform given by
$$
\widehat{f}(\tau,\xi)=\int_{\mathbb{R}^{2}} e^{-i(t\tau+x\xi)}f(t,x) \ dtdx.
$$
The restriction of the Bourgain space, denoted $X^{s,b}_\delta$, to a time slab $(0,\delta )\times\mathbb{R}$
is a Banach space when equipped with the norm
$$
\|f\|_{X^{s,b}_\delta}=\inf\big\{\|g\|_{X^{s,b}} : g=f\,\, \text{on}\,\, (0,\delta)\times\mathbb{R}\big\}.
$$
We also introduce the Gevrey-Bourgain space $X^{\sigma,s,b}=X^{\sigma,s,b}(\mathbb{R}^2)$ defined by the norm
$$
\|f\|_{X^{\sigma,s,b}}=\| e^{\sigma| \partial_x|}f\|_{X^{s,b}}.
$$
Its restriction $X^{\sigma,s,b}_\delta$ to a time slab $(0,\delta)\times\mathbb{R}$ is defined in a similar way as above,
and when $\sigma=0$ it coincides with the Bourgain space $X^{s,b}$.
The Gevrey-modification of the Bourgain spaces was used already by Bourgain \cite{B}
to study persistence of analyticity of solutions of the Kadomtsev-Petviashvili equation.
He proved that the radius of analyticity remains positive as long as the solution exists.
His argument is quite general and applies also to the KdV and Kawahara equations,
but it does not give any lower bound on the radius $\sigma(t)$ as $|t|\rightarrow\infty$.
We now list some basic properties of those spaces.
When $\sigma=0$, the proofs of the first two lemmas below can be found in Section 2.6 of \cite{T},
and the third lemma follows by the argument used for Lemma 3.1 of \cite{CKSTT2}.
But, by the substitution $f\rightarrow e^{\sigma|\partial_x|}f$, the properties of $X^{s,b}$ and its restrictions
carry over to $X^{\sigma,s,b}$.
\begin{lem}\label{lem00}
Let $\sigma \geq 0$, $s\in\mathbb{R}$ and $b>1/2$.
Then, $X^{\sigma,s,b}\subset C(\mathbb{R},G^{\sigma,s})$ and
$$\sup_{t\in\mathbb{R}}\|f(t)\|_{G^{\sigma,s}}\leq C\|f\|_{X^{\sigma,s,b}},$$
where $C>0$ is a constant depending only on $b$.
\end{lem}
\begin{lem}\label{lem2}
Let $\sigma \geq 0$, $s\in\mathbb{R}$, $-1/2<b<b^\prime<1/2$ and $\delta>0$. Then
$$
\|f\|_{X^{\sigma,s,b}_\delta} \leq C_{b,b'}\delta^{b^\prime-b}\|f\|_{X^{\sigma,s,b^\prime}_\delta},
$$
where $C_{b,b'}>0$ is a constant depending only on $b$ and $b'$.
\end{lem}
\begin{lem}\label{lem3}
Let $\sigma \geq 0$, $s\in\mathbb{R}$, $-1/2<b<1/2$ and $\delta>0$.
Then, for any time interval $I\subset[0,\delta]$,
$$
\|\chi_I f\|_{X^{\sigma,s,b}} \leq C\|f\|_{X^{\sigma,s,b}_\delta},
$$
where $\chi_I(t)$ is the characteristic function of $I$, and $C>0$ is a constant depending only on $b$.
\end{lem}
\section{Bilinear estimates in Gevrey-Bourgain spaces}\label{sec3}
In this section we present a bilinear estimate in Gevrey-Bourgain spaces, Lemma \ref{bilinear},
which plays a key role in obtaining the local well-posedness and almost conservation law in the next section.
With the aid of it, we shall also deduce an estimate, Lemma \ref{f}, which is another useful tool particularly
in obtaining the almost conservation law.
\begin{lem}\label{bilinear}
For all $\sigma\geq0$ and $s>-7/4$, there exist $1/2<b<1$ and $\varepsilon>0$ such that
\begin{equation}\label{comm}
\| \partial_x (uv)\|_{X^{\sigma,s,b'-1}}\leq C_{s,b,b'}\| u\|_{X^{\sigma,s,b}}\| v\|_{X^{\sigma,s,b}}
\end{equation}
for any $b'$ satisfying $b\leq b'<b+\varepsilon$.
Here, $C_{s,b,b'}>0$ is a constant depending only on $s$, $b$ and $b'$.
\end{lem}
It is worth comparing this lemma with the analogous result by Chen, Li, Miao and Wu \cite{CLMW} (\textit{cf}. Proposition 2.2),
where the proof for \eqref{comm} is given only in the case $b'=b$ to obtain local well-posedness in Sobolev spaces.
On the contrary, it is crucial for the present issue to have $b'$ that can range over a small interval for which \eqref{comm} holds.
Similarly as in \cite{CLMW}, we apply Tao's $[k;Z]$-multiplier norm method \cite{T2}
to our case where $b\leq b'<b+\varepsilon$.
We shall postpone the detailed proof of Lemma \ref{bilinear} until the last section, Section \ref{sec6}.
Instead here we derive the following lemma from Lemma \ref{bilinear},
which, along with the function $f$ defined here, plays a crucial role in obtaining the almost conservation law in the next section.
\begin{lem}\label{f}
Let
\begin{equation}\label{fff}
f(u)=\frac{1}{2}\partial_x\Big((e^{\sigma|\partial_x|}u)^2-e^{\sigma|\partial_x|}u^2\Big).
\end{equation}
Given $0\leq\rho\leq 1$, there exist $1/2<b<1$ and $C>0$ such that
$$
\| f(u)\|_{X^{0,b-1}} \leq C\sigma^\rho\| u\|^2_{X^{\sigma,0,b}}
$$
for all $\sigma>0$ and $v\in X^{\sigma,0,b}$.
\end{lem}
\begin{proof}
Notice first that
\begin{align*}
\widehat{v_1v_2}(\tau,\xi) &= \widehat{v_1}\ast\widehat{v_2}(\tau,\xi)\\
&= \int_{\mathbb{R}^2} \widehat{v_1}(\tau_1,\xi_1)\widehat{v_2}(\tau-\tau_1,\xi-\xi_1)\,d\tau_1d\xi_1\\
&=\int_{\mathbb{R}^2} \widehat{v_1}(\tau_1,\xi_1)\widehat{v_2}(\tau_2,\xi_2) \,d\tau_1d\xi_1
\end{align*}
with $\tau_2=\tau-\tau_1$ and $\xi_2=\xi-\xi_1$.
With this as well as the estimate\footnote{This estimate can be found in Lemma 12 of \cite{SS}.}
$$
e^{\sigma|\xi_1|}e^{\sigma|\xi_2|}-e^{\sigma|\xi_1+\xi_2|}\leq\big(2\sigma \min\{|\xi_1|,|\xi_2|\}\big)^\rho e^{\sigma|\xi_1|}e^{\sigma|\xi_2|}
$$
where $\sigma>0$, $0\leq\rho\leq 1$ and $\xi_1,\xi_2 \in \mathbb{R}$,
one can see that
\begin{align*}
&\| f(u)\|_{X^{0,b-1}}\\
&\sim\bigg \| \frac{\xi}{\langle\tau-p(\xi)\rangle^{1-b}}\int_{\mathbb{R}^2} e^{\sigma|\xi_1|}\hat{u}(\tau_1,\xi_1)e^{\sigma|\xi_2|}\hat{u}(\tau_2,\xi_2)-e^{\sigma|\xi|}\hat{u}(\tau_1,\xi_1)\hat{u}(\tau_2,\xi_2)\, d\tau_1d\xi_1\bigg \|_{L^2_{\tau,\xi}}\\
&\sim\bigg \| \frac{\xi}{\langle\tau-p(\xi)\rangle^{1-b}}\int_{\mathbb{R}^2}\big(e^{\sigma|\xi_1|}e^{\sigma|\xi_2|}-e^{\sigma|\xi|}\big)\hat{u}(\tau_1,\xi_1)
\hat{u}(\tau_2,\xi_2) \, d\tau_1 d\xi_1\bigg \|_{L^2_{\tau,\xi}}\\
&\lesssim\bigg \| \frac{\xi}{\langle\tau-p(\xi)\rangle^{1-b}}\int_{\mathbb{R}^2}\big[2\sigma \min(|\xi_1|,|\xi_2|)\big]^\rho e^{\sigma|\xi_1|}e^{\sigma|\xi_2|}\hat{u}(\tau_1,\xi_1)\hat{u}(\tau_2,\xi_2) \, d\tau_1 d\xi_1\bigg \|_{L^2_{\tau,\xi}},
\end{align*}
where $\tau_2=\tau-\tau_1$ and $\xi_2=\xi-\xi_1$.
Here, from the triangle inequality,
$$
\min(|\xi_1|,|\xi_2|)\leq 2\frac{(1+|\xi_1|)(1+|\xi_2|)}{(1+|\xi_1+\xi_2|)}=2\frac{\langle\xi_1\rangle\langle\xi_2\rangle}{\langle\xi\rangle},
$$
and therefore
\begin{align*}
\| f(u)\|_{X^{0,b-1}}&\lesssim\sigma^\rho\bigg \| \frac{\xi\langle\xi\rangle^{-\rho}}{\langle\tau-p(\xi)\rangle^{1-b}}
\int_{\mathbb{R}^2}e^{\sigma|\xi_1|}\langle\xi_1\rangle^{\rho}\hat{u}(\tau_1,\xi_1)
e^{\sigma|\xi_2|}\langle\xi_2\rangle^{\rho}\hat{u}(\tau_2,\xi_2)\, d\tau_1 d\xi_1\bigg \|_{L^2_{\tau,\xi}}\\
&=\sigma^\rho\Big\|\partial_x\Big(e^{\sigma|\partial_x|}\langle\partial_x\rangle^\rho u\cdot
e^{\sigma|\partial_x|}\langle\partial_x\rangle^\rho u\Big)\Big\|_{X^{0,-\rho,b-1}}\\
&\lesssim\sigma^\rho\big\| e^{\sigma|\partial_x|}\langle\partial_x\rangle^{\rho} u\big\|^2_{X^{0,-\rho,b}}\\
&=\sigma^\rho\| u\|^2_{X^{\sigma,0,b}}
\end{align*}
as desired. Here we used Lemma \ref{bilinear} with $\sigma=0$, $s=-\rho$ and $b'=b$ for the second inequality.
\end{proof}
\section{Local well-posedness and almost conservation law}\label{sec4}
In this section we first establish the local well-posedness and then the almost conservation law,
by making use of the bilinear estimate in the previous section.
They lie at the core of the proof of Theorem \ref{thm1} in the next section.
\subsection{Local well-posedness}\label{subsec4.1}
Based on Picard's iteration in the $X_\delta^{\sigma,s,b}$-space and Lemma \ref{lem00},
we establish the following local well-posedness in $G^{\sigma, s}$, with a lifespan $\delta > 0$.
Equally the radius of analyticity remains strictly positive in a short time interval $0 \leq t \leq\delta$,
where $\delta > 0$ depends on the norm of the initial data.
\begin{thm}\label{thm2}
Let $\sigma>0$ and $s>-7/4$.
Then, for any $u_0 \in G^{\sigma,s}$, there exist $\delta > 0$ and a unique solution u of the Cauchy problem \eqref{Kawahara} on the time interval $[0,\delta]$ such that $u \in C([0,\delta],G^{\sigma,s})$ and the solution depends continuously on the data $u_0$.
Here we have
\begin{equation}\label{deldel}
\delta = c_0(1+\| u_0 \|_{G^{\sigma.s}})^{-a}
\end{equation}
for some constants $c_0>0$ and $a>2$ depending only on $s$.
Furthermore, if\ $1/2<b<1$, the solution $u$ satisfies
\begin{equation}\label{lowe}
\| u \|_{X^{\sigma,s,b}_\delta} \leq C\| u_0 \|_{G^{\sigma,s}}
\end{equation}
with a constant $C>0$ depending only on $b$.
\end{thm}
\begin{proof}
Fix $\sigma >0$, $s>-7/4$ and $u_0\in G^{\sigma,s}$.
By Lemma \ref{lem00} we shall employ an iteration argument in the space $X^{\sigma,s,b}_\delta$ instead of $G^{\sigma,s}$.
Consider first the Cauchy problem for the linearised Kawahara equation
$$
\begin{cases}
u_t + \alpha u_{xxx} + \beta u_{xxxxx} = F(t,x),\\
u(0,x)=f(x).
\end{cases}
$$
By Duhamel's principle, the solution can be then written as
\begin{equation}\label{DF}
u(t,x)=e^{itp(-i\partial_x)}f(x)+\int_0^t e^{i(t-t')p(-i\partial_x)}F(t',\cdot)dt',
\end{equation}
where the Fourier multiplier $e^{itp(-i\partial_x)}$ with symbol $e^{itp(\xi)}$ is given by
$$e^{itp(-i\partial_x)}f(x)=\frac1{(2\pi)}\int_{\mathbb{R}}e^{ix\xi}e^{itp(\xi)}\widehat{f}(\xi)d\xi.$$
(Recall from Section \ref{sec2} that $p(\xi)=\alpha \xi^3 -\beta\xi^5$.)
Then the following $X_\delta^{\sigma,s,b}$-energy estimate follows directly from Proposition 2.1 in \cite{CLMW} (see also \cite{KPV}).
\begin{lem}\label{lem1}
Let $\sigma \geq 0$, $s\in\mathbb{R}$, $1/2<b\leq 1$ and $0<\delta\leq 1$.
Then we have
$$\| e^{itp(-i\partial_x)}f\|_{X^{\sigma,s,b}_\delta} \leq C_b\|f\|_{G^{\sigma,s}}$$
and
$$\bigg\|\int^t_0 e^{i(t-t')p(-i\partial_x)}F(t',\cdot)dt' \bigg\|_{X^{\sigma,s,b}_\delta} \leq C_b\| F\|_{X^{\sigma,s,b-1}_\delta},$$
with a constant $C_b>0$ depending only on $b$.
\end{lem}
Now, let $\left\{ u^{(n)} \right\}^{\infty}_{n=0}$ be the sequence defined by
$$
\begin{cases}
u^{(0)}_t+\alpha u^{(0)}_{xxx} + \beta u^{(0)}_{xxxxx} = 0,\\
u^{(0)}(0) = u_0(x),
\end{cases}
$$
and for $n\in\mathbb{Z}^+$
$$
\begin{cases}
u^{(n)}_t + \alpha u^{(n)}_{xxx} +\beta u^{(n)}_{xxxxx} = -\frac{1}{2}\partial_x\big(u^{(n-1)}u^{(n-1)}\big),\\
u^{(n)}(0) = u_0(x).
\end{cases}
$$
Applying \eqref{DF}, we first write
$$u^{(0)}(t,x) = e^{itp(-i\partial_x)}u_0(x)$$
and
$$u^{(n)}(t,x) =e^{itp(-i\partial_x)}u_0(x) - \frac{1}{2} \int_0^t e^{i(t-t')p(-i\partial_x)}\partial_x\big(u^{(n-1)}(t',\cdot)u^{(n-1)}(t',\cdot)\big) dt'.$$
By Lemma \ref{lem1} we have
\begin{equation}\label{0step}
\| u^{(0)}\|_{X^{\sigma,s,b}_\delta}\leq C_b\| u_0 \|_{G^{\sigma,s}},
\end{equation}
and Lemmas \ref{lem1}, \ref{lem2} and \ref{bilinear} combined imply
\begin{align}\label{nstep}
\nonumber\| u^{(n)} \|_{X^{\sigma,s,b}_\delta} &\leq C_b\| u_0\|_{G^{\sigma,s}} +
C_b\big\| \partial_x\big(u^{(n-1)}u^{(n-1)}\big)\big\|_{X^{\sigma,s,b-1}_\delta}\\
\nonumber&\leq C_b\| u_0\|_{G^{\sigma,s}} +C_bC_{b,b'}\delta^{b'-b}\big\| \partial_x\big(u^{(n-1)}u^{(n-1)}\big)\big\|_{X^{\sigma,s,b'-1}_\delta}\\
&\leq C_b\| u_0\|_{G^{\sigma,s}} +C_bC_{b,b'}C_{s,b,b'}\delta^{b'-b}\| u^{(n-1)} \|^{2}_{X^{\sigma,s,b}_{\delta}}
\end{align}
with $1/2<b<b^\prime<1$.
By induction together with \eqref{0step} and \eqref{nstep}, it follows that for all $n\geq0$
\begin{equation}\label{proof}
\| u^{(n)}\|_{X^{\sigma,s,b}_\delta} \leq 2C_b\| u_0\|_{G^{\sigma,s}},
\end{equation}
if we choose $\delta$ sufficiently small so that
\begin{equation}\label{delttt}
\delta^{b'-b}\|u_0\|_{G^{s,b}}\leq \frac1{8C_{b,b'}C_{s,b,b'}C_b^2}.
\end{equation}
Using Lemmas \ref{lem1}, \ref{lem2} and \ref{bilinear} together with \eqref{proof} and \eqref{delttt} in that order, we therefore get
\begin{align*}
\| u^{(n)}-&u^{(n-1)}\|_{X^{\sigma,s,b}_\delta}\\
&\leq C_b\big\| \partial_x\big(u^{(n-1)}u^{(n-1)}-u^{(n-2)}u^{(n-2)}\big)\big\|_{X^{\sigma,s,b-1}_\delta} \\
&\leq C_bC_{b,b'}\delta^{b'-b}\big\| \partial_x\big(u^{(n-1)}u^{(n-1)}-u^{(n-2)}u^{(n-2)}\big) \big\|_{X^{\sigma,s,b'-1}_\delta} \\
&\leq C_bC_{b,b'}C_{s,b,b'}\delta^{b'-b}\big\| u^{(n-1)} + u^{(n-2)} \big\|_{X^{\sigma,s,b}_\delta} \big\| u^{(n-1)} - u^{(n-2)} \big\|_{X^{\sigma,s,b}_\delta} \\
&\leq 4C_b^2C_{b,b'}C_{s,b,b'}\delta^{b'-b}\|u_0\|_{G^{s,b}}\big\| u^{(n-1)} - u^{(n-2)} \big\|_{X^{\sigma,s,b}_\delta}\\
&\leq \frac{1}{2}\big\| u^{(n-1)} - u^{(n-2)} \big\|_{X^{\sigma,s,b}_\delta},
\end{align*}
which guarantees the convergence of the sequence $\left\{ u^{(n)} \right\}^{\infty}_{n=0}$ to a solution $u$
with the bound \eqref{proof}.
Furthermore, \eqref{deldel} follows easily from \eqref{delttt} and $0<b'-b<1/2$.
Now assume that $u$ and $v$ are solutions to the Cauchy problem $\eqref{Kawahara}$ for initial data $u_0$ and $v_0$, respectively.
Then similarly as above, again with the same choice of $\delta$ and for any $\delta'$ such that $0<\delta^\prime<\delta$, we have
$$ \| u-v \|_{X^{\sigma,s,b}_{\delta^\prime}} \leq C_b\| u_0 -v_0 \|_{G^{\sigma,s}} + \frac{1}{2}\| u-v \|_{X^{\sigma,s,b}_{\delta^\prime}}$$
provided $\| u_0-v_0\|_{G^{\sigma,s}}$ is sufficiently small, which proves the continuous dependence of the solution on the initial data.
Finally, it remains to show the uniqueness of solutions.
Assume $u, v \in C_tG^{\sigma,s}$ are solutions to $\eqref{Kawahara}$ for the same initial data $u_0$ and let $w=u-v$. Then $w$ satisfies $w_t + \alpha w_{xxx} + \beta w_{xxxxx} + wu_x + vw_x =0$.
Multiplying both sides by $w$ and integrating in space yields
$$
\frac12\int_\mathbb{R} (w^2)_t dx +\alpha\int_\mathbb{R} ww_{xxx} dx +\beta\int_\mathbb{R} ww_{xxxxx} dx
+ \int_\mathbb{R} w^2u_x dx +\int_\mathbb{R} wvw_x dx= 0.
$$
Using $2wvw_x=(vw^2)_x-v_xw^2$ and integrating by parts, we then have
\begin{align*}
\frac{1}{2}\int_\mathbb{R} (w^2)_t dx -\frac{\alpha}{2}\int_\mathbb{R} (w_xw_x)_x dx
&+\frac{\beta}{2}\int_\mathbb{R} (w_{xx}w_{xx})_{x} dx+ \int_\mathbb{R} w^2u_x dx\\
&+\frac12\int_\mathbb{R} (vw^2)_x dx-\frac12\int_\mathbb{R} v_xw^2 dx=0.
\end{align*}
We may here assume that $w$ and its all spatial derivatives decay to zero as $|x|\rightarrow \infty$.\footnote{This property can be shown by approximation using the monotone convergence theorem
and the Riemann-Lebesgue lemma whenever $u\in X_\delta^{\sigma,1,b}$. See the argument in \cite{SS}, p. 1018.}
It follows then that
$$\frac{1}{2}\int_\mathbb{R} (w^2)_t dx =- \int_\mathbb{R} w^2u_x dx
+\frac12\int_\mathbb{R} v_xw^2 dx.$$
By H\"older's inequality, this implies
\begin{align*}
{\frac{d}{dt}}\| w(t) \|^2_{L^2_x} &\leq 2\big(\| u_x(t)\|_{L^\infty_x} +\|v_x(t)\|_{L^{\infty}_x}\big)\| w(t) \|^2_{L^2_x}\\
&\leq C\| w(t) \|^2_{L^2_x}.
\end{align*}
Here we used the fact that
$$
G^{\sigma,s} \subseteq G^{0,2}=H^2 \subseteq L^q
$$
for all $2\leq q\leq\infty$ and $\sigma>0$.
By Gr\"onwall's inequality, we now conclude that $w=0$.
\end{proof}
\subsection{Almost conservation law}\label{subsec4.2}
We have established the existence of local solutions; we would like to apply the local result repeatedly to cover time intervals of arbitrary length. This, of course, requires some sort of control on the growth of the norm on which the local existence time depends.
The following approximate conservation will allow us (see Section \ref{sec5}) to repeat the local result on successive short-time intervals to reach any target time $T>0$, by adjusting the strip width parameter $\sigma$ according to the size of $T$.
\begin{thm}\label{thm3}
Let $0\leq\rho\leq 1$, $\frac{1}{2}<b<1$ and $\delta$ be as in Theorem \ref{thm2}. Then there exists $C>0$ such that for any $\sigma > 0$ and any solution $u \in X^{\sigma,0,b}_{\delta}$ to the Cauchy problem \eqref{Kawahara} on the time interval $[0,\delta]$, we have the estimate
\begin{equation}\label{acl00}
\sup_{t\in[0,\delta]}\| u(t)\|^2_{G^{\sigma,0}} \leq \| u(0)\|^2_{G^{\sigma,0}} + C\sigma^\rho\| u \|^3_{X^{\sigma,0,b}_\delta}.
\end{equation}
\end{thm}
\begin{proof}
Let $0\leq \delta ' \leq \delta$. Setting $v(t,x)=e^{\sigma|\partial_x|}u(t,x)$ and applying $e^{\sigma|\partial_x|}$ to \eqref{Kawahara}, we obtain
$$
v_t+\alpha v_{xxx}+\beta v_{xxxxx} + vv_x = f(u),
$$
where $f(u)$ is as in \eqref{fff}:
$$
f(u)=\frac{1}{2}\partial_x\Big((e^{\sigma|\partial_x|}u)^2-e^{\sigma|\partial_x|}u^2\Big).
$$
Multiplying both sides by $v$ and integrating in space yield
$$
\int_\mathbb{R} vv_t dx +\alpha\int_\mathbb{R} vv_{xxx} dx +\beta\int_\mathbb{R} vv_{xxxxx} dx + \int_\mathbb{R} v^2v_x dx = \int_\mathbb{R} vf(u) dx.
$$
As before, we may here assume that $v$ and its all spatial derivatives decay to zero as $|x|\rightarrow \infty$.
Using this fact and integration by parts, we have
$$
\frac{1}{2}\int_\mathbb{R} (v^2)_t dx -\frac{\alpha}{2}\int_\mathbb{R} (v_xv_x)_x dx +\frac{\beta}{2}\int_\mathbb{R} (v_{xx}v_{xx})_{x} dx+ \frac{1}{3}\int_\mathbb{R} (v^3)_x dx = \int_\mathbb{R} vf(u) dx,
$$
and furthermore, the second, third and fourth terms on the left side vanish.
Subsequently integrating in time over the interval $[0,\delta ']$, we obtain
$$
\| u(\delta ')\|^2_{G^{\sigma,0}}\leq \| u(0)\|^2_{G^{\sigma,0}} + 2 \bigg | \int_{\mathbb{R}^2} \chi_{[0,\delta ']}(t)vf(u) dtdx \bigg |.
$$
Now by H\"older's inequailty, Lemma \ref{lem3} and Lemma \ref{f},
\begin{align*}
\bigg | \int_{\mathbb{R}^2} \chi_{[0,\delta ']}(t)vf(u) dtdx \bigg |
&\leq \| \chi_{[0,\delta ']}(t)v\|_{X^{0,1-b}}\| \chi_{[0,\delta ']}(t)f(u)\|_{X^{0,b-1}}\\
&\leq C\|v\|_{X^{0,1-b}_{\delta '}}\|f(u)\|_{X^{0,b-1}_{\delta '}}\\
&\leq C\|u\|_{X^{\sigma,0,1-b}_{\delta '}}\sigma^\rho\| u\|^2_{X^{\sigma,0,b}_{\delta '}}.
\end{align*}
Since $1-b<b$, we therefore get
$$
\sup_{t\in[0,\delta]}\| u(t)\|^2_{G^{\sigma,0}} \leq \| u(0)\|^2_{G^{\sigma,0}} + C\sigma^\rho\| u \|^3_{X^{\sigma,0,b}_\delta}
$$
as desired.
\end{proof}
\section{Proof of Theorem \ref{thm1}}\label{sec5}
By invariance of the Kawahara equation under the reflection $(t.x)\rightarrow(-t,-x)$, we may restrict to positive times. By the embedding \eqref{emb}, the general case $s\in\mathbb{R}$ will reduce to $s=0$ as shown in the end of this section.
\subsection{The case $s=0$}
Combining \eqref{lowe} and \eqref{acl00}, we first note that
\begin{equation}\label{acl01}
\sup_{t\in[0,\delta]}\| u(t)\|^2_{G^{\sigma,0}} \leq \| u(0)\|^2_{G^{\sigma,0}} + C\sigma^\rho\| u(0) \|^3_{G^{\sigma,0}}.
\end{equation}
Let $u_0=u(0)\in G^{\sigma_0,0}$ for some $\sigma_0>0$ and $\delta$ be as in Theorem \ref{thm2}. For arbitrarily large $T$, we want to show that the soution $u$ to \eqref{Kawahara} satisfies
$$
u(t)\in G^{\sigma(t),0} \quad\textrm{for all }\, t\in [0,T],
$$
where
\begin{equation}\label{sigma}
\sigma(t)\geq\frac{c}{T}
\end{equation}
with a constant $c>0$ depending on $\| u_0\|_{G^{\sigma_0,0}}$ and $\sigma_0$.
Now fix $T$ arbitrarily large. It suffices to show
\begin{equation}\label{label}
\sup_{t\in [0,T]}\| u(t)\|^2_{G^{\sigma,0}} \leq 2\| u_0\|^2_{G^{\sigma_0,0}}
\end{equation}
for $\sigma$ satisfying \eqref{sigma}, which in turn implies $u(t)\in G^{\sigma(t),0}$ as desired.
To prove \eqref{label}, we first choose $n\in\mathbb{Z}^+$ so that $n\delta\leq T\leq(n+1)\delta$. Using induction we shall show for any $k\in\left\{1,2,\cdots,n+1\right\}$ that
\begin{equation}\label{label1}
\sup_{t\in[0,k\delta]} \| u(t)\|^2_{G^{\sigma,0}} \leq \| u_0\|^2_{G^{\sigma,0}} + kC\sigma^\rho 2^{3/2}\| u_0\|^3_{G^{\sigma_0,0}}
\end{equation}
and
\begin{equation}\label{label2}
\sup_{t\in [0,k\delta]}\| u(t)\|^2_{G^{\sigma,0}} \leq 2\| u_0\|^2_{G^{\sigma_0,0}},
\end{equation}
provided $\sigma$ satisfies
\begin{equation}\label{label3}
\sigma\leq\sigma_0\quad\text{and}\quad\frac{2T}{\delta}C\sigma^\rho 2^{3/2}\| u_0\|_{G^{\sigma_0,0}}\leq 1.
\end{equation}
Indeed, for $k=1$, we have from \eqref{acl01} that
\begin{align*}
\sup_{t\in[0,\delta]}\| u(t)\|^2_{G^{\sigma,0}} &\leq \| u_0\|^2_{G^{\sigma,0}} + C\sigma^\rho\| u_0 \|^3_{G^{\sigma,0}}\\
&\leq 2\|u_0\|^2_{G^{\sigma_0,0}},
\end{align*}
where we used the fact that $\| u_0\|_{G^{\sigma,0}}\leq \| u_0\|_{G^{\sigma_0,0}}$ and $C\sigma^\rho \| u_0\|_{G^{\sigma_0,0}}\leq 1$
which are a direct consequence of \eqref{label3}.
Now assume \eqref{label1} and \eqref{label2} hold for some $k\in\left\{1,2,\cdots,n\right\}$. Applying \eqref{acl01}, \eqref{label2} and \eqref{label1}, we then have
\begin{align*}
\sup_{t\in[k\delta,(k+1)\delta]}\|u(t)\|^2_{G^{\sigma,0}}&\leq \| u(k\delta)\|^2_{G^{\sigma,0}} + C\sigma^\rho\| u(k\delta)\|^3_{G^{\sigma,0}}\\
&\leq \| u(k\delta)\|^2_{G^{\sigma,0}} + C\sigma^\rho 2^{3/2}\| u_0\|^3_{G^{\sigma_0,0}}\\
&\leq \| u_0\|^2_{G^{\sigma,0}} + C\sigma^\rho (k+1) 2^{3/2}\| u_0\|^3_{G^{\sigma_0,0}}.
\end{align*}
Combining this with the induction hypothesis \eqref{label1} for $k$, we get
\begin{equation}\label{label4}
\sup_{t\in[0,(k+1)\delta]}\|u(t)\|^2_{G^{\sigma,0}}\leq\| u_0\|^2_{G^{\sigma,0}} + C\sigma^\rho (k+1) 2^{3/2}\| u_0\|^3_{G^{\sigma_0,0}}
\end{equation}
which proves \eqref{label1} for $k+1$. Since $k+1\leq n+1\leq T/\delta+1\leq2T/\delta$, from \eqref{label3} we also get
$$
C\sigma^\rho (k+1) 2^{3/2}\| u_0\|_{G^{\sigma_0,0}}\leq \frac{2T}{\delta}C\sigma^\rho 2^{3/2}\| u_0\|_{G^{\sigma_0,0}}\leq 1
$$
which, along with \eqref{label4}, proves \eqref{label2} for $k+1$.
Finally, the condition \eqref{label3} is satisfied for
$$
\sigma = \bigg (\frac{\delta}{C2^{5/2}\|u_0\|_{G^{\sigma_0,0}}}\bigg )^{1/\rho}\bigg(\frac{1}{T}\bigg)^{1/\rho}.
$$
Particularly when $\rho=1$, the constant $c$ in \eqref{sigma} may be given as
$$
c=\frac{\delta}{C2^{5/2}\|u_0\|_{G^{\sigma_0,0}}}
$$
which depends only on $\| u_0 \|_{G^{\sigma_0,0}}$.
\subsection{The general case $s\in\mathbb{R}$.}
Recall that \eqref{emb} states
$$
G^{\sigma,s}\subset G^{\sigma^\prime,s^\prime}\ \textrm{ for all }\ \sigma>\sigma'\geq 0\ \textrm{ and }\ s,s'\in\mathbb{R}\textrm{.}
$$
For any $s\in\mathbb{R}$ we use this embedding to get
$$
u_0\in G^{\sigma_0,s}\subset G^{\sigma_0/2,0}.
$$
From the local well-posedness result, there is a $\delta=\delta(\| u_0\|_{G^{\sigma_0/2,0}})$ such that
$$
u(t)\in G^{\sigma_0/2,0}\quad\textrm{for }\ 0\leq t\leq\delta\textrm{.}
$$
Similarly as in the case $s=0$, for $T$ fixed greater than $\delta$, we have $u(t)\in G^{\sigma^\prime,0}$ for $t\in[0,T]$
and $\sigma^\prime\geq c/T$ with $c>0$ depending on $\| u_0 \|_{G^{\sigma_0/2,0}}$ and $\sigma_0$.
Applying the embedding again, we conclude
$$
u(t)\in G^{\sigma,s}\quad\text{for}\,\ t\in [0,T]
$$
where $\sigma =\sigma^\prime/2$.
\section{Proof of Lemma \ref{bilinear}}\label{sec6}
This last section is devoted to the proof of Lemma \ref{bilinear}.
\subsection{Preliminaries}
Before we begin the proof, we shall introduce some notations in Tao's $[k;Z]$-multiplier norm method \cite{T2}
with $k=3$ and $Z=\mathbb{R}^2$. Let $\Gamma_3(\mathbb{R}^2)$ denote the \textit{hyperplane}
$$\Gamma_3(\mathbb{R}^2) :=\{(\rho_1, \rho_2, \rho_3)\in (\mathbb{R}^2)^3 : \rho_1 + \rho_2 + \rho_3 =(0, 0)\}$$
where each $\rho_i=(\tau_i,\xi_i)$ is an ordered pair of real numbers $\tau_i,\xi_i$.
Note here that $\tau_1+\tau_2+\tau_3=\xi_1 + \xi_2 + \xi_3=0$.
We will also denote $\rho$, $\xi$ and $\tau$ the triplets $(\rho_1, \rho_2, \rho_3)$, $(\tau_1, \tau_2, \tau_3)$ and $(\xi_1, \xi_2, \xi_3)$, respectively. We endow $\Gamma_3(\mathbb{R}^2)$ with the obvious measure
$$\int_{\Gamma_3(\mathbb{R}^2)} f(\rho):=\int_{(\mathbb{R}^2)^2} f(\rho_1, \rho_2, -\rho_1-\rho_2)\,d\rho_1 d\rho_2$$
where $d\rho_i=d\tau_i d\xi_i$ for $i=1,2$.
For any function $m : \Gamma_3(\mathbb{R}^2) \rightarrow \mathbb{C}$, we define the $[3,\mathbb{R}^2]$-multiplier norm $\|m\|_{[3,\mathbb{R}^2]}$ to be the best constant $c$ such that the inequality
$$\bigg|\int_{\Gamma_3(\mathbb{R}^2)} m(\rho)\prod_{j=1}^3 f_j(\rho_j)\bigg|\leq c\prod_{j=1}^3\|f_j\|_{L^2(\mathbb{R}^{2})}$$
holds for all test functions $f_j$ on $\mathbb{R}^{2}$.
Capitalised variables such as $N_j$, $L_j$, $H$ are presumed to be dyadic, i.e., these variables range over numbers of the form $2^k$ for $k\in\mathbb{Z}$. Let $N_1$, $N_2$, $N_3>0$. It will be convenient to define the quantities $N_{max}\geq N_{med}\geq N_{min}$ to be the maximum, median and minimum of $N_1$, $N_2$, $N_3$, respectively.
Similarly we define $L_{max}\geq L_{med}\geq L_{min}$ whenever $L_1$, $L_2$, $L_3>0$. We also adopt the following summation conventions. Any summation of the form $L_{max}\sim\cdots$ is a sum over the three dyadic variables $L_1$, $L_2$, $L_3\gtrsim 1$, thus for instance
$$\sum_{L_{max}\sim H}:=\sum_{L_1, L_2, L_3\gtrsim 1:\ L_{max}\sim H}.$$
Similarly, any summation of the form $N_{max}\sim\cdots$ is a sum over the three dyadic variables $N_1$, $N_2$, $N_3>0$, thus for instance
$$\sum_{N_{max}\sim N_{med}\sim H}:=\sum_{N_1, N_2, N_3:\ N_{max}\sim N_{med}\sim H}.$$
If $\tau_j$ and $\xi_j$ are given for $j=1,2,3$, we define
$$\lambda_j:=\tau_j-p(\xi_j),$$
where $p(\xi_j)=\alpha \xi_j^3 -\beta\xi_j^5$ is as in Section $\ref{sec2}$.
Here we will let
$$h(\xi):=p(\xi_1)+p(\xi_2)+p(\xi_3)=-\lambda_1-\lambda_2-\lambda_3.$$
By dyadic decomposition of $\xi_j$, $\lambda_j$ and $|h(\xi)|$, one is led to consider
\begin{equation}\label{el1}
\|X_{N_1, N_2, N_3; H; L_1, L_2, L_3}\|_{[3;\mathbb{R}^2]}
\end{equation}
where $X_{N_1, N_2, N_3; H; L_1, L_2, L_3}$ is the multiplier
$$X_{N_1, N_2, N_3; H; L_1, L_2, L_3}(\xi, \tau):=\chi_{|h(\xi)|\sim H}\prod_{j=1}^3\chi_{|\xi_j|\sim N_j}\chi_{|\lambda_j|\sim L_j}.$$
Since the following identities
\begin{equation}\label{el2}
\xi_1+\xi_2+\xi_3=0
\end{equation}
and
\begin{equation}\label{el3}
\lambda_1+\lambda_2+\lambda_3+h(\xi)=0
\end{equation}
hold, from the support of the multiplier, one can see that $X_{N_1, N_2, N_3; H; L_1, L_2, L_3}$ vanishes unless
both
\begin{equation}\label{duck}
N_{max}\sim N_{med}\quad \text{and}\quad L_{max}\sim\max\{H, L_{med}\}
\end{equation}
are satisfied.
We then recall the following fundamental estimates on dyadic blocks for the Kawahara equation.
\begin{lem}[\cite{CLMW}]\label{el4}
Let $H,N_1,N_2,N_3,L_1,L_2,L_3>0$ obey \eqref{el2}, \eqref{el3} and $N_{max}\sim N_{med}\gtrsim 1$.
Then
\begin{equation}\label{el5}
H\sim N_{max}^4 N_{min},
\end{equation}
and
\begin{itemize}
\item If $N_{max}\sim N_{min}$ and $L_{max}\sim H$,
\begin{equation}\label{el6}
\eqref{el1}\lesssim L_{min}^{1/2}N_{max}^{-2}L_{med}^{1/2}.
\end{equation}
\item If $N_{\sigma(2)}\sim N_{\sigma(3)}\gg N_{\sigma(1)}$ and $H\sim L_{\sigma(1)}\gtrsim L_{\sigma(2)},\ L_{\sigma(3)}$,
\begin{equation}\label{el7}
\eqref{el1}\lesssim L_{min}^{1/2}N_{max}^{-2}\min\Big\{ H, \frac{N_{max}}{N_{min}}L_{med}\Big\}^{1/2}
\end{equation}
where $\sigma$ is a permutation in $\{1,2,3\}$.
\item In all other cases,
\begin{equation}\label{el8}
\eqref{el1} \lesssim L_{min}^{1/2}N_{max}^{-2}\min\{ H, L_{med}\}^{1/2}.
\end{equation}
\end{itemize}
\end{lem}
\subsection{Proof}
Now we turn to the proof of Lemma \ref{bilinear}.
By the definition of $X^{s,b}$-norms and the dual characterisation of $L^2$ space, we may show that
\begin{equation}\label{app01}
\bigg \| \frac{(\xi_1+\xi_2)\langle\xi_1\rangle^{-s}\langle\xi_2\rangle^{-s}
\langle\xi_3\rangle^s}{\langle\tau_1-p(\xi_1)\rangle^b\langle\tau_2-p(\xi_2)\rangle^b\langle\tau_3-p(\xi_3)\rangle^{1-b'}} \bigg \|_{[3;\mathbb{R}^2]}\lesssim 1.
\end{equation}
To show this, we first decompose dyadically $|\xi_j|,|\lambda_j|,|h(\xi)|$ as $|\xi_j|\sim N_j$, $|\lambda_j|\sim L_j$, $|h(\xi)|\sim H$.
By some properties of the $[k;Z]$-multiplier norm,
one may restrict the multiplier in \eqref{app01} to the region $L_j \gtrsim 1$ $(j=1,2,3)$ and $\max\{N_1, N_2, N_3\}\gtrsim 1$.
Consequently, it suffices to show that
\begin{align*}
\bigg\|\sum_{N_{max}\gtrsim 1} \sum_{H} \sum_{L_1,L_2,L_3\gtrsim1}
\frac{N_3\langle N_3\rangle^s}{\langle N_1\rangle^s\langle N_2\rangle^sL_1^bL_2^bL_3^{1-b'}} \chi_{N_1,N_2,N_3;H;L_1,L_2,L_3}\bigg\|_{[3;\mathbb{R}^2]}\lesssim 1.
\end{align*}
By Lemma 3.11 (Schur's test) in \cite{T2},
this estimate is reduced to showing that
\begin{align}\label{app02}
\nonumber\sum_{N_{max}\sim N_{med} \sim N} \sum_{H} \sum_{L_{max}\sim H}&
\frac{N_3\langle N_3\rangle^s}{\langle N_1\rangle^s\langle N_2\rangle^sL_1^bL_2^bL_3^{1-b'}}\\
&\quad\times\|\chi_{N_1,N_2,N_3;H;L_1,L_2,L_3}\|_{[3;\mathbb{R}^2]}\lesssim 1
\end{align}
and
\begin{align}\label{app03}
\nonumber\sum_{N_{max}\sim N_{med} \sim N} \sum_{H}\sum_{L_{max} \sim L_{med}\gg H}&
\frac{N_3\langle N_3\rangle^s}{\langle N_1\rangle^s\langle N_2\rangle^sL_1^bL_2^bL_3^{1-b'}}\\
&\quad\times\|\chi_{N_1,N_2,N_3;H;L_1,L_2,L_3}\|_{[3;\mathbb{R}^2]}\lesssim 1
\end{align}
for all $N\gtrsim 1$.
\subsubsection{Proof of \eqref{app03}}
Since $N_{max}\sim N_{med}\sim N\gtrsim 1$, by (\ref{el5}) and (\ref{el8}), it suffices to show
\begin{equation}\label{app04}
\sum_{N_{max}\sim N_{med} \sim N} \sum_{L_{max}\sim L_{med}\gg N^4N_{min}} \frac{N_3\langle N_3\rangle^s}{\langle N_1\rangle^s\langle N_2\rangle^sL_1^bL_2^bL_3^{1-b'}}L^{1/2}_{min}N^{1/2}_{min} \lesssim 1.
\end{equation}
We only need to consider two cases: $N_1 \sim N_2 \sim N$, $N_3 =N_{min}$ and $N_1 \sim N_3\sim N$, $N_2=N_{min}$.
(The other case $N_2 \sim N_3\sim N$, $N_1=N_{min}$ then follows by symmetry.)
In the former case, the estimate \eqref{app04} can be further reduced to
$$
\sum_{N_{min}\lesssim N} \sum_{L_{max}\sim L_{med}\gg N^4N_{min}} \frac{N^{-2s}N_{min}\langle N_{min}\rangle^s}{L^b_{min}L^b_{med}L^{1-b'}_{max}}L^{1/2}_{min}N^{1/2}_{min} \lesssim 1
$$
since $1-b'<1/2<b$.
Then performing the $L$ summations, we reduce to
$$
\sum_{N_{min}\lesssim N} \frac{N^{-2s}N_{min}^{3/2}\langle N_{min}\rangle^s}{(N^4 N_{min})^{1+b-b'}} \lesssim 1,
$$
which holds if $4(1+b-b')+2s>0$.
So we require $0\leq b'-b<1+s/2$, which is possible if $s>-2$.
Therefore, \eqref{app04} follows.
In the latter case, the estimate \eqref{app04} can be reduced to
$$
\sum_{N_{min}\lesssim N} \sum_{L_{max}\sim L_{med}\gg N^4N_{min}} \frac{N}{\langle N_{min}\rangle^sL^b_{min}L^b_{med}L^{1-b'}_{max}}L^{1/2}_{min}N^{1/2}_{min} \lesssim 1.
$$
Before performing the $L$ summations, we need to divide cases into two parts, $N_{min}\leq 1$ and $N_{min}\geq 1$, as follows:
\begin{align*}
&\sum_{N_{min}\leq 1}\sum_{L_{max}\sim L_{med}\gg N^4N_{min}} \frac{NN^{1/2}_{min}}{L^{b-1/2}_{min}L_{max}^{1-b+b'}}\\
+&\sum_{1\leq N_{min}\lesssim N} \sum_{L_{max}\sim L_{med}\gg N^4N_{min}} \frac{NN^{1/2-s}_{min}}{L^{b-1/2}_{min}L_{max}^{1-b+b'}} \lesssim 1.
\end{align*}
This holds clearly if $s>-7/2$, and therefore \eqref{app04} holds for $s>-7/2$.
\subsubsection{Proof of \eqref{el6}}
The conditions
$N_{max}\sim N_{med} \sim N\gtrsim 1$ and $L_{max}\sim H$ in \eqref{app02} are divided into three cases
corresponding to each of \eqref{el6}, \eqref{el7} and \eqref{el8}.
By (\ref{el5}) we also see that $L_{max}\sim N^4_{max}N_{min}$,
which is used repeatedly below.
\subsubsection*{The case for \eqref{el6}}
Now we shall consider the first case, which reduces to
\begin{equation}\label{app05}
\sum_{L_{max}\sim N^5} \frac{N^{-s}N}{L^b_{min}L^b_{med}L^{1-b'}_{max}}L^{1/2}_{min}N^{-2}L^{1/2}_{med}\lesssim 1.
\end{equation}
Performing the $L$ summations, we reduce to
$$
\frac{1}{N^{1+s}N^{5(1-b')}}\lesssim 1,
$$
which holds if $1+s+5(1-b')\geq0$. So we require $1/2< b'\leq(6+s)/5$, which is possible if $s>-7/2$.
Therefore, \eqref{app05} follows.
\subsubsection*{The case for \eqref{el7}}
Next we consider the second case which is divided into $N_1 \sim N_2 \gg N_3,H \sim L_3 \gtrsim L_1,L_2$ and
$N_2 \sim N_3 \gg N_1, H \sim L_1 \gtrsim L_2,L_3$.
(The other case $N_1 \sim N_3 \gg N_2, H \sim L_2 \gtrsim L_1,L_3$ then follows by symmetry.)
In the former case the estimate \eqref{app02} reduces to
\begin{equation}\label{app06}
\sum_{N_3 \ll N}\sum_{1\lesssim L_1,L_2\lesssim N^4N_3}\frac{N_3\langle N_3\rangle^s}{N^{2s}L^b_1L^b_2L^{1-b'}_3}L^{1/2}_{min}N^{-2}\min\bigg\{N^4N_3,\frac{N}{N_3}L_{med}\bigg\}^{1/2}\lesssim 1.
\end{equation}
We then decompose the left-hand side of \eqref{app06} into two parts $N_3\leq 1$ and $N_3> 1$:
\begin{align*}
&\sum_{N_3\leq 1}\sum_{1\lesssim L_1,L_2\lesssim N^4N_3} \frac{N_3\langle N_3\rangle^s}{N^{2s}L^b_1L^b_2L^{1-b'}_3}L^{1/2}_{min}N^{-2}\min\bigg\{N^4N_3,\frac{N}{N_3}L_{med}\bigg\}^{1/2}\\
&+ \sum_{1<N_3\ll N}\sum_{1\lesssim L_1,L_2\lesssim N^4N_3} \frac{N_3\langle N_3\rangle^s}{N^{2s}L^b_1L^b_2L^{1-b'}_3}L^{1/2}_{min}N^{-2}\min\bigg\{N^4N_3,\frac{N}{N_3}L_{med}\bigg\}^{1/2}\\
&=:I_1 +I_2.
\end{align*}
We first consider the part $I_1$. When $N^4N_3\geq \frac{N}{N_3}L_{med}$, which is equivalent to $N_3\geq( L_{med}/N^3)^{1/2}$,
we see that
$$
I_1\lesssim\sum_{N_3\leq1}\sum_{1\lesssim L_1,L_2\lesssim N^4N_3\atop (L_{med}/N^3)^{1/2}\leq N_3}
\frac{N_3}{N^{2s}L^{b-1/2}_{min}L^b_{med}(N^4N_3)^{1-b'}}N^{-2}N^{1/2}L^{1/2}_{med}N^{-1/2}_3.
$$
Performing the $N_3$ summation, we have the desired bound
$$
I_1 \lesssim\sum_{1\lesssim L_1,L_2\lesssim N^4}\frac{N^{-3/2}}{N^{2s}L^{b-1/2}_{min}L^{b-1/2}_{med}N^{4(1-b')}}
\lesssim 1$$
if $2s+4(1-b')+3/2>0$.
So we require $1/2<b\leq b'<(4s+11)/8$, which is possible if $s>-7/4$.
On the other hand, when $N_3<( L_{med}/N^3)^{1/2}$,
$$
I_1\lesssim\sum_{N_3\leq 1}\sum_{1\lesssim L_1,L_2\lesssim N^4N_3\atop N_3<( L_{med}/N^3)^{1/2}} \frac{N_3}{N^{2s}L^{b-1/2}_{min}L^b_{med}(N^4N_3)^{1-b'}}N^{-2}N^2 N^{1/2}_3.
$$
Performing the $N_3$ summation, we have
$$
I_{1} \lesssim \sum_{1\lesssim L_1,L_2\lesssim N^4}\frac{\min\{1,( L_{med}/N^3)^{\frac{1}{2}(\frac{1}{2}+b')}\}}{N^{2s}L^{b-1/2}_{min}L^{b}_{med}N^{4(1-b')}}
\lesssim 1
$$
if $2s+4(1-b')+\frac{3}{2}(\frac{1}{2}+b')>0$ when $(L_{med}/N^3)^{\frac{1}{2}(\frac{1}{2}+b')}\leq1$,
and if $2s+4+3b-4b'>0$ when $(L_{med}/N^3)^{\frac{1}{2}(\frac{1}{2}+b')}\geq1$.
So we require
$1/2<b\leq b'<(8s+19)/10$ and
$1/2<b'+3(b'-b)<4+2s$,
but this is possible if $s>-7/4$.
Now we consider the part $I_2$ that has the following trivial bound
\begin{equation*}
I_2\lesssim \sum_{1<N_3\ll N} \sum_{1\lesssim L_1,L_2\lesssim N^4N_3} \frac{N_3 N^s_3}{N^{2s}L^{b-1/2}_{min}L^b_{med}(N^4N_3)^{1-b'}}N^{-2}N^{1/2}L^{1/2}_{med}N^{-1/2}_3.
\end{equation*}
Performing the $N_3$ summation, we conclude that
$$
I_2\lesssim \sum_{1\lesssim L_1,L_2\lesssim N^5}\frac{\min\{1,N^{s-1/2+b'}\}}{N^{2s}L^{b-1/2}_{min}L^{b-1/2}_{med}N^{4(1-b')}N^{3/2}}\lesssim 1
$$
if $2s+4(1-b')+3/2>0$ when $s-1/2+b'< 0$, and
if $2s+4(1-b')+3/2>s-1/2+b'$ when $s-1/2+b'\geq 0$.
So we require $1/2<b\leq b'<\min\{(4s+11)/8,(s+6)/5\}$.
This is possible if $s>-7/4$.
Consequently, we get the desired estimate \eqref{app06} if $s>-7/4$.
Next we deal with the remaining case where $N_2\sim N_3 \gg N_1, H\sim L_1 \gtrsim L_2,L_3$. In this case,
applying \eqref{el7} to \eqref{app02}, we may show
\begin{equation*}
\sum_{N_1\ll N}\sum_{1\lesssim L_2,L_3 \lesssim N^4N_1} \frac{N^{1+s}L^{1/2}_{min}}{N^s\langle N_1\rangle^sL^b_2L^{1-b'}_3(N^4N_1)^b}N^{-2}\min\bigg\{ H,\frac{N}{N_1}L_{med}\bigg\}^{1/2}\lesssim 1.
\end{equation*}
We first decompose the left-hand side of this inequality into two parts $N_1\leq 1$ and $N_1> 1$:
\begin{align*}
&\sum_{N_1\leq 1}\sum_{1\lesssim L_2,L_3 \lesssim N^4N_1} \frac{N^{1+s}}{N^sL^b_{min}L^{1-b'}_{med}(N^4N_1)^b}L^{1/2}_{min}N^{1/2}_1 \\
&+\sum_{1<N_1\ll N }\sum_{1\lesssim L_2,L_3 \lesssim N^4N_1} \frac{N^{1+s}}{N^s_1N^sL^b_{min}L^{1-b'}_{med}(N^4N_1)^b}L^{1/2}_{min}N^{-2}N^{5/4}L^{1/4}_{med} \\
&=: J_1+J_2.
\end{align*}
Here we used the fact that $H\sim N^4N_{min}$ and $\min\{a,b\}\leq\sqrt{ab}$.
Note first that the $L$ summations in $J_1$ vanish unless $N^4N_1\gtrsim 1$.
Using this, we get
$$
J_1\lesssim \sum_{N^{-4}\lesssim N_1\leq 1} \frac{NN^{1/2-b}_1}{N^{4b}}\lesssim \frac{NN^{(-4)(1/2-b)}}{N^{4b}}\lesssim 1
$$
since $b>1/2$ and $N\gtrsim1$.
Performing the $L$ summations in $J_2$, we see that
$$
J_2\lesssim \sum_{1\leq N_1 \ll N} \frac{N^{1/4}N_1^{-s-b}}{N^{4b}}
$$
if we take $1/2< b\leq b'<3/4$. (Otherwise, the sum may diverge.)
When $-s-b>0$, we conclude
$$
J_2\lesssim \frac{N^{1/4}N^{-s-b}}{N^{4b}} \lesssim 1
$$
if $4b+s+b-1/4>0$.
So we require $\max\{1/2,(1/4-s)/5\}<b \leq b'<3/4$ and $b<-s$.
This is possible if $-7/2<s<-1/2$.
Furthermore, we can have the range $1/2<b \leq b'<3/4$
if $-9/4\leq s<-1/2$.
When $-s-b\leq0$,
$$
J_2\lesssim \frac{N^{1/4}}{N^{4b}}\log_2(N)\lesssim 1
$$
which holds when $1/2<b\leq b'< 3/4$.
Consequently, we get the desired estimate.
\subsubsection*{The case for \eqref{el8}}
Lastly, using \eqref{el8}, the desired estimate \eqref{app02} reduces to
$$
\sum_{N_{max}\sim N_{med} \sim N} \sum_{L_{max}\sim N^4N_{min}}
\frac{N_3\langle N_3\rangle^s}{\langle N_1\rangle^s\langle N_2\rangle^sL^b_1L^b_2L^{1-b'}_3}L^{1/2}_{min}N^{-2}\min\{H,L_{med}\}^{1/2}\lesssim 1.
$$
To show this, we need to divide the case into $N_1 \sim N_2 \sim N,$ $N_3 = N_{min}$ and $N_1 \sim N_3 \sim N,$ $N_2 = N_{min}$.
(The other case $N_2 \sim N_3 \sim N,$ $N_1 = N_{min}$ then follows by symmetry.)
In the former case, the above estimate further reduces to
$$
\sum_{ N_3\ll N} \sum_{L_{max}\sim N^4N_3} \frac{N_3\langle N_3\rangle^s}{N^{2s}L^b_{min}L^b_{med}(N^4N_3)^{1-b'}}L^{1/2}_{min}N^{-2}L^{1/2}_{med}\lesssim 1.
$$
Then performing the $L$ summations, we reduce to
$$
\sum_{N_3\ll N} \frac{N_3\langle N_3\rangle^s}{N^{2+2s}N^{4(1-b')}N^{1-b'}_3} \lesssim 1,
$$
which holds if $2+2s+4(1-b')>0$.
So we require $1/2<b\leq b'<(s+3)/2$.
This is possible if $s>-2$.
In the latter case, we reduce to
\begin{equation*}
\sum_{ N_2\ll N} \sum_{L_{max}\sim N^4N_2}\frac{N^{1+s}L^{1/2}_{min}N^{-2}}{N^s \langle N_2\rangle^s L^b_{min}L^b_{med} L^{1-b'}_{max}} \min\{H,L_{med}\}^{1/2} \lesssim 1.
\end{equation*}
Using $\min\{a,b\}\leq\sqrt{ab}$ and then performing the $L$ summations, we further reduce to
\begin{equation}\label{app11}
\sum_{ N_2\ll N} \frac{N^{1+s}N^{-2}}{N^s \langle N_2\rangle^s(N^4N_2)^{1-b'}}(N^4N_2)^{1/4} \lesssim 1.
\end{equation}
We then decompose the left-hand side of \eqref{app11} into two parts $N_2\leq 1$ and $N_2> 1$:
\begin{align*}
&\sum_{N_2\leq 1} \frac{N^{1+s}N^{-2}}{N^s \langle N_2\rangle^s(N^4N_2)^{1-b'}}(N^4N_2)^{1/4}
+ \sum_{1< N_2\leq N} \frac{N^{1+s}N^{-2}}{N^s \langle N_2\rangle^s(N^4N_2)^{1-b'}}(N^4N_2)^{1/4}\\
&=: K_1+K_2.
\end{align*}
It is easy to see that $K_1\lesssim N^{-4(1-b')}\lesssim 1$ if $b'<3/4$.
On the other hand,
$$K_2\lesssim\frac{\min\{1,N^{b'-s-3/4}\}}{N^{4(1-b')}}\lesssim1$$
if $b'\leq1$ when $b'-s-3/4\geq 0$, and if $4(1-b')-(b'-s-3/4)>0$ when $b'-s-3/4<0$.
So we get the desired bound for all $s$ with $1/2<b\leq b'<\min\{(4s+19)/20,s+3/4\}$.
This completes the proof.
|
\section{Introduction}
Why do we remember the things we do? Decades of work have provided numerous explanations: we remember things that are out of context \cite{zoya,standing}, that are emotionally salient \cite{emotions}, that involve people~\cite{Isola2014}, etc. But a picture is, as they say, worth a thousand words. What does it \emph{look like} to make an image more or less memorable? The same questions can be asked for many cognitive visual properties: what visual changes can take a bland foggy seascape and add just the right colors and tones to make it serenely beautiful.
Attributes like memorability, aesthetics, and emotional valence are of special interest because we do not have concrete definitions of what they entail. This contrasts with attributes like ``object size" and "smile". We know exactly what it means to zoom in on a photo, and it's easy to imagine what a face looks like as it forms a smile. It's an open question, on the other hand, what exactly do changes in ``memorability" look like? Previous work has built powerful predictive models of image memorability~\cite{Isola2014,Khosla_2015_ICCV} but these have fallen short of providing a fine-grained visual explanation of what underlies the predictions.
In this paper, we propose a new framework, GANalyze, based on Generative Adversarial Networks (GAN)~\cite{goodfellow}, to study the visual features and properties that underlie high-level cognitive attributes. We focus on image memorability as a case study, but also show that the same methods can be applied to study image aesthetics and emotional valence.
Our approach leverages the ability of GANs to generate a continuum of images with fine-grained differences in their visual attributes. We can learn how to navigate the GAN's latent space to produce images that have increasing or decreasing memorability, according to an off-the-shelf memorability predictor~\cite{Khosla_2015_ICCV}. Starting with a seed image, this produces a sequence of images of increasing and decreasing predicted memorability (see Figure 1). By showing this visualization for a diverse range of seed images, we come up with a catalog of different image sequences showcasing a variety of visual effects related to memorability. We call this catalog a \emph{visual definition} of image memorability. GANalyze thereby offers an alternative to the non-parametric approach in which real images are simply sorted on their memorability score to visualize what makes them memorable (example shown in Figure~\ref{fig:nonparametric}). The parametric, fine-grained visualizations generated by GANalyze provide much clearer visual definitions.
These visualizations surface several correlates of memorability that have been overlooked by prior work, including ``object size", ``circularity", and ``colorfulness". Most past work on modeling image memorability focused on semantic attributes, such as object category (e.g., ``people" are more memorable than ``trees")~\cite{Isola2014}. By applying our approach to a class-conditional GAN, BigGAN~\cite{BigGAN}, we can restrict it to only make changes that are orthogonal to object class. This reveals more fine-grained changes that nonetheless have large effects on predicted memorability. For example, consider the cheeseburgers in Figure~\ref{fig:examples1}. Our model visualizes more memorable cheeseburgers as we move to the right. The apparent changes go well beyond semantic category -- the right-most burger is brighter, rounder, more canonical, and, we think, looks tastier.
Since our visualizations are learned based on a \emph{model} of memorability, a critical step is to verify that what we are seeing really has a causal effect on human behavior. We test this by running a behavioral experiment that measures the memorability of images generated by our GAN, and indeed we find that our manipulations have a causal effect: navigating the GAN manifold toward images that are predicted to be more memorable actually results in generating images that are measurably more memorable in the behavioral experiment.
Our contributions include the following:
\begin{itemize}
\item Introducing GANalyze, a framework that uses GANs to provide a \emph{visual definition} of image properties, like memorability and aesthetics, that we can measure but are not easy, in words, to define.
\item Showing that this framework surfaces previously overlooked attributes that correlate with memorability.
\item Demonstrating that the discovered transformations have a causal effect on memorability.
\item Showing that GANalyze can be applied to provide visual definitions for aesthetics and emotional valence.
\end{itemize}
\subsection{Related work}
{\bf Generative Adversarial Networks or GANs}.
GANs ~\cite{goodfellow} introduced a revolutionary framework to synthesize natural-looking images ~\cite{stylegan,BigGAN,sagan,proggan,BigGAN}. Among the many applications for GANs are style transfer~\cite{Zhu_2017_ICCV}, visual prediction~\cite{mathieu2015deep}, and ``sim2real" domain adaptation~\cite{graspgan}. Here, we show how they can also be applied to the problem of understanding high-level, cognitive image properties, such as memorability.
{\bf Understanding CNN representations}
The internal representations of a CNN can be unveiled using methods like network dissection \cite{zhou2018,bau2017,objectdetectors} including for a CNN trained on memorability \cite{Khosla_2015_ICCV}. For instance, Khosla et al. \cite{Khosla_2015_ICCV} showed that units with strong positive correlations with memorable images specialized for people, faces, body parts, etc., while those with strong negative correlations where more sensitive to large regions in landscapes scenes. Here, our framework introduces a new way of defining what memorability, and aesthetic, variability look like.
{\bf Modifying Memorability}. The memorability of an image, like faces, can be manipulated using warping techniques \cite{Khosla_2013_ICCV}. Concurrent work has also explored using a GAN for this purpose \cite{sidorov}. Another approach is a deep style transfer ~\cite{deepstyle} which taps into more artistic qualities. Now that GANs have reached a quality that is often almost indistinguishable from real images, they offer a powerful tool to synthesize images with different cognitive qualities. As shown here, our GANalyze framework successfully modified GAN-generated images across a wide range of image categories to produce a second generation of GAN realistic photos with different mnemonic qualities.
\begin{figure*}[b]
\includegraphics[width=\textwidth]{latex/method_schematic.pdf}\qquad
\caption{\textbf{Schematic of the model}. The model learns how to transform a $\mathbf{z}$ vector such that when fed to a Generator, the resulting image's property of interest changes. The transformation is achieved by the \emph{Transformer}, who moves the $\mathbf{z}$ vector along a learned direction, $\mathbf{\theta}$, in the Generator's latent space. The property of interest (e.g., memorability) is predicted by an \emph{Assessor module} (e.g., MemNet). Finally, $\alpha$ acts as knob to set the degree of change one wants to achieve in the Assessor value (e.g., MemNet score). It tells the \emph{Transformer} how far exactly to move along $\mathbf{\theta}$.}
\label{fig:model}
\end{figure*}
\section{Model}
\subsection{Formulation}
We start with a pretrained Generator $G$, who takes a noise vector $\mathbf{z} \in \mathbb{R}^{1\times M}$ and a one-hot class vector $\mathbf{y} \in \{0;1\}^{1\times C}$ as input and generates a photo-realistic image $G(\mathbf{z},\mathbf{y})$. Assumed is also an Assessor function $A$ that assesses an image property of interest, in this case memorability. Our goal was to learn to transform any given noise vector $\mathbf{z}$ of any class $\mathbf{y}$ such that the memorability of its resulting, generated image increases (or decreases) with a certain amount $\alpha$. The transformation is achieved by a Transformer function, who moves the input $\mathbf{z}$ along a certain direction $\mathbf{\theta} \in \mathbb{R}^{1\times M}$ in the latent space. We express the objective as:
\begin{equation}
\mathcal{L}(\theta)=\mathbb{E}_{\mathbf{z},\mathbf{y},\alpha}[(A(G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y}))\\
-(A(G(\mathbf{z},\mathbf{y}))+\alpha))^2]
\end{equation}
Note that this is simply the MSE loss between the target memorability score, i.e. the seed image's score $A(G(\mathbf{z},\mathbf{y}))$ increased by $\alpha$, and the memorability score of the transformed clone image $A(G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y}))$. The scalar $\alpha$ acts as a metaphorical knob with which one can use to turn up or turn down memorability. The optimizing problem is $\theta^* = \argminD_{\theta} \mathcal{L}(\theta)$.
The Transformer $T$ is defined as:
\begin{equation}
\begin{split}
T_{\theta}(\mathbf{z},\alpha) &= \mathbf{z} + \alpha \mathbf{\theta}\\
\end{split}
\end{equation}
Figure~\ref{fig:model} presents a schematic of the model. Finally, note that when $\alpha = 0$, $T$ becomes a null operation and $G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y})$ then equals $G(\mathbf{z},\mathbf{y})$.
\subsection{Implementation}\label{implementation}
For the results presented here, we used the Generator of BigGAN \cite{BigGAN}, which generates state-of-the art GAN images and is pretrained on ImageNet \cite{Imagenet}. The Assessor was implemented as MemNet \cite{Khosla_2015_ICCV}, a CNN predicting image memorability. Note, however, that training our model with different Generators or different Assessors can easily be achieved by substituting the respective modules. We discuss an Assessor for image aesthetics in Section~\ref{other}. Furthermore, we present additional results for implementations with a StyleGAN \cite{stylegan} Generator in the supplementary materials.
To train our model and find $\theta^*$, we built a training set by randomly sampling 400K $\mathbf{z}$ vectors from a standard normal distribution truncated to the range $[-2,2]$. Each $\mathbf{z}$ was accompanied by an $\alpha$ value, randomly drawn from a uniform distribution between -0.5 and 0.5, and a randomly chosen $\mathbf{y}$. We used a batch size of 4 and an Adam optimization procedure.
In view of the behavioral experiments (see Section~\ref{experiments}), we restricted the test set to 750 randomly chosen ImageNet classes and two $\mathbf{z}$ vectors per class. Each $\mathbf{z}$ vector was then paired with five different $\alpha$ values: ${[-0.2,-0.1,0,0.1,0.2]}$. Note that this includes an $\alpha$ of 0, representing the original image $G(\mathbf{z},\mathbf{y})$. Finally, the test set consisted of 1.5K sets of five images, or 7.5K test images in total.
\section{Experiments}\label{experiments}
\subsection{Model validation}
Did our model learn to navigate the latent space such that it can increase (or decrease) the Assessor score of the generated image with positive (or negative) $\alpha$ values?
Figure~\ref{fig:memorability_graph}.A suggests the model learned. The mean MemNet score of test set images increases with every increment of $\alpha$. To test this formally, we fitted a linear mixed-effects regression model to the data and found a (unstandardized) slope ($\beta$) of 0.68 ($95\% CI = [0.66,0.70], p<0.001)$, confirming that the Memnet score increases significantly with $\alpha$.
\begin{figure}[b]
\centering
\includegraphics[width=0.95\textwidth]{latex/memnet_validation3.png}
\caption{\textbf{Model results}. A) Graph shows the mean MemNet score across the images in every $\alpha$ condition. Our model successfully learned how to modify a GAN image to decrease (negative $\alpha$) or increase (positive $\alpha$) its MemNet score. B) List of emerging factors potentially underlying the effect observed in (A), and graph of how they change in function of $\alpha$. The factors emerged from visualizations generated by the GANalyze framework (examples shown in Figures~\ref{fig:examples1}, ~\ref{fig:sup_examples_mem_1}, and ~\ref{fig:sup_examples_mem_2}). Emerging factor scores were first normalized and then averaged per $\alpha$ condition.}
\label{fig:memorability_graph}
\end{figure}
\begin{figure*}[!p]
\centering
\includegraphics[height=0.98\dimexpr\textheight\relax]{latex/examples_memnet_0.pdf}
\caption{\textbf{Examples of generated images} along the memorability dimension. The middle column represents $G(\mathbf{z},\mathbf{y})$, the generated image serving as the original seed to create a series of clone images more or less memorable.}
\label{fig:examples1}
\end{figure*}
\subsection{Emerging factors}\label{Emerging factors}
We observe that the model can successfully change the memorability of an image, given its $\mathbf{z}$ vector. Next, we ask which image factors it altered to achieve this. The answer to this question can provide further insight into what the Assessor has learned about the to-be-assessed image property, in this case what MemNet has learned about memorability. From a qualitative analysis of the test set (examples shown in Figures ~\ref{fig:examples1}, ~\ref{fig:sup_examples_mem_1}, and ~\ref{fig:sup_examples_mem_2}), a number of candidate factors stand out.
First, MemNet assigns higher memorability scores when the \textbf{size of the object} (or animal) in the image is larger, as our model is in many cases zooming in further on the object with every increase of $\alpha$.
Second it is \textbf{centering} the subject in the image frame.
Third, it seems to strive for \textbf{square} or \textbf{circular} shapes in classes where it is realistic to do so (e.g., snake, cheeseburger, necklace, and espresso in Figure~\ref{fig:examples1}).
Fourth, it is often \textbf{simplifying} the image from low to high $\alpha$, by reducing the clutter and/or number of objects, such as in the cheeseburger or flamingo, or by making the background more homogeneous, as in the snake example (see Figure~\ref{fig:examples1}).
A fifth observation is that the subject's \textbf{eyes} sometimes become more pronounced and expressive, in particular in the dog classes (see Figure~\ref{fig:teaser}).
Sixth, one can also detect color changes between the different $\alpha$ conditions. Positive $\alpha's$ often produce \textbf{brighter} and more \textbf{colorful} images, and negative $\alpha's$ often produce darker images with dull colors. Finally, for those classes where multiple object hues can be considered realistic (e.g., the the bell pepper and the necklace in Figure~\ref{fig:teaser} and Figure~\ref{fig:examples1}), the model seems to prefer a \textbf{red} hue.
To verify our observations, we quantified the factors listed above for the images in the test set (except for "expressive eyes", which is more subjective and harder to quantify). Brightness was measured as the average pixel value after transforming the image to grayscale. For colorfulness, we used the metric proposed by \cite{colorfulness}, and for redness we computed the normalized number of red pixels. Finally, the entropy of the pixel intensity histogram was taken as proxy for simplicity. For the remaining three factors, a pretrained Mask R-CNN \cite{maskrcnn,massa2018mrcnn} was used to generate an instance-level segmentation mask of the subject. To capture object size, we calculated the difference in the mask's area (normalized number of pixels) as the step size $\alpha$ varied. To measure centeredness, we computed the deviation of the mask's centroid from the center of the frame. Finally, we calculated the length of minor and major axes of an ellipse that has the same normalized second central moments as the mask, and used their ratio as a metric of squareness. Figure~\ref{fig:memorability_graph}.B shows that the emerging factor scores increase with $\alpha$.
\subsection{Realness}\label{Realness}
\begin{figure}[b]
\centering
\includegraphics[width=9.2cm]{latex/realness.png}\qquad
\caption{\textbf{Realness measures as a function of $\alpha$}. A) Two popular automatic measures for evaluating the realness of a set of GAN images. Note that lower FID values indicate higher realness. B) Human fakeness discriminability, measured as the mean proportion correct in a 2AFC-task in which AMT workers had to discriminate GAN-images (fake) from real photographs.
\label{fig:2afc}
\end{figure}
While BigGAN achieves state-of-the-art to generate highly realistic images, there remains a certain variability in the ``realness" of the generated images. How best to evaluate the realness of a set of GAN-images is still an open question. Below, we discuss two automatically computed realness measures and a human measure in relation to our data. We discuss an additional human measure, based on a different task, in the supplementary materials.
\subsubsection{Automatic measures}
In Figure~\ref{fig:2afc}.A, we plot two popular automatic measures in function of $\alpha$: the Frechet Inception Distance (FID) \cite{FID} and the Inception Score (IS) \cite{IS}. A first observation is that the FID is below 40 in all $\alpha$ conditions. An FID as low as 40 already corresponds to reasonably realistic images. Thus the effects of our model's modifications on memorability are not explained by making the images unrealistic. But we do observe interesting differences in FID- and IS-differences related to $\alpha$, suggesting that more memorable images have more interpretable semantics.
\subsubsection{Human measure}
In addition to the two automatic measures, we conducted an experiment to collect human realness scores. The experiment consisted of a two-alternative forced choice (2AFC) task, hosted on Amazon Mechanical Turk (AMT), in which workers had to discriminate GAN-images from real ones. Workers were shown a series of pairs, consisting of one GAN-image and one real image. They were presented side by side for a duration of 1.6 s. Once a pair had disappeared off the screen, workers pressed the j-key when they thought the GAN-image was shown on the right, or the f-key when they thought it was shown on the left. The position of the GAN-image was randomized across trials. The set of real images used in this experiment was constructed by randomly sampling 10 real ImageNet exemplars per GAN-image class. The set of GAN-images was the same as the one quantified on memorability in Section ~\ref{memory}. A GAN-image was randomly paired with one of the 10 real images belonging to the same class. Each series consisted of 100 trials, of which 20 were vigilance trials. For the vigilance trials, we generated GAN-images from $\mathbf{z}$ vectors that were sampled from the tails of a normal distribution (to make them look less real). For a worker's first series, we prepended 20 trials with feedback as practice (not included in the analyses). Workers could complete up to 17 series, but were blocked if they scored less than 65\% correct on the vigilance trials. Series that failed this criterion were also excluded from the analyses. The pay rate equaled \$0.50 per completed series. On average, each of our test images was seen by 2.76 workers, meaning 4137 data points per $\alpha$ condition.
We did not observe differences in task performance between different $\alpha$ (see Figure ~\ref{fig:2afc}.B). Indeed, a logistic mixed-effects regression fitted to the raw, binary data (correct/incorrect) did not reveal a statistically significant regression weight for $\alpha$ ($\beta=-0.08,95\%CI=[-0.33,0.18],p = 0.55$). In other words, the model's image modifications did not affect workers' ability to correctly identify the fake image, indicating that perceptually, the image clones of a seed image did not differ in realness.
\begin{figure*}
\includegraphics[width=480pt,height=113pt]{latex/memory_game.png}
\caption{\textbf{Schematic of the visual memory game}. Each image is shown for 600 ms, with a blank interstimulus interval of 800 ms. Workers are asked to respond whenever they recognize a repeat of a previously shown image. For a correct response, the frame around the image briefly turns green. A red frame, on the other hand, indicates a mistake.}
\label{fig:memory_game}
\end{figure*}
\subsection{Do our changes causally affect memory?}\label{memory}
In addition to the MemNet scores, is our model also successful at changing the probability of an image being recognized by participants in an actual memory experiment?
We tested people's memory for the images of a test set (see Section~\ref{implementation}) using a repeat-detection visual memory game, which was hosted on AMT (see Figure~\ref{fig:memory_game}). \cite{Isola2014,Khosla_2015_ICCV}. AMT workers watched a series of one image at the time and had to press a key whenever they saw a repeat of a previously shown image. Each series consisted of 215 images, shown each for 600 ms with a blank interstimulus interval of 800 ms. Sixty images were targets, sampled from our test set, and repeated after 34 to 139 intervening images. The remaining images were either filler or vigilance images and were sampled from a separate set. This set was created with 10 $\mathbf{z}$ vectors per class and the same five $\alpha$ values as the test set: $[-0.2,-0.1,0,0.1,0.2]$, making a total of 37.5K images. Filler images were only presented once and ensured spacing between a target and its repeat. Vigilance images were presented twice, with 0 to 3 intervening images in-between the two presentations. The vigilance repeats constituted easy trials to keep workers attentive. Care was taken to ensure that a worker never saw more than one $G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y})$ for a given $\mathbf{z}$. Workers could complete up to 25 series, but were blocked if they missed more than 55\% of the vigilance repeats in a series or made more than 30\% false positives. Series that failed this were excluded from the analyses. The pay rate was \$0.50 per completed series. On average, a test image was seen by 3.16 workers, with 4740 data points per $\alpha$ condition.
Workers could either recognize a repeated test image (hit, 1), or miss it (miss, 0).
Figure~\ref{fig:experiments_plot}.A shows the hit rate across all images and workers. The hit rate increases with every step of $\alpha$. Fitting a logistic mixed-effects regression model to the raw, binary data (hit/miss), we found that the predicted log odds of image being recognized increase with 0.19 for an increase in $\alpha$ of 0.01 ($\beta = 1.92, 95\%CI=[1.71-2.12],p<0.001$). This shows that our model can successfully navigate the BigGAN latent space in order to make an image more (or less) memorable to humans.
\begin{figure}[!h]
\centering
\includegraphics[width=12.45cm]{latex/memory_performance.png}
\caption{\textbf{Human memory performance} for images modified according to different Assessors: A) MemNet, B) Object size and C) AestheticsNet. Performance is measured as the hit rate across all images and workers in the memory game for each property.}
\label{fig:experiments_plot}
\end{figure}
\subsubsection{Emerging factors}
Given human memory data for images modified for memorability, we evaluate how the images' emerging factor scores relate to their likelihood of being recognized. We fitted mixed-effects logistic regression models, each with a different emerging factor as the predictor, see Table ~\ref{table:emerging_factors}. Except for entropy, all the emerging factors show a significant, positive relation to the likelihood of a hit in the memory game, but none fit the data as well as the model's $\alpha$. This indicates that a single emerging factor is not enough to fully explain the effect observed in Figure ~\ref{fig:experiments_plot}.A. Note that the emerging factor results are correlational and the factors are intercorrelated. This makes it hard to draw conclusions about which individual factors truly causally affect human memory performance. As an example of how this can be addressed within the GANalyze framework, we conducted an experiment focusing on the effect of one salient emerging factor: \textbf{object size}. As seen in Figure~\ref{fig:examples1}, more memorable images tend to center and enlarge the object class.
\begin {table}
\begin{center}
\begin{tabular}{
l|
S[table-format = -1.2, table-space-text-post = $^{***}$]|
>{{[}}
S[table-format = -1.2,table-space-text-pre={[}]
@{,\,}
S[table-format = -1.2,table-space-text-post={]}]
<{{]}}|
S[table-format = <2.2]|
S[table-format = 1.3]}
\toprule
{Factor} & {Log Odds} & \multicolumn{2}{c|}{CI} & {$p$} & {Tjur's D}\\
\midrule
Brightness & 0.28 & 0.24 & 0.32 & <0.001 &0.066\\
Centeredness & 0.24 & 0.19 & 0.29 & <0.001 &0.059\\
Colorfulness & 0.17 & 0.14 & 0.21 & <0.001 &0.054\\
Entropy& 0.03 & -0.04 & 0.10 & 0.441 & 0.062\\
Redness& 0.06 & 0.00 & 0.12 & 0.042 & 0.055\\
Shape & 0.19 & 0.14 & 0.24 & <0.001&0.060\\
Object size & 0.32 & 0.27 & 0.37 & < 0.001 &0.050 \\
\midrule
$\alpha$ & 1.92 & 1.71 & 2.12 & <0.001 & 0.074\\
\bottomrule
\end{tabular}
\vspace{5pt}
\caption{
\small{\textbf{Relation between emerging factors and human memory performance.} We show the output of logistic mixed-effects regressions. From left to right: the regression weight, the confidence interval (CI) for that weight, the $p$-value for statistical significance, and Tjur's coefficient of discrimination (D), being the regression model's goodness of fit \cite{tjur}. The emerging factor values were normalized before running the regression models.}
}
\end{center}
\label{table:emerging_factors}
\end{table}
We trained a version of our model with an Object size Assessor, instead of the MemNet Assessor. This is the same Object size Assessor used to quantify the object size in the images modified according to MemNet (e.g., for the results in Figure ~\ref{fig:memorability_graph}.B), now teaching the Transformer to perform ``enlarging" modifications. After training with 161750 $\mathbf{z}$ vectors, we generated a test set as described in Section ~\ref{implementation}, except that we used a different set of $\alpha$'s: $[-0.8,-0.4,0,0.4,0.8]$. We chose these values to qualitatively match the degree of object size changes achieved by the MemNet version of the model. Figure ~\ref{fig:aesthetic_validation}.A visualizes the results achieved on the test set. The model successfully enlarges the object with increasing alpha's, as confirmed by a linear mixed-effects regression analysis ($\beta=0.07,95\%CI=[0.06,0.07],p <0.001$). Figure~\ref{fig:comparison} shows example images generated by that model, after having been trained with 161750 $\mathbf{z}$ vectors. A comparison with images modified according to MemNet suggests that the latter model was doing more than just enlarging the object.
To study how the new size modifications affect memorability, we generated a new set of images (7.5K targets, 37.5K fillers) with $\alpha$'s $[-0.8,-0.4,0,0.4,0.8]$. The new images were then quantified using the visual memory game (on average 2.36 data points per image and 3540 per $\alpha$ condition). Figure ~\ref{fig:experiments_plot}.B shows the results. Memory performance increases with $\alpha$, as confirmed by a logistic mixed-effects analysis ($\beta = 0.11, 95\%CI=[0.06,0.18],p<0.001$, although mostly for positive $\alpha$ values.
\begin{figure}
\centering
\includegraphics[width=0.63\textwidth]{latex/other_assessor_validation.png}
\caption{\textbf{Model results for additional Assessors}. A) Graph shows the mean Object size Assessor score across images in every $\alpha$ condition. A different set of $\alpha$ values was chosen to qualitatively match the degree object size changes achieved by the model trained with the MemNet Assessor. B) Graph shows the mean AestheticsNet score across the images in every $\alpha$ condition. C) Graph shows the mean EmoNet score across the images in every $\alpha$ condition}
\label{fig:aesthetic_validation}
\end{figure}
\section{Other properties}\label{other}
As mentioned in Section~\ref{implementation}, the proposed method can be applied to other image properties, simply by substituting the Assessor module. To show our framework can generalize, we trained a model for aesthetics, using Kong et al's ~\cite{Kong_et_al_2016} CNN (hereinafter referred to as AestheticsNet) as the Assessor. In addition, we also trained a model for emotional valence. Emotional valence refers to the extent to which the emotions evoked by an image are experienced as positive (or negative). For this property, we trained our own Assessor by fine-tuning a ResNet50 model ~\cite{resnet}, pretrained on the Moments database ~\cite{monfort2019moments}, to the Cornell Emotion6 Image Database ~\cite{emotion6}. We refer to this Assessor as EmoNet. Finally, we generated a test set for each of the two new models, like we did for the memorability model.
Figure~\ref{fig:aesthetic_validation}.B shows the average AestheticsNet scores per $\alpha$ condition. The scores significantly increase with $\alpha$, as evidenced by the results of a linear mixed-effects regression ($\beta = 0.72, 95\% CI = [0.70,0.74],p<0.001$). We can successfully train the model to increase (or decrease) an image's aesthetic score as shown in Figure~\ref{fig:examples_aes1} (left) and Figure~\ref{fig:sup_examples_aes_1}. Similarly, Figure~\ref{fig:aesthetic_validation}.C shows the average EmoNet scores per $\alpha$ condition. Here too, the scores significantly increase with $\alpha$ ($\beta = 0.44, 95\% CI = [0.43,0.45],p<0.001$). Example visualizations generated by this model are presented in Figure~\ref{fig:examples_aes1} (right) and Figure~\ref{fig:sup_examples_emo_1}.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.37\textwidth]{latex/examples_aes_0.pdf}
\qquad
\qquad
\includegraphics[width=0.37\textwidth]{latex/examples_emo_0.pdf}
\caption{\textbf{Examples of generated images} along the aesthetics dimension (left) and the emotional valence dimension (right). Each middle column represents $G(\mathbf{z},\mathbf{y})$, the generated image serving as the original seed to create a series of clone images scoring higher or lower on the respective dimension according to the Assessor. The images' Assessor scores are presented in their top left corner.}
\label{fig:examples_aes1}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.9\textwidth]{latex/attribute_comparison_4_3.pdf}
\caption{\textbf{Comparison of examples generated according to different Assessors.} The top row represents $G(\mathbf{z},\mathbf{y})$, the generated image serving as the original seed to create a series of images with a higher or lower Assessor value. The respective Assessor values are indicated in the top left corner. Note that for object size, we used a different $\alpha$ range: \{-0.8,0.8\}.}
\label{fig:comparison}
\end{figure}
Based on a qualitative inspection of such visualizations, we observed that the aesthetics model is modifying factors like depth of field, color palette, and lighting, suggesting that the AestheticsNet is sensitive to those factors. Indeed, the architecture of the AestheticsNet includes attribute-adaptive layers to predict these factors, now highlighted by our visualizations. The emotional valence model often averts the subject's gaze away from the "camera" when decreasing valence. To increase valence, it often makes images more colorful, introduces bokeh, and makes the skies more blue in landscape images. Finally, the teddy bear in Figure~\ref{fig:teaser} (right) seems to smile more. Interestingly, the model makes different modifications for every property (see Figure~\ref{fig:comparison}), suggesting that what makes an image memorable is different from what makes it aesthetically pleasing or more positive in its emotional valence.
A final question we asked is whether an image modified to become more (less) aesthetic also becomes more (less) memorable? To test this, we quantified the images of the aesthetic test set on memorability by presenting them to workers in the visual memory game (we collected 1.54 data points per image and 2306 data points per $\alpha$ condition). Figure~\ref{fig:experiments_plot}.C shows the human memory performance in function of an $\alpha$ that is tuning aesthetics. A logistic mixed-effects regression revealed that with an 0.1 increase in the aesthetics $\alpha$, the predicted log odds of an image being recognized increase with 0.07 ($\beta = 0.72, 95\% CI = [0.44,1.00],p<0.001$). While modifying an image to make it more aesthetic does increase its memorability, the effect is rather small, suggesting that memorability is more than only aesthetics and that our model was right to modify memorability and aesthetics in different ways.
\section{Conclusion}
We introduce GANalyze, a framework that shows how a GAN-based model can be used to visualize what another model (i.e. CNN as an Assessor) has learned about its target image property. Here we applied it to memorability, yielding a kind of ``visual definition" of this high-level cognitive property, where we visualize what it looks like for an image to become more or less memorable. These visualizations surface multiple candidate features that may help explain why we remember what we do. Importantly, our framework can also be generalized to other image properties, such as aesthetics or emotional valence: by replacing the Assessor module, the framework allows us to explore the visual definition for any property we can model as a differentiable function of the image. We validated that our model successfully modified GAN images to become more (or less) memorable via a behavioral human memory experiment on manipulated images.
GANalyze's intended use is to contribute to the scientific understanding of otherwise hard to define cognitive properties. Note that this was achieved by modifying images for which the encoding into the latent space of the GAN was given. In other words, it is currently only possible to modify seed images that are GAN-images themselves, not user-supplied, real images. However, should advances in the field lead to an encoder network, this would become possible and it would open applications in graphics and education, for example, where selected images can be made more memorable. One should also be wary, though, of potential misuse, especially when applied to images of people or faces. Note that the BigGAN~\cite{BigGAN} generator used here was trained on ImageNet categories~\cite{Imagenet} which only occasionally include people, and that it does not allow to render realistically looking people. Nevertheless, with generative models yielding ever more realistic output, an increasingly important challenge in the field is to develop powerful detection methods to allow us to reliably distinguish generated, fake images from real ones~\cite{washington_post}\cite{technology_review}\cite{hill}.
\section{Acknowledgments}
This work was partly funded by NSF award 1532591 in Neural and Cognitive Systems (to A.O), by a fellowship (Grant 1108116N) and a travel grant (Grant V4.085.18N) awarded to Lore Goetschalckx by the Research Foundation - Flanders (FWO).
\bibliographystyle{unsrt}
\section{Introduction}
Why do we remember the things we do? Decades of work have provided numerous explanations: we remember things that are out of context \cite{zoya,standing}, that are emotionally salient \cite{emotions}, that involve people~\cite{Isola2014}, etc. But a picture is, as they say, worth a thousand words. What does it \emph{look like} to make an image more or less memorable? The same questions can be asked for many cognitive visual properties: what visual changes can take a bland foggy seascape and add just the right colors and tones to make it serenely beautiful.
Attributes like memorability, aesthetics, and emotional valence are of special interest because we do not have concrete definitions of what they entail. This contrasts with attributes like ``object size" and "smile". We know exactly what it means to zoom in on a photo, and it's easy to imagine what a face looks like as it forms a smile. It's an open question, on the other hand, what exactly do changes in ``memorability" look like? Previous work has built powerful predictive models of image memorability~\cite{Isola2014,Khosla_2015_ICCV} but these have fallen short of providing a fine-grained visual explanation of what underlies the predictions.
In this paper, we propose a new framework, GANalyze, based on Generative Adversarial Networks (GAN)~\cite{goodfellow}, to study the visual features and properties that underlie high-level cognitive attributes. We focus on image memorability as a case study, but also show that the same methods can be applied to study image aesthetics and emotional valence.
Our approach leverages the ability of GANs to generate a continuum of images with fine-grained differences in their visual attributes. We can learn how to navigate the GAN's latent space to produce images that have increasing or decreasing memorability, according to an off-the-shelf memorability predictor~\cite{Khosla_2015_ICCV}. Starting with a seed image, this produces a sequence of images of increasing and decreasing predicted memorability (see Figure 1). By showing this visualization for a diverse range of seed images, we come up with a catalog of different image sequences showcasing a variety of visual effects related to memorability. We call this catalog a \emph{visual definition} of image memorability. GANalyze thereby offers an alternative to the non-parametric approach in which real images are simply sorted on their memorability score to visualize what makes them memorable (example shown in Figure~\ref{fig:nonparametric}). The parametric, fine-grained visualizations generated by GANalyze provide much clearer visual definitions.
These visualizations surface several correlates of memorability that have been overlooked by prior work, including ``object size", ``circularity", and ``colorfulness". Most past work on modeling image memorability focused on semantic attributes, such as object category (e.g., ``people" are more memorable than ``trees")~\cite{Isola2014}. By applying our approach to a class-conditional GAN, BigGAN~\cite{BigGAN}, we can restrict it to only make changes that are orthogonal to object class. This reveals more fine-grained changes that nonetheless have large effects on predicted memorability. For example, consider the cheeseburgers in Figure~\ref{fig:examples1}. Our model visualizes more memorable cheeseburgers as we move to the right. The apparent changes go well beyond semantic category -- the right-most burger is brighter, rounder, more canonical, and, we think, looks tastier.
Since our visualizations are learned based on a \emph{model} of memorability, a critical step is to verify that what we are seeing really has a causal effect on human behavior. We test this by running a behavioral experiment that measures the memorability of images generated by our GAN, and indeed we find that our manipulations have a causal effect: navigating the GAN manifold toward images that are predicted to be more memorable actually results in generating images that are measurably more memorable in the behavioral experiment.
Our contributions include the following:
\begin{itemize}
\item Introducing GANalyze, a framework that uses GANs to provide a \emph{visual definition} of image properties, like memorability and aesthetics, that we can measure but are not easy, in words, to define.
\item Showing that this framework surfaces previously overlooked attributes that correlate with memorability.
\item Demonstrating that the discovered transformations have a causal effect on memorability.
\item Showing that GANalyze can be applied to provide visual definitions for aesthetics and emotional valence.
\end{itemize}
\subsection{Related work}
{\bf Generative Adversarial Networks or GANs}.
GANs ~\cite{goodfellow} introduced a revolutionary framework to synthesize natural-looking images ~\cite{stylegan,BigGAN,sagan,proggan,BigGAN}. Among the many applications for GANs are style transfer~\cite{Zhu_2017_ICCV}, visual prediction~\cite{mathieu2015deep}, and ``sim2real" domain adaptation~\cite{graspgan}. Here, we show how they can also be applied to the problem of understanding high-level, cognitive image properties, such as memorability.
{\bf Understanding CNN representations}
The internal representations of a CNN can be unveiled using methods like network dissection \cite{zhou2018,bau2017,objectdetectors} including for a CNN trained on memorability \cite{Khosla_2015_ICCV}. For instance, Khosla et al. \cite{Khosla_2015_ICCV} showed that units with strong positive correlations with memorable images specialized for people, faces, body parts, etc., while those with strong negative correlations where more sensitive to large regions in landscapes scenes. Here, our framework introduces a new way of defining what memorability, and aesthetic, variability look like.
{\bf Modifying Memorability}. The memorability of an image, like faces, can be manipulated using warping techniques \cite{Khosla_2013_ICCV}. Concurrent work has also explored using a GAN for this purpose \cite{sidorov}. Another approach is a deep style transfer ~\cite{deepstyle} which taps into more artistic qualities. Now that GANs have reached a quality that is often almost indistinguishable from real images, they offer a powerful tool to synthesize images with different cognitive qualities. As shown here, our GANalyze framework successfully modified GAN-generated images across a wide range of image categories to produce a second generation of GAN realistic photos with different mnemonic qualities.
\begin{figure*}[b]
\includegraphics[width=\textwidth]{latex/method_schematic.pdf}\qquad
\caption{\textbf{Schematic of the model}. The model learns how to transform a $\mathbf{z}$ vector such that when fed to a Generator, the resulting image's property of interest changes. The transformation is achieved by the \emph{Transformer}, who moves the $\mathbf{z}$ vector along a learned direction, $\mathbf{\theta}$, in the Generator's latent space. The property of interest (e.g., memorability) is predicted by an \emph{Assessor module} (e.g., MemNet). Finally, $\alpha$ acts as knob to set the degree of change one wants to achieve in the Assessor value (e.g., MemNet score). It tells the \emph{Transformer} how far exactly to move along $\mathbf{\theta}$.}
\label{fig:model}
\end{figure*}
\section{Model}
\subsection{Formulation}
We start with a pretrained Generator $G$, who takes a noise vector $\mathbf{z} \in \mathbb{R}^{1\times M}$ and a one-hot class vector $\mathbf{y} \in \{0;1\}^{1\times C}$ as input and generates a photo-realistic image $G(\mathbf{z},\mathbf{y})$. Assumed is also an Assessor function $A$ that assesses an image property of interest, in this case memorability. Our goal was to learn to transform any given noise vector $\mathbf{z}$ of any class $\mathbf{y}$ such that the memorability of its resulting, generated image increases (or decreases) with a certain amount $\alpha$. The transformation is achieved by a Transformer function, who moves the input $\mathbf{z}$ along a certain direction $\mathbf{\theta} \in \mathbb{R}^{1\times M}$ in the latent space. We express the objective as:
\begin{equation}
\mathcal{L}(\theta)=\mathbb{E}_{\mathbf{z},\mathbf{y},\alpha}[(A(G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y}))\\
-(A(G(\mathbf{z},\mathbf{y}))+\alpha))^2]
\end{equation}
Note that this is simply the MSE loss between the target memorability score, i.e. the seed image's score $A(G(\mathbf{z},\mathbf{y}))$ increased by $\alpha$, and the memorability score of the transformed clone image $A(G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y}))$. The scalar $\alpha$ acts as a metaphorical knob with which one can use to turn up or turn down memorability. The optimizing problem is $\theta^* = \argminD_{\theta} \mathcal{L}(\theta)$.
The Transformer $T$ is defined as:
\begin{equation}
\begin{split}
T_{\theta}(\mathbf{z},\alpha) &= \mathbf{z} + \alpha \mathbf{\theta}\\
\end{split}
\end{equation}
Figure~\ref{fig:model} presents a schematic of the model. Finally, note that when $\alpha = 0$, $T$ becomes a null operation and $G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y})$ then equals $G(\mathbf{z},\mathbf{y})$.
\subsection{Implementation}\label{implementation}
For the results presented here, we used the Generator of BigGAN \cite{BigGAN}, which generates state-of-the art GAN images and is pretrained on ImageNet \cite{Imagenet}. The Assessor was implemented as MemNet \cite{Khosla_2015_ICCV}, a CNN predicting image memorability. Note, however, that training our model with different Generators or different Assessors can easily be achieved by substituting the respective modules. We discuss an Assessor for image aesthetics in Section~\ref{other}. Furthermore, we present additional results for implementations with a StyleGAN \cite{stylegan} Generator in the supplementary materials.
To train our model and find $\theta^*$, we built a training set by randomly sampling 400K $\mathbf{z}$ vectors from a standard normal distribution truncated to the range $[-2,2]$. Each $\mathbf{z}$ was accompanied by an $\alpha$ value, randomly drawn from a uniform distribution between -0.5 and 0.5, and a randomly chosen $\mathbf{y}$. We used a batch size of 4 and an Adam optimization procedure.
In view of the behavioral experiments (see Section~\ref{experiments}), we restricted the test set to 750 randomly chosen ImageNet classes and two $\mathbf{z}$ vectors per class. Each $\mathbf{z}$ vector was then paired with five different $\alpha$ values: ${[-0.2,-0.1,0,0.1,0.2]}$. Note that this includes an $\alpha$ of 0, representing the original image $G(\mathbf{z},\mathbf{y})$. Finally, the test set consisted of 1.5K sets of five images, or 7.5K test images in total.
\section{Experiments}\label{experiments}
\subsection{Model validation}
Did our model learn to navigate the latent space such that it can increase (or decrease) the Assessor score of the generated image with positive (or negative) $\alpha$ values?
Figure~\ref{fig:memorability_graph}.A suggests the model learned. The mean MemNet score of test set images increases with every increment of $\alpha$. To test this formally, we fitted a linear mixed-effects regression model to the data and found a (unstandardized) slope ($\beta$) of 0.68 ($95\% CI = [0.66,0.70], p<0.001)$, confirming that the Memnet score increases significantly with $\alpha$.
\begin{figure}[b]
\centering
\includegraphics[width=0.95\textwidth]{latex/memnet_validation3.png}
\caption{\textbf{Model results}. A) Graph shows the mean MemNet score across the images in every $\alpha$ condition. Our model successfully learned how to modify a GAN image to decrease (negative $\alpha$) or increase (positive $\alpha$) its MemNet score. B) List of emerging factors potentially underlying the effect observed in (A), and graph of how they change in function of $\alpha$. The factors emerged from visualizations generated by the GANalyze framework (examples shown in Figures~\ref{fig:examples1}, ~\ref{fig:sup_examples_mem_1}, and ~\ref{fig:sup_examples_mem_2}). Emerging factor scores were first normalized and then averaged per $\alpha$ condition.}
\label{fig:memorability_graph}
\end{figure}
\begin{figure*}[!p]
\centering
\includegraphics[height=0.98\dimexpr\textheight\relax]{latex/examples_memnet_0.pdf}
\caption{\textbf{Examples of generated images} along the memorability dimension. The middle column represents $G(\mathbf{z},\mathbf{y})$, the generated image serving as the original seed to create a series of clone images more or less memorable.}
\label{fig:examples1}
\end{figure*}
\subsection{Emerging factors}\label{Emerging factors}
We observe that the model can successfully change the memorability of an image, given its $\mathbf{z}$ vector. Next, we ask which image factors it altered to achieve this. The answer to this question can provide further insight into what the Assessor has learned about the to-be-assessed image property, in this case what MemNet has learned about memorability. From a qualitative analysis of the test set (examples shown in Figures ~\ref{fig:examples1}, ~\ref{fig:sup_examples_mem_1}, and ~\ref{fig:sup_examples_mem_2}), a number of candidate factors stand out.
First, MemNet assigns higher memorability scores when the \textbf{size of the object} (or animal) in the image is larger, as our model is in many cases zooming in further on the object with every increase of $\alpha$.
Second it is \textbf{centering} the subject in the image frame.
Third, it seems to strive for \textbf{square} or \textbf{circular} shapes in classes where it is realistic to do so (e.g., snake, cheeseburger, necklace, and espresso in Figure~\ref{fig:examples1}).
Fourth, it is often \textbf{simplifying} the image from low to high $\alpha$, by reducing the clutter and/or number of objects, such as in the cheeseburger or flamingo, or by making the background more homogeneous, as in the snake example (see Figure~\ref{fig:examples1}).
A fifth observation is that the subject's \textbf{eyes} sometimes become more pronounced and expressive, in particular in the dog classes (see Figure~\ref{fig:teaser}).
Sixth, one can also detect color changes between the different $\alpha$ conditions. Positive $\alpha's$ often produce \textbf{brighter} and more \textbf{colorful} images, and negative $\alpha's$ often produce darker images with dull colors. Finally, for those classes where multiple object hues can be considered realistic (e.g., the the bell pepper and the necklace in Figure~\ref{fig:teaser} and Figure~\ref{fig:examples1}), the model seems to prefer a \textbf{red} hue.
To verify our observations, we quantified the factors listed above for the images in the test set (except for "expressive eyes", which is more subjective and harder to quantify). Brightness was measured as the average pixel value after transforming the image to grayscale. For colorfulness, we used the metric proposed by \cite{colorfulness}, and for redness we computed the normalized number of red pixels. Finally, the entropy of the pixel intensity histogram was taken as proxy for simplicity. For the remaining three factors, a pretrained Mask R-CNN \cite{maskrcnn,massa2018mrcnn} was used to generate an instance-level segmentation mask of the subject. To capture object size, we calculated the difference in the mask's area (normalized number of pixels) as the step size $\alpha$ varied. To measure centeredness, we computed the deviation of the mask's centroid from the center of the frame. Finally, we calculated the length of minor and major axes of an ellipse that has the same normalized second central moments as the mask, and used their ratio as a metric of squareness. Figure~\ref{fig:memorability_graph}.B shows that the emerging factor scores increase with $\alpha$.
\subsection{Realness}\label{Realness}
\begin{figure}[b]
\centering
\includegraphics[width=9.2cm]{latex/realness.png}\qquad
\caption{\textbf{Realness measures as a function of $\alpha$}. A) Two popular automatic measures for evaluating the realness of a set of GAN images. Note that lower FID values indicate higher realness. B) Human fakeness discriminability, measured as the mean proportion correct in a 2AFC-task in which AMT workers had to discriminate GAN-images (fake) from real photographs.
\label{fig:2afc}
\end{figure}
While BigGAN achieves state-of-the-art to generate highly realistic images, there remains a certain variability in the ``realness" of the generated images. How best to evaluate the realness of a set of GAN-images is still an open question. Below, we discuss two automatically computed realness measures and a human measure in relation to our data. We discuss an additional human measure, based on a different task, in the supplementary materials.
\subsubsection{Automatic measures}
In Figure~\ref{fig:2afc}.A, we plot two popular automatic measures in function of $\alpha$: the Frechet Inception Distance (FID) \cite{FID} and the Inception Score (IS) \cite{IS}. A first observation is that the FID is below 40 in all $\alpha$ conditions. An FID as low as 40 already corresponds to reasonably realistic images. Thus the effects of our model's modifications on memorability are not explained by making the images unrealistic. But we do observe interesting differences in FID- and IS-differences related to $\alpha$, suggesting that more memorable images have more interpretable semantics.
\subsubsection{Human measure}
In addition to the two automatic measures, we conducted an experiment to collect human realness scores. The experiment consisted of a two-alternative forced choice (2AFC) task, hosted on Amazon Mechanical Turk (AMT), in which workers had to discriminate GAN-images from real ones. Workers were shown a series of pairs, consisting of one GAN-image and one real image. They were presented side by side for a duration of 1.6 s. Once a pair had disappeared off the screen, workers pressed the j-key when they thought the GAN-image was shown on the right, or the f-key when they thought it was shown on the left. The position of the GAN-image was randomized across trials. The set of real images used in this experiment was constructed by randomly sampling 10 real ImageNet exemplars per GAN-image class. The set of GAN-images was the same as the one quantified on memorability in Section ~\ref{memory}. A GAN-image was randomly paired with one of the 10 real images belonging to the same class. Each series consisted of 100 trials, of which 20 were vigilance trials. For the vigilance trials, we generated GAN-images from $\mathbf{z}$ vectors that were sampled from the tails of a normal distribution (to make them look less real). For a worker's first series, we prepended 20 trials with feedback as practice (not included in the analyses). Workers could complete up to 17 series, but were blocked if they scored less than 65\% correct on the vigilance trials. Series that failed this criterion were also excluded from the analyses. The pay rate equaled \$0.50 per completed series. On average, each of our test images was seen by 2.76 workers, meaning 4137 data points per $\alpha$ condition.
We did not observe differences in task performance between different $\alpha$ (see Figure ~\ref{fig:2afc}.B). Indeed, a logistic mixed-effects regression fitted to the raw, binary data (correct/incorrect) did not reveal a statistically significant regression weight for $\alpha$ ($\beta=-0.08,95\%CI=[-0.33,0.18],p = 0.55$). In other words, the model's image modifications did not affect workers' ability to correctly identify the fake image, indicating that perceptually, the image clones of a seed image did not differ in realness.
\begin{figure*}
\includegraphics[width=480pt,height=113pt]{latex/memory_game.png}
\caption{\textbf{Schematic of the visual memory game}. Each image is shown for 600 ms, with a blank interstimulus interval of 800 ms. Workers are asked to respond whenever they recognize a repeat of a previously shown image. For a correct response, the frame around the image briefly turns green. A red frame, on the other hand, indicates a mistake.}
\label{fig:memory_game}
\end{figure*}
\subsection{Do our changes causally affect memory?}\label{memory}
In addition to the MemNet scores, is our model also successful at changing the probability of an image being recognized by participants in an actual memory experiment?
We tested people's memory for the images of a test set (see Section~\ref{implementation}) using a repeat-detection visual memory game, which was hosted on AMT (see Figure~\ref{fig:memory_game}). \cite{Isola2014,Khosla_2015_ICCV}. AMT workers watched a series of one image at the time and had to press a key whenever they saw a repeat of a previously shown image. Each series consisted of 215 images, shown each for 600 ms with a blank interstimulus interval of 800 ms. Sixty images were targets, sampled from our test set, and repeated after 34 to 139 intervening images. The remaining images were either filler or vigilance images and were sampled from a separate set. This set was created with 10 $\mathbf{z}$ vectors per class and the same five $\alpha$ values as the test set: $[-0.2,-0.1,0,0.1,0.2]$, making a total of 37.5K images. Filler images were only presented once and ensured spacing between a target and its repeat. Vigilance images were presented twice, with 0 to 3 intervening images in-between the two presentations. The vigilance repeats constituted easy trials to keep workers attentive. Care was taken to ensure that a worker never saw more than one $G(T_{\theta}(\mathbf{z},\alpha),\mathbf{y})$ for a given $\mathbf{z}$. Workers could complete up to 25 series, but were blocked if they missed more than 55\% of the vigilance repeats in a series or made more than 30\% false positives. Series that failed this were excluded from the analyses. The pay rate was \$0.50 per completed series. On average, a test image was seen by 3.16 workers, with 4740 data points per $\alpha$ condition.
Workers could either recognize a repeated test image (hit, 1), or miss it (miss, 0).
Figure~\ref{fig:experiments_plot}.A shows the hit rate across all images and workers. The hit rate increases with every step of $\alpha$. Fitting a logistic mixed-effects regression model to the raw, binary data (hit/miss), we found that the predicted log odds of image being recognized increase with 0.19 for an increase in $\alpha$ of 0.01 ($\beta = 1.92, 95\%CI=[1.71-2.12],p<0.001$). This shows that our model can successfully navigate the BigGAN latent space in order to make an image more (or less) memorable to humans.
\begin{figure}[!h]
\centering
\includegraphics[width=12.45cm]{latex/memory_performance.png}
\caption{\textbf{Human memory performance} for images modified according to different Assessors: A) MemNet, B) Object size and C) AestheticsNet. Performance is measured as the hit rate across all images and workers in the memory game for each property.}
\label{fig:experiments_plot}
\end{figure}
\subsubsection{Emerging factors}
Given human memory data for images modified for memorability, we evaluate how the images' emerging factor scores relate to their likelihood of being recognized. We fitted mixed-effects logistic regression models, each with a different emerging factor as the predictor, see Table ~\ref{table:emerging_factors}. Except for entropy, all the emerging factors show a significant, positive relation to the likelihood of a hit in the memory game, but none fit the data as well as the model's $\alpha$. This indicates that a single emerging factor is not enough to fully explain the effect observed in Figure ~\ref{fig:experiments_plot}.A. Note that the emerging factor results are correlational and the factors are intercorrelated. This makes it hard to draw conclusions about which individual factors truly causally affect human memory performance. As an example of how this can be addressed within the GANalyze framework, we conducted an experiment focusing on the effect of one salient emerging factor: \textbf{object size}. As seen in Figure~\ref{fig:examples1}, more memorable images tend to center and enlarge the object class.
\begin {table}
\begin{center}
\begin{tabular}{
l|
S[table-format = -1.2, table-space-text-post = $^{***}$]|
>{{[}}
S[table-format = -1.2,table-space-text-pre={[}]
@{,\,}
S[table-format = -1.2,table-space-text-post={]}]
<{{]}}|
S[table-format = <2.2]|
S[table-format = 1.3]}
\toprule
{Factor} & {Log Odds} & \multicolumn{2}{c|}{CI} & {$p$} & {Tjur's D}\\
\midrule
Brightness & 0.28 & 0.24 & 0.32 & <0.001 &0.066\\
Centeredness & 0.24 & 0.19 & 0.29 & <0.001 &0.059\\
Colorfulness & 0.17 & 0.14 & 0.21 & <0.001 &0.054\\
Entropy& 0.03 & -0.04 & 0.10 & 0.441 & 0.062\\
Redness& 0.06 & 0.00 & 0.12 & 0.042 & 0.055\\
Shape & 0.19 & 0.14 & 0.24 & <0.001&0.060\\
Object size & 0.32 & 0.27 & 0.37 & < 0.001 &0.050 \\
\midrule
$\alpha$ & 1.92 & 1.71 & 2.12 & <0.001 & 0.074\\
\bottomrule
\end{tabular}
\vspace{5pt}
\caption{
\small{\textbf{Relation between emerging factors and human memory performance.} We show the output of logistic mixed-effects regressions. From left to right: the regression weight, the confidence interval (CI) for that weight, the $p$-value for statistical significance, and Tjur's coefficient of discrimination (D), being the regression model's goodness of fit \cite{tjur}. The emerging factor values were normalized before running the regression models.}
}
\end{center}
\label{table:emerging_factors}
\end{table}
We trained a version of our model with an Object size Assessor, instead of the MemNet Assessor. This is the same Object size Assessor used to quantify the object size in the images modified according to MemNet (e.g., for the results in Figure ~\ref{fig:memorability_graph}.B), now teaching the Transformer to perform ``enlarging" modifications. After training with 161750 $\mathbf{z}$ vectors, we generated a test set as described in Section ~\ref{implementation}, except that we used a different set of $\alpha$'s: $[-0.8,-0.4,0,0.4,0.8]$. We chose these values to qualitatively match the degree of object size changes achieved by the MemNet version of the model. Figure ~\ref{fig:aesthetic_validation}.A visualizes the results achieved on the test set. The model successfully enlarges the object with increasing alpha's, as confirmed by a linear mixed-effects regression analysis ($\beta=0.07,95\%CI=[0.06,0.07],p <0.001$). Figure~\ref{fig:comparison} shows example images generated by that model, after having been trained with 161750 $\mathbf{z}$ vectors. A comparison with images modified according to MemNet suggests that the latter model was doing more than just enlarging the object.
To study how the new size modifications affect memorability, we generated a new set of images (7.5K targets, 37.5K fillers) with $\alpha$'s $[-0.8,-0.4,0,0.4,0.8]$. The new images were then quantified using the visual memory game (on average 2.36 data points per image and 3540 per $\alpha$ condition). Figure ~\ref{fig:experiments_plot}.B shows the results. Memory performance increases with $\alpha$, as confirmed by a logistic mixed-effects analysis ($\beta = 0.11, 95\%CI=[0.06,0.18],p<0.001$, although mostly for positive $\alpha$ values.
\begin{figure}
\centering
\includegraphics[width=0.63\textwidth]{latex/other_assessor_validation.png}
\caption{\textbf{Model results for additional Assessors}. A) Graph shows the mean Object size Assessor score across images in every $\alpha$ condition. A different set of $\alpha$ values was chosen to qualitatively match the degree object size changes achieved by the model trained with the MemNet Assessor. B) Graph shows the mean AestheticsNet score across the images in every $\alpha$ condition. C) Graph shows the mean EmoNet score across the images in every $\alpha$ condition}
\label{fig:aesthetic_validation}
\end{figure}
\section{Other properties}\label{other}
As mentioned in Section~\ref{implementation}, the proposed method can be applied to other image properties, simply by substituting the Assessor module. To show our framework can generalize, we trained a model for aesthetics, using Kong et al's ~\cite{Kong_et_al_2016} CNN (hereinafter referred to as AestheticsNet) as the Assessor. In addition, we also trained a model for emotional valence. Emotional valence refers to the extent to which the emotions evoked by an image are experienced as positive (or negative). For this property, we trained our own Assessor by fine-tuning a ResNet50 model ~\cite{resnet}, pretrained on the Moments database ~\cite{monfort2019moments}, to the Cornell Emotion6 Image Database ~\cite{emotion6}. We refer to this Assessor as EmoNet. Finally, we generated a test set for each of the two new models, like we did for the memorability model.
Figure~\ref{fig:aesthetic_validation}.B shows the average AestheticsNet scores per $\alpha$ condition. The scores significantly increase with $\alpha$, as evidenced by the results of a linear mixed-effects regression ($\beta = 0.72, 95\% CI = [0.70,0.74],p<0.001$). We can successfully train the model to increase (or decrease) an image's aesthetic score as shown in Figure~\ref{fig:examples_aes1} (left) and Figure~\ref{fig:sup_examples_aes_1}. Similarly, Figure~\ref{fig:aesthetic_validation}.C shows the average EmoNet scores per $\alpha$ condition. Here too, the scores significantly increase with $\alpha$ ($\beta = 0.44, 95\% CI = [0.43,0.45],p<0.001$). Example visualizations generated by this model are presented in Figure~\ref{fig:examples_aes1} (right) and Figure~\ref{fig:sup_examples_emo_1}.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.37\textwidth]{latex/examples_aes_0.pdf}
\qquad
\qquad
\includegraphics[width=0.37\textwidth]{latex/examples_emo_0.pdf}
\caption{\textbf{Examples of generated images} along the aesthetics dimension (left) and the emotional valence dimension (right). Each middle column represents $G(\mathbf{z},\mathbf{y})$, the generated image serving as the original seed to create a series of clone images scoring higher or lower on the respective dimension according to the Assessor. The images' Assessor scores are presented in their top left corner.}
\label{fig:examples_aes1}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.9\textwidth]{latex/attribute_comparison_4_3.pdf}
\caption{\textbf{Comparison of examples generated according to different Assessors.} The top row represents $G(\mathbf{z},\mathbf{y})$, the generated image serving as the original seed to create a series of images with a higher or lower Assessor value. The respective Assessor values are indicated in the top left corner. Note that for object size, we used a different $\alpha$ range: \{-0.8,0.8\}.}
\label{fig:comparison}
\end{figure}
Based on a qualitative inspection of such visualizations, we observed that the aesthetics model is modifying factors like depth of field, color palette, and lighting, suggesting that the AestheticsNet is sensitive to those factors. Indeed, the architecture of the AestheticsNet includes attribute-adaptive layers to predict these factors, now highlighted by our visualizations. The emotional valence model often averts the subject's gaze away from the "camera" when decreasing valence. To increase valence, it often makes images more colorful, introduces bokeh, and makes the skies more blue in landscape images. Finally, the teddy bear in Figure~\ref{fig:teaser} (right) seems to smile more. Interestingly, the model makes different modifications for every property (see Figure~\ref{fig:comparison}), suggesting that what makes an image memorable is different from what makes it aesthetically pleasing or more positive in its emotional valence.
A final question we asked is whether an image modified to become more (less) aesthetic also becomes more (less) memorable? To test this, we quantified the images of the aesthetic test set on memorability by presenting them to workers in the visual memory game (we collected 1.54 data points per image and 2306 data points per $\alpha$ condition). Figure~\ref{fig:experiments_plot}.C shows the human memory performance in function of an $\alpha$ that is tuning aesthetics. A logistic mixed-effects regression revealed that with an 0.1 increase in the aesthetics $\alpha$, the predicted log odds of an image being recognized increase with 0.07 ($\beta = 0.72, 95\% CI = [0.44,1.00],p<0.001$). While modifying an image to make it more aesthetic does increase its memorability, the effect is rather small, suggesting that memorability is more than only aesthetics and that our model was right to modify memorability and aesthetics in different ways.
\section{Conclusion}
We introduce GANalyze, a framework that shows how a GAN-based model can be used to visualize what another model (i.e. CNN as an Assessor) has learned about its target image property. Here we applied it to memorability, yielding a kind of ``visual definition" of this high-level cognitive property, where we visualize what it looks like for an image to become more or less memorable. These visualizations surface multiple candidate features that may help explain why we remember what we do. Importantly, our framework can also be generalized to other image properties, such as aesthetics or emotional valence: by replacing the Assessor module, the framework allows us to explore the visual definition for any property we can model as a differentiable function of the image. We validated that our model successfully modified GAN images to become more (or less) memorable via a behavioral human memory experiment on manipulated images.
GANalyze's intended use is to contribute to the scientific understanding of otherwise hard to define cognitive properties. Note that this was achieved by modifying images for which the encoding into the latent space of the GAN was given. In other words, it is currently only possible to modify seed images that are GAN-images themselves, not user-supplied, real images. However, should advances in the field lead to an encoder network, this would become possible and it would open applications in graphics and education, for example, where selected images can be made more memorable. One should also be wary, though, of potential misuse, especially when applied to images of people or faces. Note that the BigGAN~\cite{BigGAN} generator used here was trained on ImageNet categories~\cite{Imagenet} which only occasionally include people, and that it does not allow to render realistically looking people. Nevertheless, with generative models yielding ever more realistic output, an increasingly important challenge in the field is to develop powerful detection methods to allow us to reliably distinguish generated, fake images from real ones~\cite{washington_post}\cite{technology_review}\cite{hill}.
\section{Acknowledgments}
This work was partly funded by NSF award 1532591 in Neural and Cognitive Systems (to A.O), by a fellowship (Grant 1108116N) and a travel grant (Grant V4.085.18N) awarded to Lore Goetschalckx by the Research Foundation - Flanders (FWO).
\bibliographystyle{unsrt}
|
\section{Introduction}\label{intro}
The Mn$_3$Z (Z=Sn, Ge, Ga) compounds, known to be weak ferromagnetic at low temperatures\cite{tomiyoshi-1982,tomiyoshi-1983,kren-1970}, are in the focus of current research interest due to the exotic topological properties of their band structure. They show large anomalous Hall conductivity due to the nonvanishing Berry curvature induced by the non-collinear triangular ground state spin structure\cite{Kubler-2014,nakatsuji-2015,nayak-2016} and are
topological Weyl semimetals because of emerging Weyl nodes in the band structure near the Fermi level\cite{kubler-2017,Yang-2017}.
The Mn$_3$Z compounds are also possible candidates to replace the expensive IrMn based antiferromagnets in magnetic sensors based on the GMR effect\cite{hirohata-2017}.
The Mn$_3$Z compounds have three different structural phases: a hexagonal phase with DO$_{19}$ structure, a tetragonal phase with DO$_{22}$ structure and a cubic phase with a standard Heusler structure. These phases and the transition between them have been the subject of recent research\cite{zhang-2013,khmelevskyi-2016}. In this paper, we are going to investigate the magnetic properties in the hexagonal phase of these compounds.
The atomic positions in the DO$_{19}$ hexagonal phase are sketched in Fig.~\ref{fig:structure}. The Mn sites in each layer form a kagome lattice, i.e.\ a two-dimensional network of corner-sharing equilateral triangles. The atomic layers are shifted alternately with $\frac{2}{3}\left(\vec{a}+\vec{b}\right)$ and $\frac{1}{3}\left(\vec{a}+\vec{b}\right)$ where $\vec{a}$ and $\vec{b}$ are the primitive vectors of the kagome lattice. The unit cell marked by the black rhombus in Fig.~\ref{fig:structure} contains two layers built up from six manganese atoms and two non-magnetic Z atoms.
\begin{figure}[htb]
\includegraphics[width=0.9\linewidth]{geometry.pdf}
\caption{The atomic positions in the hexagonal phase of the Mn$_3$Z compounds: large and small circles, labelled by capital and small letters, denote sites in the atomic layers at $z=c/4$ and at $z=3c/4$, respectively, while grey and black circles stand in order for Mn and Z atoms. The rhombus encloses a possible unit cell of the system.}
\label{fig:structure}
\end{figure}
The magnetic structure of these materials was measured with polarized neutron diffraction experiments showing that the low-energy magnetic states of these compounds in the hexagonal phase are chiral antiferromagnetic (AFM) states\cite{tomiyoshi-1982,tomiyoshi-1983,kren-1970} as illustrated in Fig.~\ref{fig:spin_structure}, with a small distortion producing a tiny net magnetic moment. A simple spin-model analysis in Ref.\ \onlinecite{tomiyoshi-1982} proved the magnetocrystalline anisotropy to be the microscopic mechanism responsible for the weak ferromagnetic (WF) distortion, whereas the Dzyaloshinsky--Moriya (DM) interaction \cite{dzyaloshinskii-1958,moriya-1960} was shown to lift the chiral degeneracy of the $\Gamma_3$ and $\Gamma_5$ states.
\begin{figure}[htb]
\includegraphics[width=\linewidth]{spin_structure.pdf}
\medskip
\ $\Gamma_3$ \hskip 4.5cm $\Gamma_5$
\caption{Low-energy chiral magnetic structures of the Mn$_3$Z compounds.
}
\label{fig:spin_structure}
\end{figure}
The electronic and magnetic structure of these compounds was also investigated theoretically in terms of self-consistent field density functional calculations in the local density approximation (LDA) \cite{sandratskii-1996,Kubler-2014,zhang-2013,khmelevskyi-2016,mendive-tapia-2019}. In the hexagonal phase, the chiral $\Gamma_5$ state with a weak ferromagnetic (WF) distortion and a small net magnetic moment was found as ground state \cite{sandratskii-1996,Kubler-2014,zhang-2013}. In particular, Ref.\ \onlinecite{sandratskii-1996} discussed symmetry considerations and the role of orbital polarization on the formation of weak ferromagnetism in Mn$_3$Sn.
It should be noted that in Ref.\ \onlinecite{zhang-2013} the tetragonal phase of Mn$_3$Sn was found lower in energy than the hexagonal phase and the stability of the hexagonal phase according to experiments was attributed to structural disorder or off-stoichiometric compositions. The influence of these effects on the magnetic ordering in the Mn$_3$Ga alloy has been studied in Ref.\ \onlinecite{khmelevskyi-2016}.
In this work we present a detailed theoretical investigation of the magnetic ground state of the Mn$_3$Z alloys in the hexagonal phase. We employ the relativistic Screened Korringa--Kohn--Rostoker (SKKR) method\cite{szunyogh-1994,zeller-1995} to calculate the electronic structure and the magnetic properties. In particular, we set up a classical spin model with parameters obtained from the combination of the spin-cluster expansion (SCE) and the relativistic disordered local moment (RDLM) method\cite{szunyogh-2011}. Using the point-group symmetry of the lattice we determine the general parametric form of the exchange interaction matrices of a three-sublattice model and we provide the group theoretical reason behind the formation of the weak ferromagnetic state. By solving the spin model we quantify the weak ferromagnetic distortion in terms of the model parameters. Our results are clearly consistent with the original spin-model description of weak ferromagnetism in Mn$_3$Sn \cite{tomiyoshi-1982}, however, we exceed this approach by quantitative estimates on the weak ferromagnetic distortion being in fairly good agreement with the experiments. We also obtain the magnetic ground states of the Mn$_3$Z compounds from unconstrained self-consistent LDA calculations and investigate
the effect of spin-orbit coupling on the Mn and Z sites selectively. Finally, we give a hint to the effect of the induced moment at the Z site by performing constrained LDA calculations for the case of the Mn$_3$Ga alloy.
\section{Methods}\label{methods}
\subsection{Spin model}\label{sm}
In order to study the magnetic properties of the Mn$_3$Z alloys in the hexagonal phase we use a classical Heisenberg model for the Mn spins represented with a set of unit vectors $\{\vec{e}\}$ and neglect the effect of the induced spin moment on the Z sites. The spin model in second order of the spin variables is given by
\begin{equation}
H(\{\vec{e}\}) = \sum_{i} \vec{e}_i \mathbf K_{i} \vec{e}_i -\dfrac{1}{2}\sum_{i\neq j}\vec{e}_i \mathbf J_{ij} \vec{e}_j\;,
\label{Eq:Heisenberg}
\end{equation}
where the $i$ and $j$ indices are confined to the Mn sites, $\mathbf K_{i}$ are the second order on-site anisotropy matrices and $\mathbf J_{ij}$ are the tensorial exchange couplings. The exchange matrix can be decomposed as
\begin{equation}
\mathbf J_{ij} = J_{ij}\mathbf I + \frac{1}{2} \left( \mathbf J_{ij} - \mathbf J_{ij} ^T \right) + \frac{1}{2} \left( \mathbf J_{ij} + \mathbf J_{ij} ^T - 2 J_{ij}\mathbf I \right)\; ,
\end{equation}
where $\mathbf I$ is the unit matrix and $T$ denotes the transpose of a matrix. In the above decomposition $J_{ij} = \frac{1}{3}\mathrm{Tr} \mathbf J_{ij} $ defines the isotropic Heisenberg coupling between two spins, the antisymmetric part of the exchange tensor can be related to the Dzyaloshinsky--Moriya interaction,
\begin{equation}
\vec{e}_i\frac{1}{2}\left( \mathbf J_{ij} - \mathbf J_{ij} ^T \right)\vec{e}_j = \vec{D}_{ij} \left( \vec{e}_i \times \vec{e}_j \right),
\end{equation}
and the traceless symmetric part of $\mathbf J_{ij}$ corresponds to the two-site anisotropy.
From previous experimental\cite{tomiyoshi-1982,tomiyoshi-1983,kren-1970} and theoretical\cite{sticht-1989,sandratskii-1996} works it turns out that the ground state magnetic structure of the Mn$_3$Z compounds can be well described in terms of an effective spin model related to three Mn sublattices. This means that the A--a, B--b and C--c sublattice pairs (cf.\ Fig.\ \ref{fig:structure}) are strongly coupled ferromagnetically, ensuring that the corresponding Mn moments are parallel to each other. Consequently we only have to consider three independent sublattices to explore the low-energy magnetic configurations. As will be shown in Sec.\ \ref{sec:spinmodel} our calculated exchange interactions clearly support this observation, which leads to the following simplified Hamiltonian:
\begin{equation}
H = -\dfrac{1}{2}\sum_{\alpha,\beta=1}^{3}\vec{e}_\alpha \mathbf J_{\alpha\beta} \vec{e}_\beta\;,
\label{eq:H_sublatt}
\end{equation}
where the $\mathbf J_{\alpha\beta}$ matrices are the effective sublattice interactions. The $\mathbf J_{\alpha\beta}$ matrices can be related to the exchange matrices in Eq.\ \eqref{Eq:Heisenberg} as
\begin{equation}
\mathbf J_{\alpha\beta} = \sum_{n} \mathbf J_{0\alpha,n\beta} -2 \delta_{\alpha\beta} \mathbf K_\alpha \, ,
\label{eq:sum_sublattice}
\end{equation}
where the index $0$ stands for a fixed site in sublattice $\alpha$, while $n$ goes through the sites in sublattice $\beta$. Note that due to translation invariance the on-site anisotropy matrices at all sites in a given sublattice are identical, which explains the notation $\mathbf K_\alpha$ in Eq.\ \eqref{eq:sum_sublattice}.
By collecting the spin variables of the three sublattices into a nine-dimensional composite variable, $ \vec{e}=( \vec{e}_1, \vec{e}_2, \vec{e}_3)$ and the sublattice interactions into a composite matrix,
\begin{equation}
\mathbf J = \left(
\begin{array}{ccc}
\mathbf J_{11} & \mathbf J_{12} & \mathbf J_{13} \\
\mathbf J_{21} & \mathbf J_{22} & \mathbf J_{23} \\
\mathbf J_{31} & \mathbf J_{32} & \mathbf J_{33}
\end{array}
\right) \, ,
\label{eq:Jmat}
\end{equation}
Eq.~\eqref{eq:H_sublatt} can be rewritten into the simple form,
\begin{equation}
H = -\dfrac{1}{2} \vec{e} \, \mathbf J \, \vec{e}\; .
\label{eq:Hcomp}
\end{equation}
The structure of the matrices \eqref{eq:sum_sublattice} can be obtained by using the $D_{3h}$ point-group symmetry of the lattice. This is provided by the invariance of the energy of the spin system \eqref{eq:H_sublatt} against any point-group element $g \in D_{3h}$. Denoting the $9\times9$ matrix representation of $g$ by $\mathbf R_g$, this implies the relationships,
\begin{equation}
\mathbf J = \mathbf R_g^T \mathbf J \mathbf R_g\;.
\label{eq:symmetry}
\end{equation}
\begin{comment}
\begin{equation}
\mathbf J_{\alpha^\prime \beta^\prime} = \mathbf R_g^T \mathbf J_{\alpha \beta} \mathbf R_g\;.
\label{eq:symmerty}
\end{equation}
\end{comment}
Note that the permutation of the sublattices under the operation $g$ is also included in the representation $\mathbf R_g$.
Performing the corresponding analysis we obtain two different types of sublattice interaction matrices: the three sublattice-diagonal matrices are connected via the $C_3$ rotation, while the six sublattice off-diagonal matrices are related to each other either by $C_3$ rotation or by transposition. One representative element for each set is given by
\begin{equation}
\mathbf J_{AA} =
\left(
\begin{array}{ccc}
\frac{-K^{x} + 3K^{y}}{2} & 0 & 0 \\
0 & \frac{3K^{x} - K^{y}}{2} & 0 \\
0 & 0 & -K^{x}-K^{y}
\end{array}
\right)
\label{eq:JAA}
\end{equation}
and
\begin{equation}
\mathbf J_{BC} =
\left(
\begin{array}{ccc}
J+J^{x} & D & 0 \\- D & J+J^{y} & 0 \\0 & 0 & J-J^{x}-J^{y}
\end{array}
\right) \, ,
\label{eq:JBC}
\end{equation}
respectively.
The sublattice model has therefore six independent parameters: two sublattice-diagonal anisotropy constants, $K^{x}$ and $K^{y}$, two sublattice-off-diagonal anisotropy constants, $J^{x}$ and $J^{y}$, one Dzyaloshinsky--Moriya parameter, $D$, describing an effective DM vector parallel to the $z$ axis, and an isotropic coupling between different sublattices, $J$. Note that in principle there is an isotropic coupling parameter for the sublattice-diagonal matrices, but it only adds a constant to the energy, thus it has no effect on the magnetic ordering in the system.
\subsection{\emph{Ab initio} calculations} \label{abinitio}
We performed self-consistent electronic structure calculations for the Mn$_3$Z compounds in terms of the relativistic Screened Korringa--Kohn--Rostoker method\cite{szunyogh-1994,zeller-1995}.
The lattice constants of the different compounds were set to the experimental values shown in Table~\ref{table:lattice}.
\begin{table}[htb]
\caption{The experimental lattice constants for the Mn$_3$Z compounds}
\def1.2{1.2}
\begin{ruledtabular}
\begin{tabular}{c|ccc}
& Mn$_3$Sn\cite{zimmer-1971} & Mn$_3$Ge\cite{kadar-1971} & Mn$_3$Ga\cite{kren-1970}\\ \colrule
a$_\text{2d}$[\AA] & 5.665 & 5.36 & 5.36 \\
$c/a$ & 0.79982 & 0.80598& 0.807
\end{tabular}
\label{table:lattice}
\end{ruledtabular}
\end{table}
We used the local spin-density approximation parametrized according to Vosko \emph{et al.},\cite{vosko-1980} and we employed the atomic sphere approximation with an angular momentum cutoff of $\ell_\text{max}=2$. We used 16 energy points on a semicircular path on the upper complex half-plane for the energy integrations, and 144 points in the 2D Brillouin zone (2DBZ) for $k$-integrations.
In order to obtain the parameters of the tensorial Heisenberg model \eqref{Eq:Heisenberg}, we employed the spin-cluster expansion developed originally by Drautz and F\"ahnle\cite{drautz-2004,drautz-2005} combined with the relativistic disordered local moment method \cite{gyorffy-1985, staunton-2004,staunton-2006}. The RDLM method provides a first-principle description of a paramagnetic system according to the adiabatic decoupling of the electronic and spin degrees of freedom, while the SCE enables a systematic parametrization of the adiabatic energy surface. For the details of the SCE-RDLM method see Ref.~\onlinecite{szunyogh-2011}. In the SCE calculations we used 16 energy points on a semicircular path on the upper complex half-plane with approximately 20000 $k$ points in the 2DBZ near the Fermi energy.
\section{Results}\label{sec:results}
\subsection{Spin model parameters}\label{sec:spinmodel}
We applied the SCE-RDLM method to obtain the \emph{ab initio} spin model parameters in the paramagnetic phase of each compound. First, we discuss the isotropic couplings. In Table \ref{table:isotropic} we show the first five nearest neighbor interactions as visualized in Fig.~\ref{fig:isotrop_int}. For all the three compounds we find a similar structure of the isotropic interactions. The nearest neighbor interactions $J_1$ couple sites that belong to different sublattices and different atomic layers. They are antiferromagnetic (AFM) and the largest in magnitude among the isotropic couplings. By contrast, the third nearest neighbor out-of-plane interactions, $J_3$ and $J_3^\prime$, that connect sites in the same sublattice are strongly ferromagnetic. These interactions thus force to align the moments in the same sublattice irrespective of the atomic-layer positions. On top of this, the AFM first nearest neighbor couplings cause frustration on the kagome lattice and stabilize the triangular states as shown in Fig.\ \ref{fig:spin_structure}. The (in-plane) second nearest neighbor antiferromagnetic interactions, $J_2$, also contribute to the stabilization of the triangular spin structures, though the other type of these interactions, $J_2^\prime$,
is ferromagnetic, destabilizing the triangular state to some extent. By using the {\em ab initio} tensorial spin model we solved the Landau--Lifshitz--Gilbert equations at zero temperature (with damping term only) and for all the three compounds we indeed obtained a ground state close to the $\Gamma_5$ state. As we will demonstrate later, it is the Dzyaloshinsky--Moriya interaction which selects between the $\Gamma_3$ and $\Gamma_5$ spin states.
\begin{figure}[htb]
\includegraphics[width=0.9\linewidth]{interactions.pdf}
\caption{Schematic view of the first five nearest neighbor interactions. Different interactions for the same distance are denoted with and without primes. The $J_4$ interaction connects two sites in the same sublattice from neighboring unit cells shifted along the $z$ axis, thus, we could not illustrate it in the figure.}
\label{fig:isotrop_int}
\end{figure}
\begin{table}[htb]
\caption{Calculated isotropic couplings in the Mn$_3$Z compounds. The interactions given in units of meV are indexed according to increasing distances of the pairs, while interactions with and without prime stand for inequivalent pairs with the same distance (see Fig.~\ref{fig:isotrop_int}). For better understanding, in the second row the in-plane and out-of-plane couplings are denoted by ip and oop, respectively.}
\label{table:isotropic}
\begin{ruledtabular}
\begin{tabular}{c|rrrrrrr}
& \multicolumn{1}{c}{$J_1$} & \multicolumn{1}{c}{$J_2$} &\multicolumn{1}{c}{$J_2^{\prime}$}& \multicolumn{1}{c}{$J_3$} & \multicolumn{1}{c}{$J_3^{\prime}$}& \multicolumn{1}{c}{$J_4$} & \multicolumn{1}{c}{$J_5$} \\
& \multicolumn{1}{c}{oop} & \multicolumn{1}{c}{ip} &\multicolumn{1}{c}{ip} & \multicolumn{1}{c}{oop} & \multicolumn{1}{c}{oop} & \multicolumn{1}{c}{oop} & \multicolumn{1}{c}{oop} \\\colrule
Mn$_3$Sn & -15.3 & -3.2 & 4.5 & 13.8 & 11.3 & -2.87& -4.08 \\
Mn$_3$Ge & -22.6 & -7.4 & 10.1 & 5.9 & 7.1 & -2.54& -4.39 \\
Mn$_3$Ga & -23.7 & -15.9 & 0.9 & 10.7 & 5.5 & -3.82& -4.42
\end{tabular}
\end{ruledtabular}
\end{table}
From the $\mathbf{J}_{ij}$ and $\mathbf{K}_i$ matrices we can calculate the $\mathbf{J}_{\alpha\beta}$ matrices as defined in Eq.~\eqref{eq:sum_sublattice}. We checked that the obtained matrices satisfy to high accuracy the analytic forms \eqref{eq:JAA} and \eqref{eq:JBC} we deduced from symmetry principles, so the six parameters of the sublattice model can be read off. The values of these parameters depend on the cutoff distance of pairs in the sum in Eq.~\eqref{eq:sum_sublattice}. The dependence of the DM and the anisotropy parameters for Mn$_3$Sn is shown in Fig.~\ref{fig:cutoff}. Clearly, all the parameters converge well beyond a distance of about $2 \,a_{\text{2d}}$.
The other two compounds show a similar behavior. Based on these results, in all cases we used a cutoff of $2.51\, a_{\text{2d}}$ for the calculation of the sublattice model parameters.
\begin{figure}[htb]
\includegraphics[width=\linewidth]{cutoff_mn3sn.pdf}
\vskip -10pt
\caption{The cutoff dependence of the relativistic interactions in the sublattice matrices \eqref{eq:JAA} and \eqref{eq:JBC} for Mn$_3$Sn. Note that the DM parameter $D$ is about one order of magnitude larger than the anisotropy parameters. The isotropic coupling $J$ (not presented here) shows a similar cutoff dependence.}
\label{fig:cutoff}
\end{figure}
\begin{table}[htb]
\caption{Calculated sublattice model parameters, see Eqs.~\eqref{eq:JAA} and \eqref{eq:JBC}, for the Mn$_3$Z compounds based on the SCE-RDLM method.}
\label{table:sublattice_params}
\begin{ruledtabular}
\begin{tabular}{c|cccccc}
& $J$[meV] & $D$[meV] & $K^x$[$\mu$eV]& $K^y$[$\mu$eV] & $J^x$[$\mu$eV] & $J^y$[$\mu$eV] \\\colrule
Mn$_3$Sn & -46.7 & -0.547 & 11 & 39 & -85 & 86 \\
Mn$_3$Ge & -51.6 & -0.246 & -64 &123 &-108 & 92 \\
Mn$_3$Ga & -77.0 & -0.447 & -81 &129 & -65 &135 \\
\end{tabular}
\end{ruledtabular}
\end{table}
The calculated parameters of the sublattice model, see Eqs.~\eqref{eq:JAA} and \eqref{eq:JBC}, are summarized in Table~\ref{table:sublattice_params}. In each case a large antiferromagnetic isotropic coupling was obtained, which again explains the formation of the low-energy frustrated triangular configurations. The DM parameter being two orders of magnitude less than $J$ has negative sign, thus, it prefers the $\Gamma_5$ state against the $\Gamma_3$ state for all of the compounds. The anisotropy constants are typically one order of magnitude less than the DM parameters. Remarkably, the anisotropy parameters indexed by $x$ are positive and those indexed by $y$ are negative. The only exception is observed for $K^x$ in case of Mn$_3$Sn, which is negative in sign.
\subsection{Group-theoretical argument}\label{grouptheory}
The representation $\mathbf R_g$ of the $D_{3h}$ point group on the nine-dimensional space defined within the three-sublattice model, see Eqs.\ \eqref{eq:Hcomp} and \eqref{eq:symmetry}, can be decomposed according to irreducible representations as
\begin{equation}
\mathbf R_g = A^{\prime}_1 \oplus A^{\prime}_2 \oplus 2 E^{\prime} \oplus A^{\prime\prime}_2 \oplus E^{\prime\prime}\;,
\end{equation}
where $A^{\prime}_1$, $A^{\prime}_2$ and $A^{\prime\prime}_2$ are one-dimensional, while $E^{\prime}$ and $E^{\prime\prime}$ are two-dimensional irreducible representations. After projecting to irreducible subspaces we found that the low-energy chiral states and the ferromagnetic states correspond to the $A^{\prime}_1$, $A^{\prime}_2$ and the $2E^{\prime}$ irreducible representations as shown in Fig.~\ref{fig:irreps}. Here we distinguish the states $\Gamma_{3,x}$, $\Gamma_{3,y}$ and $\Gamma_{5,x}$, $\Gamma_{5,y}$ based on the orientation of the spin on the A sublattice.
\begin{figure}[htb]
\begin{tabular}{|c|c|}
\colrule
\quad \rotatebox[origin=c]{90}{$A^{\prime}_1 = \Gamma_{3,y}$} \quad
& \begin{minipage}[c]{.85\linewidth}
\vspace*{0.2cm}
\includegraphics[width= .3\linewidth]{gamma3y.pdf}
\vspace*{0.3cm}
\end{minipage} \\
\colrule
\quad \rotatebox[origin=c]{90}{$A^{\prime}_2 = \Gamma_{3,x}$} \quad
& \begin{minipage}[c]{.85\linewidth}
\vspace*{0.3cm}
\includegraphics[width= .25\linewidth]{gamma3x.pdf}
\vspace*{0.3cm}
\end{minipage} \\
\colrule
\quad \rotatebox[origin=c]{90}{$2E^{\prime} =\{ \Gamma_{5,x},\text{FM}_x,\Gamma_{5,y},\text{FM}_y\}$} \quad
& \begin{minipage}[c]{.85\linewidth}
\vspace*{0.3cm}
\includegraphics[width= 0.75\linewidth]{eprime.pdf}
\vspace*{0.2cm}
\end{minipage} \\ \colrule
\end{tabular}
\label{fig:irreps}
\caption{The irreducible representations corresponding to the low-energy spin configurations within the nine-dimensional subspace of the three Mn sublattices.
}
\end{figure}
From Fig.~\ref{fig:irreps} we can conclude that the states $\Gamma_{3,x}$ and $\Gamma_{3,y}$ correspond to different one-dimensional irreducible representations, therefore, they are not degenerate by symmetry. So the energy of the $\Gamma_3$ states changes under in-plane global rotations. In contrast, the $\Gamma_{5,x}$ and $\Gamma_{5,y}$ states and also the FM$_x$ and FM$_y$ states form the basis of the same two-dimensional irreducible representation ($E^\prime$), thus, they are pairwise degenerate. This means that the $\Gamma_5$ and FM states are energetically insensitive to in-plane global rotations. Furthermore, sharing the same symmetry the $\Gamma_{5,x}$ and the FM$_x$ states are coupled and the same applies to the $\Gamma_{5,y}$ and the FM$_y$ states. In order to obtain the ground state of the model we have to diagonalize the Hamiltonian \eqref{eq:Hcomp} in the corresponding subspaces which leads to the diagonalization of the following matrix:
\begin{equation}
\left(
\begin{array}{cc}
\vec{e}_{\Gamma_{5,\xi}} \;\mathbf J\;\vec{e}_{\Gamma_{5,\xi}}&
\vec{e}_{\Gamma_{5,\xi}} \;\mathbf J\;\vec{e}_{\text{FM}_\xi}\\
\vec{e}_{\text{FM}_\xi}\;\mathbf J\;\vec{e}_{\Gamma_{5,\xi}}&
\vec{e}_{\text{FM}_\xi}\;\mathbf J\;\vec{e}_{\text{FM}_\xi}
\end{array}
\right)
\end{equation}
where $\xi\in\{x,y\}$. The four resulting eigenstates can be cast into two degenerate weak ferromagnetic (WF) states that are slight distortions of the $\Gamma_5$ states and two denerate states that are modulations of the FM states. The low-energy weak ferromagnetic eigenstates can be written as
\begin{align}
\vec e_{\text{WF}_x} &=\mu \, \vec e_{\Gamma_{5,x}}+ \nu \, \vec e_{\text{FM}_x} \label{eq:WFx} \\
\vec e_{\text{WF}_y} &=\mu \, \vec e_{\Gamma_{5,y}}- \nu \, \vec e_{\text{FM}_y} \,,
\label{eq:WFy}
\end{align}
where $\mu^2+\nu^2=1$ and the explicit analytical expression for the $\nu$ parameter in terms of the sublattice model parameters indicates that
\begin{equation}
\nu \propto \left(J^x-J^y\right) + 2\left(K^x-K^y\right) \,.
\label{eq:nu}
\end{equation}
This result also suggests that the two different WF states possess a net magnetization of $\pm \nu$ coming from their FM component indicating that the direction of this WF moment with respect to the A moment is different in the $x$ and $y$ cases. Moreover, the weak ferromagnetic moment appears only if the $x$ and $y$ on-site and/or two-site anisotropy parameters differ from each other, but it does not occur if only the Dzyaloshinsky--Moriya interaction (DMI) is present in the system on top of the AFM isotropic interactions. Thus, in terms of group theoretical analysis, we regained the result of Tomiyoshi and Yamaguchi \cite{tomiyoshi-1982} stating that the formation of the weak ferromagnetism in the Mn$_3$Sn compound happens due to magnetic anisotropy rather than DMI. We note that the WF states \eqref{eq:WFx} and \eqref{eq:WFy} are stationary states of the Hamiltonian \eqref{eq:Hcomp} with the lowest energy,
but they do not refer to a set of spin vectors of unit length assumed in the classical Heisenberg model. In the next section we therefore revisit our investigation of the WF states within the space of classical spin states.
\subsection{Classical spin-model study}
As discussed in the previous section in terms of group-theoretical arguments, the energy of the $\Gamma_3$ state shows anisotropic behavior under global rotations around the $z$ axis, while the energy of the $\Gamma_5$ state is invariant to such rotations. This can be easily shown by calculating the rotational energies directly from
Eq.\ \eqref{eq:H_sublatt} and using the parametric forms of the sublattice exchange matrices, Eqs.~\eqref{eq:JAA} and \eqref{eq:JBC}. For the $\Gamma_3$ state this gives
\begin{align}
E_{\Gamma_3}(\phi)=&\dfrac{3}{2}\left(J-\sqrt{3}D+\dfrac{3J^x-J^y}{2}-\dfrac{3K^x-K^y}{2}\right)\nonumber\\&- 3\left((J^x-J^y)-(K^x-K^y)\right)\sin^{2}\phi \, ,
\label{eq:E_g3}
\end{align}
where $\phi$ is the rotation angle around the $z$ axis with respect to the $y$ direction of the magnetic moment at the A atom (see the left panel of Fig.\ \ref{fig:spin_structure}). From the data of Table \ref{table:sublattice_params} it can be inferred that the in-plane anisotropy constant defined as the coefficient of the $\sin^2\phi$ term is negative, so the $\phi=0$ state ($\Gamma_{3,y}$) is the lowest in energy. For the energy of the $\Gamma_5$ states we get a constant indeed, \begin{equation}
E_{\Gamma_5} = \dfrac{3}{2}\left(J+\sqrt{3}D+\dfrac{J^x+J^y}{2}-\dfrac{K^x+K^y}{2}\right)\,.
\end{equation}
The energy difference between the $\Gamma_5$ and the $\Gamma_{3,y}$ states is then given by
\begin{align}
E_{\Gamma_5}-E_{\Gamma_{3,y}} = 6\sqrt{3} D &-\dfrac{3}{2}(J^x-J^y-K^x+K^y) \, .
\label{eq:Ediff_G5-G3}
\end{align}
Since for all considered Mn$_3$Z compounds the DMI is negative and is much larger in magnitude than the in-plane anisotropy term entering Eq.\ \eqref{eq:Ediff_G5-G3}, in each case the $\Gamma_5$ state has lower energy. This should be contrasted with the $L1_{2}$-type Mn$_3$Ir alloy, where the $\Gamma_3$ state is stabilized due to the magnetic anisotropy of about 10~meV \cite{szunyogh-2009}.
\begin{figure}[htb]
\includegraphics[width=0.9\linewidth]{wf_structure.pdf}
\caption{Weak ferromagnetic distortions of the $\Gamma_5$ states: the WF$_x$ state on the left and the WF$_y$ state on the right. The shaded arrows show the spin directions in the original $\Gamma_5$ state. The distortion is parametrized by the tilting angle $\Delta\phi$.}
\label{fig:wf}
\end{figure}
Within the classical spin model the weak ferromagnetic distortions can be parametrized by a tilting angle $\Delta\phi$ for two of the sublattices as illustrated in Fig.\ \ref{fig:wf}.
The weak ferromagnetic moment is then related to $\Delta\phi$ as
\begin{equation}
m_{\text{WF}}(\Delta\phi) = 1 - 2\sin\left(\frac{\pi}{6}-\Delta\phi\right)\;.
\label{eq:wf_mom}
\end{equation}
We calculated the energy of the weak ferromagnetic states as a function of the tilting angle $\Delta\phi$ based on the sublattice spin model.
The corresponding energy curves are shown in Fig.~\ref{fig:sm_wf} for both the WF$_x$ and WF$_y$ distortions and for all the three compounds. As expected, for each alloy a clear parabolic minimum is obtained with
positive $\Delta\phi$ for WF$_x$ distortions and with negative $\Delta\phi$ for WF$_y$ distortions (cf.\ Fig.\ \ref{fig:wf}).
The weak ferromagnetic moments and the distortion angles obtained from the minima of the energy curves in Fig.~\ref{fig:sm_wf} are summarized in Table~\ref{table:sublattice_results}. Interestingly, the magnitudes of the WF$_x$ tilting angle are systematically smaller than those for WF$_y$, although the relative difference is about or less than 1~\%. Correspondingly, the size of the WF moments also somewhat differ for the two kinds of WF distortions, which contradicts the prediction of the group-theoretical analysis. It should be recalled again that the ground state obtained from group theory is outside the space of classical spin states.
\begin{figure}[htb]
\includegraphics[width=1\linewidth]{sm_wf_curves.pdf}
\caption{Calculated energy curves for the Mn$_3$Z compounds as a function of the $\Delta\phi$ weak ferromagnetic distortion angle. Left: WF$_x$ state, right: WF$_y$ state. The minimum positions $\Delta\phi_{x/y}$ were determined from parabolic fits.}
\label{fig:sm_wf}
\end{figure}
\begin{table}[htb]
\caption{Calculated weak ferromagetic distortions and net magnetic moments for the Mn$_3$Z compounds from the sublattice spin model with parameters obtained using the SCE-RDLM method.}
\label{table:sublattice_results}
\begin{ruledtabular}
\begin{tabular}{c|cccc}
& $\Delta\phi_x$[deg] & $m_{\text{WF}_x}[10^{-3}]$ & $\Delta\phi_y$[deg] & $m_{\text{WF}_y}[10^{-3}]$ \\ \colrule
Mn$_3$Sn & 0.0785 & 2.37 & -0.0795 & -2.40 \\
Mn$_3$Ge & 0.1821 & 5.51 & -0.1825 & -5.51 \\
Mn$_3$Ga & 0.1009 & 3.05 & -0.1018 & -3.06 \\
\end{tabular}
\end{ruledtabular}
\end{table}
In order to gain deeper insight into the results obtained above and on the relationship of the WF distortion angle and the parameters of the spin model, we repeated the search for the minimum of the WF distortion energy analytically. The energy of the WF$_x$ distortion can be expressed as
\begin{align}
E_{\text{WF}_x}(\Delta\phi) &=
J \left( \cos(2\Delta\phi-\pi/3) + 2 \cos(\Delta\phi+\pi/3) \right) \nonumber \\
&+ D \left( \sin(2\Delta\phi-\pi/3) + 2 \sin(\Delta\phi+\pi/3) \right) \nonumber \\
&+ \dfrac{J^x}{2} \left( \cos(2\Delta\phi-\pi/3) + 2\cos(\Delta\phi)- 1 \right) \nonumber \\
&+ \dfrac{J^y}{2} \left( \cos(2\Delta\phi-\pi/3)-2\sqrt{3}\sin(\Delta\phi)+1 \right) \nonumber \\
&- (K^x-K^y) \cos(2\Delta\phi+\pi/3) - \dfrac{K^x}{4} - \dfrac{5K^y}{4}\;,
\label{eq:dphi_x}
\end{align}
\begin{comment}
\begin{align}
E(\Delta\phi_x) &=
J \left( \cos(2\Delta\phi_x-\pi/3) + 2 \cos(\Delta\phi_x+\pi/3) \right) \nonumber \\
&+ D \left( \sin(2\Delta\phi_x-\pi/3) + 2 \sin(\Delta\phi_x+\pi/3) \right) \nonumber \\
&+ \dfrac{J^x}{2} \left( \cos(2\Delta\phi_x-\pi/3) + 2\cos(\Delta\phi_x)- 1 \right) \nonumber \\
&+ \dfrac{J^y}{2} \left( \cos(2\Delta\phi_x-\pi/3)-2\sqrt{3}\sin(\Delta\phi_x)+1 \right) \nonumber \\
&- (K^x-K^y) \cos(2\Delta\phi_x+\pi/3) - \dfrac{K^x}{4} - \dfrac{5K^y}{4}\;,
\label{eq:dphi_x}
\end{align}
\end{comment}
and similar for $E_{\text{WF}_y}(\Delta\phi)$ by interchanging indices $x$ and $y$.
After expanding the above expression up to second order in $\Delta\phi$ it is easy to find its minimum yielding
\begin{equation}
\Delta\phi_{x} = -\dfrac{\sqrt{3}}{2}\dfrac{J^{x} - J^{y} + 2(K^{x} - K^{y})}{-3J - 3\sqrt{3}D - 2J^{x} - J^{y} + 2K^{x} - 2K^{y}}\;.
\label{eq:analytic_dphi}
\end{equation}
and again similar for $\Delta\phi_y$ by interchanging the indices $x$ and $y$.
This result has some important implications. Firstly, the numerator on the right-hand side of Eq.\ \eqref{eq:analytic_dphi} changes sign between the WF$_{x}$ and WF$_{y}$ distortions, while the denominator is always positive due to the large negative $J$ and $D$ parameters, which uniquely explains the sign change between the distortion angles $\Delta\phi_x$ and $\Delta\phi_y$. In addition, there is a small change in the denominator for the $x$ and $y$ cases, so we also found an analytic explanation for the deviation in the size of the corresponding tilting angles in Table~\ref{table:sublattice_results}. Secondly, the numerator is identical to the multiplicative factor we obtained from group theory for the weak ferromagnetic moment, see Eq.\ \eqref{eq:nu}. If we expand Eq.~\eqref{eq:wf_mom} around $\Delta\phi=0$, we can see that for small distortions the weak ferromagnetic moment $m_\text{WF}$ is proportional to $\Delta\phi$ thus also with $J^{x} - J^{y} + 2(K^{x} - K^{y})$ as proposed by group theory.
The weak ferromagnetism in the Mn$_3$Z alloys is, therefore, qualitatively explained in the same way from group theory and from the classical spin model as being the consequence of nonzero on-site and/or two-site anisotropies,
$K^{x} - K^{y}$ and $J^{x} - J^{y}$, respectively.
\subsection{Self-consistent calculations}
\begin{table*}[htb]
\caption{The properties of the self-consistent weak ferromagnetic ground states of the Mn$_3$Z compounds calculated from the SKKR method. The experimental results are taken from Refs.~\onlinecite{cable-1993}, \onlinecite{kadar-1971} and \onlinecite{kren-1970} for Mn$_3$Sn, Mn$_3$Ge and Mn$_3$Ga, respectively. The sign of the weak ferromagnetic moments refers to their orientation relative to the orientation of the moments in the A sublattice, see Fig.~\ref{fig:wf}.}
\label{table:scf_results}
\begin{ruledtabular}
\begin{tabular}{c|ccc}
& Mn$_3$Sn & Mn$_3$Ge & Mn$_3$Ga \\ \colrule
$\Delta\phi_x$& spin: -$0.006^\circ$ orbital: -$0.74^\circ$ & spin: $ 0.052^\circ$ orbital: $ 6.9^\circ$ & spin: $ 0.22^\circ$ orbital: $ 10.5^\circ$ \\
$\Delta\phi_y$& spin: $ 0.010^\circ$ orbital: $ 0.78^\circ$ & spin: -$0.065^\circ$ orbital: -$5.9^\circ$ & spin: -$0.39^\circ$ orbital: -$8.9^\circ$\\
m$_s^{\text{Mn}}[\mu_{B}]$ & 3.15 (3.17 $\pm$ 0.07 exp.) & 2.61 ($2.4\pm0.2$ exp.) & 2.60 ($2.4 \pm 0.2$ exp.) \\
m$_{\text{WF}}[\mu_{B}]$ & x: -$0.003$, y: 0.003 (0.009 exp.)& x: 0.016, y: -$0.017$ (0.06 exp.)& x: 0.030, y: -$0.041$ (0.045 exp.)\\
$E_{\text{WF}_x}-E_{\text{WF}_y}$[eV/f.u.] & $8.9\cdot 10^{-8}$ & -$3\cdot 10^{-9}$ & -$4.7\cdot 10^{-6}$
\end{tabular}
\end{ruledtabular}
\end{table*}
We also performed self-consistent calculations to determine the ground states of the Mn$_3$Z alloys by using the relativistic SKKR approach. We use the setup for the SKKR calculations discussed in Sec.~\ref{abinitio} and we let the magnetic moments relax from the $\Gamma_{5,x}$ and $\Gamma_{5,y}$ states to the corresponding weak ferromagnetic states.
This way we include the effect of the induced moment on the Z sites and also orbital-polarization effects into the calculations similarly as in the work of Sandratskii and K\"{u}bler\cite{sandratskii-1996} who investigated the weak ferromagnetism of Mn$_3$Sn from {\em ab initio} calculations.
\begin{figure}[htb]
\includegraphics[width=\linewidth]{scf_wf.pdf}
\caption{Self-consistently calculated spin and orbital contributions to the weak ferromagnetic moments for the Mn$_3$Z alloys. The moments are shown for both the WF$_x$ and WF$_y$ states.}
\label{fig:scf_wf}
\end{figure}
In Table~\ref{table:scf_results} we summarize the main results of the SKKR calculations for the weak ferromagnetic states of the Mn$_3$Z compounds.
First, we observe that the distortion of the spin vectors for Mn$_3$Sn is opposite for both the WF$_x$ and WF$_y$ states as proposed by the spin model, see Table~\ref{table:sublattice_results}. Regarding the expressions we derived for the distortion angles, Eq.~\eqref{eq:analytic_dphi}, this would mean that the self-consistent calculations predicted a magnetic anisotropy of opposite sign as compared to the spin model obtained from the SCE method. From Fig.~\ref{fig:scf_wf} presenting the spin and orbital contributions to the weak ferromagnetic moments we, however, see that in case of Mn$_3$Sn the WF moment is dominated by the orbital moment to which the spin model does not apply. In this case the tilting of spin moments seems to follow that of the orbital moments. Reassuringly, Sandratskii and K\"{u}bler\cite{sandratskii-1996} also obtained distortions of the same rotational sense for the WF$_x$ and WF$_y$ states, resulting in a net moment antiparallel and parallel with the moment of the A atom, respectively. In contrast to our work, the spin-moment contribution reported in Ref.~\onlinecite{sandratskii-1996} is almost twice as large as the orbital contribution, though the total weak ferromagnetic moment, 0.004 $\mu_\text{B}$, is very close to our value (0.003 $\mu_\text{B}$). Remarkably, the distortion angle of the orbital moments is about two orders of magnitude larger than that of the spin moments, which, at least to a somewhat smaller extent, was also found in Ref.~\onlinecite{sandratskii-1996}. It is worth to note that the SKKR codes rely on the solution of the Kohn--Sham--Dirac equation, while in Ref.~\onlinecite{sandratskii-1996} the spin-orbit coupling is treated as an additive term to a scalar-relativistic Hamiltonian \cite{takeda-1978}.
For Mn$_3$Ge and Mn$_3$Ga, the self-consistent calculations yield weak ferromagnetic distortions of the same direction as found in the spin-model studies, although for Mn$_3$Ge the orbital contribution is still nearly three times larger than the orbital contribution, see Fig.~\ref{fig:scf_wf}. Only in case of Mn$_3$Ga, the spin-moment contributions become dominant and, curiously, this contribution shows a large difference between the WF$_x$ and WF$_y$ states which can not be understood based on the sublattice spin model. By contrast, the orbital contributions do not show this asymmetry for any of the systems under investigation,
even though the quite enhanced distortion angles of the orbital moments for Mn$_3$Ge and Mn$_3$Ga display a remarkable anisotropy.
As also indicated in Table~\ref{table:scf_results}, our {\em ab initio} calculations can not resolve with a reliable accuracy which of the two kinds of weak ferromagnetic states is energetically preferred. In case of Mn$_3$Ga, the WF$_x$ seems to be lower in energy, but from the very small energy difference of 4.7$\cdot10^{-6}$~eV/f.u.\ we rather conclude that the two weak ferromagnetic states are degenerate within the precision of the method we use.
A comparison with the experimental results is also shown in Table~\ref{table:scf_results}. The calculated spin moments of the Mn atoms are within the error range of the experiments. We note that the Mn moments slightly differ on the A and B (or C) sublattices as also reported in Ref.~\onlinecite{sandratskii-1996}. The calculated weak ferromagnetic moments are also in the range of the experimental values. The smallest value is obtained for Mn$_3$Sn as in the experiment, while for Mn$_3$Ga we find a very good quantitative agreement with the measured value. The largest deviation from the experiment is found in the case of Mn$_3$Ge. A fair comparison between theory and experiment is, however, hardly possible for this alloy, since only off-stoichiometric samples could be prepared \cite{kadar-1971}, where Mn atoms can occupy Z positions leading to an enhanced net magnetic moment. Zhang \emph{et al.}\cite{zhang-2013}\ also performed density functional calculations using VASP\cite{vasp} for the hexagonal Mn$_3$Z alloys. As compared to their results, our calculated Mn spin moments are systematically larger by about 0.1 $\mu_\text{B}$, which might be attributed to the fact that in Ref. \onlinecite{zhang-2013} slightly smaller, optimized lattice constants and the generalized gradient approximation for the exchange-correlation functional \cite{PBE-1996} were used as opposed to the experimental lattice constants and the local density functional we employed, respectively.
Reassuringly, however, they reported the weak ferromagnetic moments, 0.01 $\mu_\text{B}$ for Mn$_3$Sn and Mn$_3$Ge and 0.03 $\mu_\text{B}$ for Mn$_3$Ga, that are consistent with our values.
As we noted already, the orbital moments have a significant weight in the weak ferromagnetic moment. Interestingly, the orbital moments of the Mn atoms show the behavior we found from group theory, namely that the weak ferromagnetic state can be decomposed as the linear combination of a $\Gamma_5$ and an FM state with the same mixing coefficients for the $x$ and $y$ state as given in Eqs.~\eqref{eq:WFx} and \eqref{eq:WFy}. In correspondence with these relationships, the net orbital moments for the WF$_x$ and WF$_y$ states are $m^\text{Mn}_{\ell} = 3\nu$ and $m^\text{Mn}_{\ell} = -3\nu$, respectively, while subtracting $\vec m^\text{Mn}_{\ell}/3$ from the orbital moments of each sublattice, a perfect $\Gamma_5$ state is obtained with local orbital moments of $\mu$.
The corresponding parameters for the three alloys are collected in Table~\ref{table:scp_orbital_states}. The opposite sign of $\nu$ for Mn$_3$Sn as compared to Mn$_3$Ge and Mn$_3$Ga reflects the opposite sign of the tilting angle as discussed above. Moreover, the increased magnitudes of $\nu/\mu$ for Mn$_3$Ge and Mn$_3$Ga correspond to the enhanced tilting angles for these alloys with respect to Mn$_3$Sn, see Table~\ref{table:scf_results}. Note that for Mn$_3$Ga
we found a slight deviation from the decomposition based on Eqs.~\eqref{eq:WFx} and \eqref{eq:WFy}. We believe that this impressive agreement between the distortion of the orbital moments and the group-theoretical prediction is due to the strongly non-rigid character of the orbital moments. The lack of a constraint of a constant magnitude allows the orbital moments to assume the superimposed WF configuration preferred by group theory.
\begin{table}[htb]
\caption{The $\mu$ and $\nu$ parameters for the self-consistently calculated orbital moments, see Eqs.~\eqref{eq:WFx} and \eqref{eq:WFy}, using the convention where the $\Gamma_5$ and FM states are constructed from dimensionless unit vectors and the parameters are measured in units of $10^{-3}\mu_\text{B}$.}
\label{table:scp_orbital_states}
\begin{ruledtabular}
\begin{tabular}{c|ccc}
& Mn$_3$Sn & Mn$_3$Ge & Mn$_3$Ga \\ \colrule
$\mu$& 40 & 33 & 26 \\
$\nu$& -0.6 & 4.3 & x: 5.07 \quad y: -5.10
\end{tabular}
\end{ruledtabular}
\end{table}
Within the KKR formalism it is possible to scale down the spin-orbit coupling (SOC) selectively on different sites using a scalar relativistic framework for the treatment of the single-site scattering \cite{takeda-1978,ebert-1996}. In order to gain more insight into the effect of the spin-orbit coupling, we performed self-consistent calculations for each compound where we switched off the SOC either on the Z or on the Mn sites. The results for the calculated spin and orbital parts of the net moments and their resolution into the Mn- and Z-atom contributions are summarized in Table~\ref{table:soc} together with the case where the SOC is included at all sites.
\begin{table}[htb]
\caption{Self-consistently calculated net magnetic moments of the weak ferromagnetic states in the Mn$_3$Z alloys decomposed into spin and orbital contributions, as well as into contributions related to the Mn and Z atoms. All values are given in units of $10^{-3} \mu_\text{B}$. Both the WF$_x$ and WF$_y$ states are considered (see the second column). The cases when the SOC is included on all sites, only on the Mn sites and only on the Z sites, are indicated in the third column by `all', Mn and Z, respectively (Z=Sn, Ge, Ga).}
\label{table:soc}
\begin{ruledtabular}
\begin{tabular}{ccc||rrr|rrr}
& & SOC & $m_s^{\text{Mn}}$ & $m_s^{\text{Z}}$ & $m_{s}$ & $m_\ell^{\text{Mn}}$ & $m_\ell^{\text{Z}}$ & $m_{\ell}$ \\ \hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{Mn$_3$Sn}} & \multirow{3}{*}{\rotatebox[origin=c]{90}{WF$_x$}} & all & -0.13 & -0.01 & -0.14 & -1.80 & -1.04 & -2.84 \\
& & Mn & 1.58 & -0.03 & 1.56 & 9.73 & -0.99 & 8.74 \\
& & Sn & 0.24 & 0.01 & 0.25 & -11.47 & -0.07 & -11.53\\ \cline{2-9}
& \multirow{3}{*}{\rotatebox[origin=c]{90}{WF$_y$}}& all & 0.49 & 0.01 & 0.49 & 1.80 & 1.04 & 2.85\\
& & Mn & -1.66 & 0.03 & -1.63 & -9.73 & 0.99 & -8.74 \\
& & Sn & -0.24 & -0.01 & -0.25 & 11.47 & 0.07 & 11.53 \\\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{Mn$_3$Ge}} & \multirow{3}{*}{\rotatebox[origin=c]{90}{WF$_x$}}& all & 4.43 & -0.08 & 4.36 & 12.79 & -1.32 & 11.47 \\
& & Mn & 6.13 & -0.09 & 6.04 & 18.65 & -1.23 & 17.42 \\
& & Ge & -0.14 & 0.00 & -0.13 & -5.89 & -0.09 & -5.98 \\ \cline{2-9}
& \multirow{3}{*}{\rotatebox[origin=c]{90}{WF$_y$}}& all & -5.39 & 0.10 & -5.30 & -12.81 & 1.32 & -11.49 \\
& & Mn & -7.98 & 0.11 & -7.86 & -18.69 & 1.23 & -17.45 \\
& & Ge & 0.11 & -0.00 & 0.11 & 5.89 & 0.09 & 5.98 \\\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{Mn$_3$Ga}} & \multirow{3}{*}{\rotatebox[origin=c]{90}{WF$_x$}}& all & 16.33 & -0.33 & 16.00 & 15.22 & -1.11 & 14.11 \\
& & Mn & 24.38 & -0.50 & 23.88 & 20.41 & -1.03 & 19.38 \\
& & Ga & 33.83 & -0.70 & 33.13 & -5.21 & -0.08 & -5.29 \\ \cline{2-9}
& \multirow{3}{*}{\rotatebox[origin=c]{90}{WF$_y$}}& all & -27.57 & 0.55 & -27.02 & -15.33 & 1.11 & -14.22 \\
& & Mn & -38.97 & 0.77 & -38.20 & -20.57 & 1.03 & -19.54 \\
& & Ga & -21.69 & 0.43 & -21.26 & 5.21 & 0.08 & 5.29 \\
\end{tabular}
\end{ruledtabular}
\end{table}
First of all, Table~\ref{table:soc} indicates that net orbital moments induced by the SOC at the Mn and at the Z atoms add up almost perfectly when the SOC is switched on at each of the sites. Moreover, this is valid for the site-resolved contributions of the net orbital moments. Apparently, such an additivity of the SOC induced net spin moments can not be established.
In case of Mn$_3$Sn, the SOC at the Mn sites induces a sizable net spin moment mainly with contributions from the Mn atoms, but superimposed with the SOC of Sn this spin moment considerably reduces in size and it also changes sign. By contrast, the SOC of Sn raises an orbital moment of about 0.01 $\mu_\text{B}$ in size originating from the Mn atoms, which is, however, largely compensated by the orbital moment of opposite sign induced by the SOC of the Mn atoms. Although some orbital moment is induced also at the Sn sites being parallel to the orbital moment of the Mn atoms induced by the SOC of Sn, the total orbital moment remains parallel to that induced by the SOC of Mn. This peculiar cancellation of the orbital moments was also noticed and discussed by Sandratskii and K\"{u}bler\cite{sandratskii-1996}.
In case of the Mn$_3$Ge and the Mn$_3$Ga alloys, it remains valid that both the net spin- and orbital moment mostly have contributions from the Mn atoms, while the Ge and Ga atoms add a negligible amount especially to the spin moment.
The SOC of Mn plays a dominant role in the formation of the net moments in Mn$_3$Ge and this effect is only compensated in about 30~\% by the SOC of Ge. This explains the opposite direction of the weak ferromagnetic distortion, and correspondingly, the opposite direction of the net moment as compared to Mn$_3$Sn, where the effect of the SOC of Sn dominates. The SOC of both the Mn and Ga atoms have a large effect in inducing a net spin moment in Mn$_3$Ga, but, interestingly, when switching on the SOC simultaneously on both sites, the total spin moment is much less than the sum of the spin moments for selectively switched SOC. Moreover, the size of the spin moments induced by the SOC of Mn and Ga follow an opposite order for the WF$_x$ and WF$_y$ states, which might be connected with the large asymmetry of the spin moments for these weak ferromagnetic states in Mn$_3$Ga.
\begin{figure}[htb]
\includegraphics[width=1\linewidth]{constraining_field.pdf}
\vskip -15pt
\caption{Magnitudes of the spin and the orbital contributions to the weak ferromagnetic moment in Mn$_3$Ga as a function of the induced spin moment at the Ga site in Mn$_3$Ga obtained from self-consistent constraining field calculations.}
\label{fig:cf}
\end{figure}
The origin of the large asymmetry of the weak ferromagnetic spin moments for the WF$_x$ and WF$_y$ states obtained from the self-consistent density functional calculations is an unresolved issue. The largest anisotropy is found for Mn$_3$Ga, where also the largest induced spin moment at the Z atom is observed, likewise with a remarkably large asymmetry. In order to see whether there is a connection between these two effects we performed self-consistent calculations by applying a longitudinal constraining field at the Ga site by which the spin moment of Ga could be set arbitrarily. In Fig.~\ref{fig:cf} we plotted the net spin and orbital moments as a function of the induced spin moment on the Ga site. Or results indicate that the induced moment of Ga affects the net orbital moment only very moderately. By contrast, the net spin moments increase nearly linearly with the induced moment of Ga. The corresponding lines for the WF$_x$ and WF$_y$ states have a different slope which implies that the asymmetry is also increasing with increasing induced moment of Ga.
Extrapolating the lines to $m^\text{Ga}_s=0$ there still remains a difference, $m_{\text{WF}_y}- m_{\text{WF}_x} \simeq 10^{-2} \mu_\text{B}$, therefore, we can at best conclude that the induced moment of Ga is one of the sources of the asymmetry of the weak ferromagnetic moment in Mn$_3$Ga. The linear increase of the net spin moment with $m^\text{Ga}_s$ can qualitatively be understood in terms of an isotropic exchange coupling between the Mn and Ga spin moments, while the asymmetry of the net spin moments can be attributed to the anisotropy, i.e to the tensorial nature of this coupling.
\section{Conclusions}
We studied the weak ferromagnetism in the Mn$_3$Z (Z=Sn, Ge, Ga) alloys using a combination of \emph{ab initio} and spin model calculations. Using the point-group symmetry of the systems we set up a model for the three Mn sublattices including the relativistic terms of the Heisenberg Hamiltonian. The parameters of this model were obtained from \emph{ab initio} calculations relying on a spin-cluster expansion. Based on a group-theoretical analysis we showed that there are two degenerate ground states, WF$_x$ and WF$_y$, being the mixture of the $\Gamma_5$ and FM states. The analytical forms of the mixing coefficients imply that the weak ferromagnetic states will only form if the anisotropy parameters distinguish between the $x$ and $y$ spin directions. We recovered this result from analytical expressions for the energy within the classical spin model. Our presented results are fully consistent with the original spin model description of weak ferromagnetism in Mn$_3$Sn by Tomiyoshi and Yamaguchi \cite{tomiyoshi-1982}.
We also performed self-consistent relativistic local density functional calculations in order to investigate the weak ferromagnetic states in a more involved way. In agreement with the seminal work for Mn$_3$Sn by Sandratskii and K\"{u}bler\cite{sandratskii-1996}, our results highlighted the significance of the orbital moments in the weak ferromagnetism of the Mn$_3$Z alloys. A key observation from our calculations is that, as opposed to the spin moments, the orbital moments almost strictly follow the decomposition according to irreducible representations as predicted by group theory. As far as the spin moments of the Mn atoms and the weak ferromagnetic moments are concerned, we found a good agreement with the experiments for all the three alloys. The only exception is the net magnetic moment of Mn$_3$Ge, where presumably due to the off-stoichiometry of the samples the measured moment is significantly higher than the theoretical one.\cite{zhang-2013} By switching on the spin-orbit coupling at different sites selectively we discussed the role of the SOC on the formation of weak ferromagnetic moments and recovered the peculiar cancellation of orbital moments for Mn$_3$Sn reported in Ref.~\onlinecite{sandratskii-1996}, while we also found an argument for the opposite weak ferromagnetic distortion as compared to Mn$_3$Ge and Mn$_3$Ga.
By using constrained local density functional calculations
we established that the induced spin moment of Ga plays an important role in the formation of the weak ferromagnetic moment in Mn$_3$Ga and also it is one of the reasons for the observed large asymmetry of the net moment concerning the WF$_x$ and WF$_y$ states.
\begin{acknowledgments}
The authors are grateful for the financial support by the Hungarian National Scientific Research Fund (NKFIH) under project Nos.\ K115575 and PD124380, as well as by the BME Nanonotechnology FIKP grant of EMMI (BME FIKP-NAT).
\end{acknowledgments}
|
\section{Introduction}
Visual analytics is the science of analytical reasoning facilitated by interactive visualization \cite{Thomas2006}. It integrates human perception, cognition, and reasoning abilities into the data analysis process through visualization and interactions. With the explosion of data, one of its main challenges is to develop scalable interactive visualization techniques for large-scale data to reduce the growing impedance mismatch between data volume/complexity and the human ability to explore and interact with data.
To address this challenge with a focus on data presentation, many techniques have been proposed for exploratory visualization of multi-dimensional data, targeted at both generic and specific application domains \cite{Ware2004,Keim2002}. Parallel coordinates plot (PCP) has become a standard tool for visualizing multi-dimensional data \cite{Heinrich2013}. It has been incorporated into several visual analytics applications, including iPCA \cite{Jeong2009} and iVisClassifier \cite{Choo2010}. In PCPs, axes, that corresponding to dimensions of the original data, are aligned parallel to each other and data points are mapped to lines intersecting the axes at the respective value. The embedding of an arbitrary number of parallel axes into the plane allows the simultaneous display of many dimensions, providing a good overview of the data. However, with classic PCP, although the representation of data points as lines may reveal trends and patterns contained in the data, it also tends to create visual clutter caused by overplotting, especially for large-scale data sets. This visual clutter is one of the main challenges of PCP, which hamper the recognition of patterns in the data. Furthermore, the zigzagging polygonal lines (polylines) used for data representation leads to the cross-over problem, which generally loose visual continuation across the axes in PCP, because they are only $C^{0}$ continuous \cite{Luo2008}. With the cross-over problem, users are difficult to follow lines that share a common point along an axis.
Edge-bundling is a widely used approach to reduce visual clutter in PCP. The data will be aggregated first with data clustering algorithms, then rendered in an illustrative fashion using different forms of edge-bundling or to distinguish between major trends and outliers. However, existing data clustering and edge-bundling techniques used in PCP are not scalable well for large-scale data. Their computing performance depends on the number of dimensions and data points. Their rendering performance depends on the number of lines. This significantly hinders the integration of human judgment into the data analysis process because they have been too slow for interactive use when the number of dimensions or the volume of data sets is scaled up. Furthermore, without human judgment, the clusters generated by automatic algorithms may not reveal the real trends and patterns in the data.
In this paper, we propose an edge-merging PCP technique for large-scale data sets, which includes a one-dimensional clustering method and an edge-merging layout for PCPs. It is scalable for large-scale data sets and supports interactive use in the visual analytics process with acceptable near-real-time performance. The experiments on several data sets are conducted to demonstrate the effectiveness of our edge-merging PCP technique. Our edge-merging PCP technique enhances the visual analytics for large-scale multi-dimensional data sets in the following ways:
\begin{enumerate}
\item It uses a one-dimensional clustering method to cluster the data in each dimension independently. The clusters are shown on each axis. The clustering is highly configurable, including configuring the number and the size of the clusters, to integrate human judgment in the visual analytics process.
\item It reduces the number of lines that need to be rendered by merging the lines belong to the same clusters in two neighboring axes together as one single B\'ezier spline. This edge-merging layout reduces the rendering time. It makes the rendering process scalable for large-scale data sets since the rendering performance is independent of the number of data points.
\item To overcome the cross-over problem, it renders the merged lines as B\'ezier splines.
\item It computes the densities of merged lines to render their widths to reveal the trends, patterns and data distributions between axes.
\end{enumerate}
Based on the edge-merging PCP, we implement a web-based visual analytics application. The application enhances the visual analytics for large-scale multi-dimensional data sets in the following ways:
\begin{enumerate}
\item Its web-based architecture separates the rendering process and clustering process. The clustering process is performed on the server side. The clustering result is rendered in browsers, which supports the interactive exploration of large-scale data with an edge-merging PCP.
\item The application integrates human perception and reasoning abilities into the data analysis process by supporting the configuration of the clustering in each axis to continuously improve the visualization results.
\end{enumerate}
The paper is organized as follows: Section~\ref{section:related work} reviews related work. Section~\ref{section:method} presents the method used in the proposed visual analytics framework. Section~\ref{section:experiment} conducts the experiments and Section~\ref{section:discussion} discusses the result. Section~\ref{section:future} draws conclusions and discusses future work.
\section{Related Work}\label{section:related work}
Since parallel coordinates were introduced in the seminal work of Inselberg \cite{inselberg1985plane,inselberg1987parallel}, and later extended by Wegman \cite{wegman1990hyperdimensional} as a visualization tool, it has become a common and widely used visualization technique for visualizing high-dimensional data. It aligns parallel axes to each other and data points are mapped to coordinates intersecting the axes at the respective value. It allows the simultaneous display of a number of dimensions by embedding of the corresponding number of parallel axes into a plane to reveal trends and patterns in the data.
For large data sets, visual clutter caused by heavy overplotting of lines is often cited as a major challenge for parallel coordinates \cite{Heinrich2013}. Preprocessing the data with various clustering algorithms is a possible approach to reduce the visual clutter \cite{jain1988algorithms}. The clusters can then be displayed and highlighted in extensions of PCPs by using color or edge-bundling. Fua et al. \cite{fua1999hierarchical} develop a multiresolutional view of the data via hierarchical clustering, and use a variation on parallel coordinates to convey aggregation information for the resulting clusters. To reveal details of the internal structure of clusters, Johansson et al. \cite{johansson2005revealing} combine clusters and high-precision textures in a PCP to represent data, in which specific transfer functions that operate on the high-precision textures are used to highlight different aspects of the cluster characteristics. These clusters generated by multidimensional clustering algorithms may not reflect the relation between the two axes because not all dimensions have a strong correlation. Palmas et al. \cite{Palmas2014} present an edge-bundling layout for PCPs using density-based clustering for each dimension independently so that the clustering is directly related to the shown dimensions in every part of the plot. However, these clustering methods are not suitable for a visual analytics application because they are not configurable: the number or size of clusters is generated automatically by algorithms. In contrast, our edge-merging technique takes advantage of human expertise by supporting the configuration of the number and size of clusters. Furthermore, existing clustering-based PCPs are not scalable for large-scale data. The long computation time makes them too slow for interactive use in the visual analytics process. For example, in Palmas et al.'s \cite{Palmas2014} method, to reduce the computation time for the clustering, they sample the parameters 2000 times between 25\% and 1\% of the data range instead of computing it on the whole data range. However, this precomputation still needs almost one minute for $10^{5}$ data points on current desktop hardware in a single computing thread. We propose a one-dimensional clustering method to cluster the data in each dimension independently, which reduces much computation time.
The edge-bundling method that uses cubic B-splines to show adjacency relations atop different tree visualization methods was proposed for the first time by Holten et al. \cite{holten2006hierarchical}. It has been adapted to PCPs based on data clustering by McDonnell et al. \cite{McDonnell2008}. With edge bundling, polylines in PCP become polycurves, which are bent between two axes towards centroid of the corresponding clusters. While this approach reduces visual clutter by freeing up screen space, the resulting polycurves are still $C^{0}$ continuous, which makes them difficult to be visually traced. To avoid this, Heinrich et al. \cite{heinrich2011evaluation} use a $C^{1}$ continuous bundling. Furthermore, Zhou et al. \cite{Zhou2008} present a bundling method that creates visual clusters without the prerequisite of data clustering. However, the bundling brings the curves closer to each other, which may increase the overplotting within a bundle and hide the internal structure of clusters \cite{Palmas2014}. In addition, its rendering performance depends on the number of curves, which hinders the interactive exploration for large-scale data. To reduce the overplotting in bundles and make the rendering performance independent of the number of data points, we propose an edge-merging layout for PCP by combining the idea of confluent drawing and edge bundling. Confluent drawing \cite{dickerson2003confluent} merges groups of edges and draws them as tracks, which allows even very dense and non-planar graphs to be drawn in a planar way. It was first proposed for node-link diagrams. Figure~\ref{fig:cd} illustrates the visual improvement of a simple confluent drawing over the corresponding node-link diagram. Our edge-merging method merges bundles between clusters in the same way.
\begin{figure}[htb]
\centering
\begin{subfigure}{0.45\linewidth}
\centering
\includegraphics[width=.45\linewidth]{imgs/cd1}
\caption{\label{fig:cd1}
\textit{A graph drawn as a traditional node-link diagram.}}
\end{subfigure}
\hspace*{\fill}
\begin{subfigure}{0.45\linewidth}
\centering
\includegraphics[width=.45\linewidth]{imgs/cd2}
\caption{\label{fig:cd2}
\textit{The same graph, with confluent drawing.}}
\end{subfigure}
\caption{\label{fig:cd}
\textit{The visual improvement of a simple confluent drawing over the corresponding node-link diagram} \cite{Eppstein2016}.}
\end{figure}
Another popular solution to the overplotting problem is to replace opaque lines with a density representation in PCPs \cite{miller1991construction, wegman1997high}. Artero et al. \cite{Artero2004} highlight significant relationships between axes in a PCP by constructing a density plot based on density information computed from the data set. Heinrich et al. \cite{Heinrich2009} derive a mathematical density model to generate a continuous parallel coordinates plot. To reveal the internal structure of clusters and the data distribution between axes, we compute the densities of merged bundles to render their widths. It is effective even with a purely geometry-based visualization, which leaves the color and opacity perceptual channels free for visualizing other information.
\section{Method} \label{section:method}
With our edge-merging PCP technique, the data for each dimension is clustered based on a one-dimensional clustering method which is evolved from K-means clustering (Section~\ref{section:dataclustering}). The data points between axes are merged together according to the clusters in each axis (Section~\ref{section:edgemerging}). The merged data points are rendered as B\'ezier splines with specific widths which are determined by their densities (Section~\ref{section:representation}). Our technique is implemented as a web-based visual analytics application with an intention to integrate it with other big data platforms (Section~\ref{section:implementation}). It supports several interactions for taking advantage of human perception and reasoning abilities in the visual analytics process (Section~\ref{section:interaction}).
\subsection{Data Clustering} \label{section:dataclustering}
We cluster the data points for each dimension (continuous variable) independently. For each dimension, there will be several clusters. One cluster represents that the data points in the cluster are close to each other in the dimension. These clustered points will be merged on the respective axis of a PCP.
We propose a one-dimensional clustering method especially for clustering the large-scale data in each dimension in a PCP, which is evolved from K-means clustering. K-means \cite{Wagstaf2001} automatically partition a data set into $k$ groups for a given data set $D=\left\lbrace d_{1}, d_{2}, . . ., d_{n} \right\rbrace$ as follows:
\begin{enumerate}
\item Select $k$ initial cluster centers, $C_{1}, C_{2}, . . ., C_{k}$.
\item For each point $d_{i}$ in the data set $D$, assign it to the closest cluster $C_{j}$ by computing the least squared Euclidean distance.
\item For each cluster $C_{i}$, update its center by averaging all of the points $d_{j}$ that have been assigned to it.
\item Iterate between (2) and (3) until convergence.
\item Return $k$ clusters, $\left\lbrace C_{1}, C_{2}, . . ., C_{k}\right\rbrace$.
\end{enumerate}
The K-means algorithm converges when there is no further change in assignment of instances to clusters by using an iterative refinement technique. For one-dimensional data, K-means will generate several ordered clusters with different size and without overlap. Arthur and Vassilvitskii have proved that the worst-case running time of the method is exponential in the number of points \cite{arthur2006slow}. Therefore, it is hard to be applied in the visual analytics process directly for a large-scale data set.
To address this challenge, instead of selecting $k$ random initial cluster centers, we specifically set $k$ selected fixed centers. Then we modify K-means clustering with the following two considerations to eliminate the iterations in the algorithm:
\begin{enumerate}
\item In a parallel coordinates plot, the data points in each axis are ordered and one point will be only grouped into one cluster.
\item For a complex data set, to take advantage of human reasoning, perception, and cognition, users may prefer to specify the clusters based on their own knowledge than an automated algorithm.
\end{enumerate}
With the first consideration, for a specified cluster number $k$, the clusters in one axis are also ordered and there is no overlap between clusters, which is illustrated in Figure~\ref{fig:clustersinaxis}. In Figure~\ref{fig:clustersinaxis}, we use a center and a radius to define a cluster in one axis, where the gray rectangles represent clusters, the blue dots represent the centers of the clusters, and the green lines represent the radius of the clusters. With this definition, $k$ clusters in one axis for a given data set $D$ are expressed as follows:
\[
\left\lbrace (C_{1},R_{1}), (C_{2},R_{3}), . . ., (C_{k},R_{k})\right\rbrace
\]
which fulfills the following constrains:
\[
C_{i}<C_{i+1}, R_{i} \leq C_{i+1} - R_{i+1} - C_{i}, \sum_{i=1}^{k} R_{i} \leq (d_{max}-d_{min})/2
\]
where $C_{i}$ is the center of the $i$-th cluster, $R_{i}$ is the radius of the $i$-th cluster, $d_{max}$ is the maxima of the data set $D$, and $d_{min}$ is the minima of the data set $D$.
In our method, to cluster the data in a given data set $D$ into $k$ groups, we first specifically select $k$ fixed centers which evenly divides the data set. Accordingly, for a given data set $D$, we can compute $k$ ordered clusters with the same radius without iterations through the following computation:
\[
C_{i}=d_{max}-R\times(1+2\times i), R = \frac{(d_{max}-d_{min})}{2 \times k}
\]
This computation only relies on the maxima, minima of the data set and the number of the clusters. This will reduce much computation time that needed for other clustering method \cite{Palmas2014} used in edge-bundling PCPs.
Furthermore, for complex data sets, the clusters with the same size may be not enough for exploring the data and gaining insight into the data. Based on the second consideration, for taking advantage of human perception and reasoning abilities, our method allow users to specify the number of clusters, and the centers and the various radiuses of the clusters in one axis, as long as the the specified parameters fulfill the condition in the previously mentioned formula. This visual analytics process is demonstrated in Section~\ref{section:experiment}.
Categorical variables are not clustered using the above method. Instead, we treat each category as a cluster. Using this method, the time of computing arbitrary clusters for a large-scale data set could be ignored with the known maxima and minima. The efficient process of obtaining maxima and minima of a extreme large-scale data set is discussed in Section~\ref{section:implementation}.
\begin{figure}[tb]
\centering
\includegraphics[width=.8\linewidth]{imgs/axis-cluster}
\caption{\label{fig:clustersinaxis}
\textit{Clusters in one (horizontal) axis of a PCP.}}
\end{figure}
\subsection{Edge Merging} \label{section:edgemerging}
In a classic PCP, the data points are drawn as polylines. Our method merges these polylines according to the clusters in each axis into one single line and renders the merged line as a cubic B\'ezier spline with a specific width. The width is determined by the density information of the merged line. The purpose of our edge-merging method is to provide a more informative visualization of the internal structure of the clusters and the data distribution between axes, and reduce rendering time in order to interactively and visually explore a large-scale data set. The idea of our edge-merging method combines edge bundling in PCP and confluent drawing which is illustrated in Section~\ref{section:related work}. We use B\'ezier splines because they are widely used in edge-bundling PCP to guarantee $C^{1}$ continuous, which are also easy to follow with the eyes.
To formulate and discuss our edge-merging method, we first denote the clusters in one axis as: $C_{A}^{k}$, where $A$ refers to the name of the axis, $k$ represents that it is the $k$-th cluster in this axis. A B\'ezier spline that connects the centers of two clusters in two axes is denoted as: $B(C_{A}^{i}, C_{B}^{j})$. A pair of two clusters in two axes are denoted as: $P(C_{A}^{i}, C_{B}^{j})$. If the data points $a$ in axis $A$ and $b$ in axis $B$, and $a\in C_{A}^{i}$ and $b\in C_{B}^{j}$, we define it as data points belong to a pair of clusters, denoted as $(a,b) \in P(C_{A}^{i}, C_{B}^{j})$. The number of data points that belong to $P(C_{A}^{i}, C_{B}^{j})$ is denoted as: $N(C_{A}^{i}, C_{B}^{j})$. Then, we describe our edge-merging method by actually drawing an example PCP with two axes, two clusters, and several data points, as illustrated in Figure~\ref{fig:edgemerging}. Consider there are two neighboring axes A and B, and two clusters $C_{A}^{1}$ and $C_{B}^{1}$ in the two axes. With a classic PCP, data points that belong to the two clusters are drawn as polylines, as illustrated in Figure~\ref{fig:edgemergingC}. In Figure~\ref{fig:edgemergingO}, with our edge-merging method, all polylines belong to these two clusters are merged together as one single B\'ezier spline with a specific width, which starts from the center of cluster $C_{A}^{1}$ and ends at the center of cluster $C_{B}^{1}$ and denoted as $B(C_{A}^{1}, C_{B}^{1})$.
\begin{figure}[tb]
\centering
\begin{subfigure}{0.6\linewidth}
\centering
\includegraphics[width=.6\linewidth]{imgs/edgemerging1}
\caption{\label{fig:edgemergingC}
\textit{Several polylines between two neighboring axes in a classic PCP.}}
\end{subfigure}
\begin{subfigure}{0.6\linewidth}
\centering
\includegraphics[width=.6\linewidth]{imgs/edgemerging2}
\caption{\label{fig:edgemergingO}
\textit{Several polylines merged as one single B\'ezier spline.}}
\end{subfigure}
\caption{\label{fig:edgemerging}
\textit{Edge merging between two axes A and B.}}
\end{figure}
In our approach, there is a maximum of $k_{A}\times k_{B}$ merged lines between two neighboring axes, where $k_{A}$ and $k_{B}$ are the number of clusters in either axis. For each merged line, we also compute its density, which is used to specify its width, as follows:
\begin{enumerate}
\item For each two neighboring axes $A$ and $B$, compute clusters in either axis: $C_{A}^{i}$ and $C_{B}^{j}$.
\item For $C_{A}^{i}$ and $C_{B}^{j}$, compute all pairs of clusters between the two axes with data points: $P(C_{A}^{i}, C_{B}^{j})$.
\item For two data points $a$ and $b$ in the two axes, if $(a,b) \in P(C_{A}^{i}, C_{B}^{j})$, $N(C_{A}^{i}, C_{B}^{j})$ add one.
\item Iterate the step (3) for all data points in the two axes.
\item Compute the density of $B(C_{A}^{i}, C_{B}^{j})$ by: $\frac{N(C_{A}^{i}, C_{B}^{j})}{\sum_{i=1}^{i}\sum_{j=1}^{j}N(C_{A}^{i}, C_{B}^{j})}$
\end{enumerate}
As shown in Figure~\ref{fig:edgemerging}, $N(C_{A}^{1}, C_{B}^{1})$ is 5, and the density of $B(C_{A}^{1}, C_{B}^{1})$ is 1. In our edge-merging method, $B(C_{A}^{i}, C_{B}^{j})$ and $N(C_{A}^{i}, C_{B}^{j})$ are used for rendering the merged lines (see in Section~\ref{section:representation}).
\subsection{Representation} \label{section:representation}
In classic PCPs, drawing the polylines or curves for each data point has two main disadvantages for a large-scale data set:
\begin{itemize}
\item Visual clutter and overplotting which hampers the recognition of patterns and trends in the data.
\item The rendering performance depends on the number of data points so that large-scale data sets cannot be explored interactively.
\end{itemize}
These disadvantages are also true for an edge-bundling PCP since they still need to render each data point as a polyline or a curve. Furthermore, with edge bundling, the polylines or curves may be closer, which actually somehow increases the visual clutter and overplotting of the bundled lines or curves.
To address these challenges, we merge data points that belong to a pair of clusters in two axes as one single B\'ezier spline, which connects the centers of two clusters. Based on the clustering result, we render a merged B\'ezier spline between axes as illustrated in Figure~\ref{fig:edgemergingO}. We compute the density of a merged B\'ezier spline to determine its width as follows:
\[
W_{b} = D \times W_{max}
\]
where, $W_{b}$ is the width of the merged B\'ezier spline $B(C_{A}^{i}, C_{B}^{j})$, $D$ is the density of the merged B\'ezier spline, $W_{max}$ is the max line width in the whole visualization. For different data sets, $W_{max}$ is configurable. Figure~\ref{fig:densityRepresent} shows the comparison of a classic PCP with the PCP using our edge-merging layout and density based rendering methods. Since the total density of all merged lines between two axes is 1, in our edge-merging layout, the width of a merged line reflects the distribution of data points between the axes.
\begin{figure}[tb]
\centering
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/car1}
\caption{\label{fig:densityRepresentC}
\textit{Classic parallel coordinate plot.}}
\end{subfigure}
\par\bigskip
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/car2}
\caption{\label{fig:densityRepresentO}
\textit{PCP using our edge-merging with density-based rendering.}}
\end{subfigure}
\caption{\label{fig:densityRepresent}
\textit{Comparison of the classic PCP and the edge-merging PCP on a car data set.} }
\end{figure}
It is easy to see that our edge-merging layout has a much smaller visual footprint than a classic PCP. With fewer B\'ezier splines, it not only reduces visual clutter and overplotting but also reduces the rendering time. More importantly, the rendering performance of our edge-merging layout is independent of the number of data points since we only render the merged B\'ezier spline, which is demonstrated in Section~\ref{section:experiment}. For two given axes, the maximum of merged lines between them only depends on the number of clusters in each axis. Compared with the number of data points in a large-scale data set, the number of merged lines between two axes with a give configuration of clustering is much smaller and constant. Therefore, our edge-merging layout is scalable for large-scale data sets.
\subsection{Implementation} \label{section:implementation}
We implement our edge-merging technique in a web-based visual analytics application with an intention to integrate it with other big data platforms. It is implemented in Java using spring MVC framework \cite{Ladd2006}, and JavaScript using Data-Driven Documents (D3) \cite{bostock2011d3}. It supports the visual analytics process by providing several interactions including configuring the clustering and changing the order of axes, which are described in Section~\ref{section:interaction}. The visual analytics and knowledge generation process of the application is shown in Figure~\ref{fig:process}.
\begin{figure}[tb]
\centering
\includegraphics[width=1\linewidth]{imgs/process}
\caption{\label{fig:process}
\textit{The visual analytics and knowledge generation process of the visual analytics application.}}
\end{figure}
According to Figure~\ref{fig:process}, the web-based architecture of our application separates the rendering process and clustering process. The clustering process, which takes most of the computation time, is performed on the server side. It will reduce the transmission time of data in the network since the data are stored in a relational database or big data platforms in servers. The rendering process is performed in clients' browsers lightweight based on the clustering result. This separation allows users to interactively explore large-scale data with an edge-merging PCP in their browsers, because the clustering result, which contains few graphical elements, could be rendered in browsers in real time.
One additional advantage of our implementation is that it could take advantage of the state-of-the-art advances in data processing and management technologies on the server side. First, the computation is independent for each data item. Therefore, the process can be parallelized: the clustering can be run either on a GPU (which can handle more than 1000 items in parallel) or on a distributed platform such as Apache Hadoop. Second, the results of the clustering process are small in size. Therefore, extensive caching and precomputation can be used. When the user chooses a particular cluster configuration, the clustering must be done only once. Afterwards, the result of the clustering can be cached on the server in a relational database or memory. Next time when a visualization with the same cluster configuration is needed, it can be loaded from the cache immediately. In addition, the server can precompute likely configurations that the user will explore. In this way, the user would experience real-time interaction without any delays.
\subsection{User Interactions} \label{section:interaction}
Our application allows users to interactively explore the visualization space through a control panel that is shown in Figure~\ref{fig:controlPanel}. The application provides the following interactions:
\begin{description}
\item [(a)] Users will select one data set from a list of data sets in the system. These data sets can be stored in a relational database, such as MySQL, or a big data platform, such as Apache Hadoop.
\item [(b)] For each axis, users will specify and save the clustering configuration. Users will first select one axis from a list of all dimensions. Then input the number of clusters $n$, and the center and the radius of each cluster in order by the format: $C_{1},R_{1};C_{2},R_{2};...;C_{n},R_{n}$, where $C$ is the center of the cluster and $R$ is the radius of the cluster. Finally, users will save the configuration.
\item[(c)] Users will input the names of dimensions (axes) by the format: $name_{1},name_{2},...,name_{n}$, and save the configuration. The application only visualizes the data for the input dimensions. The data for each dimension will be visualized in order of the input names. Therefore, the input information is used for adding and deleting axes and changing the order of axes.
\item[(d)] System will update the visualization according to the users' interactions and the saved configurations.
\end{description}
\begin{figure}[tb]
\centering
\includegraphics[width=0.8\linewidth]{imgs/control_panel_areas}
\caption{\label{fig:controlPanel}
\textit{The control panel of the visual analytics application. (a) Data sets selection. (b) Clustering configuration. (c) Coordinates configuration. (d) Visualization update.}}
\end{figure}
\section{Experimental Results} \label{section:experiment}
In this section, we demonstrate the effectiveness of our edge-merging PCP technique through experiments on several data sets.
First, we examined the effectiveness of our edge-merging PCP on a wholesale customer data set from the UCI Machine Learning Repository. The data set refers to customers of a wholesale distributor, which includes the annual spending in monetary units (m.u.) on 6 product categories \cite{Baudry2015}. In Figure~\ref{fig:wholesale}, 3 selected variables of the data set are visualized. Figure~\ref{fig:wholesale2} is the clustering and edge-merging result over the classic PCP shown in Figure~\ref{fig:wholesale1} of the data set. In Figure~\ref{fig:wholesale1}, due to the visual clutter and the overlap of data items, it hides the patterns in the data set and reveals some incorrect results. For example, it shows that the numbers of customers for delicatessen products are almost the same from the Horeca and the Retail channel. The distribution of delicatessen products into geographical regions (Lisbon, Oporto and Other) is not clear from the plot. In contrast, our edge-merging PCP shown in Figure~\ref{fig:wholesale2} successfully reveals the hidden patterns in the data set and shows the trends and the data distribution between axes.
It reduces visual clutter and the overlap of data items by rendering the width of the merged curves according to their densities.
For example, the number of customers for delicatessen products from the Horeca channel is about twice that from the Retail channel (while it seemed to be similar in Figure~\ref{fig:wholesale1}). It also shows data distribution between the different regions in detail. In addition, our result also successfully shows the major clusters and outliers (rare items or observations which raise suspicions by differing significantly from the majority of the data \cite{Zimek2017}) even without the assistance of the color and opacity channels. In our result, the data items with extreme small proportions are visualized as dashed lines. For example,
the red dashed lines show that very few costumers spend more than 17980.5 m.u. on delicatessen products annually.
\begin{figure}[tb]
\centering
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/wholesale}
\caption{\label{fig:wholesale1}
\textit{Classic parallel coordinate plot.}}
\end{subfigure}
\par\bigskip
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/wholesale2}
\caption{\label{fig:wholesale2}
\textit{With the data clustering and the edge-merging layout.}}
\end{subfigure}
\caption{\label{fig:wholesale}
\textit{Experiments on a wholesale customers data set with 3 dimensions and 440 data items.}}
\end{figure}
Next, we tested our proposed methods on a more sophisticated multi-dimensional data set within the visual analytics process. A widely used car information data set with seven variables is visualized in Figure~\ref{fig:experiment}.
Original plot, based on a classic PCP, is shown in Figure~\ref{fig:experiment1}. Figure~\ref{fig:experiment2} shows PCP with edge-merging and automatic clustering, based on the same data set. The clusters in each dimension have the same size, which is computed without human judgment according to the default number of clusters for each dimension. Comparing Figure~\ref{fig:experiment1} and Figure~\ref{fig:experiment2}, our edge-merging layout for PCP successfully shows the major clusters, outliers, and the data distribution between axes even without the assistance of the color and opacity channels. For example, in Figure~\ref{fig:experiment2}, the red dashed lines could be seen as outliers because their proportions are too small compared with others. The well-organized B\'ezier splines and clear distribution between the axes of mpg, cylinders, and displacement reveal the following patterns and trends about the data set:
\begin{enumerate}
\item Most cars' mpg attributes are equally distributed around 27.79 and 15.25, only a relatively small proportion of this attribute values lie around 40.33.
\item Cylinders attributes are near equally distributed around 6.75 and 4.25.
\item Most cars with mpg around 27.79 have cylinders around 4.25 and most cars with mpg around 15.25 have cylinders around 6.75.
\item All cars with cylinders around 4.25 have displacement around 132.5.
\item Cars with cylinders around 6.75 have displacement mainly around 261.5 and 390.5.
\end{enumerate}
In Figure~\ref{fig:experiment2}, the trends and data distributions between the axes of horsepower, weight, and acceleration are still not clear enough. In the visual analytics process, this could be improved by taking advantage of human perception and reasoning abilities. After obtaining a perception of data trends and the distributions between axes of horsepower, weight, and acceleration based on Figure~\ref{fig:experiment2}, we configure the new clustering of these axes, which is shown in Figure~\ref{fig:experiment3}. Comparing Figure~\ref{fig:experiment3} and Figure~\ref{fig:experiment2}, the pattern and data distribution in Figure~\ref{fig:experiment3} made with human judgment are better revealed compared to Figure~\ref{fig:experiment2}. It shows that a big proportion of cars have horsepower around 73 and a relatively small proportion of cars have horsepower around 150. Most cars weigh between 2000 and 4000, and have acceleration attributes around 15.
\begin{figure}[tb]
\centering
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/ex}
\caption{\label{fig:experiment1}
\textit{Classic parallel coordinate plot.}}
\end{subfigure}
\par\bigskip
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/e1}
\caption{\label{fig:experiment2}
\textit{With edge-merging layout and automatic data clustering.}}
\end{subfigure}
\par\bigskip
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/e2}
\caption{\label{fig:experiment3}
\textit{With edge-merging layout and manually clustering.}}
\end{subfigure}
\caption{\label{fig:experiment}
\textit{Experiments on a car data set with seven dimensions and 392 data items.}}
\end{figure}
Our technique is scalable for large-scale data. Figure~\ref{fig:carm} shows a synthesized data set based on the car data set, which has seven dimensions and $10^{5}$ data items. In Figure~\ref{fig:carm1}, the data set is visualized in a classic PCP. In Figure~\ref{fig:carm2}, the data set is visualized in a PCP with edge-merging layout and three clusters in each dimension. The comparison between Figure~\ref{fig:carm1} and~\ref{fig:carm2} clearly shows that our technique reveals the data trends, patterns and distributions between clusters, and greatly reduces visual clutter.
To further illustrate the effectiveness of our edge-merging technique, we analyze its running time of its clustering and rendering process with several synthesized five-dimensional data sets. The synthesized data sets were randomly generated based on the wholesale customers data set and the car data set. The application was implemented on a MacBook Pro desktop with 3.1 GHz Intel Core i5 CPU, 8GB Memory and Intel Iris Plus 650 Graphics Processing Unit (GPU).
\begin{figure}[tb]
\centering
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/carm}
\caption{\label{fig:carm1}
\textit{Classic parallel coordinate plot.}}
\end{subfigure}
\par\bigskip
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{imgs/carm2}
\caption{\label{fig:carm2}
\textit{With edge-merging layout and three clusters in each dimension.}}
\end{subfigure}
\caption{\label{fig:carm}
\textit{Experiments on a synthesized car data set with seven dimensions and $10^{5}$ data items.}}
\end{figure}
First, a run-time analysis of the clustering algorithm with the synthesized data sets was performed. The clustering algorithm includes the process of computing the clusters and the densities (described in Section~\ref{section:dataclustering} and~\ref{section:edgemerging}). In the test, the default number of clusters was set to five. For each data set with different data items, the clustering algorithm was executed ten times and the average running time was calculated. According to the result shown in Table~\ref{tab:run-time}, the proposed clustering algorithm is scalable for large-scale data. Its running time increases linearly with the number of data items. Furthermore, the running time of the proposed clustering algorithm is much shorter compared with other clustering algorithms used in PCP. For example, Palmas et al. \cite{Palmas2014} present a density-based clustering for each dimension independently in edge-bundling PCP, which takes about one minute for clustering a one-dimensional data set with $10^{5}$ data items. In contrast, our algorithm takes 1.20980 seconds for clustering a five-dimensional data set with $10^{5}$ data items, which allows the interactive use of it in the visual analytics process.
\begin{table}[tb]
\caption{\textit{Run-time analysis of the clustering algorithm.}}
\label{tab:run-time}
\centering%
\begin{tabu}{%
c c
}
\toprule
Number of data items & Run-time (s) \\
\midrule
$10^{3}$ & 0.06820 \\
$10^{4}$ & 0.14610 \\
$5\times10^{4}$ & 0.63520 \\
$10^{5}$ & 1.20980 \\
$5\times10^{5}$ & 6.08090 \\
$10^{6}$ & 15.01720 \\
\bottomrule
\end{tabu}%
\end{table}
Next, we analyze the rendering time of our edge-merging PCP based on the synthesized data sets. The clustering time described above depends on the performance of the server where data processing is running. In contrast, rendering time needed to draw the chart in the web browser depends on the computing power of the user's computer. As described earlier, the proposed solution is scalable - the clustering and rendering processes can be distributed on different computers. To make the comparison of rendering time in our edge-merging PCP and a classic PCP fair, only time for rendering the data items is included. The time for rendering the axes, labels, and stickers are not included - this part is static and requires a constant rendering time regardless of the number of data items. The result is shown in Table~\ref{tab:render-time}. According to Table~\ref{tab:render-time}, the rendering time of our edge-merging PCP is much shorter compared to that of classic PCP. It allows our edge-merging PCP to be interactively used in the visual analytics process. More importantly, the rendering performance of our edge-merging PCP remains nearly constant when the number of data items increases. In fact, its rendering depends only on the cluster configuration, not the number of data items. It makes our edge-merging technique scalable for large-scale data sets. For data sets with more than $10^{5}$ items, the classic PCP takes more than one minute and often causes a crash in browsers. In contrast, our edge-merging PCP still takes less than 0.001 seconds for rendering these data sets.
\begin{table}[tb]
\caption{\textit{Rendering time analysis of the edge-merging layout.}}
\label{tab:render-time}
\centering%
\begin{tabu}{%
M{2cm} M{2.5cm} M{2.5cm}
}
\toprule
Number of data items & Rendering time (s) of the classic PCP & Rendering time (s) of the edge-merging PCP \\
\midrule
$10^{3}$ & 0.03279 & 0.00097 \\
$10^{4}$ & 0.28104 & 0.00088 \\
$5\times10^{4}$ & 1.50314 & 0.00092 \\
$10^{5}$ & 3.00769 & 0.00094 \\
\bottomrule
\end{tabu}%
\end{table}
\section{Discussions} \label{section:discussion}
From the experiments, we can see that our edge-merging PCP has clear advantages for the visual analytics process, especially when the size of data is scaled up. Firstly, it greatly reduces visual clutter. It helps to reveal the trends, hidden patterns and data distributions between axes. Furthermore, compared with other edge-bundling PCP, it supports the interactive use in the visual analytics process for large-scale data. The performance is acceptable for near-real-time interaction. Its computing time of the clustering process is linearly increased with the number of data items, which is much shorter than other edge-bundling PCP optimizations. Its rendering performance is independent of the number of data items. Therefore, our edge-merging PCP is scalable for large-scale data. Finally, it allows users to integrate their perception and reasoning abilities into the data analysis process by supporting the configurable clustering.
Although our application supports the interactions of changing the order of axes, adding and deleting axes, and configuring the number and size of clusters, one limitation of the current version is that all interactions are performed through a control panel, which is not intuitive for users.
\section{Conclusion and Future Work}\label{section:future}
In this paper, we present a novel edge-merging PCP for visual exploration of large-scale multi-dimensional data sets. Based on the edge-merging PCP, we implement a web-based visual analytics application, which could be extended with other big data platforms to supports extreme large-scale data sets. In the edge-merging technique, we propose a one-dimensional clustering method which is evolved from K-means. The edge-merging PCP merges data points based on the clustering to reduce visual clutter caused by overplotting. It uses density information to render the merged B\'ezier splines to reveal the trends and data distributions between axes, which leaves the color and opacity perceptual channels free for visualizing other information. Furthermore, its clustering algorithm has linear complexity and the rendering performance of its edge-merging layout is independent of the number of data points, which makes it available for the interactive usage in the visual analytics process. More importantly, this visual analytics application integrates human perception and reasoning abilities into the data analysis process by supporting the configuration of the clustering in each axis to continuously improve the visualization results.
The work in this paper can be further improved in several ways. In the current implementation, the interactions are performed through a control panel. We plan to develop more intuitive interactions for this framework. In addition, a controlled user study could be done in the future to evaluate the usability of the framework. To improve the visualization, the color and opacity perceptual channels could be investigated for display of additional information.
\section{Introduction}
Parallel coordinates plots (PCP) has become a standard tool for visualizing multi-dimensional data~\cite{Hei2013}. It has been incorporated into several visual analytics applications for exploring multi-dimensional data sets, such as iPCA~\cite{Jeo2019} and iVisClassifier~\cite{Cho2010}. In PCP, axes, that corresponding to dimensions of the original data, are aligned parallel to each other and data points are mapped to lines intersecting the axes at the respective value. The embedding of an arbitrary number of parallel axes into the plane allows the simultaneous display of many dimensions, providing a good overview of the data, which reveals patterns and trends in the data. However, it often creates visual clutter, especially for large-scale data sets. This hampers the recognition of patterns and trends in the data. In addition, another related issue of PCP for large-scale data sets is overplotting, in which lines are plotted on top of each other. This hides information in the data.
Edge bundling is a widely used approach to reduce visual clutter in PCP. The data will be aggregated first with data clustering algorithms, then rendered in an illustrative fashion using different forms of edge bundling. However, existing data clustering and edge-bundling techniques used in PCP are not scalable well for large-scale data. Their rendering performance depends on the number of data items. For millions of data items, they have been too slow for interactive use without any hardware acceleration, for example, in web-based visual analytics, which hinders the integration of human judgment into the data analysis process. Furthermore, they do not support to modify both the number and the size of the clusters generated by automatic algorithms, which hinders their use in visual analytics. Moreover, they do not address the issue of overplotting.
Web-based interactive visual analytics is a widely used in various application domains. Many researches have been conducted to explore web-based interactive visual analytics for various types of data, such as miRTarVis+~\cite{Seh2017}, GraphVis~\cite{Ahm2015}, and a web-based visual analytics system for real estate data~\cite{Sun2013}. However, existing web-based visual analytics methods are not scalable well for large data. The time to process data and render visualizations hinders their interactive use in web browsers when the size of data sets is scaled up.
In this paper, we propose a confluent-drawing-based PCP approach for web-based visual analytics of large multi-dimensional data sets. It maps multi-dimensional data to node-link diagrams to reduce unambiguous visual clutter through a data binning-based clustering for each dimension. It visualizes the data in density-based confluent drawing to address overplotting. It is scalable for large data sets and supports interactive use in the web-based visual analytics. Based on the proposed confluent-drawing PCP, we implement a web-based visual analytics application for large multi-dimensional data sets. The experiments on several data sets are conducted to demonstrate the scalability of the proposed approach. In addition, a user study is conducted to compare its effectiveness against another parallel coordinates bundling technique. Results show that it enhances the web-based interactive visual analytics of large multi-dimensional data in the following ways:
\begin{enumerate}
\item To the best of our knowledge, our confluent-drawing PCP is the first PCP approach that supports interactive use on large data sets (e.g. millions of data items) in web browsers without hardware-accelerated rendering and big data infrastructure-based data processing.
\item It reduces visual clutter and overplotting through data binning-based clustering for each dimension and density-based confluent drawing, which eliminates the ambiguity in other bundling techniques that bundle edges by spatial proximity.
\item The rendering time of our method is independent of the number of data items. By applying confluent drawing on PCP, it merges the "links" before rendering them and uses their densities to render their widths to reveal the data distributions between axes.
\item Our method can be used to detect anomalies that have extremely small proportions in large multi-dimensional data sets.
\item To support visual analytics of large multi-dimensional data, our method provides interactions to directly modify both the number and the size of the clusters in the visualization.
\end{enumerate}
The paper is organized as follows:~\autoref{section:related work} reviews related work. \autoref{section:method} presents the method. \autoref{section:experiment} conducts the experiments and user study, and discusses the result. \autoref{section:future} draws conclusions.
\section{Related Work}\label{section:related work}
Since parallel coordinates were introduced in the seminal work of Inselberg~\cite{Ins1985,Ins1987}, and later extended by Wegman~\cite{Weg1990} as a visualization tool, it has become a common and widely used visualization technique for visualizing multi-dimensional data. For large-scale data sets, visual clutter and overplotting are often cited as major challenges for parallel coordinates~\cite{Hei2013}. Use data clustering algorithms to group data items and display the clusters in extensions of PCP with techniques such as edge bundling is a usual way to reduce visual clutter. Fua et al.~\cite{Fua1999} develop a multiresolutional view of the data via hierarchical clustering, and use a variation on parallel coordinates to convey aggregation information for the resulting clusters. To visualize a large number of data items without hiding the inherent structure they constitute, Johansson et al. \cite{Joh2005} combine clusters and high-precision textures in PCP to represent data, in which specific transfer functions that operate on the high-precision textures are used to highlight different aspects of the cluster characteristics.
The edge-bundling method that uses cubic B-splines to show adjacency relations atop different tree visualization methods was proposed for the first time by Holten et al. \cite{Hol2006}. It has been adapted to PCP based on data clustering by McDonnell and Mueller~\cite{Mcd2008}. With edge bundling, polylines in PCP become polycurves, which are bent between two axes towards centroid of the corresponding clusters. While this approach reduces visual clutter by freeing up screen space, the resulting polycurves are still $C^{0}$-continuous, which makes them difficult to be visually traced. To avoid this, Heinrich et al.~\cite{Hei2011} use a $C^{1}$-continuous bundling. Furthermore, Zhou et al.~\cite{Zho2008} present a bundling method that creates visual clusters without the prerequisite of data clustering.
Since edge bundling was introduced, multiple approaches have been proposed for diverse purposes, such as ink saving, reducing ambiguity or revealing information hidden in the visualization~\cite{Lhu2017}. To visualize non-planar graphs in a planar way and reduce ambiguity, Dickerson et al.~\cite{Dic2003} propose an unambiguous edge bundling called confluent drawing, which allows groups of edges to be merged together and drawn as "tracks" to facilitate tasks where the user needs to follow lines. Luo et al.~\cite{Luo2008} introduce a model for curve bundles in PCP to reduce visual clutter and display inherent structure in the data by vertically redistributing and separating the clusters. Palmas et al.~\cite{Pal2014} present an edge-bundling layout for PCP using density-based clustering for each dimension independently so that the clustering is directly related to the shown dimensions in every part of the plot. Do Amor Divino Lima et al.~\cite{Do2018} propose an edge-bundling technique to visually encode the clusters information of each dimension, such as variance, means, and quartiles, into the curvature of lines.
Existing work on edge bundling concentrates on combining various clustering algorithms and different forms of edge bundling with parallel coordinates. There is a lack of works in the literature about edge bundling and PCP that focus on supporting their interactive use for visual analytics of large-scale data sets, especially, without hardware accelerations. The long computation and rendering time makes them too slow for interactive use in the visual analytics process. For example, in Palmas et al.'s~\cite{Pal2014} method, to reduce the computation time for the clustering, they sample the parameters 2000 times between 25\% and 1\% of the data range instead of computing it on the whole data range. However, this precomputation still needs almost one minute for $10^{5}$ data points on desktop hardware in a single computing thread. In addition, its rendering performance depends on the number of data items, which hinders the interactive exploration for visual analytics when the number of data items is scaled up. In contrast, our method takes advantage of human expertise by supporting directly modifying both the number and the size of clusters of millions of data items in the visualization. Its rendering performance is independent of the number of data items.
Density representation is another usual way to address overplotting in PCP~\cite{Mil1991,Weg1997}. Artero et al.~\cite{Art2004} highlight significant relationships between axes in a PCP by constructing a density plot based on density information computed from the data set. Heinrich et al.~\cite{Hei2009} derive a mathematical density model to generate a continuous parallel coordinates plot. Kosara et al.~\cite{Kos2006} introduce a method called parallel sets that uses parallelograms to visualize categorical data in PCP, where the thickness of the parallelogram reveals the number of data points that are in both of the two connected categories. To reveal the internal structure of clusters and the data distribution between axes, we follow~\cite{Kos2006} in using the densities of bundles to render their widths. It is effective even with a purely geometry-based visualization, which leaves the color and opacity perceptual channels free for visualizing other information.
The ever-increasing amount of data poses a significant challenge for interactive visual analytics, especially, for the web-based interactive visual analytics. Sansen et al.~\cite{San2017} present a system for visual exploration of large multi-dimensional data with parallel coordinates based on a big data infrastructure, which employs a hybrid computing method to accommodate pre-computing time and hardware-accelerated rendering to reduce rendering time. However, their method is hard to use in the web-based interactive visual analytics without the big data infrastructure and hardware-accelerated rendering.
\section{Method} \label{section:method}
With our confluent-drawing PCP technique, the data is clustered through a data binning-based clustering method for each dimension (\autoref{section:dataclustering}). After clustering, it first maps the multi-dimensional data into a node-link diagram and then uses density-based confluent drawing to visualize the data within B\'ezier splines in PCP (\autoref{section:edgemerging}). In addition, it provides interactions to support the web-based interactive visual analytics (\autoref{section:interaction}).
\subsection{Data Binning-Based Clustering} \label{section:dataclustering}
Data binning is the process to group a number of more or less continuous values into a smaller number of given data intervals (also called "bins"), which transforms numerical variables into categorical counterparts. We use data binning to cluster the data points for each dimension (continuous variable) with the following two considerations:
\begin{enumerate}
\item In a parallel coordinates plot, the data points in each axis are ordered and one data point will be grouped into only one cluster.
\item For a complex data set, to take advantage of human reasoning, perception, and cognition, users may prefer to specify the clusters based on their own knowledge than an automated algorithm.
\end{enumerate}
With the first consideration, for a specified cluster number $k$, the clusters in one axis are also ordered and there is no overlap between clusters, which is illustrated in \autoref{fig:clustersinaxis}. In \autoref{fig:clustersinaxis}, we use a center and a radius to define a cluster in one axis, where the gray rectangles represent clusters, the blue dots represent the centers of the clusters, and the green lines represent the radius of the clusters. With this definition, $k$ clusters in one axis for a given data set $D$ are expressed as follows:
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{imgs/axis-cluster}
\caption{\label{fig:clustersinaxis}
\textit{Clusters in one (horizontal) axis of a PCP.}}
\end{figure}
\[
\left\lbrace (C_{1},R_{1}), (C_{2},R_{3}), . . ., (C_{k},R_{k})\right\rbrace
\]
which fulfills the following constrains:
\[
C_{i}<C_{i+1}, R_{i} \leq C_{i+1} - R_{i+1} - C_{i}, \sum_{i=1}^{k} R_{i} \leq (d_{max}-d_{min})/2
\]
where $C_{i}$ is the center of the $i$-th cluster, $R_{i}$ is the radius of the $i$-th cluster, $d_{max}$ is the maxima of the data set $D$, and $d_{min}$ is the minima of the data set $D$.
To use this data binning method in our confluent-drawing PCP, we first specifically select $k$ centers which evenly group the data points into $k$ clusters (bins) for each dimension. Accordingly, in each dimension, the $k$ ordered clusters with the same radius can be obtained without iterations through the following computation:
\[
C_{i}=d_{max}-R\times(1+2\times i), R = \frac{(d_{max}-d_{min})}{2 \times k}
\]
This computation only relies on the maxima, minima of the data points and the number of the clusters for each dimension. This will eliminate most of the computational time required for other clustering methods used in edge-bundling PCP, such as Gaussian kernel density estimation [19] and DBSCAN [4], especially for large data sets.
Furthermore, with the second consideration, to take advantage of human perception and reasoning abilities, our method allows users to directly modify the number and size of clusters in the visualization with designed interactions, which is demonstrated in \autoref{section:interaction}.
Categorical variables are not clustered using the above method. Instead, we treat each category as a cluster. Using this method, the time of computing arbitrary clusters for a large data set could be ignored with the known maxima and minima.
\subsection{Density-Based Confluent Drawing on PCP} \label{section:edgemerging}
Based on the result of the clustering process, our method maps multi-dimensional data into a node-link diagram, where the clusters are mapped as the "nodes" and the data points that are in both of the two connected clusters are mapped as the "links". Then, confluent drawing is applied on the node-link diagram to merge the links. Finally, the node-link diagram with confluent drawing is visualized on PCP, where the merged links are rendered as B\'ezier splines to guarantee $C^{1}$ continuous across axes. During the confluent drawing process, the number of the links between two nodes is counted and used to calculate the density of the merged link. When visualizing the node-link diagram on PCP with confluent drawing, the densities of merged links are used to render their widths to reveal the data distributions between axes. This complete process is shown in~\autoref{fig:vis_process}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{imgs/vis_process}
\caption{The complete process of confluent drawing on PCP.}
\label{fig:vis_process}
\end{figure}
As the process of mapping clusters to the node-link diagram shown in~\autoref{fig:vis_process}, the clusters in each axis are mapped as nodes, and the data points that belong to a pair of clusters are mapped as links connecting two nodes. To group the data points within two neighbor axes into cluster pairs, we use the following algorithm:
\begin{align*}
& for \; all \; data \; points \; in \; two \; neighbor \; axes \; :\\
& \quad if \; a \in C_{A}^{i} \; and \; b \in C_{B}^{j}\\
& \quad then \; (a,b) \in P(C_{A}^{i}, C_{B}^{j})
\end{align*}
where, $A$ and $B$ are the names of the axes, $a$ and $b$ are the data points in $A$ and $B$, $C_{A}^{i}$ is the $i$-th cluster in the axis $A$, $C_{B}^{j}$ is the $j$-th cluster in the axis $B$, and $P(C_{A}^{i}, C_{B}^{j})$ is the pair of two clusters.
Confluent drawing is a technique for bundling links in node-link diagrams. It coalesce groups of lines into common paths or bundles based on network connectivity to reduce edge clutter in node-link diagrams~\cite{Dic2003, Bac2017}. This eliminates the ambiguity in other bundling techniques that bundle edges by spatial proximity. Our method uses confluent drawing to merge the links that belong to cluster pairs in the node-link diagram, which is shown in \autoref{fig:vis_process} as confluent drawing process.
In addition, our method computes the density for each bundle as:
\[
D_{i,j} = \frac{N(C_{A}^{i}, C_{B}^{j})}{\sum_{i=1}^{i}\sum_{j=1}^{j}N(C_{A}^{i}, C_{B}^{j})}
\]
where, $D_{i,j}$ is the density of one bundle, $N(C_{A}^{i}, C_{B}^{j})$ is the number of data points (links) that belong to the cluster pair $P(C_{A}^{i}, C_{B}^{j})$.
The density of a bundle is used to render its width to reveal the data distribution between axes, which is shown in \autoref{fig:vis_process} as the process of rendering bundles' widths. The width of each bundle is computed as:
\[
W_{i,j} = D_{i,j} \times W_{max}
\]
where, $W_{i,j}$ is the width of one bundle, and $W_{max}$ is the max bundle width in the whole visualization. $W_{max}$ is configurable for different data sets. Therefore, for two neighbor axes, the width of a bundle represents the proportion of data items that are in both of the two connected clusters. For each pair of clusters, the larger the proportion of data items is, the wider the bundle is. The smaller the proportion of data items is, the narrower the bundle is. This allows our method to detect anomalies (rare items or observations which raise suspicions by differing significantly from the majority of the data~\cite{Liu2009}) in a large multi-dimensional data set through visualizing the extremely narrow bundles in a different style, for example, a dashed line, which is demonstrated in~\autoref{section:experiment_result}.
Finally, the node-link diagram with confluent drawing is visualized on PCP, which is shown in \autoref{fig:vis_process} as the process of mapping to PCP. In classic PCP, data points are drawn as polylines, which loses visual continuation across axes. To guarantee the $C^{1}$ continuous across axes, our method displays the bundles as B\'ezier splines, which is a common method in PCP.
\subsection{Web-Based Interactive Visual Analytics} \label{section:interaction}
In addition to the typical interactions of classic PCP, such as changing the order of axes, our method provides interactions that are specifically designed for its data binning-based clustering process to support interactive visual analytics. To support these interactions, it adds control points between clusters in each axis, which is shown in \autoref{fig:control_points}. The interactions are described as follows:
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\linewidth]{imgs/control_points}
\caption{\label{fig:control_points}
The interactions for interactive visual analytics.}
\end{figure}
\begin{description}
\item [Change the order of the axes.] One-click on the label of an axis to select the axis and drag it to a new position. Then the order of the axes will be rearranged and the visualization will be re-rendered.
\item [Modify the size of clusters.] One-click on a control point and drag it along the axis to a new position on the axis. Then the size of the clusters that are adjacent to it will be modified and the visualization will be re-rendered.
\item[Add a cluster.] Double-click on a cluster. Then a control point will be added at the clicked position to split the original cluster and the visualization will be re-rendered based on the new clustering results.
\item[Delete a cluster.] Double-click on a control point. Then the control point will be deleted, the two clusters that are adjacent to it will be merged into a bigger cluster, and the visualization will be re-rendered based on the new clustering result.
\end{description}
Without hardware-accelerated rendering and big data infrastructure-based data processing, the scalability of our method supports its interactive use in the web-based visual analytics of large multi-dimensional data. Its scalability is demonstrated in \autoref{section:experiment_result}. We implement our method in a web-based visual analytics application using Java with Spring MVC on the server side and Data-Driven Documents (D3.js) on the client side. It supports visual analytics by combining data clustering, visualization and human judgment through user interactions. Its architecture and knowledge generation model of visual analytics are shown in \autoref{fig:va_process}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{imgs/process}
\caption{\label{fig:va_process}
The visual analytics and knowledge generation process of the visual analytics application.}
\end{figure}
\autoref{fig:system} is the screen-shot of the web-based visual analytics application. It shows the confluent-drawing PCP on four dimensions from the cars data set~\cite{Cars}. In this PCP layout, the total density of all bundles between each two neighbor axes is 1, therefore, the widths of the bundles reflect the distribution of the data points within the axes.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{imgs/system}
\caption{\label{fig:system}
Confluent-drawing PCP on four dimensions from the cars data set.}
\end{figure}
Please see the video in the supplemental material for an impression of our application and the interactions mentioned above.
\section{Experimental Results and User Study} \label{section:experiment}
In this section, we evaluate the scalability and the effectiveness of the proposed approach through experiments and a user study on the office occupancy detection data set (five attributes and 20560 data points for each attribute)~\cite{Can2016} and the cars data set (seven attributes and 392 data points for each attribute)~\cite{Cars}. The office occupancy detection data set uses the measurement data of temperature, humidity, light, and CO2 to detect the occupancy of an office room. The cars data set consists of data on cylinders, horsepower, weight, etc. for cars. In addition, to examine the scalability of our method on large data sets, we also synthesized several larger data sets based on the office occupancy detection data set. All the experiments were performed on a MacBook Pro desktop with 3.1 GHz Intel Core i5 CPU, 8GB Memory and Intel Iris Plus 650 Graphics Processing Unit (GPU) without hardware-accelerated computing and rendering.
\subsection{Experimental Results}\label{section:experiment_result}
Two experiments are performed to evaluate the scalability of the proposed method. First, a run-time analysis of the data binning-based clustering method was performed. The clustering process includes calculation of the clusters for each dimension and the densities of bundles. The result of the run-time analysis is shown in \autoref{tab:run_time_clustering}, which shows the computation time of our method is linearly dependent on the number of dimensions and the number of data points, and independent on the number of clusters. Furthermore, the computation time is much shorter compared with other clustering algorithms used in PCP. For example, Palmas et al. \cite{Pal2014} present a density-based clustering for each dimension independently in edge-bundling PCP, which takes about one minute for clustering a one-dimensional data set with $10^{5}$ data items. In contrast, our algorithm takes 0.023 seconds for clustering a four-dimensional data set with $10^{5}$ data items, which allows the interactive use of it in visual analytics.
\begin{table}[tb]
\caption{Run-time analysis of the data binning-based clustering.}
\label{tab:run_time_clustering}
\scriptsize%
\centering%
\begin{tabu}{%
*{3}{c}%
l%
}
\toprule
\rotatebox{90}{Number of} \rotatebox{90}{dimensions} & \rotatebox{90}{Number of} \rotatebox{90}{data points} & \rotatebox{90}{Number of} \rotatebox{90}{clusters} & \rotatebox{90}{Run-time} \rotatebox{90}{(in seconds)} \\
\midrule
2 & $10^4$ & 3 & 0.002 \\
2 & $10^4$ & 4 & 0.002 \\
2 & $10^5$ & 3 & 0.017 \\
2 & $10^5$ & 4 & 0.018 \\
3 & $10^4$ & 3 & 0.0045 \\
3 & $10^4$ & 4 & 0.004 \\
3 & $10^5$ & 3 & 0.022 \\
3 & $10^5$ & 4 & 0.022 \\
4 & $10^5$ & 3 & 0.023 \\
4 & $10^5$ & 4 & 0.03 \\
4 & $10^6$ & 3 & 0.12 \\
4 & $10^6$ & 4 & 0.123 \\
\bottomrule
\end{tabu}%
\end{table}
Next, we examined the efficiency of our method by comparing its rendering time with both classic PCP and do Amor Divino Lima et al.'s edge-bundling PCP~\cite{Do2018} using the office occupancy detection data set. All methods are implemented in D3.js in Chrome (version: 73.0.3683.103, 64-bit). \autoref{fig:pcp_comparison} shows the comparison of these three different PCP layouts. Particularly, \autoref{room3} is an example of anomaly detection by using our method. In~\autoref{room3}, the extremely narrow bundles are visualized in dashed lines to represent anomalous data items that has extremely small proportions between each two neighbor axes.
To make the comparison fair, only time for rendering the data items is included. The time for rendering the axes, labels and stickers are not included - this part is static and requires a constant rendering time regardless of the number of data items. The result is shown in \autoref{tab:run_time_rendering}, in which the rendering time is counted in seconds.
\begin{table}[tb]
\caption{Comparison of rendering time of different PCP layouts.}
\label{tab:run_time_rendering}
\scriptsize%
\centering%
\begin{tabu}{%
*{2}{c}%
*{2}{l}%
}
\toprule
\rotatebox{90}{Number of} \rotatebox{90}{data points} & \rotatebox{90}{Rendering time of} \rotatebox{90}{our method} & \rotatebox{90}{Rendering time of} \rotatebox{90}{classic PCP} & \rotatebox{90}{Rendering time of} \rotatebox{90}{edge-bundling PCP} \\
\midrule
$10^3$ & 0.00115 & 0.0261 & 0.0735 \\
$10^4$ & 0.00136 & 0.1505 & 0.92712 \\
$5\times10^4$ & 0.00137 & 0.65785 & 1.53715 \\
$10^5$ & 0.00117 & 1.44726 &26.3704 \\
$10^6$ & 0.00136 & N/A & N/A \\
\bottomrule
\end{tabu}%
\end{table}
\begin{figure*}[tb]
\centering
\subfloat[The office occupancy data set visualized in the classic PCP\label{room1}]{%
\includegraphics[width=0.48\linewidth]{imgs/room1}}
\hfill
\subfloat[The office occupancy data set visualized in do Amor Divino Lima et al.~\cite{Do2018}'s edge-bundling PCP.\label{room2}]{%
\includegraphics[width=0.48\linewidth]{imgs/room2}}
\hfill
\subfloat[The office occupancy data set visualized in our confluent-drawing PCP without anomalies (dashed lines)\label{room3_na}]{%
\includegraphics[width=0.48\linewidth]{imgs/room3_na}}
\hfill
\subfloat[The office occupancy data set visualized in our confluent-drawing PCP with anomalies (dashed lines)\label{room3}]{%
\includegraphics[width=0.48\linewidth]{imgs/room3}}
\hfill
\caption{The comparison of different PCP layouts with the office occupancy data set.}
\label{fig:pcp_comparison}
\end{figure*}
According to \autoref{tab:run_time_rendering}, the classic PCP and do Amor Divino Lima et al.'s edge-bundling PCP take more than 1.4 and 26 seconds for visualizing $10^5$ data items and cause a crash in the browser with $10^6$ data items. In contrast, our method takes about 0.001 seconds for rendering the data sets. The rendering time of our method is much shorter compared to that of classic PCP and do Amor Divino Lima et al.'s edge-bundling PCP. This is because our method maps the data into the node-link diagram and merges data points before rendering them by applying confluent drawing on the data. For given two axes, the maximum of bundles between them only depends on the number of clusters in each axis, which is much smaller and constant when comparing with the number of data points in a large data set. The rendering performance of our method is independent on the number of data items, which makes our method is scalable for large data sets.
\subsection{User Study}
To evaluate the effectiveness of our method, we conducted a comparative user study of our method with do Amor Divino Lima et al.'s edge-bundling PCP~\cite{Do2018}. Two typical tasks, correlation estimation and subset estimation were used in the user study, in which accuracy was used as the sole dependent measure. The study was performed using static images. \autoref{correlationEB} and \autoref{correlationCD} were used in the correlation estimation. \autoref{room2} and \autoref{room3_na} were used in the clusters estimation. The two tasks are described as follows:
\begin{description}
\item [Correlation Estimation.] Participants are asked to estimate the correlation of two variables on a discrete scale from -1 to +1 with the interval of 0.1. +1 denotes positive correlation. -1 denotes negative correlation. 0 denotes that there is no correlation. Example task: Estimate the correlation coefficient of displacement and horsepower in \autoref{correlationEB} and \autoref{correlationCD}.
\item [Subset Estimation.] Participants are asked to estimate the number of subsets over several axes. A subset is a set of clusters that have common data items on different axes, which are shown as a common path over the axes. For example, between the axes of temperature and light, there are 4 subsets in \autoref{room3_na} and 3 subsets in \autoref{room2}.
\end{description}
\begin{figure} [tb]
\centering
\subfloat[The car data set visualized in do Amor Divino Lima et al.~\cite{Do2018}'s edge-bundling PCP.\label{correlationEB}]{%
\includegraphics[width=0.9\linewidth]{imgs/correlationEB}}
\hfill
\subfloat[The car data set visualized in our confluent-drawing PCP\label{correlationCD}]{%
\includegraphics[width=0.9\linewidth]{imgs/correlationCD}}
\caption{The comparison of different PCP layouts with the car data set. Visualization used in the user study for the correlation tasks.}
\label{fig:correlationVis}
\end{figure}
16 students and researchers at a local university participated the user study. Each task was carefully explained to the participants. They divided into two groups. Both groups did the two tasks with the two methods but in different order. One group first did the tasks with our method. Another group first did the tasks with the edge-bundling PCP. In the questionnaire, for each method, there are 5 questions of correlation estimation, and 4 questions of subset estimation. To preprocess the data, we removed the answers of participants who did not complete the tasks, and the answers with an error larger than the mean error plus 2 times standard deviations. For the correlation estimation task, 1 participant's answers are removed (5 out of 80 answers). For the subset estimation task, 10 out of 64 answers are removed.
For each method, to assess the accuracy of the two tasks, we calculated the mean error of all the answers in each task. For each task, the error of an answer is computed as follows:
\begin{description}
\item [Correlation Estimation.] The error is the absolute difference between the participant's answer and the ground-truth correlation (Pearson correlation coefficient).
\item [Subset Estimation.] The error is the absolute difference between the participant's answer and the ground-truth subset number.
\end{description}
\autoref{fig:mean_error} shows the mean errors of the two tasks for each method. According to~\autoref{correlation_error} and~\autoref{subset_error}, in both of the correlation estimation and subset estimation tasks, our method has better performance (smaller mean errors) than do Amor Divino Lima et al.'s edge-bundling PCP. This is because our method eliminates overlaps on clusters and edges, which reveals all information of the data. In contrast, do Amor Divino Lima et al.'s edge-bundling PCP creates the obvious overlaps on clusters on each axis, which is shown in~\autoref{room2} and~\autoref{correlationEB}. \autoref{room2} also shows the overlaps on the edges between the axes of light and occupancy, which is the reason why do Amor Divino Lima et al.'s edge-bundling PCP has a larger mean error for the subset estimation task. Moreover, do Amor Divino Lima et al.'s edge-bundling PCP eliminates much data points in the data set by the clustering method they used, which may lead to the misunderstanding of the whole data set.
\begin{figure} [tb]
\centering
\subfloat[The mean error of the correlation estimation task.\label{correlation_error}]{%
\includegraphics[width=0.8\linewidth]{imgs/correlationT}}
\hfill
\subfloat[The mean error of the subset estimation task. \label{subset_error}]{%
\includegraphics[width=0.8\linewidth]{imgs/subsetT}}
\caption{The comparison of the mean errors.}
\label{fig:mean_error}
\end{figure}
Based on the experiments and user study, we can see that our method has clear advantages for the web-based interactive visual analytics of large multi-dimensional data. With its scalability, it supports real-time interactions on millions of data items without hardware acceleration in a normal web browser. For the typical tasks of PCP, it has better performances than the compared edge-bundling PCP. In addition, based on the proposed clustering method, it can display the major clusters and outliers without the assistance of the color and opacity channels.
\section{Conclusion}\label{section:future}
In this paper, we proposed a confluent-drawing PCP for the web-based visual analytics of large multi-dimensional data. Based on the proposed approach, we implement a web-based visual analytics application. We evaluated the scalability and effectiveness of the proposed approach through the experiments and user study of comparing it with another edge-bundling PCP. Results show that the proposed approach significantly enhances the web-based interactive visual analytics of large multi-dimensional data.
The work in this paper can be further improved in several ways. To improve the visualization, the color and opacity perceptual channels could be investigated to improve its performance. For example, use color to highlight the related subsets (common paths) when hovering over one bundling.
\bibliographystyle{abbrv-doi}
|
\section{Introduction}
Let $\Omega \subset \bbR^n$ be a bounded domain with sufficiently smooth boundary. Denote the Dirichlet and Neumann eigenvalues of $-\Delta$ by
\[
0 < \lambda_1 < \lambda_2 \leq \lambda_3 \leq \cdots
\]
and
\[
0 = \mu_1 < \mu_2 \leq \mu_3 \leq \cdots
\]
respectively. In the one-dimensional case, the interlacing property
\[
\mu_k < \lambda_k \leq \mu_{k+1}
\]
holds for every $k$. The extent to which these inequalities generalize to higher dimensions has received much attention over the years. However, most results impose rather strong assumptions (e.g. convexity) on the geometry on $\p\Omega$, and the optimal results for general domains remain unknown.
\subsection{Survey of known results}
The variational formulation of the eigenvalue problem immediately gives $\mu_k \leq \lambda_k$ for all $k$. It was shown by P\'olya that $\mu_2 < \lambda_1$ always holds \cite{P52}. In the case of a convex planar domain with $C^2$ boundary, Payne \cite{P55} showed that $\mu_{k+2} < \lambda_k$ for any $k$.
By imposing assumptions on various combinations of the principal curvatures of the boundary, Levine and Weinberger \cite{LW86} showed that $\mu_{k+r} < \lambda_k$ for a $C^{2,\alpha}$ domain, where $1 \leq r \leq n$ depends on the geometric assumptions. In particular, $\mu_{k+n} < \lambda_k$ when $\Omega$ is convex. Through a limiting argument, they obtained $\mu_{k+n} \leq \lambda_k$ for a general (not necessarily smooth) convex domain. They also obtained (as a special case) a result of Aviles \cite{A86}, that $\mu_{k+1} < \lambda_k$ when the mean curvature is nonnegative.
More generally, it was shown by Friedlander \cite{F91} that $\mu_{k+1} \leq \lambda_k$ on any domain with $C^1$ boundary. This result was extended to Lipschitz domains by Arendt and Mazzeo \cite{AM12} and even further by Filinov \cite{F04} to domains of finite Lebesgue measure for which the embedding $H^1(\Omega) \subset L^2(\Omega)$ is compact. In fact, the latter two references obtain the strict inequality $\mu_{k+1} < \lambda_k$.
\subsection{Open problems and conjectures}
The question of whether or not $\mu_{k+n} \leq \lambda_k$ holds for non-convex domains remains open.\footnote{A purported counterexample in \cite{LW86} is easily seen to be wrong, as has been pointed out in \cite{BLP09}.} More precisely, given a Dirichlet eigenvalue $\lambda_k$, one can ask how many Neumann eigenvalues, $\mu$, exist with $\mu \leq \lambda_k$, and how this number depends on the geometry of $\Omega$. We investigate this when $k=1$, where little is known beyond the general results described above.
The case $k=1$ is particularly important when studying nodal domains of eigenfunctions. This is because any eigenvalue $\lambda$ is necessarily the first Dirichlet eigenvalue on any nodal domain of its associated eigenfunction. Several applications of this idea are described in Section \ref{sec:apply}.
Given a domain $\Omega$, we define the number
\begin{align}\label{eq:Ndef}
N(\Omega) = \#\{k \in \bbN : \mu_k \leq \lambda_1\}.
\end{align}
It follows from the results surveyed above that $N(\Omega) \geq 2$ for any domain, and $N(\Omega) \geq n+1$ when $\Omega$ is convex.
To obtain further intuition, we use the finite-element method (FEM) to approximate $N(\Omega)$ for a wide variety of planar domains. In every case, our finite-element approximations satisfy $N(\Omega) \geq 3$, whether or not $\Omega$ is convex. In fact, the farther $\Omega$ is from a ball, the larger $N(\Omega)$ is seen to be, bringing to mind the remark of Osserman \cite{O78} that ``One has the somewhat ironic situation that the more irregular the boundary, the stronger will be the isoperimetric inequality, but the harder it is to prove."
We quantify this irregularity using the isoperimetric ratio
\begin{align}
{\mathcal{I}}(\Omega) = \frac{|\p\Omega|^2}{|\Omega|},
\end{align}
where $|\Omega|$ denotes the area of the domain and $|\p\Omega|$ the perimeter. Computing $N(\Omega)$ and ${\mathcal{I}}(\Omega)$ numerically for a wide variety of examples, we arrive at the following conjecture.
\begin{figure}
\includegraphics[scale=0.5]{combined_isoplot_1_20pt}
\caption{$N$ vs. ${\mathcal{I}}$ for planar domains of varying geometry and topology}
\label{fig:combined}
\end{figure}
\begin{conj}\label{con:2D}
There exist constants $c_1,c_2 > 0$ so that
\[
c_1 {\mathcal{I}}(\Omega) \leq N(\Omega) \leq c_2 {\mathcal{I}}(\Omega)
\]
for any bounded Lipschitz domain $\Omega \subset \bbR^2$.
\end{conj}
The computations leading to this conjecture are described in Sections \ref{sec:analytical} and \ref{sec:numerical} and summarized in Figure \ref{fig:combined}.
While our numerical investigations focus on $n=2$, we conjecture that the same result also holds in higher dimensions. For $\Omega \subset \bbR^n$ we define the isoperimetric ratio
\begin{align}
{\mathcal{I}}(\Omega) = \frac{|\p\Omega|^n}{|\Omega|^{n-1}},
\end{align}
where $|\Omega|$ is the Lebesgue measure of $\Omega$, and $|\p\Omega|$ is the $(n-1)$-dimensional Hausdorff measure of the boundary.
\begin{conj}\label{con:general}
There exist constants $c_1,c_2 > 0$, depending only on $n$, so that
\[
c_1 {\mathcal{I}}(\Omega) \leq N(\Omega) \leq c_2 {\mathcal{I}}(\Omega)
\]
for any Lipschitz domain $\Omega \subset \bbR^n$.
\end{conj}
In Theorem \ref{thm:rectangle}, we verify this conjecture for $n$-dimensional rectangles. For the unit ball, $B^n \subset \bbR^n$, we prove in Theorem \ref{thm:ball} that $N(B^n)$ grows faster than any polynomial function of $n$. The isoperimetric ratio satisfies
\[
{\mathcal{I}}(B^n) = \frac{(n\omega_n)^{n}}{\omega_n^{n-1}} = n^n \omega_n,
\]
where $\omega_n = \pi^{n/2} / \Gamma(n/2+1)$ is the volume of the unit ball in $\bbR^n$, and Stirling's approximation for the Gamma function implies $\omega_n \approx C n^{-(n+1)/2}$ for large $n$. Therefore, ${\mathcal{I}}(B^n) \approx C n^{(n-1)/2}$ also has superpolynomial growth in $n$.
To the best of our knowledge, none of the known isoperimetric inequalities for Dirichlet and Neuman eigenvalues are of use in resolving Conjecture \ref{con:general}. Instead, we suggest that more detailed information be sought in the spectrum of the Dirichlet-to-Neumann map, building on the work of Friedlander \cite{F91}. As described in Sections \ref{sec:nodal} and \ref{sec:size}, such information would also improve our understanding of Courant's nodal domain theorem and Yau's conjecture of the size of the nodal set.
\subsection*{Acknowledgments}
The authors would like to thank Ram Band for helpful discussions and comments on this work. G.C. acknowledges the support of NSERC grant RGPIN-2017-04259. The work of S.M. was partially supported by NSERC discovery grants RGPIN-2014-06032 and RGPIN-2019-05692. L.S. was supported by an NSERC Undergraduate Summer Research Award and the aforementioned NSERC grants.
\section{Motivation and implications}\label{sec:apply}
In this section, we explain our motivation for considering the quantity $N(\Omega)$ in the first place, and describe the consequences Conjecture \ref{con:general} would have if true.
\subsection{Relation to the nodal deficiency}\label{sec:nodal}
Our investigation into the quantity $N(\Omega)$ was inspired by a recent result in \cite{CJM17} (see also \cite{BCM19}) on the nodal deficiency of Laplacian eigenfunctions. Combining this with Friedlander's lemma for the Dirichlet-to-Neumann map, we obtain an upper bound on the spectral position of an eigenvalue.
To state the result in its simplest form, let $\phi$ be a Dirichlet eigenfunction of the Laplacian on $\Omega$, corresponding to a simple eigenvalue $\lambda$, and define the regions $\Omega_+ = \{\phi(x) > 0\}$ and $\Omega_- = \{\phi(x) < 0\}$. Since $\lambda$ is simple, it equals $\lambda_k$ for a unique $k \in \bbN$, which we call the spectral position of $\lambda$.
\begin{theorem}\label{thm:position}
The spectral position of $\lambda$ satisfies
\begin{align}\label{kbound}
k \leq N(\Omega_+) + N(\Omega_-),
\end{align}
where the indices on the right-hand side count eigenvalues with either Dirichlet or Neumann boundary conditions on the interior boundaries $\p\Omega_\pm \cap \Omega$ and Dirichlet boundary conditions on $\p\Omega$.
\end{theorem}
This formula remains valid if $\Omega$ is replaced by a compact Riemannian manifold, and a modified version holds for degenerate eigenvalues; see \cite{BCM19} for details.
\begin{proof}
Let $\nu_\pm$ denote the number of connected components of $\Omega_\pm$, so the total number of nodal domains is $\nu = \nu_+ + \nu_-$. On $\Omega_\pm$, the first Dirichlet eigenvalue is $\lambda$, with multiplicity $\nu_\pm$. Thus, for sufficiently small $\epsilon > 0$, $0$ is not an eigenvalue of the operator $\Delta + (\lambda+\epsilon)$ on $\Omega_\pm$, so the Dirichlet-to-Neumann maps $\Lambda_\pm$ are well defined (see \cite{CJM17} for details).
The nodal deficiency of $\phi$ is defined to be $\delta(\phi) = k - \nu$. From \cite[Corollary 2.5]{CJM17}, we have $\delta(\phi) = \Mor(\Lambda_+ + \Lambda_-) \leq \Mor(\Lambda_+) + \Mor(\Lambda_-)$, where $\Mor$ denotes the Morse index, i.e. the number of negative eigenvalues. Moreover, \cite[Lemma 1.2]{F91} implies $\Mor(\Lambda_\pm) = N(\Omega_\pm) - \nu_\pm$ when $\epsilon$ is sufficiently small, so we obtain $
k - \nu \leq \big(N(\Omega_+) - \nu_+\big) + \big(N(\Omega_-) - \nu_-\big)$ and the result follows.
\end{proof}
If Conjecture \ref{con:general} is true, then Theorem \ref{thm:position} implies that highly deficient eigenfunctions have nodal domains with large isoperimetric ratios. For instance, if $\phi_k$ is an eigenfunction with only two nodal domains, then one of them, say $\Omega_+$, must have $N(\Omega_+) \geq k/2$ and hence ${\mathcal{I}}(\Omega_+) \geq c_1k/2$, where the constant $c_1$ only depends on the dimension. This is particular useful when one has a subsequence of eigenfunctions $(\phi_{j_{_k}})$ each having just two nodal domains; see \cite{CH53,L77,S25} for classical examples of such eigenfunctions, and \cite{JZ18} for a recent construction.
In general, denoting the nodal domains by $\Omega_1, \ldots, \Omega_\nu$, we have
\[
k \leq N(\Omega_1) + \cdots + N(\Omega_\nu) = \nu \overline{N(\Omega_\ast)},
\]
where $\overline{N(\Omega_\ast)}$ denotes the mean of the numbers $N(\Omega_1), \ldots, N(\Omega_\nu)$. Generalizing the two-dimensional result of Pleijel \cite{P56}, B\'erard and Meyer showed in \cite{BM82} that
\[
\limsup_{k \to \infty} \frac{\nu(\phi_k)}{k} \leq \gamma(n)
\]
where $\gamma(n) < 1$ is a universal constant that only depends on the dimension. From this we obtain
\begin{align}\label{eq:Pleijel}
\liminf_{k \to \infty} \overline{N(\Omega_\ast)} \geq \frac{1}{\gamma(n)}.
\end{align}
When $n=2$ we have $\gamma(2) = (2/j_{0,1})^2 < 0.7$, so \eqref{eq:Pleijel} yields $\liminf \overline{N(\Omega_\ast)} \geq 1.4$. This is not anything new, since Friedlander has already shown that $N(\Omega_i) \geq 2$ for every $i$, so the same is true of the mean. Thus, for \eqref{eq:Pleijel} to be of any use, one must have $\gamma(n) < 1/2$, as this would imply $\liminf \overline{N(\Omega_\ast)} > 2$. It was shown in \cite{HP16} that the constant $\gamma(n)$ is strictly decreasing in $n$, with $\gamma(n) < 1/2$ for $n \geq 3$. In fact, one has $\gamma(n) \leq C (2/e)^n$ for some uniform constant $C>0$.
\subsection{The size of the nodal set}\label{sec:size}
Recall that $\Sigma = \{\phi(x) = 0\}$ denotes the nodal set of a fixed eigenfunction. A well-known conjecture of Yau suggests that
\begin{align}\label{Yau}
c_1 \sqrt\lambda \leq |\Sigma| \leq c_2 \sqrt\lambda,
\end{align}
where $\lambda$ is the eigenvalue corresponding to $\phi$. This result is known to be true when $n=2$ \cite{B78}, and on real analytic manifolds \cite{DF88}. Recently, Logunov showed that the lower bound on $|\Sigma|$ holds on a smooth manifold of any dimension \cite{L18lower}, and proved the upper bound $|\Sigma| \leq c_2 \lambda^\alpha$ for some constant $\alpha>1/2$ that depends only on $n$ \cite{L18upper}.
We observe here that the upper bound in \eqref{Yau} follows from Conjecture \ref{con:general} and an additional assumption on the nodal domains and nodal sets. Recall that $\Omega_1, \ldots, \Omega_\nu$ denote the nodal domains of an eigenfunction. If we assume that there exist constants $a$ and $b$ such that $\nu |\Omega_i| \geq a$ and $\nu |\p\Omega_i| \leq b |\Sigma|$, uniformly in $\lambda$, then it follows from Theorem \ref{thm:position} and Conjecture \ref{con:general} that
\begin{align*}
k &\leq N(\Omega_1) + \cdots + N(\Omega_\nu) \\
&\leq c_2 \left( \frac{|\p\Omega_1|^n}{|\Omega_1|^{n-1}} + \cdots + \frac{|\p\Omega_\nu|^n}{|\Omega_\nu|^{n-1}} \right) \\
&\leq c_2 b^n a^{1-n} |\Sigma|^n.
\end{align*}
Furthermore, Weyl's law implies $\lambda_k^{n/2} \leq C k$ for some uniform constant $C$, so we obtain $\lambda_k^{n/2} \leq C |\Sigma|^n$, as desired. However, the assumed uniform bounds on the nodal domains and nodal sets are not known.
We conclude by stating an equivalent form of Yau's conjecture in terms of the nodal deficiency.
It follows from the argument of Pleijel \cite{P56} that $c_1 \lambda^{n/2} \leq \delta(\phi) \leq c_2 \lambda^{n/2}$, from which we see that \eqref{Yau} is, in fact, equivalent to the existence of uniform constants $c_1,c_2>0$ such that
\begin{align}
c_1 |\Sigma|^n \leq \delta(\phi) \leq c_2 |\Sigma|^n
\end{align}
for all eigenfunctions $\phi$.
From this equivalence, Logunov's upper and lower bounds on the size of the nodal set immediately imply there exist positive constants $c_1,c_2$ and $\beta = \beta(n)$ such that $c_1 |\Sigma|^{n/\beta} \leq \delta(\phi) \leq c_2 |\Sigma|^n$, where $\beta(2)= 1$ and $\beta(n) > 1$ for $n > 2$.
We find this formulation of Yau's conjecture particularly appealing in light of the formula $\delta(\phi) = \Mor(\Lambda_+ + \Lambda_-)$ for the nodal deficiency, which was proved in \cite{CJM17} and used in the proof of Theorem \ref{thm:position} above. In particular, it suggests an alternate route to studying the conjecture through the spectra of the Dirichlet-to-Neumann maps, $\Lambda_\pm$.
\section{Analytical results}\label{sec:analytical}
In this section, we consider a few separable examples, where the calculations are relatively explicit.
\subsection{Rectangles}\label{sec:rectangle}
We first consider an $n$-dimensional rectangle with side lengths $\ell_1, \ldots, \ell_n$. In this case, we are able to verify Conjecture \ref{con:general}.
\begin{theorem}\label{thm:rectangle}
There exist positive constants, $c_1(n)$ and $c_2(n)$, such that
\[
c_1 {\mathcal{I}}(R) \leq N(R) \leq c_2 {\mathcal{I}}(R)
\]
for any rectangle $R \subset \bbR^n$.
\end{theorem}
\begin{proof}
The Dirichlet eigenvalues are $(m_1\pi/\ell_1)^2 + \cdots + (m_n\pi/\ell_n)^2$
for $m_i \in \bbN$, while the Neumann eigenvalues are given by the same formula but with $m_i \in \bbN_0 = \bbN \cup \{0\}$. Therefore
\begin{align}
N(R) = \#\left\{ (m_1, \ldots, m_n) \in \bbN_0^n : \frac{m_1^2}{\ell_1^2} + \cdots + \frac{m_n^2}{\ell_n^2} \leq \frac{1}{\ell_1^2} + \cdots + \frac{1}{\ell_n^2}\right\}.
\end{align}
This equals the number of nonnegative lattice points in the $n$-ellipsoid with axes $a_i = \ell_i\rho$, where $\rho^2 = \ell_1^{-2} + \cdots + \ell_n^{-2}$.
Denote this ellipsoid by ${\mathcal{E}}(a)$, and let $L = {\mathcal{E}}(a) \cap \bbN_0^n$, so $N(R) = \# L$. For each $m = (m_1,\ldots,m_n) \in L$ let $C_m$ denote the unit cube with smallest vertex at $m$,
and define the set
\[
E = \bigcup_{m \in L} C_m,
\]
so that $N(R) = \# L = |E|$. Finally, let $P \subset \bbR^n$ denote the first quadrant, where all coordinates are nonnegative. We claim that
\begin{align}\label{Econtain}
{\mathcal{E}}(a) \cap P \subset E \subset {\mathcal{E}}(2a) \cap P,
\end{align}
where ${\mathcal{E}}(2a)$ is the ellipsoid with axes $2 a_i = 2 \ell_i \rho$.
Let $x = (x_1,\ldots,x_n) \in {\mathcal{E}}(a) \cap P$ and define $m = \lfloor x \rfloor \in \bbN^n_0$ componentwise. Then
\[
\frac{m_1^2}{\ell_1^2} + \cdots + \frac{m_n^2}{\ell_n^2} \leq \frac{x_1^2}{\ell_1^2} + \cdots + \frac{x_n^2}{\ell_n^2} \leq \rho^2,
\]
hence $m \in L$ and $x \in C_m \in E$. On the other hand, for $m \in L$ we have
\[
\frac{(m_1 + 1)^2}{\ell_1^2} + \cdots + \frac{(m_n + 1)^2}{\ell_n^2} \leq 4 \rho^2,
\]
hence $C_m \subset {\mathcal{E}}(2a) \cap P$. It follows that $E \subset {\mathcal{E}}(2a) \cap P$, so we have verified \eqref{Econtain}, and conclude that the volume $|E|= N(R)$ satisfies
\begin{align}\label{rec:Nbound}
\frac{\omega_n}{2^n} a_1 \cdots a_n \leq |E| \leq \omega_n a_1 \cdots a_n,
\end{align}
where $\omega_n$ is the volume of the unit ball in $\bbR^n$.
We next estimate the isoperimetric ratio ${\mathcal{I}}(R)$. The rectangle has volume $|R| = \ell_1 \cdots \ell_n$, and the boundary has area
\[
|\p R| = 2 \sum_{i=1}^n \ell_1 \cdots \widehat \ell_i \cdots \ell_n,
\]
where $\ell_1 \cdots \widehat \ell_i \cdots \ell_n$ denotes the product with the $i$th term omitted. Note that
\[
\frac{1}{\ell_1^2} + \cdots + \frac{1}{\ell_n^2} = \frac{\sum_{i=1}^n (\ell_1 \cdots \widehat \ell_i \cdots \ell_n)^2}{(\ell_1 \cdots \ell_n)^2}
\]
and hence
\[
a_1 \cdots a_n = \frac{\left(\sum_{i=1}^n (\ell_1 \cdots \widehat \ell_i \cdots \ell_n)^2\right)^{n/2}}{(\ell_1 \cdots \ell_n)^{n-1}}.
\]
Using the inequalities
\[
\frac1n \left( \sum_{i=1}^n x_i \right)^2 \leq \sum_{i=1}^n x_i^2 \leq \left( \sum_{i=1}^n x_i \right)^2
\]
for nonnegative numbers $(x_i)$, we obtain
\begin{align}\label{rec:Ibound}
\frac{1}{2^n n^{n/2}} \frac{|\p R|^n}{|R|^{n-1}} \leq a_1 \cdots a_n \leq \frac{1}{2^n} \frac{|\p R|^n}{|R|^{n-1}}.
\end{align}
Combining \eqref{rec:Nbound} and \eqref{rec:Ibound} completes the proof.
\end{proof}
For a rectangle in $\bbR^2$ we can obtain a more precise result. Assume without loss of generality that the side lengths are $1$ and $\ell \geq 1$. The inequality
\[
m^2 + \frac{n^2}{\ell^2} \leq 1 + \frac{1}{\ell^2}
\]
is only satisfied by $(0,0)$, $(1,0)$, $(1,1)$ and $(0,n)$ with $1 \leq n^2 \leq 1 + \ell^2$, so we have
\begin{align}\label{rec:2D}
\begin{split}
N(R) &= 3 + \lfloor \sqrt{\ell^2+1}\rfloor \\
{\mathcal{I}}(R) &= 4(1+\ell)^2/\ell.
\end{split}
\end{align}
In particular, we see that $N(R) / {\mathcal{I}}(R) \to 1/4$ as $\ell \to \infty$.
\subsection{The unit ball}
Levine and Weinberger observed in \cite{LW86} that $\mu_3 < \lambda_1 < \mu_4$ on the disc in $\bbR^2$, and $\mu_4 < \lambda_1 < \mu_5$ for the ball in $\bbR^3$. Here, we investigate extensions of these inequalities to the unit ball, $B^n \subset \bbR^n$, for $n>3$. Since $B^n$ is strictly convex, \cite[Theorem 2.1]{LW86} implies $\mu_{n+1}(B^n) < \lambda_1(B^n)$. However, we observe numerically that the inequality $\lambda_1(B^n) < \mu_{n+2}(B^n)$ is false for $n > 3$, and prove that the number of Neumann eigenvalues below $\lambda_1(B^n)$ grows faster than any polynomial function of $n$.
The eigenvalue equation $\Delta u + \lambda u$ is separable, with radial solutions given by the so-called ultraspherical Bessel functions, $r^{1-n/2} J_{n/2+\ell-1}(\sqrt\lambda r)$, for $\ell \in \bbN \cup\{0\}$; see \cite{AB93,LS94}. The positive Neumann eigenvalues are, thus, of the form $(p_{n/2,k}^{(\ell)})^2$, where $p_{\nu,k}^{(\ell)}$ denotes the $k$th positive zero of $[x^{1-\nu} J_{\nu+\ell-1}(x)]'$. The corresponding angular solution is a spherical harmonic of degree $\ell$, so eigenvalues corresponding to $\ell=0$ are simple, those corresponding to $\ell=1$ have multiplicity $n$, and those corresponding to $\ell > 1$ have multiplicity
\begin{align}\label{multiplicity}
\binom{n+\ell-1}{n-1} - \binom{n+\ell-3}{n-1}.
\end{align}
(It is immediate that $(p_{n/2,k}^{(\ell)})^2$ has multiplicity at least \eqref{multiplicity}; the fact that the multiplicity is no greater is a consequence of the fact that, for any $m \in \bbN$, the functions $[x^{1-\nu} J_{\nu+\ell-1}(x)]'$ and $[x^{1-\nu} J_{\nu+\ell+m-1}(x)]'$ have no zeros in common, which was established in \cite[Lemma 2.5]{HP16}.)
Similarly, the first Dirichlet eigenvalue is given by $\lambda_1 = \big(j_{n/2-1,1}\big)^2$, where $j_{\nu,1}$ denotes the first positive zero of $J_{\nu}$. It follows from standard properties of Bessel functions (see \cite{LS94}) that $j_{\nu,k} = p_{\nu,k}^{(0)}$.
Since $j_{\nu,1}$ is an increasing function of $\nu$ (provided $\nu>0$, see \cite[p. 508]{W44}), we have $\sqrt{\lambda_1} = j_{n/2-1,1} < j_{n/2,1} = p_{n/2,1}^{(0)}$, and hence $p_{n/2,k}^{(0)} > \sqrt{\lambda_1}$ for all $k$. Moreover, it follows from Dixon's interlacing theorem \cite[p. 480]{W44} that the zeros of $[x^{1-\nu} J_{\nu+\ell-1}(x)]'$ interlace with those of $J_{\nu+\ell-1}(x)$. In particular, $J_{n/2}$ must have a zero between $p_{n/2,1}^{(1)}$ and $p_{n/2,2}^{(1)}$. Therefore $p_{n/2,2}^{(1)} > j_{n/2,1} > \sqrt{\lambda_1}$, and so $p_{n/2,k}^{(1)} > \sqrt{\lambda_1}$ for $k \geq 2$.
On the other hand, the inequality
\begin{align}\label{LSineq}
\frac{2\ell(\nu+\ell)(\nu+\ell+1)}{\nu+2\ell+1} < \big(p_{\nu,1}^{(\ell)}\big)^2 < 2\ell(\nu + \ell)
\end{align}
of Lorch and Szego \cite{LS94} implies that $p_{\nu,1}^{(1)} < p_{\nu,1}^{(\ell)}$ for $\nu>0$ and $\ell \geq 2$. Therefore $(p_{n/2,1}^{(1)})^2$ is the smallest positive Neumann eigenvalue of $B^n$, with multiplicity $n$, and the only other Neumann eigenvalues potentially less than $\lambda_1$ are $(p_{n/2,k}^{(\ell)})^2$ for $\ell \geq 2$. Numerically, we observe that $p_{n/2,1}^{(2)} < j_{n/2-1,1}$ for $n \geq 4$; see Table \ref{table:zeros}. Thus there are more than $n+1$ Neumann eigenvalues below the first Dirichlet eigenvalue whenever $n \geq 4$. In terms of the quantity $N$ defined in \eqref{eq:Ndef}, this means $N(B^n) > n+1$ when $n \geq 4$.
\begin{table}
\begin{center}
\begin{tabular}{c|c|c|c|c}
$n$ & $p_{n/2,1}^{(1)}$ & $p_{n/2,1}^{(2)}$ & $p_{n/2,1}^{(3)}$ & $j_{n/2-1,1}$\\ \hline
2 & 1.84 & 3.05 & 4.42 & 2.40 \\
3 & 2.08 & 3.34 & 4.51 & 3.14 \\
4 & 2.30 & 3.61 & 4.81 & 3.83 \\
5 & 2.52 & 3.86 & 5.09 & 4.49 \\
6 & 2.69 & 4.10 & 5.37 & 5.14 \\
7 & 2.86 & 4.33 & 5.63 & 5.76 \\ \hline
\end{tabular}
\end{center}
\caption{Comparing zeros of ultraspherical Bessel functions with the first zero of $J_{n/2-1}$}
\label{table:zeros}
\end{table}
More precisely, when $p_{n/2,1}^{(2)} < j_{n/2-1,1}$ we obtain an additional Neumann eigenvalue below $\lambda_1$. The angular part of the eigenfunction is a spherical harmonic of degree $\ell$, so the multiplicity is
\[
\binom{n+1}{n-1} - \binom{n-1}{n-1} = \frac12 n(n+1) - 1.
\]
Adding this to the $n+1$ Neumann eigenvalues already known to exist below $\lambda_1$, we find that $N(B^n) \geq \frac12 n(n+3)$.
This argument can be generalized, using the fact that for any fixed $\ell$, we have $p_{n/2,1}^{(\ell)} < j_{n/2-1,1}$ once $n$ is sufficiently large. (For $\ell=3$ we observe numerically that $n=7$ suffices; see Table \ref{table:zeros}.) As a result, we find that $N(B^n)$ grows faster than any polynomial function in $n$.
\begin{theorem}\label{thm:ball}
For any $\ell \in \bbN$ there exists $n_0 \in \bbN$ such that $N(B^n) \geq n^\ell$ for all $n \geq n_0$.
\end{theorem}
\begin{proof}
From the upper bound in \eqref{LSineq} and the lower bound $j_{\nu,1} > \nu$ (see \cite[p. 485]{W44}), we find that $\big(p_{n/2,1}^{(2)}\big)^2 < \lambda_1$ if $2\ell(n/2 + \ell) \leq (n/2-1)^2$. For fixed $\ell$, this will be satisfied by all sufficiently large $n$. For any such $n$, there is thus a Neumann eigenvalue below $\lambda_1$, of multiplicity
\[
\binom{n+\ell-1}{n-1} - \binom{n+\ell-3}{n-1} = \frac{n^\ell}{\ell!} + \text{ lower order terms},
\]
so for large enough $n$, we can guarantee $N(B^n) \geq n^\ell/(2 \ell!)$.
Applying the above argument to $\ell+1$, we obtain $N(B^n) \geq n^{\ell+1}/(2 (\ell+1)!)$ when $n$ is sufficiently large, say $n \geq n_1$. Thus, for $n \geq \max\{n_1,2(\ell+1)!\}$ we have $N(B^n) \geq n^\ell$.
\end{proof}
\section{Numerical results}\label{sec:numerical}
In this section, we present the numerical results that motivated Conjecture \ref{con:2D}.
Using a finite-element method (FEM), we compute $N(\Omega)$ and plot it against ${\mathcal{I}}(\Omega)$ for different choices of $\Omega$. Note that the size of $\Omega$ is irrelevant as both quantities are scale invariant.
\subsection{Overview of FEM}
In general, a finite-element method approximates solutions to either a PDE or the associated eigenvalue problem by projecting the weak form of the problem onto a finite-dimensional subspace \cite{DBraess_2001, SCBrenner_LRScott_1994a}. For a symmetric and positive-definite operator, as considered here, a standard Galerkin projection is typically chosen, with the finite-dimensional approximation space determined by low-order piecewise polynomial basis functions on an appropriate discretization of the domain, $\Omega$. Considering the weak form of the Laplacian eigenvalue problem, finding $u\in \mathcal{V} \subset H^1(\Omega)$ such that
\[
\int_\Omega \nabla u \cdot \nabla v = \lambda \int_\Omega uv \text{ for all }v\in\mathcal{V},
\]
we define a space $\mathcal{V}^h \subset \mathcal{V}$ in terms of a finite-dimensional basis, and simply solve the eigenvalue problem restricted to $\mathcal{V}^h$ in place of the continuum weak form. Once restricted to a finite-dimensional basis set, the approximate eigenvalue calculation becomes a standard matrix eigenvalue problem, for which efficient numerical methods are well-known \cite{YSaad_2011a}. The form above is equally valid for both the Dirichlet and Neumann eigenvalue problems (or combinations of these boundary conditions), simply by appropriately choosing $\mathcal{V} = H^1_0(\Omega)$ and $\mathcal{V} = H^1(\Omega)$, respectively. For the Dirichlet problem, the boundary condition is enforced strongly, as it is automatically satisfied by all test functions in $\mathcal{V}$, whereas, for the Neumann problem, the boundary condition is enforced weakly.
In the calculations that follow, we consider only polygonal domains, $\Omega$. As such, we directly discretize the domain by considering triangulations, $\mathcal{T}^h$, such that $\Omega = \cup_{T\in\mathcal{T}^h} T$, where the union is taken over planar triangles. The superscript $h$ is given to indicate the discrete nature of the triangulation (and the associated approximation space, $\mathcal{V}^h$), and may be taken to be the maximum edge length (or diameter) of any triangle in $\mathcal{T}^h$, for example. The approximation space, $\mathcal{V}^h$, is then defined as
\[
\mathcal{V}^h = \{ u \in C^0(\Omega) \mid u_T \in \mathcal{P}^k(T) \text{ for every }T\in\mathcal{T}^h \},
\]
where $u_T$ is the restriction of $u$ to triangle $T$, and $\mathcal{P}^k(T)$ is the space of polynomials of total degree no more than $k$ on $T$. The calculations below are computed using the FEniCS finite-element package \cite{MAlnaes_etal_2015a}, with eigenvalue calculations performed using the SLEPc package \cite{Hernandez:2005:SSF}.
\subsection{Rectangles}
Let $R_\ell$ denote the rectangle with side lengths of $1$ and $\ell \geq 1$. Our numerical results, shown in Figure \ref{fig:rectanglecomb}, are consistent with the explicit formulas for $N(R_\ell)$ and ${\mathcal{I}}(R_\ell)$ given in \eqref{rec:2D}. In particular, the graph of $N$ vs. ${\mathcal{I}}$ is asymptotic to a line of slope $1/4$ as ${\mathcal{I}} \to \infty$.
\begin{figure}
\includegraphics[scale=0.45]{rectangle_isoplot20pt}
\includegraphics[scale=0.45]{comb_isoplot20pt}
\caption{$N$ vs. ${\mathcal{I}}$ for rectangles (left) and combs (right)}
\label{fig:rectanglecomb}
\end{figure}
\subsection{Combs}
We next consider a family of so-called ``comb" domains. The comb with $m$ teeth, denoted $C_m$, is the union of $m$ $1\times 2$ rectangles with $m-1$ squares with unit side length, as shown in Figure \ref{fig:CW}. It has area $3m-1$ and perimeter $6m$; hence, the isoperimetric ratio ${\mathcal{I}}(C_m) = 36 m^2/(3m-1)$ is approximately linear in $m$. The corresponding values of $N(C_m)$ are shown in Figure \ref{fig:rectanglecomb}.
\begin{figure}
\begin{tikzpicture}[scale=0.9]
\draw[thick] (0,0) -- (5,0) -- (5,2) -- (4,2) -- (4,1) -- (3,1) --
(3,2) -- (2,2) -- (2,1) -- (1,1) -- (1,2) -- (0,2) -- (0,0) -- (5,0);
\end{tikzpicture}
\hspace{2cm}
\begin{tikzpicture}[scale=0.9]
\draw[thick] (0,0) -- (5,0) -- (5,5) -- (0,5) -- (0,0);
\draw[thick] (1,1) -- (1,2) -- (2,2) -- (2,1) -- (1,1);
\draw[thick] (3,1) -- (3,2) -- (4,2) -- (4,1) -- (3,1);
\draw[thick] (1,3) -- (1,4) -- (2,4) -- (2,3) -- (1,3);
\draw[thick] (3,3) -- (3,4) -- (4,4) -- (4,3) -- (3,3);
\end{tikzpicture}
\caption{The comb $C_3$ (left) and the waffle $W_2$ (right)}
\label{fig:CW}
\end{figure}
\begin{figure}
\includegraphics[scale=0.5]{reg_poly_isoplot20pt}
\includegraphics[scale=0.5]{rand_poly_indep_isoplot20pt}
\caption{$N$ vs. ${\mathcal{I}}$ for regular polygons (left) and random polygons (right)}
\label{fig:poly}
\end{figure}
\subsection{Regular polygons}
Let $P_m$ be a regular $m$-sided polygon. One easily calculates the isoperimetric ratio, ${\mathcal{I}}(P_m) = 4m \tan(\pi/m)$, which decreases towards $4\pi$ as $m\to\infty$. On the other hand, it is also known (see, for instance, \cite[Theorem VI.10]{CH53}) that the Dirichlet and Neumann eigenvalues satisfy
\[
\lambda_k(P_m) \to \lambda_k(D), \quad \mu_k(P_m) \to \mu_k(D)
\]
as $m \to \infty$, where $D \subset \bbR^2$ is the unit disc. Since $\mu_3(D) < \lambda_1(D) < \mu_4(D)$, we obtain $\mu_3(P_m) < \lambda_1(P_m) < \mu_4(P_m)$, and hence $N(P_m) = 3$, for all sufficiently large $m$. We observe numerically that it suffices to take $m = 5$; see Figure \ref{fig:poly}.
\subsection{Random polygons}\label{sec:random}
We next consider a family of random polygons, generated by the following algorithm.
First, three distinct random points within the unit square are chosen and ordered counter clockwise in a vertex list: $(v_1,v_2,v_3)$. Then, until the desired number of vertices is achieved, new vertices are added as follows:
\renewcommand{\labelenumi}{\arabic{enumi})}
\begin{enumerate}
\item Generate a random point $P$ distinct from the existing vertices $(v_1,\ldots,v_m)$, and select a random index $i \in \{1,\ldots,m\}$ for a position in the current list of vertices.
\item Check if creating edges between $P$ and the vertices $v_i$ and $v_{i+1}$ would cause any edges in the new polygon to intersect. If no edges intersect, then $P$ is added to the vertex list between $v_i$ and $v_{i+1}$.
\item If an intersection occurs, replace $i$ with $i+1 \ (\operatorname{mod} m)$ and go to step 2.
\item If $P$ cannot be added between any adjacent vertices without causing an intersection, return to step 1. (The necessity of this step is demonstrated by Figure \ref{fig:counter}.)
\end{enumerate}
\begin{figure}
\begin{tikzpicture}[scale=1.4]
\draw[thick] (-0.8,-0.85) -- (1,2) -- (0.5,1) -- (2,-2) -- (1.35,-1) -- (-2,-1)
-- (-0.8,-0.85);
\fill (0.4,-0.2) circle[radius=2pt];
\node at (0.4,0.1) {$P$};
\end{tikzpicture}
\caption{The point $P$ cannot be added between any adjacent vertices without causing an intersection in the resulting seven-sided polygon}
\label{fig:counter}
\end{figure}
Figure \ref{fig:randomsteps} tracks the evolution of $N$ and ${\mathcal{I}}$ throughout this process, for two different realizations of this algorithm. The resulting domains tend to be highly non-convex; some representative 30-sided examples are shown in Figure \ref{fig:random}. In Figure \ref{fig:poly}, we plot $N$ vs. ${\mathcal{I}}$ for twelve such random polygons, generating two each with 5, 10, 15, 20, 25 and 30 sides.
\begin{figure}
\includegraphics[scale=0.5]{rand_poly_step_1_isoplot20pt}
\includegraphics[scale=0.5]{rand_poly_step_2_isoplot20pt}
\caption{Evolution of $N$ and ${\mathcal{I}}$ (as vertices are added) for two realizations of the random polynomial generator}
\label{fig:randomsteps}
\end{figure}
\begin{figure}
\includegraphics[scale=0.5]{rand_poly_30_no1}
\includegraphics[scale=0.5]{rand_poly_30_no2}
\caption{Two of the randomly generated polygons considered in Section \ref{sec:random}}
\label{fig:random}
\end{figure}
\subsection{Non-simply connected domains}
The domains considered above were all simply connected. In this section, we consider two families of domains with holes. In both cases, we observe behaviour consistent with Conjecture \ref{con:2D}, providing evidence that the conjectured bounds on $N(\Omega)$ hold uniformly for $\Omega \subset \bbR^2$, independent of its topology.
The first is the so-called ``square annulus," consisting of a unit square with a smaller concentric square removed from its interior. The results are shown in Figure \ref{fig:annulus}, where the interior side length ranges from 0.1 to 0.9.
\begin{figure}
\includegraphics[scale=0.5]{square_one_hole_isoplot20pt}
\includegraphics[scale=0.5]{square_uniform_holes_isoplot20pt}
\caption{$N$ vs. ${\mathcal{I}}$ for square annuli (left) and waffles (right)}
\label{fig:annulus}
\end{figure}
The next example is the family of ``waffle" domains. The $m$th waffle domain, $W_m$, consists of a square of side length $2m+1$ with $m^2$ evenly spaced unit squares removed from its interior; see Figure \ref{fig:CW} for an illustration of a waffle and Figure \ref{fig:annulus} for the numerical results with $1 \leq m \leq 8$.
\section{Conclusion}
In this paper, we numerically approximated the quantity $N(\Omega)$ for many planar domains, with varying geometry and topology. In all cases, we observed that $N(\Omega)$ is controlled by the isoperimetric ratio, ${\mathcal{I}}(\Omega)$. Based on these observations, we hypothesized that this relationship always holds (Conjecture~\ref{con:2D}). We also suggested that this holds in higher dimensions (Conjecture \ref{con:general}). We also discussed some implications these conjectures would have if they are indeed true. In particular, our conjectures, combined with results from \cite{CJM17}, yield a direction connection between the spectral position of an eigenfunction and the isoperimetric ratio of its nodal sets.
We suggest that these conjecture be approached using the Dirichlet-to-Neumann map formalism in \cite{CJM17,F91}. This approach seems promising in light of results giving isoperimetric control on Steklov eigenvalues; see, for instance, \cite{CEG11}.
In terms of the finite-element experiments presented in this paper, a natural generalization would be the numerical study of domains in higher dimensions, where the geometry and topology can be much more complicated than in the planar case studied here.
\bibliographystyle{amsplain}
|
\section{Introduction}
Gabber's presentation lemma, initially proved by O. Gabber for the base, spectrum of an infinite field in \cite{gabber} (see also \cite{colliot}, \cite{hogadi2018}) plays a fundamental role in the study of ${\mathbb A}^1$- homotopy theory, especially as developed by Morel in \cite{morel2012}.
This lemma may be thought of as an algebro-geometric analogue of the tubular neighbourhood theorem in differential geometry. In \cite{schmidt2018}, this lemma was generalized by J. Schmidt and F. Strunk to the case where the base is a spectrum of a Dedekind domain with infinite residue fields. The goal of this paper is to show that the arguments given in \cite{schmidt2018} can, in fact, be modified to obtain a proof of Gabber's presentation lemma over a general noetherian domain with all its residue fields infinite. The following is the main result of this paper.
\begin{thm}\label{main}
Let $S={\rm Spec \,} (R)$ be the spectrum of a noetherian domain with all its residue fields infinite. Let $X$ be a smooth, irreducible, equi-dimensional $S$-scheme of relative dimension $d$. Let $Z\subset X$ be a closed subscheme, $z$ be a closed point in $Z$ lying over $s\in S$, such that that $dim (Z_s) < dim (X_s)$. Then after possibly replacing $S$ by a Nisnevich neighbourhood of $s$ and $X$ by a Nisnevich neighbourhood of $z$, there exists a map $\Phi=(\Psi,\nu):X\rightarrow {\mathbb A}_{S}^{d-1}\times {\mathbb A}_{S}^1$, an open subset $V\subset{\mathbb A}_{S}^{d-1}$ and an open subset $U\subset \Psi^{-1}(V)$ containing $z$ such that
\begin{enumerate}
\item $Z\cap U = Z\cap \Psi^{-1}(V)$
\item $\Psi_{|Z}: Z\rightarrow {\mathbb A}_{S}^{d-1}$ is finite
\item $\Phi_{|U}: U\rightarrow {\mathbb A}_{S}^d$ is \'{e}tale{}
\item $\Phi_{|Z\cap U}:Z\cap U\rightarrow {\mathbb A}^1_V$ is a closed immersion
\item $\Phi^{-1}(\Phi(Z\cap U))\cap U=Z\cap U$.
\end{enumerate}
\end{thm}
In \cite{schmidt2018} J. Schmidt and F. Strunk, use the presentation lemma to generalize the ${\mathbb A}^1$-connectivity result of F. Morel ( \cite[Theorem 6.1.8]{morel2005}) over Dedekind schemes with infinite residue fields. As an application of Theorem \ref{main}, we observe that the connectivity result holds over any noetherian domain with all its residue fields infinite. To state this result we recall the following standard notation: For a base scheme $S$, let ${\mathcal S}{\mathcal H}^s_{S^1}(S)$
be the model category of sheaves of $S^1$-spectra over $S$. For an integer $i$, let ${\mathcal S}{\mathcal H}^{s}_{S^1 \geq i}(S)$ be
the full subcategory of $i$-connected spectra. Let
$ {\mathcal S}{\mathcal H}^s_{S^1}(S) \xrightarrow{L^{{\mathbb A}^1}} {\mathcal S}{\mathcal H}^s_{S^1}(S)$
be the ${\mathbb A}^1$-fibrant replacement functor. Then
\begin{theorem}\label{connectivity}
Let $S={\rm Spec \,} (R)$ be the spectrum of a noetherian domain of dimension $d$ with all its residue fields infinite. Then S has the shifted stable ${\mathbb A}^1$-connectivity property, that is, if $E \in {\mathcal S}{\mathcal H}^{s}_{S^1 \geq i}(S)$ then $L^{{\mathbb A}^1} E \in {\mathcal S}{\mathcal H}^{s}_{S^1 \geq i-d}(S).$
\end{theorem}
The proof of Theorem \ref{connectivity} is exactly the same as the proof of its analogue in \cite{schmidt2018} except for the input from Gabber's presentation lemma, the required generality of which is available once Theorem \ref{main} is proved. We present a sketch of the proof of Theorem \ref{connectivity} in Section 4.\\
An important ingredient of the proof of the Gabber's presentation lemma of \cite{schmidt2018} is \cite[Theorem 4.1]{kai2015}, which states that given an equi-dimensional scheme $Y$ over a Dedekind scheme $B$ with infinite residue fields, Nisnevich locally on $B$ there exists a projective closure $\overline{Y}$ of $Y$ in which $Y$ is fiber-wise dense. Unfortunately, we are unable to prove such a result over a general base. However, we observe that a slightly weaker result (see Theorem \ref{ydense}) can be proved which suffices for our purpose. As in Gabber's original proof of the presentation lemma, as well as in \cite{schmidt2018}, the condition of residue fields being infinite in Theorem \ref{main} is required in order to make suitable generic choices. We are currently working on removing the condition of residue fields being infinite taking inputs from \cite{hogadi2018}.
\noindent{\bf Acknowledgments.} The first-named author was supported by the INSPIRE fellowship of the Department of Science and Technology, Govt.\ of India during the course of this work. The last-named author was supported by NBHM fellowship of Department of Atomic Energy, Govt.\ of India during the course of this work. We thank F. Strunk and J. Koll\'ar for their comments on the draft of this paper. We thank the anonymous referees for their helpful comments and suggestions.
\section{Fiber-wise denseness}\label{section_dense}
In this section, we prove a technical result which is crucial to the proof of our main theorem. It is essentially \cite[Theorem 4.1]{kai2015} with minor modifications (see also \cite[Theorem 10.2.2]{levine2006} ). Throughout this section, $\dim_B(Y)$ denotes the supremum of dimensions of all the fibers of $Y\rightarrow B$.
\begin{theorem}\label{ydense}
Let $B$ be the spectrum of a noetherian domain. Let $Y/B$ be either a smooth scheme or a divisor in a smooth scheme $X$. Let $y\in Y$ be a point lying over a point $b\in B$ with $\dim_B(Y_b) = n$. Assume $k(b)$ is an infinite field. Then there exist Nisnevich neighborhoods $(Y',y)\rightarrow(Y,y)$ and $(B',b)\rightarrow(B,b)$, fitting into the following commutative diagram
\begin{center}
\begin{tikzcd}
& Y' \arrow[d] \arrow[r]&Y\arrow{d}\\
& B'\arrow[r]&{B}
\end{tikzcd}
\end{center}
and a closed immersion $Y'\rightarrow {\mathbb A}_{B'}^N$ for some $N \geq 0$ such that if $\overline{Y'}$ is its closure in $\P_{B'}^N$ then $Y'_y$ is dense in the union of $n$-dimensional irreducible components of $(\overline{Y'})_y$.
\end{theorem}
\begin{remark}\label{Kaiapp}
The above theorem is a weaker statement than \cite[Theorem 4.1]{kai2015} (see also \cite[Theorem 10.2.2]{levine2006}) but over a general base. In the proof of \cite[Theorem 4.1]{kai2015} the author mentions that the base is assumed to be Dedekind to ensure that the projective closure of an equi-dimensional scheme remains equi-dimensional over $B$.
\end{remark}
We begin with an intermediary lemma which will be used repeatedly (see also \cite[Lemma 10.1.4]{levine2006}).
\begin{lemma}\label{1-d}
Let $X$ be an affine scheme. Choose a closed embedding $X\rightarrow{\mathbb A}^N_B$ and a point $x\in X$. Let $\overline{X}$ be the projective closure of $X$ in $\P^N_B$, and assume that it has fiber dimension $n$. Then, there exists
\begin{enumerate}
\item a projective scheme $\tilde{X}$,
\item an open neighbourhood $X_0$ of $x$ (in $X$),
\item an open immersion $X_0\hookrightarrow \tilde{X}$ and
\item a projective morphism $\psi: \tilde{X}\rightarrow \P^{n-1}_B$
\end{enumerate}
such that $\psi$ has fiber dimension one.
\end{lemma}
\begin{proof}
We follow the arguments given in \cite[Theorem 4.1]{kai2015} verbatim (see also \cite[Theorem 10.1.4]{levine2006}).
After possibly shrinking $B$, we can find $n$ hyperplanes $\Psi=\lbrace\psi_1,\ldots,\psi_n\rbrace$ which are part of a basis of $\Gamma(\P^N_B,{\mathcal O}(1))$ as a $B$-module. The choice is such that $V(\Psi)$, which denotes the common zeros of all $\psi_i$, does not contain $x$ and it meets $X$ fiber-wise properly over $B$, so that $\overline{X}\cap V(\Psi)$ is finite over $B$. Let $p:\tilde{\P^N}\rightarrow \P^N$ be the blowup of $\P^N$ along $V(\Psi)$, and $\tilde{X}$ the strict transform of $\overline{X}$ in the blowup. Let $\psi: \tilde{\P^N_B}\rightarrow \P^{n-1}_B$ denote the map induced by the rational map defined by a projection from $V(\Psi)$. Let $X_0:=\overline{X}\setminus V(\Psi)$. We have the following commutative diagram:
\begin{center}
\begin{tikzcd}[row sep=tiny]
& \tilde{X} \arrow[dd] \arrow[r,hook,"cl."]&\tilde{\P^N_B}\arrow{dd}\arrow[r,"\psi"]&\P^{n-1}_B\\
X_0 \arrow[ur,hook] \arrow[dr,hook] & \\
& \overline{X}\arrow[r,hook,"cl."]&{\P^N_B}
\end{tikzcd}
\end{center}
We claim that $\psi:\tilde{X}\rightarrow\P^{n-1}_B$ has fiber dimension one.
To see this, choose any point $y\in\P^{n-1}_B$, and consider the composite $a:{\rm Spec \,}(\Omega)\overset{y}{\rightarrow}\P^{n-1}_B\rightarrow B$. Then, the fiber of $\psi$ over $y$ may be identified with a linear subscheme $V(y)$ of $\P^N_a$, of dimension $N-n+1$. Furthermore, $V(y)$ contains the base change $V(\Psi)_a$, which has dimension $N-n$, by construction. Again by construction, the intersection $V(y)\cap \overline{X}\cap V(\Psi)_a$ is finite in $\P^N_a$. This means that $V(y)\cap \overline{X}$ has dimension $1$ in the projective space $V(y)$.
Further note that for $x\in V(\Psi)$, $p^{-1}(x)\simeq\P^{n-1}$. Also, the exceptional divisor of $\tilde{X}$ is an irreducible subscheme. Therefore, for any point $x\in V(\Psi)\cap X$, the fiber $\tilde{X}_x$ is an irreducible subscheme of $\P^{n-1}$ of dimension $n-1$. Therefore, $p^{-1}(\overline{X})=\tilde{X}$, so that $p:\psi^{-1}(y)\cap \tilde{X}\rightarrow V(y)\cap \overline{X}$ is a bijection. Thus, $\psi: \tilde{X}\rightarrow \P^{n-1}_B$ has 1-dimensional fibers.
\end{proof}
\begin{proof}[Proof of \ref{ydense}] We first prove the result in the case when $Y=X$ is a smooth scheme. The proof is by induction on $n$. The case $n=0$ follows from a version of Hensel's lemma.\\
\noindent \underline{Step 1}: As $X$ is smooth, Zariski locally on $B$, we write $X$ as a hypersurface in some ${\mathbb A}^N_B$. Let $\overline{X}$ denote its reduced closure in $\P^N_B$. Note that $\overline{X}$ also has fiber-dimension $n$ over $B$. By applying Lemma \ref{1-d}, we get a projective morphism $\psi: \tilde{X}\rightarrow \P^{n-1}_B$ with 1-dimensional fibers.
\noindent\underline{Step 2}: Set $T= \P^{n-1}_B$ and $t=\psi(x)$. Choose any projective embedding $\tilde{X} \hookrightarrow \P^{N_2}_T$. Let $(\tilde{X})_t$ and $(X_0)_t$ denote the fibers over $t$ of $\tilde{X}$ and $X_0$ respectively. Then choose a hypersurface $H_t \subset \P^{N_2}_t$ satisfying the next three conditions.
\begin{enumerate}
\item $x\in H_t$ (if $x$ is a closed point in $(X_0)_t$)
\item $(\tilde{X})_t$ and $H_t$ meet properly in $\P_t^{N_2}$.
\item $H_t$ does not meet $\overline{({X_0})_t}\setminus{({X_0})_t}$.
\end{enumerate}
Now after restricting to a suitable Nisnevich neighbourhood of $T$, which we denote again by $T$ (and after base changing everything to $T$), using the hyperplane $H_t$, we can choose a Cartier divisor ${\mathcal D}$ which fits into the following diagram
\begin{center}
\begin{tikzcd}[row sep=tiny,column sep=large]
& \tilde{X} \arrow{r}{projective}[swap]{1-dim}&T\arrow[r,"Nis"]&\P^{n-1}_B\\
X_0 \arrow[ur,hook] & \\
& {\mathcal D}\arrow[uu]\arrow[ul,hook,"{Cartier. div}"]\arrow{uur} [swap]{finite}
\end{tikzcd}
\end{center}
For sufficiently large $m$ we can find a section $s_0$ of $\Gamma(\tilde{X},{\mathcal O}_{\tilde{X}}(m{\mathcal D}))$ which maps to a nowhere vanishing section of $\Gamma({\mathcal D},{\mathcal O}_{{\mathcal D}})$. Let $s_1: {\mathcal O}_{\tilde{X}} \rightarrow {\mathcal O}_{\tilde{X}}(m{\mathcal D})$ be the canonical inclusion. Since the zero-loci of $s_0$ and $s_1$ are disjoint, we get a map
$$f=(s_0,s_1):\tilde{X} \rightarrow \P^1_{T}.$$
Since the quasi-finite locus of a morphism is open, shrink $T$ around $t$ such that ${\mathcal D}$ is contained in the quasi finite locus of $f$ after the base change. Let $X_0'$ be the quasi-finite locus of the base change.
\begin{center}
\begin{tikzcd}[column sep=large]
& f^{-1}(\infty_T)={\mathcal D} \arrow[d] \arrow[r,hook]&X_0'\arrow{d}[swap]{{quasi-finite}}\arrow[r,hook]&\tilde{X}\arrow[dl,"f"]\\
& \infty_T\arrow[r,hook]&{\P^1_T}
\end{tikzcd}
\end{center}
Then the subset $W=f(\tilde{X}\setminus X_0')\subset \P^1_T$ is proper over $T$ and is contained in $\P^1_T\setminus f(H_t)={\mathbb A}^1_T$. Hence, it is finite over $T$. The map $\tilde{X}\setminus f^{-1}(W)\rightarrow \P^1_T\setminus W$, being proper and quasi-finite, is finite. By condition $(1)$, we see that $\tilde{X}\setminus f^{-1}(W)$ contains $x$.\\
\noindent\underline{Step 3}: Now by induction there exist Nisnevich neighborhoods $B_1\rightarrow B$ and $T_1 \rightarrow T$ such that the projective compactification $T_1\rightarrow \overline{T_1}$ is fiber-wise dense in the union of $n$-dimensional irreducible components over $B_1$. Take a factorization of $f$ of the form $\tilde{X} \hookrightarrow \P^{N_3}_{T_1}\times_{T_1}\P^1_{T_1} \rightarrow \P^1_{T_1}$. Let $\overline{X_1}$ denote the reduced closure of $\tilde{X}$ in $\P^{N_3}_{\overline{T_1}}\times_{\overline{T_1}}\P^1_{\overline{T_1}}$. We get the following diagram where every square is Cartesian
\begin{center}
\begin{tikzcd}[row sep=large]
X_2:=\tilde{X}\setminus f^{-1}(W) \arrow[d,hook] \arrow[r,hook]&\tilde{X}\arrow[d,hook]\arrow[r,hook]&\overline{X_1}\arrow[d,hook]\\
\P^{N_3}_{T_1}\times_{T_1}(\P^1_{T_1}\setminus W) \arrow[r,hook]\arrow{d}&\P^{N_3}_{T_1}\times_{T_1}\P^1_{T_1} \arrow[r,hook]\arrow{d}&\P^{N_3}_{\overline{T_1}}\times_{\overline{T_1}}\P^1_{\overline{T_1}}\arrow{d}\\
\P^1_{{T_1}}\setminus W \arrow[r,hook] &\P^1_{{T_1}}\arrow[r,hook]&\P^1_{\overline{T_1}}
\end{tikzcd}
\end{center}
By Stein factorization we decompose the map $\overline{f_1}:\overline{X_1}\rightarrow \P^1_{\overline{T_1}}$ as
$$ \overline{f_1}:\overline{X_1}\rightarrow \overline{X_2}\xrightarrow{finite} \P^1_{\overline{T_1}},$$ where the first map has geometrically connected fibers. Since $\overline{f_1}$ is finite over the open set $\P^1_{T_1}\setminus W$, $\overline{X_2}\times_{\P^1_{\overline{T_1}}}(\P^1_{T_1}\setminus W)$ is isomorphic to $X_2 :=\tilde{X}\setminus f^{-1}(W) $. Since ${X_2}$ is open in $\overline{X_2}$, the fiber dimension of $\overline{X_2}$ is at least $n$. Combining this with the fact that $\overline{X_2}$ is finite over $\P^1_{\overline{T_1}}$, we conclude that the fiber dimension of $\overline{X_2}$ over $B_{1}$ is exactly $n$.
We observe that since $T_{1}$ is fiberwise dense in the union of $n$-dimensional irreducible components of $\overline{T_{1}}$, so is $\P^1_{{T_1}}$ (in $\P^1_{\overline{T_1}}$). Also as $W$ is finite over $T_{1}$, $\P^1_{T_1}\setminus W$ is fiberwise dense in $\P^1_{T_1}$. Hence it is dense in the union of $n$-dimensional irreducible components of $\P^1_{\overline{T_1}}$.
Now we claim that $X_2 $ intersects the fiber of $\overline{X_2}$ over any point $b_{1}$ of $B_{1}$. Let $X_{2}'$ be an $n$-dimensional irreducible component of the fiber $(\overline{X_2})_{{b}_{1}}$. Then the induced map $X_{2}' \rightarrow( \P^1_{\overline{T_1}})_{{b}_{1}}$ is a finite morphism of schemes of the same dimension. Hence it is a surjection to an irreducible component say, $U$ of $( \P^1_{\overline{T_1}})_{{b}_{1}}$. Further $\P^1_{T_1}\setminus W$ intersects $U$ by denseness. Taking inverse image of its intersection with irreducible component proves that $X_{2}$ intersects $Y$.
As $\overline{X_2}$ is projective over $B_{1}$, we choose any embedding of it in projective space $\P^{N}_{{B}_{1}}$. Then for the closed subscheme $\overline{X_2} \setminus X_2$ (with reduced structure) there exists a hypersurface $H$ of $\P^{N}_{{B}_{1}}$ of degree, say $d$, containing $\overline{X_2} \setminus X_2$, not containing the point $x$ and such that $H_{{b}_{1}}$ intersects $({X_2})_{{b}_{1}}$ properly in $\P^{N}_{{b}_{1}}$. Hence by discussion in previous paragraph, $H_{{b}_{1}}$ also intersects $(\overline{X_2})_{{b}_{1}}$ properly. Replacing $X_{2}$ by $\overline{X_2} \setminus H$ and taking $d$ fold Veronese embedding we may assume $H$ to be $\P^{N-1}_{\infty}$. Now we have the embedding $\overline{X_2} \setminus H \hookrightarrow {\mathbb A}^{N}_{{B}_{1}} = \P^{N}_{{B}_{1}} \setminus \P^{N-1}_{\infty} $ thereby proving the smooth case.
We shall now consider the case when $Y$ is a divisor in a smooth scheme.
\noindent\underline{Step 4}: Let $Y$ be a divisor in a smooth scheme $X$. We will produce a map, $\psi:\tilde{Y}\rightarrow \P^{d-1}$ whose fibers are $1$-dimensional.
Since $X$ is smooth, by Steps 1-3, Nisnevich locally, we have a closed embedding $Y\rightarrow X\rightarrow{\mathbb A}^N_B$ such that all fibers of $\overline{Y}\rightarrow B$ are $n$-dimensional. Then by Lemma \ref{1-d}, we have a commutative diagram,
\begin{center}
\begin{tikzcd}[row sep=tiny]
& \tilde{Y} \arrow[dd] \arrow[r,hook,"cl."]&\tilde{\P^N_B}\arrow{dd}\arrow[r,"\psi"]&\P^{n-1}_B\\
Y_0 \arrow[ur,hook] \arrow[dr,hook] & \\
& \overline{Y}\arrow[r,hook,"cl."]&{\P^N_B}
\end{tikzcd}
\end{center}
Then as in Step 2 above, we obtain a morphism Nisnevich locally on $Y$, $\phi : Y \rightarrow \P^{1}_{T}$, where $T$ is a Nisnevich neighbourhood of $\P^{n-1}$.
Since $T$ is a smooth $B$-scheme, our theorem holds for $T$.
Remaining proof is the same as in Step 3.
\end{proof}
\section{Relative version of Gabber's Presentation Lemma }
We now prove Theorem \ref{main}. We follow \cite{schmidt2018} to prove Theorem \ref{main}, the only difference being, that in their version of Theorem \ref{ydense} (which is for Henselian DVR), they have the stronger condition of fiberwise denseness, which they use to construct a finite map $\Psi_{|Z}: Z\rightarrow {\mathbb A}_{S}^{d-1}$. However, we observe that their proof still goes through with our weaker condition of denseness in $n$-dimensional components, which we illustrate in Propositions \ref{misses} and \ref{finite}. The rest of the proof does not require any new inputs and we just state those results from \cite{schmidt2018} which are essentially an application of the proof from \cite{colliot}.\\
First we reduce to the case that $z$ is a closed point and $Z$ is a principal divisor.
\begin{lemma}\label{reduction}(See \cite[Lemma 3.2.1]{colliot})
With the notation as in Theorem \ref{main}, there exists a closed point $z'\in X$ such that $z'$ is a specialization of $z$ and there exists a non-zero $f\in \Gamma(X,{\mathcal O}_X)$ such that $Z\subset V(f)$.
\end{lemma}
\begin{remark}
Since in Theorem \ref{main}, we assume that $dim (Z_s) < dim (X_s)$, in the Lemma \ref{reduction} we furthermore assume $f$ is such that $dim((V(f)_{s})< dim (X_{s})$.
\end{remark}
\begin{remark}
Since Theorem \ref{main} is a Nisnevich local statement, henceforth we assume that the ring $R$ is Henselian local with closed point $\sigma$ and infinite residue field $k$.
\end{remark}
Let $S= Spec(R)$ with ${\mathbb A}^{n}_S = R[x_1,\ldots,x_n]$. Let $E$ be the $R$-span of $\{x_1,\ldots,x_n\}$ and consider ${\mathcal E}:=\underline{{\rm Spec \,}}(Sym^{\bullet}E^{\vee})$ (note that ${\mathcal E}(R)=E$). For any integer $d > 0$ and any $R$-algebra $A$, ${\mathcal E}^d(A)$ parametrizes all linear morphisms $v= (v_1,\ldots,v_{d}):{\mathbb A}^{n}_T\rightarrow {\mathbb A}^{d}_T$, where $T= Spec(A)$. Considering ${\mathbb A}^n_S \hookrightarrow \P^{n}_S = $ Proj $S[X_0,\ldots,X_n]$, as a distinguished open subscheme $D(X_{0})$, we extend such a linear morphism to a rational map $ \overline{v}:\P^{n}_S \dashrightarrow \P^{d}_S$ whose locus of indeterminacy $L_{v}$ is given by the vanishing locus of $v_1,\ldots,v_{d}$ and $X_0$, $V_{+}(X_0,v_1,\ldots,v_{d}) \subseteq \P^{n}_S$ (We will use this notation throughout this section). Given any closed subscheme $Y$ in ${\mathbb A}^n_S$, we denote by $\overline{Y}$ its projective closure in $ \P^{n}_S $. For the following lemma we refer to \cite[Lemma 2.3]{schmidt2018}
\begin{lemma}(see \cite[Lemma 2.3]{schmidt2018})\label{shafar}
In the setting of the previous paragraph if $L_{v} \cap \overline{Y} = \emptyset$, then $\overline{v}:\overline{Y} \rightarrow \P^{d}_S$ and $v: Y \rightarrow {\mathbb A}^d_S$ are finite maps.
\end{lemma}
Following lemma is standard.
\begin{lemma}\label{dimdrop}
Let $W$ be a closed subscheme of $\P^N_k$. Then there exists a hyperplane $H \subset \P^N_k$ such that $\dim_k(H\cap W)=\dim_k(W)-1$.
\end{lemma}
\begin{proof}
Let $\zeta_1,\ldots\zeta_r$ be the generic points of $W$ corresponding to the homogeneous prime ideals $\wp_{1},\ldots \wp_{r} $. Viewing the $\wp_{i}$'s and $\varGamma({\mathcal O}(1),\P^N_k )$ as vector spaces over the infinite field $k$, we can find a hyperplane $H$ not containing $\zeta_i$'s: as no non trivial vector space over an infinite field can be written as a finite union of proper subspaces. Hence by Krull's principal ideal theorem $\dim_k(H\cap W)=\dim_k(W)-1$.
\end{proof}
\begin{proposition}\label{misses}
Let $Y $ be as in Theorem \ref{ydense} and $\overline{Y}$ be its projective closure, then there exist $v_1,\ldots,v_{n}$ in the $k$-span of $\{X_1,\ldots,X_N\}$ such that $(\overline{Y})_{\sigma}\cap L_v= \emptyset$, where $L_v= V_{+}(X_0,v_1,\ldots,v_{n})$.
\end{proposition}
\begin{proof}
Without loss of generality, we assume ${\mathbb A}^{n}_k = D(X_{0})$. Let $H_{\infty} = V_{+}(X_{0})$ denote the hyperplane at infinity of $\P^{N}_{k}$. Generic points of irreducible components of $\overline{Y_{\sigma}}$ lie in ${\mathbb A}^{n}_k = D(X_{0})$. Therefore $\dim(\overline{Y_{\sigma}} \cap H_{\infty} ) = n-1$. By Theorem \ref{ydense}, we have $\dim((\overline{Y})_{\sigma} \cap H_{\infty} ) = n-1$. Now applying Lemma \ref{dimdrop} repeatedly proves the claim.
\end{proof}
\begin{theorem}\label{fer}
Let $X=Spec(A)/S$ be a smooth, equi-dimensional, affine, irreducible scheme of relative dimension $d$. Let $Z=Spec(A/f)$, $z$ be a closed point in $Z$ lying over $s\in S$, where $f$ is such that $\dim (Z_s) < \dim (X_s)$. Then there exists an open subset $\Omega\subset {\mathcal E}^d$ with $\Omega(k)\neq \emptyset$ such that for all $\Phi=(\Psi,\nu)\in \Omega(k)$ the following hold
\begin{enumerate}
\item $\Psi_{|Z}:Z\rightarrow {\mathbb A}^{d-1}_S$ is finite.
\item $\Psi$ is \'{e}tale{} at all points of $F:=\psi^{-1}(\psi(z))\cap Z$.
\item $\Phi_{|F}:F\rightarrow \Phi(F)$ is radicial.
\end{enumerate}
\end{theorem}
Recall that $\Phi:F\rightarrow \Phi(F)$ is said to be radicial \cite[Tag 01S2]{stacksradicial} if $\Phi$ is injective and for all $x \in F$ the residue field extension $k(x)/k(\Phi(x))$ is trivial.\\
To prove this theorem, we first get an open set of finite maps in Proposition \ref{finite}. Then we get a non-empty open set of \'{e}tale{} and radicial maps in Lemma \ref{er}.
\begin{remark}\label{rem}
By \cite[Prop. 2.6 and Lemma 2.7]{schmidt2018} we have a closed embedding $X \hookrightarrow {\mathbb A}^{N}_S$ such that $Z$ (Nisnevich locally around $z$) satisfies Theorem \ref{ydense}.
\end{remark}
\begin{proposition}\label{finite}
Let $X$ and $Z$ be as in Theorem $\ref{fer}$ with $S$ a spectrum of a Henselian local ring $R$. Then there is an open subset $\Omega\subset {\mathcal E}^{d}$ with $\Omega(R)\neq \emptyset$ such that for all $\Psi \in \Omega(R)$, $\Psi_{|Z}:Z\rightarrow {\mathbb A}^{d-1}_S$ is finite.
\end{proposition}
\begin{proof}
We proceed as in \cite[Lemma 2.11]{schmidt2018}. By Remark \ref{rem} we have closed embedding $X \hookrightarrow {\mathbb A}^{N}_S$. Viewing ${\mathcal E}^{d-1}$ as a closed subscheme of ${\mathcal E}^{d}$ by taking the first $d-1$ factors we consider the closed subscheme
\begin{center}
$V = {\mathcal E}^{d-1} \times_{S} H_{\infty} \hookrightarrow {\mathcal E}^{d} \times_{S} H_{\infty}$
\end{center}
where $H_{\infty}$ is the hyperplane at infinity in $\P^{N}_{S}$. Note that $V \rightarrow {\mathcal E}^{d}$ has fiber $V_{v}= L_{(v_1,\ldots,v_{d-1})}$ for any $v= (v_1,\ldots,v_{d}) \in {\mathcal E}^{d}(R)$. Consider the open subscheme $\Omega$ of ${\mathcal E}^{d}$ defined as
$${\mathcal E}^{d} \setminus p_{1}(V \cap ({\mathcal E}^{d} \times _{S}(\overline{Z} \cap H_{\infty})) ), $$
\noindent where $p_{1}$ is projection of ${\mathcal E}^{d-1} \times_{S} H_{\infty}$ onto the first factor. By construction every point in $\Omega(R)$ consists of a linear map $v=(v_1,\ldots,v_{d}) : {\mathbb A}^N_{S} \rightarrow {\mathbb A}^d_{S} $ such that $L_{v'} \cap \overline{Z} = \emptyset $, where $v'= (v_1,\ldots,v_{d-1})$. By Lemma \ref{shafar}, this will be our required finite map, thus proving $\Omega(R) \neq \emptyset$ will finish the proposition. As $R$ is Henselian local, the induced map from $\Omega(R)$ to $\Omega(k)$ is surjective, hence it suffices to prove $\Omega(k) = \Omega_{\sigma}(k) \neq \emptyset $. By construction we have, $\Omega_{\sigma}(k) = {\mathcal E}^{d}_{\sigma} \setminus p_{1}(V_{\sigma} \cap ({\mathcal E}^{d}_{\sigma} \times _{S}((\overline{Z} )_{\sigma}\cap H_{\infty})) )$ and any point in $\Omega(k)$ gives a linear map $u=(u_1,\ldots,u_{d}) : {\mathbb A}^N_{k} \rightarrow {\mathbb A}^d_{k} $ such that $L_{u'} \cap (\overline{Z})_\sigma = \emptyset $, where $u'= (u_1,\ldots,u_{d-1})$. By Lemma \ref{misses} such a map exists.
\end{proof}
\begin{proposition}\label{er} Let $\phi=(\psi,\nu)=(u_1,\ldots,u_d): X\rightarrow {\mathbb A}^{d-1}_S\times {\mathbb A}^1_S$ and $F:=\psi^{-1}(\psi(z))\cap Z$.
There exists an open set $\Omega_2\subset {\mathcal E}^d$ such that $\Omega_2(R)\neq \emptyset$ and for any $\phi\in\Omega_2(R)$
\begin{enumerate}
\item $\phi$ is \'{e}tale{} at all points of $F$.
\item $\phi_{|F}:F\rightarrow \phi(F)$ is radicial.
\end{enumerate}
\end{proposition}
\begin{proof} See \cite[Lemma 2.12]{schmidt2018}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{fer}.]
Let $\Omega_1$ and $\Omega_2$ be as in the Propositions \ref{finite} and \ref{er}. Then the set $\Omega=(\Omega_1\times {\mathcal E})\cap\Omega_2$ satisfies all the required conditions.
\end{proof}
Now we obtain the sets $U$ and $V$. The sets $U$ and $V$ are constructed to satisfy all the conditions of Theorem \ref{main}.
\begin{lemma}\label{setv}
Let $\Phi=(\Psi,\nu)$ satisfy conditions (1)-(3) of Theorem \ref{fer}. Then there exists an open neighborhood $V\subset {\mathbb A}^{d-1}_S$ of $\Psi(z)$ such that
\begin{enumerate}
\item $\Phi$ is \'{e}tale{} at all points of $Z\cap \Psi^{-1}(V)$.
\item $\Phi|_{Z\cap \Psi^{-1}(V)}:Z\cap \Psi^{-1}(V)\rightarrow {\mathbb A}^1_V$ is a closed immersion.
\end{enumerate}
\end{lemma}
\begin{proof} See \cite[Lemma 2.13]{schmidt2018}.
\end{proof}
\begin{lemma}\label{setu}
There exists a closed subset $\mathfrak{U}\subset \Psi^{-1}(V)$ such that
\begin{enumerate}
\item $U_1=\Psi^{-1}(V)\setminus\mathfrak{U}$ contains $z$
\item $U_1$ satisfies $Z\cap \Psi^{-1}(V)=Z\cap U_1 $
and $\Phi^{-1}(\Phi(Z\cap U_1)) \cap U_1=Z\cap U_1$.
\end{enumerate}
\end{lemma}
\begin{proof} See \cite[Lemma 2.14]{schmidt2018}
\end{proof}
\begin{proof}[Proof of Theorem \ref{main}]
Let $U_2$ be the open locus where $\Phi$ is \'{e}tale{}. From Lemma \ref{setv} $z\in U_2$ and $Z\cap \Psi^{-1}(V)\subset U_2$. Now let $U=U_1\cap U_2$, with $U_1$ as in Lemma \ref{setu}. Then $U$ also satisfies conditions $(2)$ and $(3)$ of Lemma \ref{setu}. Furthermore $\Psi_{U}$ is \'{e}tale{}. Hence we get $\Phi,\Psi,U,V$ satisfying all the conditions of Theorem \ref{main}.
\end{proof}
\section{Stable Connectivity} In this section we give a sketch of the proof of Theorem \ref{connectivity}. We do not claim any originality here and all the proofs of the statements in this section can be found in \cite[\S 4]{schmidt2018}. Throughout this section $Sm_{S}$ will denote the category of smooth schemes over a given scheme $S$.
\begin{lemma} \label{zero}
Let $Spec(R)=S$ be a noetherian scheme of finite Krull dimension with a codimension $d$ point $s \in S$. Let $E \in {\mathcal S}{\mathcal H}^{s}_{S^1 \geq d+1}(S)$. Then for any $X \in Sm_{S}$ with $X_{s} \neq \emptyset $ and any $f \in [ \Sigma^{\infty}_{S^{1}} X_{+} , L^{{\mathbb A}^1} E ]$ in $ {\mathcal S}{\mathcal H}^{s}_{S^1}(S)$, there exists an open subscheme $U$ in $X$ such that $U$ intersects each irreducible component of $X_{s} $ non trivially and $f|_{\Sigma^{\infty}_{S^{1}} U_{+} } = 0$.
\end{lemma}
\begin{proof}
Let $Z_{i}$'s be the irreducible components of $X_{s}$ . From the proof of \cite[Lemma 4.9 ]{schmidt2018} we obtain open subschemes $U_{i}$'s of $X$ such that $U_{i} \cap Z_{i} \neq \emptyset$ and $f|_{\Sigma^{\infty}_{S^{1}} (U_{i })_+} = 0$. Define $U$ to be union of all such $U_{i}$'s. Since the left Quillen functor $\Sigma^{\infty}_{S^{1}}$ gives an adjunction at the level of homotopy category and $L^{{\mathbb A}^1} E$ (apart from being a fibrant object in $ {\mathcal S}{\mathcal H}^{s}_{S^1}(S)$ ) is a spectrum of Nisnevich sheaves, we have $f|_{\Sigma^{\infty}_{S^{1}} U_{+} } = 0$.
\end{proof}
\begin{lemma}\label{dist}
Let $X$ be a smooth irreducible scheme over $S$ and $U$ be a non-empty open subscheme of $X$. Denote by $Z$ the reduced closed subscheme $ X \setminus U$. Suppose Nisnevich locally on $X$ we have the following Nisnevich distinguished square
\begin{center}
\begin{tikzcd}
& U \arrow[d] \arrow[r]&X\arrow[d,"p"]\\
& {\mathbb A}^{1}_V \setminus p(Z) \arrow[r]&{{\mathbb A}^{1}_V}
\end{tikzcd}
\end{center}
where the map $p : X \rightarrow {\mathbb A}^{1}_V$ in $Sm_{S}$ is \'{e}tale{}, with $Z \rightarrow V$ finite. Then $\pi^{{\mathbb A}^{1}}_{0}(X/U)=0$.
\end{lemma}
\begin{proof}
See \cite[Lemma 4.6 and Cor. 4.7]{schmidt2018}.
\end{proof}
Lemmas \ref{zero} and \ref{dist} together give a bound on the drop in connectivity, which is sufficient to prove the connectivity result Theorem \ref{connectivity}, for details see \cite[Prop. 4.5]{schmidt2018}.
\begin{remark}
Note that if $X, Z $ and $S$ in previous lemma satisfy the conditions stated in Theorem \ref{main}, we obtain the distinguished square of previous lemma and hence $\pi^{{\mathbb A}^{1}}_{0}(X/U)=0$.
\end{remark}
\begin{remark} \label{Hensel} To prove stable connectivity we can assume $S$ to be Henselian local by \cite[ Lemma 4.10]{schmidt2018}
\end{remark}
\begin{proof}[Proof of Theorem \ref{connectivity}]
We proceed by induction on the dimension of $S$. The case, $dim(S)=0$ follows from \cite{morel2005}. By Remark \ref{Hensel}, we may assume $S$ to be Henselian local with closed point $\sigma$. Further we can assume $X_{\sigma} \neq \emptyset$, where $X \in Sm_{s}$. Consider $f \in [ \Sigma^{\infty}_{S^{1}} X_{+} , L^{{\mathbb A}^1} E ]$, then by Lemma \ref{zero} we obtain an open subscheme $U$ such that $U$ intersects each irreducible component of $X_{\sigma} $ non-trivially and $f|_{\Sigma^{\infty}_{S^{1}} U_{+} } = 0$. Take the reduced closed subscheme $Z = X \setminus U$. Then dim $Z_{\sigma} < $ dim $X_{\sigma} $. Hence by Gabber presentation lemma we have Nisnevich distinguished square of Lemma \ref{dist} which proves $\pi^{{\mathbb A}^{1}}_{0}(X/U)=0$. Now connectivity follows from \cite[Prop. 4.5]{schmidt2018}
\end{proof}
\bibliographystyle{alphanum}
|
\section{Abstract}
Nonlinear processes in integrated, guiding systems are fundamental for both classical and quantum experiments. Integrated components allow for compact, modular and stable light-processing systems and as such their use in real-world systems continues to expand.
In order to use these devices in the most demanding applications, where efficiency and/or spectral performance are critical, it is important that the devices are fully optimized. In order to achieve these optimizations, it is first necessary to gain a thorough understanding of current fabrication limits and their impact on the devices final performance.
In this paper we investigate the current fabrication limits of titanium in-diffused lithium niobate waveguides produced using a masked photolithographic method.
By dicing a long ($\sim$8cm) sample into smaller pieces and recording the resulting phasematching spectra, the fabrication error present in the UV photolithographic process is characterized.
The retrieved imperfections fit well with theoretical expectations and from the measured imperfection profile it is shown that one can directly reconstruct the original distorted phasematching spectrum.
Therefore, our measurements directly quantify the intrinsic limitations of the current standard UV photolitographic technique for the realization of integrated waveguides in lithium niobate.
\section{Introduction}
Guiding systems fabricated in nonlinear materials are employed in a wide variety of contexts for both classical and quantum applications. In comparison to bulk systems, guiding systems provide simple integration into fiber networks and provide a tighter spatial mode confinement, which generally strengthens the nonlinear interaction. Such systems are already in use in a myriad of classical applications, for example in second harmonic and sum frequency generation (SHG/SFG) to efficiently produce light in spectral ranges otherwise inaccessible \cite{Lim1989, Eger1994, Sun2012}. More recently these systems are also finding applications in quantum systems, for example in generation and manipulation of quantum states \cite{Tanzilli2001,Tanzilli2005, Rutz2017, Maring2017, Stefszky2017}.
While waveguiding systems offer many advantages, these systems also generally exhibit higher losses and are much more sensitive to device imperfections, when compared to their bulk counterparts. It is known that small device imperfections can dramatically reduce device performance \cite{Helmfrid1991, Helmfrid1992, Pelc2010, Cui2012, Phillips2013, Chang2014, FrancisJones2016, Santandrea2019}.
For this reason, it is critical to assess the limits of the current fabrication technology to identify the classes of nonlinear processes physically achievable and to devise strategies to overcome such limitations.
Despite the importance of these investigations, very little experimental work has been undertaken \cite{Chang2014, FrancisJones2016}.
In \cite{Chang2014}, both the amplitude and the phase of a phasematched process in reverse proton exchanged lithium niobate (RPE-LN) waveguides was characterized, allowing the complete determination of the inhomogeneity profile of their waveguides. However, such a scheme is not always possible as it relies on a chirped broadband conversion process to map spatial inhomogeneities to the amplitude and phase of the second harmonic field, which can be characterized using frequency resolved optical gating (FROG).
In \cite{FrancisJones2016}, a destructive approach was used to reconstruct the variation of the fabrication parameters of a photonic crystal fibre (PCF).
In this method, the sample under analysis is diced into smaller sections and the phasematching of each section is used to infer the local properties of the system.
In this paper, the dicing technique is used to retrieve the fabrication errors of titanium indiffused lithium niobate (Ti:LN) waveguides.
An 83mm-long sample is diced down into $\sim$10mm-long pieces, whose individual phasematching, as well as the phasematching profile of intermediate lengths, are characterized.
The shift of the phasematching spectrum along the waveguide is mapped and used to retrieve the phasematching variation.
Via numerical modelling, we are able to relate the measured variation to fabrication parameter errors and from this show that the estimated errors agree with previous predictions \cite{Santandrea2019}.
Finally, from the measured waveguide inhomogeneities, we are able to reconstruct the original phasematching of the sample.
\section{Experiment}
The system under investigation is a set of seven 83mm-long nonlinear waveguides fabricated by in-diffusing Ti ions in a z-cut LiNbO$_3$ crystal.
The waveguides are designed to be single-mode in the telecom C band ($7\mu$m in width) and are periodically poled with a period $\Lambda$=16.8$\mu$m.
This allows a type 0, $ee\rightarrow e$ degenerate SHG process pumped at 1528.4nm at room temperature.
The nonlinear process is completely determined by the phase mismatch $\Delta\beta$ of the involved light fields
\begin{equation}
\Delta\beta(z, \lambda) = 2\pi\left(\frac{n_e(z, \lambda/2)}{\lambda/2} -2\frac{n_e(z, \lambda)}{\lambda} -\frac{1}{\Lambda} \right),\label{eq:deltabeta}
\end{equation}
where $n_e(z,\lambda)$ is the extraordinary refractive index of LiNbO$_3$ at the position $z$ and at the pump wavelength $\lambda$.
We explicitly consider the variation of the refractive index along the propagation axis $z$ to include the effect of fabrication imperfections, such as inhomogeneities in the waveguide width, depth, operating temperature, poling pattern or a combination thereof.
In the case of small refractive index variations, one can approximate variations in the momentum mismatch to be wavelength independent, i.e. $\Delta\beta(z, \lambda) \approx \Delta\beta_0(\lambda) + \delta\beta(z)$. In this case, the output intensity spectrum of the SH process is given by \cite{Helmfrid1992, Chang2014, Santandrea2019}
\begin{equation}
I(\lambda) \propto \left|\int_0^L \mathrm{e}^{-i\Delta\beta_0(\lambda) z}\mathrm{e}^{-i\int_0^z \delta\beta(\xi)\mathrm{d}\xi}\mathrm{d}z \right|^2 \label{eq:pm_integral}
\end{equation}
and thus depends on the specific $\delta\beta(z)$ profile.
However, if the waveguide is sufficiently short, the impact of the variations $\delta\beta(z)$ are not strong enough to appreciably distort the phasematching spectrum, as shown in \cite{Santandrea2019}. They are still expected, however, to affect positioning of the centre of the spectrum.
Previous work in this waveguide system has shown that we expect to see nearly ideal phasematching spectra for waveguides shorter than 1cm in length \cite{Santandrea2019}.
Therefore, it should be possible to retrieve the variation of $\delta\beta(z)$ along the sample by dicing a long sample down to $\sim$1cm-long pieces and monitoring the shift in the position (in wavelength) of the phasematching spectrum.
\begin{figure}[bt]
\centering
\includegraphics[width = 0.7\textwidth]{setup.pdf}
\caption{Measurement setup. Full description in text.}
\label{img:setup}
\end{figure}
At first, the phasematching spectra of the initial 83cm-long waveguides are measured in the setup illustrated in Figure \ref{img:setup}. The sample is first stabilised in temperature to around 25$\pm$0.1$^\circ$C as unwanted temperature shifts are indistinguishable from true waveguide imperfections, since both will act to shift the centre of the phasematching profile. The waveguides are then pumped with approximately 3mW from a tunable 1550nm laser (\textit{EXFO Tunics}) and the wavelength of the pump is scanned in 1pm steps. The generated SH field is then detected using a Si-PIN photodiode. To increase the signal-to-noise ratio of the measurement the pump field is run through a chopping blade, and the photocurrent from the photodiode is fed into a lock-in amplifier (\textit{Ametek Scientific Instrument 7265}).
As expected, the spectra of the long waveguides are distorted due to the presence of waveguide inhomogeneities, as can be seen in Figure \ref{img:ph_wg16}.
A $\sim$1cm-long piece was then cut from one end of the original sample and the resulting surfaces were polished. Similar to the measurement for the original long sample, both resulting lengths were then temperature stabilised and the phasematching spectra of the waveguides in each section were again recorded. This process was repeated until the full sample was finally cut down into 7 pieces of approximately 1cm long, as shown in Figure \ref{img:chopping}. Note that some of the length is lost in the dicing and polishing of the sample as well as a small piece that was damaged during dicing.
\begin{figure}[bt]
\centering
\includegraphics[width = 0.7\textwidth]{chopping_scheme.pdf}
\caption{Sequence of sample dicing. All the final sections are $\sim$10mm long.
While polishing the 70mm-long sample, roughly 8mm of sample snapped from the left side and thus no phasematching spectrum could be recorded for S2 and for the 70mm-long sample.}
\label{img:chopping}
\end{figure}
Theoretical calculations show that, for a given inhomogeneity profile, the phasematching spectrum should become less distorted as the length of the waveguides is reduced, as waveguide imperfections become less critical \cite{Santandrea2019}. The phasematching spectra measured from a single waveguide as it was gradually cut into shorter and shorter pieces is shown in Figure \ref{img:ph_wg16}. As expected, one can see that the measured phasematching spectra gradually approach the ideal sinc$^2$-like shape as the sample is shortened down to 1cm.
\begin{figure}[bt]
\centering
\includegraphics[width = 0.7\textwidth]{phasematching_spectra.pdf}
\caption{Type 0 SHG phasematching spectra of a single waveguide for different waveguide lengths. Note that, as expected, the phasematching spectrum broadens as the sample becomes shorter and its shape tends towards the ideal sinc$^2$ profile. The oscillations are due to the presence of Fabry-Perot oscillations from the uncoated end facets.}
\label{img:ph_wg16}
\end{figure}
From the measured phasematching profile of each 1cm-long piece a more quantitative investigation is undertaken. Figure \ref{img:ph_wg16_1cm} shows the measured phasematching profiles for all 1cm pieces across a single waveguide. For reference, the dashed curve shows the expected sinc$^2$ profile fitted only for the central wavelength. The central phasematching wavelength $\lambda_{pm}$ is found from the data as the average of the measured wavelengths weighted by the corresponding measured intensities.
Since the sample is temperature stabilized, the deviation of $\lambda_{pm}$ from the target phasematching wavelength $\bar{\lambda}_{pm}=$1528.4nm is a measure of the waveguide inhomogeneities. With the help of Eq. (\ref{eq:deltabeta}), one can quantify the phase mismatch $\delta\beta$ with respect to the ideal phasematching for each 1cm-long section.
This characterisation is repeated for all remaining waveguides and the central phasematching wavelengths found in the 1cm samples for all waveguides are shown in Figure \ref{img:phasematching_shift}.
It is immediately apparent that all the measured waveguides show a similar trend, whereby $\lambda_{pm}$ increases along the sample length.
Additionally, $\lambda_{pm}$ changes quite dramatically between the different sections, with an average maximum variation of 0.6nm.
This corresponds to a maximum $\delta\beta$ variation of $\sim$442 m$^{-1}$.
Using the model presented in \cite{Strake1988}, one can relate the phasematching shift to a variation of the waveguide properties.
For simplicity, we assume that only variations in the waveguide width contribute to phasematching variations along the sample, since this parameter is likely to be the major contributor to waveguide imperfections \cite{Santandrea2019}.
For the waveguides under consideration, a variation of $\delta\beta\sim$ 440m$^{-1}$ corresponds to a width inhomogeneity $\delta w\sim 0.25\mu$m, in close agreement with the theoretical estimation presented in \cite{Santandrea2019}.
Interestingly, the retrieved $\delta w$ is close to the width error estimated in \cite{Chang2014} ($\delta w \sim 0.2\mu$m) with a completely different measurement technique, for a very different process realized in RPE-LN waveguides. The only similarity between the two systems is the employment of a wet-etching based photolithographic step for waveguide patterning. This is perhaps an indication that the main cause of waveguide imperfections resides in the photolithographic step.
\begin{figure}[bt]
\centering
\includegraphics[width = 0.7\textwidth]{phasematching_spectra_1cm_pieces.pdf}
\caption{Measured phasematching spectra for the different 1cm-long pieces in a single waveguide (WG1). The dashed lines correspond to the theoretical spectra fitted only for central wavelength. The shift of the phasematching centre is due to local variation of the waveguide properties and is used to derive the phasematching variation along the sample. Section S2 is missing since it broke during the dicing stage.}
\label{img:ph_wg16_1cm}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics[width = 0.7\textwidth]{phasematching_shift2.pdf}
\caption{Measured variation in the phasematching wavelength $\lambda_{pm}$ across all waveguides for each 1cm-long sample. The black solid line is the average $\lambda_{pm}$ shift. The corresponding phase mismatch variation $\delta\beta$ and waveguide width $w$ can be read on the right axes. Here, we assume that only the width variation is responsible of the whole phase mismatch $\delta\beta$ variation. }
\label{img:phasematching_shift}
\end{figure}
Finally, it should be possible to reconstruct the original phasematching spectrum given the measured $\delta w$ profile along the sample. The performance of this reconstruction will of course be limited by the resolution of the measured $\delta w$ profile and hence only the profile of the full length sample is reconstructed.
We interpolate the measured data points with a piecewise linear interpolation to avoid numerical artifacts that are seen to arise from the choice of higher order polynomials fits.
Thus, we obtain an approximate waveguide width profile $w(z)$ and the respective phase mismatch variation $\delta\beta(z)$. Eq. (\ref{eq:pm_integral}) is then used to calculate the expected phasematching spectrum of the original, full-length sample.
In approximately half of cases, the reconstruction of the phasematching was successful and showed very similar behavior to the phasematching that was initially measured from the full-length sample. One such reconstruction and the measured phasematching spectrum is shown in Figure \ref{img:pm_recon}. It is believed that the reconstruction failed in half of cases due to an insufficient spatial resolution for $\lambda_{pm}$. Note that, in contrast to previous work in PCF fibres \cite{FrancisJones2016}, no fitting has been performed to reproduce the measurement of the original sample. This shows that, in most cases, the phasematching properties in Ti:LN waveguide can be considered constant within 1cm.
\begin{figure}[hbt]
\centering
\includegraphics[width = 0.8\textwidth]{waveguide14_sectionL1__interpol}
\caption{Reconstructed phasematching spectrum from the measured data points (solid blue line) and comparison with the measured one (dashed orange line). In the inset, the inferred waveguide widths and the linear interpolation used to reconstruct the phasematching are shown.}
\label{img:pm_recon}
\end{figure}
\section{Conclusion}
In this paper, the inhomogeneity present in Ti:LN waveguides produced via UV photolithograpy was characterised by investigating the performance of a type 0 SHG process in these waveguides. The phasematching spectrum of seven waveguides was measured for the full-length sample (83mm) and for each resulting smaller section as $\sim$10mm-long pieces were cut from the ends of the initial sample.
The variation in the central phasematching wavelength along the sample was recorded for the 1cm pieces. From these measurements the variation in the phasematching across the sample were inferred, showing that the maximum phasematching deviation is $\delta\beta \leq$ 440m$^{-1}$. Under the assumption that variations in the phasematching are primarily due to imperfections in the width of the waveguides $\delta w$, a maximum width error of $\delta w\leq 0.25\mu$m was found, consistent with previous theoretical predictions \cite{Santandrea2019}.
It was also shown that the measured phasematching variation along the sample could be used to directly reconstruct the measured, highly distorted phasematching profile of the full-length sample.
These results confirm the theoretical model presented in \cite{Santandrea2019}, thus verifying the validity of the conclusions presented therein.
As the retrieved waveguide width error $\delta w$ is very close to values reported from other groups for different types of samples, we can thus conclude that this error is most likely the current ultimate resolution for waveguide fabrication using state-of-the-art UV masked photolithography.
This provides crucial information that can be fed into the design of new samples. Knowing the fabrication limitations allows one to calculate the expected phasematching degradation for any desired process, particularly for high demand applications. Furthermore, understanding these limitations provides the foundation for new developments towards improved sample design and fabrication technologies.
\section{Acknowledgment}
The work was supported by the European Union via the EU quantum flagship project UNIQORN (Grant No. 820474) and by the DFG (Deutsche Forschungsgemeinschaft).
|
\section{Introduction}
Binary stars are ubiquitous: roughly half of all field stars have binary companions \citep{Duquennoy_1991, Duchene_2013, Moe_2017}, and the binary fraction is even higher in star forming regions \citep[e.g.][]{Ghez_1993, Leinert_1993, Mathieu_1994, Connelley_2008, Sadavoy_2017, Duchene_2018}.
Significant progress has been made in developing theoretical models to explain the population statistics of observed binaries \citep[e.g.][]{Kroupa_1995, Marks_2011, Bate_2012, Lomax_2015}, but fundamental aspects of the binary formation process remain imperfectly understood.
The distribution of binary mass ratios has been a subject of interest for at least a century \citep[e.g.][]{Biesbroeck_1916, Opik_1924, Kuiper_1935}. As a final outcome of the binary formation process, the mass ratio distribution provides useful constraints on theoretical models of star formation. Unlike the distributions of orbital separation and eccentricity, the mass ratio distribution has been suggested to be insensitive to dynamical evolution after formation (such that binaries of different mass ratios are disrupted at similar rates; e.g. \citealt{Parker_2013}). Mapping the mass ratio distribution over a range of binary masses and separations has thus been the focus of many studies \citep[e.g.][]{Trimble_1974, Trimble_1987, Trimble_1990, Eggleton_1989, Mazeh_1992, Hogeveen_1992, Shatsky_2002, Burgasser_2007, Soderhjelm_2007, Raghavan_2010, Tokovinin_2014, Gullikson_2016}.
Observational studies of the mass ratio distribution are complicated by incompleteness. All binary detection methods are biased against low-mass ratio companions, which produce weaker radial-velocity shifts at fixed separation \citep[e.g.][]{Shahaf_2019}, contribute less light to the observed spectra of unresolved binaries \citep[e.g.][]{ElBadry_2018_mock, Elbadry_2018a}, cause weaker eclipses \citep[e.g.][]{Moe_2013}, and are less likely to be detected as visual companions \citep[e.g.][]{Tokovinin_2011, ElBadry_2018}. The detection efficiency also varies with primary mass and orbital separation. This complicates measurement of the mass ratio distribution, because the distributions of separation, primary mass, and mass ratio are not independent \citep[e.g.][]{Moe_2017}. It is thus important for demographic studies of binaries that the selection function of observed samples is well understood. If possible, binaries of different masses and physical separations should be considered independently.
A puzzling feature of the mass ratio distribution identified by previous works is the so-called ``twin'' phenomenon, which refers to a purported statistical excess of nearly equal-mass binaries with mass ratios $0.95 \lesssim q < 1$. Most studies that find an excess of equal-mass twins have focused on spectroscopic binaries with close separations \citep[$a \ll 1$\,AU, e.g. ][]{Lucy_1979, Hogeveen_1992, Tokovinin_2000, Halbwachs_2003, Lucy_2006, Pinsonneault_2006, Simon_2009, Kounkel_2019}. Indeed, several studies have reported a sharp drop-off in the fractional excess of twin binaries beyond periods of $40$\, days \citep[$a \lesssim 0.2$\,AU; e.g.][]{Lucy_1979, Tokovinin_2000, Simon_2009}. Recent studies have not confirmed such a sharp drop-off in the twin excess for solar-type stars \citep{Tokovinin_2014, Moe_2017}, but have still found it to decline monotonically with increasing separation. At least for $M_1 \gtrsim 1 M_{\odot}$ (the mass range on which most previous studies have focused), the twin excess has been found to decrease with increasing primary mass at fixed separation and to extend to wider separations for lower-mass primaries \citep{Moe_2017}.
Some works have also argued that the twin phenomenon may be a selection effect \citep{Mazeh_2003, Cantrell_2014}, since equal-mass binaries are brighter, can be detected at larger distances, and are preferentially selected by several binary detection methods \citep[e.g.][]{Branch_1976}. Such biases are minimized for binary samples that are nearly volume-complete \citep[e.g.][]{Raghavan_2010, Tokovinin_2014} and/or have well-understood completeness.
If it is a real effect, the physical origin of the twin phenomenon is not fully understood. Several mechanisms have been proposed that could lead to preferential formation of equal-mass binaries, including fragmentation during the late stages of protostellar collapse, mass transfer between pre-main sequence stars, and competitive accretion (see \citealt{Tokovinin_2000} for discussion of different formation mechanisms). Several simulations have predicted that accretion of high-angular momentum gas, particularly from a circumbinary disk, tends to drive binary mass ratios toward unity \citep[e.g.][]{Bate_2000, Farris_2014, Young_2015}.
However, it is not obvious why, when averaged over a large population of binaries, accretion from a circumbinary disk would produce a sharp peak in the mass ratio distribution at $q\gtrsim 0.95$ as opposed to a gradual increase.
There have also been hints of an excess of twins among spatially resolved wide binaries with separations ranging from tens to thousands of AU \citep{Trimble_1987, Soderhjelm_2007}. The selection functions of the wide binary samples studied in these works were poorly understood, causing investigators to remain agnostic of whether the excess of equal-brightness pairs in their catalogs was of astrophysical origin or rooted in selection biases. Recently, \citet{Moe_2017} measured the twin excess at different separations in a small but relatively complete sample of solar-type binaries within 25\,pc of the Sun. They found the twin excess to decline with separation, but found it inconsistent with 0 out to separations of 200\,AU. At even wider separations, they set an upper limit of $\sim$5\% on the excess twin fraction.
High-quality parallaxes and proper motions from the recent \textit{Gaia} data releases \citep{Gaia_2016, Gaia_2018} have simplified the process of constructing samples of wide binaries with (a) little contamination from chance alignments and (b) a well-understood selection function. Using data from {\it Gaia} DR2, \citet[][hereafter ER18]{ElBadry_2018} constructed a high-purity catalog of wide binaries within 200 pc of the Sun consisting mainly of AFGKM dwarfs. In this paper, we use a subset of that catalog to constrain the mass ratio distribution over a wide range of primary masses ($0.1 \lesssim M/M_{\odot} \lesssim 2.5$), mass ratios ($0.1 \lesssim q < 1$) and separations ($50 \lesssim s/{\rm AU} < 50,000$). The large size of the catalog allows us to constrain $p(q)$ in narrow bins of primary mass and separation independently. This approach make it possible to measure variation in $p(q)$ with mass and separation, and it minimizes the sensitivity of our results to imperfectly known inputs such as the initial mass function (IMF) and separation distribution.
A striking result of our investigation is the unambiguous evidence that twins are not purely a close-binary phenomenon: a significant excess of equal-mass ($q \gtrsim 0.95$) binaries persists out to separations as wide as 10,000\,AU. We derive constraints on the excess fraction of twins and the width of the twin excess as a function of mass and separation. We also provide constraints on the full mass ratio distribution over $0.1 \lesssim q \leq 1$ in all bins.
The remainder of this paper is organized as follows. Section~\ref{sec:data} describes the binary catalog and tests we have done to verify that the twin excess is real. In Section~\ref{sec:modeling}, we describe how we forward-model synthetic binary populations to fit for the intrinsic mass ratio distribution. Results of this fitting are presented in Section~\ref{sec:results}. In Section~\ref{sec:discussion}, we compare to previous work and discuss implications of our results for models of binary star formation and dynamical evolution. The appendices provide additional details about several aspects of the data and model. There we discuss sensitivity to the adopted parametric form of $p(q)$, (Appendix~\ref{sec:func_form}), systematic uncertainties in our model (Appendix~\ref{sec:systmatics}), evidence for a twin excess in archival wide binary catalogs (Appendix~\ref{sec:other_catalogs}), empirical determination of the selection function (Appendix~\ref{sec:selection_function_details}), and validation of the adopted Galactic model and selection function (Appendix~\ref{sec:model_validation}). Constraints on fitting parameters are tabulated in Appendix~\ref{sec:full_constraints}.
\section{Data}
\label{sec:data}
\begin{figure*}
\includegraphics[width=\textwidth]{cmd_kde_all.pdf}
\caption{Color-magnitude diagram of all stars (both primaries and secondaries) in the \citetalias{ElBadry_2018} catalog of resolved wide binaries. Right y-axes and upper x-axes show the mass and effective temperature corresponding to a given magnitude and color for main-sequence stars with $\rm [Fe/H]=0$.
In the upper left panel, points are colored by a Gaussian kernel density estimate of the local density.
We compare PARSEC isochrones with a range of metallicity (upper right) and age (lower left) to the data.
The bottom-right panel divides the CMD into three regions.
Black points (main-sequence components with no bright unresolved sub-components) comprise the bulk of our sample, and their masses are estimated using single-star isochrones.
Yellow points (components of wide binaries with a bright unresolved companion) are kept, but their masses are estimated using unresolved binary isochrones.
Red points (pre-main sequences stars, unresolved triples, and evolved stars) are discarded.}
\label{fig:cmds}
\end{figure*}
Our primary analysis uses the binary catalog assembled in \citetalias{ElBadry_2018}. This catalog contains $\sim$55,000 spatially resolved wide binaries with main sequence (MS), giant, and white dwarf components, and projected physical separations of $50 \lesssim s/{\rm AU} < 50,000$. We refer to \citetalias{ElBadry_2018} for a full description of the catalog's contents. In brief, it was constructed by searching {\it Gaia} DR2 for nearby ($d < 200$\,pc) pairs of stars whose positions, proper motions, and parallaxes are consistent with being gravitationally bound. Resolved higher-order multiples and suspected members of bound and dissolving clusters were removed. The catalog is designed to be pure but not complete: cuts on photometric and astrometric precision ensure that the contamination rate from chance alignments is low ($\ll1$\%), but they also reduce the number of faint binaries.
We do not use the full catalog from \citetalias{ElBadry_2018}, but impose the following additional cuts:
\begin{itemize}
\item We only consider MS/MS binaries, removing binaries in which either component is suspected to be a white dwarf, subgiant, giant, or pre-main sequence star. We identify non-MS components from the color-magnitude diagram (CMD; see below).
\item We require {\it both} components to have \texttt{parallax} > 5, \texttt{parallax\_over\_error} > 20, \texttt{phot\_bp\_mean\_flux\_over\_error} > 20, and \texttt{phot\_rp\_mean\_flux\_over\_error} > 20. \citetalias{ElBadry_2018} used these same cuts for the primary, but used less stringent cuts for the secondary. Here we apply the same cuts to both components in order to symmetrize the selection function.
\item We reject binaries in which the CMD-inferred mass of the primary falls outside $0.1 < M_1/M_{\odot} < 2.5$. The \citetalias{ElBadry_2018} catalog contains fewer than 100 binaries with estimated primary masses above 2.5 $M_{\odot}$, and none with estimated primary masses below $0.1 M_{\odot}$.
\end{itemize}
These additional cuts remove 23\% of the \citetalias{ElBadry_2018} catalog, leaving us with a sample of 42,338 MS/MS binaries.
Figure~\ref{fig:cmds} shows the CMD of all stars in the \citetalias{ElBadry_2018} catalog, with primaries and secondaries included on the same axes. Overplotted PARSEC isochrones show that the spread on the lower main sequence is primarily attributable to metallicity (upper right panel), while that on the upper main sequence and red giant branch is primarily due to age (lower left panel). A secondary sequence consisting mainly of unresolved binaries is visible above the main sequence. Wide ``binaries'' with one component in this sequence are primarily hierarchical triples. We do not remove these from our sample but account for them in our model when fitting for the mass ratio distribution in Section~\ref{sec:modeling}.\footnote{We also experimented with removing binaries with suspected unresolved components from the sample; doing so does not qualitatively change any of our results.}
\begin{figure*}
\includegraphics[width=\textwidth]{examples_sdss_binaries.pdf}
\caption{SDSS images of a selection of binaries from our catalog. Top panel shows binaries in which the primary is a solar-type star ($0.9 < M_1/M_{\odot}<1.1$) and the mass ratio varies from $\approx 0.1$ to 1. The primary is at the center of each image, and the secondary is circled. Middle panel shows binaries in which the primary is a late K dwarf ($0.4 < M_1/M_{\odot}<0.6$), again for a range of mass ratios.
Bottom panels shows examples of ``twin'' binaries with near-identical components, each with $0.4 < M_1/M_{\odot}<0.6$.
Each image is 100 (top panel) or 45 (bottom panels) arcsec on a side.
Our analysis uses photometry from {\it Gaia}, not SDSS. Because the SDSS photometry is ground-based, blending and source contamination affect it more severely. }
\label{fig:sdss_pictures}
\end{figure*}
MS components that fall below a main sequence PARSEC isochrone with $[\rm Fe/H]=0.5$ (black points in the lower right panel of Figure~\ref{fig:cmds}) are treated as single stars. Those that fall above this isochrone but below an unresolved binary isochrone for two equal-mass stars with $[\rm Fe/H]=0.5$ (yellow points) are treated as unresolved binaries. Finally, sources that fall above this binary isochrone (red points; likely a mix of unresolved triples, pre-MS stars, and giants/subgiants) are rejected, as are white dwarfs. Of the 42,382 wide binaries in our catalog, 35,087 have two components consistent with having no unresolved sub-components (black points), and 7,295 have at least one component suspected to be an unresolved binary (yellow points).
We estimate masses for both components of each binary based on their location in the CMD. The ratios of these masses are not used explicitly in fitting the mass ratio distribution, but they are used to assign the primary vs. secondary components and to assign binaries to bins of primary mass. For MS stars that are suspected to be single, we estimate masses by interpolating from a grid of single-star PARSEC isochrones. This method is reasonably effective for single components, but it would yield biased results for the unresolved sub-components.
To estimate masses for components suspected to be unresolved binaries, we construct a population of synthetic unresolved binaries (see Section~\ref{sec:modeling}) and, for the subset of this population that falls in the region of the CMD colored in yellow in the bottom right panel of Figure~\ref{fig:cmds}, we calculate the median {\it primary} mass as a function of $\rm M_{G}$ of the unresolved binary. For observed sources in that region of the CMD, we estimate the primary mass by interpolating from $\rm M_{G}$ on this median relation. That is, the mass assigned to unresolved components represents the mass of the primary of the unresolved component, not the total mass.
This method of assigning masses is not without drawbacks: masses assigned to unresolved binaries are imprecise, because the mass ratio is not known. In addition, some pre-MS stars may be mistaken for unresolved binaries, and some low-mass ratio unresolved binaries may be mistaken for higher-metallicity single stars. However, we expect the typical accuracy to be $\lesssim 0.1 M_{\odot}$, which is good enough for our purpose of assigning binaries to different bins of primary mass prior to fitting. In modeling the mass ratio distribution, it is not critical that the mass ratio of any one binary be measured accurately, but rather that the {\it distribution} of magnitude difference be predicted self-consistently.
Figure~\ref{fig:sdss_pictures} shows ``postage-stamp'' images\footnote{\href{http://skyserver.sdss.org/dr14/en/tools/chart/listinfo.aspx}{skyserver.sdss.org/dr14/en/tools/chart/listinfo.aspx}} of some of the binaries in our catalog from the Sloan Digital Sky Survey (SDSS; \citealt{York_2000}). We note that SDSS photometry is not used in our analysis or in fitting the mass ratio distribution; we show it here because raw images from {\it Gaia} are not publicly available. Source contamination in the SDSS images is generally expected to be more severe than in {\it Gaia} photometry due to atmospheric seeing. To showcase the diversity of binaries in the catalog, we choose a selection of binaries with roughly solar-mass primaries and a range of mass ratios (top panel), binaries with primary masses of $\approx 0.5\,M_{\odot}$ and a range of mass ratios (middle panel), and some examples of equal-mass ``twin'' binaries with component masses of $\approx 0.5 M_{\odot}$ (bottom panel). The twins are easily recognized because their magnitudes and colors are very similar.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{delta_G_points.pdf}
\caption{Apparent $G$-band magnitude difference between the two components of binaries in our catalog as a function of the apparent magnitude of the primary (left), angular separation (middle), and projected physical separation (right). The sign of $\Delta G$ is randomized for easier visualization. There is a clear excess of equal-brightness binaries with $\Delta G \approx 0$. These ``twin'' binaries are found over a large range of apparent magnitudes, preferentially at closer physical and angular separations. The middle panel shows the contrast sensitivity limit for our sample; at fixed angular separation, the probability of a companion passing our photometric quality cuts drops rapidly outside this limit due to source contamination (Appendix~\ref{sec:sensitivy}). The lack of binaries with large $\Delta G$ at small separations is a selection effect; the narrow excess at $\Delta G\approx 0$ is not.}
\label{fig:dG_points}
\end{figure*}
An observable quantity that is closely related to the mass ratio is the difference in apparent magnitude of the two components, $\Delta G = |G_1 - G_2|$.
Figure~\ref{fig:dG_points} shows the distribution of $\Delta G$ as a function of apparent magnitude of the primary, angular separation, and projected physical separation. The sign of $\Delta G$ is randomized, such that the distribution about $\Delta G=0$ is symmetric. This is helpful in making the equal-brightness population stand out, since it would otherwise be squished against the x-axis. An excess of equal-brightness binaries is evident over a wide range of apparent magnitudes; it is strongest at close separations. As we will show, the feature is quite narrow: the density of binaries is enhanced primarily at magnitude differences of $\Delta G < 0.25$\,mag. This is much narrower than the range over which the selection function varies significantly: at $\theta > 2.5$\,arcsec, the contrast sensitivity is basically constant over $0< \Delta G < 2$ (see \citetalias{ElBadry_2018}; their Appendix A).
The middle panel of Figure~\ref{fig:dG_points} shows that there are no binaries with small separations and large $\Delta G$ in the catalog. This is a consequence of photometric contamination at small angular separations. The dashed line shows the contrast sensitivity limit derived in Appendix~\ref{sec:sensitivy}; this is the value of $\Delta G$ at which the sensitivity is 50\% of its value at asymptotically large separations for a given $\theta$. The contrast limit is derived from the correlation function of chance alignment sources subject to similar quality cuts as real binaries. The fall-off in sensitivity is quite steep beyond the contrast limit, leading to an envelope in $\Delta G (\theta)$ beyond which no binaries are found. The drop-off toward larger $\Delta G$ is less steep as a function of physical separation $s$, since binaries with similar $s$ have different angular separations at different distances.
\begin{figure*}
\includegraphics[width=\textwidth]{chance_alignments_vs_real_binaries.pdf}
\caption{Difference in $G$-band magnitude (top) and color (bottom) between the two components of pairs with a range of angular separations. We compare genuine binaries (left) to chance-alignments (right). The chance alignments are required to pass the same quality cuts as the true binaries. A clear excess of pairs with $\Delta G\approx 0$ and $\Delta (G_{\rm BP} - G_{\rm RP})\approx 0$ is evident for the true binaries, but not for the chance alignments. The absence of the excess among chance alignments bolsters our confidence that the feature found in the real binaries is due to a true excess of equal-mass twins and is not a selection effect or data artifact. }
\label{fig:chance_align}
\end{figure*}
\subsection{Is the twin excess real?}
\label{sec:sanity_checks}
To test whether the narrow excess of binaries with $\Delta G \approx 0$ is a real astrophysical effect (as opposed to a data artifact), we produced a control sample of chance alignments with similar observable properties to the real binaries. This sample was produced by repeating the procedure used to produce the real binary catalog (applying the same quality cuts and limits on photometric and astrometric precision), but requiring that the two stars have parallaxes and proper motions that are {\it inconsistent} rather than consistent. This selects pairs of stars that are close together on the sky (and thus are affected by contamination and blending in the same way as real binaries) but are not physically associated. We applied the same procedure for removing resolved higher-order multiples and potential members of bound and dissolving clusters that was used for real binaries. Because chance alignments are intrinsically rare at close angular separations, we searched out to 400 pc (rather than 200 pc for the fiducial binary catalog) to obtain better statistics. We verified that our conclusions are unchanged when only the sample within 200 pc is considered.
In Figure~\ref{fig:chance_align}, we compare the distributions of magnitude and color difference for real binaries in our catalog (left panels) and chance alignments (right panels). Chance alignments are more common at large angular separations. The broad distribution and outer envelope of $\Delta G$ and $\Delta(G_{\rm BP}-G_{\rm RP})$ at a given $\theta$ is similar for binaries and chance alignments, reflecting {\it Gaia's} contrast sensitivity. For the chance alignments, there is no sharp excess of pairs with nearly-equal magnitude and color. Because chance alignments are subject to same cuts on astrometric and photometric quality and signal-to-noise as the real binaries, any aspects of the {\it Gaia} source detection algorithm that might be expected to produce a bias toward equal-brightness pairs should affect real binaries and chance alignments very similarly. We therefore interpret the lack of a thin excess at $\Delta G \approx 0$ and $\Delta(G_{\rm BP}-G_{\rm RP}) \approx 0 $ for chance alignments as strong evidence that the twin excess among real binaries is astrophysical.
An initial worry was that apparent twins might be duplicate {\it Gaia} sources that were not properly removed: if a source were observed twice in two different scans without being identified as a duplicate, it could manifest in our catalog as an apparent binary pair in which the two components had essentially identical astrometry and photometry. We verified that unrecognized duplicate sources are {\it not} the source of the twin signal using the SDSS images: we visually inspected the postage stamps of several hundred equal-brightness binaries in the SDSS footprint, and all the systems indeed contain two stars.
\begin{figure*}
\includegraphics[width=\textwidth]{three_samples_excess_vs_theta_and_s.pdf}
\caption{Ratio of the number of binaries with nearly equal magnitudes ($\Delta G < 0.25$) to the number with slightly different magnitude ($0.25 < \Delta G < 0.5$) as a function of physical separation (left) and angular separation (right). This ratio quantifies the excess of equal-brightness ``twins''. Error bars are 1 sigma. We compare binaries at different distances, as well as a control sample of chance-alignments.
At fixed \textit{physical} separation, the twin excess is consistent with being independent of distance. At fixed \textit{angular} separation, it decreases with distance.
Together, these trends strongly suggest that the $\Delta G\approx 0$ excess (Figures~\ref{fig:dG_points} and~\ref{fig:chance_align}) reflects a real excess of equal-mass binaries, not a selection effect or photometric issue at close angular separations.}
\label{fig:three_distances}
\end{figure*}
Another useful test to verify that the $\Delta G\approx 0$ excess is physical is to determine whether its strength depends primarily on physical or angular separation. This can be accomplished by comparing the $\Delta G$ distributions of binaries at different distances. If the excess of equal-brightness pairs were due to an observational bias against binaries that are not nearly equal-mass (or if there were an issue with the {\it Gaia} photometry causing close pairs to erroneously have the same reported magnitude), one would expect the twin excess to depend on angular separation, manifesting itself at different {\it physical} separations for binaries at different distances. If the excess is due to an astrophysical preference for equal-mass binaries, then it should be primarily a function of $s$, manifesting itself at the same physical separation but different $\theta$ for binary samples at different distances.\footnote{An implicit assumption here is that the fractional excess of twins is independent of the intrinsic properties of a binary (e.g. mass), since binaries that pass our quality cuts at larger distances will be more massive on average than those that are nearby. We show in Section~\ref{sec:results} that this assumption does not hold up in detail, which could lead to modest variation with distance in the observed twin excess at fixed physical separation.}
In Figure~\ref{fig:three_distances}, we compare the excess of equal-brightness pairs for binaries in three different distance bins as a function of physical separation (left) and angular separation (right). We measure the excess as the ratio of the number of binaries with $0 < \Delta G < 0.25$ (nearly-equal magnitudes) to the number with $0.25 < \Delta G < 0.5$ (slightly different, but still similar magnitudes). In order to include binaries at distance larger than 200\,pc, we repeated the binary search from \citetalias{ElBadry_2018}, this time searching out to 400\,pc. The additional binaries with $200 < d/{\rm pc} < 400$ are considered for this test only and are not included in the sample used to fit for the mass ratio distribution.
The left panel of Figure~\ref{fig:three_distances} shows that at fixed {\it physical} separation, the fractional excess of twins increases toward closer separations in a manner that is consistent across the three different distance samples. The right panel shows that as expected, the twin excess at fixed {\it angular} separation varies with distance. At fixed $\theta$, larger distances correspond to wider physical separation. Because the twin excess decreases with physical separation, it decreases with distance at fixed $\theta$. The excess of equal-brightness chance alignments is also shown as a function of angular separation in the right panel of Figure~\ref{fig:three_distances}. As expected, this is nearly consistent with 0 at all angular separations, meaning that there is no strong bias toward equal-brightness pairs compared to pairs with slightly different brightness. At the closest angular separations ($\theta \lesssim 5\,\rm arcsec$), there is a slight excess of equal-brightness chance alignments, suggesting that contrast sensitivity begins to play a role. The excess for the chance alignments is less than that found for the true binaries at all distances and is self-consistently accounted for in the selection function (Appendix~\ref{sec:sensitivy}).
\begin{figure}
\includegraphics[width=\columnwidth]{delta_G_mass_bins.pdf}
\caption{Magnitude difference as a function of projected physical separation for binaries with different primary masses. The sign of $\Delta G$ is randomized for easier visualization, and masses are computed from the CMD. The thin band of ``twin'' binaries with $\Delta G \approx 0$ is clearly visible in all but the highest-mass bin. In the top three bins, there is a clear preference for unequal-mass (large $\Delta G$, low-$q$) binaries. Incompleteness to faint companions prevents the detection of large $\Delta G$ companions to low-mass secondaries. }
\label{fig:delta_G_mass_bins}
\end{figure}
Figure~\ref{fig:delta_G_mass_bins} shows distributions of $\Delta G$ for binaries in 5 different bins of primary mass. Both the overall shape of the distribution and the strength of the excess at $\Delta G\approx 0$ vary substantially between mass bins. For the highest-mass bin, $1.2 < M_1/M_{\odot} < 2.5$, the excess at $\Delta G \approx 0$ is weak, and the observed distribution (without accounting for incompleteness) peaks at $\Delta G\approx 8$, corresponding to $q\approx 0.3$ (see Figure~\ref{fig:delta_G_vs_q}). For $0.8 < M_1/M_{\odot} < 1.2$ and $0.6 < M_1/M_{\odot} < 0.8$, there is a clear excess of twins out to $s\approx 1000$\,AU, with the observed peaks in the $\Delta G$ distribution corresponding to $q\approx 0.45$ and $q\approx 0.6$, respectively. In the two lowest-mass bins, the visible twin excess appears to extend to larger separations, and there is no secondary peak in the distributions of $\Delta G$. For all mass bins, the binaries with the largest $\Delta G$ have secondaries with $M \approx 0.1 M_{\odot}$. The maximum $\Delta G$ in each panel is probably set largely by observational incompleteness, since at moderately old ages ($\gg 100\,\rm Myr$), objects below the hydrogen burning limit are several magnitudes fainter than those just above it. Incompleteness at small separations due to the angular resolution and contrast sensitivity of our catalog becomes significant at different projected physical separations for different mass bins, because the low-mass binaries that pass our quality cuts are on average at closer distances than the higher-mass binaries.
\section{Modeling}
\label{sec:modeling}
We now turn to the intrinsic mass ratio distribution, $p(q)$. Because we expect that $p(q)$ may vary with primary mass and/or separation, we split the observed sample into 5 bins of primary mass and 7 bins of projected physical separation, fitting each of the resulting 35 subsamples independently. Our choice of bins is designed to balance the number of binaries in each bin while still covering a large dynamic range of mass and separation.
We constrain $p(q)$ for each subsample by forward-modeling a simulated population of binaries with a given distribution of primary mass, age, metallicity, distance, physical separation, and mass ratio, passing them through the selection function, and comparing to the data. The ``data'' we consider is the distribution of the observed binaries in the 3-dimensional space of angular separation $\theta$, magnitude difference $\Delta G$, and parallax $\varpi$. The best-fitting $p(q)$ is then the one that best matches the observed data, and uncertainties in $p(q)$ are estimated via Markov chain Monte Carlo from the range of $p(q)$ that adequately reproduce the observed data. This approach requires a parameterized form of $p(q)$ (Section~\ref{sec:functional_form}), knowledge of the selection function (Section~\ref{sec:selection_function}), and a Galactic model from which simulated binaries are drawn (Section~\ref{sec:model_inputs}).
To generate a model prediction for a given set of fitting parameters, we generate a population of $N=10^6$ synthetic binaries and forward-model their distribution into the space of observables. Masses, ages, metallicities, and distances are Monte Carlo sampled from their respective distribution functions. For mass ratios and separations, we use a regular $1000 \times 1000$ grid, weighting the synthetic binary at each gridpoint by the mass ratio and separation distributions. This approach is chosen because in the fiducial model, we fit for the mass ratio and separations distributions but leave the distributions of age, mass, distance, and metallicity fixed. The number of synthetic binaries generated must be large enough that Poisson noise is negligible. We verified that $N=10^6$ binaries in each mass and separation bin is large enough that our constraints are converged and insensitive to the random seed.
Synthetic photometry is calculated for both components from isochrones (including the effects of unresolved companions in hierarchies; see Section~\ref{sec:model_inputs}), and the observable properties of each binary are passed through the selection function. In constructing the distribution of mock-observables to be compared with the data, each synthetic binary is weighted by the selection function evaluated for its observables. Finally, the observed and simulated distributions are binned on a regular 3D grid. We default to using 100 bins in $\Delta G$, 100 bins in $\log \theta$, and 5 bins in $\varpi$ (because typical errors in $\varpi$ are larger than those in $\Delta G$ or $\theta$). The maximum value of the $\Delta G$ grid for a given primary mass bin is chosen to include the largest $\Delta G$ value for the data in that bin, and the 5 bins in $\varpi$ are chosen so that roughly the same number of observed binaries fall in each bin. Our constraints are not sensitive to the choice of bins, since they are small compared to the scale on which the data exhibit substructure.
We re-scale the binned model prediction such that the total counts match the observed data. We calculate the likelihood for a particular set of model parameters by summing over all cells in the 3D distribution, assuming that the distribution of counts in each cell is set by a Poisson process. The log-likelihood function is
\begin{equation}
\ln L=\sum_{m_{i}\neq0} \left[d_{i}\ln m_{i}-m_{i}-\ln(d_{i}!)\right],
\label{eq:lnL}
\end{equation}
where $m_i$ and $d_i$ are the counts in the $i$th cell of the binned model and data (``!'' denotes a factorial). We sample from the posterior using \texttt{emcee} \citep{ForemanMackey_2013}, using priors described in Section~\ref{sec:functional_form}. We use 200 walkers and draw 20,000 samples for each bin of mass and separation after a burn-in period of 200 steps per walker. Inspecting the chains, we find this to be sufficient for convergence in all cases. We carried out tests with mock data that was drawn from a known mass ratio distribution assigned with realistic observational uncertainties in order to verify that our approach yields unbiased constraints on $p(q)$.
This fitting procedure is qualitatively very similar to the method commonly used to constrain population properties such as the IMF, star formation history, unresolved binary fraction, or initial-final mass relation from CMDs \citep[e.g.][]{Dolphin_2002,Bonatto_2012, Geha_2013, ElBadry_2018_IFMR}. The difference between our approach and these studies is that we are forward-modeling the distribution of binaries in the space of angular separation, magnitude difference, and parallax rather than color and magnitude. The approach can in principle be generalized to include other observables, such as color difference or apparent magnitude of the primary, but the computational expense increases rapidly with the dimension of the space in which the likelihood function is calculated.
We note that the data uncertainties do not enter Equation~(\ref{eq:lnL}). The implicit assumption (which does hold for our problem setup) is that the uncertainties are small compared to the scale of the bins in all quantities. We also note that for a fine grid, a majority of grid cells will contain 0 or 1 real binaries. This is not a problem; Equation~(\ref{eq:lnL}) does not make any assumptions about the magnitude of $d_i$.
\subsection{Parameterization}
\label{sec:functional_form}
\begin{figure}
\includegraphics[width=\columnwidth]{schematic_q_distribution.pdf}
\caption{Parameterized mass ratio distribution. The broad part of the distribution is modeled with a broken power law of logarithmic slope $\gamma_{\rm smallq}$ at $q<q_{\rm break}$ and $\gamma_{\rm largeq}$ at $q>q_{\rm break}$. $F_{\rm twin}$ is the \textit{excess} fraction of nearly equal-mass binaries with $q > q_{\rm twin}$, relative to the underlying power-law distribution for $q>0.3$. For this particular example, $F_{\rm twin} = 0.04$, $q_{\rm twin} = 0.95$, $\gamma_{\rm smallq} = 0.3$, $\gamma_{\rm largeq}=-1.3$, and $q_{\rm break} = 0.5$.}
\label{fig:schematic_q_dist}
\end{figure}
We fit the mass ratio distribution by assuming a parameterized form of $p(q)$ and then obtaining constraints on the parameters. Our fiducial parameterization is motivated by the one used by \citet{Moe_2017} and is shown in Figure~\ref{fig:schematic_q_dist}. $p(q)$ is parameterized as a broken power law with logarithmic slope $\gamma_{\rm smallq}$ at $q < q_{\rm break}$ and $\gamma_{\rm largeq}$ at $q > q_{\rm break}$. An possible excess (or deficit) of twins is added on top of the power law component at $q>q_{\rm twin}$. This excess is modeled as a step function, with the magnitude such that the integrated excess of twins divided by the total number of binaries with $q > 0.3$ is $F_{\rm twin}$. The reason for this choice (as opposed to normalizing relative to {\it all} binaries) is that, compared to the constraints at $q > 0.3$, the constraints on $p(q)$ at small $q$ are often weak due to incompleteness. $p(q)$ is set to 0 at $q < 0.05$ to prevent divergence when $\gamma_{\rm smallq}<-1$. Because we are not sensitive to companions with $q < 0.05$, this has little effect on our results.
There is no a priori motivation for this particular parameterization: because the physics that set $p(q)$ are imperfectly understood, we simply require a functional form that is sufficiently flexible to reproduce the observed distributions of $\Delta G$. We have experimented with other forms of $p(q)$, including adding an additional break point to the power law and modeling the twin excess as ramping up linearly instead of increasing stepwise. However, we find the form shown in Figure~\ref{fig:schematic_q_dist} to provide a good fit in all mass and separation bins, with more complicated models providing only marginal improvements. We thus use this functional form for our main analysis. In Appendix~\ref{sec:func_form}, we show results of using alternative parameterizations, including one that smoothly transitions between the two power laws instead of including a sharp break and one that leaves the shape of the twin excess flexible.
There are two differences between our parameterization of $p(q)$ and the one used by \citet{Moe_2017}, who focused primarily on binaries with $M_1 \gtrsim M_{\odot}$. First, they fixed $q_{\rm break} = 0.3$, a value that was appropriate for their analysis because several studies have found the mass ratio distribution to peak at $q \approx 0.2-0.3$ for binaries with $M_1 \gtrsim M_{\odot}$ \citep{Duquennoy_1991, Gullikson_2016, Murphy_2018}. We also use $q_{\rm break} = 0.3$ for our highest-mass bin, but we find $q_{\rm break} = 0.5$ to provide a better fit at lower masses (see Appendix~\ref{sec:qbreak}). Second, they fixed $q_{\rm twin} = 0.95$, roughly the value found for spectroscopic binaries \citep[e.g.][]{Tokovinin_2000}. In order to identify or rule out trends with mass and separation, we leave $q_{\rm twin}$ as a free parameter.
\begin{table*}
\centering
\caption{Summary of priors adopted in each primary mass bin. We use the same priors for all separation bins. $\mathcal{U}(a,b)$ represents a uniform distribution over $[a,b]$, and $\mathcal{N}(\mu,\sigma)$ represents a normal distribution with mean $\mu$ and dispersion $\sigma$. $\gamma_s$ is the logarithmic slope of the local separation distribution; i.e., $p(s) \propto s^{\gamma_s}$. Other parameters are described in Figure~\ref{fig:schematic_q_dist}.}
\label{tab:priors}
\begin{tabular}{ |c|c|c|c|c|c|}
\hline
& $0.1 < M_1/M_{\odot} < 0.4$ & $0.4 < M_1/M_{\odot} < 0.6$ & $0.6 < M_1/M_{\odot} < 0.8$ & $0.8 < M_1/M_{\odot} < 1.2$ & $1.2 < M_1/M_{\odot} < 2.5$ \\
\hline
$F_{\rm twin}$ & $\mathcal{U}(-1, 1)$ & $\mathcal{U}(-1, 1)$ & $\mathcal{U}(-1, 1)$ & $\mathcal{U}(-1, 1)$ & $\mathcal{U}(-1, 1)$ \\
$q_{\rm twin}$ & $\mathcal{U}(0.93, 1)$ & $\mathcal{U}(0.93, 1)$ & $\mathcal{U}(0.93, 1)$ & $\mathcal{U}(0.93, 1)$ & $\mathcal{U}(0.93, 1)$ \\
$\gamma_{\rm largeq}$ & $\mathcal{N}(0.5, 1)$ & $\mathcal{N}(-1,1)$ & $\mathcal{N}(-1,1)$ & $\mathcal{N}(-1,1)$ & $\mathcal{N}(-1,1)$ \\
$\gamma_{\rm smallq}$ & --- & $\mathcal{N}(0.5, 0.5)$ & $\mathcal{N}(0, 0.5)$ & $\mathcal{N}(0, 0.5)$ & $\mathcal{N}(0, 0.5)$ \\
$\gamma_{\rm s}$ & $\mathcal{N}(-1.5, 1)$ & $\mathcal{N}(-1.5, 1)$ & $\mathcal{N}(-1.5, 1)$ & $\mathcal{N}(-1.5, 1)$ & $\mathcal{N}(-1.5, 1)$ \\
$q_{\rm break}$ & --- & 0.5 & 0.5 & 0.5 & 0.3 \\
\hline
\end{tabular}
\end{table*}
Along with the parameters of the mass ratio distribution, we also fit for $\gamma_s$, the local logarithmic slope of the separation distribution in each bin of primary mass and separation. Opik's law (a uniform distribution of $\log(s)$) corresponds to $\gamma_s=0$.
Our adopted priors are listed in Table~\ref{tab:priors}.
For the lowest-mass bin, the smooth component of $p(q)$ is modeled as a single power law, because no useful constraints can be obtained at $q \lesssim 0.5$. Our priors on $\gamma_{\rm smallq}$ and $\gamma_{\rm largeq}$ are loosely motivated by constraints from the literature (particularly \citealt{Tokovinin_2014} and \citealt{Moe_2017}) but are fairly weak. We use a uniform flat prior on $F_{\rm twin}$. We require $q_{\rm twin}>0.93$ in order to prevent cases where a large $F_{\rm twin}$ is combined with $q_{\rm twin}\ll 1$ such that a broad ``twin excess'' simply modifies the overall shape of $p(q)$ but does not actually correspond to a sharp increase near $q=1$. We show in Appendix~\ref{sec:how_sharp} that, where a twin excess is significant, it is always narrow ($q_{\rm twin} \gtrsim 0.94$).
\subsection{Selection Function}
\label{sec:selection_function}
Because both components of a binary must pass astrometric and photometric quality cuts, the binaries in our catalog are relatively bright. The median apparent magnitude of all stars in our fitting sample (considering primaries and secondaries together) is $\left\langle G\right\rangle =14.5$, and 90\% (99\%) of stars fall in the range $9.1 < G < 17.4$ ($6.6 < G < 18.1$). For stars in this magnitude range, {\it Gaia} DR2 is nearly complete outside of crowded fields \citep{Arenou_2018, Sollima_2019}. The completeness is not quite 100\% due to a variety of issues, but it is primarily a function of position on the sky, not color or magnitude. The selection function for our sample is thus determined by the cuts imposed on astrometric and photometric precision.
In order for a binary to appear in the catalog, (a) both components must be bright enough that they individually pass the cuts we impose on \texttt{parallax\_over\_error} and photometric signal to noise, and (b) they must not be so close on the sky that the photometry of either component is significantly contaminated. The selection function for binaries can thus be expressed as a product of the two components' single-star detection probabilities and a contrast sensitivity cross term:
\begin{align}
\label{eq:s_12}
s_{\rm binary}=s_{1}\times s_{2}\times s_{\Delta G}(\theta).
\end{align}
Here $s_1$ and $s_2$ represent the independent probabilities of detecting an isolated star with the observable properties of star 1 and star 2; they depend primarily on apparent magnitude and color. $s_{\Delta G}(\theta)$ quantifies the reduction in the probability of detecting star 1 and star 2 together, {\it relative to the probability of detecting them at asymptotically large separation}. It is primarily a function of the angular separation of the two stars and their flux ratio. For example, $s_{\Delta G}(\theta)\approx 0$ at $\theta < 2\,{\rm arcsec}$ and 1 at $\theta > 10\,{\rm arcsec}$; at intermediate separations, it depends strongly on $\Delta G$. We calculate the single-star selection function given our quality cuts in Appendix~\ref{sec:single_star_term} and $s_{\Delta G}(\theta)$ in Appendix~\ref{sec:sensitivy}. The derived selection functions are then validated in Appendix~\ref{sec:model_validation}.
\subsection{Model inputs}
\label{sec:model_inputs}
We draw primary masses and system ages assuming a \citet{Kroupa_2001_IMF} IMF and a constant star formation history over the last 10\,Gyr. Because suspected members of bound and dissolving clusters are removed from our binary catalog, we remove synthetic binaries with $\rm age< 100$\,Myr. We assume that the wide binary fraction, $f_{\rm wb}$, scales with mass as $f_{\rm wb}\propto M_1^{\alpha_{\rm wb}}$, where $\alpha_{\rm wb} = 0.4$ is a constant (i.e., higher-mass primaries are more likely to have wide binary companions). The effect of this assumption is that primary masses are drawn from a distribution with logarithmic slope $\alpha_{\rm IMF} + \alpha_{\rm wb}$, where $\alpha_{\rm IMF}$ is the local logarithmic slope of the IMF ($\alpha_{\rm IMF} = -2.3$ for $M_1/M_{\odot} > 0.5$ and $\alpha_{\rm IMF} = -1.3$ for $M_1/M_{\odot} < 0.5$). We find that with this choice of $\alpha_{\rm wb}$, our model predicts a distribution of primary magnitudes in reasonably good agreement with that of the binary catalog when the selection function is taken into account.
We model the intrinsic spatial distribution of all stars as a plane-parallel exponential with the Sun at the midplane. The exponential scale height increases with stellar age \citep[e.g.][]{Nordstrom_2004, Seabroke_2007}, because older stars (a) were born from kinematically hotter gas and (b) have been dynamically heated more since their formation \citep[e.g][]{Ting_2018}. We use a fit to the empirical age-scale height relation recently measured by \citet{Sollima_2019} using \textit{Gaia} star counts:
\begin{align}
\log\left(h_{z}/{\rm pc}\right)=0.53\log\left({\rm age}/{\rm yr}\right)-2.65.
\label{eq:hz}
\end{align}
I.e., the scale height increases from 40\,pc for stars of age 100 Myr, to 130 pc at age 1 Gyr, to 450 pc at age 10 Gyr. We show in Appendix~\ref{sec:single_star_validation} that this leads to a predicted distance distribution in good agreement the data.
We use the tabulated empirical metallicity distribution function (MDF) for our binary catalog that was measured in \citet{ElBadry_2019} by considering a subset of $\sim$8000 binaries in the catalog for which $\rm [Fe/H]$ was measured spectroscopically for at least one component.
Most of the binaries in our catalog are disk stars, with a median metallicity of $\left\langle \left[{\rm Fe/H}\right]\right\rangle \approx-0.1$ and tails extending to $[\rm Fe/H] \approx -1.0$ and $[\rm Fe/H] \approx +0.4$. We do not include any variation in the MDF with age or distance.
We generate synthetic photometry in the {\it Gaia} DR2 bands from \citet{Evans_2018} using PARSEC\footnote{\href{http://stev.oapd.inaf.it/cgi-bin/cmd}{http://stev.oapd.inaf.it/cgi-bin/cmd}} isochrones \citep{Bressan_2012, Chen_2014}. Just as for the real data, we remove synthetic binaries in which either component has evolved off the main sequence.
For companions with $M_2 < 0.1\,M_{\odot}$, we supplement the PARSEC models with \texttt{BT-Settl} models for very low-mass stars and brown dwarfs \citep{Allard_2012, Allard_2014}.\footnote{Synthetic photometry in {\it Gaia} DR2 bands is computed for the \texttt{BT-Settl} models using the Pheonix web simulator, available at \href{https://phoenix.ens-lyon.fr/simulator-jsf22-26/index.faces}{phoenix.ens-lyon.fr/simulator-jsf22-26/index.faces}.} We include companions with masses as low as $0.01\,M_{\odot}$ in our model for completeness. However, these have very little effect on our results, because substellar companions are too faint and red to pass our quality cuts except at very young ages.
\begin{figure}
\includegraphics[width=\columnwidth]{delta_G_vs_q.pdf}
\caption{Magnitude difference vs. mass ratio for binaries of different primary mass, computed from PARSEC isochrones. We assume both components have $[\rm Fe/H]=0$, adopting an age of 1 Gyr for the highest primary mass and 5 Gyr for the others. Because the mass-luminosity relation varies with mass, the relation between $\Delta G$ and $q$ does as well. A magnitude difference of 0.25 mag (roughly where the twin excess becomes significant in our catalog) corresponds to a mass ratio of 0.93 to 0.97, depending on primary mass.}
\label{fig:delta_G_vs_q}
\end{figure}
Figure~\ref{fig:delta_G_vs_q} shows the relation between $q$ and $\Delta G$ predicted for PARSEC isochrones. We show relations for 5 different primary masses, corresponding approximately to the median masses of primaries in each of our five mass bins, and use a minimum companion mass of $0.1M_{\odot}$ in all cases.
Comparing to Figure~\ref{fig:delta_G_mass_bins}, it is evident that the lowest-mass secondaries have $M_2\approx 0.1M_{\odot}$ in all bins of primary mass. As for the excess of equal-brightness binaries, it is primarily manifest over $0 < \Delta G < 0.25$ (though there is some variation with mass; see Figure~\ref{fig:posterior_predictive}), which corresponds to $0.93 \lesssim q < 1$.
Some components of wide binaries have their own spatially-unresolved close companions. We incorporate these in our models following observed binary statistics. The probability that a star has a close companion is taken to be a function of its mass, increasing from 20\% at $M < 0.2 M_{\odot}$, to 30\% at $0.2 < M/M_{\odot} < 0.5$, to 35\% at $0.5 < M/M_{\odot} < 0.8$, to 45\% at $M/M_{\odot} > 0.8$ \citep{Duchene_2013}. For components that are assigned a binary companion, we draw the unresolved companion mass from a mass ratio distribution that is uniform between $q_{\rm min} = 0.1 M_{\odot}/ M_1$ and 1. We assume that the two components' probabilities of having a close companion are independent and neglect dynamical stability constraints. This prescription reproduces the morphology of the observed CMD reasonably well. In particular, the fraction of stars within 200\,pc that fall in the ``suspected unresolved binary'' region of the CMD (yellow points in the bottom right panel of Figure~\ref{fig:cmds}) at a given $\rm M_G$ is reproduced within $\sim$25\% over $0.1 < M/M_{\odot} < 1.0$.
We do not include extinction or reddening due to dust in our model and do not attempt to correct the data for it. Because the stars in our catalog are nearby $(d < 200\,\rm pc)$, the effects of extinction are expected to be modest. The morphology of the CMD (e.g. the compactness of the red clump and main sequence; see Figure~\ref{fig:cmds}) validates this assumption. Moreover, because the two components of a binary have similar position on the sky (within an arcminute in most cases) and similar distance, the extinction toward both components is expected to be similar.
We discuss the sensitivity of our constraints on the mass ratio distribution to various model ingredients in Appendix~\ref{sec:systmatics}. The largest systematic uncertainties come from the choice of stellar models. All the systematics we consider primarily affect constraints at low $q$. This is also true for uncertainties in the completeness function: because the two components of binaries with $q\approx 1$ have similar magnitudes and colors, incompleteness affects them similarly. The translation between the distribution of $\Delta G$ and $p(q)$ is thus more straightforward at $q\approx 1$ than at $q \ll 1$.
\section{Results}
\label{sec:results}
Our fitting produces samples from the posterior distribution of free model parameters for each bin of mass and separation. These samples translate to marginalized constraints on each parameter of the mass ratio distribution and the covariances between them. An example for a single mass and separation bin is shown in Figure~\ref{fig:corner}. Marginalized constraints for all bins are listed in Appendix~\ref{sec:full_constraints}.
\begin{figure*}
\includegraphics[width=\textwidth]{qdist_ranges.pdf}
\caption{Mass ratio distribution constraints. Each panel corresponds to a single bin of primary mass (increasing left to right) and physical separation (increasing top to bottom). Gray shaded regions show 95.4\% probability. Normalization is arbitrary; the scale of the y-axis is linear and begins at 0. Vertical dashed lines show the lowest mass ratio of observed binaries in each bin. }
\label{fig:qdist_ranges}
\end{figure*}
The 35 panels of Figure~\ref{fig:qdist_ranges} show median and 2 sigma constraints on the mass ratio distribution for 35 bins of primary mass and separation. We do not show the y-axis ticks to avoid cluttering the figure, as the limits are different in each panel. The uncertainties (shaded regions) are derived by drawing 1000 samples from the posterior, calculating a normalized $p(q)$ for each sample, and then shading the middle 95.4\% range of these samples at each $q$. Solid black lines show the median at each $q$. Because $p(q)$ is normalized, uncertainty in the mass ratio distribution at {\it any} $q$ translates to uncertainty in the normalization of $p(q)$ at {\it all} $q$. This reflects the fact that it is impossible to know the total fraction of all binaries that fall in some mass ratio range if one does not know how many binaries with low mass ratios are missed. However, $\gamma_{\rm largeq}$, the slope of the mass ratio distribution at $q > q_{\rm break}$, is usually well-constrained. Dashed vertical lines in each panel show the lowest-mass ratio observed binary included in that bin. This roughly corresponds to the observational completeness limit and marks the mass ratio below which meaningful constraints cannot be obtained.
The fact that $p(q)$ is modeled as a double power law with a sharp break leads to some unphysical artifacts in Figure~\ref{fig:qdist_ranges}, including a sharp change in slope at $q=q_{\rm break}$ and artificially low uncertainty just below $q_{\rm break}$, which is caused by $p(q)$ ``pivoting'' about this point as $\gamma_{\rm smallq}$ varies and $p(q)$ is renormalized. We show in Appendix~\ref{sec:smoothly_broken} that these features are not present when we fit a more flexible ``smoothly-broken'' power law model. However, doing so introduces parameter covariances that are not present for the fiducial form of $p(q)$. We use the simpler sharply broken power law as our fiducial model to facilitate easier comparison between different mass and separation bins and comparison with the literature.
\begin{figure*}
\includegraphics[width=\textwidth]{posterior_predictive.pdf}
\caption{Magnitude difference between the two components of a binary (linear y-axis; normalization is arbitrary). Black histograms show real binaries, split into bins of primary mass (increasing left to right) and physical separation (increasing top to bottom). The twin excess can be seen in many panels as an excess of binaries with $\Delta G \approx 0$. Red histograms show Monte Carlo populations generated from the best-fit model. The number of observed binaries (black) and the marginalized 1 sigma constraints on $F_{\rm twin}$ (red) are listed in each panel. Overall, the model is quite successful in matching the observed distributions. }
\label{fig:posterior_predictive}
\end{figure*}
Figure~\ref{fig:posterior_predictive} compares the observed 1D distributions of $\Delta G$ to the predictions of the best-fit model; this is useful for assessing the quality of the fits. In each panel, we note the number of observed binaries and the marginalized 1 sigma constraints on $F_{\rm twin}$. In generating the model predictions, we use the median of the 1D marginalized posterior distribution for each free parameter. We then generate a Monte Carlo realization of the binary population, drawing primary masses, mass ratios, separations, ages, metallicities, and distances from the appropriate joint distributions, calculating synthetic photometry, and weighting each binary by the selection function evaluated for its observables.
Overall, the model predictions are in good agreement with the observed distributions of $\Delta G$. This indicates that our parameterization of $p(q)$ is suitable and sufficiently flexible. In panels where $F_{\rm twin}$ is inconsistent with 0, there is a clear excess of equal-brightness binaries. The distributions of $\Delta G$ near $\Delta G=0$ are also adequately reproduced by the model, indicating that the simple ``step function'' model for the twin excess is consistent with the data. Although it is not shown here, we also find the best-fit models to predict distributions of other observables (angular separation, parallax, and apparent magnitude of the primary) in good agreement with the data.
\begin{figure}
\includegraphics[width=\columnwidth]{Ftwin_vs_separation.pdf}
\caption{Marginalized constraints on $F_{\rm twin}$, the fractional excess of binaries with nearly equal mass, vs. physical separation. Each panel shows a separate bin of primary mass. Dark and light error bars show 1 and 2 sigma constraints. In all mass bins, $F_{\rm twin}$ declines with increasing separation and is consistent with 0 at the largest separations. The twin excess reaches the widest separations for $0.4 < M_1/M_{\odot} < 0.6$, where $F_{\rm twin}$ is $\sim$5\% out to 15,000 AU. We compare to constraints from the nearby solar-type binary sample of \citet[][panel 4]{Raghavan_2010} and the adaptive optics imaging survey from \citet[][panel 5]{DeRosa_2014}; these were derived by \citet{Moe_2017}. }
\label{fig:Ftwin_vs_separation}
\end{figure}
Figure~\ref{fig:Ftwin_vs_separation} shows the constraints on $F_{\rm twin}$ as a function of mass and projected physical separation. Light and dark error bars show middle 95.4\% and 68.2\% of the marginalized posterior distributions. Note that the y-axis scale is different in the top panel. In all bins of primary mass, the excess twin fraction falls with increasing separation and is negligible at the largest separations we consider ($s > 15,000$\,AU). At fixed separation, the magnitude of the excess varies with primary mass. For close separations ($50 <s/{\rm AU} < 350$), $F_{\rm twin}$ is largest in the lowest bin of primary mass ($0.1 < M_1/M_{\odot} < 0.4$). This may be in part because low-mass primaries in our sample are at closer distances on average, such that the median separation within the ($50 <s/{\rm AU} < 350$) bin is smaller than for higher-mass primaries.
The maximum separation out to which there is a significant twin excess also varies with primary mass. For the highest-mass bin, $F_{\rm twin}$ is consistent with 0 at $s>600$\,AU. For solar-type primaries, $F_{\rm twin}$ is only consistent with 0 at $s>5000$\,AU, but it is $<3\%$ for $s > 1000$\,AU. The fall-off in $F_{\rm twin}$ with increasing separation is shallowest for $0.4 < M_1/M_{\odot} < 0.6$ and $0.6 < M_1/M_{\odot} < 0.8$, with a larger normalization for $0.4 < M_1/M_{\odot} < 0.6$. Here $F_{\rm twin}$ is almost independent of separation over $350<s/{\rm AU}<15,000$. Finally, the fall-off steepens again in the lowest primary mass bin, where $F_{\rm twin}\approx 0$ beyond 2500\,AU. We discuss possible interpretations of these trends in Section~\ref{sec:widening}.
In the bottom two panels of Figure~\ref{fig:Ftwin_vs_separation}, we compare our constraints on $F_{\rm twin}$ to the 1 sigma constraints obtained by \citet[][their Tables 8 and 11]{Moe_2017} for binaries in similar mass and separation ranges. The constraints for solar-type stars ($0.8 < M_1/M_{\odot} < 1.2$) were obtained from the solar neighborhood sample of \citet{Raghavan_2010}. Those shown in the bottom panel were obtained from the AO-assisted survey of visual binaries with A-type primaries ($1.7 < M_1/M_{\odot} < 2.3$) described in \citet{DeRosa_2014}. Reassuringly, constraints for both mass bins are consistent with those obtained from our catalog. Because the binary sample we analyze is larger than the \citet{Raghavan_2010} and \citet{DeRosa_2014} samples, we can tighten the uncertainties on $F_{\rm twin}$ at large separations, showing, for instance, that $F_{\rm twin}$ for solar-type primaries is inconsistent with 0 out to $s\approx 5000$\,AU.
\begin{figure}
\includegraphics[width=\columnwidth]{gamma_vs_separation.pdf}
\caption{Marginalized constraints on $\gamma_{\rm largeq}$, the logarithmic slope of the power law component of the mass ratio distribution at large $q$, vs. projected separation. Dark and light error bars show 1 and 2 sigma constraints. The range of $q$ over which $\gamma_{\rm largeq}$ is fit in each mass bin is noted in the upper left of each panel.
Dashed lines show the slope expected if binary component masses were drawn from the IMF and paired randomly. Except in the lowest mass bin, $p(q)$ is more bottom-heavy than uniform, but it is always more top-heavy than expected for random IMF pairings.}
\label{fig:gamma_vs_separation}
\end{figure}
Figure~\ref{fig:gamma_vs_separation} shows constraints on $\gamma_{\rm largeq}$ as a function of separation, again separating binaries by primary mass. The dashed line in each panel shows the slope that would be expected for random pairings from the IMF. This is obtained by fitting a power law mass ratio distribution over the same range of primary masses and mass ratios to a simulated population of binaries in which the masses of both components are drawn from a \citet{Kroupa_2001_IMF} IMF and paired randomly.
Consistent with previous work \citep[e.g.][]{Lepine_2007, Reggiani_2011, Duchene_2013, Moe_2017}, we find that the mass ratio distribution is {\it not} consistent with random pairings from the IMF, but is weighted toward higher mass ratios than would be expected in such a scenario. Whether this is an imprint of the binary formation process or in part reflects the fact that binaries with higher mass ratio have higher binding energy and are thus more difficult to disrupt is an open question. Any trends in $\gamma_{\rm largeq}$ with separation are weak over the separation range we probe: at the 2 sigma level, our constraints are consistent with a separation-independent $\gamma_{\rm largeq}$ over $100 \lesssim s/{\rm AU} < 50,000$ in all mass bins. However, they are not consistent with being independent of primary mass: $p(q)$ becomes increasingly bottom-heavy (lower $\gamma_{\rm largeq}$) with increasing $M_1$.
The fact that $\gamma_{\rm largeq}$ does not vary much with separation can serve as a strong constraint on formation models for wide binaries.
It has frequently been argued that while binaries with separations of $100 \lesssim s/{\rm AU} \lesssim 5000$ formed primarily by core fragmentation, those with $s \gtrsim 5000$\,AU (the size of typical cloud cores) formed by another process. Candidate processes include cluster dissolution \citep{Kouwenhoven_2010, Moeckel_2010}, unfolding of hierarchical triples \citep{Reipurth_2012}, or pairing of adjacent cores \citep{Tokovinin_2017b}. One might naively expect a change in the mass ratio distribution at $s\sim 5000$\,AU if the binary formation mechanism changes there, but none is observed.
The mass ratio distribution for wide solar-type binaries is not uniform, but is weighted towards low mass ratios. For example, companions with $q\approx 0.5$ are roughly twice as common as those with $q \approx 0.9$. The dominance of low-mass ratio companions can be seen clearly in the data at wide separations (Figure~\ref{fig:posterior_predictive}) and cannot be due to selection effects, which all work {\it against} low-mass ratio binaries. $p(q)$ is thus more bottom-heavy at wide separations than at close separations, where it is basically uniform \citep{Mazeh_1992, Tokovinin_2014}. Analyzing the 25-pc \citet{Raghavan_2010} sample of solar-type binaries, \citet{Moe_2017} found the mass ratio distribution to transition from $\gamma_{\rm largeq} = -0.4 \pm 0.3$ (close to uniform) across $10<s/{\rm AU}<200$ to $\gamma_{\rm largeq} = -1.1 \pm 0.3$ across $200<s/{\rm AU}<5000$.\footnote{Because we use $q_{\rm break}=0.5$ for solar-type primaries and \citet{Moe_2017} used $q_{\rm break} = 0.3$, our measurements should not be directly compared. $p(q)$ flattens at $q < q_{\rm break}$, so at fixed $\gamma_{\rm largeq}$, a lower $q_{\rm break}$ corresponds to a more bottom-heavy mass ratio distribution.} Combined with our constraints at wide separations, this implies that the transition between a uniform mass ratio distribution at close separations and a bottom-heavy distribution at wide separations occurs relatively abruptly at $s\sim 100$\,AU. Several other binary population properties are observed to change at $s\sim 100$\,AU (see \citealt{ElBadry_2019} and references therein), perhaps due to a transition in the dominant binary formation mechanism at this separation. This sharp transition, and the fact that $\gamma_{\rm largeq}$ is nearly constant over $300\lesssim s/{\rm AU} \lesssim 50,000$, provides a useful constraint for star formation models. Because the effects of dynamical processing after formation on the mass ratio distribution are imperfectly understood, similar constraints obtained in star-forming environments will prove useful for disentangling the primordial mass ratio distribution from the effects of dynamical processing.
\begin{figure*}
\includegraphics[width=\textwidth]{fit_parameters_vs_separation.pdf}
\caption{Constraints on $q_{\rm twin}$ (left; the mass ratio above which the twin excess begins), $\gamma_{\rm smallq}$ (middle; the logarithmic slope of the mass ratio distribution at small $q$), and $\gamma_s$ (right; the logarithmic slope of the local separation distribution). Dark and light error bars show 1 and 2 sigma constraints. Primary mass increases from top to bottom. Constraints on $q_{\rm twin}$ are only shown for mass and separation bins where $F_{\rm twin}$ is inconsistent with 0 at the 2 sigma level. Dashed line in the left panels shows $q=0.93$, the lower limit set on $q_{\rm twin}$ by the prior. Dashed lines in the right panels show $\gamma_{s}=-1.6$, the value found by \citetalias{ElBadry_2018} when averaging over all masses and separations. }
\label{fig:other_parameters}
\end{figure*}
We show constraints on other parameters obtained from our fitting in Figure~\ref{fig:other_parameters}. The left column shows $q_{\rm twin}$. This parameter is only meaningful if $F_{\rm twin}$ is nonzero, so we do not show constraints for bins of mass and separation where $F_{\rm twin}$ is consistent with 0 at the 2 sigma level. There are no strong trends in $q_{\rm twin}$ with mass or separation; the typical best-fit value is $q_{\rm twin}\approx 0.95$. In a few bins (e.g. $1.2 < M_1/M_{\odot} < 2.5$ and $350 < s/{\rm AU} < 600$), the constraint runs up against the prior, implying that a broader excess may be preferred. However, the width of the observed $\Delta G \approx 0$ excess is still reproduced reasonably well in these cases (Figure~\ref{fig:posterior_predictive}).
Constraints on $\gamma_{\rm smallq}$ are shown in the middle column. Most constraints are consistent with $\gamma_{\rm smallq} = 0$ (a flat distribution at small $q$), and trends with separation are weak. For the lowest-mass bin ($M_1 < 0.4 M_{\odot}$), we fit a single power law, so $\gamma_{\rm smallq} = \gamma_{\rm largeq}$. Unlike for $\gamma_{\rm largeq}$, there are no strong trends with primary mass. We note that our data cannot strongly constrain $\gamma_{\rm smallq}$ for low-mass primaries, so the constraints are influenced somewhat by the priors (Section~\ref{sec:functional_form}).
Finally, the right column of Figure~\ref{fig:other_parameters} shows constraints on the local slope of the separation distribution $\gamma_s$, which is defined such that $p(s)\propto s^{\gamma_s}$. The dashed line shows $\gamma_s = -1.6$, which is approximately the value that has been found for wide binaries when marginalizing over a larger range of primary masses and separations (\citealt{Andrews_2017}; \citetalias{ElBadry_2018}). The constraints we find here are similar to this value on average but show some evidence for a steepening in $p(s)$ with increasing separation. Any trends with primary mass at fixed separation are weak. We emphasize that these constraints come from the gradient in binary counts as a function of $s$ {\it measured within a narrow bin of $s$}, which is necessarily noisy: trends in $\gamma_s$ with separation represent the second derivative of $p(s)$.
\section{Discussion}
\label{sec:discussion}
\subsection{Comparison to previous work}
\label{sec:literature}
Although most work on twin binaries to date has focused on spectroscopic binaries, hints of a twin excess at wide separation have been reported in several previous works. \citet{Trimble_1987} found a significant excess of equal-brightness pairs among a sample of bright, nearby visual binaries. She suggested that this excess might point toward a formation mechanism that favors equal-mass systems but found significantly different distributions of magnitude difference when comparing different samples of wide binaries and thus did not rule out the possibility that the $q\sim 1$ peak was the result of selection effects. Similar conclusions were reached by \citet{Giannuzzi_1987}. \citet{Halbwachs_1988} argued that the mass ratio distribution of wide, common proper motion binaries was likely consistent with random pairings from the IMF once selection effects were corrected for.
Larger and more homogeneous samples of bright wide binaries were identified using astrometry from the Hipparcos satellite for one or both components \citep[e.g.][]{Soderhjelm_2000, Soderhjelm_2007, Eggenberger_2004, Lepine_2007, Shaya_2011}. The mass ratio distribution of Hipparcos binaries with A and F star primaries was modeled in detail by \citet{Soderhjelm_2007}, who found evidence for a $q\approx 1$ peak at $100 \lesssim s/\rm AU \lesssim 1000$. The strength of the peak decreased with primary mass. He argued that the twin excess was not the result of any known selection effect, although he did not reject the possibility that an unknown bias in the Hipparcos input catalog could explain it. The twin feature identified by \citet{Soderhjelm_2007} is likely the same feature apparent in our catalog. We note however, that most of the binaries observed by Hipparcos fall in our highest-primary mass bin (where the twin excess is weaker than at lower masses), as Hipparcos only observed bright stars ($G\lesssim 13$).
A preference for equal-brightness pairs was also noticed among wide common proper motion disk and halo binaries identified by \citet{Chaname_2004} and \citet{Dhital_2010}. The authors could not rule out the possibility that it was the result of selection effects, which were not well understood for their samples. The twin excess in our catalog is visually quite striking (e.g. Figure~\ref{fig:chance_align}), so a natural question is why it was not as as clear in earlier binary catalogs. In Appendix~\ref{sec:other_catalogs}, we show that a twin excess {\it is} apparent in the large, low-mass wide binary catalog produced by \citet{Dhital_2015} using SDSS photometry, but it only becomes obvious once objects with blended photometry are removed. It is also clearer in the {\it Gaia} photometry than in the ground-based SDSS photometry.
More broadly, the twin excess we identify is visually striking because it is narrow, but it represents only a few percent of the total twin population. This means that it will only become obvious when (a) the photometry is sufficiently precise and uncontaminated that the difference in the components' apparent magnitude can be measured with precision that is good compared to the intrinsic width of the twin excess, and (b) the sample considered is large, containing (at least) hundreds of objects. A twin excess among solar-type visual binaries is also observed in the \citet{Raghavan_2010} 25-pc and \citet{Tokovinin_2014} 67-pc samples, extending out to $s\approx 200$\,AU at a statistically significant level \citet{Moe_2017}. This excess is consistent with our constraints (e.g. Figure~\ref{fig:Ftwin_vs_separation}); at wider separations, these samples did not contain enough objects to detect or rule out a few-percent excess with high significance.
\subsubsection{Width of the twin excess}
\label{sec:qtwin}
Among spectroscopic binaries, there has been some disagreement in the literature over whether the twin excess is limited to a narrow peak in the mass ratio distribution at $q \gtrsim 0.95$ \citep[e.g.][]{Tokovinin_2000} or is a broader feature, corresponding simply to a positive slope in $p(q)$ at $q\gtrsim 0.8$ (e.g. \citealt{Halbwachs_2003}; see \citealt{Lucy_2006} for further discussion). In our sample, the twin feature is unambiguously narrow, only becoming significant above $q_{\rm twin}\approx 0.95$ (Figure~\ref{fig:other_parameters}). In Appendix~\ref{sec:how_sharp}), we show that $p(q)$ is always consistent with a flat or negative power law at $q< 0.94$; for the majority of the mass and separation bins, the twin excess only becomes strong at $q > 0.97$ (see Figure~\ref{fig:hist_model}). To allow better comparison between close and wide binaries, we now re-examine the twin excess among spectroscopic binaries.
\begin{figure}
\includegraphics[width=\columnwidth]{qtwin_tokovinin.pdf}
\caption{Cumulative distribution function of short-period double-lined spectroscopic binaries with $q > 0.76$ in the volume-limited 67-pc sample of FG dwarfs from \citet{Tokovinin_2014}. We compare predictions for models with three values of $q_{\rm twin}$ (see Section~\ref{sec:qtwin}). For this sample, we find $q_{\rm twin} = 0.964 \pm 0.013$, consistent with the values we find for wide binaries with periods $P=10^{5-9}$\,days.}
\label{fig:sb2s}
\end{figure}
In the volume-limited 67-pc sample of FG dwarfs, \citet{Tokovinin_2014} identified 98 double-lined spectroscopic binaries (SB2s) with periods $P = 1-100$\,days, 70 of which have dynamical mass ratios $0.76 < q < 1$. Nearly all binaries with $P < 100$\,days and $q > 0.76$ will appear as SB2s, so the 70 observed systems in this parameter space represent a relatively complete subsample. In Figure~\ref{fig:sb2s}, we plot the cumulative mass ratio distribution of the 70 short-period SB2s with $q > 0.76$, about half of which have $q > 0.95$. We model a uniform mass ratio distribution across $0.76 < q < 1$ with an excess twin fraction $F_{\rm twin}$ above $q > q_{\rm twin}$. We use the maximum-likelihood method described in \citet{Moe_2017} to fit the two free parameters $F_{\rm twin}$ and $q_{\rm twin}$ and draw 1,000 bootstrap samples to estimate their uncertainties. We show in Figure~\ref{fig:sb2s} the best-fit models obtained when fixing $q_{\rm twin}$ = 0.93, 0.95, and 0.97. The value $q_{\rm twin}$ = 0.93 is inconsistent with the data ($p = 0.014$), while $q_{\rm twin} = 0.95$ and 0.97 both provide reasonable fits. We formally measure $q_{\rm twin} = 0.964 \pm 0.013$ (1 sigma uncertainties).\footnote{We do not present constraints on $F_{\rm twin}$ for this sample, because $F_{\rm twin}$ depends on $p(q)$ at $0.3 < q < 1$, and many lower-mass ratio binaries will not be double-lined.} By using a larger and more complete sample of short-period SB2s, we thus confirm the conclusions of \citet{Tokovinin_2000} and \citet{Moe_2017} that close solar-type binaries with $a < 0.5$\,AU exhibit a large excess twin fraction and that the twins are narrowly distributed above $q_{\rm twin} \gtrsim 0.95$.
\begin{figure}
\includegraphics[width=\columnwidth]{Ftwin_qtwin_full_range.pdf}
\caption{$F_{\rm twin}$ (top) and $q_{\rm twin}$ (bottom) for solar-type binaries ($0.8 \lesssim M_1/M_{\odot} \lesssim 1.2$). We compare results from {\it Gaia} wide binaries (dark and light error bars show 1 and 2 sigma constraints) to 1 sigma constraints at closer separations from the samples of \citet{Raghavan_2010} and \citet[][see Figure~\ref{fig:sb2s}]{Tokovinin_2014}. $F_{\rm twin}$ decreases with separation and is consistent with 0 at $a\gtrsim 5,000$\,AU. However, $q_{\rm twin}\approx 0.95$ is basically constant at all separations.}
\label{fig:full_period}
\end{figure}
In Figure~\ref{fig:full_period}, we show constraints on $F_{\rm twin}$ and $q_{\rm twin}$ for solar-type binaries across a wide range of periods and separations. At wide separations, the constraints from {\it Gaia} wide binaries are reproduced from Figure~\ref{fig:Ftwin_vs_separation} and~\ref{fig:other_parameters}. At closer separations, we show the constraints on $F_{\rm twin}$ obtained by \citet{Moe_2017} from the \citet{Raghavan_2010} sample, as well as the constraint on $q_{\rm twin}$ at $P=1-100$\,days derived above for the \citet{Tokovinin_2014} sample (Figure~\ref{fig:sb2s}). At intermediate periods ($P=10^{2-6}$\,days), we show $q_{\rm twin}\approx 0.95\pm 0.02$ based on the \citet{Moe_2017} analysis of the \citet{Raghavan_2010} sample. This constraint is not the result of formal fitting but provided a good fit to the data (see Figure 30 of \citealt{Moe_2017}). Figure~\ref{fig:full_period} shows that while $F_{\rm twin}$ decreases with separation, $q_{\rm twin}\approx 0.95$ is consistent with being constant over all separations in this mass range. Similar values of $q_{\rm twin}$ are also found for massive binaries at the short periods where there is a significant twin excess \citep{Moe_2013}.
\subsection{Origin of twin binaries}
\label{sec:twin_origin}
It is typically assumed that the components of binaries wider than a few hundred AU formed nearly independently of one another \citep[e.g.][]{White_2001, Moe_2017, Tokovinin_2017b, Moe_2018, ElBadry_2019} during turbulent core fragmentation (for binaries with separation less than a few thousand AU; e.g. \citealt{Offner_2010}) or by becoming bound at slightly later times (for those with the widest separations; e.g. \citealt{Parker_2009}; \citealt{Moeckel_2010}; \citealt{Kouwenhoven_2010}; \citealt{Tokovinin_2017}).
The existence of a narrow twin excess at $q\gtrsim 0.95$ suggests that the components of a fraction of binaries with $s\gg 100$\,AU formed at closer separations in a highly correlated way. We do expect that the dynamical process of becoming and staying bound may lead to a preference for roughly equal-mass binaries (say, $q > 0.5$), because these have higher binding energy. This could quite reasonably explain, at least in part, why the power-law component of the mass ratio distribution is shallower than expected for random pairings from the IMF (Figure~\ref{fig:gamma_vs_separation}). But we do not expect dynamical processes to produce a {\it sharp} twin feature like what is observed: at fixed primary mass, the binding energy at $q=0.9$ is not much less than that at $q=1$. And indeed, simulations of binary formation during cluster dissolution find larger typical mass ratios at wide separations than predicted for random pairings, but they do not predict a narrow excess of twins \citep{Kroupa_1998, Moeckel_2010, Kouwenhoven_2010}.
The excess twin fraction uniformly decreases with separation and eventually goes to 0 at $s> 15,000$\,AU in all mass bins. The shape of the twin excess (i.e. $q_{\rm twin}$ and the slope of $p(q)$ at $q>q_{\rm twin}$) does not vary much between 50 and 15,000 AU in our catalog. Moreover, it is effectively the same for spectroscopic binaries (with separations as close as $0.01$\,AU) and wide binaries (Figure~\ref{fig:full_period}). Invoking Occam's razor, it seems more likely that the wide binary twin phenomenon is an extension of the phenomenon that has previously been observed at $s \lesssim 100$\,AU than that it is produced by a qualitatively different process.
Even for close binaries, there is not a clear consensus in the literature about the physical origin of the twin phenomenon. Some models for the formation of twins can only apply to very close binaries. In the first paper to highlight the twin phenomenon for spectroscopic binaries, \citet{Lucy_1979} suggested that twins were formed by fragmentation of rapidly rotating pre-main sequence stars during the late stages of dynamical collapse, at scales of $a \ll 1$\,AU. Alternatively, \citet{Krumholz_2007} proposed that twins could be produced by mass transfer between stars of initially different masses during pre-main sequence evolution.\footnote{These authors sought to explain the observed twin excess in massive stars. The specific mechanism they proposed, which relies on deuterium shell burning causing protostars to expand and overflow their Roche lobes, cannot operate in solar-type or lower mass stars. Moreover, the twin excess for massive stars appears to be limited to close separations \citep{Moe_2017}, so mass transfer may adequately explain it. Here we simply suppose, for the sake of argument, that there is some mechanism through which stable mass transfer in lower mass stars could drive the mass ratio to unity.} It seems implausible that such mechanisms can explain the twin phenomenon among wide binaries, because there is no known mechanism to widen the orbits of twins from the separations at which they operate -- a few, or at most a few tens of, solar radii -- to the separations at which they are observed today. Such widening would require a very strong velocity kick, the magnitude of which would have to be fine-tuned in order to not unbind the binaries completely.
A more plausible formation mechanism for equal-mass twins at wider separations is through competitive accretion from a circumbinary disk. Many studies have found that the accretion rate from a circumbinary disk is usually higher for the secondary than the primary \citep[e.g.][]{Bate_1997, Bate_2000, Farris_2014, Young_2015, Nelson_2016, Matsumoto_2019}. Because the secondary's orbit is larger than that of the primary, it sweeps out a larger radius in the disk and can accrete more rapidly than the primary, unless the material being accreted has low angular momentum.\footnote{It is worth noting that there has not been a full consensus in the literature whether this mechanism works: some simulations of accretion from a circumbinary disk have actually predicted the opposite trend, with accretion favoring the primary \citep{Ochi_2005, Hanawa_2010, de_val_borro_2011}. These simulations assumed a higher gas temperature than those which have found accretion to favor the secondary; \citet{Young_2015} showed that accretion only favors the secondary when accreted gas is cold, as is appropriate for stellar binaries.} Preferential accretion onto the secondary will necessarily drive the mass ratio towards unity. If such accretion continues for long enough, binaries within circumbinary disks should thus end up with $q\approx 1$. An appeal of this formation mechanism is that it can operate at scales comparable to the size of circumbinary disks, $s \lesssim 100$\,AU.
It is plausible but not obvious that preferential accretion onto the secondary will give rise to a sharp twin feature like the one found observationally. In order to end up at $q \approx 1$, a binary that initially had an intermediate mass ratio must accrete a large fraction of its mass from a circumbinary disk, such that there is enough time to drive the mass ratio to 1 even while the primary continues to accrete. If twins are formed by accretion from circumbinary disks, then the width of the twin feature (i.e., $q_{\rm twin}$) can tell us about the fraction of the total mass accreted from the disk, as well as the mass ratio above which the accretion rate onto the two components becomes nearly equal.
A sharp twin feature could be expected if a fraction of binaries accrete most of their mass from a circumbinary disk (becoming twins) and the rest either do not develop circumbinary disks or only accrete a subdominant fraction of their total mass from them.
In accretion-driven explanations of the twin excess, the increase in $F_{\rm twin}$ towards close separations has been interpreted as evidence that gravitational torques within circumbinary disks \citep[e.g.][]{Artymowicz_1991, Shi_2012} shrink the orbits of twins \citep{Young_2015}. Indeed, the observed twin excess is largest at separations $a < 0.2$\,AU (Figure~\ref{fig:full_period}), too close for binaries to have formed at their current separations. This implies that some combination of gravitational torques, viscous dissipation, and dynamical interactions \citep[e.g.][]{Bate_2012, Moe_2018} must have shrunk the orbits of twins at very close separations.
However, several recent simulations of circumbinary disks have found that, contrary to previous results in the literature, accretion can also {\it widen} binaries within circumbinary disks, when the advective torque dominates over the gravitational torque \citep{Miranda_2017, Munoz_2019, Moody_2019}. Whether gravitational or advective torques dominate depends on details such as the sink prescription used for accretion \citep{Tang_2017}. Further work is needed to clarify the effects of circumbinary disks on orbital evolution. However, the fact that the twin excess extends to very wide separations suggests that orbit shrinkage is not a ubiquitous outcome of accretion from circumbinary disks.
High-resolution studies of the dynamics of individual circumbinary disks are generally too idealized, run for too short a time, and are focused on too narrow a range of initial conditions to make ab-initio predictions of the full mass ratio distribution. However, they do typically find that gas is preferentially accreted onto the secondary for realistic accretion geometries once a steady state is reached. On the other hand, global simulations of the fragmentation of molecular clouds \citep[e.g.][]{Bate_2009, Bate_2014, Bate_2019} are reaching the point where they can make realistic predictions of the mass ratio distribution with minimal fine-tuning. Such studies have lower resolution than idealized simulations of individual binaries, so it is not guaranteed that the dynamics within disks are well resolved, but they are able to predict the accretion rate and angular momentum distribution of accreted material, and the mass and size distribution of disks \citep[e.g.][]{Bate_2018}. Such global calculations predict an excess of equal-mass binaries out to separations of order 100 AU (see e.g. \citealt{Bate_2014}, Figure 18). Because they typically only form a few dozen binaries in a cloud, such calculations do not yet have the statistical power to predict or rule out a few-percent effect at wider separations.
Accretion can plausibly explain an excess of twins out to significantly larger separations than mass transfer or late-stage fragmentation, but it alone cannot explain a signal reaching out to thousands of AU. Observed circumstellar and circumbinary gas disks have typical radii of order $100$\,AU \citep{Ansdell_2018, Eisner_2018}.
The {\it largest} observed circumbinary disks have radii of order 500 AU and host binaries with separations of 50-200 AU \citep[e.g.][]{Hioki_2007, Brinch_2016, Tobin_2016, Takakuwa_2017, Comeron_2018, Czekala_2019}; these preferentially host relatively massive binaries. There are no observed disks with radii exceeding 1000 AU, and simulations also predict the largest circumbinary disks to have radii of several hundred AU \citep{Bate_2018}. It thus seems exceedingly unlikely that twin binaries with $s\gtrsim 200$\,AU formed at their present-day separation by accretion from a circumbinary disk. This implies that either (a) twin binaries formed at closer separations and their orbits were subsequently widened, or (b) some other process is responsible for producing twins at very wide separations. Lacking a good candidate mechanism for (b), we here consider the plausibility of orbit widening.
\subsubsection{Dynamical orbit widening in young clusters}
\label{sec:widening}
In the Galactic field, dynamical interactions have a negligible effect on most binaries with $s\lesssim 10,000$ AU \citep[e.g.][]{Weinberg_1987}. However, dynamical interactions are expected to be more efficient in binaries' birth environments, where the typical stellar density is higher. The dynamical evolution of binaries within their birth clusters has been the subject of considerable study. As a general rule, interactions within birth clusters are expected to widen the orbits of binaries with orbital velocities lower than the cluster velocity dispersion, and to tighten the orbits of those with orbital velocities greater than it \citep{Heggie_1975, Hills_1975}. There are several complicating factors in real clusters. For example, the mass distribution within clusters at early times is not smooth but clumpy \citep{Dorval_2017}, both stars and gas can be dynamically important, and cluster density and velocity dispersion change as clusters age, in part because energy is redistributed among binaries \citep[see e.g.][]{Kroupa_2001, Parker_2009, Goodwin_2010}. Observed binary populations provide constraints on models for binaries' dynamical evolution, but a unified model to explain the diversity of binary populations found in different environments does not exist. We summarize some relevant observational constraints on disruption below.
\begin{enumerate}
\item Over a wide range of separations ($10 \lesssim s/{\rm AU} < 3000$), the binary fraction in low-density star forming regions (e.g. Taurus) is higher than in the field by roughly a factor of 2 \citep[e.g.][]{Leinert_1993, Duchene_1999}. This is true especially at wider separations, where the separation distribution of young binaries is roughly log-uniform, but that of field stars declines more steeply \citep[e.g.][]{Connelley_2008, Kraus_2011}.
\item In dense young clusters (e.g. the ONC), the binary fraction at separations of $s\gtrsim 100$\,AU is similar to the field (lower than in low-density star forming regions) and declines steeply at wide separations \citep{Reipurth_2007}. At closer separations ($10 \lesssim s/{\rm AU} \lesssim 60$) the binary fraction in dense clusters is comparable to that in low-density star forming regions, and higher than that found in the field \citep{Duchene_2018}.
\item At very close separations ($s\lesssim 5\,\rm AU$), the binary fraction in star forming regions (over a range of densities) is consistent with that in the field \citep[e.g.][]{Kounkel_2019}.
\end{enumerate}
Some models \citep[e.g.][]{Kroupa_1995, Marks_2011, Marks_2012} postulate that the initial binary fraction and separation distribution are insensitive to local properties, such that observed variation in binary populations must be due to disruption. These models interpret the higher wide binary fraction in low-density clusters as the primordial population, which is transformed into the field population by dynamical widening and disruption. Because disruption of wide systems is more rapid in dense clusters, these models also predict the binary fraction at $s \gtrsim 100$\,AU to decrease with cluster density, in agreement with observations. However, if such models are correct, it is not clear what happens to the excess of relatively tight binaries ($10 \lesssim s/{\rm AU} \lesssim 100$), which are over-represented relative to the field in both high- and low-density star forming regions. Such systems have high enough binding energies that they can only be disrupted in very dense clusters. Thus the lower binary fraction in the field at $10 \lesssim s/{\rm AU} \lesssim 100$ would imply that a large fraction of field stars formed in dense environments.
In any case, the fact that the binary fraction in the field is lower than that in star forming regions down to fairly close separations implies that a significant fraction of young binaries undergo quite energetic interactions, the cumulative effect of which is sufficient to disrupt binaries with initial separations as close as 10-100\,AU. In many cases, binaries will be disrupted by such interactions, but in some cases, they will only be widened \citep[e.g.][]{Kroupa_2001}. It is in this latter context that the twin fraction at wide separations is informative about the fraction of stars at a given present-day separation that formed at significantly closer separations.
The fact that the dependence of $F_{\rm twin}$ on separation varies with binary mass in a non-monotonic way (Figure~\ref{fig:Ftwin_vs_separation}) suggests that, if the twin excess at $s\gg 100$\,AU is due to dynamical widening of orbits, then the primordial twin statistics (e.g. the twin fraction and the range of separations over which twins are form) must also vary with mass. This may not be unreasonable, since the physical properties of disks do vary with mass \citep{Bate_2018, Eisner_2018}, but it means that disentangling the effects of dynamical widening and the primordial separation distribution of twins is nontrivial.
\begin{figure}
\includegraphics[width=\columnwidth]{schematic_Ftwin_vs_separation.pdf}
\caption{Predictions of the toy model for orbit widening described in Section~\ref{sec:widening}. We assume that the initial twin fraction is 20\% for $a<a_{\rm twin,\,max}$ (varied in the top panel; 100\,AU by default) and 0 at larger separations. We consider an initial separation distribution that is flat in log space. We then widen the orbits of all binaries by adding energy, representing the cumulative effect of gravitational interactions within a birth cluster. The total added energy is drawn from a uniform distribution between 0 and $E_{{\rm max,interaction}}$. Binaries that are disrupted are discarded. We plot the final twin fraction as a function of separation, varying $E_{{\rm max,interaction}}$ (middle panel; default is $G\left(0.5M_{\odot}\right)^{2}/\left(100\,{\rm AU}\right)$) and mass (bottom panel; default is 0.5\,$M_{\odot}$).}
\label{fig:toy_model_widening}
\end{figure}
We use a simple toy model to explore how the separation-dependence of the twin excess may depend on the initial twin population and the strength of dynamical interactions. We suppose that twin formation is efficient out to a separation of $a_{\rm twin,\,max}\approx 100$\,AU and does not operate at wider initial separations. We consider primordial populations of equal-mass binaries with components of mass $M$ and a uniform distribution of $\log a$ between 10 and $10^4$\,AU. At $a < a_{\rm twin,\,max}$, we assign a random subset of 20\% of the binaries the label of ``twin''. We then assume that dynamical interactions add an energy $E_{\rm int}$ to each orbit (for both twins and non-twins), such that the new orbital energy is $E_{{\rm new}}=E_{{\rm initial}}+E_{\rm int}$, where $E_{{\rm initial}}=-GM^{2}/a_{{\rm initial}}$. Values of $E_{\rm int}$ for each binary are drawn from a uniform distribution between 0 and $E_{\rm max,\,interactions}$. If $E_{\rm new}$ is positive, then the binary is considered unbound and is discarded. The new separation after interactions is $a_{{\rm new}}=-GM^{2}/E_{{\rm new}}$. Finally, we measure what fraction of the surviving binaries bear the twin label as a function of final separation.
This fraction is proportional to the fraction of binaries at a given present-day separation that formed with $a < a_{\rm twin,\,max}$. We do not model dynamical hardening of close binaries, and thus implicitly assume that all binaries we consider are soft; i.e., that they reside in clusters with velocity dispersion larger than than the highest orbital velocities of binaries being widened. For an initial separation of 100\,AU and binaries with typical component mass of 0.5\,$M_{\odot}$, this corresponds to a dispersion of $\sigma \approx 1.5\,\rm km\,s^{-1}$; for an initial separation of 10\,AU, to $\sigma \approx 5\,\rm km\,s^{-1}$.
We plot the results of this experiment in Figure~\ref{fig:toy_model_widening}, varying $a_{\rm twin,\,max}$ (top), $E_{\rm max,\,interactions}$ (middle), and $M$ (bottom). The final separation-dependence of the twin fraction depends significantly on all of these parameters. The twin excess extends to larger separations, and falls off less steeply with separation, when (a) the initial separation out to which twins form is larger, (b) dynamical interactions are more energetic, or (c) the binding energy of twins is lower. In order to obtain a nonzero twin fraction at very wide separations, it is necessary that a fraction of binaries undergo dynamical interactions energetic enough to unbind binaries at separation at which twins form. Once this is satisfied, twins can contribute significantly to the binary population at wider separations, because the same interactions that widen close binaries will unbind a large fraction of initially wider non-twin binaries. Wide twins are produced most efficiently in clusters where the velocity dispersion is comparable to the orbital velocity at a separation of $a_{\rm twin,\,max}$: significantly denser clusters produce few wide binaries, since most binaries that are not close are disrupted completely.
The toy model generically predicts that $F_{\rm twin}$ decreases with separation and that low-mass twins can be more efficiently widened than high-mass twins due to their lower binding energy. It thus predicts that the excess twin fraction will fall off less steeply for lower-mass binaries. This trend {\it is} found in the observed binaries over $0.4 < M_1/M_{\odot} < 2.5$ (Figure~\ref{fig:Ftwin_vs_separation}), but it is reversed for the lowest-mass subsample: $F_{\rm twin}$ falls off significantly more steeply for $0.1 < M_1/M_{\odot} < 0.4$ than for $0.4 < M_1/M_{\odot} < 0.6$. This is not easily explained in the context of the toy model.
It is thought that a fraction of wide binaries (likely at $a \gg 1000$\,AU) form at later times (perhaps during cluster dissolution) than wide binaries at closer separations. Dilution of the twin excess due to these binaries is not accounted for in the toy model. If the fraction of binaries at fixed separation that formed during cluster dissolution were higher at lower masses, this could explain the observed steeper decline in $F_{\rm twin}$ with separation at low masses.
An addition complication is that the observed population of field binaries is an average over a wide range of formation environments, from low-density regions to dense clusters. The trends in Figure~\ref{fig:toy_model_widening} rely on a number of crude approximations and should not be used to directly interpret the observed trends in $F_{\rm twin}$ (and we have not attempted to tune the model to match observed trends). Here we simply emphasize that the separation distribution of twins is quite sensitive to dynamical processing. We conclude that dynamical orbit widening provides a plausible explanation for the existence of the wide twin excess, but more realistic theoretical modeling is needed to determine whether the observed trends in $F_{\rm twin}$ with primary mass and separation can be reproduced when averaging over a realistic population of star-forming environments.
\subsubsection{Widening of unstable triple systems?}
\label{sec:triples}
Another possible mechanism for dynamically widening twins is through unfolding of unstable triple systems, in which initially close companions can be scattered to much wider orbits \citep[e.g.][]{Reipurth_2012}. Triples are not uncommon: more than a third of all wide binaries contain subsystems \citep[e.g.][]{Tokovinin_2002, Tokovinin_2014}, and a large fraction of binaries are thought to have formed as higher-order multiples that subsequently decayed \citep[e.g.][]{Sterzik_1998}.
\begin{figure}
\includegraphics[width=\columnwidth]{dG_triples.pdf}
\caption{Magnitude difference between the two resolved components of all wide binaries in which one or both components is suspected to have a bright, unresolved companion ($N=7,295$; yellow points in Figure~\ref{fig:cmds}). There is no obvious excess at $\Delta G\approx 0$ (compare to the right panel of Figure~\ref{fig:dG_points}). }
\label{fig:dG_triples}
\end{figure}
Widening of the outer orbit in triples generally comes at the expense of shrinking of the inner orbit. Since the outer orbit is the more fragile one, a triple-origin of the wide twin excess would imply that one of the component stars in wide twins had, and in most cases still has, a close companion, which in most cases would be spatially unresolved. However, the twin excess at $\Delta G \lesssim 0.25$\,mag cannot be due to systems in which either resolved component has an unresolved companion any less than $\approx 2.5$ magnitudes fainter than it; otherwise the light from the unresolved companion would make it more than 0.25 mag brighter than the other resolved component. Figure~\ref{fig:dG_triples} shows that there is indeed no obvious excess at $\Delta G \approx 0$ for wide binaries in which either component has a bright unresolved companion. There is thus a photometric upper limit on the mass ratio of any unresolved companions to twins with $\Delta G \approx 0$; it ranges from $q < 0.45$ to $q < 0.7$, depending on primary mass (see Figure~\ref{fig:delta_G_vs_q}).
Unresolved companions with lower mass ratios cannot be ruled out based on photometry, but at close separations, they can be detected from radial velocity (RV) variability. In Appendix~\ref{sec:rv_variability}, we show that the {\it Gaia} visit-to-visit radial velocity scatter implies a {\it lower} close binary fraction for the components of wide twins than for components of non-twins with similar separations and masses. This speaks against a triple-origin of wide twins. It is, however, consistent with our expectations if twins are formed through accretion in a circumbinary disk, since the presence of a third close companion in the disk would complicate the mechanism through which accretion drives the mass ratio of two stars in a disk toward unity.
\section{Summary and conclusions}
\label{sec:conclusions}
We have analyzed a pure and homogeneous sample of $\sim$42,000 main-sequence wide binaries selected from {\it Gaia} DR2 to constrain the mass ratio distribution of binaries with projected separations $50\lesssim s/{\rm AU} < 50,000$ and primary masses $0.1 < M_1/M_{\odot} < 2.5$. High-precision photometry allows us to measure mass ratios with unprecedented accuracy, and a well-understood selection function makes it possible to account for biases arising from the magnitude and contrast sensitivity limits of {\it Gaia} DR2 given our quality cuts.
A striking feature of the catalog is a high-significance excess of ``twin'' binaries with nearly equal brightness (Figure~\ref{fig:dG_points}). The excess is present over a narrow range of magnitude differences, $0 < \Delta G \lesssim 0.25$, corresponding to mass ratios $0.95 \lesssim q \lesssim 1$, and extends over a wide range of masses and separations. The twin excess is reminiscent of the excess of equal-mass binaries historically reported at very close separations ($a < 0.2$\,AU) and recently found to extend to $a\sim 100$\,AU, but it extends to separations of several thousand AU, where binary formation models do not predict strongly correlated component masses.
We have done a variety of tests to confirm that the twin feature is caused by an astrophysical excess of equal-mass binaries, not selection effects or data artifacts. We first repeated the search that produced the binary catalog, but required that the two components of a pair have {\it inconsistent} rather than consistent astrometry. This yields a catalog of physically unassociated ``chance alignments'' that are subject to essentially the same selection function as the real binaries and have similar distributions of most observable properties. A narrow excess of equal-brightness pairs is not found among chance alignments (Figure~\ref{fig:chance_align}).
By considering binaries at different distances, we verified that the strength of the twin feature varies primarily with physical, not angular, separation (Figure~\ref{fig:three_distances}). This speaks to the physicality of the twin feature, since most observational biases are expected to depend on angular separation, which is the observable quantity. Finally, we verified that the twin feature is not an artifact of the {\it Gaia} photometry; it is visible in photometry from other surveys. We also identified the same twin excess in archival data from another binary catalog after removing objects with contaminated photometry (Appendix~\ref{sec:other_catalogs}).
In order to measure the intrinsic mass ratio distribution, we forward-model the joint distribution of magnitude difference, angular separation, and parallax given a Galactic model, a selection function, and a parameterized mass ratio distribution. We measure the selection function for our catalog empirically, using chance alignments to quantify the contrast sensitivity as a function of angular separation. Our primary results are as follows:
\begin{enumerate}
\item {\it Twin fraction}: we quantify the twin excess as $F_{\rm twin}$, the fractional excess of binaries with $q > q_{\rm twin}$ relative to the full population with $q>0.3$ (Figure~\ref{fig:schematic_q_dist}; $q_{\rm twin}\approx 0.95$ quantifies the width of the excess). Typical values of $F_{\rm twin}$ are 10\% at 100\,AU and 3\% at 1000\,AU, with some dependence on mass (Figure~\ref{fig:Ftwin_vs_separation}). These values are lower than the value found for spectroscopic binaries (Figure~\ref{fig:full_period}), which is $F_{\rm twin}\approx 25\%$, but they are clearly inconsistent with 0.
\item {\it Width of the twin excess}: at all masses and separations where it is statistically significant, the twin excess we find is quite narrow, only becoming significant at $q \gtrsim 0.95$ (Figure~\ref{fig:other_parameters}). We experimented with different functional forms for the enhancement near $q\approx 1$, but we find a step function increase at $q \gtrsim 0.95$ to perform as well as more complicated models (Figure~\ref{fig:hist_model}). We also re-measured $q_{\rm twin}$ at close separations using a volume-limited sample of spectroscopic binaries (Figure~\ref{fig:sb2s}). The width of the twin excess, $q_{\rm twin}$, is basically constant over 6 orders of magnitude in separation, from the closest spectroscopic binaries to wide binaries at $s\sim 10,000$\,AU.
\item {\it Mass and separation dependence}: the fractional excess of twins declines with separation and is consistent with 0 at $s > 15,000$\,AU for all mass bins (Figure~\ref{fig:Ftwin_vs_separation}). The twin excess at $s>1,000$\,AU is strongest for primary masses $0.4 < M_1/M_{\odot} < 0.6$. At closer separations ($50 < s/{\rm AU} < 1,000$), it is strongest for low-mass primaries ($M_1 < 0.4 M_{\odot}$) and roughly consistent for other mass bins. The slope of the fall-off in $F_{\rm twin}$ with separation varies non-monotonically with mass; it is shallowest at $0.4 < M_1/M_{\odot} < 0.6$.
\item {\it Mass ratio distribution at lower q:} we provide broken power law fits to the full mass ratio distribution for all mass and separation bins (Figure~\ref{fig:qdist_ranges}; Appendix~\ref{sec:full_constraints}). These fits reproduce the observed data well (Figure~\ref{fig:posterior_predictive}). For solar-type stars, the power law slope is $\gamma_{\rm largeq}\approx -1.3$ for $q > 0.5$ (i.e., weighted toward lower mass ratios than the uniform distribution found at closer separations) and $\gamma_{\rm smallq}\approx 0$ for $q < 0.5$. $p(q)$ becomes more bottom-heavy with increasing primary mass but is always flatter than expected for random pairings from the IMF. Besides variation in the excess twin fraction, $p(q)$ does not vary much with separation over $100 \lesssim s/{\rm AU} < 50,000$ (Figures~\ref{fig:gamma_vs_separation} and~\ref{fig:other_parameters}).
\item {\it Origin of the twin excess}: no theoretical models that have been proposed to explain twin binaries at close separations predict them to form at $s \gtrsim 100$\,AU. Dynamical processes may lead to a formation bias against low-$q$ binaries, but they are not expected to produce a sharp peak at $q \approx 1$. Given the monotonic fall-off in $F_{\rm twin}$ with separation, the similar width of the twin feature between close and wide binaries, and the lack of a plausible mechanism for forming twins at very wide separations, we conjecture that the excess twins must have formed at closer separations ($s \lesssim 100$\,AU; likely through accretion from a circumbinary disk) and subsequently been widened by dynamical interactions.
In this scenario, the separation-dependence of the twin fraction is an imprint of dynamical orbit widening in binaries' birth environments (see Section~\ref{sec:discussion}). The plausibility of this explanation is not straightforward to assess because (a) present-day field binaries formed in a wide range of environments and (b) existing models for widening and disruption of binaries in star forming regions do not fully explain the diversity of observed binary populations in young clusters and in the field. A simple toy model suggests that a separation-dependence in $F_{\rm twin}$ comparable to that which is observed is predicted if typical dynamical interactions are strong enough to disrupt binaries at the separation inside which twins are expected to form ($a \lesssim 100$\,AU). However, the mass-dependence of $F_{\rm twin}$ at wide separations is not fully explained in such models. Further theoretical work is required to (a) predict the primordial separation distribution of twins at different masses and (b) constrain the efficiency of dynamical orbit widening for a realistic ensemble of star forming environments. In future work, we will search for a wide twin excess in observed star forming regions to shed more light on the primordial separation distribution of twins and the density-dependence of orbit widening.
\end{enumerate}
\section*{Acknowledgements}
We are grateful to the referee, Andrei Tokovinin, for thoughtful comments.
We thank Matthew Bate, Anthony Brown, Eugene Chiang, Ian Czekala, Paul Duffel, Morgan Fouesneau, Harshil Kamdar, Tomoaki Matsumoto, Chris Mckee, Eliot Quataert, and Dan Weisz for helpful discussions. KE was supported in part by an NSF graduate research fellowship and by SFB 881. HJT acknowledges the National Natural Science Foundation of China (NSFC) under grants 11873034. MM acknowledges financial support from NASA under Grant No. ATP-170070. This project was developed in part at the 2019 Santa Barbara Gaia Sprint, hosted by the Kavli Institute for Theoretical Physics at the University of California, Santa Barbara. This research was supported in part at KITP by the Heising-Simons Foundation and the National Science Foundation under Grant No. NSF PHY-1748958. This work has made use of data from the European Space Agency (ESA) mission {\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia} Data Processing and Analysis Consortium (DPAC, \url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the {\it Gaia} Multilateral Agreement.
\bibliographystyle{mnras}
|
\section{Recapitulation of the mathematical framework}\label{sectionmathframework}
In this section, we briefly review the model for ligand binding based on the grand canonical ensemble of statistical mechanics and recall the notion of minimal absolute interaction from \cite{martini2017measure}. We refer the reader to \cite{ben2013cooperativity,wyman1990binding,hill2013cooperativity} for detailed exposition on the model.
\begin{definition}[Molecule]
A \emph{molecule} ${\rm W}$ with $n$ sites is a positive real point, whose coordinates are indexed by the subsets of $[n]:=\{1,\ldots,n\}$ and with $w_\emptyset=1${\rm :}
\[ {\rm W} = (w_I)_{I\subseteq [n]} \in (\mathbb \RR_{>0})^{2^n}. \]
We refer to $w_I$ as \emph{binding energies} if $|I|=1$ and \emph{interaction energies} if $|I|>1$.
For the sake of brevity, we abbreviate $w_{\{i_1,\ldots,i_r\}}$ to $w_{i_1\ldots i_r}$ for $i_1<\dots<i_r$ (see Fig.~\ref{fig:moleculeFourSites}).
\end{definition}
The binding energy $w_i$ encodes the difference between free energy of the
molecule in the completely unoccupied state and the state with only a ligand
bound to site $i$. The interaction energy $w_{i_1 i_2}$ is the
difference between free energy when sites $i_1,i_2$ are occupied and the sum of
the differences of energy encoded by $w_{i_1}$ and $w_{i_2}$. Analogously, $w_{i_1 \dots i_r}$ encodes the discrepancy between the energy required for $r$ ligands to bind to the sites $i_1,\dots,i_r$ and the energy already defined by the interactions of lower degree.
Note that we may call $w_{i_1 \dots i_r}$ energies even though they actually represent exponentials of the physical quantity ``energy'' \cite{ben2013cooperativity,wyman1990binding,schellman1975macromolecular}.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=1,every node/.style={font=\footnotesize}]
\pgfmathsetseed{13371338}
\coordinate (o) at (0,0);
\draw[rounded corners=.5mm] (o) \irregularcircle{1.75cm}{1mm};
\draw (25:1.75cm) node[draw,fill=red!30] (s1) {1};
\draw (105:1.75cm) node[draw,fill=red!30] (s2) {2};
\draw (195:1.75cm) node[draw,fill=red!30] (s3) {3};
\draw (285:1.75cm) node[draw,fill=red!30] (s4) {4};
\node[anchor=east,xshift=-0.5cm] (aText) at (s3) {binding energy $w_3$};
\draw[->,densely dotted] ($(aText.east)+(-0.15,0)$) -- (s3);
\draw[<->,densely dotted] (s1) to[out=60,in=60] node[anchor=south west,xshift=0cm,yshift=-2mm] {interaction energy $w_{12}$} (s2);
\path[fill=blue!20] (s3.east) to[bend left=15] (s4.north) to[bend left=15] (s1.south west) to[bend left=15] (s3.east);
\node[anchor=south,xshift=-4mm,yshift=8mm] at (s4) {$w_{134}$};
\end{tikzpicture}
\caption{A molecule with 4 binding sites.}
\label{fig:moleculeFourSites}
\end{figure}
\begin{definition}[Binding Polynomial]
Given a molecule ${\rm W}=(w_I)_{I\subseteq [n]}$ with $n$ sites, its {binding polynomial} $\Phi({\rm W})$ is a univariate polynomial of degree $n$,
\begin{equation*}\label{bindingpol}
\Phi({\rm W}):=a_n \Lambda^n + \dots + a_1 \Lambda + a_0,
\end{equation*}
whose positive real coefficients $a_k$ are given by
\begin{equation*}
\label{eq:bindingPolynomialCoefficients}
a_k := \sum_{|I|=k} \prod_{I'\subseteq I} w_{I'} \in \RR_{>0}.
\end{equation*}
\end{definition}
\begin{example}\label{ex:background}\rm
Let ${\rm W}=(w_I)_{I\subseteq[3]}\in (\RR_{>0})^{2^3}$ be a molecule with $3$ sites. The binding polynomial $\Phi({\rm W})=a_3\Lambda^3+a_2\Lambda^2+a_1\Lambda+a_0$ is a real univariate polynomial of degree~$3$ whose coefficients are given by
\begin{align*}
a_0 &= w_\emptyset = 1,\\
a_1 &= w_1+w_2+w_3,\\
a_2 &= w_1w_2w_{12}+w_1w_3w_{13}+w_2w_3w_{23},\\
a_3 &= w_1w_2w_3w_{12}w_{13}w_{23}w_{123}.
\end{align*}
Note that two different molecules may have the same binding polynomial and thus the same binding curve. Hence, the map $\Phi$ which maps a molecule with $n$ sites to its binding polynomial is not injective.
\end{example}
\begin{definition}[Absolute interaction]\label{def:norm}
The \emph{absolute interaction} of a molecule ${\rm W}=(w_I)\in (\RR_{>0})^{2^n}$ is given by
\begin{equation*}
\| {\rm W} \| := \prod_{|I|>1} \max\Big(w_I,w_I^{-1}\Big).
\end{equation*}
\end{definition}
Since $w_I$ represents the exponential of a difference of actual energies, which can be positive or negative, $\max(w_I,w_I^{-1})$ represents the exponential of the absolute value of that difference. Hence, the absolute interaction $\|{\rm W}\|$ represents the exponential of the sum of the absolute values of all energy differences.
The minimal value $\| {\rm W} \| $ can adopt is $1$ which corresponds to a minimal physical interaction energy of $0$.
\begin{definition}[Minimal absolute interaction]
Given a binding polynomial $P$ of degree $n$, the \emph{minimal absolute interaction} is
\begin{equation*}
\|P\| := \min \Big\{ \| {\rm W} \| \left.\vphantom{\Big\{} \suchthat \vphantom{\Big\}}\right. {\rm W}\in (\RR_{>0})^{2^n} \text{ with } \Phi({\rm W})=P \Big\}.
\end{equation*}
We call a molecule $W$ \emph{minimal}, if $\|W\| = \|\Phi(W)\|$.
\end{definition}
In other words, the minimal absolute interaction $\|P\|$ is the minimum of all absolute interactions of molecules with a fixed binding polynomial $P$, or equivalently with a fixed binding curve.
This notion is well defined by the following proposition.
\begin{proposition}[{\cite[\S4]{martini2017measure}}]
For any univariate polynomial $P$ of degree $n$ with positive coefficients there exists a molecule ${\rm W}\in (\RR_{>0})^{2^n}$ with $n$ sites such that $\Phi({\rm W})=P$ and $\|{\rm W}\|=\|P\|$.
\end{proposition}
Since cooperativity is an emerging property of the interaction between the sites and deeply connected to the binding curve, the minimal absolute interaction of a binding polynomial is a natural candidate for quantifying cooperativity. In addition, \cite[\S4]{martini2017measure} shows that it has several properties that could be considered desirable in such a quantifier.
\begin{example}\label{example3site}\rm
Consider the following binding polynomials:
\begin{align*}
P_1:=&4\Lambda^3+3\Lambda^2+2\Lambda+1,\\
P_2:=&6\Lambda^3+7\Lambda^2+4\Lambda+1=(2\Lambda+1)(3\Lambda^2+2\Lambda+1),\\
P_3:=&6\Lambda^3+11\Lambda^2+6\Lambda+1=(\Lambda+1)(2\Lambda+1)(3\Lambda+1).
\end{align*}
A brief computation reveals that the minimal absolute interactions are
\[ \|P_1\|=13.50,\qquad \|P_2\|=3, \qquad \|P_3\|=1. \]
The computation for $\|P_1\|$ is explicitly shown in Example~\ref{example3dim}. Figure~\ref{fig:minimalMolecules} illustrates the minimal molecules for each binding polynomial. Since $P_3$ factorizes into three real linear factors, its minimal molecule has only trivial interactions.
\end{example}
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.9]
\pgfmathsetseed{42424242}
\coordinate (o) at (0,0);
\draw[blue!50,rounded corners=.5mm] (o) \irregularcircle{1.75cm}{1mm};
\draw (0:1.75cm) node[draw,fill=white,font=\scriptsize] (s1) {$0.667$};
\draw (120:1.75cm) node[draw,fill=white,font=\scriptsize] (s2) {$0.667$};
\draw (240:1.75cm) node[draw,fill=white,font=\scriptsize] (s3) {$0.667$};
\node[fill=white,font=\scriptsize] (w12) at (60:1.75cm) {$1.00$};
\node[fill=white,font=\scriptsize] (w23) at (180:1.75cm) {$4.75$};
\node[fill=white,font=\scriptsize] (w13) at (300:1.75cm) {$1.00$};
\path[fill=blue!20] (s3.north) to[bend left=15] (s1.west) to[bend left=15] (s2.south) to[bend left=15] (s3.north);
\node[font=\scriptsize] (w123) {$2.83$};
\node[anchor=north] at (0,-2) {$\|W_1\|=13.50$};
\end{tikzpicture}\hfill
\begin{tikzpicture}[scale=0.9]
\pgfmathsetseed{42424242}
\coordinate (o) at (0,0);
\draw[blue!50,rounded corners=.5mm] (o) \irregularcircle{1.75cm}{1mm};
\draw (0:1.75cm) node[draw,fill=white,font=\scriptsize] (s1) {$2.00$};
\draw (120:1.75cm) node[draw,fill=white,font=\scriptsize] (s2) {$1.00$};
\draw (240:1.75cm) node[draw,fill=white,font=\scriptsize] (s3) {$1.00$};
\node[fill=white,font=\scriptsize] (w12) at (60:1.75cm) {$1.00$};
\node[fill=white,font=\scriptsize] (w23) at (180:1.75cm) {$3.00$};
\node[fill=white,font=\scriptsize] (w13) at (300:1.75cm) {$1.00$};
\path[fill=blue!20] (s3.north) to[bend left=15] (s1.west) to[bend left=15] (s2.south) to[bend left=15] (s3.north);
\node[font=\scriptsize] (w123) {$1.00$};
\node[anchor=north] at (0,-2) {$\|W_2\|=3.00$};
\end{tikzpicture}\hfill
\begin{tikzpicture}[scale=0.9]
\pgfmathsetseed{42424242}
\coordinate (o) at (0,0);
\draw[blue!50,rounded corners=.5mm] (o) \irregularcircle{1.75cm}{1mm};
\draw (0:1.75cm) node[draw,fill=white,font=\scriptsize] (s1) {$1.00$};
\draw (120:1.75cm) node[draw,fill=white,font=\scriptsize] (s2) {$2.00$};
\draw (240:1.75cm) node[draw,fill=white,font=\scriptsize] (s3) {$3.00$};
\node[fill=white,font=\scriptsize] (w12) at (60:1.75cm) {$1.00$};
\node[fill=white,font=\scriptsize] (w23) at (180:1.75cm) {$1.00$};
\node[fill=white,font=\scriptsize] (w13) at (300:1.75cm) {$1.00$};
\path[fill=blue!20] (s3.north) to[bend left=15] (s1.west) to[bend left=15] (s2.south) to[bend left=15] (s3.north);
\node[font=\scriptsize] (w123) {$1.00$};
\node[anchor=north] at (0,-2) {$\|W_3\|=1.00$};
\end{tikzpicture} \caption{Minimal molecules for binding polynomial $P_1$, $P_2$, $P_3$.}
\label{fig:minimalMolecules}
\end{figure}
\section{The Algebraic Optimization Problem}\label{sectionalgopt}
In this section, we consider the computation of the minimal absolute interaction as an optimization problem with the absolute interaction as objective function and the set of all molecules sharing the same binding polynomial as feasible set:
\begin{equation}\label{orgopt}
\begin{array}{lll}
\displaystyle\mini_{{\rm W}} & \quad & \| {\rm W} \|\\[3mm]
\st & & \Phi({\rm W})=P
\end{array}
\end{equation}
This problem seems simple and its concept can be understood easily. However, it
becomes quickly complicated if larger systems are considered. For small $n$ it
can be worked out by methods of algebraic optimization (e.g. using sums of squares
approaches \cite{Lasserre01, Parrilo03} or critical point methods \cite{SG14}).
For $n=4$ explicitly,
given the binding polynomial $P=a_4\Lambda^4+a_3\Lambda^3+a_2\Lambda^2+a_1\Lambda$ we have:
\begin{equation*}
\begin{array}{lcl}
\displaystyle\mini_{{\rm W}}
&& \displaystyle\prod_{|I|>1} \max\Big(w_I,w_I^{-1}\Big)\\[3mm]
\st
&& a_1 = w_1+w_2+w_3+w_4,\\[1mm]
&& a_2 = w_1w_2w_{12}+w_1w_3w_{13}+w_1w_4w_{14}\\
&&\qquad\qquad +w_2w_3w_{23}+w_2w_4w_{24}+w_3w_4w_{34},\\[1mm]
&& a_3 = w_1w_2w_3w_{12}w_{13}w_{23}w_{123}+w_1w_2w_4w_{12}w_{14}w_{24}w_{124}\\
&& \qquad\qquad +w_1w_3w_4w_{13}w_{14}w_{34}w_{134}+w_2w_3w_4w_{23}w_{24}w_{34}w_{234},\\[1mm]
&& a_4 = w_1w_2w_3w_4w_{12}w_{13}w_{14}w_{23}w_{24}w_{34}w_{123}w_{124}w_{134}w_{234}w_{1234}.
\end{array}
\end{equation*}
One intuitive way to compute the minimal absolute interaction $\|P\|$ is to decompose the space of molecules $(\RR_{>0})^{2^n}$. We look at $2^{2^n-n-1}$ regions on which for $|I|>1$ either $w_I\geq 1$ or $w_I\leq 1$ holds.
Explicitly, for each set of interactions $\mathcal I\subseteq \{I\subseteq [n]\mid |I|>1\}$, we define a region
\[ O_{\mathcal I}:= \big\{(w_I)_{I \subseteq [n]}\mid \forall |I|>1: w_I\geq 1 \;\forall I\notin\mathcal I \text{ and } w_I\leq 1 \; \forall I\in\mathcal I \big\} \]
so that
\[ (\RR_{>0})^{2^n} = \bigcup_{\mathcal I\subseteq \{I\subseteq [n]\mid |I|>1\}} O_{\mathcal I}. \]
On each region $O_{\mathcal I}$, the absolute interaction $\|W\|$ is a rational function in the interaction energies which we denote by $f_{\mathcal I}(W)$,
\begin{equation}
\label{eq:fI}
\|W\| = \prod_{I\in\mathcal I} w_I^{-1} \prod_{\substack{|I|>1\\I\not\in\mathcal I}} w_I=: f_{\mathcal I}(W) \quad \text{for } {\rm W} \in \mathcal{O}_{\mathcal{I}}.
\end{equation}
Finding the minimal absolute interaction inside $O_{\mathcal I}$ is a straight-forward problem in polynomial optimization:
\begin{equation}
\label{eq:OptInO}
\begin{array}{lll}
\displaystyle\mini_{{\rm W}\in\mathcal O_{\mathcal I}} & \quad & f_{\mathcal I}({\rm W})\\[3mm]
\st & & \Phi({\rm W})=P.
\end{array}
\end{equation}
However, this approach requires solving $2^{2^n-n-1}$ polynomial optimization problems, one for each region $\mathcal O_{\mathcal I}$, which is not easily feasible. \\
We can reduce the number of regions that require consideration by exploiting the
intrinsic symmetry of the problem.
To do this, we need to do a small journey in algebra and use the notion of group.
In mathematics, a (commutative) group is a set $S$ equipped with a binary
operation $\diamond$ that combines any two elements to form a third element of
that set and such that this operation is associative (and commutative), the set
contains an identity element ($\exists e \in S$ s.t. for all $x \in S$, $e
\diamond x = x$) and any element of $S$ is invertible (for all $x \in S$, there
exists $y\in S$ s.t. $x\diamond y = e$).
The basic example for groups is the set of integers $\mathbb{Z}$ equipped with the binary operation of addition.
In our setting, we want to exploit symmetry. We will consider the symmetric
group $S_n$ which is the set of all bijections from $\{1, \ldots, n\}$ to $\{1,
\ldots, n\}$ equipped with composition of maps. In the sequel, we use the following
standard notation: $(a b)$, for $a,b$ in $\{1, \ldots, n\}$, denotes the
bijection that permutes $a$ and $b$ in $\{1, \ldots, n\}$.
There exists a natural action of the symmetric group $S_n$ on the space of molecules $(\RR_{>0})^{2^n}$ which permutes the binding sites and the corresponding interaction energies,
\[ S_n\times (\RR_{>0})^{2^n} \longrightarrow (\RR_{>0})^{2^n},\quad (\sigma,(w_I)) \longmapsto \sigma\cdot (w_I) := (w_{\sigma(I)}). \]
This action permutes the regions $\mathcal O_{\mathcal I}$ and preserves binding polynomials as well as absolute interactions:
\begin{proposition}
The binding polynomial and the minimal absolute interaction of a molecule are invariant under the group action of $S_n$ i.e., $\Phi(\sigma\cdot {\rm W})=\Phi({\rm W})$ and $\|\sigma\cdot {\rm W}\| = \|{\rm W}\|$ for all ${\rm W}\in (\RR_{>0})^{2^n}$ and all $\sigma \in S_n$.
\end{proposition}
For example for any $1\leq i<j\leq n$, the permutation $(1i)(2j)\in S_n$ induces a bijection between the regions $O_{\{12\}}$ and $O_{\{ij\}}$ and hence
\begin{align*}
&\min\{ \|{\rm W}\| \mid {\rm W} \in O_{\{12\}} \text{ with }\Phi({\rm W})=P\} \\
&\qquad\qquad \qquad= \min\{ \|{\rm W}\| \mid {\rm W}\in O_{\{ij\}} \text{ with }\Phi({\rm W})=P\}.
\end{align*}
Therefore, it suffices to consider one representative per $S_n$-orbit of
regions. However, the number of required regions remains infeasibly large: We
can identify each region $\mathcal O_{\mathcal I}$ with a hypergraph with
vertices $[n]$ and edges $\mathcal I$ (a hypergraph is a generalization of a
graph in which an edge can join any number of vertices). The number of required
regions then coincides with the number of such hypergraphs up to isomorphism.
Included in the hypergraphs of the regions are all graphs with vertices $[n]$,
and the number of graphs on $n$ vertices up to isomorphism is well known to grow
faster than any exponential function \cite[\S 6.4]{GYZ14}.
Due to this problem, we preprocess the optimization problem. Note that we can rewrite Problem \eqref{orgopt} in the following minimax form, as on each region $\mathcal{O}_{\mathcal{I}}$ the function $f_{\mathcal I}(W)$, as defined in~\eqref{eq:fI}, dominates all other $f_{\mathcal I'}(W)$ with $\mathcal{I}' \neq \mathcal{I}$:
\begin{equation*}
\begin{aligned}
&\mini_{W} \qquad \max_{\mathcal I\subseteq \{I\subseteq [n]\mid |I|>1\}} f_{\mathcal I}(W) \\
&\st \qquad \quad \Phi(W)=P.
\end{aligned}
\end{equation*}
This can then be lifted to a problem with linear objective function and non-linear constraints:
\begin{equation}\label{newoptprob}
\begin{aligned}
&\mini_{W}\qquad r \\
&\st \quad \begin{cases} \quad f_{\mathcal I}(W)\leq r \quad \text{for all } \mathcal I\subseteq \{I\subseteq [n]\mid |I|>1\},\\ \quad \Phi(W)=P. \end{cases} \\
\end{aligned}
\end{equation}
Given a fixed binding polynomial $P$, the polynomial system $\Phi(W)=P$ poses a serious computational problem. To sidestep this issue we now introduce new coordinates $s_I:=\prod_{I'\subseteq I}w_{I'}$. Here, $s_I$ represents the free energy difference of microstate $I$ to the fully unoccupied state. On the positive orthant $(\RR_{>0})^{2^n}$, this defines a bijection
\begin{flushright}
\begin{tikzpicture}
\node (topLeft) {$(\RR_{>0})^{2^n}$};
\node[anchor=base west,xshift=35mm] (topRight) at (topLeft.base east) {$(\RR_{>0})^{2^n}$};
\node[anchor=base,yshift=15mm] (midLeft) at (topLeft.base) {$(w_I)$};
\node[anchor=base east,xshift=2mm] at (midLeft.base west) {$W=$};
\node[anchor=base,yshift=15mm] (midRight) at (topRight.base) {$\Big(\displaystyle\prod_{I'\subseteq I} w_{I'}\Big)$};
\node[anchor=base,yshift=-15mm] (bottomRight) at (topRight.base) {$(s_I)$};
\node[anchor=base west,xshift=-2mm] at (bottomRight.base east) {$=S$};
\node[anchor=base,yshift=-15mm] (bottomLeft) at (topLeft.base) {$\Big(\displaystyle\prod_{I'\subseteq I} s_{I'}^{(-1)^{|I\setminus I'|}}\Big)$};
\draw[<->,shorten >=7mm,shorten <=7mm] (topLeft) -- (topRight);
\draw[|->] ($(topLeft)+(1.5,1.5)$) -- node[below] {$\varphi$} ($(topRight)+(-1.5,1.5)$);
\draw[<-|] ($(topLeft)+(1.5,-1.5)$) -- node[above] {$\phantom{{}^{-1}}\varphi^{-1}$} ($(topRight)+(-1.5,-1.5)$);
\draw[draw opacity=0]
(midLeft) -- node[sloped] {$\in$} (topLeft)
(bottomLeft) -- node[sloped] {$\in$} (topLeft)
(midRight) -- node[sloped] {$\in$} (topRight)
(bottomRight) -- node[sloped] {$\in$} (topRight);
\node[anchor=base west,text width=35mm,align=right,yshift=7mm] at (topRight.base east) {};
\end{tikzpicture}\label{eq:s_I}
\end{flushright}
In the new coordinates, the formerly polynomial constraints $a_k = \sum_{|I|=k} \prod_{I'\subseteq I} w_{I'}$ are simplified to linear constraints $a_k=\sum_{|I|=k} s_I$ for $k=1,\ldots,n$. However, the functions $f_{\mathcal I}$ become more complicated. For example, for $n=4$ and $\mathcal I=\{ \{1,2,3\} \}$
\[ f_{\mathcal I} (W) = w_{123}^{-1}\cdot \prod_{\substack{|I|>1,\\ I\neq\{1,2,3\}}} w_I \quad \text{becomes} \quad
f_{\mathcal I}\circ\varphi^{-1} (S) = \frac{s_{1234}s_{12}^2s_{13}^2s_{23}^2}{s_{123}^2s_1^3s_2^3s_3^3s_4}. \]
Since the $f_{\mathcal I}\circ\varphi^{-1} (S)$ remain monomials (with possibly negative exponents), this complication is of little consequence for the resulting problem:
\begin{equation}\label{eq:optInS}
\begin{aligned}
&\mini_{S}\qquad r \\
&\st \quad \begin{cases} \quad f_{\mathcal I}\circ\varphi^{-1}(S)\leq r \quad \text{for all } \mathcal I\subseteq \{I\subseteq [n]\mid |I|>1\},\\ \quad \sum_{|I|=k} s_I= a_k \quad \text{ for all } k=1,\ldots,n. \end{cases} \\
\end{aligned}
\end{equation}
\begin{example}\label{example3dim}
Consider the polynomial $P_1$ of Example~\ref{example3site}. We use {\tt SCIP}~\cite{SCIP}, which is currently one of the fastest solvers for non-linear programming, to solve the resulting polynomial optimization problem \eqref{eq:optInS}. It uses a branch and bound method to solve the optimization problem with non-linear constraints.
Figure~\ref{scipformat} shows the full input on the left and the partial output on the right. In the input, the first constraints \texttt{c1}, \texttt{c2}, \texttt{c3} enforce $\sum_{|I|=k} s_I= a_k$ for $i=1,2,3$. The remaining constraints {\tt c4} to {\tt c19} enforce $f_{\mathcal{I}}\circ\varphi^{-1}(S)\leq r$. For example, \texttt{c19} states that $s_1s_2s_3 / s_{123}\leq r$ which is equivalent to $f_{\emptyset}(W)=(w_{123}w_{12}w_{13}w_{23})^{-1} \leq r$ in the coordinates $w_I$. The output states an approximate optimal value of $y=13.5$ and clearly lists all values of $s_I$ of the minimal molecules, which gave rise to the values of $w_I$ in Example~\ref{example3site}.
\end{example}
\begin{figure}
\begin{minipage}[t]{0.475\linewidth}
\begin{lstlisting}[showlines=true, basicstyle=\scriptsize\ttfamily]
Minimize
obj: r
Subject to
c1: s1+s2+s3=2
c2: s12+s13+s23=3
c3: s123=4
c4: s123-r*s1*s2*s3<=0
c5: s1*s2*s123-r*s3*s12^2<=0
c6: s1*s3*s123-r*s2*s13^2<=0
c7: s2*s3*s123-r*s1*s23^2<=0
c8: s12^2*s13^2*s23^2
-r*s1^3*s2^3*s3^3*s123<=0
c9: s1^3*s2*s3*s123-r*s12^2*s13^2<=0
c10: s1*s2^3*s3*s123-r*s12^2*s23^2<=0
c11: s13^2*s23^2-r*s1*s2*s3^3*s123<=0
c12: s1*s2*s3^3*s123-r*s13^2*s23^2<=0
c13: s12^2*s23^2-r*s1*s2^3*s3*s123<=0
c14: s12^2*s13^2-r*s1^3*s2*s3*s123<=0
c15: s1^3*s2^3*s3^3*s123
-r*s12^2*s13^2*s23^2<=0
c16: s1*s23^2-r*s2*s3*s123<=0
c17: s2*s13^2-r*s1*s3*s123<=0
c18: s3*s12^2-r*s1*s2*s123<=0
c19: s1*s2*s3-r*s123<=0
Bounds
0<s1<2
0<s2<2
0<s3<2
0<s12<3
0<s13<3
0<s23<3
4<=s123<=4
1<=r
End
\end{lstlisting}
\end{minipage}\hfill%
\begin{minipage}[t]{0.475\linewidth}
\begin{lstlisting}[showlines=true, basicstyle=\scriptsize\ttfamily]
SCIP version 6.0.0
Copyright (C) 2002-2018 ZIB Berlin
SCIP> read input.pip
SCIP> opt
SCIP Status : problem is solved
Solving Time : 2.32
Solving Nodes: 587
Primal Bound : +1.349997004816e+01
Dual Bound : +1.349997004816e+01
Gap : 0.00 %
SCIP> display solution
objective value: 13.4999700481621
y 13.4999700481621
s1 0.666630230040994
s2 0.66617295525322
s3 0.667196814705786
s12 0.444091987464541
s13 0.444773633975962
s23 2.1111343785595
s123 4
[...]
\end{lstlisting}
\end{minipage}\vspace{-3mm}
\caption{Computing the minimal absolute interaction for a molecule with 3 sites using {\tt SCIP}.}\label{scipformat}
\end{figure}
Since the number of constraints are not too many when $n=3$, {\tt SCIP} works well. However, for $n=4$ there are $2^{11}$ constraints which makes it a hard problem to solve directly.
To simplify our problem computationally, we compute an upper bound $b_{+}$ and a lower bound $b_{-}$ for the minimal absolute interaction of a given binding polynomial. If both bounds are identical, the minimum is found.
To compute the upper bound we minimize the minimal absolute interaction in a single region. For the experiments in Subsection~\ref{sectionexperimhemog} we consider the orthant $\mathcal{O}_{\emptyset}$. The result may not be the global minimum - since the minimum may lie in another orthant - but it gives an upper bound.
For the lower bound, we consider the relaxed problem by looking at a problem with less constraints. We pick a representative for each of the orbit of the group action $S^{n}$ and consider the corresponding region. We then only consider the constraints given by functions $f_{\mathcal{I}}$ dominant in those regions. Considering less constraints means that the feasible set of the optimization problem is bigger than originally allowed.
Thus, this new problem gives the lower bound on the true optimization problem.
When the gap between upper and lower bound is sufficiently small, we assume to have found a molecule minimizing the problem.
\section{Discussion and Outlook}
\subsection{Cooperativity and minimal absolute interaction}
A commonly used macroscopic conceptualization of cooperativity is that ``It's all or nothing'' \cite{hunter2009cooperativity}, meaning for hemoglobin that the probability weights of the extreme macrostates $0$ and $4$ are relatively big. This definition also aligns well with using the maximal Hill slope as criterion of cooperativity, since it detects a variance which is higher than that of any independent system.
However, observations made on the macroscopic level have often been interpreted with a view on microscopic properties, for instance that ``binding of a ligand molecule increases the receptor's apparent affinity, and hence increases the chance of another ligand molecule binding'' \cite{stefan2013cooperative}. Connecting macroscopic and microscopic properties is a non-trivial problem if no additional restriction on the underlying molecule are made. Microscopic properties can be easily inferred from the macroscopic behavior if additional assumptions are made, for instance the molecule being composed of physically indistinguishable sites. In particular, given a binding polynomial, we can calculate all interaction energies easily when we assume that $w_{ij} = w_{kl}$, $w_{ijk} = w_{lmn}$, etc.
However, without additional knowledge or restrictions, it is difficult to relate macroscopic and microscopic properties of larger, asymmetric systems \cite{martini2016cooperative}.
The minimal interaction energy required to generate the observed binding curve is a good candidate to connect macroscopic and microscopic characteristics and additional side conditions based on knowledge of the molecule structure could also be incorporated \cite{martini2017measure}. \\
As illustrated by the examples of $P_4$ and $P_6$, the minimal interaction energy criterion captures a conceptually different type of cooperativity than $n_{max}$. The polynomial $P_6$ requires a relatively high interaction energy due to giving the highest weights to macrostates $0$, $2$ and $4$, but stabilizing towards the intermediate value $2$ prevents an extreme variance. Contrary, $P_4$ has monotonously decreasing coefficients $a_1$ to $a_3$. Here, modeling the distribution requires less interaction energy to capture the relation of the probability on macrostates $1$, $2$ and $3$, and a higher value of the variance of the distribution can be realized.
\subsection{The mathematical problem}
Ligand binding raises many intriguing questions in applied algebra and geometry, and stands to benefit equally from the provided insight into the intrinsic geometry of the problem.
While previous works have explored the application of polynomial system solving techniques to a situation with different types of ligands \cite{Ren2019,martini2013interaction}, this work explores the use of polynomial optimization.
Using the minimal interaction energy as a criterion for cooperativity without assuming the target molecule to be composed of physically identical units, we deal with an algebraic optimization problem.
Even though there exists dedicated software to solve this kind of problem, we had to apply a transformation of the coordiantes (from $w_I$ to $s_I$), a transformation of the optimization problem (Eq.(\ref{newoptprob})) and approximations by upper and lower bounds to approach the considered examples. This effort illustrates that the setup of looking for the minimal absolute interaction by this method becomes rapidly complicated with increasing number of binding sites.
Whereas the case of three binding sites can be solved easily with \texttt{SCIP}, tetrameric hemoglobin with four binding sites already posed computational problems.
A way to facilitate the problem may be to exploit its intrinsic structure. For instance, there are results for strongly symmetric systems which might be generalized and tailored to the system with which we are dealing \cite{RienerSafey18}.
In addition, other mathematical fields such as tropical geometry \cite{maclagan2015introduction} may provide interesting tools to approach this problem.
\subsection{Other mathematical perspectives}
\subsubsection{Tensors}
Tensors \cite{Landsberg12} are multidimensional arrays of real numbers which generalize matrices. We can regard molecules as (positive) tensors in $\RR^{2\times \dots\times 2} = (\RR^2)^{\otimes n}$, whose entries are indexed by $\{0,1\}^n$ which is in bijection with the subsets of $[n]$.
From this point of view, many concepts in tensor theory have direct analogues in ligand binding, such as symmetric tensors and molecules with indistinguishable sites. Notably, the central concepts of tensor rank and tensor rank decomposition have a special meaning in ligand binding:
\begin{proposition}
Let $S=(s_I)_{I\subseteq[n]}\in\RR^{2^n}$ be a molecule with $n$ sites and whose corresponding tensor $(s_I)_{I\in\{0,1\}^n}$ has (positive) rank $k$. Then $S$ can be written as a weighted average of $k$ molecules with $n$ independent sites. In particular, a corresponding mixture of the $k$ molecules will have the same binding curve as $S$.
Moreover, this is not possible with $k-1$ molecules with $n$ independent sites.
\end{proposition}
\begin{proof}
Rank-one tensors are of the form $v_1\otimes \dots \otimes v_n$ for some $v_i=(s_i,t_i)\in\RR^2$ with entries
\begin{equation*}
\label{eq:tensor}
(1,0,\dots)\!: s_1\cdot\prod_{i\neq 1} t_i, \quad (0,1,0,\dots)\!: s_2 \cdot\prod_{i\neq 2} t_i, \quad (1,1,0,\dots)\!: s_1s_2 \cdot\prod_{i\neq 1,2} t_i, \quad \ldots
\end{equation*}
They can be regarded as molecules with particularly simple interaction, as molecules with independent sites are rank-one tensors with $t_i=1$. A tensor $S$ of rank $k$ can be expressed as
\[ S = \sum_{j=1}^k v_1^{(j)} \otimes \dots \otimes v_n^{(j)} \]
for $k$ rank-one tensors $v_1^{(j)} \otimes \dots \otimes v_n^{(j)}$ and cannot be expressed as a sum of $k-1$ rank-one tensors. Supposing $v_i^{(j)}=(s_i^{(j)},t_i^{(j)})\in\RR^2$, we can reformulate this into
\[ S = \sum_{j=1}^k \Big(\prod_{i=1}^k t_i^{(j)} \Big)
\cdot \left(
\begin{smallmatrix}
s_1^{(j)}/t_1^{(j)} \\ 1
\end{smallmatrix}
\right) \otimes \dots \otimes \left(
\begin{smallmatrix}
s_n^{(j)}/t_n^{(j)} \\ 1
\end{smallmatrix}
\right), \]
thus $S$ is a weighted average $k$ molecules with $n$ independent sites.
\end{proof}
\subsubsection{Tropical Geometry}
Tropical geometry \cite{maclagan2015introduction} revolves around functions on the max-plus semiring. One example of such as function is the absolute interaction in logarithmic coordinates:
\begin{center}
\begin{tikzpicture}
\node (topLeft) {$(\RR_{>0})^{2^n}$};
\node[anchor=base west,xshift=45mm] (topRight) at (topLeft.base east) {$\RR^{2^n}$};
\node[anchor=base,yshift=-15mm] (bottomRight) at (topRight.base) {$(\log(w_I))$};
\node[anchor=base west,xshift=-2mm] at (bottomRight.base east) {$=:(x_I)$};
\node[anchor=base,yshift=-15mm] (bottomLeft) at (topLeft.base) {$(w_I)$};
\node[anchor=base,yshift=-15mm] at (bottomLeft.base) {$\displaystyle\prod_{I\subseteq[n]}\max(w_I,w_I^{-1})$};
\node[anchor=base,yshift=-15mm] (bottombottomRight) at (bottomRight.base) {$\displaystyle\sum_{I\subseteq[n]}\max(x_I,-x_I)$};
\node[anchor=base west,xshift=-2mm] at (bottombottomRight.base east) {$=:\|(x_I)\|$};
\draw[<->] ($(topLeft.base east)+(1.25,0.1)$) -- ($(topRight.base west)+(-1.25,0.1)$);
\draw[|->] ($(topLeft.base east)+(1.25,-1.4)$) -- ($(topRight.base west)+(-1.25,-1.4)$);
\draw[|->] ($(topLeft.base east)+(1.25,-2.9)$) -- ($(topRight.base west)+(-1.25,-2.9)$);
\draw[draw opacity=0]
(bottomLeft) -- node[sloped] {$\in$} (topLeft)
(bottomRight) -- node[sloped] {$\in$} (topRight);
\end{tikzpicture}
\end{center}
In tropical geometry, the function $\|\cdot\|$ is referred to as a tropical rational function. It is a piecewise-linear function, linear on the image of each $\mathcal O_{\mathcal I}$, and the boundary separating its regions of linearity is called its tropical hypersurface $\text{Trop}(\|\cdot\|)$. The gradient of $\|\cdot\|$ is constant on each region of linearity, which is why understanding the intersection of a set $L\subseteq\RR^{2^n}$ with $\text{Trop}(\|\cdot\|)$ is paramount to computing the minimum of $\|\cdot\|$ on $L$.
In our setting, $L$ is the image of the feasible set given by the conditions $\Phi(W)=P$ which is an exponential variety of codimension $n$. Since the number of regions of linearity is finite, there exist only a finite number of intersection patterns of $L$ and $\text{Trop}(\|\cdot\|)$. Classifying these intersection patterns depending on the coefficients of $P$ will contribute greatly towards computing the minimal interactions and minimizing molecules.
\subsubsection{Dependency measures}
From a stochastic or statistical point of view, the problem of cooperativity is a problem of quantifying (minimal) dependency, which is a classical research topic in these fields \cite{renyi1959measures,schweizer1981nonparametric,koyak1987measuring}. In particular copulas \cite{schweizer1991thirty,durante2010copula,nelsen2007introduction} have become a common tool to describe dependencies between random variables. Copulas cannot be applied directly to the problem since they are defined on continuous, not discrete, random variables. Moreover, they aim at modeling the dependency of given random variables which in our scenario corresponds to the dependency of the sites of a given molecule. The minimization problem of finding the minimal dependency required to generate a given distribution of a sum of dependent variables is usually not included. Nevertheless, the problem of cooperativity should also be treated from a more probabilistic point of view and the available concepts should be considered in more detail in the context of ligand binding.
\subsection{Conclusion}
We illustrated how searching for the minimal interaction energy required to generate a binding curve leads to a problem of algebraic optimization. Even though, there are computational tools available and the structure of the problem seems relatively simple on first sight, it becomes complicated when more than $3$ binding sites are considered. In particular the fact that we had to calculate a lower bound by a relaxed problem and not having been able to close the gap between upper and lower bound completely, indicates that our approach can currently not be enveloped in a simple ready-to-use tool to quantify cooperativity. We showed that the concept of minimally present interaction energy is different from using the maximal Hill slope as a measure of cooperativity. The first also detects required changes in affinity from one macrostate to the other, whereas the latter only focuses on the macroscopic variance of the distribution. Future work may use other mathematical concepts such as tensor theory, tropical geometry and probabilistic measures of dependencies to approach questions related to cooperativity.
\section{Experimental Results}\label{sectionexperimhemog}
In this section, we run computational experiments on data from \cite{imai1973analyses,ikeda1983thermodynamic} which is also summarized in \cite{connelly1986analysis}.
\subsection{Data}
The first data set consists of eight oxygen binding polynomials of human adult hemoglobin in its native form and of three chemically treated versions. The oxygen binding to each of the four hemoglobin variants has been observed under two different environmental conditions \cite{imai1973analyses}.
The coefficients $a_1$, $a_2$ and $a_3$ of the corresponding polynomials $P_1$ to $P_8$ are listed in Figure~\ref{tableconnelly}. The coefficients $a_0$ and $a_4$ are standardized to $1$. Figure~\ref{fig:polynomials1To8} describes the experimental conditions under which binding polynomials $P_1$ to $P_8$ were obtained and illustrates the ranking of the degree of cooperativity as defined by the maximal slope of the Hill plot $n_{max}$ (as reported by \cite{imai1973analyses}). The slope to the Hill plot relates the variance of the probability distribution on the macrostates $\{0,1,2,3,4\}$ at the respective ligand activity to the variance of a binomial distribution with the same mean \cite{abeliovich2005empirical,martini2016cooperative}. A value larger than 1 indicates a variance higher than the maximal value an independent system can generate.\\
\begin{figure}[h]
\centering
\begin{tabular}{l|ccc}
\multirow{2}{*}{Human Hb} & in the absence && in the presence \\
& of 2mM DPG && of 2mM DPG \\[1mm] \hline
\multirow{2}{*}{untreated} & \multirow{2}{*}{$P_1$} &\multirow{2}{*}{$\prec$} & \multirow{2}{*}{$P_2$} \\
& & & \\[-2mm]
& $\curlyvee$ & & $\curlyvee$ \\[-2mm]
treated with & \multirow{2}{*}{$P_3$} & \multirow{2}{*}{$\prec$} & \multirow{2}{*}{$P_4$} \\
iodoacetamide & & & \\[-2mm]
& $\curlyvee$ & & $\curlyvee$ \\[-2mm]
treated with & \multirow{2}{*}{$P_5$} &\multirow{2}{*}{$\prec$} & \multirow{2}{*}{$P_6$} \\
N-ethylmaeimide & & & \\[-2mm]
& $\curlyvee$ & & $\curlyvee$ \\[-2mm]
treated with & \multirow{2}{*}{$P_7$} &\multirow{2}{*}{$\prec$} & \multirow{2}{*}{$P_8$} \\
carboxypeptidase A & & &
\end{tabular}\vspace{-3mm}
\caption{Conditions under which binding polynomials 1 to 8 were derived and the relation of their degree of cooperativity according to the maximal slope of the Hill plot $n_{max}$ ($\succ$ compares the hill slope $n_{max}$, DPG = 2,3-diphosphoglycerate).}
\label{fig:polynomials1To8}
\end{figure}
The second data set \cite{ikeda1983thermodynamic} contains five binding polynomials ($P_9$ to $P_{13}$) of native hemoglobin HbII of {\it Scapharca inaequivalvis} measured at different temperatures (Figure~\ref{fig:polynomials9To25}) \cite[Table~3]{ikeda1983thermodynamic}.
\begin{figure}[h]
\centering
\begin{tabular}{l|ccccc}
Clam HbII & at 10${}^\circ$ & at 15${}^\circ$ & at 20${}^\circ$ & at 25${}^\circ$ & at 30${}^\circ$\\[3mm] \hline
& $P_{9}$ & $P_{10}$ & $P_{11}$ & $P_{12}$ & $P_{13}$
\end{tabular}
\caption{Temperature in degree Celsius at which binding polynomials 9 to 13 were determined.}
\label{fig:polynomials9To25}
\end{figure}
\subsection{Minimal absolute interactions}
Unlike the case $n=3$ in Example~\ref{example3site}, solving Problem~(\ref{eq:optInS}) for $n=4$ proved to be too hard for \texttt{SCIP}, which is why we list upper ($b_+$) and lower ($b_-$) bounds for the minimal absolute interaction $\|P_i\|$ instead. The upper bound is the minimal absolute interaction inside the region $\mathcal O_{\emptyset}$, i.e., it is obtained by solving Problem \eqref{eq:OptInO} for $\mathcal I=\emptyset$. In other words, the upper bound represents the minimal value on the region in which all interactions are equal to or larger than 1. The lower bound is computed through a relaxation of Problem~(\ref{eq:optInS}). Inspired by the action of the symmetric group $S_n$ on the regions $\mathcal O_{\mathcal I}$, we picked a set of $180$ representatives and discarded all constraints $f_{\mathcal I}(W)\leq r$ with $\mathcal I$ not in them, see Section~\ref{app:representatives}.
\begin{figure}[h]
\centering
\begin{tabular}{l|rrr|r|r|r}
& $a_1$ & $a_2$ & $a_3$ & $b_{+}$ & $b_{-}$ & $n_{max}$\\ \hline
$P_{1}$ & 0.835 & 0.379 & 0.541 & 527 & 527& 2.51\\
$P_{2}$ & 0.789 & 0.154 & 0.0648 & 3322 & 1600&3.09 \\
$P_{3}$ & 1.42 & 2.42 & 0.752 & 111 & 63 &1.63\\
$P_{4}$ & 0.647 & 0.568 & 0.0986 & 2002& 1460 &2.71\\
$P_{5}$ & 2.0 & 2.31 & 2.04 & 16 & 16 &1.44\\
$P_{6}$ & 0.539 & 0.909 & 0.554 & 3033& 3030 &2.27\\
$P_{7}$ & 3.47 & 4.74 & 2.76 & 2.27 & 1.68 &1.15\\
$P_{8}$ & 3.26 & 5.36 & 2.23 & 7.64 & 2.19 &1.23\\[3mm]
$P_{9}$ & 1.4 & 1.0 & 0.62 & 66 & 66 & 2.08\\
$P_{10}$ & 1.4 & 0.96 & 0.60 & 66 & 66 &2.10\\
$P_{11}$ & 1.2 & 0.93 & 0.70 & 123 & 123 &2.08\\
$P_{12}$ & 1.4 & 0.95 & 0.62 & 66 & 66 &2.10\\
$P_{13}$ & 1.1 & 0.98 & 0.59 & 175 & 175 &2.12
\end{tabular}
\caption{Binding polynomials, bounds on the minimal absolute interaction and $n_{max}$ as reported by \cite{ikeda1983thermodynamic,imai1973analyses}. Coefficients $a_0$ and $a_4$ are equal to $1$ for all polynomials.}\label{tableconnelly}
\end{figure}
Figure~\ref{tableconnelly} illustrates that for most binding polynomials our upper and lower bound allows for a sufficient approximation of the absolute minimal interaction.
However, there remains a nontrivial gap between both bounds for $P_2$, $P_3$, $P_4$.
Let us assume that the correct value for $P_2$ is close to the upper bound which we determined. Then, nearly all relations of the degree of cooperativity of polynomials $P_1$ to $P_8$ described by the maximal slope of the Hill plot $n_{max}$ (Figure~\ref{fig:polynomials1To8}) are also found with the minimal interaction criterion. The only exception is the ranking of $P_4$ and $P_6$.
Whereas $P_4$ describes a more cooperative system than $P_6$ according to $n_{max}$, we obtain a different picture when using the minimal interaction criterion which states that $P_6$ is almost as cooperative as $P_2$. The question arises why the two measures of cooperativity are so different.
An answer may be found considering the coefficients of $P_4$ and $P_6$. The coefficient $a_2$ of $P_6$ is larger than $a_1$ and $a_3$, which indicates that the macrostate of having $2$ sites occupied --at sufficiently high ligand activity-- is more probable than the states of having $1$ or $3$ ligands bound. To generate this type of distribution, we may need absolute interaction energy at each macrosotate level, for instance negative and positive interaction energy. Moreover, a family of distributions with a relatively high weight on macrostate $2$ has a reduced possibility to generate a high variance since an extraordinary high variance requires more weights on the macrostates $0$ and $4$.\\
The polynomials $P_9$ to $P_{13}$ were derived from the same protein at different temperatures \cite{ikeda1983thermodynamic}. We see that the reported $n_{max}$ is not behaving monotonously with increasing temperature, but that it varies between $2.08$ and $2.12$. Since one might expect a more monotonous development, we interpret this observation as a result of measuring imprecision of the coefficients. Comparing the minimal interaction to $n_{max}$, we see the same problem on a larger scale, which is a results of the minimal energy criterion being defined on an exponential scale of actual physical energies.
We consider the molecules realizing the upper bound in more detail.
\subsection{Molecules solving the optimization problem}
Figure~\ref{fig:minimalMolecules4Sites} presents the molecules realizing the upper bound of the minimal absolute interaction for all polynomials.
We see that the major part of the estimated interactions is very close to $1$. This is not surprising since the considered optimization problem minimizes the absolute values of the physical energies required to generate the binding curve. $L_1$ optimizations --for instance such as the least absolute shrinkage and selection operator (Lasso) regression \cite{tibshirani1996regression}-- are known to provide sparse solutions, meaning that many of the physical energies of the solution will be zero. Thus, this observation fits well to the context. Considering the structure of the solutions in more detail, we see that most solutions include a non-trivial weight for a second degree interaction ($w_{34}$).
The only polynomial whose minimizing molecule in $\mathcal{O}_{\emptyset}$ does not possess an interaction of pair-wise interaction is $P_2$
which has the highest $n_{max}$. This observation seems plausible since the weights should be more towards the extreme values, that is to the macrostates $0$ and $4$ to obtain a high variance of the macroscopic distribution. Compared to $P_1$, $P_2$ puts more probability on the macrostates $3$ and $4$. With this view, we can also explain the apparent contradiction that we find when using $n_{max}$ or the minimal absolute interaction to compare polynomials $P_4$ and $P_6$. As described above, $P_6$ has a coefficient $a_2$ which is larger than $a_1$ and $a_3$ and almost of same size as $a_4$, which is $1$.
This means that the distribution on the macrostates is stabilized towards the macrostate $2$, which hinders
very large variances and thus leads to a smaller $n_{max}$ than that of $P_4$.
However, this circumstance leads to a high interaction weight for a microstate of macrostate $2$. While the main weight for $P_4$ is on macrostate $4$ -which helps to achieve a higher value of $n_{max}$, $P_6$ requires more interaction energy to model the behavior at macrostate $2$. \\
\begin{figure}[h]
\centering
\setlength{\tabcolsep}{2.5pt}
\begin{tabular}{l|c c c c c c c c c c c}
& $w_1$ & $w_2$ & $w_3$ & $w_{4}$ & $w_{12}\cdots w_{24}$ & $w_{34}$ & $w_{123}$ & $w_{124}$ & $w_{134}$ & $w_{234}$ & $w_{1234}$ \\ \hline
$P_{1}$ & 0.209 & 0.209 & 0.209 & 0.209 & 1.00 & \color{blue} 3.70 & 1.00 & 1.00 & \color{blue}14.5 & 1.00 & \color{blue} 9.79 \\
$P_{2}$ & 0.0822 & 0.0822 & 0.0822 & 0.543 & 1.00 & 1.00 & \color{blue} 6.60 & 1.00 & \color{blue} 14.7 & 1.00 & \color{blue} 34.3 \\
$P_{3}$ & 0.179 & 0.179 & 0.531 & 0.531 & 1.00 & \color{blue} 7.12 & 1.00 & 1.00 & 1.00 & 1.00 &\color{blue} 15.6 \\
$P_{4}$ & 0.101 & 0.101 & 0.223 & 0.223 & 1.00 & \color{blue} 9.42 & 1.00 & 1.00 & 1.00 & 1.00 & \color{blue}212 \\
$P_{5}$ & 0.500 & 0.500 & 0.500 & 0.500 & 1.00 & \color{blue} 4.24 & 1.00 & 1.00 & \color{blue} 1.69 & \color{blue} 1.69 & \color{blue} 1.32 \\
$P_{6}$ & 0.135 & 0.135 & 0.135 & 0.135 & 1.00 & \color{blue} 45.1 & 1.00 & 1.00 & \color{blue} 2.49 & \color{blue} 2.49 &\color{blue} 10.9 \\
$P_{7}$ & 0.569 & 0.569 & 1.17 & 1.17 & 1.00 & \color{blue} 1.30 &1.00 & 1.00 & 1.00 & 1.00 &\color{blue} 1.75 \\
$P_{8}$ & 0.265 & 0.265 & 1.36 & 1.36 & 1.00 & \color{blue} 2.06 & 1.00 & 1.00 & 1.00 & 1.00 & \color{blue} 3.70 \\[3mm]
$P_{9}$ & 0.350 & 0.350 & 0.350 & 0.350 & 1.00 & \color{blue} 3.16 & 1.00 & 1.00 &\color{blue} 1.97 & \color{blue} 1.97 & \color{blue} 5.43 \\
$P_{10}$ & 0.350 & 0.350 & 0.350 & 0.350 & 1.00 & \color{blue} 2.84 &1.00 & 1.00 &\color{blue} 2.11 &\color{blue} 2.11 &\color{blue} 5.26\\
$P_{11}$ & 0.300 & 0.300 & 0.300 & 0.300 & 1.00 & \color{blue} 5.33 & 1.00 & 1.00 & \color{blue} 2.24 &\color{blue} 2.24 & \color{blue} 4.60 \\
$P_{12}$ & 0.350 & 0.350 & 0.350 & 0.300 & 1.00 & \color{blue} 2.76 & 1.00 & 1.00 &\color{blue} 2.26 &\color{blue} 2.26 & \color{blue} 4.73 \\
$P_{13}$ & 0.275 & 0.275 & 0.275 & 0.275 & 1.00 &\color{blue} 7.96 & 1.00 & 1.00 & \color{blue} 1.66 & \color{blue} 1.66 &\color{blue} 8.01
\end{tabular}
\caption{Molecules of polynomials $P_1$ to $P_{13}$ realizing the upper bound $b_+$. The $w_I$s are rounded to four digits. The incidences of a $w_I$ being smaller than $1$ are numerical impressions when resolving from the $s_I$ coordinates. All interactions larger than 1 are highlighted in blue.} \label{fig:minimalMolecules4Sites}
\end{figure}
\section{Introduction}
Interaction between components is a key concept of biological systems.
While characteristics of a system of independent subunits are determined by the description of each element alone, a complex system with interactions is more than the sum of its parts. A classical example of a small biological system with non-trivial interaction is hemoglobin with its four binding sites for oxygen or other ligands such as carbon monoxide \cite{bohr1904ueber,barcroft1913combinations,hill1913combinations}. The overall binding curve, as a function of oxygen concentration at constant temperature (``isotherm''), is
much steeper than an overall binding curve generated by a system of independent sites can be. Hemoglobin ``prefers'' the extreme states of being fully (un)occupied to intermediate states with one, two or three ligands bound.
This dependent binding behavior or an ``abnormal" steepness of a response curve is described by the concept of ``cooperativity''. Cooperativity is not only considered to be important for
(oxygen) transport, but for instance also for the formation of multi-protein complexes \cite{roy2017cooperative}, general signal transduction processes \cite{salakhieva2016kinetic,lenaerts2009information,martini2016model} and the regulation of noise \cite{gutierrez2009role,monteoliva2013noise}.
Observations on the macroscopic level, that is on how many sites are occupied have often been connected to microscopic behavior that is which sites are occupied and how certain sites interact (for a review see \cite{stefan2013cooperative}).
However, it has been shown that in the framework of the grand canonical ensemble of statistical mechanics, different definitions of cooperative binding are not coinciding when general asymmetric systems with more than two binding sites are considered \cite{martini2016cooperative,martini2017relation}. In particular, macroscopic phenomena which have been attributed to positive cooperativity in literature can be caused by a system being negative cooperative on the microcopic level \cite{martini2016cooperative}.
\\
For the characterization of the macroscopic binding behavior, roots of the grand partition function (also called the ``binding polynomial'') have been the objects of investigations \cite{briggs1983new,briggs1984cooperativity,briggs1985relationship}. In particular, it has been pointed out that the criterion of having non-real roots is generally the relevant aspect indicating that a binding process relies on non-negligible interaction \cite{martini2016cooperative}. In case that the binding polynomial has only real roots, these roots define hypothetical single-site-systems whose sum give the initial overall curve \cite{onufriev2001novel,onufriev2004decomposing,martini2013mathematical}. The macroscopic behavior of the system could thus theoretically be engineered by using these independent sites which suggests that the potentially present intrinsic interaction is not essential for the macroscopic functionality of the system. Conversely, in case that the binding polynomial has non-real roots, the overall system cannot be represented as a decoupled system consisting of independent ``real'' binding sites suggesting that the intrinsic interaction is essential for the macroscopic behavior of the system. This qualitative definition raises the question of how to quantify the cooperativity of a system. \\
The quantification of cooperativity has predominantly been discussed for symmetric systems consisting of physically indistinguishable sites \cite{abeliovich2016hill,rong2018quantification,rong19}. In particular, in the framework of the grand canonical ensemble, the molecule consisting of identical sites can be easily inferred from the coefficients of the binding polynomial and the corresponding interaction energies can be determined \cite{martini2016cooperative}. Moreover, it has been proposed to use the absolute minimal interaction which is required to generate the overall binding behavior as a general measure of cooperativity when the binding sites are not considereed to be physically identical \cite{martini2017measure}. In this general situation without additional side conditions, an observed overall titration behavior can be explained by an infinite number of different hypothetical molecules. A central question to characterize cooperativity can then be what the minimal absolute interaction across all these molecules is and in particular whether it is zero.
This quantitative measure of cooperativity has been defined and upper bounds based on polynomial factorization or on the interaction of a symmetric system, have been derived \cite{martini2017measure}.
However, it has not been clear how to explicitly calculate this measure of cooperativity for larger systems. \\
In this work, we show that looking for the minimal interaction energy required to generate a given overall titration behavior, leads to an algebraic optimization problem which can be tackled using the software {\tt SCIP} \cite{SCIP}. We describe the mathematical problem and calculate examples for binding polynomials from native hemoglobins and some modified variants \cite{connelly1986analysis,ikeda1983thermodynamic,imai1973analyses}. For our examples, the minimal interaction reduces when the molecule structure of hemoglobin is disturbed.
However, our results rank the degree of cooperativity of chemically treated hemoglobines differently than the maximal Hill slope.
\section{List of Constraints}\label{app:representatives}
Below is a compact representation of the list of $180$ constraints we considered to compute the lower bounds $b_{-}$ of minimal absolute interaction in Figure~\ref{tableconnelly}.
There, $\{12,34,123\}$ represents the constraint $f_{\mathcal I}\leq r$ for $\mathcal I=\{12,34,123\}$, which is the absolute interaction $\|W\|$ for $W\in\mathcal O_{\mathcal I}$ where $w_{12},w_{34},w_{123}\leq 1$ and $w_{I} \geq 1$ otherwise.
\vspace{0.7cm}
\noindent
\begin{scriptsize}
$\emptyset$, $\!\{12\}$, $\!\{123\}$, $\!\{1234\}$, $\!\{12$, $\!\!13\}$, $\!\{12$, $\!\!34\}$, $\!\{12$, $\!\!123\}$, $\!\{12$, $\!\!134\}$, $\!\{123$, $\!\!124\}$, $\!\{12$, $\!\!1234\}$, $\!\{123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14\}$, $\!\{12$, $\!\!13$, $\!\!23\}$, $\!\{12$, $\!\!13$, $\!\!24\}$, $\!\{12$, $\!\!13$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!124\}$, $\!\!\!\{12$, $\!\!34$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!234\}$, $\!\{12$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!123$, $\!\!134\}$, $\!\{12$, $\!\!134$, $\!\!234\}$, $\!\{123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!1234\}$, $\!\{12$, $\!\!34$, $\!\!1234\}$, $\!\{12$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!134$, $\!\!1234\}$, $\!\{123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!134\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!234\}$, $\!\{12$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!123$, $\!\!134$, $\!\!234\}$, $\!\{123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\!\!\{12$, $\!\!123$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!123$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$, $\!\{12$, $\!\!13$, $\!\!14$, $\!\!23$, $\!\!24$, $\!\!34$, $\!\!123$, $\!\!124$, $\!\!134$, $\!\!234$, $\!\!1234\}$
\end{scriptsize}
|
\section{Introduction}
This paper discusses how relativistic invariance is realized in
Euclidean formulations of relativistic quantum theory. In a quantum
theory relativistic invariance means that quantum observables, which
are probabilities, expectation values and ensemble averages, have the
same value for equivalent experiments that are performed in different
inertial coordinate systems. This means that experiments performed in an
isolated system cannot be used to distinguish inertial coordinate
systems. In special relativity different inertial coordinate systems
are related by the subgroup of Poincar\'e group connected to the
identity. In 1939 Wigner \cite{Wigner:1939cj} showed that a necessary
and sufficient condition for a quantum system to be relativistically
invariant is that vectors representing equivalent quantum states in
different inertial coordinate systems are related by a unitary ray
representation of this subgroup on the Hilbert space of the quantum
theory.
Relativistically invariant quantum theories are needed to study
physics on distance scales that are small enough to be sensitive to
the internal structure of a nucleon. This is because in order to get
wavelengths short enough to resolve the internal structure of a
nucleon it is necessary to transfer a momentum to the nucleon that is
comparable to or larger than its mass scale.
In quantum theories time evolution is generated by a one-parameter
unitary group. The infinitesimal generator of this group is the
Hamiltonian, which is a positive self-adjoint operator on the Hilbert
space of the quantum theory. Because the spectrum of the Hamiltonian,
time can be analytically continued to the lower-half complex time
plane. For imaginary times, $t \to -i \tau$, the unitary time
evolution group becomes a contractive Hermitian semigroup. For any
fixed $\tau>0$, $e^{-H\tau}$ has the same eigenvectors as the
Hamiltonian, and the eigenvalues $\lambda$ of $H$ are related to the
eigenvalues $\eta$ of $e^{-H\tau}$ by $\lambda = - \ln (\eta)/\tau$.
This implies that it is possible to solve dynamical problems directly
in a Euclidean representation. For some applications it is enough to
replace $H$ by $e^{-\tau H}$. This is a well-behaved bounded operator
with a spectrum on the unit interval $[0,1]$; the parameter $\tau >0$
can be adjusted to be sensitive to different parts of the spectrum of $H$.
Relativistic invariance normally requires an analytic continuation back to real
time. These observations provide the motivation for investigating
Euclidean approaches to relativistic quantum field theory and quantum
mechanics.
Euclidean approaches were first advocated by Schwinger
\cite{Schwinger:pna}\cite{Schwinger:1959zz} who used the spectral
condition in time-ordered Green's functions to establish the existence
of an analytic continuation to imaginary times. Independently,
axiomatic treatments of quantum field theory
\cite{Wightman:1980}\cite{jost} led to an understanding of the
analytic properties of vacuum expectation values of products of
fields, also based on the spectral condition. The Euclidean approach
to quantum field theory was advocated by Symanzik
\cite{Symanzik:1966}\cite{Symanzik:1968zz}, and developed by Nelson
\cite{Nelson:1973}. Osterwalder and Schrader
\cite{Osterwalder:1973dx}\cite{Osterwalder:1974tc} identified
properties of Euclidean covariant distributions that are sufficient to
reconstruct a relativistic quantum field theory. Two observations
that are contained in the work of Osterwalder and Schrader are (1)
that an explicit analytic continuation is not necessary to construct a
relativistic quantum theory and (2) the reconstruction of a
relativistic quantum theory is not limited to local field theories.
The discussion that follows is motivated by these two observations.
The Poincar\'e and four-dimensional Euclidean groups are related
because the parameters of both groups can be analytically
continued and the covering group of the resulting complex groups are identical,
$SL(2,\mathbb{C})\times SL(2,\mathbb{C})$. What this means is that
the real Poincar\'e group can be considered to be a complex subgroup
of the complex Euclidean group, or conversely, the real Euclidean group
can be considered to be a complex subgroup of the complex Poincar\'e group.
These identifications imply formal relations between the infinitesimal
generators of the Poincar\'e group and the real Euclidean group \cite{palle}\cite{olafsson}.
Specifically, if $ P^0_e,\mathbf{P}_e,J^{ij}_e,J^{0i}_e$ satisfy the
commutation relations of the Euclidean Lie Algebra, then the operators
$P^0_m=-i P^0_e,\mathbf{P}_m=\mathbf{P}_e, J^{ij}_m=J^{ij}_e,J^{0i}_m
:= -iJ^{0i}_e$, will satisfy the commutation relations of the
Poincar\'e Lie Algebra. However, because of the factors of $i$, both
sets of operators cannot be self-adjoint on the same representation of
the Hilbert space.
Osterwalder and Schrader construct a new Hilbert space representation
where the Poincar\'e generators become self adjoint. Osterwalder and
Schrader start with a representation of a Hilbert space defined with a
Euclidean covariant kernel. On this space the Euclidean
transformations are norm preserving which defines a unitary
representation of the Euclidean group. Next they choose an arbitrary
time direction and multiply the final Euclidean time variables in this
kernel by an operator that reverses the sign of all of the final
Euclidean times. Introducing this time reflection in the Euclidean
kernel breaks the Euclidean invariance and has the effect of making
the Poincar\'e generators constructed from the Euclidean generators
Hermitian on this space. The integration variables remain unchanged -
they include the Euclidean times. The problem is that the resulting
quadratic form cannot be positive for all Euclidean test functions.
This is easily seen by taking functions with positive time support and
extending them to be even or odd under time reflection. Since the
quadratic forms will have opposite signs, they cannot both have positive
norm with this new inner product. This flaw is fixed by projecting
the test functions on a suitable subspace. The subspace identified by
Osterwalder and Schrader is the subspace of functions of Euclidean
space-time variables with support for positive absolute and relative
Euclidean times. The Euclidean kernels are called {\it reflection
positive} if the norms with respect to the inner product with the
Euclidean time reflection is non-negative on this subspace.
Reflection positivity is a constraint on the Euclidean distributions
\cite{Jaffe:2018ftu}. This construction is a specific application of a
general construction based on an abstract notion of reflection
positivity \cite{palle}\cite{olafsson}.
Because this projection is independent of the form of the Euclidean
kernel, cluster properties, which are an important physical
requirement, can be expressed entirely in terms of properties of the
kernel - the range of the projector does not change. Cluster
properties can be a difficult constraint to satisfy in some representations of
relativistically quantum
mechanics \cite{Sokolov:1977}\cite{Coester:1982vt}\cite{Keister:1991sb},
but it can be easily achieved in the Euclidean approach.
The restriction to positive relative time is because the Euclidean
kernels for irreducible representations of the Poincar\'e group become
singular for zero relative times. Since identical particles have an
exchange symmetry, this is reflected in the symmetry properties of the
Euclidean kernel. As long as the relative time supports are disjoint,
the symmetry can be used to reorder the variables so the support
satisfies the positive relative-time condition. What separates
relativistic quantum theory from local quantum field theory is whether
the symmetries involve all of the coordinates in the kernel or just
separately involve the initial and final coordinates. This will be
discussed in more detail below.
Reflection positivity is a strong constraint, particularly when it is
combined with Euclidean covariance and cluster properties. One
consequence is that it implies the spectral condition that Schwinger
originally used to justify the existence of an analytic continuation.
The advantage of the Osterwalder-Schrader reconstruction is that this
analytic continuation is never explicitly needed.
In this paper a Euclidean relativistic theory is defined by a
finite or infinite collection of Euclidean covariant tempered
distributions
\begin{equation}
S_{m:n} (x_m, \cdots ,x_1 ; y_1, \cdots, y_n).
\label{a.1}
\end{equation}
These kernels contain the dynamics.
The kernels satisfy the permutation symmetry,
\[
S_{m:n} (x_m, \cdots ,x_1 ; y_1, \cdots, y_n) =
(\pm)^{\vert \sigma \vert}
S_{m:n} (x_{\sigma (m)} \cdots ,x_{\sigma (1)} ; y_1, \cdots, y_n)
\]
\begin{equation}
= (\pm)^{\vert \sigma \vert}
S_{m:n} (x_m, \cdots ,x_1 ; y_{\sigma (1)}, \cdots, y_{\sigma (n)} )
\label{a.2}
\end{equation}
where $\sigma()$ is a permutation on $m$ or $n$ objects,
$\vert \sigma \vert$ is 0 if $\sigma$ is an even permutation and 1 if
it is an odd permutation. The $+$ sign is for Bosons and the $(-)$
sign is for Fermions. For local quantum field theories the collection
must be infinite, $S_{m:n} = S_{k:l}$ whenever $m+n=k+l$, and the
permutation symmetry is with respect to all $n+m$ variables. In
(\ref{a.1}-\ref{a.2}) the $x_n$ can also include spin degrees of
freedom.
The symmetry in the local field theory case arises because the domain
of analyticity, that comes from the spectral condition, can be
extended by complex Lorentz transformations. The extended domain of
analyticity includes real space-like separated points (Jost points)
that allow the local
fields to be reordered \cite{Wightman:1980}, relating Green's functions with
permuted arguments. This symmetry is not assumed in this work. One
consequence of relaxing this condition is that it is possible to have
different $N$-point Green's functions for different numbers of initial
and final coordinates.
The setting for a quantum theory is a Hilbert space. A dense set of
vectors in the Euclidean representation of the Hilbert space are
sequences of Schwartz test functions of Euclidean space-time variables
\begin{equation}
\{\psi_n (x_1 \cdots x_n)\}_{n=0}^N
\label{a.3}
\end{equation}
that vanish unless the Euclidean times satisfy
$0 < x_1^0 < x_1^0 < \cdots < x_n^0$.
The Hilbert space inner product is
\[
\langle \psi \vert \phi \rangle =
\sum_{mn}
\int d^4x_m \cdots d^4x_1 d^4y_1\cdots d^4 y_n
\psi_m^* (\theta x_m, \cdots , \theta x_1) \times
\]
\begin{equation}
S_{m:n} (x_m, \cdots ,x_1 ; y_1, \cdots, y_n)
\phi_n (y_1, \cdots , y_n)
\label{a.4}
\end{equation}
where $\theta$ represents Euclidean time reflection,
$\theta (x_i^0,\mathbf{x}_i) := (-x_i^0,\mathbf{x}_i)$ .
Because of the
assumed symmetry properties of the Euclidean distributions, as long as
the Euclidean time supports in the functions are ordered for one set
of Euclidean times, the permutation symmetry can be used to replace
them for one that is ordered as above.
One property of this representation of the Hilbert space, where the
inner products has a non-trivial kernel, is that distributions like
delta functions represent normalizable vectors.
For this to be a Hilbert space scalar product, this quantity must be
non-negative whenever $\{ \psi_m \} = \{\phi_n \}$. This condition is
called reflection positivity. In general there can be $0$-norm
vectors. The Hilbert space vectors are Cauchy sequences of
equivalence classes of vectors, where two vectors are in the same
class if the norm of their difference vanishes. This distinction will be
ignored in what follows.
For free particles, reflection positivity restricts the form of the
allowed distributions
\cite{Widder:1931}\cite{Widder:1934}\cite{Widder:1941}. They are
singular when the relative Euclidean coordinates vanish. The restriction picks a
domain where the scalar products are finite.
Because the Euclidean time reflection breaks the Euclidean invariance,
both Euclidean time translation and rotations in Euclidean space-time
planes are no longer unitary on this space. These transformations are
nevertheless defined on this space with restricted domains; they
represent translations in imaginary time and boosts with imaginary
rapidity. The infinitesimal forms of these elementary Euclidean
transformations can be used to construct both the Hamiltonian and
Lorentz boost generators.
The purpose of this work is to give a detailed discussion of how
relativistic invariance is realized in these theories. Rather than
consider a general set of Euclidean covariant kernels,
this work is limited to Euclidean representations of irreducible
representation of the Poincar\'e group \cite{Polyzou:2019a}. There are
two motivations for this. The first is that the kernels for these
representations are known, so it is possible to understand domain
issues related to the properties of the kernel and give explicit
representations for the Poincar\'e generators.
The second motivation
is that any unitary representation of the Poincar\'e group can be
decomposed into a direct integral of irreducible representations. In
a relativistic quantum theory these can be identified with the
complete set of one-body states plus multi-particle in or out
scattering states. These states either transform irreducibly or as
tensor products of irreducible representations. The kernel of a
general interacting model should be related to the direct integral of
irreducible kernels by a unitary transformation. The construction of
this direct integral from a general set of Euclidean covariant
distributions is the relativistic analog of diagonalizing the
Hamiltonian in non-relativistic quantum mechanics. This will not be
considered in this work.
In the next section the Poincar\'e group and its relation to the
Euclidean group is discussed. In section three unitary
representations of the Poincar\'e group are discussed, along with
structure of positive mass irreducible representations. Section 4
contains explicit forms of Euclidean covariant kernels of irreducible
representations of the Poincar\'e group for any mass and spin. They
are shown to be reflection positive. Explicit forms for all the
Poincar\'e generators are constructed, commutation relations are
verified, and the generators are shown to be symmetric with respect
the inner product with the Euclidean time reflection. Section 5
discuss the self-adjointness of the Hamiltonian and rotationless boost
generators. Section 6 has a brief discussion of finite Poincar\'e
transformations. The results are summarized in section 7.
\section{Background}
The Poincar\'e group is the group of space-time transformations that
relate different inertial reference frames in the theory of special relativity.
It is the symmetry group that preserves the proper time $\tau_{ab}$,
or proper distance, $d_{ab}$, between any two events with space-time
coordinates $x_a^{\mu} , x_b^{\mu}$
\begin{equation}
-\tau_{ab}^2=d_{ab}^2=\eta_{\mu\nu}(x_a-x_b)^{\mu}(x_a-x_b)^{\nu},
\label{b.1}
\end{equation}
where $\eta_{11}=\eta_{22}=\eta_{33}=-\eta_{00}=1$, $\eta_{\mu\nu}=0$ for
$\mu \not=\nu$ is the Minkowski metric tensor. Repeated indices
are assumed to be summed.
The most general point transformation,
$x'^\mu=f^\mu(x)$ satisfying (\ref{b.1}) has the form
\begin{equation}
x^{\mu}\to x'^{\mu}=\Lambda^{\mu}{}_{\nu} x^\nu+a^\mu
\label{b.2}
\end{equation}
where $\Lambda^{\mu}{}_{\nu}$ is a Lorentz transformation satisfying
\[
\eta_{\mu\nu}=\Lambda^{\alpha}{}_\mu\eta_{\alpha\beta} \Lambda^{\beta}{}_\nu
\]
or in matrix form
\begin{equation}
\eta=\Lambda^t\eta\Lambda .
\label{b.3}
\end{equation}
Equations (\ref{b.2}) and (\ref{b.3}) are relativistic generalizations of the
fundamental theorem of rigid body motion, which asserts that any motion
that preserves the distance between points in a rigid-body in a
composition of an orthogonal transformation and a translation.
The full Poincar\'e group contains discrete transformations that are
not associated with special relativity.
Equation (\ref{b.3}) implies that
\begin{equation}
\mbox{det}(\Lambda)^2=1 \qquad \mbox{and} \qquad (\Lambda_0^0)^2=
1+\sum_{i}(\Lambda_i^0)^2 .
\label{b.4}
\end{equation}
This means that the Lorentz group can be decomposed
into four topologically disconnected components
\begin{itemize}
\item
$\mbox{det}(\Lambda)=1, \qquad(\Lambda_0^0) \geq 1$; includes identity
\item
$\mbox{det}(\Lambda)=-1, \qquad(\Lambda_0^0) \geq 1$; includes space reflection
\item
$\mbox{det}(\Lambda)=-1, \qquad (\Lambda_0^0) \leq-1 $; includes time reversal
\item
$\mbox{det}(\Lambda)=1, \qquad(\Lambda_0^0) \leq -1$; includes space-time reversal .
\end{itemize}
Since the discrete symmetries of space reflection and time reversal
are not symmetries of the weak
interaction, the symmetry group associated with special relativity is
normally considered to be the subgroup of Poincar\'e transformations
that is continuously connected to the identity. This subgroup contains
the active transformations that can be experimentally realized.
The relation between the Lorentz group and the
four-dimensional orthogonal group can be understood by
expressing Minkowski, $x^{\mu}$, and Euclidean, $x_e^{\mu}$, four
vectors as $2\times 2$ matrices:
\begin{equation}
X_m=x^\mu\sigma_\mu=
\begin{pmatrix}
x^0+x^3 &x^1-ix^2 \\
x^1+ix^2 &x^0-x^3
\end{pmatrix}\; \qquad x^\mu=\frac{1}{2}\mbox{Tr}(X\sigma_\mu)
\label{b.5}
\end{equation}
\begin{equation}
X_e=x_e^\mu\sigma_{e\mu}=
\begin{pmatrix}
ix_e^0+x^3 &x^1-ix^2 \\
x^1+ix^2 &ix_e^0-x^3
\end{pmatrix}\; \qquad x_e^\mu=\frac{1}{2}\mbox{Tr}(X_e\sigma^{\dagger}_{e\mu}).
\label{b.6}
\end{equation}
where $\sigma_i=\sigma_{ei}$ are the Pauli matrices, $\sigma_0$ is the identity and
$\sigma_{e0}= i \sigma_0$. The determinants of these matrices are
related to the Minkowski and Euclidean line elements respectively:
\begin{equation}
\mbox{det}(X_m) = (x^0)^2 - \mathbf{x}\cdot \mathbf{x}
\qquad
\mbox{det}(X_e) = - \left ( (x_e^0)^2 + \mathbf{x}\cdot \mathbf{x}
\right ).
\label{b.7}
\end{equation}
The linear transformations that preserve the determinant and hermiticity of
$X_m$ have the form
\begin{equation}
X_m \to X_m' = \pm A X_m A^{\dagger}
\qquad \mbox{det} (A)=1.
\label{b.8}
\end{equation}
The (-) sign represents a space-time reflection, which is not
considered part of the symmetry group of special relativity. The
group of complex $2 \times 2$ matrices with $\mbox{det}(A)=1$ is
$SL(2,\mathbb{C})$. Similarly linear transformations
corresponding to real four-dimensional orthogonal transformations have
the general form
\begin{equation}
X_e \to X_e' = A X_e B^t \qquad A,B \in SU(2).
\label{b.9}
\end{equation}
Transformations of the form
\begin{equation}
X_e \to X_e' = A X_e B^t \qquad
X_m \to X_m' = A X_m B^t
\label{b.10}
\end{equation}
with both $A$ and $B$ in $SL(2,\mathbb{C})$ preserve both the
Minkowski and Euclidean line elements. However they do not
preserve the reality of the four vectors. They represent complex
Lorentz or orthogonal transformations.
This shows that the covering group of both the complex Lorentz and
complex orthogonal group is
$SL(2,\mathbb{C}) \times SL(2,\mathbb{C})$. This means that the real
Lorentz group can be considered to be a subgroup of the complex
orthogonal group; similarly the real orthogonal group can be
considered to be a complex subgroup of the Poincar\'e group. The
relevant relation that will be exploited in this work is that
Euclidean rotations that involve a space and the Euclidean time
coordinate can be identified with Lorentz boosts with complex
rapidity.
For the full Poincar\'e group it is necessary to include translations.
Euclidean time translations by $\tau$ are identified with Minkowski
time translations with $t=-i\tau$.
\section{Unitary representations of the Poincar\'e group}
In this section Poincar\'e group elements are labeled by $(\Lambda,A)$
where $\Lambda$ is a $SL(2,\mathbb{C})$ matrix and $A$ is a $2 \times 2$
Hermitian matrix representing a translation. In this representation
Poincar\'e transformations have the form
\begin{equation}
X' = \Lambda X \Lambda^{\dagger} +A
\label{c.1}
\end{equation}
where the group multiplication law is
\begin{equation}
(\Lambda_2 , A_2) (\Lambda_1 , A_1)= (\Lambda_2 \Lambda_1 , \Lambda_2
A_1 \Lambda_2^{\dagger} + A_2).
\label{c.2}\end{equation}
Four vector representations of these equations are
\begin{equation}
x^{\mu \prime} = \Lambda^{\mu}{}_{\nu} x^{\nu} + a^{\mu}
\label{c.3}
\end{equation}
\begin{equation}
(\Lambda^{\mu}{}_{\nu},a^{\mu} ) =
(\Lambda_2^{\mu}{}_{\alpha} \Lambda_1^{\alpha}{}_{\nu},
\Lambda_2^{\mu}{}_{\alpha}a_1^{\alpha} + a_2^{\mu}),
\label{c.4}
\end{equation}
where the
four vector and $2\times 2$ representations are related by
\begin{equation}
a^{\mu}:= {1 \over 2} \mbox{Tr}(\sigma_{\mu}A)
\qquad
\Lambda^{\mu}{}_{\nu} := {1 \over 2} \mbox{Tr}(\sigma_{\mu}\Lambda \sigma_{\nu} \Lambda^{\dagger}).
\label{c.5}
\end{equation}
$SL(2,\mathbb{C})$ is a six parameter group. It has six independent
one-parameter subgroups
\begin{equation}
\Lambda_r (\pmb{\theta})= e^{{i \over 2 }\pmb{\sigma}\cdot \pmb{\theta}}
\qquad
\Lambda_b (\pmb{\rho}) = e^{{1 \over 2 }\pmb{\sigma}\cdot \pmb{\rho}}
\label{c.6}
\end{equation}
corresponding to rotations about three different axes and rotationless
Lorentz boosts in three different directions. In these expressions
$\pmb{\theta}$ represents the angle and axis of a rotation while
$\pmb{\rho}$ represents the rapidity and direction of a rotationless boost.
The polar decomposition expresses a general $SL(2,\mathbb{C})$ matrix $\Lambda$
as a product of a rotation ($\Lambda_r$ unitary) followed by
rotationless boost ($\Lambda_b$ positive Hermitian):
\begin{equation}
\Lambda = \Lambda_{b}\Lambda_{r}
\label{c.7}
\end{equation}
where
\begin{equation}
\Lambda_{b} :=
(\Lambda \Lambda^{\dagger})^{1/2} = \Lambda_b (\pmb{\rho})
\qquad
\Lambda_{r} :=
(\Lambda \Lambda^{\dagger})^{-1/2}\Lambda =\Lambda_r (\pmb{\theta}).
\label{c.8}
\end{equation}
A unitary representation of the Poincar\'e group (inhomogeneous
$SL(2,\mathbb{C})$) is a set of unitary operators
$U(\Lambda ,A)$, labeled by elements of $SL(2,\mathbb{C}$)
satisfying
\begin{equation}
U(\Lambda_2,A_2) U(\Lambda_1,A_1)=
U(\Lambda_2 \Lambda_1,\Lambda_2 A_1 \Lambda_2^{\dagger} +A_2)
\label{c.9}
\end{equation}
\begin{equation}
U(I,0) =I
\label{c.10}
\end{equation}
\begin{equation}
U^{\dagger}(\Lambda ,A) = U^{-1}(\Lambda ,A)=
U (\Lambda^{-1} ,-\Lambda^{-1} A (\Lambda^{\dagger})^{-1} ) .
\label{c.11}
\end{equation}
The Poincar\'e group is a 10 parameter group. Infinitesimal generators
are the 10 self-adjoint operators defined by
\begin{equation}
H= i {d \over da^0} U(I,a^0\sigma_0)_{\vert_{a_0=0}}
\label{c.12}
\end{equation}
\begin{equation}
P^i= - i {d \over da^j} U(I,a^j\sigma_j)_{\vert_{a_i=0}}
\label{c.13}
\end{equation}
\begin{equation}
J^i= - i {d \over d\theta} U(e^{i{\theta\over 2} \sigma_j},0)_{\vert_{\theta=0}}
\label{c.14}
\end{equation}
\begin{equation}
K^i= - i {d \over d\rho} U(e^{{\rho\over 2} \sigma_j},0)_{\vert_{\rho=0}}
\label{c.15}
\end{equation}
where there is no sum in (\ref{c.13}) over the repeated $j$, and
$j\in \{1,2,3\}$ in (\ref{c.13}-\ref{c.15}).
The group representation property (\ref{c.9}) implies that these generators
satisfy commutation relations
\begin{equation}
[J^i ,J^j] = i \epsilon_{ijk} J^k
\qquad
[J^i ,P^j] = i \epsilon_{ijk} P^k
\qquad
[J^i ,K^j] = i \epsilon_{ijk} K^k
\label{c.16}
\end{equation}
\begin{equation}
[K^i ,K^j] = -i \epsilon_{ijk} J^k
\qquad
[J^i ,H] = 0
\qquad
[P^i,H] = 0
\label{c.17}
\end{equation}
\begin{equation}
[K^j,H] =i P^j
\qquad
[K^i ,P^j] = i \delta_{ij}H .
\label{c.18}
\end{equation}
These operators are components of a four vector, $P^{\mu}$, and an anti-symmetric
tensor operator, $J^{\mu \nu}$,
\begin{equation}
P^{\mu} = (H,\mathbf{P})
\qquad
J^{\mu \nu} =
\left (
\begin{array}{cccc}
0 & -K^1 & -K^2 & -K^3 \\
K^1 & 0 & J^3 & -J^2 \\
K^2 & -J^3 & 0 & J^1 \\
K^3 & J^2 & -J_1 & 0\\
\end{array}
\right ) .
\label{c.19}
\end{equation}
There are two independent polynomial invariants
\begin{equation}
M^2 = (P^0)^2 -\mathbf{P}^2 = - P^{\mu}P_{\mu}
\label{c.20}
\end{equation}
and
\begin{equation}
W^2 = W^{\mu}W_{\mu}
\qquad
W^{\mu} = -{1 \over 2} \epsilon^{\mu \nu \alpha \beta} P_{\nu}J_{\alpha \beta}.
\label{c.21}
\end{equation}
where $W^{\mu}$ is called the Pauli-Lubanski vector.
When $M\not=0$ the spin is defined by
\begin{equation}
S^2 = W^2/M^2 .
\label{c.22}
\end{equation}
A spin vector can be defined by an {\it operator}
rotationless (canonical)
boost that
transforms the angular momentum tensor to the rest frame:
\begin{equation}
s^i = \epsilon_{ijk}
\Lambda_c^{-1} (P)^{j}{}_{\mu}
\Lambda_c^{-1} (P)^{k}{}_{\nu}J^{\mu \nu}
\label{c.23}
\end{equation}
where
\begin{equation}
\Lambda_c (P)^{\mu}{}_{\nu} =
\left (
\begin{array}{cc}
V^0 & \mathbf{V} \\
\mathbf{V} & I + {\mathbf{V}\otimes \mathbf{V} \over 1+ V^0} \\
\end{array}
\right )
\qquad V^{\mu}= P^{\mu}/M
\label{c.24}
\end{equation}
and $P^{\mu}$ and $M$ are
considered operators. Note that $\Lambda_c (p) = \Lambda_b (\pmb{\rho})$
with
\begin{equation}
\mathbf{V} = \mathbf{P}/M = \hat{\pmb{\rho}} \sinh (\rho).
\end{equation}
This spin vector is called the canonical spin; other types
of spin vectors (helicity, light-front spin) are
related to the canonical spin by momentum
dependent rotations. For the purpose of this work it is sufficient to
consider the canonical spin.
The canonical spin can also be expressed in terms of the Pauli Lubanski vector:
\begin{equation}
\left
(
\begin{array}{c}
0 \\
\mathbf{s}_c \\
\end{array}
\right ) =
- {1 \over 2M } \Lambda_c^{-1} (P)^{\mu}{}_{\nu} W^{\mu}.
\label{c.25}
\end{equation}
The components of the spin satisfy $SU(2)$ commutations relations:
\begin{equation}
[s_i,s_j] = i \epsilon_{ijk} s^k .
\label{c.27}
\end{equation}
With these definitions, for $M>0$, $M^2,s^2,\mathbf{P},s_z$ are a
maximal set of commuting self-adjoint functions of the Poincar\'e
generators. The spectrum of each component of $\mathbf{P}$ is the
real line since each component of $\mathbf{P}$ can be boosted to any
value. Similarly the spectrum of spins are restricted to be integral
or half integral as a consequence of the $SU(2)$ commutations
relations. In a general system these commuting observables are not
complete; they can be supplemented by additional Poincar\'e-invariant
degeneracy quantum numbers, which will be denoted by $\alpha$. A
basis for the Hilbert space are the simultaneous eigenstates of
$M,S^2,\alpha,\mathbf{P},s_z $,
\begin{equation}
\vert (m,s,\alpha)
\mathbf{p},\mu \rangle .
\label{c.28}
\end{equation}
Because these vectors are constructed out of eigenvalues of functions of
$P^{\mu}$ and $J^{\mu \nu}$, which have well-defined the Poincar\'e
transformation properties, the Poincar\'e transformation properties
of these basis state follow from the definitions
\begin{equation}
U(\Lambda,a)\vert(m,s,\alpha) p,\mu \rangle =e^{i\Lambda p\cdot a}\vert
(m,s,\alpha)\Lambda p,\nu
\rangle
D_{\nu\mu}^{j}[R_{ws}(\Lambda,p)] \sqrt{\frac{\omega_{m}(\Lambda p)}{\omega_{m}(p)}}
\label{c.29}
\end{equation}
where, $R_{cw}(\Lambda,p):=\Lambda_c^{-1}(\Lambda p)\Lambda \Lambda_c(p)$ is the
canonical-spin Wigner rotation, $\Lambda_c(p)= e^{{1\over 2}\pmb{\rho}\cdot \pmb{\sigma}}$
where $\pmb{\rho}$ is the rapidity of a particle of mass $m$ and momentum
$\mathbf{p}$,
and $\omega_{m}(p):=\sqrt{m^2+\mathbf{p}^2}$ is the energy of the particle.
The Wigner $D$-function is the finite dimensional unitary representation of the
rotation group in the $\vert s,\mu\rangle $ basis \cite{rose}:
\[
D_{m,m'}^{s}[R]=\langle s,\mu\vert U(R) \vert s,\mu'\rangle =
\]
\[
\sum_{k=0}^{s+\mu}\frac{\sqrt{(s+\mu)!(s+\mu')!(s-\mu)!(s-\mu')!}}{k!
(s+\mu'-k)!(s+\mu-k)!(k-\mu-\mu')!}R_{++}^{k}R_{+-}^{s+\mu'-k}
R_{-+}^{s+\mu-k}R_{--}^{k-\mu-\mu'}
\]
where
\begin{equation}
R=
\left (
\begin{array}{cc}
R_{++} & R_{+-}\\
R_{-+} & R_{--}
\end{array}
\right ) = e^{{i\over 2}\pmb{\theta}\cdot \pmb{\sigma}}=
\sigma_0 \cos ({\theta \over 2}) + i \hat{\pmb{\theta}}\cdot \pmb{\sigma}
\sin ({\theta \over 2})
\label{c.30}
\end{equation}
is a $SU(2)$ matrix. Because $D^s_{\mu \nu}[R]$ is a degree $2s$
polynomial in the matrix elements of $R$, and
$R=e^{i {\pmb{\theta}\cdot \pmb{\sigma}\over 2}}$ is an entire function of the
angles, $\pmb{\theta}$, it follows that
$D_{\mu,\mu'}^{s}[e^{i {\pmb{\theta}\cdot \pmb{\sigma}\over 2}}]$ is
an entire function of all three components of $\pmb{\theta}$. This means
that the group representation property
\begin{equation}
\sum_{\mu''} D_{\mu,\mu''}^{s}[R_2] D_{\mu'',\mu'}^{s}[R_1] -
D_{\mu,\mu'}^{s}[R_2R_1]=0,
\label{c.31}
\end{equation}
and the formulas for adding angular momenta
\begin{equation}
D_{\mu,\mu'}^{s}[R] -
\sum_{ \mu_1 \mu_2 \mu_1' \mu_2'}
\langle s,\mu\vert s_1, \mu_1, s_2, \mu_2 \rangle
D_{\mu_1,\mu_1'}^{s_1}[R] D_{\mu_2,\mu_2'}^{s_2}[R]
\langle s_1, \mu_1', s_2', \mu_2' \vert s , \mu'\rangle =0
\label{c.32}
\end{equation}
and
\begin{equation}
D_{\mu_1,\mu_1'}^{s_1}[R] D_{\mu_2,\mu_2'}^{s_2}[R] -
\sum_{s \mu \mu'} \langle s_1, \mu_1, s_2, \mu_2 \vert s , \mu\rangle
D_{\mu,\mu'}^{s}[R]
\langle s,\mu'\vert s_1, \mu_1', s_2, \mu_2' \rangle =0,
\label{c.33}
\end{equation}
which hold for real angles, can be analytically continued to complex
angles. This means that (\ref{c.31}-\ref{c.33}) also hold when the
$SU(2)$ matrices $R$ are
replaced by $SL(2,\mathbb{C})$ matrices.
In these expressions, $\langle s,\mu\vert s_1, \mu_1, s_2,
\mu_2 \rangle$, are $SU(2)$ Clebsch-Gordan coefficients. While the
analytic continuation preserves the group representation and angular
momentum addition properties, it does not preserve unitarity.
\section{Euclidean Formulation}
The common property of any relativistic quantum theory is that it can
be decomposed into a direct integral of irreducible representations.
The structure of irreducible representations of the Poincar\'e group
in the Euclidean representation can be understood by starting with
Minkowski-space irreducible representations of the Poincar\'e group.
This work considers only positive-mass positive-energy
representations. These can be expressed in a basis of simultaneous
eigenstates of the mass, spin, linear momentum and $z$-component of
the canonical spin. The action of the unitary representation of the
Poincar\'e group on this basis is given by (\ref{c.29}).
This is unitary for basis vectors with the normalization:
\begin{equation}
\langle (m',s')p',\mu'
\vert (m,s) p,\mu \rangle = \delta_{m'm}\delta_{s's} \delta
(\mathbf{p}'-\mathbf{p}) \delta_{\mu' \mu}.
\label{d.1}
\end{equation}
Because of the unitarity of $R_{wc}(\Lambda,p)$, the $SU(2)$
Wigner rotation can be
expressed in two equivalent ways:
\begin{equation}
R_{wc}(\Lambda,p) =
\Lambda_c^{-1} (\Lambda p) \Lambda \Lambda_c (p) = \Lambda_c^{\dagger}
(\Lambda p) (\Lambda^{\dagger})^{-1} \Lambda_c^{\dagger -1} (p) .
\label{d.2}
\end{equation}
The $SL(2,\mathbb{C})$ group representation property (\ref{c.31})
implies that the unitary representation of the Wigner rotation can be
factored into a product of three finite-dimensional representations of
$SL(2,\mathbb{C})$ in two different ways:
\begin{equation}
D_{\nu\mu}^{s}[R_{wc}(\Lambda,p)]=
\sum_{\alpha \beta} D_{\nu\alpha}^{s}[\Lambda_c^{-1} (\Lambda p)]
D_{\alpha\beta}^{s}[\Lambda] D_{\beta\mu}^{s}[\Lambda_c (p)]
\label{d.3}
\end{equation}
or
\begin{equation}
D_{\nu\mu}^{s}[R_{wc}(\Lambda,p)]= \sum_{\alpha \beta}
D_{\nu\alpha}^{s}[\Lambda_c^{\dagger} (\Lambda p)]
D_{\alpha\beta}^{s}[(\Lambda^{\dagger})^{-1}]
D_{\beta\mu}^{s}[(\Lambda_c^{\dagger})^{-1} (p)] .
\label{d.4}
\end{equation}
These relations can be used to rewrite equation (\ref{c.29}) in
terms of new Lorentz covariant basis states:
\[
U(\Lambda,a) \underbrace{ \sum_\alpha
\vert (m,j) p,\alpha \rangle
D_{\alpha\mu}^{s}[\Lambda_b^{-1} (p)] \sqrt{\omega_m(p)}}_
{\vert (m,j) p,\mu \rangle_{cov}}
=
\]
\begin{equation}
e^{i\Lambda p\cdot a}\sum_\beta
\underbrace{\sum_\alpha \vert (m,j)\Lambda p,\alpha \rangle
D_{\alpha\beta}^{s}[\Lambda_c^{-1} (\Lambda p)]
\sqrt{\omega_{m}(\Lambda p)}}_
{\vert (m,j) \Lambda p,\beta \rangle_{cov}}
D_{\beta\mu}^{s}[\Lambda]
\label{d.5}
\end{equation}
or
\[
U(\Lambda,a) \underbrace{\sum_\alpha
\vert (m,j) p,\alpha \rangle
D_{\alpha\mu}^{s}[\Lambda_c^{-\dagger} (p)] \sqrt{\omega_m(p)}}_
{\vert (m,j) p,\mu \rangle_{cov*}}
=
\]
\begin{equation}
e^{i\Lambda p\cdot a}\sum_\beta
\underbrace{\sum_\alpha \vert (m,j)\Lambda p,\alpha \rangle
D_{\alpha\beta}^{s}[\Lambda_c^{\dagger} (\Lambda p)]
\sqrt{\omega_{m}(\Lambda p)}}_
{\vert (m,j) \Lambda p,\beta \rangle_{cov*}}
D_{\beta\mu}^{s}[(\Lambda^{\dagger})^{-1}].
\label{d.6}
\end{equation}
These expressions replace the states (\ref{c.28}) that transform
covariantly with respect to the Poincar\'e group with states that
transform covariantly with respect to $SL(2,\mathbb{C})$. The
transformations relating the Lorentz and Poincar\'e covariant
representations are
invertible, however there are two distinct Lorentz covariant
representations, because while $R=(R^{\dagger})^{-1}$ for $R\in
SU(2)$, the corresponding representations in $SL(2,\mathbb{C})$ are
inequivalent. These two representations are called right and left
handed representations for reasons that will become apparent.
In the Lorentz covariant representations, (\ref{d.5}) and (\ref{d.6}),
this equivalence can be used to show that the Hilbert space inner
product of two $SL(2,\mathbb{C})$ covariant wave functions has a
non-trivial kernel
\[
\langle \psi \vert \phi \rangle =
\sum_\mu \int
\langle \psi \vert (m,j) p,\mu \rangle
d\mathbf{p} \langle (m,j) p,\mu \vert \phi \rangle=
\]
\begin{equation}
\int \sum_{\mu \nu} \langle \psi \vert (m,j) p,\mu \rangle_{cov}{}
D_{\mu\nu}^{j}[p\cdot\sigma]
2 \delta (p^2+m^2)\theta (p^0) d^4 p{}
_{cov}\langle (m,j) p,\nu \vert \phi \rangle
\label{d.7}
\end{equation}
\[
\langle \psi \vert \phi \rangle =
\int \sum_\mu
\langle \psi \vert (m,j) p,\mu \rangle
d\mathbf{p} \langle (m,j) p,\mu \vert \phi \rangle=
\]
\begin{equation}
\int \sum_{\mu \nu} \langle \psi \vert (m,j) p,\mu \rangle_{cov*}{}
D_{\mu\nu}^{j}[\Pi p\cdot\sigma]
2 \delta (p^2+m^2)\theta (p^0) d^4p{}
_{cov*}\langle (m,j) p,\nu \vert \phi \rangle
\label{d.8}
\end{equation}
where $\Lambda_c (p) \Lambda^{\dagger}_c (p)= {\sigma}\cdot {p}$ and
$\Lambda^{-1}_c (p) (\Lambda^{\dagger}_c)^{-1}(p) = \Pi {p} \cdot {\sigma}$,
was used in these equations.
$\Pi$ is the space reflection
operator and $p\cdot\sigma = \omega_m(p)\sigma_0 + \mathbf{p}\cdot
\pmb{\sigma}$. These equations explain why
(\ref{d.7}) and (\ref{d.8}) are called right and left handed
representations. These kernels are, up to normalization,
spin-$s$ two-point Wightman functions \cite{Wightman:1980}.
The motivation for considering these $SL(2,\mathbb{C})$ covariant
representations is that they are naturally related to the
corresponding Euclidean covariant representations.
To show this let $f(x_e,\mu)$ and $g(y_e,\nu)$ be functions of
Euclidean space-time variables and spins with positive Euclidean-time support.
Consider the following Euclidean covariant kernel:
\begin{equation}
S^s_{e}(x_e,\mu;y_e,\nu):=
\int d^4p
{2 \over (2 \pi)^4} {e^{i p_e \cdot (x_e -y_e)}
\over p_e^2 + m^2}D^s_{\mu \nu}(p_e \cdot \sigma_e) .
\label{d.9}
\end{equation}
The physical Hilbert space inner product (\ref{a.4}) for
this Euclidean Green's function
has the form
\[
\int\sum_{\mu \nu} d^4x_ed^4y_e f^*(\theta x_e,\mu)
S^s_{e}(x_e,\mu;y_e,\nu)g(y_e,\nu) =
\]
\[
\int\sum_{\mu \nu} d^4p_e f^*(\theta x_e,\mu)
{2 \over (2 \pi)^4} {e^{i p_e \cdot (x_e -y_e)}
\over p_e^2 + m^2}D^s_{\mu \nu}(p_e \cdot \sigma_e)
g(y_e,\nu) =
\]
\begin{equation}
\int\sum_{\mu\nu} \psi^* (\mathbf{p},\mu) {d \mathbf{p} \over \omega_m (\mathbf{p})}
D^s_{\mu \nu}(p \cdot \sigma) \phi(\mathbf{p},\nu)
\label{d.10}
\end{equation}
where
\begin{equation}
\psi^*(\mathbf{p},\mu):=
{1 \over (2 \pi)^{3/2}} \int d\mathbf{x}d\tau e^{i \mathbf{p} \cdot \mathbf{x} - \omega_m (\mathbf{p}) \tau} f^*(\mathbf{x},\tau,\mu)
\label{d.11}
\end{equation}
and
\begin{equation}
\phi(\mathbf{p},\nu):=
{1 \over (2 \pi)^{3/2}} \int d\mathbf{x}d\tau e^{-i \mathbf{p} \cdot \mathbf{x} - \omega_m (\mathbf{p}) \tau} g(\mathbf{x},\tau,\nu).
\label{d.12}
\end{equation}
The Euclidean time-support condition ensures that the Laplace
transforms with respect to the Euclidean times in (\ref{d.11}) and
(\ref{d.12}) are well defined. The resulting kernel in (\ref{d.10})
is identical to the covariant kernel in (\ref{d.7}) after performing
the integrals over the $p^0_e$.
This shows that the ``Euclidean'' inner product
(\ref{d.10}) can be identified with the corresponding Lorentz covariant
inner product, which itself is identical to the original Poincar\'e
covariant inner product.
This means that
\begin{equation}
S^s_r(x_e, \mu ; y_e, \nu):= \int
{2 d^4p \over (2 \pi)^4} {e^{i p_e \cdot (x_e -y_e)}
\over p_e^2 + m^2}D^s_{\mu \nu}(p_e \cdot \sigma_e)
\label{d.13}
\end{equation}
is a Euclidean covariant reflection positive kernel for right
handed representations of mass $m$ and spin $s$ respectively. The corresponding
kernel for left-handed representations is
\begin{equation}
S^s_l(x_e, \mu ; y_e, \nu):= \int
{2 d^4 p\over (2 \pi)^4} {e^{i p_e \cdot (x_e -y_e)}
\over p_e^2 + m^2}D^s_{\mu \nu}(\Pi p_e \cdot \sigma_e).
\label{d.14}
\end{equation}
Space reflection interchanges right and left-handed representations.
The space reflection operator does not commute with the Euclidean
covariant kernel. This implies that space reflected states will not
transform correctly under Lorentz transformations in these Lorentz
covariant representations. Kernels for systems that allow a linear
representation of space reflection can be constructed by taking direct
sums of right and left handed kernels.
The kernels (\ref{d.13}-\ref{d.14}) can be evaluated analytically
using the methods in
\cite{bogoliubov}.
The results are
\[
S^s_{r} (z_e, \mu , \nu) :=
{2 \over (2\pi)^4} \int {d^4p \over p_e^2 + m^2}
D^s_{\mu \nu}(p_e \cdot \sigma_e)
e^{i p \cdot z_e} =
\]
\begin{equation}
D^s_{\mu \nu}(-i \nabla_{ze} \cdot \sigma_e)
{2m^2
\over (2 \pi)^2}{K_1 (m\sqrt{z_0^2 +\mathbf{z}^2})
\over m\sqrt{z_0^2 +\mathbf{z}^2}}
\label{d.15}
\end{equation}
\[
S^s_{l} (z_e ,\mu , \nu) :=
{2 \over (2\pi)^2} \int {d^4p \over p_e^2 + m^2}
D^s_{\mu \nu}(\Pi p_e \cdot \sigma_e)e^{i p_e \cdot z_e} =
\]
\begin{equation}
D^s_{\mu \nu}(-i \Pi \nabla_{ze} \cdot \sigma_e)
{2m^2
\over (2 \pi)^2}{K_1 (m\sqrt{z_0^2 +\mathbf{z}^2})
\over m\sqrt{z_0^2 +\mathbf{z}^2}}
\label{d.16}
\end{equation}
where $z_e = x_e-y_e$.
Note that ${K_1 (\eta)
\over \eta}$ behaves like $1/\eta^2$ near the origin.
Since
$D^s_{\mu \nu}(-i \nabla_{ze} \cdot \sigma_e )$
is a degree $2s$ polynomial in $-i \nabla_{ze}$, these kernels
have power law singularities at the origin, but fall off exponentially
for large values of $z_e^2$, The restriction of the support of the
vectors to positive Euclidean time ensures that $z_e^2 >0$, so the
singularity at $z_e=0$ never causes a problem. These Green's
functions are reflection positive on this space.
This is because $D^s_{\mu \nu}(p\cdot \sigma)
$ factors into a product of a matrix and its adjoint:
\begin{equation}
D^s_{\mu \nu}(p\cdot \sigma)
= \sum_{\alpha}D^s_{\mu\alpha}(\Lambda_c(p))D^s_{\alpha\nu}(\Lambda_c(p))^{\dagger} .
\label{d.17}
\end{equation}
The treatment of relativity follows from the relation between
the four dimensional Euclidean group and the associated complex
subgroup of the Lorentz group discussed in section 2. Consider the two matrices
\begin{equation}
p \cdot \sigma :=
\left (
\begin{array}{cc}
p^0+p^2 & p^1 -i p^2\\
p^1+i p^2 & p^0 - p^3\\
\end{array}
\right)
\qquad
p_e \cdot \sigma_e :=
\left (
\begin{array}{cc}
i p_e^0+p_e^2 & p_e^1 -i p_e^2\\
p_e^1+i p_e^2 & i p_e^0 - p_e^3\\
\end{array}
\right).
\label{d.18}
\end{equation}
The $SL(2,\mathbb{C})\times SL(2,\mathbb{C})$ transformation properties of
these matrices (denoted by $P$) are
\begin{equation}
P\to P' = A P B^t .
\label{d.19}
\end{equation}
The associated complex $4 \times 4 $ Lorentz and four-dimensional orthogonal
transformation
matrices are
\begin{equation}
\Lambda (A,B)^{\mu}{}_{\nu} = {1 \over 2} \mbox{Tr}(\sigma_{\mu}A\sigma_{\nu}B^t)
\qquad
{\cal O} (A,B)^{\mu}{}_{\nu} = {1 \over 2} \mbox{Tr}(\sigma_{e\mu}^{\dagger}A\sigma_{e\nu}B^t) .
\end{equation}
For ordinary rotations
$A=B^*=e^{i{\lambda \over 2} \hat{\mathbf{n}}}$.
For rotations about the $\hat{\mathbf{z}}$ axis
\begin{equation}
{\cal O}(A,A^*)( \lambda) =
\left (
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & \cos(\lambda)&\sin (\lambda)&0\\
0 & -\sin(\lambda)&\cos (\lambda)&0\\
0 & 0 & 0& 1\\
\end{array}
\right )
\label{d.20}
\end{equation}
and
\begin{equation}
\Theta {\cal O}(A,A^*)( \lambda) \Theta = {\cal O}(A,A^*)( \lambda).
\label{d.21}
\end{equation}
For real rotations in Euclidean space-time planes $A=B^t=e^{i{\lambda \over 2} \hat{\mathbf{n}}\cdot \pmb{\sigma}}$.
For the case of the $x^0_e-\hat{\mathbf{z}}$ plane
\begin{equation}
{\cal O}(A,A^t)( \lambda) x=
\left (
\begin{array}{cccc}
\cos(\lambda) & 0 & 0 & \sin (\lambda) \\
0 & 1 & 0 &0\\
0 & 0 & 1 &0\\
-\sin (\lambda) & 0 & 0& \cos(\lambda)\\
\end{array}
\right )
\label{d.22}
\end{equation}
\and
\begin{equation}
\Theta {\cal O}^t(A,A^t)( \lambda) \Theta = {\cal O}(A,A^t)( \lambda).
\label{d.23}
\end{equation}
While ordinary 3-dimensional rotations are the same for
$p\cdot \sigma$ or $p_e \cdot \sigma_e$, real rotations in
Euclidean space time planes are interpreted as rotationless Lorentz boosts
with imaginary rapidity.
These identifications imply the following algebraic relations between the
infinitesimal generators of the four dimensional orthogonal group and
the Lorentz group:
\begin{equation}
\mathbf{P}_m = \mathbf{P}_e \qquad J^{ij}_{m} = J^{ij}_{e}
\label{d.24}
\end{equation}
\begin{equation}
H_m = i H_e \qquad K^{i}_{m} = -i J^{0i}_{e}
\label{d.25}
\end{equation}
Because of the factor of $i$, if the Euclidean generators are
self-adjoint operators on a representation of the Hilbert space, the
constructed Poincar\'e generators cannot be self-adjoint on that
representation of the Hilbert space.
In the spinless case ($s=0$) the identifications (\ref{d.20}-\ref{d.23})
result in the following expressions for the infinitesimal generators
of the Poincar\'e group on the Hilbert space with the time reflection:
\begin{equation}
H\Psi (x_e)= {\partial \over \partial x^0_e} \Psi (x_e)
\qquad
\mathbf{P}\Psi (x_e)= -i {\partial \over \partial \mathbf{x}_e } \Psi (x_e)
\label{d.26}
\end{equation}
\begin{equation}
\mathbf{J}\Psi (x_e)= -i \mathbf{x}
\times \pmb{\nabla}_x \Psi (x_e)
\qquad
K^j \Psi (x_e)=( x^j {\partial \over \partial x^0_e}-
x^0_e {\partial \over \partial x^j}) \Psi (x_e).
\label{d.27}
\end{equation}
It is straightforward to demonstrate that these operators satisfy
the Poincar\'e
commutations relations (\ref{c.16}-\ref{c.18}). For example
\begin{equation}
[K^i,H] = [
x^i {\partial \over \partial x_e^0} -
x^0_e {\partial \over \partial x^i},
{\partial \over \partial x^0_e}]=
i (-i {\partial \over \partial x^i}) =
i P^i
\label{d.28}
\end{equation}
which agrees with (\ref{c.18}). The other commutators can be checked similarly.
The Euclidean time reversal of the final state makes both the
Hamiltonian $H$ and the boost generators $\mathbf{K}$ formally
Hermitian with respect to the scalar product (\ref{d.7}). The
non-trivial observation is that even an infinitesimal rotation in a Euclidean
space time plane can map functions with positive Euclidean time support
to functions that violate the support condition. This maps
Hilbert space vectors out of the Hilbert space.
The resolution of this problem will be discussed in section 6.
To show the hermiticity of the rotationless boost generators
(\ref{d.27})
note that rotational invariance of the Euclidean Green's function in
Euclidean space-time planes means that the Euclidean rotation
generators commute with the Euclidean Green's function:
\begin{equation}
(-i x^i {\partial \over \partial x_e^0} +i
x^0_e {\partial \over \partial x^i})
S^0_e(x-y) =
S^0_e(x-y)
(-i y^i {\partial \over \partial y_e^0} +i
y^0_e {\partial \over \partial y^i}).
\label{d.29}
\end{equation}
Multiplying both sides by $i$ gives
\begin{equation}
( x^i {\partial \over \partial x_e^0} -
x^0_e {\partial \over \partial x^i})
S^0_e(x-y) =
S^0_e(x-y)
(y^i {\partial \over \partial y_e^0} -
y^0_e {\partial \over \partial y^i}).
\label{d.30}
\end{equation}
Next consider the inner product
\[
\langle f \vert K^i \vert g \rangle =
\]
\begin{equation}
\int d^4x d^4y f^* (\mathbf{x},-x_e^0 ) S^0_e(x-y)
(y^i {\partial \over \partial y_e^0} -
y^0_e {\partial \over \partial y^i}) g(\mathbf{y},y^0_e).
\label{d.31}
\end{equation}
Using (\ref{d.30}) in (\ref{d.31}) gives
\begin{equation}
= \int d^4x d^4y f^* (\mathbf{x},-x_e^0 )
(x^i {\partial \over \partial x_e^0} -
x^0_e {\partial \over \partial x^i})
S^0_e(x-y)
g(\mathbf{y},y^0_e).
\label{d.32}
\end{equation}
Integrating by parts again gives
\begin{equation}
= -\int d^4x d^4y
(x^i {\partial \over \partial x_e^0} +
x^0_e {\partial \over \partial x^i})
(\theta f)^* (\mathbf{x},x_e^0 )
S^0_e(x-y)
g(\mathbf{y},y^0_e).
\label{d.33}
\end{equation}
Finally factoring the time reversal out of $f$ gives
\begin{equation}
-(x^i {\partial \over \partial x_e^0} +
x^0_e {\partial \over \partial x^i}) \theta
f^* (\mathbf{x},x_e^0)=
\theta
\left ((x^i {\partial \over \partial x_e^0} -
x^0_e {\partial \over \partial x^i})
f^* (\mathbf{x},x_e^0) \right )
\label{d.34}
\end{equation}
which when used in (\ref{d.33}) gives
\[
\langle f \vert K^i \vert g \rangle =
\int d^4x d^4y f^* (\mathbf{x},-x_e^0 ) S^0_e(x-y)
(y^i {\partial \over \partial y_e^0} -
y^0_e {\partial \over \partial y^i}) g(\mathbf{y},y^0_e) =
\]
\begin{equation}
\int d^4x d^4y
\theta ((x^i {\partial \over \partial x_e^0} -
x^0_e {\partial \over \partial x^i}) f (\mathbf{x},x_e^0 ))^*
S^0_e(x-y)
g(\mathbf{y},y^0_e) =
\langle K^i f \vert g \rangle.
\label{d.35}
\end{equation}
This shows that $K^i$ is a Hermitian operator on this representation
of the Hilbert space.
The other non-trivial operator is the Hamiltonian (\ref{d.26}).
In this case
\[
\langle f \vert H \vert g \rangle =
\int d^4x d^4y f^* (\mathbf{x},-x_e^0 ) S^0_e(x-y)
{\partial \over \partial y_e^0}
g(\mathbf{y},y_e^0) =
\]
\[
-\int d^4x d^4y f^* (\mathbf{x},-x_e^0 )
{\partial \over \partial y_e^0}
S^0_e(x-y)
g(\mathbf{y},y^0_e) =
\int d^4x d^4y f^* (\mathbf{x},-x_e^0 )
{\partial \over \partial x_e^0}
S^0_e(x-y)
g(\mathbf{y},y^0_e) =
\]
\begin{equation}
-\int d^4x d^4y
{\partial \over \partial x_e^0}
f^* (\mathbf{x},-x_e^0 )
S^0_e(x-y)
g(\mathbf{y},y^0_e) =
\int d^4x d^4y
{\partial f^* \over \partial x^0}
(\mathbf{x},-x_e^0 )
S^0_e(x-y)
g(\mathbf{y},y^0_e) =
\langle Hf \vert g \rangle .
\label{d.36}
\end{equation}
The Euclidean time reversal does not change the linear or angular
momentum operators. These methods can be used to demonstrate that all
of the $s=0$ generators (\ref{d.26}-\ref{d.27}) are Hermitian in this
representation of the Hilbert space and satisfy the Poincar\'e Lie
algebra.
\section{Spin}
In this section explicit formulas for generators for particles with
arbitrary spin are derived, generalizing the method used in the
previous section for scalar particles.
In the original Poincar\'e covariant theory the spin is associated
with the observable that is the $\hat{\mathbf{z}}$-component of the
spin that would be
measured in the particle's rest frame if it was transformed to the
rest frame with a rotationless Lorentz transformation. The spin in
the covariant wave function is related to this spin by multiplying
by one of the $SL(2,\mathbb{C})$ matrices, $D^s_{\mu \nu}(\Lambda_c(p)^{-1})$ or
$D^s_{\mu \nu}(\Lambda_c(p)^{\dagger})$. These transformations lead to distinct
right or left handed spinors. In discussing spin it is important to
understand that the Poincar\'e covariant spinors and the Lorentz
covariant spinors are related, but different. Representations of the
Poincar\'e generators for each type of covariant spin must be
considered separately. In addition, for each type of covariant spinor
there are invariant linear functionals that define dual spinors. The
dual spinors are spinor analogs of covariant and contravariant
vectors. In conventional treatments
\cite{Wightman:1980}\cite{Wightman} \cite{berestetskii} the
right-handed spinors are denoted by $\xi^a$, left handed spinors are denoted
by $\xi^{\dot{a}}$ and their duals are denoted by $\xi_a$ and
$\xi_{\dot{a}}$ respectively.
The first step is to consider the $SL(2,\mathbb{C})$ transformation
properties of the Euclidean kernels for right and left handed
covariant spinors and their duals.
Euclidean four vectors can be represented by any of the four matrices:
\begin{equation}
p_e \cdot \sigma_e = p_e^\mu\sigma_{e\mu}
\qquad
p_e \cdot ( \sigma_2 \sigma_e \sigma_2 ) = p_e^\mu\sigma_2 \sigma_{e\mu} \sigma_2
\qquad
p_e \cdot \sigma^t_e = p_e^\mu\sigma^t_{e\mu}
\qquad
p_e \cdot ( \sigma_2 \sigma_e^t \sigma_2 )
= p_e^\mu\sigma_2 \sigma^t_{e\mu} \sigma_2 .
\label{s.1}
\end{equation}
The determinant of each of these matrices is (-) the square of the
Euclidean length of $p_e$, which is preserved under linear
transformations of the form
\begin{equation}
P' = APB^t
\label{s.2}
\end{equation}
where $P$ represents any of the matrices in (\ref{s.1}), and $A,B \in
SL(2,\mathbb{C})$. Real four-dimensional orthogonal transformations
are obtained by restricting $A$ and $B$ to be elements of $SU(2)$.
The $4 \times 4$ orthogonal matrix $\mathbb{O}(A,B)^{\mu}{}_{\nu}$
is related to the pair $(A,B)$ by
\begin{equation}
\mathbb{O}(A,B)^{\mu}{}_{\nu}:=\frac{1}{2}Tr(\sigma_{e\mu}^\dagger A\sigma_{e\nu}B^t).
\label{s.3}
\end{equation}
It follows that
\begin{equation}
A p_e^{\mu}\sigma_{e\mu} B^t= \sigma_{e\mu}\mathbb{O}(A,B)^{\mu}{}_{\nu}p_e^{\nu}
= \sigma_{e\mu} (\mathbb{O}(A,B)p_e)^{\mu}.
\label{s.4}
\end{equation}
Multiplying
(\ref{s.4}) by $\sigma_2$ on both sides gives
\begin{equation}
A^*(p_e \cdot \sigma_2\sigma_e\sigma_2)B^\dagger=(\mathbb{O}(A,B)p)_e\cdot\sigma_2\sigma_e\sigma_2 .
\label{s.5}
\end{equation}
Taking transposes of the $2 \times 2$ matrices (\ref{s.4}) and (\ref{s.5})
give
\begin{equation}
B(p_e\sigma_e^t)A^t= \sigma_e^{t}\cdot(\mathbb{O}(A,B)p_e)
\label{s.6}
\end{equation}
and
\begin{equation}
B^*(p_e \cdot \sigma_2\sigma_e^{t}\sigma_2)A^\dagger=
\sigma_2\sigma_e^{t}\sigma_2 \cdot
(\mathbb{O}(A,B)p_e)
\label{s.7}
\end{equation}
where $\sigma_2 A \sigma_2 = A^*$ for $A\in SU(2)$ was used in
(\ref{s.5}-\ref{s.7}). In all four of these expressions
$A$, $B$ and the orthogonal matrix
$\mathbb{O}(A,B)$ are unchanged. All four of the
matrices (\ref{s.1}) become positive when $p_e$ is replaced by the
on-shell Minkowski four momentum,
$p^{\mu}_m = (\sqrt{\mathbf{p}^2+m^2},\mathbf{p})$
and $\sigma_e^{\mu}$ is replaced by $\sigma^{\mu}$.
These matrices appear in the Euclidean covariant
kernels for the right and left-handed representations
and their duals. The spin $s$ Euclidean covariant kernels
for each type of covariant spinor are:
\begin{equation}
S_e^s (x_e;\mu,\nu)={2 \over (2\pi)^4} \int {D^s_{\mu\nu} [p_e \cdot \sigma_{e}]\over p_e^2 +m^2}
e^{i p_e \cdot x_e}d^4p_e
\label{s.8}
\end{equation}
\begin{equation}
S_{ed}^s (x_e;\mu,\nu) ={2 \over (2\pi)^4} \int
{D^s_{\mu\nu} [p_e \cdot (\sigma_2 \sigma_{e} \sigma_2)]\over p_e^2 +m^2}
e^{i p_e \cdot x_e}d^4p_e
\label{s.9}
\end{equation}
\begin{equation}
S_{e*}^s (x_e;\mu,\nu) ={2 \over (2\pi)^4} \int
{D^s_{\mu\nu} [p_e \cdot \sigma^t_{e}]\over p_e^2 +m^2}
e^{i p_e \cdot x_e}d^4p_e
\label{s.10}
\end{equation}
\begin{equation}
S_{ed*}^s (x_e;\mu,\nu) ={2 \over (2\pi)^4} \int
{D^s_{\mu\nu} [p_e \cdot (\sigma_2 \sigma^t_{e} \sigma_2)]\over p_e^2 +m^2}
e^{i p_e \cdot x_e}d^4p_e .
\label{s.11}
\end{equation}
The physical Hilbert space
inner product associated with each of these kernels is
\begin{equation}
\langle\psi_{e} \vert \phi_{e} \rangle=
\int \sum_{\mu \nu}\psi_e^*(\theta x,\mu)
S_e^s (x_e-y_e;\mu,\nu) \phi_e(y,\nu)d^4x d^4y
\label{s.12}
\end{equation}
\begin{equation}
\langle\psi_{ed} \vert \phi_{ed} \rangle=
\int \sum_{\mu \nu} \psi^*_{ed}(\theta x,\mu)
S_{ed}^s (x_e-y_e;\mu,\nu) \phi_{ed}(y,\nu)d^4x d^4y
\label{s.13}
\end{equation}
\begin{equation}
\langle\psi_{e*} \vert \phi_{e*} \rangle=
\int \sum_{\mu \nu} \psi^*_{e*}(\theta x,\mu)
S_{e*}^s (x_e-y_e;\mu,\nu) \phi_{e*}(y,\nu)d^4x d^4y
\label{s.14}
\end{equation}
\begin{equation}
\langle\psi_{ed*} \vert \phi_{ed*} \rangle=
\int \sum_{\mu \nu} \psi^*_{ed*}(\theta x,\mu)
S_{ed*}^s (x_e-y_e;\mu,\nu) \phi_{ed*}(y,\nu)d^4x d^4y .
\label{s.15}
\end{equation}
For wave functions with positive Euclidean time support, the $p^0_e$
integral can be evaluated by the residue theorem, closing the contour
in the lower half plane. This replaces $p^0_e$ by
$-i\omega_m(\mathbf{p})$. The kernels become the two-point
Minkowski Wightman functions \cite{Wightman:1980} for
mass $m$ spin $s$ irreducible
representations of the Lorentz group. Equations (\ref{s.13}) and
(\ref{s.15}) are dual representations of the right-handed kernel,
while (\ref{s.12}) and (\ref{s.14}) are dual representations of the
left-handed kernel. $\sigma_2$ behaves like a metric tensor for the
Lorentz covariant spinors, relating the representations
(\ref{s.12}) and (\ref{s.13}) or (\ref{s.14}) and (\ref{s.15}).
Contraction of the two types of right
or left handed spinors are Lorentz invariant. The results of performing
the
$p^0_e$ integral are
\begin{equation}
\langle \psi_{e} \vert \phi_{e}\rangle=
\int \sum_{\mu \nu} f^*_m(\mathbf{p},\mu)
{d\mathbf{p}
D_{\mu \nu}^s[p_m\cdot
\sigma] \over \omega_m (\mathbf{p})}
g_m(\mathbf{p},\nu)
\label{s.16}
\end{equation}
\begin{equation}
\langle \psi_{ed} \vert \phi_{ed} \rangle=
\int \sum_{\mu \nu} f^*_m(\mathbf{p},\mu)
{d\mathbf{p}
D_{\mu \nu}^s[p_m\cdot \sigma_2 \sigma
\sigma_2] \over \omega_m (\mathbf{p})}
g_m(\mathbf{p},\nu)
\label{s.17}
\end{equation}
\begin{equation}
\langle \psi_{e*}\vert \phi_{e*}\rangle=
\int \sum_{\mu \nu}
f^*_m(\mathbf{p},\mu)
{d\mathbf{p}
D_{\mu \nu}^s[p_m\cdot
\sigma^*] \over \omega_m (\mathbf{p})}
g_m(\mathbf{p},\nu)
\label{s.18}
\end{equation}
\begin{equation}
\langle \psi_{ed*} \vert \phi_{ed*}\rangle=
\int \sum_{\mu \nu} f^*_m(\mathbf{p},\mu)
{d\mathbf{p}
D_{\mu \nu}^s[p_m\cdot \sigma_2
\sigma^* \sigma_2 ]
\over \omega_m (\mathbf{p})}
g_m(\mathbf{p},\nu)
\label{s.19}
\end{equation}
where
\begin{equation}
f^*_m(\mathbf{p},\mu):=
\int {d^4x \over (2 \pi)^{3/2}}
\psi^*(x,\mu)e^{i \mathbf{p}\cdot \mathbf{x} - \omega_m (\mathbf{p})x^0}
\label{s.20}
\end{equation}
\begin{equation}
g_m(\mathbf{p},\nu):=
{d^4y \over (2 \pi)^{3/2}}
\psi(y,\nu)e^{-i \mathbf{p}\cdot \mathbf{y} - \omega_m (\mathbf{p})y^0}
\label{s.21}
\end{equation}
for each type of spinor wave function.
Each of the spin matrices,
$D_{\mu \nu}^s[p_m\cdot \sigma]$,
$D_{\mu \nu}^s[p_m\cdot \sigma_2 \sigma \sigma_2]$,
$D_{\mu \nu}^s[p_m\cdot \sigma^*]$ and
$D_{\mu \nu}^s[p_m\cdot \sigma_2 \sigma^* \sigma_2 ]$ are
positive Hermitian matrices,
so the Euclidean Green's functions (\ref{s.8}-\ref{s.11})
are all reflection positive.
The first step to find the spinor parts of the Poincar\'e generators
in the Euclidean representation is
to use the identities (\ref{s.4}-\ref{s.7}) which lead to
\[
\int \sum_{\mu \nu} \psi_e^*(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[\mathbb{O}p\cdot \sigma_e]\phi_e(y,\nu)d^4x d^4y d^4p=
\]
\begin{equation}
\int \sum_{\mu \nu} \psi_e^*(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot A\sigma_e B^t]\phi_e(y,\nu)d^4x d^4y d^4p
\label{s.22}
\end{equation}
\[
\int \sum_{\mu \nu} \psi_{ed}^*(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}
D_{\mu \nu}^s[\mathbb{O}p\cdot \sigma_2 \sigma_e \sigma_2]
\phi_{ed}(y,\nu)d^4x d^4y d^4p=
\]
\begin{equation}
\int \sum_{\mu \nu} \psi_{ed}^*(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot A^*\sigma_2 \sigma_e \sigma_2B^\dagger]\phi_{ed}(y,\nu)d^4x d^4y d^4p
\label{s.23}
\end{equation}
\[
\int \sum_{\mu \nu} \psi^*_{e*}(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[\mathbb{O}p\cdot \sigma_e^t]\phi_{e*}(y,\nu)d^4x d^4y d^4p=
\]
\begin{equation}
\int \sum_{\mu \nu} \psi^*_{e*}(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot B\sigma_e^tA^t]\phi_{e*}(y,\nu)d^4x d^4y d^4p
\label{s.24}
\end{equation}
\[
\int \sum_{\mu \nu} \psi^*_{ed*}(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[\mathbb{O}p\cdot \sigma_2 \sigma_e^t \sigma_2 ]\phi_{ed*}(y,\nu)d^4x d^4y d^4p=
\]
\begin{equation}
\int \sum_{\mu \nu} \psi_{ed*}^*(\theta x,\mu) \frac{e^{ip\cdot(x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot B^*\sigma_2 \sigma_e^t \sigma_2A^\dagger ]\phi_{ed*}(y,\nu)d^4x d^4y d^4p .
\label{s.25}
\end{equation}
The next step is to move the transformations from the kernels to the
wave functions. The Euclidean invariance of the measures and scalar
products, the group representation properties of the Wigner functions,
and re-definitions of the wave functions can be used to show that
(\ref{s.22}-\ref{s.25}) are equivalent to
\[
\int \sum (D^s_{\mu\alpha}[A^\dagger]^{-1}{\psi}_e(\theta \mathbb{O}^t \theta x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_e]
{\phi}_e(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_e^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_e]D^s_{\alpha \nu}[B^t]{\phi}_e(\mathbb{O}y,\nu)d^4x d^4y d^4p
\label{s.26}
\end{equation}
\[
\int \sum ( D^s_{\mu \alpha}[A^t]^{-1}{\psi}_{ed}(\theta \mathbb{O}^t \theta x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_2 \sigma_e \sigma_2]{\phi}_{ed}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}^*_{ed}( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_2 \sigma_e \sigma_2]
D^s_{\alpha \nu}[B^\dagger]{\phi}_{ed}(\mathbb{O}y,\nu)d^4x d^4y d^4p
\label{s.27}
\end{equation}
\[
\int \sum (D^s_{\mu \alpha}[B^\dagger]^{-1} {\psi}_{e*}(\theta \mathbb{O}^t \theta x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_e^t]{\phi}_{e*}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_{e*}^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_e^t]D^s_{\alpha \nu}[A^t]\Tilde{\phi}_{e*}(\mathbb{O}y,\nu)d^4x d^4y d^4p
\label{s.28}
\end{equation}
\[
\int \sum (D^s_{\mu \alpha}[B^t]^{-1} {\psi}_{ed*}(\theta \mathbb{O}^t\theta x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_2 \sigma_e^t \sigma_2 ]{\phi}_{ed*}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
= \int \sum {\psi}_{ed*}^*( x,\mu) \frac{e^{ip\cdot(\theta
x-y)}}{p^2+m^2}D_{\mu \alpha}^s[p\cdot \sigma_2 \sigma_e^t \sigma_2]
D^s_{\alpha \nu}[A^\dagger]{\phi}_{ed*}(\mathbb{O}y,\nu)d^4x d^4y d^4p .
\label{s.29}
\end{equation}
To derive expressions for the generators for each type of spinor,
check the hermiticity and verify the commutation relations
the first step is to replace $A$ and $B$ with
the pairs of $SU(2)$ matrices representing
one-parameter groups for both ordinary rotations about a fixed
axis and rotations in a Euclidean space time plane.
For ordinary rotations about the $\hat{\mathbf{n}}$ axis,
the one-parameter group is
\begin{equation}
A(\lambda ) = B^*(\lambda) = e^{i {\lambda \over 2}\hat{\mathbf{n}}
\cdot \pmb{\sigma}}
\label{s.30}
\end{equation}
and
$(\theta \mathbb{O}^t(\lambda)\theta) = \mathbb{O}^t(\lambda)$,
while for rotations in Euclidean $\hat{\mathbf{n}}$-$x^0$ space-time planes
the one-parameter group is
\begin{equation}
A(\lambda ) = B^t(\lambda) = e^{i {\lambda \over 2}\hat{\mathbf{n}}
\cdot \pmb{\sigma}}
\label{s.31}
\end{equation}
and
$(\theta \mathbb{O}^t(\lambda)\theta) = \mathbb{O}(\lambda)$.
The $4 \times 4$ orthogonal transformations, $\mathbb{O}(\lambda)$
associated with each
type of transformation are shown explicitly for
rotations about the $\hat{\mathbf{z}}$ axis and for
rotations in the $\hat{\mathbf{z}}$-$x^0$
plane:
For rotations about the $\hat{\mathbf{z}}$ axis
\begin{equation}
\mathbb{O}(A,A^*)( \lambda) =
\left (
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & \cos(\lambda)&\sin (\lambda)&0\\
0 & -\sin(\lambda)&\cos (\lambda)&0\\
0 & 0 & 0& 1\\
\end{array}
\right )
\label{s.32}
\end{equation}
and
\begin{equation}
\Theta \mathbb{O}(A,A^*)( \lambda) \Theta = \mathbb{O}(A,A^*)( \lambda).
\label{s.33}
\end{equation}
For rotations in the $\hat{\mathbf{z}}$-$x^0$ plane
\begin{equation}
\mathbb{O}(A,A^t)( \lambda)=
\left (
\begin{array}{cccc}
\cos(\lambda) & 0 & 0 & \sin (\lambda) \\
0 & 1 & 0 &0\\
0 & 0 & 1 &0\\
-\sin (\lambda) & 0 & 0& \cos(\lambda)\\
\end{array}
\right )
\label{s.34}
\end{equation}
and
\begin{equation}
\theta \mathbb{O}^t(A,A^t)( \lambda) \theta = \mathbb{O}(A,A^t)( \lambda)
\label{s.35}
\end{equation}
For the case of ordinary rotations
$A=B^*$ and equations (\ref{s.26}-\ref{s.29}) become
\[
\int \sum (D^s_{\mu\alpha}[A]{\psi}_e ( \mathbb{O}^t(\lambda) x,\alpha) )^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_e]{\phi}_e(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_e^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_e]D^s_{\alpha \nu}[A^{\dagger}]{\phi}_e(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p
\label{s.36}
\end{equation}
\[
\int \sum ( D^s_{\mu \alpha}[A^*]{\psi}_{ed}( \mathbb{O}^t(\lambda) x,\alpha))^*
\frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_2 \sigma_e \sigma_2]{\phi}_{ed}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_{ed}^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_2 \sigma_e \sigma_2]
D^s_{\alpha \nu}[A^t] {\phi}_{ed}(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p
\label{s.37}
\end{equation}
\[
\int \sum (D^s_{\mu \alpha}[A^*] {\psi}_{e*}( \mathbb{O}^t(\lambda) x,\alpha))^*
\frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_e^t]{\phi}_{e*}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_{e*}^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_e^t]D^s_{\alpha \nu}[A^t]{\phi}_{e*}(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p
\label{s.38}
\end{equation}
\[
\int \sum (D^s_{\mu \alpha}[A] {\psi}_{ed*}( \mathbb{O}^t(\lambda) x,\alpha))^*
\frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_2
\sigma_e^t \sigma_2 ]{\phi}_{ed*}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
= \int \sum {\psi}_{ed*}^*( x,\mu) \frac{e^{ip\cdot(\theta
x-y)}}{p^2+m^2}D_{\mu \alpha}^s[p\cdot \sigma_2 \sigma_e^t \sigma_2
]D^s_{\alpha \nu}[A^\dagger]{\phi}_{ed*}(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p
\label{s.39}
\end{equation}
For the case of rotations in Euclidean space-time planes
for $A=B^t$ equations (\ref{s.26}-\ref{s.29}) become
\[
\int \sum (D^s_{\mu\alpha}[A]{\psi}_e(\mathbb{O}(\lambda) x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_e]{\phi_e}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi_e}^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_e]D^s_{\alpha \nu}[A]
{\phi_e}(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p
\label{s.40}
\end{equation}
\[
\int \sum ( D^s_{\mu \alpha}[A^*]{\psi}_{ed}(\mathbb{O}(\lambda) x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_2 \sigma_e \sigma_2]{\phi}_{ed}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_{ed}^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_2 \sigma_e \sigma_2]
D^s_{\alpha \nu}[A^*]{\phi}_{ed}(\mathbb{O}y,\nu)d^4x d^4y d^4p
\label{s.41}
\end{equation}
\[
\int \sum (D^s_{\mu \alpha}[A^t] {\psi}_{e*}(\mathbb{O}(\lambda) x,\alpha))^* \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_e^t]{\phi}_{e*}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
=\int \sum {\psi}_{e*}^*( x,\mu) \frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \alpha}^j[p\cdot \sigma_e^t]D^s_{\alpha \nu}[A^t]{\phi}_{e*}(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p
\label{s.42}
\end{equation}
\[
\int \sum (D^s_{\mu \alpha}[A^\dagger ] {\psi}_{ed*}(\mathbb{O}(\lambda) x,\alpha))^*
\frac{e^{ip\cdot(\theta x-y)}}{p^2+m^2}D_{\mu \nu}^s[p\cdot \sigma_2 \sigma_e^t \sigma_2 ]{\phi}_{ed*}(y,\nu)d^4x d^4y d^4p
\]
\begin{equation}
= \int \sum {\psi}_{ed*}^*( x,\mu) \frac{e^{ip\cdot(\theta
x-y)}}{p^2+m^2}D_{\mu \alpha}^s[p\cdot \sigma_2 \sigma_e^t \sigma_2
]D^s_{\alpha \nu}[A^\dagger]{\phi}_{ed*}(\mathbb{O}(\lambda)y,\nu)d^4x d^4y d^4p.
\label{s.43}
\end{equation}
Note that the transformations above represent inverse Lorentz
transformations since
\begin{equation}
\langle x, \nu \vert U(\Lambda,0) \vert \psi \rangle =
\langle \psi \vert U^{\dagger}(\Lambda,0)\vert x,\nu \rangle^* =
\langle \mathbb{O}^t x,\nu \vert \psi \rangle^*=
\langle \psi \vert \mathbb{O}^t x, \nu \rangle .
\label{s.44}
\end{equation}
To construct generator of ordinary rotations differentiate the
right hand side of
(\ref{s.36}-\ref{s.39}) by $\lambda$, set $\lambda=0$, and multiply the result
by $i$. To construct the generators of Euclidean space-time rotations
differentiate the right hand side of
(\ref{s.40}-\ref{s.43}) by $\lambda$, set $\lambda=0$, and multiply the
result by $i$
to get expressions for the generators. To get expressions for the
Lorentz Boost generators multiply the Euclidean space-time rotation
generators by an additional factor of $-i$.
The derivatives of the
Wigner functions can be computed using
\begin{equation}
{d \over d\lambda}D^s_{\mu \nu} [A(\lambda)]_{\vert_{\lambda=0}} =
{d \over d\lambda}\langle s, \mu \vert
e^{i \lambda \hat{\mathbf{n}}\cdot \mathbf{S}}
\vert s, \nu \rangle_{\vert_{\lambda=0}}=
i \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \nu \rangle
\label{s.45}
\end{equation}
\begin{equation}
{d \over d\lambda}D^s_{\mu \nu} [A(\lambda)^{\dagger}]_{\vert_{\lambda=0}} =
{d \over d\lambda}\langle s, \mu \vert e^{-i \lambda \hat{\mathbf{n}}\cdot \mathbf{S}} \vert s, \nu \rangle_{\vert_{\lambda=0}} =
- i \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \nu \rangle
\label{s.46}
\end{equation}
\begin{equation}
{d \over d\lambda}D^s_{\mu \nu} [A^*(\lambda)]_{\vert_{\lambda=0}} =
{d \over d\lambda}(D^s_{\mu \nu} [A(\lambda)])^*_{\vert_{\lambda=0}} =
-i \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \nu \rangle^*=
-i \langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \mu \rangle
\label{s.47}
\end{equation}
\begin{equation}
{d \over d\lambda}D^s_{\mu \nu} [A^t(\lambda)]_{\vert_{\lambda=0}} =
{d \over d\lambda}(D^s_{\mu \nu} ([A(\lambda)])^*)^{-1}_{\vert_{\lambda=0}} =
i\langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \nu \rangle^*=
i\langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \mu \rangle .
\label{s.48}
\end{equation}
These can be evaluated using $S_z$ and angular momentum
raising and lowering operators.
The rotation generators for each type of spinor representation can be read off
of (\ref{s.36}-\ref{s.39}):
\begin{equation}
\langle x,s,\nu \vert \mathbf{J} \vert \psi_e \rangle =
\sum_\nu \left (\delta_{\mu \nu} \mathbf{x} \times (-i {\partial \over \partial \mathbf{x}})
+ \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \nu \rangle
\right ) \langle x,s,\nu \vert \psi_e \rangle
\label{s.49}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert \mathbf{J} \vert \psi_{ed} \rangle =
\sum_\nu \left ( \delta_{\mu \nu} \mathbf{x} \times (-i {\partial \over \partial \mathbf{x}})
- \langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \mu \rangle
\right ) \langle x,s,\nu \vert \psi_{ed} \rangle
\label{s.50}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert \mathbf{J} \vert \psi_{e*} \rangle =
\sum_\nu \left ( \delta_{\mu \nu} \mathbf{x} \times (-i {\partial \over \partial \mathbf{x}})
- \langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \mu \rangle
\right ) \langle x,s,\nu \vert \psi_{e*} \rangle
\label{s.51}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert \mathbf{J} \vert \psi_{ed*} \rangle =
\sum_\nu \left ( \delta_{\mu \nu} \mathbf{x} \times (-i {\partial \over \partial \mathbf{x}})
+ \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s, \nu \rangle
\right ) \langle x,s,\nu \vert \psi_{ed*} \rangle
\label{s.52}
\end{equation}
The first and fourth term are representations of standard rotation generators
. In the second and third terms the spin generator matrix elements are
transposed and multiplied by
with a (-) sign. To show that these operator satisfy $SU(2)$
commutation relations,
consider matrices satisfying $SU(2)$ commutation relations:
\begin{equation}
[ M_i, M_j] = i \epsilon_{ijk} M_k .
\label{s.53}
\end{equation}
The transposes satisfy
\begin{equation}
[ M^t_j, M^t_i] = i \epsilon_{ijk} M^t_k
\label{s.54}
\end{equation}
\begin{equation}
[ (-M^t_i),(- M^t_j)] = i \epsilon_{ijk} (-M^t_k)
\label{s.55}
\end{equation}
which shows that the negative transpose of these matrices also
satisfy $SU(2)$ commutation relations. This shows that all of the
spin generator satisfy $SU(2)$ commutation relations.
Generators for rotations in
Euclidean space-time planes are constructed the same way from
\begin{equation}
\langle x,s,\nu \vert J^{0\hat{n}} \vert \psi_{e} \rangle =
\sum_\nu \left (i \delta_{\mu \nu} (\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}})
- \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\nu \rangle
\right ) \langle x,s,\nu \vert \psi_e \rangle
\label{s.56}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert J^{0\hat{n}} \vert \psi_{ed} \rangle =
\sum_\nu \left (i \delta_{\mu \nu}(\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}}) +
\langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\mu \rangle
\right ) \langle x,s,\nu \vert \psi_{ed} \rangle
\label{s.57}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert J^{0\hat{n}} \vert \psi_{e*} \rangle =
\sum_\nu \left ( i \delta_{\mu \nu}(\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}}) -
\langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\mu \rangle
\right ) \langle x,s,\nu \vert \psi_{e*} \rangle
\label{s.58}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert J^{0\hat{n}} \vert \psi_{ed*} \rangle =
\sum_\nu \left ( i \delta_{\mu \nu}(\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}})
+ \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\nu \rangle
\right ) \langle x,s,\nu \vert \psi_{ed*} \rangle
\label{s.59}
\end{equation}
In order to construct the boost generators it is necessary to multiply these
expression by an additional factor of (-i)
\begin{equation}
\langle x,s,\nu \vert \mathbf{K} \vert \psi_{e} \rangle =
\sum_\nu \left ( \delta_{\mu \nu}(\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}}) +
i \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\nu \rangle
\right ) \langle x,s,\nu \vert \psi_e \rangle
\label{s.60}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert \mathbf{K} \vert \psi_{ed} \rangle =
\sum_\nu \left ( \delta_{\mu \nu}(\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}}) -i
\langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\mu \rangle
\right ) \langle x,s,\nu \vert \psi_{ed} \rangle
\label{s.61}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert \mathbf{K} \vert \psi_{e*} \rangle =
\sum_\nu \left ( \delta_{\mu \nu}(\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}}) +i
\langle s, \nu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\mu \rangle
\right ) \langle x,s,\nu \vert \psi_{e*} \rangle
\label{s.62}
\end{equation}
\begin{equation}
\langle x,s,\nu \vert \mathbf{K} \vert \psi_{ed*} \rangle =
\sum_\nu \left ( \delta_{\mu \nu} (\mathbf{x} {\partial \over \partial x^0} -
x^0 {\partial \over \partial \mathbf{x}})
-i \langle s, \mu \vert \hat{\mathbf{n}}\cdot \mathbf{S} \vert s,\nu \rangle
\right ) \langle x,s,\nu \vert \psi_{ed*} \rangle
\label{s.63}
\end{equation}
The continuous part of these expressions agree with (\ref{d.26}-\ref{d.27})
for spinless operators. The relevant commutators involving
the spin parts of the boost generators in each of the
four representations are
\begin{equation}
[K_i, K_j ] = [i S_i,i S_j] = -i \epsilon_{ijk} S_k
\label{s.65}
\end{equation}
\begin{equation}
[K_i, K_j ] = [-i S^t_i,-i S^t_j] = - \epsilon_{ijk} (-S^t_k)
\label{s.65}
\end{equation}
\begin{equation}
[K_i, K_j ] = [i S^t_i,i S^t_j] = - \epsilon_{ijk} (-S^t_k)
\label{s.66}
\end{equation}
\begin{equation}
[K_i, K_j ] = [-i S_i,-i S_j] = -i \epsilon_{ijk} S_k
\label{s.67}
\end{equation}
\begin{equation}
[K_i, S_j ] = [i S_i, S_j] = \epsilon_{ijk} iS_k = \epsilon_{ijk} K_k
\label{s.68}
\end{equation}
\begin{equation}
[K_i, S_j ] = [-iS_i^t, -S^t_j] = \epsilon_{ijk} -iS^t_k = \epsilon_{ijk} K_k
\label{s.69}
\end{equation}
\begin{equation}
[K_i, S_j ] = [i S^t_i, -S^t_j] = \epsilon_{ijk} iS^t_k = \epsilon_{ijk} K_k
\label{s.70}
\end{equation}
\begin{equation}
[K_i, S_j ] = [-i S_i,S_j] = -i \epsilon_{ijk} (-iS_k) = \epsilon_{ijk} K_k
\label{s.71}
\end{equation}
where the spin generators in (\ref{s.65},\ref{s.66},\ref{s.69}) and
(\ref{s.70}) are (-) the transposes of the matrices satisfying $SU(2)$
commutation relations, which were shown in (\ref{s.53}-\ref{s.55}) to
satisfy $SU(2)$ commutation relations. It follows that the
expressions (\ref{s.49}-\ref{s.52}) and (\ref{s.60}-\ref{s.63}) for the
Lorentz generators in each of the spinor representations satisfy the
Poincar\'e commutation relations.
The hermiticity of these generators follow from the expressions
(\ref{s.36}-\ref{s.39}) and (\ref{s.40}-\ref{s.43}).
Each of equations (\ref{s.36}-\ref{s.39}) has the form
\begin{equation}
\langle U^{\dagger}(\lambda)\psi \vert \phi \rangle =
\langle \psi \vert U(\lambda) \vert \phi \rangle
\label{s.72}
\end{equation}
so the rotation operators, which are generators of
unitary one-parameter groups \cite{riesz} are self-adjoint in the
Hilbert spaces with inner products (\ref{s.12}-\ref{s.15}).
For the boost generators hermiticity follows from
(\ref{s.40}-\ref{s.43}). In this case all of
these equations have the form
\begin{equation}
\langle T(\lambda)\psi \vert \phi \rangle =
\langle \psi \vert T(\lambda) \vert \phi \rangle
\label{s.73}
\end{equation}
In these cases $T(\lambda)$ is Hermitian, but the generators are
constructed by multiplying the $\lambda$ derivative $1=(i)(-i)$
rather than $i$, resulting in Hermitian operators.
In these covariant representations the spin does not enter in
the Hamiltonian or the linear momentum operators. These operators
all commute with the spin operators and commutators with these
operators follow from the scalar case.
The main result of this section the expressions
(\ref{s.49}-\ref{s.52}) and (\ref{s.60}-\ref{s.63})
for the Poincar\'e generators. The construction relates the
Euclidean spinors to the Lorentz covariant spinors.
\section{Self Adjointness}
While the self-adjointness of the generators of ordinary rotations
follows from the unitarity of the one-parameter group of rotations on
the Hilbert spaces (\ref{s.12}-\ref{s.15}), this argument does not
apply to either the Hamiltonian or the boost generators. In both
cases the operators were derived from the corresponding Euclidean
generators by multiplication by an imaginary constant. The Euclidean
generators and corresponding Lorentz generators act on different
Hilbert space representations. The problem is that the corresponding
finite Euclidean transformations can map functions with positive time
support to functions that violate this condition.
For the Hamiltonian this can be treated by only considering
translations in the positive Euclidean time direction.
These translations map functions with positive Euclidean time
support into functions with positive Euclidean time support.
Reflection positivity can be used to show that
translations in the positive Euclidean time direction define a
contractive Hermitian semigroup on the Hilbert space with the scalar products
(\ref{d.7}-\ref{d.10}).
The argument \cite{glimm} uses the Schwartz inequality on both
the physical and Euclidean Hilbert spaces. One application of
the Schwartz inequality on the physical Hilbert space gives
\begin{equation}
\Vert \vert e^{-H x^0} \vert \phi \rangle \Vert
=
\langle e^{-H x^0} \phi \vert e^{-H x^0} \vert \phi \rangle^{1/2}
=
\langle \phi \vert e^{-H 2x^0} \vert \phi \rangle^{1/2} \leq
\Vert \vert e^{-H 2x^0} \vert \phi \rangle \Vert^{1/2}
\Vert \vert \phi \rangle \Vert^{1/2} .
\label{sa.1}
\end{equation}
Repeating these steps $n$-times gives
\begin{equation}
\Vert \vert e^{-H x^0} \vert \phi \rangle \Vert \leq
\Vert \vert e^{-H 2^nx^0} \vert \phi \rangle \Vert^{1/2^n}
\Vert \vert \phi \rangle \Vert^{1-1/2^n} .
\label{sa.2}
\end{equation}
The quantity
\begin{equation}
\Vert \vert e^{-H 2^nx^0} \vert \phi \rangle \Vert \leq
\Vert \theta U_e(2^nx^0) \vert \psi \rangle \Vert_e <
\Vert \vert \psi \rangle \Vert_e <
\infty
\label{sa.3}
\end{equation}
is bounded by the Euclidean norm, $\Vert \cdot \Vert_e$,
since $U_e(2^nx^0)$ is unitary and $\Vert \theta \Vert_e= 1$ on
that Hilbert space.
Since this is finite and independent of $n$, taking the limit as
$n \to \infty$ gives
\begin{equation}
\Vert \vert e^{-H x^0} \vert \phi \rangle \Vert \leq
\Vert \vert \phi \rangle \Vert .
\label{sa.4}
\end{equation}
It follows that positive Euclidean time translations define a
contractive Hermitian semigroup on the Hilbert spaces (\ref{s.12}-\ref{s.15}).
The generator is a positive self-adjoint operator \cite{riesz}\cite{simon}.
Boosts present additional complications. Even an infinitesimal
rotation in a Euclidean space time plane will map a general function
with positive Euclidean time support to one that violates this
condition. The self-adjointness of the boost generator cannot be
demonstrated by showing that it defines a unitary one-parameter group
or contractive semigroup, however it turns out that rotations in
Euclidean space time planes, which are interpreted as boosts with
complex rapidity, define local symmetric semigroups
\cite{Klein:1981}\cite{Klein:1983} \cite{Frohlich:1983kp} on the
Hilbert spaces (\ref{s.12}-\ref{s.15}). These have self-adjoint
generators, which are exactly the boost generators.
The conditions for a local symmetric semigroup
\cite{Klein:1981}
are
\begin{itemize}
\item[1.] For each $\theta \in [0 ,\theta_0]$, there is a linear subset
${\cal D}_{\theta}$ such that ${\cal D}_{\theta_1}\supset {\cal D}_{\theta_2}$
if $\theta_1< \theta_2$, and $\cup_{0<\theta<\theta_0}
{\cal D}_{\theta_2}$ is dense.
\item[2.] For each $\theta \in [0,\theta_0]$, $E(\theta)$ is a linear operator
on the Hilbert space with domain ${\cal D}_{\theta}$
\item[3.] $E(0)=I$,
$E(\theta_1):{\cal D}_{\theta_2}\to {\cal D}_{\theta_2-\theta_1}$,
and $E(\theta_1)E(\theta_2)= E(\theta_1+\theta_2)$ on
${\cal D}_{\theta_1+\theta_2}$ for $\theta_1,\theta_2,\theta_1+\theta_2 \in
[0,\theta_0]$
\item[4.] $E(\theta)$ is Hermitian for $\theta \in [0,\theta_0]$
\item[5.] $E(\theta)$ is weakly continuous on $[0,\theta_0]$
\end{itemize}
When these conditions are satisfied there is a unique self-adjoint
operator $K$ such that ${\cal D}_{\theta} \subset {\cal D}_{e^{-K\theta}}$
and $E(\theta)$ is the restriction of $e^{-K\theta}$ to
${\cal D}_{\theta}$.
In this case $E(\theta)$ represents Euclidean space time rotations
considered as operators on the Hilbert space (\ref{a.4})
restricted to domains that will be described below.
The domains are Schwartz functions with space Euclidean time support the
wedge shaped region defined by
\begin{equation}
\mathbf{x} \cdot \hat{\mathbf{n}} - {x^0_e \over \epsilon} + \epsilon <0
\label{sa.5}
\end{equation}
\begin{equation}
\mathbf{x} \cdot \hat{\mathbf{n}} + {x^0_e \over \epsilon} - \epsilon >0
\label{sa.6}
\end{equation}
The wedge shaped region becomes the positive Euclidean time
half plane in the limit
that $\epsilon\to 0$. Schwartz functions with support on this
half plane are dense. In addition,
if this domain is rotated by an angle less than
$\theta_\epsilon :=\pm \tan^{-1}(\epsilon)$, it
will still be contained in the positive Euclidean time
half plane. Schwartz functions with support in
these wedge shaped regions can be constructed from
Schwartz functions that have support or positive Euclidean time
by multiplying the function by
$g(x^0,\mathbf{x}\cdot \hat{\mathbf{n}},\epsilon)$
where
\begin{equation}
g(x^0,\mathbf{x}\cdot \hat{\mathbf{n}},\epsilon)) =
h({x^0_e \over \epsilon} - \epsilon +\mathbf{x} \cdot \hat{\mathbf{n}})
h({x^0_e \over \epsilon} - \epsilon -\mathbf{x} \cdot \hat{\mathbf{n}}).
\label{sa.7}
\end{equation}
and
\begin{equation}
h (\lambda) =
\left \{
\begin{array}{cc}
e^{- {1 \over (\lambda)^2}} & \lambda >0 \\
0 & \lambda \leq 0 \\
\end{array}
\right . .
\label{sa.8}
\end{equation}
is a smoothed Heaviside function.
$g(x^0,\mathbf{x}\cdot \hat{\mathbf{n}},\epsilon))$ is a Schwartz function with
support in the wedge shaped region (\ref{sa.5}-\ref{sa.6}) that
approaches $1$ as $\epsilon (\theta)$ approaches 0.
The domain ${\cal D}_{\theta}$ is taken as the space of Schwartz functions with
positive time support multiplied by the function
$g(x^0,\mathbf{x}\cdot \hat{\mathbf{n}},\epsilon))$ where $\theta=\theta_\epsilon$. The Euclidean space time rotations restricted to
these domains have all of the properties of local symmetric semigroup.
It follows that the boost generators $\mathbf{K}$ are self-adjoint
on the physical Hilbert space.
\section{Finite transformations}
Finite Poincar\'e transformations are needed for most applications.
While the generators for space translations and rotations were
constructed from the associated unitary one-parameter groups, the
construction of the boost generators and Hamiltonian was not as
direct. Because self adjointness was established for the Hamiltonian
and boost generators, each one of these generators has a dense set of
analytic vectors \cite{simon1} where exponential series for the unitary
one parameter groups converge. This ensures that the differential
operators (\ref{d.27}) that define the boost generators
applied to a dense set of functions with positive time support have positive
time support.
Directly summing the exponential series is inefficient. On the other
hand, the structure of the finite unitary transformations is fixed by
(\ref{c.29}) when they act on irreducible basis states. The situation
is analogous to non-relativistic quantum mechanics - time evolution
becomes trivial once the Hamiltonian is diagonalized. In the
Euclidean framework, the analogous problem is to diagonalize the mass
squared operator. This is a dynamical problem that depends on the
choice of reflection positive Euclidean covariant distributions. For
the Green functions discussed in sections 4-5 the mass operator is the
four dimensional Euclidean Laplacian. For more general Euclidean
covariant Green's functions is it second order differential operator.
The spectral condition ensures that there are no negative energy
states.
The Euclidean Green's functions are manifestly covariant
with respect to space translations and rotations. Given a mass
eigenstate, the translational and rotational covariance can
be used to decompose the mass eigenstate into a
linear superposition of simultaneous eigenstates of linear momentum,
and spins. On these states the unitary representation of the
Poincar\'e group acts irreducibly.
In the Euclidean formalism, since the dynamics
is in the Green function, mass eigenstates are solutions to
\begin{equation}
\langle \phi \vert (M^2-m^2) \vert \psi \rangle =0
\label{ft.1}
\end{equation}
for all $\phi$ satisfying the support condition.
Methods for constructing mass eigenstates satisfying the support
condition are discussed in \cite{Aiello:2015jgc}.
Mass-momentum eigenstates can be constructed using
\begin{equation}
\vert m,\mathbf{p} \rangle :=
\int e^{-i \mathbf{p}\cdot \mathbf{a}}U(\mathbf{a}) \vert \psi \rangle d\mathbf{a}.
\label{ft.2}
\end{equation}
Applying the translation operator $U(\mathbf{a}')$ to this vector gives
\[
U(\mathbf{a}')\vert m,\mathbf{p} \rangle =
U(\mathbf{a}') \int e^{-i \mathbf{p}\cdot \mathbf{a}}U(\mathbf{a}) \vert \psi \rangle d\mathbf{a} =
\int e^{-i \mathbf{p}\cdot \mathbf{a}}U(\mathbf{a}+\mathbf{a}')
\vert \psi \rangle d\mathbf{a}=
\]
\begin{equation}
\int e^{-i \mathbf{p}\cdot (\mathbf{a}''-\mathbf{a}')}U(\mathbf{a}'')
\vert \psi \rangle d\mathbf{a}''=
e^{i \mathbf{p}\cdot \mathbf{a}'}
\int e^{-i \mathbf{p}\cdot \mathbf{a}''}U(\mathbf{a}'')
\vert \psi \rangle d\mathbf{a}'' =
e^{i\mathbf{p}\cdot \mathbf{a}'}
\vert m,\mathbf{p} \rangle
\label{ft.3}
\end{equation}
which shows that (\ref{ft.2}) is either 0 or an eigenstate of
linear momentum with eigenvalue $\mathbf{p}$.
The mass-momentum eigenstates can be decomposed into spin eigenstates.
Right and left handed kernels with the covariance properties
(\ref{s.12}) or (\ref{s.15}) after integration become kernels for the
covariant representations (\ref{d.13}) and (\ref{d.14}). For
Green's functions with these rotational covariance properties
the covariant basis states, up to normalization, can be constructed
as follows,
\begin{equation}
\vert (m,s) \mathbf{p},\mu \rangle :=
\int U(R) \vert m, R^{-1}\mathbf{p} \rangle D^{s*}_{\mu 0} [R] dR
\label{ft.4}
\end{equation}
where the integral is over the $SU(2)$ Haar measure.
For a fixed rotation $R'$:
\[
U(R') \vert (m,s) R^{\prime -1}\mathbf{p},\mu \rangle =
U(R')\int U(R) \vert m, R^{-1} R^{\prime -1} \mathbf{p} \rangle D^{s*}_{\mu 0} [R] dR =
\int U(R' R) \vert m, (R'R)^{-1}\mathbf{p} \rangle D^{s*}_{\mu 0 } [R] dR =
\]
\begin{equation}
\int U(R'') \vert m, R^{\prime\prime -1}\mathbf{p} \rangle D^{s*}_{\mu \nu} [R^{'-1}]
D^{s*}_{\nu 0}[R''] dR'' =
\int U(R'') \vert m, R^{\prime \prime -1}R' \mathbf{p} \rangle
D^{s*}_{\nu 0}[R''] dR''D^{s}_{\nu \mu} [R'] =
\vert (m,s) \mathbf{p},\nu \rangle D^{s}_{\nu \mu} [R'].
\label{ft.5}
\end{equation}
If $R'$ is a rotation about the $z$ axis, it follows
that the resulting vector is an eigenstate of
$s^2$ and $s_z$.
This shows how mass eigenstates can be
decomposed into a superposition of Lorentz covariant states that
transform irreducibly with respect to the Poincar\'e group.
\section{Summary and Conclusion}
The purpose of this paper is to provide explicit representations for
Poincar\'e generators for systems of particles of any spin in
Euclidean representations of relativistic quantum mechanics,
demonstrate that these generators satisfy the commutation relations of
the Poincar\'e Lie Algebra and are self-adjoint with respect to a
reflection positive scalar product. This was done by starting with
irreducible unitary representations of the Poincar\'e group and
expressing them in a manifestly Lorentz covariant form. The inner
product in the Lorentz covariant representation necessarily had a
non-trivial kernel, which could be expressed in terms of reflection
positive Green functions. Expressions for the generators for any spin
were derived based on these relations.
While the results are specifically for positive mass irreducible
representations, they apply more generally since any unitary
representation of the Poincar\'e group can be decomposed into a direct
integral of positive-mass positive-energy irreducible representations.
Two consequences of the Osterwalder-Schrader reconstruction theorem
are (1) the locality axiom is logically independent of the other
Euclidean axioms and (2) the Hilbert space representation of the
quantum theory does not require explicit analytic continuation. These
observations suggest the possibility of formulating phenomenological
non-local relativistic quantum mechanical models in a purely Euclidean
representation
\cite{Kopp:2011vv}\cite{Polyzou:2013nga}\cite{Aiello:2015jgc}. The
new feature is that the dynamics appears in model Euclidean Green's
functions rather than in the Hamiltonian, which is a simple
differential operator. One of the advantages of the Euclidean
formulation is that Euclidean Green's functions are moments of a
Euclidean path integral, which provides a formal connection to the
dynamics of Lagrangian field theories. Models can be formulated by
perturbing products of free Green Euclidean functions with Euclidean
covariant interactions that preserve reflection positivity.
The authors would like to acknowledge Palle J{\o}rgensen for helpful discussions
on reflection positivity.
|
\section{Introduction}
Machine learning (ML) methods reach ever deeper into quantum chemistry and materials simulation, delivering predictive models of interatomic potential energy surfaces~\cite{behler2007generalized,braams2009permutationally, bartok2010gaussian,smith2017ani,podryabinkin2017active,podryabinkin2019accelerating}, molecular forces~\cite{chmiela2017machine,chmiela2018towards}, electron densities~\cite{Ryczko2018}, density functionals~\cite{brockherde2017bypassing}, and molecular response properties such as polarisabilities~\cite{Wilkins2019}, and infrared spectra~\cite{gastegger2017machine}.
Large data sets of molecular properties calculated from quantum chemistry or measured from experiment are equally being used to construct predictive models to explore the vast chemical compound space~\cite{rupp2012fast,eickenberg2017solid,doi:10.1002/anie.201709686,gilmer2017neural,jha2018elemnet} to find new sustainable catalyst materials~\cite{Kitchin2018}, and to design new synthetic pathways~\cite{doi:10.1002/anie.201803562}. Several works have also explored the potential role of machine learning in constructing approximate quantum chemical methods~\cite{doi:10.1021/acs.jctc.8b00873, hegde2017machine}.
\begin{figure}
\includegraphics[width=0.8\columnwidth]{figures/flowchart}
\caption{\small \textbf{Synergy of quantum chemistry and machine learning.}
(\textbf{a}) Forward model: ML predicts chemical properties based on reference calculations. If another property is required, an additional ML model has to be trained.
(\textbf{b}) Hybrid model: ML predicts the wavefunction. All ground state properties can be calculated and no additional ML is required. The wavefunctions can act as an interface between ML and QM.
\label{fig:flowchart}}
\end{figure}
Most existing ML models have in common that they learn from quantum chemistry to describe molecular properties as scalar, vector, or tensor fields~\cite{Grisafi2018, Thomas2018}.
Fig.~\ref{fig:flowchart}\textbf{a} shows schematically how quantum chemistry data of different electronic properties, such as energies or dipole moments, is used to construct individual ML models for the respective properties.
This allows for the efficient exploration of chemical space.with respect to these properties.
Yet, these ML models do not explicitly capture the electronic degrees of freedom in molecules that lie at the heart of quantum chemistry. All chemical concepts and physical molecular properties are determined by the electronic Schr\"odinger equation and derive from the ground-state wavefunction.
Thus, an electronic structure ML model that directly predicts the ground-state wavefunction (see Fig.~\ref{fig:flowchart}\textbf{b}) would not only allow to obtain all ground-state properties, but could open avenues towards new approximate quantum chemistry methods based on an interface between ML and quantum chemistry.
In this work, we develop a deep learning framework that provides an accurate ML model of molecular electronic structure via a direct representation of the electronic Hamiltonian in a local basis representation.
The model provides a seamless interface between quantum mechanics and ML by predicting the eigenvalue spectrum and molecular orbitals (MOs) of the Hamiltonian for organic molecules close to 'chemical accuracy' ($\sim$0.04~eV). This is achieved by training a flexible ML model to capture the chemical environment of atoms in molecules and of pairs of atoms.
Thereby, it provides access to electronic properties that are important for chemical interpretation of reactions such as charge populations, bond orders, as well as dipole and quadrupole moments without the need of specialised ML models for each property.
We demonstrate how this model retains the conceptual strength of quantum chemistry at force field efficiency by performing an ML-driven molecular dynamics simulation of malondialdehyde showing the evolution of the electronic structure during a proton transfer.
As we obtain a symmetry-adapted and analytically differentiable representation of the electronic structure, we are able to optimise electronic properties, such as the HOMO-LUMO gap, in a step towards inverse design of molecular structures.
Beyond that, we show that the electronic structure predicted by our approach may serve as input to further quantum chemical calculations. For example, wavefunction restarts based on this ML model provide a significant speed-up of the self-consistent field procedure (SCF) due to a reduced number of iterations, without loss of accuracy. The latter showcases that quantum chemistry and machine learning can be used in tandem for future electronic structure methods.
\section{Results}
\subsection{Atomic Representation of Molecular Electronic Structure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figures/schnorb2}
\caption{\small \textbf{Prediction of electronic properties with SchNOrb}
(\textbf{a})~Illustration of the network architecture. The neural network architecture consists of three steps (gray boxes) starting from initial representations of atom types and positions (top), continuing with the construction of representations of chemical environments of atoms and atom pairs (middle) before using these to predict energy and Hamiltonian matrix respectively (bottom). The left path through the network to the energy prediction $E$ is rotationally invariant by design, while the right pass to the Hamiltonian matrix $\mat{H}$ allows for a maximum angular momentum $L$ of predicted orbitals by employing a multiplicative construction of the basis $\vec{\omega}_{ij}$ using sequential interaction passes $l=0 \dots 2L$. The onsite and offsite blocks of the Hamiltonian matrix are treated separately.
The prediction of overlap matrix $\mat{S}$ is performed analogously.
(\textbf{b})~Illustration of the SchNet interaction block~\cite{schutt2018schnet}.
(\textbf{c})~Illustration of SchNorb interaction block. The pairwise representation $\vec{h}^l_{ij}$ of atoms $i,j$ is constructed by a factorised tensor layer $f_\text{tensor}$ from atomic representations as well as the interatomic distance.
Using this, rotationally invariant interaction refinements $\mathbf{v}_{i}^m$ and basis coefficients $\vec{p}^l_{ij}$ are computed.
(\textbf{d})~Loewdin population analysis for uracil based on the density matrix calculated from the predicted Hamiltonian and overlap matrices.
(\textbf{e})~Mean abs. errors of lowest 20 orbitals (13 occupied + 7 virtual) of ethanol for Hartree-Fock and DFT@PBE.
(\textbf{f})~The predicted (solid black) and reference (dashed grey) orbital energies of an ethanol molecule for DFT. Shown are the last four occupied and first four unoccupied orbitals, including HOMO and LUMO.
The associated predicted and reference molecular orbitals are compared for four selected energy levels.
}
\label{fig:schnorb}
\end{figure*}
In quantum chemistry, the wavefunction associated with the electronic Hamiltonian $\hat{H}$ is typically expressed by anti-symmetrised products of single-electron functions or molecular orbitals.
These are represented in a local atomic orbital basis of spherical atomic functions $\ket{\psi_m}=\sum_i c_m^i \ket{\phi_i}$ with varying angular momentum.
As a consequence, one can write the electronic Schr\"odinger equation in matrix form
\begin{equation}
\mat{H} \vec{c}_m = \epsilon_m \mat{S} \vec{c}_m,
\end{equation}
where the Hamiltonian matrix $\mat{H}$ may correspond to the Fock or Kohn--Sham matrix, depending on the chosen level of theory~\cite{cramer2013essentials}.
In both cases, the Hamiltonian and overlap matrices are defined as:
\begin{equation}
H_{ij} = \braket{\phi_i|\hat{H}|\phi_j}
\end{equation}
and
\begin{equation}
S_{ij} = \braket{\phi_i|\phi_j}.
\end{equation}
The eigenvalues $\epsilon_m$ and electronic wavefunction coefficients $c_m^i$ contain the same information as $\mathbf{H}$ and $\mathbf{S}$ where the electronic eigenvalues are naturally invariant to rigid molecular rotations, translations or permutation of equivalent atoms.
Unfortunately, as a function of atomic coordinates and changing molecular configurations, eigenvalues and wavefunction coefficients are not well-behaved or smooth.
State degeneracies and electronic level crossings provide a challenge to the direct prediction of eigenvalues and wavefunctions with ML techniques.
We address this problem with a deep learning architecture that directly describes the Hamiltonian matrix in local atomic orbital representation.
\subsection{SchNOrb deep learning framework}
SchNOrb (SchNet for Orbitals) presents a framework that captures the electronic structure in a local representation of atomic orbitals that is common in quantum chemistry.
Fig.~\ref{fig:schnorb}a gives an overview of the proposed architecture.
SchNOrb extends the deep tensor neural network SchNet~\cite{schutt2017quantum} to represent electronic wavefunctions. The core idea is to construct symmetry-adapted pairwise features $\vec{\Omega}_{ij}^l$ to represent the block of the Hamiltonian matrix corresponding to atoms $i,j$.
They are written as a product of rotationally invariant ($\lambda=0$) and covariant ($\lambda>0$) components $\vec{\omega}_{ij}^{\lambda}$ which ensures that -- given a sufficiently large feature space -- all rotational symmetries up to angular momentum $l$ can be represented:
\begin{align}
\vec{\Omega}_{ij}^l &= \prod_{\lambda=0}^{l} \vec{\omega}_{ij}^{\lambda} \quad \text{with } 0 \leq l \leq 2L \label{eq:basis}\\
\vec{\omega}_{ij}^{\lambda} &=
\begin{cases}
\,\, \vec{p}_{ij}^\lambda \otimes \mathds{1}_{D} & \text{for } \lambda=0\\
\left[ \vec{p}_{ij}^\lambda \otimes \frac{\mathbf{r}_{ij}}{\|\mathbf{r}_{ij}\|} \right] \mat{W}^\lambda & \text{for } \lambda>0\\
\end{cases},
\end{align}
Here, $\mathbf{r}_{ij}$ is the vector pointing from atom $i$ to atom $j$, $\vec{p}_{ij}^\lambda \in \mathbb{R}^{B}$ are rotationally invariant coefficients and $\mat{W}^\lambda \in \mathds{R}^{3 \times D}$ are learnable parameters projecting the features along $D$ randomly chosen directions.
This allows to rotate the different factors of $\vec{\Omega}_{ij}^l \in \mathbb{R}^{B \cdot D}$ relative to each other and further increases the flexibility of the model for $D > 3$.
In case of $\lambda=0$, the coefficients are independent of the directions due to rotational invariance.
We obtain the coefficients $\vec{p}_{ij}^\lambda$ from an atomistic neural network as shown in Fig.~\ref{fig:schnorb}a.
Starting from atom type embeddings $\mathbf{x}_i^0$, rotationally invariant representations of atomistic environments $\mathbf{x}_i^T$ are computed by applying $T$ consecutive interaction refinements.
These are by construction invariant with respect to rotation, translation and permutations of atoms.
This part of the architecture is equivalent to the SchNet model for atomistic predictions (see Refs.~\cite{schutt2017schnet,schutt2018schnet}).
In addition, we construct representations of atom pairs $i,j$ that will enable the prediction of the coefficients $\vec{p}_{ij}^\lambda$.
This is achieved by $2L+1$ SchNOrb interaction blocks, which compute the coefficients $\vec{p}_{ij}^\lambda$ with a given angular momentum $\lambda$ with respect to the atomic environment of the respective atom pair $ij$. This corresponds to adapting the atomic orbital interaction based on the presence and position of atomic orbitals in the vicinity of the atom pair.
As shown in Fig.~\ref{fig:schnorb}b, the coefficient matrix depends on pair interactions $\mlp_\text{pair}$ of atoms $i,j$ as well as environment interactions $\mlp_\text{env}$ of atom pairs $(i,m)$ and $(n,j)$ for neighboring atoms $m,n$.
These are crucial to enable the model to capture the orientation of the atom pair within the molecule for which pair-wise interactions of atomic environments are not sufficient.
The Hamiltonian matrix is obtained by treating on-site and off-site blocks separately.
Given a basis of atomic orbitals up to angular momentum $L$, we require pair-wise environments with angular momenta up to $2L$ to describe all Hamiltonian blocks
\begin{align}
\tilde{\mat{H}}_{ij} &=
\begin{cases}
\mat{H}_\text{off} \left( \left[ \vec{\Omega}_{ij}^{l} \right]_{0 \leq l \leq 2L+1} \right) & \text{for } i \neq j\\ \\
\mat{H}_\text{on}\left( \left[ \vec{\Omega}_{im}^{l} \right]_{\substack{m \neq i \\ 0 \leq l \leq 2L+1}} \right) & \text{for } i = j
\end{cases},
\end{align}
The predicted Hamiltonian is obtained through symmetrisation $\mathbf{H} = \frac{1}{2} (\mathbf{\tilde{H}} + \mathbf{\tilde{H}}^\intercal )$. $\mat{H}_\text{off}$ and $\mat{H}_\text{on}$ are modeled by neural networks that are described in detail in the methods section. The overlap matrix $\mat{S}$ can be obtained in the same manner. In addition to the Hamiltonian and overlap matrices, we predict the total energy separately as a sum over atom-wise energy contributions, in analogy with the conventional SchNet treatment~\cite{schutt2018schnet}.
\subsection{Learning electronic structure and derived properties}
The proposed SchNOrb architecture allows us to perform predictions of total energies, Hamiltonian and overlap matrices in end-to-end fashion using a combined regression loss.
We train neural networks for several data sets of the molecules water, ethanol, malondialdehyde, and uracil.
The reference calculations were performed with Hartree-Fock (HF) and density functional theory (DFT) with the PBE exchange correlation functional~\cite{perdew1996generalized}. The employed Gaussian atomic orbital bases include angular momenta up to $l=2$ (d-orbitals).
Detailed model and training settings for each data set are listed in Supplementary Table~S1.
As Supplementary Table S2 shows, the total energies could be predicted up to a mean absolute error below 2~meV for the molecules. The predictions show mean absolute errors below 8~meV for the Hamiltonian and below $1 \cdot 10^{-4}$ for the overlap matrices.
We examine how these errors propagate to orbital energy and coefficients.
Fig.~\ref{fig:schnorb}e shows mean absolute errors for energies of the lowest 20 molecular orbitals for ethanol reference calculations using DFT as well as HF.
The errors for the DFT reference data are consistently lower.
Beyond that, the occupied orbitals (1-13) are predicted with higher accuracy ($<$20~meV) than the virtual orbitals ($\sim$100~meV).
We conjecture that the larger error for virtual orbitals arises from the fact that these are not strictly defined by the underlying data from the HF and Kohn-Sham DFT calculations. Virtual orbitals are only defined up to an arbitrary unitary transformation.
Their physical interpretation is limited and, in HF and DFT theory, they do not enter in the description of ground-state properties.
For the remaining data sets, the average errors of the occupied orbitals are $<$10~meV for water and malondialdehyde as well as 48~meV for uracil. This is shown in detail in Supplementary Fig. S1. The orbital coefficients are predicted with cosine similarities $\geq$~90\% (see Supplementary Fig.~S2).
Fig.~\ref{fig:schnorb}f depicts the predicted and reference orbital energies for the frontier MOs of ethanol (solid and dotted lines, respectively), as well as the orbital shapes derived from the coefficients.
Both occupied and unoccupied energy levels are reproduced with high accuracy, including the highest occupied (HOMO) and lowest unoccupied orbitals (LUMO).
This trend is also reflected in the overall shape of the orbitals.
Even the slightly higher deviations in the orbital energies observed for the third and fourth unoccupied orbital only result in minor deformations.
The learned covariance of molecular orbitals for rotations of a water molecule is shown in Supplementary Fig. S3.
As SchNOrb learns the electronic structure of molecular systems, all chemical properties that are defined as quantum mechanical operators on the wavefunctions can be computed from the ML prediction without the need to train a separate model.
We investigate this feature by directly calculating electronic dipole and quadrupole moments from the orbital coefficients predicted by SchNOrb.
The corresponding mean absolute errors are reported in Supplementary Tab.~S4.
Excellent agreement with the electronic structure reference is observed for the majority of molecules ($<$0.054 D for dipoles and $<$0.058 D\,\AA\ for quadrupoles).
The only deviation from this trend is observed for uracil, which exhibits a more complicated electronic structure due to its delocalised $\pi$-system (see above).
In this case, a similar accuracy as the other methods could in principle be reached upon the addition of more reference data points.
The above results demonstrate the utility of combining a learned Hamiltonian with quantum operators. This makes it possible to access a wide range of chemical properties without the need for explicitly developing specialised neural network architectures.
\subsection{Chemical insights from electronic deep learning}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{figures/orbital_curves_narrow}
\caption{
\small \textbf{Proton transfer in malondialdehyde.} (\textbf{a}) Excerpt of the MD trajectory showing the proton transfer, the electron density as well as the relevant MOs HOMO-2 and HOMO-3 for three configurations (I, II, III).
(\textbf{b}) Forces exerted by the MOs on the transferred proton for configurations I and II.
(\textbf{c}) Density of states broadened across the proton transfer trajectory. MO energies of the equilibrium structure are indicated by gray dashed lines. The inset shows a zoom of HOMO-2 and HOMO-3.
}
\label{fig:trajectories}
\end{figure*}
Recently, a lot of research has focused on explaining predictions of ML models~\cite{Bach2015,montavon2017methods,kindermans2018learning} aiming both at the validation of the model~\cite{kim2017interpretability,lapuschkin2019unmasking} as well as the extraction of scientific insight~\cite{de2016comparing,schutt2017quantum,jha2018elemnet}.
However, these methods explain ML predictions either in terms of the input space, atom types and positions in this case, or latent features such as local chemical potentials~\cite{schutt2017quantum,schutt2018quantum}.
In quantum chemistry however, it is more common to analyse electronic properties in terms of the MOs and properties derived from the electronic wavefunction, which are direct output quantities of the SchNOrb architecture.
Molecular orbitals encode the distribution of electrons in a molecule, thus offering direct insights into its underlying electronic structure. They form the basis for a wealth of chemical bonding analysis schemes, bridging the gap between quantum mechanics and abstract chemical concepts, such as bond orders and atomic partial charges~\cite{cramer2013essentials}.
These quantities are invaluable tools in understanding and interpreting chemical processes based on molecular reactivity and chemical bonding strength.
As SchNOrb yields the MOs, we are able to apply population analysis to our ML predictions.
Fig.~\ref{fig:schnorb}d shows Loewdin partial atomic charges and bond orders for the uracil molecule.
Loewdin charges provide a chemically intuitive measure for the electron distribution and can e.g. aid in identifying potential nucleophilic or electrophilic reaction sites in a molecule.
The negatively charged carbonyl oxygens in uracil, for example, are involved in forming RNA base pairs.
The corresponding bond orders provide information on the connectivity and types of bonds between atoms.
In the case of uracil, the two double bonds of the carbonyl groups are easily recognizable (bond order 2.12 and 2.14, respectively).
However, it is also possible to identify electron delocalisation effects in the pyrimidine ring, where the carbon double bond donates electron density to its neighbors.
A population analysis for malondialdehyde, as well as population prediction errors for all molecules can be found in Supplementary Fig. S4. and Table S3.
The SchNOrb architecture enables an accurate prediction of the electronic structure across molecular configuration space, which provides for rich chemical interpretation during molecular reaction dynamics. Fig.~\ref{fig:trajectories}a shows an excerpt of a molecular dynamics simulation of malondialdehyde that was driven by atomic forces predicted using SchNOrb.
It depicts the proton transfer together with the relevant MOs and the electronic density.
The density paints an intuitive picture of the reaction as it migrates along with the hydrogen.
This exchange of electron density during proton transfer is also reflected in the orbitals. Their dynamical rearrangement indicates an alternation between single and double bonds.
The latter effect is hard to recognise based on the density alone and demonstrates the wealth of information encoded in the molecular wavefunctions.
Fig.~\ref{fig:trajectories}b depicts the forces the different MOs exert onto the hydrogen atom exchanged during the proton transfer.
All forces are projected onto the reaction coordinate, where positive values correspond to a force driving the proton towards the product state. In the initial configuration I, most forces lead to attraction of the hydrogen atom to the right oxygen. In the intermediate configuration II, orbital rearrangement results in a situation where the majority of orbitals forces on the hydrogen atom become minimal, representing mostly non-bonding character between oxygens and hydrogen.
One exception is MO 13, depicted in the inset of Fig.~\ref{fig:trajectories}b.
Due to a minor deviation from a symmetric O-H-O arrangement, the orbital represents a one-sided O-H bond, exerting forces that promote the reaction.
The intrinsic fluctuations during the proton transfer molecular dynamics are captured by the MOs as can be seen in Fig.~\ref{fig:trajectories}c. This shows the distribution of orbital energies encountered during the reaction.
As would be expected, both HOMO-2 and HOMO-3 (inset, orange and blue respectively), which strongly participate in the proton transfer, show significantly broadened peaks due to strong energy variations in the dynamics. This example nicely shows the chemically intuitive interpretation that can be obtained by the electronic structure prediction of SchNOrb.
\subsection{Deep learning-enhanced quantum chemistry}
\begin{figure*}
\centering
\includegraphics[width=0.85\textwidth]{figures/fig3}
\caption{\small \textbf{Applications of SchNOrb.} (\textbf{a}) Optimisation of the HOMO-LUMO gap. HOMO and LUMO with energy levels are shown for a randomly drawn configuration of the malonaldehyde dataset (centre) as well as for configurations that were obtained from minimising or maximising the HOMO-LUMO gap prediction using SchNOrb (left and right, respectively). For the optimised configurations, the difference of the orbitals are shown in green (increase) and violet (decrease). The dominant geometrical change is indicated by the black arrows.
(\textbf{b}) The predicted MO coefficients for the uracil configurations from the test set are used as a wavefunction guess to obtain accurate solutions from DFT at a reduced number of self-consistent-field (SCF) iterations. This reduces the required SCF iterations by an average of 77\%.
}
\label{fig:applications}
\end{figure*}
An essential paradigm of chemistry is that the molecular structure defines chemical properties. Inverse chemical design turns this paradigm on its head by enabling property-driven chemical structure exploration.
The SchNOrb framework constitutes a suitable tool to enable inverse chemical design due to its analytic representation of electronic structure in terms of the atomic positions. We can therefore obtain analytic derivatives with respect to the atomic positions, which provide the ability to optimise electronic properties.
Fig.~\ref{fig:applications}a shows the minimisation and maximisation of the HOMO-LUMO gap $\epsilon_\text{gap}$ of malondialdehyde as an example.
We perform gradient descent and ascent from a randomly selected configuration $\mathbf{r}_\text{ref}$ until convergence at $\mathbf{r}_\text{min}$ and $\mathbf{r}_\text{max}$, respectively.
We are able to identify structures which minimise and maximise the gap from its initial $3.15$~eV to $2.68$~eV at $\mathbf{r}_\text{min}$ and $3.59$~eV at $\mathbf{r}_\text{max}$.
While in this proof of concept these changes were predominantly caused by local deformations in the carbon-carbon bonds indicated in Fig.~\ref{fig:applications}a, they present an encouraging prospect how electronic surrogate models such as SchNOrb can contribute to computational chemical design using more sophisticated optimisation methods, such as alchemical derivatives~\cite{to2016guiding} or reinforcement learning~\cite{you2018graph}.
ML applications in chemistry have traditionally been one-directional, i.e. ML models are built on quantum chemistry data. Models such as SchNOrb offer the prospect of providing a deeper integration and feedback loop of ML with quantum chemistry methods. SchNOrb directly predicts wavefunctions based on quantum chemistry data, which in turn, can serve as input for further quantum chemical calculations. For example, in the context of HF or DFT calculations, the relevant equations are solved via a self-consistent field approach (SCF) that determines a set of MOs.
The convergence with respect to SCF iteration steps largely determines the computational speed of an electronic structure calculation and strongly depends on the quality of the initialisation for the wavefunction. The coefficients predicted by SchNOrb can serve as such an initialisation of SCF calculations.
Fig.~\ref{fig:applications}b depicts the SCF convergence for three sets of computations on the uracil molecule:
using the standard initialisation techniques of quantum chemistry codes, and the SchNorb coefficients with or without a second order solver.
Nominally, only small improvements are observed using SchNorb coefficients in combination with a conventional SCF solver.
This is due to the various strategies employed in electronic structure codes in order to provide a numerically robust SCF procedure.
Only by performing SCF calculations with a second order solver, which would not converge using a less accurate starting point than our SchNorb MO coefficients, the efficiency of our combined ML and second order SCF approach becomes apparent.
Convergence is obtained in only a fraction of the original time, reducing the number of cycles by $\sim77\%$.
Similarly, Supplementary Fig.~S5 shows the reduction of SCF iteration by $\sim73\%$ for malondialdehyde.
It should be noted, that this significantly faster, combined approach does not introduce any approximations into the electronic structure method itself and yields exactly the same results as the full computation.
Another example of integration of the SchNOrb deep learning framework with quantum chemistry is the use of predicted wavefunctions and MO energies based on Hartree--Fock as starting point for post-Hartree--Fock correlation methods such as M{\o}ller-Plesset perturbation theory (MP2). Supplementary Table S5 presents the mean absolute error of an MP2 calculation for ethanol based on wavefunctions predicted from SchNOrb. The associated prediction error for the test set is 83 meV.
Compared to the overall HF and MP2 energy, the relative error of SchNOrb amounts to 0.01~\% and 0.06~\%, respectively.
For the MP2 correlation energy, we observe a deviation of 17~\%, the reason of which is inclusion of virtual orbitals in the calculation of the MP2 integrals.
However, even in this case, the total error only amounts to a deviation of 93~meV.
\section{Discussion}
The SchNOrb framework provides an analytical expression for the electronic wavefunctions in a local atomic orbital representation as a function of molecular composition and atom positions.
The rotational covariance of atomic orbital interactions is encoded by basis functions with angular momentum $l>0$ which constitutes significant progress over previous work~\cite{hegde2017machine}.
As a consequence, the model provides access to atomic derivatives of wavefunctions, which include molecular orbital energy derivatives, Hamiltonian derivatives, which can be used to approximate nonadiabatic couplings~\cite{Maurer2016}, as well as higher order derivatives that describe the electron-nuclear response of the molecule.
Thus, the SchNOrb framework preserves the benefits of interatomic potentials while enabling access to the electronic structure as predicted by quantum chemistry methods.
SchNOrb opens up completely new applications to ML-enhanced molecular simulation.
This includes the construction of interatomic potentials with electronic properties that can facilitate efficient photochemical simulations during surface hopping trajectory dynamics or Ehrenfest-type mean-field simulations, but also enables the development of new ML-enhanced approaches to inverse molecular design via electronic property optimisation.
On the basis of the SchNOrb framework, intelligent preprocessing of quantum chemistry data in the form of effective Hamiltonians or optimised minimal basis representations~\cite{Lu2004} can be developed in the future. Such preprocessing will also pave the way towards the prediction of the electronic structure based on post-HF correlated wavefunction methods and post-DFT quasiparticle methods.
This work serves as a first proof of principle that direct ML models of electronic structure based on quantum chemistry can be constructed and used to enhance further quantum chemistry calculations. We have presented an immediate consequence of this by reducing the number of DFT-SCF iterations with wavefunctions predicted via SchNOrb. The presented model delivers derived electronic properties that can be formulated as quantum mechanical expectation values. This provides an important step towards a full integration of ML and quantum chemistry into the scientific discovery cycle.
\section{Methods}
\subsection{Reference data}
All reference calculations were carried out with the ORCA quantum chemistry code~\cite{Neese2012WCMS} using the def2-SVP basis set~\cite{Weigend2005PCCP}.
Integration grid levels of 4 and 5 were employed during SCF iterations and the final computation of properties, respectively.
Unless stated otherwise, the default ORCA SCF procedure was used, which is based on the Pulay method~\cite{PULAY1980393}.
For the remaining cases, the Newton--Raphson procedure implemented in ORCA was employed as a second order SCF solver.
SCF convergence criteria were set to VeryTight.
DFT calculations were carried out using the PBE functional~\cite{Perdew1996PRL}.
For ethanol, additional HF computations were performed.
Molecular dynamics simulations for malondialdehyde were carried out with SchNetPack~\cite{schuett2018schnetpack}.
The equations of motions were integrated using a timestep of 0.5~fs.
Simulation temperatures were kept at 300~K with a Langevin thermostat~\cite{PhysRevE.75.056707} employing a time constant of 100~fs.
Trajectories were propagated for a total of 50~ps, of which the first 10~ps were discarded.
\subsection{Details on the neural network architecture}
In the following we describe the neural network depicted in Fig.~\ref{fig:schnorb} in detail.
\subsubsection{Basic building blocks and notation}
We use shifted softplus activation functions
\begin{align}
\ssp(x) = \ln\left(\frac{1}{2} e^x + \frac{1}{2}\right)
\end{align}
throughout the architecture.
Linear layers are written as
\begin{align}
\linear(\vec{x}) = \mat{W}^\intercal \vec{x} + \vec{b}
\end{align}
with input $\vec{x} \in \mathbb{R}^{n_\text{in}}$, weights $\mat{W} \in \mathbb{R}^{n_\text{in} \times n_\text{out}}$ and bias $\vec{b} \in \mathbb{R}^{n_\text{out}}$.
Fully-connected neural networks with one hidden layer are written as
\begin{align}
\mlp(\vec{x}) =
\mat{W}_2^\intercal \ssp\left(\mat{W}_1^\intercal \vec{x} + \vec{b}_1\right) + \vec{b}_2
\end{align}
with weights $\mat{W}_1 \in \mathbb{R}^{n_\text{in} \times n_\text{hidden}}$ and $\mat{W}_2 \in \mathbb{R}^{n_\text{hidden} \times n_\text{out}}$ and biases $\vec{b}_1, \vec{b}_2$ accordingly.
Model parameters are shared within layers across atoms and interactions, but never across layers.
We omit layer indices for clarity.
\subsubsection{Atomic environments}
The representations of atomic environments are constructed with the neural network structure as in SchNet.
In the following, we summarise this first part of the model. For further details, please refer to~\citet{schutt2018schnet}.
First, each atom is assigned an initial element-specific embedding
\begin{align}
\mathbf{x}_i^0 &= \mathbf{a}_{Z_i} \in \mathbb{R}^{B},
\end{align}
where $Z_i$ is the nuclear charge and $B$ is the number of atom-wise features.
In this work, we use $B=1000$ for all models.
The representations are refined using SchNet interaction layers (Fig.~\ref{fig:schnorb}\textbf{b}).
The main component is a continuous-filter convolutional layer (cfconv)
\begin{multline}
\cfconv((\mathbf{x}_1, r_1), \dots, (\mathbf{x}_n, r_n)) = \\
\left[ \sum_{j \neq i} \mathbf{x}_j \circ \mat{W}_\text{filter}(r_{ij}) \right]_{i=1 \dots n}
\end{multline}
which takes a spatial filter $\mat{W}_\text{filter}: \mathbb{R} \rightarrow \mathbb{R}^{B}$
\begin{align}
\mat{W}_\text{filter}(r_{ij}) &= \mlp(\vec{g}(r_{ij})) \, \text{f}_\text{cutoff}(r_{ij}) \label{eq:filter}
\end{align}
with
\begin{align}
\vec{g}(r_{ij}) &= \left[ \exp ( -\gamma (r_{ij} - k \Delta\mu )^2 ) \right]_{0 \leq k \leq r_{\text{c}}/\Delta \mu} \label{eq:radialbasis} \\
\text{f}_\text{cutoff}(r_{ij}) &= \begin{cases}
0.5 \times \left[1 + \cos\left(\frac{\pi r}{r_\text{c}}\right)\right]
& r < r_\text{c} \\
0 & r \geqslant r_\text{c}
\end{cases},
\end{align}
where $r_\text{c}$ is the cutoff radius and $\Delta\mu$ is the grid spacing of the radial basis function expansion of interatomic distance $r_{ij}$.
While this adds spatial information to the environment representations for each feature separately, the crosstalk between features is performed atom-wise by fully-connected layers (linear and $\text{mlp}_\text{atom}$ in Fig.~\ref{fig:schnorb}\textbf{b}) to obtain the refinements $\vec{v}^t_i$, where $t$ is the current interaction iteration.
The refined atom representations are then
\begin{align}
\vec{x}_i^t &= \vec{x}_i^{t-1} + \vec{v}^t_i.
\end{align}
These representations of atomic environments are employed by SchNet to predict chemical properties via atom-wise contributions.
However, in order to extend this scheme to the prediction of the Hamiltonian, we need to construct representations of pair-wise environments in a second interaction phase.
\subsubsection{Pair-wise environments and angular momenta}
The Hamiltonian matrix is of the form
\begin{align}
\mat{H} &=
\left[\begin{array}{ccccc}
\mat{H}_{11} & \cdots & \mat{H}_{1j} &\cdots & \mat{H}_{1n}\\
\vdots & \ddots & \vdots & & \vdots\\
\mat{H}_{i1} & \cdots & \mat{H}_{ij} & \cdots & \mat{H}_{in}\\
\vdots & & \vdots & \ddots &\vdots \\
\mat{H}_{n1} & \cdots & \mat{H}_{nj} & \cdots & \mat{H}_{nn}\\
\end{array}\right] \label{eq:hamblock}
\end{align}
where a matrix block $\mat{H}_{ij} \in \mathbb{R}^{n_{\text{ao},i} \times n_{\text{ao},j}}$ depends on the atoms $i,j$ within their chemical environment as well as on the choice of $n_{\text{ao},i}$ and $n_{\text{ao},j}$ atomic orbitals, respectively.
Therefore, SchNOrb builds representations of these embedded atom pairs based on the previously constructed representations of atomic environments.
This is achieved through the SchNOrb interaction module (see Fig.~\ref{fig:schnorb}\textbf{c}).
First, a raw representation of atom pairs is obtained using a factorised tensor layer~\cite{schutt2017quantum,sutskever2011generating}:
\begin{align}
\vec{h}_{ij}^\lambda &= \ftens(\mathbf{x}_i^\lambda, \mathbf{x}_j^\lambda, r_{ij}) \\
& = \ssp ( \, \linear_2 [ \nonumber \\
&\quad\quad \linear_{1} (\vec{x}_i^\lambda) \circ \linear_{1}(\vec{x}_j^\lambda) \circ \mat{W}_\text{filter}(r_{ij}) \nonumber \\
& ] ). \nonumber
\end{align}
The layers $\linear_{1}: \mathbb{R}^{B} \mapsto \mathbb{R}^{B}$ map the atom representations to the factors, while the filter-generating network $\mat{W}_\text{filter}: \mathbb{R} \rightarrow \mathbb{R}^{B}$ is defined analogously to Eq.~\ref{eq:filter} and directly maps to the factor space.
In analogy to how the SchNet interactions are used to build atomic environments, SchNOrb interactions are applied several times, where each instance further refines the atomic environments of the atom pairs with additive corrections
\begin{align}
\vec{x}_i^{T+\lambda} &= \vec{x}_i^{T+\lambda-1} + \vec{v}_i^{T+\lambda} \\
\vec{v}_i^{T+\lambda} &= \mlp_\text{atom} \left( \sum_{j \neq i} \vec{h}_{ij}^l \right)
\end{align}
as well as constructs pair-wise features:
\begin{align}
\vec{p}_{ij}^\lambda &= \vec{p}_{ij}^{\lambda,\text{pair}} + \sum_{m \neq i} \vec{p}_{mj}^{\lambda,\text{env}} + \sum_{n \neq j} \vec{p}_{in}^{\lambda,\text{env}} \\
\vec{p}_{ij}^{\lambda,\text{pair}} &= \mlp_\text{pair}(\vec{h}_{ij}^\lambda) \\
\vec{p}_{ij}^{\lambda,\text{env}} &= \mlp_\text{env}(\vec{h}_{ij}^\lambda)
\end{align}
where $\mlp_\text{pair}$ models the direct interactions of atoms $i,j$ while $\mlp_\text{env}$ models the interactions of the pair with neighboring atoms.
As described above, the atom pair coefficients are used to form a basis set
\begin{align*}
\vec{\omega}_{ij}^{\lambda} &=
\begin{cases}
\,\, \vec{p}_{ij}^0 \otimes \mathds{1}_{D} & \text{for } \lambda=0\\
\left[ \vec{p}_{ij}^\lambda \otimes \frac{\mathbf{r}_{ij}}{\|\mathbf{r}_{ij}\|} \right] \mat{W} & \text{for } \lambda>0 \\
\end{cases}
\end{align*}
where $\lambda$ corresponds to the angular momentum channel and $\mat{W} \in \mathds{R}^{3 \times D}$ are learnable parameters to project along $D$ directions.
For all results in this work, we used $D=4$.
For interactions between s-orbitals, we consider the special case $\lambda=0$ where the features along all directions are equal due to rotational invariance.
At this point, $\vec{\omega}_{ij}^{0}$ is rotationally invariant and $\vec{\omega}_{ij}^{\lambda>0}$ is covariant.
On this basis, we obtain features with higher angular momenta using:
\begin{align*}
\vec{\Omega}_{ij}^l &= \prod_{\lambda=0}^{l} \vec{\omega}_{ij}^{\lambda} \quad \text{with } 0 \leq l \leq 2L,
\end{align*}
where features $\vec{\Omega}_{ij}^l$ possess angular momentum $l$.
The SchNOrb representations of atom pairs embedded in their chemical environment, that were constructed from the previously constructed SchNet atom-wise features, will serve in a next step to predict the corresponding blocks of the ground-state Hamiltonian.
\subsubsection{Predicting the target properties}
Finally, we assemble the Hamiltonian and overlap matrices to be predicted.
Each atom pair block is predicted from the corresponding features $\vec{\Omega}_{ij}^l$:
\begin{align*}
\tilde{\mat{H}}_{ij} &=
\begin{cases}
\mat{H}_\text{off} \left( \left[ \vec{\Omega}_{ij}^{l} \right]_{0 \leq l \leq 2L+1} \right) & \text{for } i \neq j\\ \\
\mat{H}_\text{on}\left( \left[ \vec{\Omega}_{im}^{l} \right]_{\substack{m \neq i \\ 0 \leq l \leq 2L+1}} \right) & \text{for } i = j
\end{cases},
\end{align*}
where we restrict the network to linear layers in order to conserve the angular momenta:
\begin{align*}
\mat{H}_\text{off}(\cdot) &= \sum_{l} \linear_\text{off}^l\left(\vec{\Omega}_{ij}^l \right)[: \! n_{\text{ao},i}, :\!n_{\text{ao},j}] \\
\mat{H}_\text{on}(\cdot) &= \sum_{j,l} \linear_\text{on}^l\left( \vec{\Omega}_{ij}^l \right) [: \! n_{\text{ao},i}, :\!n_{\text{ao},i}]
\end{align*}
with $\linear_\text{off}, \linear_\text{on}: \mathbb{R}^{2L+1 \times D} \rightarrow \mathbb{R}^{n_\text{ao,max} \times n_\text{ao,max}}$, i.e. mapping to the maximal number of atomic orbitals in the data.
Then, a mask is applied to the matrix block to yield only $\tilde{\mat{H}}_{ij} \in \mathbb{R}^{n_{\text{ao},i} \times n_{\text{ao},j}}$.
Finally, we symmetrise the prediced Hamiltonian:
\begin{align}
\mathbf{H} = \frac{1}{2} (\mathbf{\tilde{H}} + \mathbf{\tilde{H}}^\intercal)
\end{align}
The overlap matrix is obtained similarly with blocks
\begin{align*}
\mat{S}_\text{off}\left( \cdot \right) &= \linear_\text{S,on}(\vec{\Omega}_{ij}) [: \! n_{\text{ao},i}, :\!n_{\text{ao},j}]\\
\mat{S}_{ii} &= \mat{S}_{Z_i}.
\end{align*}
The prediction of the total energy is obtained anologously to SchNet as a sum over atom-wise energy contributions:
\begin{align*}
E &= \sum_i \mlp_\text{E}(\mathbf{x}_i).
\end{align*}
\subsection{Data augmentation}
While SchNOrb constructs features $\vec{\Omega}_{ij}^l$ and $\vec{\Omega}_{ii}^l$ with angular momenta such that the Hamiltonian matrix can be represented as a linear combination of those, it does not encode the full rotational symmetry a priori.
However, this can be learned by SchNOrb assuming the training data reflects enough rotations of a molecule.
To save computing power, we reduce the amount of reference calculations by randomly rotating configurations before each training epoch using Wigner $\mathcal{D}$ rotation matrices.~\cite{Schober2016}
Given a randomly sampled rotor $R$, the applied transformations are
\begin{align}
\tilde{\mathbf{r}}_{i} &= \mat{\mathcal{D}}^{(1)}(R) \mathbf{r}_i \\
\tilde{\vec{F}}_{i} &= \mat{\mathcal{D}}^{(1)}(R) \vec{F}_i \\
\tilde{\mathbf{H}}_{\mu \nu} &= \mat{\mathcal{D}}^{(l_\mu)}(R) \mathbf{H}_{\mu \nu} \mat{\mathcal{D}}^{(l_\nu)}(R) \\
\tilde{\mathbf{S}}_{\mu \nu} &= \mat{\mathcal{D}}^{(l_\mu)}(R) \mathbf{S}_{\mu \nu} \mat{\mathcal{D}}^{(l_\nu)}(R)
\end{align}
for atom positions $\mathbf{r}_i$, atomic forces $\vec{F}_i$, Hamiltonian matrix $\mathbf{H}$, and overlap $\mathbf{S}$.
\subsection{Neural network training}
For the training, we used a combined loss to train on energies $E$, atomic forces $\mat{F}$, Hamiltonian $\mat{H}$ and overlap matrices $\mat{S}$ simultaneously:
\begin{multline}
\ell\left[(\tilde{\mat{H}}, \tilde{\mat{S}}, \tilde{E}), (\mat{H}, \mat{S}, E, \mathbf{F}) \right] = \\
\|\mat{H} - \tilde{\mat{H}} \|^2_F + \|\mat{S} - \tilde{\mat{S}} \|^2_F \, + \rho \, \|E - \tilde{E} \|^2 \\
+ \frac{1}{n_\text{atoms}} \sum_{i=0}^{n_\text{atoms}} \left \| \mathbf{F}_i - \left (-\frac{\partial \tilde{E}}{\partial \mathbf{r}_i}\right ) \right\|^2
\label{eq:loss}
\end{multline}
where the variables marked with a tilde refer to the corresponding predictions.
The neural networks were trained with stochastic gradient descent using the ADAM optimiser~\cite{kingma2014adam}.
We reduced the learning rate using a decay factor of $0.8$ after $t_\text{patience}$ epochs without improvement of the validation loss.
The training is stopped at $\text{lr} \leq 5 \cdot 10^{-6}$.
The mini-batch sizes, patience and data set sizes are listed in Supplementary Table~S1.
Afterwards, the model with lowest validation error is selected for testing.
\begin{acknowledgments}
All authors gratefully acknowledge support by the Institute of Pure and Applied Mathematics (IPAM) at the University of California Los Angeles during a long program workshop. RJM acknowledges funding through a UKRI Future Leaders Fellowship (MR/S016023/1).
KTS and KRM acknowledge support by the Federal Ministry of Education and Research (BMBF) for the Berlin Center for Machine Learning (01IS18037A).
This project has received funding from the European Unions Horizon 2020 research and innovation program under the Marie Sk\l{}odowska-Curie grant agreement No. 792572.
Computing resources have been provided by the Scientific Computing Research Technology Platform of the University of Warwick, and the EPSRC-funded high end computing Materials Chemistry Consortium (EP/R029431/1).
KRM acknowledges partial financial support by the German Ministry for Education and Research (BMBF) under Grants 01IS14013A-E, 01GQ1115 and 01GQ0850; Deutsche Forschungsgesellschaft (DFG) under Grant Math+, EXC 2046/1, Project ID 390685689 and by the Technology Promotion (IITP) grant funded by the Korea government (No. 2017-0-00451, No. 2017-0-01779). Correspondence to RJM, KRM and AT.
\end{acknowledgments}
|
\section{Introduction}
Sequence labeling is one of the most fundamental NLP models, which is used for many tasks such as named entity recognition (NER), chunking, word segmentation and part-of-speech (POS) tagging. It has been traditionally investigated using statistical approaches~\cite{Lafferty:2001}, where conditional random fields (CRF) ~\cite{Lafferty:2001} has been proven to be an effective framework, by taking discrete features as the representation of input sequence~\cite{Keerthi:2007}. With the advances of deep learning, neural sequence labeling models have achieved state-of the-art results for many tasks~\cite{Peters:2017}.
For the purpose of this paper, we consider neural network solution for multilingual named entity recognition for Bulgarian, Czech, Polish and Russian languages for the BSNLP 2019 Shared Task \cite{shared-task-2019}. Our solution is based on BERT language model~\cite{bert}, use bidirectional LSTM~\cite{lstm}, Multi-Head attention \cite{isalluneed}, NCRFpp~\cite{ncrfpp} (being neural network version of CRF++framework for sequence labelling) and Pooling Classifier (for language classification) on the top as additional information.
\section{Task Description}
\subsection{Data Format}
The data consists of raw documents and the annotations, separately provided by the organizers. Each annotation contains a set of extracted entities and their types without duplication. We convert each raw document and corresponding annotations to labeled sequence and predict named entity label for each token in the input sentence.
The documents are categorized into topics. There are two topics in the dataset released first: named ``brexit'' and ``asia\_bibi''.
\subsection{Tasks}
The BSNLP Shared Task has three parts \cite{shared-task-2019}:
\begin{enumerate}
\item Named Entity Mention Detection and Classification;
\item Name Lemmatization;
\item Cross-lingual entity Matching.
\end{enumerate}
For more details about the dataset and the task refer to the description on the web page\footnote{Full BSNLP Shared Task description available at http://bsnlp.cs.helsinki.fi/shared\_task.html.}.
We focused on Named Entity Mention Detection (Named Entity Recognition) in this work.
\section{System Description}
We propose modeling the task as both sequence labeling and language classification jointly with a neural architecture to learn additional information about text. The model consists of one encoder, which on its own is build from the pretrained multilingual BERT model, followed by several trainable layers and two decoders. While the first decoder generates output tags, the second decoder identifies the language of the input sentence\footnote{Our code is available at https://github.com/anonymize/slavic-ner. This code is based on https://github.com/sberbank-ai/ner-bert.}. The system architecture is presented in Figure~\ref{figure-arch} and consists of seven parts:
\begin{enumerate}
\item BERT Embedder as pretrained multilingual language model;
\item Weighted aggregation of BERT output;
\item Recurrent BiLSTM layer to be trained for the NER task;
\item Multi-Head attention to take shorter dependencies between words into account;
\item linear layer as the head of the encoder part;
\item NCRF++ inference layer for decoding, i.e. final sequence labelling;
\item Concatenation operation of Max Pooling, Average Pooling and last output of Multi-Head attention layer, later passed to linear layer for classification as a second decoder for language identification.
\end{enumerate}
\begin{figure*}[htbp]
\centering
\includegraphics[width=14cm]{BSNLP_arch.png}
\caption{\label{figure-arch}The system architecture}
\end{figure*}
\subsection{Neural Network Architecture}
\subsubsection{BERT Embedder}
The BERT embeddings layer contains Google's original implementation of multilingual BERT language model. Each sentence is preprocessed as described in BERT paper \cite{bert}:
\begin{enumerate}
\item Process input text sequence to WordPiece embeddings \cite{wordpiece} with a 30,000 token vocabulary and pad to 512 tokens.
\item Add first special BERT token marked ``[CLS]''.
\item Mark all tokens as members of part ``A'' of the input sequence.
\end{enumerate}
But instead of BERT's original paper \cite{bert} we keep ``B'' (``Begin'') prefix for labels and do a prediction for ``X'' labels on training stage. BERT neural network is used only to embed input text and don't fine-tune on the training stage. We freeze all layers except dropout here, that decreases overfitting.
We take hidden outputs from all BERT layers as the output of this part of the neural network and pass to the next level of the neural network. So the shape of output is $12\times 768$ for each token of $512$ length's padded input sequence.
\subsubsection{BERT Weighting}
Here we sum all of BERT hidden outputs from previous part:
\begin{equation}
o_i = \gamma \times \sum_{i=0}^{m-1} b_i s_i
\end{equation}
where
\begin{itemize}
\item $o_i$ is output vector of size $768$;
\item $m=12$ is the number hidden layers in BERT;
\item $b_i$ is output from $i$ BERT hidden layer;
\item $\gamma$ and $s_i$ is trainable task specific parameters.
\end{itemize}
As we do not fine-tune BERT, we should adapt its outputs for our specific sequence labeling task. The suggested weighting approach is similar to ELMo \cite{elmo}, with a lower number of weighting vectors parameters $s_i$. This approach can help to learn importance of each BERT output layer for this task and and network doesn't lose too much information about text, that was stored in all BERT outputs.
\subsubsection{Recurrent Part}
This part contains two LSTM networks for forward and backward passes with $512$ hidden units so that the output representation dim is $1024$ for each token. We use a recurrent layer for learning dependencies between tokens in an input sequence \cite{lstm}.
\subsubsection{Multi-Head Attention}
After applying the recurrent layer, we use Self-attention mechanism to learn any other dependencies in a sequence for each token.
This can be denoted as $D(d_h|S)$, where $D$ is some hidden dependency; $d_h$ is the $h$ head of attention, and $S$ is all sequence. each head can learn its dependencies such as morphological, syntactic or semantic relationships between words (tokens). Presumably, dependencies may look as shown at Figure 2. Also, mechanism attention can compensate limitations of the recurrent layer when working with long sequences \cite{bandanau}. In our architecture, we use multihead-attention block as proposed in the paper ``attention is all you need'' \cite{isalluneed}. We took $6$ heads and value and key dim $64$.
\subsubsection{Inference for NER Task}
After the input sequence was encoded, we achieve the final representation of each token in a sequence. This representation is passed to Linear layer with $tanh$ activation function and gets a vector with $14$ dim, that equals to the number of entities labels (include supporting labels ``pad'' and ``[CLS]''). The inference layer takes the extracted token sequence representations as features and assigns labels to the token sequence. As the inference layer, we use Neural CRF++ layer instead of vanilla CRF. That captures label dependencies by adding transition scores between neighboring labels. NCRF++ supports CRF trained with the sentence-level maximum log-likelihood loss. During the decoding process, the Viterbi algorithm is used to search the label sequence with the highest probability. But also, NCRF++ extends the decoding algorithm with the support of $nbest$ output \cite{ncrfpp}. We chose the $nbest$ parameter equal to $11$, because we have $11$ meaningful labels. In this decision we followed the original article \cite{ncrfpp}.
\subsubsection{Inference for Language Classification}
We train our system for language classification. For the classification inference, we use Pooling Linear Classifier block as proposed in ULMFiT paper \cite{ulmfit}. We pass output sequence representation $H$ from Multihead-attention part to different Poolings and concat (as shown in Figure~\ref{figure-arch}):
\begin{equation}
h_c = [h_0, maxpool(H), meanpool(H)]
\end{equation}
where $[]$ is concatenation;
$h_0$ is first output significant vector of Multihead-attention part (which does have ``[CLS]'' label).
The result of concat Pooling ($3\times1024$) is passed to Linear layer, and that predicts probability for four language classes (Bulgarian, Czech, Polish and Russian).
\subsection{Postprocessing Prediction}
After getting labels for the sequence of WordPiece tokens, we should convert prediction to word level labels extraction named entities. Each WordPiece token in the word is matched with neural network label prediction. We use ensemble classifier on labels by count all predicted labels for one word except ``X'' and select label for a word with the higher number of votes.
For final prediction we unite token's sequences which have not ``O'' (``Other'') label to spans and write to result of entities set.
\section{Training the System}
\subsection{Data Conversion}
On the training stage we divide the input data into two parts: the training set (named ``brexit'') and development set (named ``asia\_bibi''). Hence we train the system on one topic and evaluate the system on another topic. Because the input contains raw text and annotation, but BERT take words sequence as input, we convert data to word level IOB markup \cite{iob}. After that, each word was tokenized by WordPiece tokenizer and word label matched with IOBX labels.
On the prediction stage result, labels were received by voice classifier. After this, we transform word predictions to spans markup. The results of develop evaluation stage described in Table~\ref{dev-metrics-table}.
After evaluation stage we train our network on all input data (``brexit'' and ``asia\_bibi'') to make final predictions on the blind test set.
\subsection{Training Procedure}
The proposed neural network was trained with joint loss:
\begin{equation}
\mathcal{L} = \mathcal{L}_{SL} + \mathcal{L}_{clf}
\end{equation}
where $\mathcal{L}_{SL}$ is maximum log-likelihood loss \cite{ncrfpp} for the sequence labeling task and $\mathcal{L}_{clf}$ is Cross Entropy Loss for the language classification.
We use Adam with a learning rate of $1e-4$, $\beta_1=0.8$, $\beta_2=0.9$, $L2$ weight decay of $0.01$, learning rate warm up, and linear decay of the learning rate. Also, gradient clipping was applied for weights with $clip=1.0$.
Training of proposed neural network architecture was performed on one GPU with the batch size equal to $16$, the number of epochs equal to $150$, but stopped at epoch number $80$ because the loss function has ceased to decrease. The model required only around $3$ GB of memory instead of fine-tuning all BERT model, which would have required more than $8$ GB GPU memory. All training procedure lasted around five hours on one GPU with the evaluation of development set on each epoch.
The final model was trained on unit of training and development datasets.
\section{Results and Discussion}
\subsection{Evaluation Results}
As baseline for BSNLP Shared Task we use a simple CRF tagger and obtain exact word level f1-score $0.372$ on the development dataset.
Finally we use joint model for named entity recognition task and language classification task because the model without part of the classification gave a result by several percent less than proposed final model. This means that the joint model pays attention to a specific language morphology and some connections between words within one language.
\begin{table}[htp]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
label & precision & recall & f1-score \\\hline
PER & 0.733 & 0.725 & 0.729 \\\hline
PRO & 0.384 & 0.547 & 0.451 \\\hline
EVT & 0.385 & 0.370 & 0.377 \\\hline
LOC & 0.648 & 0.872 & 0.744 \\\hline
ORG & 0.550 & 0.630 & 0.587 \\\hline
avg/total & 0.540 & 0.629 & 0.578 \\\hline
\end{tabular}
\caption{\label{dev-metrics-table} Evaluation metrics on development dataset}
\end{table}
For proposed neural network architecture the evaluation of the training stage was produced on development dataset. Table~\ref{dev-metrics-table} shows span-level metrics precision, recall, and f1-measure. For development set, we obtained the following scores: language classification quality (f1-score): $0.998$ and Multilingual Named Entity Recognition quality (f1-score): $0.70$ for exact word level matching and $0.578$ for exact full entities matching. Also we train model without language classification, which resulted in f1-score equal to $0.66$ . This confirms the impact of language classification. Our model significantly outperforms the CRF baseline.
The evaluation of test dataset presented in Table~\ref{partial-matching-table} (relaxed partial matching) and Table~\ref{partial-matching-table} (relaxed exact matching) is measured by the BSNLP Shared Task organizers.
\begin{table}[htp]
\centering
\begin{tabular}{|c|c|c|c|}
\multicolumn{4}{c}{ Relaxed partial matching} \\ \hline \hline
label & precision & recall & f1-score \\\hline
PER & 0.84955 & 0.87119 & 0.86023 \\\hline
LOC & 0.77526 & 0.93197 & 0.84642 \\\hline
ORG & 0.62642 & 0.87170 & 0.72898 \\\hline
PRO & 0.42079 & 0.81416 & 0.55483 \\\hline
EVT & 0.24074 & 0.15476 & 0.18841 \\\hline
All & 0.90142 & 0.69917 & 0.78752 \\\hline
\multicolumn{4}{c}{ Relaxed exact matching} \\ \hline \hline
label & precision & recall & f1-score \\\hline
PER & 0.76835 & 0.74023 & 0.73317 \\\hline
LOC & 0.87747 & 0.73014 & 0.79705 \\\hline
ORG & 0.71390 & 0.52295 & 0.60369 \\\hline
PRO & 0.34439 & 0.18506 & 0.24075 \\\hline
EVT & 0.10714 & 0.16667 & 0.13043 \\\hline
All & 0.56225 & 0.46901 & 0.50102 \\\hline
\end{tabular}
\caption{\label{partial-matching-table} Evaluation metrics on test dataset}
\end{table}
\subsection{Error Analysis}
First of all, we face some errors with converting from origin data format (raw and annotations) to word markup and back to origin format after predictions were made. This problems stand for extra spaces, bad Unicode symbols and symbols, absent in WordPiece vocabulary.
Other errors are caused by neural network prediction failures. The model turns to be overfitted on the negative label ``O'' so that there are many false positives in the prediction. Lastly, the infrequent labels ``PRO'' and ``EVT'' are often confused.
\section{Related Work}
The related work has several parts: firstly, our work follows the recent trend of using pretrained neural languages models, such as \cite{bert, elmo, ulmfit}. The main difference between original BERT's approach for named entity recognition task \cite{bert} we use its only as input embeddings of sequence without fine-tuning. From ELMo paper \cite{elmo} we use weighting approach for different outputs from network and getting final representation of sequence. From ULMFiT work we took part which is related to the final decoding for classification (Pooling Classifier) without proposed language model \cite{ulmfit}. Secondly we model the task of NER as a joint sequence labeling and classification task following other joint architectures \cite{joint1, joint2}.
\section{Conclusion and Future Work}
We have proposed neural network architecture that solves Multilingual Named Entity Recognition without any additional labeled data for Bulgarian, Czech, Polish and Russian languages. This implementation allows to train the model even on a modern personal computer with GPU. This neural network architecture can be used for other tasks, that can be reformulated as a sequence labeling task for any other language.
As the next steps in the study of the underlying architecture, we can increase or decrease the number of units on each layer or remove the recurrent layer or multihead-attention layer. As improvements of the system, we can fine-tune BERT embeddings and put additional layers on top of BERT or pass other modern language models as an input.
\section*{Acknowledgments}
The article was prepared within the framework of the HSE University Basic Research Program and funded by the Russian Academic Excellence Project ``5-100''. We are thankful to the Muppets and to the BSNLP shared task organizers.
|
\section{INTRODUCTION}
We consider the transport of neutral particles as described by the linear Boltzmann equation,
\begin{equation}\label{RadiativeTransfer}
\frac{1}{c}\frac{\partial \psi(z,\mu,t) }{\partial t} + \mu \frac{\partial \psi(z,\mu,t)}{\partial z} + \sigma_\mathrm{t} \psi(z,\mu,t) = \frac{\sigma_\mathrm{s} }{2}
\int_{-1}^1\psi(z,\mu',t) \,d\mu'+S.
\end{equation}
The total and isotropic scattering macroscopic cross-sections are denoted as $\sigma_t(z)$ and $\sigma_s(z)$, respectively, $c$ is the particle speed and $S(z,t)$ is a prescribed source. The angular flux $\psi(z,\mu, t)$ is a function of position $z$, time $t$, and the cosine of the polar angle $\mu \in [-1,1]$. We also write the scalar flux, $\phi(z,t)$, as the integral of the angular flux
\begin{equation}
\phi(z,t) = \int_{-1}^1\psi(z,\mu,t) \,d\mu.
\end{equation}
In slab geometry we have a two-dimensional phase space and in principle there is a general function space that can describe solutions to the transport problem. Intuitively, it is known that many transport problems require only a subspace of this function space (called a manifold in mathematical parlance) to describe the transport. In other words, the solution is not any possible function of two variables, rather only a subset of functions. An example of this are problems in the diffusion limit: these problems require only a linear dependence on $\mu$. One can also formulate problems where this manifold over which the solution depends evolves over time: a beam entering a scattering medium would be described by a delta-function in space and angle at time zero, but eventually relax to much smoother distribution that we could characterize using a simple basis expansion.
We desire to generalize this idea, and possibly automatically discover the manifold that describes the system evolution. We accomplish this task by expressing the solution to a transport problem as a basis expansion in space and angle, and using techniques to determine what subspace of those bases are needed to describe the solution and how that subspace evolves. We use the dynamical low-rank approximation (DLRA) of Koch and Lubich to evolve time-dependent matrices by tangent-space projection \cite{Koch2007}. DLRA has been extended to tensors \cite{Koch2010} and further results can be found in \cite{Nonnenmacher2008}. DLRA has been used to reduce the computational complexity of quantum propagation \cite{Kloss2017} by restricting the evolution to lower-rank amongst other work \cite{Jahnke2008,Boiveau2017,Einkemmer2018,Markovsky2008}. In this work, we apply DLRA to neutral particle transport.
Here we give a brief mathematical introduction of a robust and accurate projector-splitting method developed by Lubich \cite{Lubich2014} to perform the DLRA for matrix differential equations of the form
\[
\frac{\partial}{\partial t} A(t) \equiv \dot{A}(t)= F(A(t)),
\]
for $A(t) \in \mathbb{R}^{m \times n}$.
DLRA seeks to find an approximating matrix $Y(t)$ of rank $r$ that minimizes the error in the Frobenius norm $||{Y}(t)-{A}(t)||_\mathrm{f}$.
Then we note that rank $r$ matrices are a manifold, $\mathcal{M}_r$, of the space $\mathbb{R}^{m \times n}$. The solution to this minimization problem can be found using the singular value decomposition (SVD). However, to use the SVD this way we would need to have the solution $A(t)$.
We would prefer a way to evolve the solution on $\mathcal{M}_r$ directly. We reformulate the problem as minimizing the difference between the time derivative of the approximation and the solution
$||\dot{Y}(t)-\dot{A}(t)||_\mathrm{f},$
where the derivative $\dot{Y}(t)$ is in the tangent space of $\mathcal{M}_r$. With a Galerkin condition the minimization problem is equivalent to an orthogonal projection. With the decomposition $Y(t) = U(t)S(t)V^T(t)$, this minimization problem can be solved using time splitting.
\section{NUMERICAL METHOD}
In this study we write the solution to Eq.~\eqref{RadiativeTransfer} as
\begin{equation}\label{LRA1}
\psi(z,\mu,t) \approx \sum_{i,j = 1}^r X_i(z,t)S_{ij}(t)W_j(\mu,t)
\end{equation}
as the best approximation with rank $r$ of the solution for the equation (\ref{RadiativeTransfer}), where we have written
$X_i$ as an orthonormal basis for $z$ and $W_j$ as an orthonormal basis for $\mu$.
We define the inner products \[\langle f,g\rangle _z = \int_{0}^{Z} f(z) g(z)\, dz, \qquad \langle f,g\rangle _\mu = \int_{-1}^1 f(\mu) g(\mu) \, d\mu.\] Due to orthonormality we also have $\langle X_i, {X}_j \rangle_z = \langle W_i, {W}_j \rangle_\mu = \delta_{ij}.$ Then $
\bar{X} = \{X_1, X_2, ..., X_r\}
$
and
$
\bar{W} = \{W_1, W_2, ..., W_r\}
$ are constructed as ansatz spaces.
The expansion in Eq.~\eqref{LRA1} is not unique and we choose as gauge conditions $\langle X_i, \dot{X}_j \rangle_z = 0$ and $\langle W_i, \dot{W}_j \rangle_\mu = 0$.
We now define orthogonal projectors using the bases:
\begin{equation}\label{Projection1}
P_{\bar{X}} g = \sum_{i=1}^{r}X_i \langle X_i g \rangle_{z}
\end{equation}
\begin{equation}\label{Projection2}
P_{\bar{W}} g = \sum_{j=1}^{r}W_j \langle W_j g \rangle_{\mu}
\end{equation}
We apply the projectors to define a split of the original equations into three steps and each of these is solved for a short time step
\begin{equation}\label{MainEq1}
\partial_t \psi_1(z,\mu,t) = P_{\bar{W}}\left(-\mu\partial_z \psi_1(z,\mu,t)+\frac{\sigma_\mathrm{s}}{2}\int_{-1}^{1}\psi_1(z,\mu^{'},t)\,d\mu^{'}-\sigma_t \psi_1(z,\mu,t)+\frac{1}{2}S \right),
\end{equation}
\begin{equation}\label{MainEq2}
\partial_t \psi_2(z,\mu,t) = -P_{\bar{X}}P_{\bar{W}}\left(-\mu\partial_z \psi_2(z,\mu,t)+\frac{\sigma_\mathrm{s}}{2}\int_{-1}^{1}\psi_2(z,\mu,t)d\mu^{'}-\sigma_\mathrm{t} \psi_2(z,\mu,t)+\frac{1}{2}S\right),
\end{equation}
\begin{equation}\label{MainEq3}
\partial_t \psi_3(z,\mu,t) = P_{\bar{X}}\left(-\mu\partial_z \psi_3(z,\mu,t)+\frac{\sigma_\mathrm{s}}{2}\int_{-1}^{1}\psi_3(z,\mu^{'},t)d\mu^{'}-\sigma_\mathrm{t} \psi_3(z,\mu,t)+\frac{1}{2}S\right).
\end{equation}
The $\psi_2$ step uses $\psi_1$ as an initial condition, and the $\psi_3$ uses $\psi_2$ as an initial condition.
It can be shown that the above evolution is contained in the low rank manifold $\mathcal{M}_r$, if the initial value is in $\mathcal{M}_r$ \cite{Lubich2014} because the right-hand side of each step remains in the tangent space $\mathcal{T} \mathcal{M}_r$.
To make the splitting more concrete we write $\psi_1$ as
\begin{equation}
\psi_1(z,\mu,t) = \sum_{j=1}^r K_j(z,t) W_j(\mu,t),
\end{equation}
where $K_j(z,t) = \sum^{r}_{i} X_{i}(z,t)S_{ij}(t)$. We plug this solution into Eq.~\eqref{MainEq1} and multiply by $W_\ell(\mu,t)$ and integrate over $\mu$ to get
\begin{equation}\label{eq:psi1}
\partial_t K_j + K_j \cancelto{0}{\langle W_\ell \dot{W}_j \rangle_\mu} = - \sum_{j'=1}^r \langle \mu W_j W_{j'} \rangle_\mu \partial_z K_{j'} + \frac{\sigma_\mathrm{s}}{2} \sum_{j'=1}g^r \langle W_j \rangle_\mu \langle W_{j'} \rangle_\mu K_{j'} - \sigma_\mathrm{t} K_j + \frac{\langle W_{j}\rangle_\mu}{2}S.
\end{equation}
Notice that there is no change in $W_j$ bases in this equation. Equation \eqref{eq:psi1} resembles the standard P$_N$ equations, a point we will return to later. It is a system of advection problems in $z$.
We can then factorize $K_j$ into $X_{i}^{(1)}$ and $S_{ij}^{(1)}$. This is used to define an initial condition for $\psi_2 = \sum_{i,j=1}^r X_{i}^{(1)} S_{ij}^{(1)} W_j.$ Then, we can perform similar calculations on Eq.~\eqref{MainEq2} to get
\begin{multline}\label{eq:s_eq}
\partial_t S_{ij} = \sum_{kl}^{r}\langle \partial_z X_k X_i\rangle_z S_{kl} \langle \mu W_l W_j\rangle_{\mu} - \frac{1}{2}\sum_{kl}^{r}\langle \sigma_s X_k X_i\rangle_z S_{kl} \langle W_l \rangle_{\mu} \langle W_j \rangle_{\mu} + \\ \sigma_t S_{ij} - \frac{1}{2}\langle X_i \rangle_{z}\langle W_j \rangle_{\mu}.
\end{multline}
We call this solution $S_{ij}^{(2)}$. Equation \eqref{eq:s_eq} is a set of $r^2$ ordinary differential equations. The solution is used to create an initial condition for $\psi_3 = \sum_{i,j=1}^r X_{i}^{(1)} S_{ij}^{(2)} W_j.$
Writing $L_i = S_{ij}(t)W_{j}(t,\mu)$ we can multiply Eq.~\eqref{MainEq3} by a spatial basis function and integrate over space to get
\begin{equation}\label{eq:L_eq}
\partial_t L_{i} = -\mu \sum_{k}^{r} \langle \partial_z X_k X_i \rangle_{z} L_k + \frac{1}{2} \langle \sigma_s X_i\rangle_z \langle L_i \rangle_{\mu} - \langle \sigma_t X_i\rangle_z L_i + \frac{1}{2}\langle X_i S \rangle_{z},
\end{equation}
which evolves the solution in $\mu$ space. Upon factoring $L_i = S^{(3)}_{ij}(t)W_j^{(3)}(\mu,t)$, and write the solution as $\psi = \sum_{i,j,=1}^r X_i^{(1)}(x,t) S^{(3)}_{ij}(t)W_j^{(3)}(\mu,t).$
\subsection{Discretization Details}
The procedure outlined above of solving Eqs.~\eqref{eq:psi1}, \eqref{eq:s_eq}, and \eqref{eq:L_eq} in that order is accomplished by using a first-order explicit integration. The bases we use are based on a finite volume discretization in space with a constant mesh spacing $\Delta z$ and $m$ zones, and $n$ Legendre polynomials in angle. To make orthonormal bases we define
\begin{equation}\label{Discretization1}
X_i(t,z) = \sum_{k=1}^{m}Z_k(z)u_{ki}(t)
\end{equation}
\begin{equation}\label{Discretization2}
W_j(t,\mu) = \sum_{l=1}^{n}P_l(\mu) v_{lj}(t)
\end{equation}
Noted that $Z_i(z) = \frac{1}{\sqrt{\Delta z}}$ with $z \in [z_{i-\frac{1}{2}},z_{i+\frac{1}{2}}]$ where $i$ is the cell number, $P_j(\mu) = \sqrt{\frac{2n-1}{2}}\bar{P}_{n-1}(\mu)$, where $\bar{P}_{n}(\mu)$ is the $n_{th}$ order Legendre polynomial, $u_{ki}$ and $v_{lj}$ are components of the time dependent matrix $U(t) \in \mathbb{R}^{m \times r}$ and $V(t) \in \mathbb{R}^{n \times r}$.
After the first and last step in the split the matrices $U$ and $V$ found by a QR decomposition to either $K_j$ or $L_i$.
The memory footprint required to compute the solution is the based on storing the matrices $U$, $V$ and $S$. Therefore, the memory required is \begin{equation}\label{Memory}
\mathrm{memory} = 2(mr+r^2+nr),
\end{equation} the factor 2 assumes that we need to store the previous step solution as well as the new step. The full solution to this problem without splitting would require a memory footprint of $2mn$. Therefore, for $r \ll m,n$ there will be large memory savings.
In the solution procedure we needed to calculate $ \langle \mu W_j W_{j'} \rangle_\mu$. Using our angular basis this term becomes.
\begin{equation}\label{Term1}
\begin{aligned}
\langle \mu W_j W_{j'} \rangle_\mu= & \left \langle \mu \sum_{i=1}^{n}P_i(\mu) v_{ij}(t) \sum_{k=1}^{n}P_k(\mu) v_{kj'}(t) \right\rangle_{\mu} = \left\langle \sum_{i=1}^{n}\sum_{k=1}^{n}v_{ij}(t)\mu P_i(\mu) P_k(\mu) v_{kj'}(t) \right\rangle_{\mu} \\
= & \sum_{i=1}^{n}\sum_{k=1}^{n}v_{il}(t) \langle \mu P_i(\mu) P_k(\mu) \rangle_{\mu} v_{kj}(t)
\end{aligned}
\end{equation}
Note that $ \langle \mu P_i(\mu) P_k(\mu) \rangle_{\mu}$ forms a $n \times n$ matrix, $C$, that can be precomputed. Thus Eqs.~\eqref{Term1} requires $\mathcal{O}(n^2 r)$ operations, which is affordable because usually $n$ is not large and $C$ is sparse. Alternatively, we could calculate $ \langle \mu W_j W_{j'} \rangle_\mu$ on-the-fly by choosing $\mathcal{O}(n)$ quadrature points in angle, and it requires $\mathcal{O}(n r^2)$ operations for all the $r^2$ entries.
Additionally, using the standard upwinding technique for the spatial derivative terms leads to
\begin{align}\label{Term2}
\partial_z K \langle \mu W^T W \rangle_{\mu} =& \frac{1}{\sqrt{\Delta z}}(K_{i+\frac{1}{2}} - K_{i-\frac{1}{2}}) \langle \mu W^T W \rangle_{\mu} \\
&=\frac{1}{2\sqrt{\Delta z}}[(K_{i+1}-K_{i-1})V^T CV - (K_{i+1}-2K_{i}+K_{i-1})V^T \Sigma V],
\end{align}
where $\Sigma$ is a stabilization matrix that we take to be a diagonal matrix with the singular values of $C$. Other stabilization terms could be used, including Lax-Friedrichs where $V^T \Sigma V$ is replaced by a constant times an identity matrix.
The spherical harmonic we used in the angular expansion can yield oscillatory or negative solutions. To address this issue we implemented angular filtering \cite{McClarren2010,radice2013new} which can significantly increase the performance of $P_n$ method in solving radiative transfer equation by removing the oscillations. We implemented the filter into our explicit solver and combined it with the low-rank approximation algorithm. The filtered equation adds anisotropic scattering. In this study we use a Lanczos filter.
\subsection{Conservation}
The low-rank algorithm we have described does not conserve the number of particles. This loss of conservation is a result of information lost in the algorithm when restricting the solution to low rank descriptions. We have addressed this by globally scaling the solution after each time step to correct for any particles lost. This point is discussed in further detail in the conclusion section.
\section{NUMERICAL RESULTS}
\subsection{Plane source problem}
\begin{figure}[h!]
\begin{subfigure}{1\textwidth}\caption{Low-rank solutions without a filter}
\centering
\includegraphics[scale=0.25]{./Figures/PP_generalresults_T1.eps}
\end{subfigure}
\begin{subfigure}{1\textwidth}\caption{Comparison of P$_7$ solutions of rank $4$ with and without a filter.}
\centering
\includegraphics[scale=0.25]{./Figures/PP_Filter_T1.eps}
\end{subfigure}
\caption{Solutions to the plane source problem using the low rank method compared to the analytic solution.}
\label{PPresults}
\end{figure}
First, we solve the plane source problem of a delta function source in space and time in a purely scattering medium with $\sigma_\mathrm{t} = \sigma_\mathrm{s} = 1$; the analytical benchmark solution was given by Ganapol \cite{Ganapol2008}. For this problem we fix the spatial resolution to be $\Delta z = 0.01$ (this corresponds to $m = 301$ for the $t=1$ solution and $m = 1201$ for the $t=5$ results), and vary the number of angular basis functions, $n$, and the rank $r$. When used, the filter strength is set to $50$.
Figure \ref{PPresults} shows the solutions of varying rank and Legendre polynomial orders with and without a filter. We can see the low-rank solution using a P$_{15}$ basis matches the analytic solution to the scale of the graph in the middle of the problem. We also observe that the low-rank solution can be improved by the filter: P$_7$ solutions of reduced rank improve when a filter is used.
For a more quantitative comparison, the error of the numerical results with different $n$ and $r$ is shown in Figure \ref{PlanePulseErrors}. In this figure the colors for the dotted lines correspond to the rank used in a calculation and different values of the $n$, the number of angular basis functions, are corresponding dots. For each color the value of $n$ ranges from $r$ to $100$. The large points are the value of the error using the standard full rank method with $r=n$. We can observe that the low-rank solution is more accurate than the full rank with the same memory usage. For example, the error of full rank solution $n=12$ using with a memory footprint of $8000$ is about $0.07$. With less memory, the error can be reduced to $0.02$. We can also use $70\%$ of the memory to achieve the same accuracy. Increasing the resolution and rank will contribute to the accuracy of solutions. Given the way we performed this study with a fixed spatial mesh and time step and the conservation fix we used, we can see some error stagnation in the low-rank solution at $t=5$. Other numerical experiments indicate that increasing the number of spatial zones can further decrease the error.
\begin{figure}[h!]
\begin{subfigure}{1\textwidth}
\centering
\includegraphics[scale=0.25]{./Figures/PP_Errors_T1.eps}
\end{subfigure}%
\begin{subfigure}{1\textwidth}
\centering
\includegraphics[scale=0.25]{./Figures/PP_Errors_T5.eps}
\end{subfigure}
\caption{The comparison of errors on the plane source problem with different memory usage are shown. Each dotted line represents the error with a fixed rank that varies the number of angular basis functions $n$. The bold dot denotes the full rank solution.}
\label{PlanePulseErrors}
\end{figure}
\subsection{Reed's problem}
\begin{figure}
\centering
\includegraphics[scale=0.5]{./Figures/reeds.eps}
\caption{The material layout in Reed’s problem.}
\label{Reeds}
\end{figure}
The second problem is Reed's problem \cite{Reed,Mcclarren2007,Zheng:2016ei}, which is a multi-material problem, and its set-up is detailed in Fig.~\ref{Reeds}.
Because Reed's problem does not have an analytical solution, a numerical result with high resolution and full rank, where $\Delta z = 0.01$ ($m=1600$), P$_{99}$ ($n= 100$) and $CFL = \Delta t / \Delta x = 0.1$, is set as a benchmark for memory analysis. It can be observed in Figure \ref{Reedserror} that the low rank solutions (solid lines with small dots) can give solutions with comparable errors to the full rank solutions (large dots) with much larger memory. For example the rank $8$ solutions obtain a solution error better than the full rank $P_{19}$ solution with less memory.
\begin{figure}
\begin{subfigure}{1\textwidth}
\centering
\includegraphics[scale=0.25]{./Figures/Reed_Errors_T1.eps}
\end{subfigure}%
\begin{subfigure}{1\textwidth}
\centering
\includegraphics[scale=0.25]{./Figures/Reed_Errors_T5.eps}
\end{subfigure}
\caption{The comparison of errors for Reed's problem with different memory usage are shown. Each dotted line represents the error with a fixed rank that varies the number of angular basis functions $n$. The bold dot denotes the full rank solution.}
\label{Reedserror}
\end{figure}
\section{CONCLUSIONS}
We have developed a practical algorithm to find the low-rank solution of the slab geometry transport equation using explicit time integration. The method is based on projecting the equation to low-rank manifolds and numerically integrating in three steps. The numerical simulations show that on several test problems the memory savings of the low-rank method can be on the order of a factor of 2-3. Given that these are only slab geometry problems we expect even larger memory savings on 2- and 3-D problems due to their larger size. Exploring this is ongoing work. Furthermore, we will be investigating other means for correcting the loss of conservation in the method, including posing the problem as a high-order/low-order problem.
\setlength{\baselineskip}{12pt}
\bibliographystyle{mandc}
|
\section{Introduction}
Ordinary Cherenkov radiation in an optical medium
is emitted when a charged, massive particle
travels faster than the phase velocity of light
in that particular medium.
If the latter is the case,
the polarized atoms or molecules
in the vicinity of the particle trajectory
emit their wave trains in phase
whereupon these interfere constructively.
As a result,
coherent radiation is produced
that can be detected far away from the source.
The modified light cone or mass-shell structure
in Lorentz-violating theories
may be responsible for an energy loss of a charged, massive particle in vacuum.
As such a process shares certain characteristics
with Cherenkov radiation in macroscopic media,
it is known in the community as vacuum Cherenkov radiation.
In principle,
any vacuum endowed with a Lorentz-violating background field
in the context of the Standard-Model Extension\cite{Colladay:1998fq} (SME)
can be interpreted as a vacuum with a nontrivial refractive index.
This property explains the analogy.
Two very different scenarios exist for modified fermions\cite{Kostelecky:2013rta}
that may render vacuum Cherenkov radiation possible.
For spin-degenerate Lorentz violation,
photon emission in vacuum is possible
when the slope of the mass shell at a certain energy is larger than one.
Multiple photons can be emitted subsequently
as long as the previously mentioned condition is satisfied.
The condition fails when the particle energy
drops under a certain threshold,
whereupon the process ceases.
As long as the energy is sufficiently above the threshold,
the momentum of the emitted photon is large enough
such that the resulting recoil reverses the momentum direction
of the emitting particle.
A reversal of the momentum is directly connected to a helicity flip,
which is why photons of helicity $h=\pm 1$ can be emitted.
These photons are circularly polarized.
However,
when the energy of the radiating fermion is only slightly above the threshold,
the photons emitted are too soft to enable a helicity change.
As a consequence,
they have helicity $h=0$,
i.e.,
they are linearly polarized.
The alternative scenario of modified fermions
that allows for vacuum Cherenkov radiation
is that of spin-nondegenerate Lorentz violation.
In this case,
photon emission is possible for a certain helicity of the fermion,
i.e.,
the fermion loses energy and switches its mass shell.
The latter corresponds to a flip of spin or helicity
without a change of the momentum direction.
Therefore,
these particular processes are sometimes called helicity decays
and photons emitted in such processes are always circularly polarized.
Also,
in contrast to the previous class of processes,
helicity decays can occur for arbitrarily low energies of the radiating fermion.
In what follows,
the decay rates\cite{Schreck:2017isa} for both scenarios
will be presented and discussed.
These phenomenological results rest on a technical work\cite{Reis:2016hzu}
on the modified Dirac theory based on the nonminimal SME.
\section{Spin-degenerate Lorentz violation}
Let us first consider a fermion
modified by spin-degenerate, isotropic Lorentz violation.
The decay rates of such processes
are presented in Fig.~\ref{fig:decay-rates-vacuum-cherenkov-spin-conserving}
as functions of the incoming fermion momentum.
Minuscule values were inserted for the controlling coefficients,
as those are the most interesting ones from a phenomenological perspective.
Two curves are contained in the plot.
The first shows the behavior of the decay rates
$\Gamma_{\ring{c},\ring{d}}$
for the isotropic $c$ and $d$ coefficients,
whereas the second illustrates $\Gamma_{\ring{e},\ring{f},\ring{g}}$
for the isotropic $e$, $f$, and $g$ coefficients.
The thresholds for both curves are clearly visible
and correspond to the momenta where the rates go to zero.
For large momenta,
the decay rates grow linearly with the momentum.
In principle,
the second curve is a scaled version of the first.
The values at the axes reveal
that the $\Gamma_{\ring{e},\ring{f},\ring{g}}$ are suppressed
by one additional power of Lorentz violation
in comparison to $\Gamma_{\ring{c},\ring{d}}$.
The thresholds are correspondingly larger.
Thus,
a process of this kind can occur
for the dimensionless spin-nondegenerate coefficients $d$, $g$
whereas it is forbidden for the dimensionful coefficients $b$, $H$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.30]{decay-rate-1-bw.pdf}
\caption{Decay rates $\Gamma$ for the isotropic dimensionless coefficients (indicated by rings on top of the symbols) in the SME fermion sector as a function of the fermion momentum $q$ where $m_{\psi}$ is the fermion mass and $\alpha=e^2/(4\pi)$ the fine-structure constant. The plot is double-logarithmic, whereby both axes cover a large range of values.}
\label{fig:decay-rates-vacuum-cherenkov-spin-conserving}
\end{figure
A reasonable partial crosscheck of the results can be carried out
by taking into account the coordinate transformation
that maps CPT-even, spin-degenerate Lorentz violation in the fermion sector
to CPT-even, nonbirefringent coefficients in the photon sector.\cite{Kostelecky:2009zp}
At leading order in the coefficients,
the map\cite{Altschul:2006zz} for the isotropic fermion coefficient $\ring{c}$
and the isotropic photon sector coefficient $\widetilde{\kappa}_{\mathrm{tr}}$
reads $\widetilde{\kappa}_{\mathrm{tr}}=-(4/3)\ring{c}+\dots$.
Hence,
for small enough Lorentz violation,
the decay rate\cite{Klinkhamer:2008ky} for vacuum Cherenkov radiation
in the isotropic photon sector for $\widetilde{\kappa}_{\mathrm{tr}}>0$
should correspond to the result in the isotropic fermion sector
with $\ring{c}<0$.
It was checked that the curve
that relies on the isotropic photon sector
perfectly matches $\Gamma_{\ring{c}}$.
This finding strengthens our confidence in the correctness of the results,
as both computations were carried out completely independently
from each other.
Based on the fact that ultrahigh-energy cosmic rays reach Earth,
constraints\cite{Schreck:2017isa} on Lorentz violation of quarks
embedded in a modified quantum electrodynamics were also obtained.
\section{Spin-nondegenerate Lorentz violation}
\begin{figure}[t]
\centering
\includegraphics[scale=0.30]{decay-rate-2-bw.pdf}
\caption{Double-logarithmic plot of the decay rate for vacuum Cherenkov radiation for the isotropic $b$ coefficient where $\ring{b}/m_{\psi}=10^{-10}$. The quantities used are the same as those in Fig.~\ref{fig:decay-rates-vacuum-cherenkov-spin-conserving}.}
\label{fig:helicity-decay-b-isotropic}
\end{figure
Finally,
we review the behavior of a fermion modified
by spin-nondegenerate Lorentz violation.
In particular,
the isotropic $b$ coefficient will be discussed.
The decay rate of the corresponding helicity decay
is presented in Fig.~\ref{fig:helicity-decay-b-isotropic}.
It is evident that the process does not have a threshold.
The decay rate grows polynomially
as a function of the fermion momentum
until it reaches a maximum
beyond which it decreases again logarithmically.
Furthermore,
the rate is strongly suppressed by Lorentz violation.
For high energies,
the concept of helicity becomes more and more equivalent to chirality.
Thus,
the probability of a spin flip is reduced for large energies,
which explains the decrease of the decay rate
in the high-momentum regime.
If data were available on the polarization of ultrahigh-energy cosmic rays
arriving on Earth,
fermion coefficients could,
in fact,
be constrained based on such processes.
\section*{Acknowledgments}
It is a pleasure to thank the Brazilian agencies CNPq and FAPEMA
for financial support
under grant numbers CNPq Universal 421566/2016-7,
CNPq Produtividade 312201/2018-4,
and FAPEMA Universal 01149/17.
The author greatly acknowledges the hospitality
of the Indiana University Center for Spacetime Symmetries (IUCSS).
|
\section{Introduction and Preliminaries}
Let $G=(V,E)$ be an undirected graph. A (proper vertex) coloring of $G$ is
a surjective function $\sigma: V\to S$ such that $xy\in E$ implies
$\sigma(x)\ne \sigma(y)$. The minimum number $|S|$ of colors such that
there is a coloring of $G$ is known as the \emph{chromatic number}
$\chi(G)$. A \emph{greedy coloring} of $G$ is obtained by ordering the set
of colors and coloring the vertices of $G$ in a random order with the first
available color. The \emph{Grundy number} $\gamma(G)$ is the maximum number
of colors required in a greedy coloring of $G$
\cite{Christen:79}. Obviously $\gamma(G)\ge\chi(G)$. Determining $\chi(G)$
\cite{Karp:72} and $\gamma(G)$ \cite{Zaker:06} are NP-complete problems. A
graph $G$ is called \emph{well-colored} if $\chi(G)=\gamma(G)$
\cite{Zaker:06}. It is \emph{hereditarily well-colored} if every induced
subgraph is well-colored.
\begin{definition}[\cite{Corneil:81}]
\label{def:Corneil}
A graph $G$ is a \emph{cograph} if $G=K_1$, $G$ is the disjoint union
$G=\bigcupdot_i G_i$ of cographs $G_i$, or $G$ is a join
$G=\bigop{\triangledown}_i G_i$ of cographs $G_i$.
\end{definition}
This recursive construction induces a rooted tree $T$, whose leaves are
individual vertices corresponding to a $K_1$ and whose interior vertices
correspond to the union and join operations. We write $L(T)$ for the leaf
set and $V^0(T)$ for the set of inner vertices of $T$. The set of children
of $u$ is denoted by $\mathsf{child}(u)$. For edges $e=uv$ in $T$ we adopt the
convention that $v$ is a child of $u$. We define a labeling function
$t:V^0(T)\rightarrow\{0,1\}$, where an interior vertex $u$ of $T$ is
labeled $t(u)=0$ if it is associated with a disjoint union, and $t(u)=1$
for joins. The set $L(T(u))$ denotes the leaves of $T$ that are descendants
of $u$. To simplify the notation we will write $G(u):=G[L(T(u))]$ for the
subgraph of $G$ induced by the vertices in $L(T(u))$. Note that $G(u)$ is
the graph consisting of the single vertex $u$ if $u$ is a leaf of $T$.
Given a cograph $G$, there is a unique \emph{discriminating}
cotree\footnote{In \cite{Corneil:81} the discriminating cotree is defined
as \emph{the} cotree associated with $G$. Here we call every tree arising
from Def.~\ref{def:Corneil} \emph{a} cotree of $G$.} in which adjacent
operations are distinct, i.e., $t(u)\neq t(v)$ for all interior edges
$uv\in E(T)$. It is possible to refine the discriminating cotree by
subdiving a disjoint union or join into multiple disjoint unions or joins,
respectively \cite{Boecker:98,Corneil:81}. It is well known that every
induced subgraph of a cograph is again a cograph. A graph is a cograph if
and only if it does not contain a path $P_4$ on four vertices as an induced
subgraph \cite{Corneil:81}. The cographs are also exactly the hereditarily
well-colored graphs \cite{Christen:79}. The chromatic number of a cograph
$G$ can be computed recursively, as observed in
\cite[Tab.1]{Corneil:81}. Starting from $\chi(K_1)=1$ as base case we have
\begin{equation}
\begin{split}
\chi(G) &= \chi\left(\bigcupdot_{i} G_i\right) = \max_{i} \chi(G_i)
\textrm{ or }\\
\chi(G) &= \chi\left(\bigop{\triangledown}_{i} G_i\right) = \sum_{i} \chi(G_i)
\end{split}
\label{eq:rec1}
\end{equation}
\emph{Hierarchically colored cograph} (\emph{hc-cographs}) were introduced
in \cite{Geiss:19x} as the undirected colored graphs recursively defined by
\begin{description}
\item[\AX{(K1)}] $(G,\sigma)=(K_1,\sigma)$, i.e., a colored vertex, or
\item[\AX{(K2)}] $(G,\sigma)= (H_1,\sigma_{H_1}) \join (H_2,\sigma_{H_2})$
and $\sigma(V(H_1))\cap \sigma(V(H_2))=\emptyset$, or
\item[\AX{(K3)}]
$(G,\sigma)= (H_1,\sigma_{H_1}) \charfusion[\mathbin]{\cup}{\cdot} (H_2,\sigma_{H_2})$ and
$\sigma(V(H_1))\cap \sigma(V(H_2)) \in
\{\sigma(V(H_1)),\sigma(V(H_2))\}$,
\end{description}
where $\sigma(x) = \sigma_{H_i}(x)$ for every $x\in V(H_i)$, $i\in\{1,2\}$
and $(H_1,\sigma_{H_1})$ and $(H_2,\sigma_{H_2})$ are hc-cographs.
This recursive construction of an hc-cograph $G$ implies a binary \emph{cotree} $(T,t)$. Its inner vertices can be associated with the
intermediate graphs in the construction. We say that $\sigma$ is an
\emph{hc-coloring w.r.t.\ $(T,t)$.}
Obviously, the graph $G$ underlying an hc-cograph is a cograph.
\begin{definition}\label{def:hc-coloring}
Let $G=(V,E)$ be a cograph. A coloring $\sigma:V\to S$ of $G$ is an
\emph{hc-coloring} of $G$ if there is binary cotree $(T,t)$ of $G$ such
that $(G,\sigma)$ is hc-cograph w.r.t.\ $(T,t)$.
\end{definition}
This contribution aims to investigate the properties of hc-colorings and
their relationships with other types of cograph colorings.
\section{Existence of hc-Colorings}
As noticed in \cite{Geiss:19x}, a coloring $\sigma$ of a cograph $G$ may be
an hc-coloring w.r.t.\ some cotree $(T,t)$ but not w.r.t.\ to another
cotree $(T',t')$ that yields the same cograph. An example is shown in
Fig. \ref{Fig:diffcotree}.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.55]{cotree-coloring.pdf}
\end{center}
\caption{The induced cotree of a cograph $G$ affects the hc-coloring
property of $\sigma$, where
$\sigma(a)=\sigma(d)\neq\sigma(b)=\sigma(c)$. In $(T,t)$, the first
tree from left to right, \AX{(K1)}-\AX{(K3)} are satisfied making
$\sigma$ an hc-coloring. However the second tree $(T',t')$ does not
satisfy \AX{(K3)} since the parent node of $c\simeq K_1$ and
$d\simeq K_1$ corresponds to a disjoint union operation and
$\sigma(c)\cap\sigma(d)=\emptyset$. Thus $\sigma$ is not an hc-coloring
w.r.t. $(T',t')$.}
\label{Fig:diffcotree}
\end{figure}
\begin{theorem}
Let $\sigma:V\to S$ be an hc-coloring of a cograph $G$. Then
$|S|=\chi(G)$.
\label{thm:hc-implies-chrom}
\end{theorem}
\begin{proof}
We proceed by induction w.r.t.\ $|V|$. The statement is trivially true
for $|V|=1$, i.e. $G=K_1$, since $\chi(K_1)=1$. Now suppose $|V|>1$. Thus
$G=G_1\join G_2$ or $G=G_1 \charfusion[\mathbin]{\cup}{\cdot} G_2$ for some graphs $G_1=(V_1,E_1)$
and $G_1=(V_2,E_2)$ with $1\leq|V_1|,|V_2|<|V|$. By induction
hypothesis we have $|\sigma(V_1)|=\chi(G_1)$ and
$|\sigma(V_2)|=\chi(G_2)$.
First consider $G=G_1\join G_2$. Since $xy\in E(G)$ for all
$x\in V_1$ and $y\in V_2$ we have $\sigma(x)\ne\sigma(y)$, and hence
$\sigma(V_1)\cap\sigma(V_2)=\emptyset$. Thus,
$\sigma(V)=\sigma(V_1)\charfusion[\mathbin]{\cup}{\cdot}\sigma(V_2)$ and therefore,
\begin{equation*}
|\sigma(V)| = |\sigma(V_1)| + |\sigma(V_2)| \stackrel{\text{IH.}}{=}
\chi(G_1)+\chi(G_2) \stackrel{\text{Equ.\ }\eqref{eq:rec1}}{=} \chi(G).
\end{equation*}
We note that the coloring condition in \AX{(K2)} therefore only enforces
that $\sigma$ is a proper vertex coloring.
Now suppose $G=G_1\charfusion[\mathbin]{\cup}{\cdot} G_2$. Axiom \AX{(K3)} implies
$|\sigma(V)|=\max( |\sigma(V_1)|,|\sigma(V_2)|)$. Hence,
\begin{equation*}
|\sigma(V)| = \max( |\sigma(V_1)|,|\sigma(V_2)|)
\stackrel{\text{IH.}}{=}
\max(\chi(G_1),\chi(G_2)) \stackrel{\text{Equ.\ }\eqref{eq:rec1}}{=}
\chi(G).
\end{equation*}
\end{proof}
As detailed in \cite{Christen:79}, we have $\chi(G)=\gamma(G)$. Thus, it
seems natural to ask whether every greedy coloring is an
hc-coloring. Making use of the fact that $\chi(G)=\gamma(G)$, we assume
w.l.o.g.\ that the color set is $S=\{1,2,\dots,\chi(G)\}$ whenever we
consider greedy colorings of a cograph. By definition of greedy colorings
and the fact that cographs are hereditarily well-colored, we immediately
observe
\begin{lemma}
Let $\sigma$ be a greedy coloring of a cograph $G$ and
$G_1=(V_1,E_1)$ a connected component of $G$. Then $G_1$ is colored by
$\sigma(V_1)=\{1,\dots,\chi(G_1)\}$.
\end{lemma}
We shall say that a cograph $G$ is a \emph{minimal counterexample for some
property $\mathcal{P}$} if (1) $G$ does not satisfy $\mathcal{P}$ and (2)
every induced subgraph of $G$ (i.e., every ``smaller'' cograph) satisfies
$\mathcal{P}$.
\begin{lemma}
\label{lem:grehc}
Let $G$ be a cograph, $(T,t)$ an arbitrary binary cotree for $G$
and $\sigma$ a greedy coloring of $G$. Then $\sigma$ is an
hc-coloring w.r.t.\ $(T,t)$.
\end{lemma}
\begin{proof}
Assume $G$ is a minimal counterexample, i.e., $G$ is a minimal cograph
for which a coloring $\sigma$ exists that is a greedy coloring but not an
hc-coloring. If $G$ is connected, then $G=\bigop{\triangledown}_{i=1}^n G_i$, for some
$n>1$ and $\sigma(V)=\bigcupdot_{i=1}^n \sigma(V_i)$, i.e., \AX{(K2)} is
satisfied. By assumption, $\sigma$ is not an hc-coloring, hence $\sigma$
must fail to be an hc-coloring on at least one of the connected
components $G_i$, contradicting the assumption that $G$ is a minimal
counterexample. Thus, $G$ cannot be connected.
Therefore, assume $G=\bigcupdot_{i=1}^n G_i$ for some $n>1$. Since $G$ is
represented by a binary cotree $(T,t)$, the root of $T$ must have exactly
two children $u$ and $v$. Hence, we can write $G=G(u)\charfusion[\mathbin]{\cup}{\cdot} G(v)$.
Since $G$ is a minimal counterexample, we can conclude that $\sigma$
induces an hc-coloring on $G(u)$ and $G(v)$. However, since $\sigma$ is,
in particular, a greedy coloring of $G(u)$ and $G(v)$,
$\sigma(V(G(u)))\subseteq \sigma(V(G(v)))$ or
$\sigma(V(G(v)))\subseteq \sigma(V(G(u)))$ most hold. But this
immediately implies that $(G,\sigma)$ satisfies $\AX{(K3)}$ and thus
$\sigma$ is an hc-coloring of $G$. Therefore, $G$ is not a minimal
counterexample, which completes the proof.
\end{proof}
As an immediate consequence we find
\begin{corollary}
Every cograph has an hc-coloring.
\end{corollary}
The converse of Lemma \ref{lem:grehc} is not true. Fig.\
\ref{Fig:hcnogrundy} shows an example of an hc-coloring that is not a greedy
coloring.
\begin{figure}
\begin{tabular}{ccc}
\begin{minipage}{0.3\textwidth}
\includegraphics[width=\textwidth]{hc-no-grundy.pdf}
\end{minipage} &&
\begin{minipage}{0.6\textwidth}
\caption{Both colorings \textsf{A} and \textsf{B} of
$K_2\charfusion[\mathbin]{\cup}{\cdot} K_1\charfusion[\mathbin]{\cup}{\cdot} K_1$ are hc-colorings. However, only
\textsf{A} is a greedy coloring since the two single-vertex
components have different colors in \textsf{B}.}
\end{minipage}
\end{tabular}
\label{Fig:hcnogrundy}
\end{figure}
\begin{theorem}
A coloring $\sigma$ of a cograph $G$ is greedy coloring if and only if it
is an hc-coloring w.r.t.\ every binary cotree $(T,t)$ of $G$.
\end{theorem}
\begin{proof}
By Lemma~\ref{lem:grehc}, every greedy coloring is an hc-coloring for
every binary cotree $(T,t)$. Now suppose $\sigma$ is an hc-coloring for
every binary cotree $(T,t)$ and let $G$ be a minimal cograph for which
$\sigma$ is not a greedy coloring. As in the proof of Lemma
\ref{lem:grehc} we can argue that $G$ cannot be a minimal counterexample
if $G$ is connected: in this case, $G=\bigop{\triangledown}_{i=1}^n G_i$ and
$\sigma(V)=\bigcupdot_{i=1}^n \sigma(V_i)$ for all colorings, and thus
$\sigma$ is a greedy coloring if and only if it is a greedy coloring with
disjoint color sets for each $G_i$. Hence, a minimal counterexample must
have at least two connected components.
Let $G=\bigcupdot_{i=1}^n G_i$ for some $n>1$ and define a partition of
$\{1,\dots,n\}$ into sets $I_1,\dots,I_m$, $m\geq 1$, such that for every
$r\in \{1,\dots,m\}$, $i,j\in I_r$ if and only if
$\chi(G_i) = \chi(G_j)$. Since every
$G^r \coloneqq \bigcupdot_{i\in I_r} G_i$, $1\leq r\leq m$, is a cograph,
each $G^r$ can be represented by a (not necessarily unique) binary cotree
$(T^r,t^r)$. Note, we have $\chi(G^r)=\chi(G_i)$ for all $i\in
I_r$. Now, we can construct a binary cotree $(T,t)$ for $G$ as follows:
let $T^*$ be a caterpillar with leaf set $L(T^*)=\{l_1,\dots,l_\ell\}$.
We choose $T^* = (\ldots((l_1,l_2),l_3), \ldots l_\ell)$ (in Newick
format). Note, if $\ell=1$, then $T^*\simeq K_1$. Now, the root of every
tree $T^1,\dots,T^r$ is identified with a unique leaf in $L(T^*)$ such
that the root of $T_i$ is identified with $l_i\in L(T^*)$ and the
root of $T_j$ is identified with $l_j\in L(T^*)$, where $i<j$ if and
only if $\chi(G^i)<\chi(G^j)$. This yields the tree $T$. The labeling
$t$ for $(T,t)$ is provided by keeping the labels of each $(T^r,t^r)$ and
by labeling all other inner vertices of $T$ by $0$. It is easy to see
that $(T,t)$ is a binary cotree for $G$. By hypothesis, this in
particular implies that $\sigma$ is an hc-coloring w.r.t.\ $(T,t)$. We
denote by $C(T^*)\subseteq V(T)$ the set of inner vertices of
$T^*$. Since $\sigma$ is an hc-coloring w.r.t.\ $(T,t)$ and thus in
particular w.r.t.\ any subtree $(T^r,t^r)$, we have
$\sigma(V(G_i))\cap \sigma(V(G_j))\in \{\sigma(V(G_i)),\sigma(V(G_j))\}$
for any $i,j\in I_r$, $1\le r\le m$, by \AX{(K3)}. Hence, as
$\chi(G^r)=\chi(G_i)=\chi(G_j)$, it must necessarily hold
$\sigma(V(G_i))= \sigma(V(G_j))$ for all $i,j\in I_r$, i.e., all
connected components $G_i$ with the same chromatic number are colored by
the same color set. By construction, at every node $v\in C(T^*)$ of the
caterpillar structure, with children $v'$ and $v''$, the components
$G':=G(v')$ and $G'':=G(v'')$ satisfy $\chi(G')<\chi(G'')$. Invoking
\AX{(K3)} we therefore have $\sigma(G')\subset\sigma(G'')$ and
$G^*:=G(v)$ is colored by the color set $\sigma(G^*)=\sigma(G'')$. These
set inclusions therefore imply a linear ordering of the colors such that
colors in $\sigma(G')$ come before those in
$\sigma(G'')\setminus \sigma(G')$. Thus $\sigma$ is a greedy coloring
provided that the restriction of $\sigma$ to each of the connected
components $G_i$ of $G$ is a greedy coloring, which is true due to the
assumption that $G$ is a minimal counterexample. Thus no minimal
counterexample exists, and the coloring $\sigma$ is indeed a greedy
coloring of $G$.
\end{proof}
Given an hc-cograph, it is not difficult to recover a corresponding binary
cotree. To this end, we proceed top down. Denote the root of $(T,t)$ by
$r$. It is associated with the graph $G(r)=G$. In the general step we
consider an induced subgraph $G(u)$ of $G$ associated with a vertex $u$ of
$T$. If $G(u)$ is connected, then $t(u)=1$ and $G(u)$ is the joint of pair
of induced subgraphs $G(v_1)$ and $G(v_2)$. To identify these graphs,
consider the connected components $\overline{G_1},\dots,\overline{G_k}$ of
the complement $\overline{G(u)}$ of $G(u)$. We have
\begin{equation}
G(u) = \overline{\bigcupdot_{i=1}^k \overline{G_i}} =
\bigop{\triangledown}_{i=1}^k \overline{\overline{G_i}} =
\bigop{\triangledown}_{i=1}^k G_i \,.
\end{equation}
We therefore set $G(v_1)=G_1$ and
$G(v_2)=\overline{\bigcupdot_{i=2}^k \overline{G_i}}=\bigop{\triangledown}_{i=2}^k
G_i$. By construction, we therefore have $G(u)=G(v_1)\join G(v_2)$ with
disjoint color sets $\sigma(V(G(v_1)))$. If $G(u)$ is disconnected, define
$t(u)=0$, identify one of the components, say $G_1$, with the smallest
numbers of colors $|\sigma(V(G_1))|$ and set $G(v_1)=G_1$ and
$G(v_2)=G(u)\setminus G(v_1)$. The fact that $G(u)$ is an hc-cograph
ensures that $\sigma(V(G(v_1)))\subseteq \sigma(V(G(v_2)))$. In both the
connected and the disconnected case we attach $v_1$ and $v_2$ as the
children of $u$ in $T$. The reconstruction of $(T,t)$ can be preformed in
linear time.
\section{Recursively-Minimal Colorings}
Not every minimal coloring of a cograph is an hc-coloring. For instance, if
$G$ is a disconnected cograph, i.e., $G=\bigcupdot_{i=1}^n G_i$, then it
suffices that $|\sigma(V_i)|\le\chi(G)$. In this case, we may use more
colors than necessary on a connected component $G_i$ of $G$, resulting in
$\chi(G_i)\neq |\sigma(V_i)|$. Thus, Theorem \ref{thm:hc-implies-chrom}
implies that $\sigma$ is not an hc-coloring of $G_i$ and hence, by
definition, not an hc-coloring of $G$. This suggests to consider another
class of colored cographs.
\begin{definition}
A color-minimal cograph $(G,\sigma)$ is either a $K_1$, the disjoint
union of color-minimal cographs or the join of color-minimal cographs,
and satisfies $|\sigma(V)|=\chi(G)$. A coloring $\sigma$ of a
color-minimal cograph will be called \emph{recursively minimal}.
\end{definition}
Color-minimal cographs thus are those colorings for which every constituent
in their construction along \emph{some} binary cotree is colored with the
minimal number of colors. Since every greedy coloring of every cograph
satisfies this condition, every cograph has a recursively minimal coloring.
\begin{theorem}
Let $G$ be a cograph. A coloring $\sigma$ of $G$ is recursively minimal
if and only it is an hc-coloring.
\label{thm:mrc}
\end{theorem}
\begin{proof}
Since every hc-coloring $\sigma$ of a cograph $G$ uses exactly $\chi(G)$
colors, the recursive definition of hc-colorings immediately implies that
$\sigma$ is recursively minimal.
Now suppose there is a minimal cograph $G$ with a coloring $\sigma$ that
is recursively minimal but not an hc-coloring. If $G$ is connected, then
$G=\bigop{\triangledown}_{i=1}^n G_i$ for some $n\ge 2$ and the restrictions of
$\sigma$ to the connected components $G_i$ use disjoint color
sets. Hence, $\sigma$ is an hc-coloring whenever the restriction to each
$G_i$ is an hc-coloring. Thus a minimal counterexample cannot be
connected. Now suppose $G=\bigcupdot_{i=1}^n G_i$ for some $n\ge
2$. Since $(G,\sigma)$ is by assumption a minimal counterexample, each
connected component $(G_i,\sigma_i)$ is an hc-cograph. By Equ.\
\eqref{eq:rec1} there is a connected component, say w.l.o.g.\ $G_1$, such
that $\chi(G)=\chi(G_1)$. By definition, $\sigma$ induces a recursively
minimal coloring $\sigma_1$ on $G_1$ and $\sigma'$ on
$G'=\bigcupdot_{i=2}^n G_i$. Since $G$ is a minimal counterexample,
$\sigma_1$ and $\sigma'$ are hc-colorings of $G_1$ and $G'$,
respectively, in other words $(G_1,\sigma_1)$ and $(G',\sigma')$ are
hc-cographs. Moreover, $\chi(G)=\chi(G_1)$ implies
$\sigma'(V(G'))\subseteq \sigma_1(V(G_1))$. In summary, therefore,
$(G,\sigma) = (G_1,\sigma_1) \charfusion[\mathbin]{\cup}{\cdot} (G',\sigma')$ satisfies \AX{(K3)},
thus it is a cograph with hc-coloring $\sigma$. Hence, there cannot
exist a minimal cograph with a coloring $\sigma$ that is recursively
minimal but not an hc-coloring.
\end{proof}
Recursively minimal colorings can be constructed in a very simply manner by
stepwisely relabeling colors of disconnected subgraphs as outlined in
Alg.~\ref{alg:cBMG}.
\begin{algorithm}
\caption{Recursively minimal coloring of a cograph}
\label{alg:cBMG}
\algsetup{linenodelimiter=}
\begin{algorithmic}[1]
\REQUIRE Cograph $G$
\STATE $(T_G,t_G) \leftarrow$ discriminating cotree of $G$
\STATE initialize all $v \in V(G)$ with different colors
\FORALL {$v\in V^{0}(T_G)$, from bottom to top}
\IF {$t(v) = 0$}
\STATE $\mathcal{G} \leftarrow$ set of connected components of $G(v)$
\STATE $G^* \leftarrow \argmax_{G_i\in\mathcal{G}} \chi(G_i)$
\STATE $S \leftarrow \sigma(V(G^*))$
\FORALL {$G_j\in\mathcal{G}, G_j\neq G^*$}
\STATE randomly choose an injective map
$\phi:\sigma(V(G_j))\to S$
\FORALL {$x\in V(G_j)$}
\STATE $\sigma(x)\leftarrow \phi(\sigma(x))$
\ENDFOR
\ENDFOR
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{theorem}\label{Thm:Algo1}
Given a cograph $G$, Algorithm \ref{alg:cBMG} returns a recursively
minimal coloring of $G$. Moreover, every recursively minimal coloring of
a cograph $G$ can be constructed with this algorithm.
\end{theorem}
\begin{proof}
The bottom-up traversal of the cotree ensures that for every inner vertex
$v$ of $T_G$, the subgraphs induced by its children ${v_1,...,v_j}$ are
color-minimal cographs. In particular this means that
$|\sigma(V(G(v_i))|=\chi(G(v_i))$ for all $i\in \{1,..,j\}$. Take an
arbitrary $v\in V^0(G)$. Suppose $t(v)=0$. The fact that $T_G$ is a
discriminating cotree implies that all $t(v_i)=1$, for all
$i\in \{1,..,j\}$, and therefore $G_i := G(v_i)$ are connected
components. Futhermore, there exists a $G_k$ such that
$\chi(G_k)=\max_{i}\chi(G_i)$. These observations and Equ.\
\eqref{eq:rec1} guarantee that lines 5-7 obtain a set of color $S$ such
that $|S|=\chi(G(v))$. Explicitly, $\chi(G(v))=\chi(G_k)$ and
$S=\sigma(V(G_k))$. An injective recoloring
$\phi:\sigma(V(G_i)) \rightarrow S$, in lines 8-11, assures that
$\sigma(G_i)\subseteq S$ and therefore $\sigma(V(G(v))=S$. This implies
that $G(v)$ is a color-minimal cograph with $\sigma$ a recursively
minimal coloring. The converse is followed by the fact that every
recursively minimal coloring can be obtained with a particular injection
$\phi$.
\end{proof}
Algorithm \ref{alg:cBMG} can be modified easily to construct a recursively
minimal coloring of $G$ with respect to a user defined cotree $(T,t)$. It
suffices to replace the connected components of $G(v)$ by the (not
necessarily connected) induced subgraphs $G(w)$ corresponding to the
children of $v$. Since $\chi(G(v))=\max_{w\in\mathsf{child}(v)}\chi(G(w))$, it
suffices to choose the color set of the child that uses the largest number
of colors and re-color all other child-graphs with this color set. For
completeness, we summarize this variant in Algorithm \ref{alg:cBMG2}.
\begin{algorithm}
\caption{Recursively minimal coloring of a cograph w.r.t.\ a given
cotree.}
\label{alg:cBMG2}
\algsetup{linenodelimiter=}
\begin{algorithmic}[1]
\REQUIRE Cograph $G$ and co-tree $(T,t)$
\STATE initialize all $v \in V(G)$ with different colors
\FORALL {$v\in V^{0}(T)$, from bottom to top}
\IF {$t(v) = 0$}
\STATE $\mathcal{G} \leftarrow \{G(w)\colon w\in\mathsf{child}(v)\}$
\STATE $G^* \leftarrow \argmax_{w\in\mathsf{child}(v)} |\chi(G(w))|$
\STATE $S \leftarrow \sigma(V(G^*))$
\FORALL {$G_j\in\mathcal{G}, G_j\neq G^*$}
\STATE randomly choose an injective map
$\phi:\sigma(G_j)\to S$
\FORALL {$x\in V(G_j)$}
\STATE $\sigma(x)\leftarrow \phi(\sigma(x))$
\ENDFOR
\ENDFOR
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
The recursive structure of hc-cographs can also be used to count the number
of distinct hc-colorings of a cograph $G$ that is explained by a cotree
$(T,t)$. For an inner vertex $u$ of $T$ denote by $N(G(u))$ the number of
hc-colorings of $G(u)$. If $u$ is a leaf, then $N(u)=1$. Recall that $T$
is binary by Def.\ \ref{def:hc-coloring}, i.e.,
$\mathsf{child}(u)=\{v_1,v_2\}$. For $t(u)=1$, we have $N(G(u))=N(G(v_1))N(G(v_2))$
since the color sets are disjoint. If $t(u)=0$, assume, w.l.o.g.\
$s_1:=|\sigma(G(v_1))|\le|\sigma(G(v_2))|=:s_2$,
$N(G(u))=N(G(v_1)) N(G(v_2)) g(s_1,s_2)$, where $g(s_1,s_2)$ is the number
of injections between a set of size $s_1$ into a set of size $s_2$, i.e.,
$g(s_1,s_2)=\binom{s_2}{s_1} s_1!$.
The total number of hc-colorings can be obtained by considering a
caterpillar tree for the step-wise union of connected components. For
each connected component $G_i$ with $\chi(G_i)=s_i$, and $s=\max_i s_i$
there are $\binom{s}{s_i}$ choices of the colors, i.e., $g(s,s_i)$
injections and thus $N(G)=\prod_i g(s,s_i) N(G_i)$ colorings. We
note in passing that the chromatic polynomial of a cograph, and thus the
number of colorings using the minimal number of colors, can be computed
in polynomial time \cite{Makowsky:06}. There does not seem to be an obvious
connection between the hc-colorings and the chromatic polynomial,
however.
\section{Concluding Remarks}
The cotrees $(T,t)$ associated with a cograph are a special case of the
modular decomposition tree \cite{Gallai:67}, which in addition to disjoint
unions and joins also contains so-called prime nodes. The latter have a
special structure known as spiders, which also admit a well-defined unique
decomposition in so-called $P_4$-sparse graphs \cite{Jamison:92}. For this
type of graphs it also makes sense to consider recursively minimal
colorings. More generally, many interesting classes of graphs admit
recursive constructions \cite{Proskurowski:81,Noy:04}. For every graph
class that has a recursive construction, one can ask whether minimal
colorings can be constructed from optimal colorings, i.e., whether
recursively minimal colorings exist. In some cases, such Cartesian products
of graphs, where $\chi(G)$ equals the maximum of the chromatic numbers of
the factors \cite{Sabidussi:57}, this seems rather straightforward. In
general, however, the answer is probably negative.
\subsection*{Acknowledgments}
This work was support in part by the German Federal Ministry of Education
and Research (BMBF, project no.\ 031A538A, de.NBI-RBC) and the Mexican
Consejo Nacional de Ciencia y Tecnolog{\'i}a (CONACyT, 278966 FONCICYT2).
\bibliographystyle{plain}
|
\section{Introduction}
Driving the state of a dynamical system to a given desired point or a desired set is a problem of major practical importance, particularly in the fields of robot motion planning and safety-critical control. Various types of constraints are present in real-world applications due to structural and operational requirements. Such constraints impose limitations on several aspects of the control synthesis. For example, spatial constraints, i.e., constraints requiring the system trajectories to evolve in some \textit{safe} set at all times, while visiting some goal set(s), are common in safety-critical applications. Furthermore, temporal constraints, i.e., constraints pertaining to convergence within a {fixed} time, appear in time-critical applications, for instance when a task must be completed within a given time interval. \textit{Spatiotemporal} specifications impose spatial as well as temporal or time constraints on the system trajectories.
Various approaches have been developed to achieve convergence of system trajectories to desired set points while satisfying safety and/or control input constraints. Model predictive control (MPC) methods \cite{saska2014motion,grancharova2015uavs}, rapidly-exploring random tree (RRT) based methods \cite{svenstrup2010trajectory,palmieri2016rrt,salzman2016asymptotically}, as well as combinations of them \cite{svenstrup2010trajectory} have been studied extensively in the literature. In addition, vector field approaches \cite{kovacs2016novel,han2019robust} and solutions based on control Lyapunov functions (CLF) \cite{srinivasan2018control,li2018formally,ames2012control} are also popular, in part because they are inherently amenable to Lyapunov analysis.
Safety is typically realized as \textit{forward invariance} of a \textit{safe} set. The control objective reduces to designing a control law such that the closed-loop system trajectories remain always in the safe set. The authors in \cite{tee2009barrier} use a Lyapunov-like barrier function approach to guarantee asymptotic tracking and that the system output always remains inside a given set. More recently, in \cite{ames2014control,ames2017control}, conditions using zeroing control barrier functions (ZCBF) are presented to ensure forward invariance of a desired set. The authors in \cite{barry2012safety} propose a sum-of-squares formulation to find a barrier certificate that guarantees forward invariance of a given set. Forward invariance of a safe set in addition to convergence to a desired set (or point) can be achieved via a combination of CLFs and CBFs \cite{ames2017control,ames2014control,romdlony2016stabilization}. Input constraints, such as actuator saturation, is another class of constraints that is inevitable in practice. Since a limited control input can affect the region of convergence, addressing spatiotemporal and input constraints simultaneously is a challenging control problem.
In recent years, online optimization, particularly, quadratic program (QP) based approaches have gained popularity for control synthesis; see \cite{li2018formally,srinivasan2018control,ames2017control,rauscher2016constrained, cortez2019control,xu2015robustness}. These methods are suitable for real-time implementation as QPs can be solved efficiently \cite{cortez2019control,glotfelter2017nonsmooth,glotfelter2018boolean} {(see \cite[Chapter 5]{gill2019practical} regarding numerical methods for solving such constrained optimization problems)}. The authors in \cite{nguyen2016exponential} use an exponential barrier function in the QP formulation to guarantee safety of the closed-loop trajectories (see also \cite{ames2014rapidly}).
In \cite{wang2017safety}, safety barrier certificates are presented to ensure scalable collision-free behavior in multi-robot systems. The control input is realized as the solution to a QP that encodes the minimization of the deviation from a given \textit{nominal} controller while guaranteeing safety. A similar approach is used in \cite{wang2016safety}, where the dynamics of the neighboring agents are assumed to be unknown.
{{ Most of the aforementioned contributions address control design that achieves convergence to a desired goal set (or point), but without explicitly considering control input constraints. Such constraints are considered in \cite{ames2017control}, where performance and safety objectives are represented using CLFs and CBFs, respectively, along with control input constraints in a single QP. More recently, the authors in \cite{cortez2019control} address the issue of sampling effects and additive bounded disturbances in guaranteeing safety using robust control barrier functions; more specifically, they formulate a QP to design a control input for a zero-order hold mechanical system in the presence of component-wise input constraints.
}}
{In contrast to asymptotic stability (AS), which pertains to convergence of system trajectories as time tends to infinity, finite-time stability (FTS) is a concept that guarantees convergence in finite time. Hence, for specifications involving temporal constraints and time-critical systems, the theory of finite- or fixed-time stability can be leveraged in the control design to guarantee that such specifications are met.} In the seminal work \cite{bhat2000finite}, the authors introduce necessary and sufficient conditions in terms of a Lyapunov function for the equilibrium of a continuous-time, autonomous system to exhibit finite-time stability (FTS). Fixed-time stability (FxTS) \cite{polyakov2012nonlinear} is a stronger notion than FTS, where the time of convergence does not depend upon the initial conditions.
{It has also been shown that a faster rate of convergence generally implies that the closed-loop system has better disturbance rejection properties \cite{bhat2000finite}, which further motivates the study of finite- and fixed-time stable systems.
}
The authors of \cite{srinivasan2018control} formulate a QP for finite-time convergence to a desired set, yet without considering input constraints. This limitation is removed in \cite{li2018formally}, where the authors consider a QP formulation incorporating input constraints as well, in addition to the safety and convergence constraints. The authors in \cite{lindemann2019control} use CBFs to encode signal-temporal logic (STL) based specifications, involving reaching to a goal set within a finite time, and formulate a QP to compute the control input.
In this paper, we study the problem of reaching a given goal set $S_G$ {within a fixed time $T_{ud}$}, while remaining in a given safe set $S_S$ at all times, for a general class of nonlinear control-affine systems with input constraints. {In the preliminary conference version \cite{garg2019control}, a QP formulation is proposed to compute the control input for fixed-time convergence with input and safety constraints, yet without any guarantees on the feasibility of the proposed method. The current paper has several new results in comparison to \cite{garg2019control} and prior literature.}
We formulate a QP framework for control synthesis of nonlinear, control-affine systems, that guarantees forward invariance of a safe set and {fixed-time (FxT) convergence of the system trajectories to a goal set. We use slack terms in the safety and convergence constraints to guarantee that the QP is always feasible, even in the presence of component-wise control input constraints.
We discuss the continuity of the solution of the QP as a function of the state variables so that the closed-loop solutions are well-defined, and safety can be guaranteed. We also characterize the region of convergence as a function of the slack variables used in the QP. It is worth noting that in the presence of control input constraints, it is not possible to guarantee FxTS for any initial condition. This limitation is discussed in \cite{garg2019prescribed}, where a set of initial conditions is characterized for guaranteeing FxTS under input constraints (see Remark \ref{remark: d1 T U}).}
{
The proposed approach applies to a general class of nonlinear, control-affine systems, and makes no assumptions on the convexity of the safe set or the goal set. The fact that QPs can be solved efficiently enables the use of the proposed method in real-time safety- and time-critical applications. Furthermore, the approach advances the results in \cite{srinivasan2018control,li2019finite,li2018formally} in terms of the achieved time of convergence. More specifically, the aforementioned papers utilize the notion of FTS \cite{bhat2000finite} and therefore, the convergence time depends upon the initial conditions; in contrast, our approach guarantees that the closed-loop system trajectories reach the goal set within a \textit{fixed} time that can be chosen independently of the initial conditions.} The authors in \cite{ames2017control} show feasibility of the proposed QP formulation in the absence of control input constraints under certain conditions on the existence of a particular class of CBFs.
{In contrast to the aforementioned work, we consider control input constraints in addition to the safety and convergence requirements, and guarantee the feasibility of the proposed QP. Furthermore, we generalize the results of \cite{ames2017control,xu2015robustness,glotfelter2017nonsmooth}, where the regularity properties of the solution of the QP is discussed in the absence of input constraints; we show continuity of the solution of the proposed QP under the presence of component-wise input constraints, and under relaxed regularity assumptions on the CLFs, CBFs, and the system dynamics.}
\section{Mathematical Preliminaries}\label{sec: math prelim}
In the rest of the paper, ${\mathbb R}$ denotes the set of real numbers and $\mathbb R_+$ denotes the set of non-negative real numbers. We use
$\|\cdot\|_p$ to denote the $p-$norm, and
$\|\cdot\|$ to denote the Euclidean norm. We write $\partial S$ for the boundary of a closed set $S\in \mathbb R^n$, $\textnormal{int}(S) \coloneqq S\setminus \partial S$ for its interior, and $|x|_S = \inf_{y\in S}\|x-y\|$ for the distance of a point $x\in \mathbb R^n\setminus S$ from the set $S$. We use $\mathcal C^k$ to denote the set of $k$ times continuously differentiable functions. The Lie derivative of a $\mathcal C^1$ function $V:\mathbb R^n\rightarrow \mathbb R$ along a vector field $f:\mathbb R^n\rightarrow\mathbb R^n$ at a point $x\in \mathbb R^n$ is denoted as $L_fV(x) \coloneqq \frac{\partial V}{\partial x} f(x)$. For two vectors $x, y\in \mathbb R^n$, we use $x\leq y$ to represent element-wise inequalities $x_i\leq y_i, i = 1, 2, \ldots, n$. A continuous function $\alpha:\mathbb R_+\rightarrow\mathbb R_+$ is a class-$\mathcal K$ function if it is strictly increasing and $\alpha(0) = 0$. It belongs to $\mathcal K_\infty$ if in addition, $\lim_{r\to \infty}\alpha(r) = \infty$. A function $\kappa:\mathbb R_+\times\mathbb R_+\rightarrow\mathbb R_+$ is a class-$\mathcal{KL}$ function if 1) for all $t\geq 0$, the map $\kappa(r,t)$ belongs to class-$\mathcal K$ and 2) for all $r\geq 0$, the map $\kappa(r,t)$ is decreasing in $t$ with $\kappa(r,t)\to 0$ as $t\to \infty$. We drop the arguments $t, x$ whenever clear from the context.
Next, we present the necessary definitions for the problem formulation. Consider the nonlinear control affine system
\begin{align}\label{ex sys}
\dot x(t) = f(x(t)), \quad x(0) = x_0,
\end{align}
where $x\in \mathbb R^n$ and $f: \mathbb R^n \rightarrow \mathbb R^n$ is continuous with $f(0)=0$. The origin is said to be an FTS equilibrium of \eqref{ex sys} if it is Lyapunov stable and \textit{finite-time convergent}, i.e., for all $x(0) \in \mathcal N \setminus\{0\}$, where $\mathcal N$ is some open neighborhood of the origin, $\lim_{t\to T} x(t)=0$, where $T = T(x(0))<\infty$ depends upon the initial condition $x(0)$ \cite{bhat2000finite}. The authors in \cite{polyakov2012nonlinear} presented the following result for fixed-time stability, where the time of convergence does not depend upon the initial condition.
\begin{Theorem}[\cite{polyakov2012nonlinear}]\label{FxTS TH}
Suppose there exists a positive definite function $V$ for system \eqref{ex sys} such that
\begin{align}\label{eq: dot V FxTS old }
\dot V(x) \leq -aV(x)^p-bV(x)^q,
\end{align}
with $a,b>0$, $0<p<1$ and $q>1$. Then, the origin of \eqref{ex sys} is FxTS with continuous settling time function
\begin{align}\label{eq: T bound old}
T \leq \frac{1}{a(1-p)} + \frac{1}{b(q-1)}.
\end{align}
\end{Theorem}
\noindent We need the following lemma to prove one of the main results of the paper.
\begin{Lemma}\label{lemma:int dot V}
Let $V_0, \alpha_1, \alpha_2>0$, $\delta_1\in \mathbb R$, $\gamma_1 = 1+\frac{1}{\mu}$ and $\gamma_2 = 1-\frac{1}{\mu}$, where $\mu>1$. Define
\begin{align}\label{eq:int dot V}
I \triangleq \int_{V_0}^{0}\frac{dV}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}.
\end{align}
Then, the following holds:
\begin{itemize}
\item[(i)] If $0\leq \delta_1<2\sqrt{\alpha_1\alpha_2}$, we have for all $V_0\geq 0$
\begin{align}\label{eq: I bound 1}
I\leq \frac{\mu}{\alpha_1k_1}\left(\frac{\pi}{2}-\tan^{-1}k_2\right),
\end{align}
where $k_1 = \sqrt{\frac{4\alpha_1\alpha_2-\delta_1^2}{4\alpha_1^2}}$ and $k_2 = -\frac{\delta_1}{\sqrt{4\alpha_1\alpha_2-\delta_1^2}}$;
\item[(ii)] If $\delta_1\geq 2\sqrt{\alpha_1\alpha_2}$ and $V_0^\frac{1}{\mu}\leq k\frac{\delta_1-\sqrt{\delta_1^2-4\alpha_1\alpha_2}}{2\alpha_1}$ with $0<k<1$, we have for all $V_0\geq 0$
\begin{align}\label{eq: I bound 2}
I\leq \frac{\mu}{\alpha_1(b-a)}\left(\log\left(\frac{b-ka}{a(1-k)}\right)-\log\left(\frac{b}{a}\right)\right),
\end{align}
where $a\leq b$ are the roots of $\gamma(z) \triangleq \alpha_1z^2-\delta_1z+\alpha_2 = 0$;
\end{itemize}
\end{Lemma}
\begin{proof}
We have
\begin{align*}
I & = \int_{V_0}^{0}\frac{dV}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}\\
& = \int_{V_0}^{0}\frac{dV}{V(-\alpha_1V^{a_1-1}-\alpha_2V^{a_2-1}+\delta_1)}\\
& = \int_{V_0}^{0}\frac{dV}{V(-\alpha_1V^{\frac{1}{\mu}}-\alpha_2V^{\frac{-1}{\mu}}+\delta_1)}.
\end{align*}
Substitute $m = V^{\frac{1}{\mu}}$, so that $dm = \frac{1}{\mu}V^{\frac{1}{\mu}-1}dV$, which implies that $\frac{1}{\mu}\frac{dV}{V} = \frac{dm}{V^\frac{1}{\mu}} = \frac{dm}{m}$. Using this, we obtain that
{
\begin{align*}
I & = \mu\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{m(-\alpha_1m-\alpha_2\frac{1}{m}+\delta_1)} = \mu\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{(-\alpha_1m^2-\alpha_2+\delta_1m)}.
\end{align*}
}
Now, we consider the three cases, namely, $\delta_1< 2\sqrt{\alpha_1\alpha_2}$, $\delta_1= 2\sqrt{\alpha_1\alpha_2}$ and $\delta_1> 2\sqrt{\alpha_1\alpha_2}$ separately.
First, consider the cases when $\delta_1< 2\sqrt{\alpha_1\alpha_2}$. In the case, we can re-write $I$ as
\begin{align*}
I & = \mu\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{-\alpha_1\left((m-\frac{\delta_1}{2\alpha_1})^2+\frac{4\alpha_1\alpha_2-\delta_1^2}{4\alpha_1^2}\right)}
\end{align*}
Evaluating the integral, we obtain
\begin{align*}
I & = \frac{\mu}{-\alpha_1k_1}(\tan^{-1}k_2-\tan^{-1}k_3),
\end{align*}
where $k_1 = \sqrt{\frac{4\alpha_1\alpha_2-\delta_1^2}{4\alpha_1^2}}$, $k_2 = -\frac{\delta_1}{\sqrt{4\alpha_1\alpha_2-\delta_1^2}}$ and $k_3 = \frac{2\alpha_1V_0^{\frac{1}{\mu}}-\delta_1}{\sqrt{4\alpha_1\alpha_2-\delta_1^2}}$. Hence, we have that
\begin{align*}
I & = \frac{\mu}{\alpha_1k_1}(\tan^{-1}k_3-\tan^{-1}k_2)\leq \frac{\mu}{\alpha_1k_1}(\frac{\pi}{2}-\tan^{-1}k_2),
\end{align*}
since $\tan^{-1}(\cdot)\leq \frac{\pi}{2}$.
Next, we consider the case when $\delta_1> 2\sqrt{\alpha_1\alpha_2}$. In this case, the roots of $\gamma(m) = 0$ are real. Let $a\leq b$ be the such that $\alpha_1m^2-\delta_1m+\alpha_2 = \alpha_1(m-a)(m-b)$. This substitution allows us to factorize the denominator to evaluate the integral $I$. Note that since $ab = \alpha_2>0$ and $a+b = \delta_1$, we have $0<a\leq b$. Since $V_0^\frac{1}{\mu}\leq k\frac{\delta_1-\sqrt{\delta_1^2-4\alpha_1\alpha_2}}{2\alpha_1} = ka$ where $k<1$, we have that $\frac{1}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}<0$ for all $V\leq V_0$, i.e., the denominator $\delta_1V-\alpha_1V^{\gamma_1}+\alpha_2V^{\gamma_2}$ does not vanish for $V\in [0, V_0]$. Thus, we obtain that
{
\begin{align*}
I & = \mu\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{(-\alpha_1m^2-\alpha_2+\delta_1m)} = -\frac{\mu}{\alpha_1}\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{(m-a)(m-b)}\\
& = -\frac{\mu}{\alpha_1(a-b)}\left(\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{m-a}-\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{m-b}\right).
\end{align*}
}
Evaluating the integrals, we obtain
\begin{align*}
I & = \frac{-\mu}{\alpha_1(a-b)}\left(\log\left(\frac{a}{|V_0^\frac{1}{\mu}-a|}\right)-\log\left(\frac{b}{|V_0^\frac{1}{\mu}-b|}\right)\right)\\
& = \frac{\mu}{\alpha_1(a-b)}\left(\log\left(\frac{b}{a}\right)+\log\left(\frac{|V_0^\frac{1}{\mu}-a|}{|V_0^\frac{1}{\mu}-b|}\right)\right)\\
& \leq \frac{\mu}{\alpha_1(b-a)}\left(\log\left(\frac{b-ka}{a(1-k)}\right)-\log\left(\frac{b}{a}\right)\right),
\end{align*}
Finally, for the case when $\delta_1 = 2\sqrt{\alpha_1\alpha_2}$, we have $a = b = \frac{-\delta_1}{2\alpha_1} = \sqrt{\frac{\alpha_2}{\alpha_1}}$, and thus,
\begin{align*}
I & = -\frac{\mu}{\alpha_1}\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{(m-a)(m-b)} = -\frac{\mu}{\alpha_1}\int_{V_0^{\frac{1}{\mu}}}^{0}\frac{dm}{(m-a)^2}.
\end{align*}
It is easy to see that for $V_0^\frac{1}{\mu}\leq ka<a = \frac{\delta_1}{2\alpha_1} = \sqrt{\frac{\alpha_2}{\alpha_1}}$, the integral $I$ evaluates to a finite value. Thus, for all $V_0^\frac{1}{\mu}\leq k\sqrt{\frac{\alpha_2}{\alpha_1}}<\sqrt{\frac{\alpha_2}{\alpha_1}}$ for $0<k<1$, we have that
\begin{align*}\
I & = \frac{\mu}{\alpha_1}\Big(-\frac{1}{a}-\frac{1}{-a+V_0^\frac{1}{\mu}}\Big) \leq \frac{\mu}{\alpha_1}\Big(-\frac{1}{a}-\frac{1}{-a+k\sqrt{\frac{\alpha_2}{\alpha_1}}}\Big)\\
& = \frac{\mu}{\alpha_1}\sqrt{\frac{\alpha_1}{\alpha_2}}\Big(-1-\frac{1}{-1+k}\Big) = \frac{\mu}{\sqrt{\alpha_1\alpha_2}}\Big(\frac{k}{1-k}\Big).
\end{align*}
This completes the proof.
\end{proof}
\subsection{Problem formulation}
In this section, we present the main problem considered in this paper. Specifically, let the dynamical control system be given by
\begin{align}\label{cont aff sys}
\dot x(t) = f(x(t)) + g(x(t))u, \quad x(t_0) = x_0,
\end{align}
where $x\in \mathbb R^n$ is the state vector, $f:\mathbb R^n\rightarrow \mathbb R^n$ and $g:\mathbb R^n\rightarrow \mathbb R^{n\times m}$ are system vector fields, continuous in their arguments, and $u\in \mathbb R^m$ is the control input vector. In addition, define a safe set $S_s = \{x\; |\; h_s(x)\leq 0\}$, and consider a goal set to be reached in a user-defined {fixed} time $T_{ud}$ defined as $S_g = \{x\; |\; h_g(x)\leq 0\}$, where $h_s, h_g:\mathbb R^n\rightarrow\mathbb R$ are user-defined functions. {We make the following standing assumption.
\begin{Assumption}\label{assum: sets Ss Sg}
The functions $h_S(x),h_G(x)\in \mathcal C^1$ , $S_G\bigcap S_S\neq \emptyset$, the set $S_G$ is compact, and the sets $S_S$ and $S_G$ have non-empty interiors. Furthermore, the function $h_G$ is proper with respect to set $S_G$, i.e., {there exists a class-$\mathcal K_\infty$} function $\alpha_g$ such that $h_G(x)\geq \alpha_g(|x|_{S_G})$, for all $x\notin S_G$.
\end{Assumption}
\noindent The main problem considered in this paper is as follows:
\begin{Problem}\label{P reach S}
Design a control input $u(t)\in \mathcal U = \{v\in \mathbb R^m\; |\; u_{min_i}\leq v_i\leq u_{max_i}, i = 1, 2, \ldots, m\}$, so that the closed-loop trajectories of \eqref{cont aff sys} reach the set $S_g = \{x\; |\; h_g(x)\leq 0\}$ in a {fixed} time $T_{ud}>0$ with $x(t) \in S_s = \{x\; |\; h_s(x)\leq 0\}$ for all $t\geq 0$ and $x_0\in S_s$.
\end{Problem}
\noindent Here, $u_{min_i}$ and $u_{max_i}$ where $u_{min_i}< u_{max_i}$, are the lower and upper bounds on the individual input $u_i$ for $i = 1, 2, \ldots, m$, respectively. Input constraints of this form are very commonly considered in the literature \cite{cortez2019control,ames2017control,ames2014rapidly}. We can write the constraint set in a compact form as $\mathcal U = \{v \; | \; A_uv\leq b_u\}$, where $A_u\in \mathbb R^{2m\times m}, b_u\in \mathbb R^{2m}$.
Problem \ref{P reach S} can be readily translated into a temporal logic formula for the form of specifications that are encountered, for instance, in mission planning problems,
The signal temporal logic (STL) specifications, given by formula $\phi$ include the following semantics:
\begin{itemize}
\item $(x,t)\models\phi \iff h(x(t))\leq 0$;
\item $(x,t)\models\lnot \phi \iff h(x(t)) > 0$;
\item $(x,t)\models \phi_1\land \phi_2 \iff (x,t)\models \phi_1\land (x,t)\models \phi_2$;
\item $(x,t)\models G_{[a,b]}\phi \iff h(x(t))\leq 0, \forall t\in [a,b]$;
\item $(x,t)\models F_{[a,b]}\phi \iff \exists t\in [a,b]$ such that $h(x(t))\leq 0$,
\end{itemize}
where $\phi = \textrm{true}$ if $h(x)\leq 0$ and $\phi = \textrm{false}$ if $h(x)>0$. The temporal functions $G_{[a,b]}\phi$ and $F_{[a,b]}\phi$ stand for satisfaction of the formula $\phi$ \textit{always} in the interval $[a,b]$, i.e., for all $t\in [a,b]$ and \textit{eventually} in the interval $[a,b]$, i.e., for some $t\in [a,b]$, respectively. So, Problem \ref{P reach S} can be written in the STL semantics as follows.
\begin{Problem}
Design control input $u\in \mathcal U$ so that the closed-loop trajectories satisfy
\begin{align}
(x,0)\models & G_{[0, T]}\phi_s\land F_{[0, T]}\phi_g,
\end{align}
where $\phi_s$ (respectively, $\phi_g$) = true if $h_s(x)$ (respectively, $h_g(x)$) $\leq 0$, and false otherwise.
\end{Problem}
Note that the requirements involving more complex STL formula, for examples, requirements of the form
\begin{align}\label{eq: STL req new}
(x,0)\models & G_{[t_0, t_N]}\phi_s\land G_{[t_0, t_1]}\phi_0\land F_{[t_0, t_1]}\phi_1\land G_{[t_1, t_2]}\phi_1 \nonumber\\
& \land F_{[t_1, t_2]}\phi_2\land \cdots \land G_{[t_{N-1},t_N]}\phi_{N-1}\land F_{[t_{N-1}, t_N]}\phi_N,
\end{align}
can also be considered in the proposed framework. {Here $[t_0, t_1], [t_1, t_2], \ldots, [t_{N-1}, t_N]$} are intervals such that $t_{i+1}-t_i\geq T_{ud}$ for some $0<T_{ud}<\infty$, for all $0\leq i\leq N-1$, and $\phi_i$ is true if $h_i(x)\leq 0$ (see \cite{garg2019control} for more details). The requirements \eqref{eq: STL req new} translate to the problem of designing control input $u$ such that the closed-loop trajectories satisfy
\begin{subequations}\label{eq: STL req new set}
\begin{align}
x(t)\in S_s,\; & \forall \; t \geq t_0,\\
x(t)\in S_i,\; & \forall\; t\in [t_i, t_{i+1}]\label{s_i inv constr},
\end{align}
\end{subequations}
where $S_i = \{x\; |\; h_i(x)\leq 0\}$, i.e., the closed-loop trajectories should stay in the set $S_s$ at all times, and visit the sets $S_i$ in the given time sequence, can also be considered in the proposed framework. The requirements of the form \eqref{eq: STL req new set} can be satisfied by solving Problem \ref{P reach S} sequentially with safe set defined as $\bar S_s = S_s\bigcap S_i$ and goal set as $\bar S_g =S_s\bigcap S_{i+1}$ for the time interval $[t_i, t_{i+1}]$ with $t_{i+1}-t_i\geq T_{ud}>0$, and requiring the time of convergence to satisfy $T\leq T_{ud}$. It is evident that we need $S_i\bigcap S_{i+1}\neq \emptyset$ for the problem to be well-posed
(Section \ref{sec sim B} presents an instance of such an example for a two-robot motion planning case).
}
\begin{Remark}
If the STL-based specifications satisfy certain assumptions, then these specifications can be addressed in the framework of Problem \ref{P reach S}. For illustration, consider Example 2 from \cite{lindemann2019control}. The STL specification $\phi= \phi_1\land\phi_2$, where $\phi_1 = F_{[5,15]}(\|x-[10 \; 0]^T\|\leq 5)$ and $\phi_2 = G_{[5,15]}(\|x-[10 \; 5]^T\|\leq 10)$, means that the closed loop trajectories should reach the set $S_1 = \{x\; |\; \|x-[10 \; 5]^T\|\leq 10\}$ on or before $t = 5$ sec, remain in the set $S_1$ for $t\in [5, 15]$ and reach the set $S_2 = \{x\; |\; \|x-[10 \; 0]^T\|\leq 5\}$ on or before $t = 15$. Since $S_1\bigcap S_2\neq\emptyset$, we can use the problem set of Problem \ref{P reach S} to address these specifications. In Section \ref{sec sim B}, we present an example on how to address problems that do not satisfy this setup, i.e., if the functions $h(x)$ or $h_i(x)$ are non-smooth or $S_i\bigcap S_{i+1} = \emptyset$, e.g., the case study in \cite{lindemann2017robust}.
\end{Remark}
\section{Safety and FxT convergence}\label{sec: safe FxT conv}
\subsection{Forward-invariance of safe set}
\noindent Since the system trajectories are required to stay in the set $S_s$ at all times, the set $S_s$ can be thought of as a safe set, or its compliment set $\mathbb R^n \setminus S_s$, an unsafe set. Following Nagumo's Theorem on set invariance \cite{blanchini1999set}, we make the following assumption to guarantee that the safe set $S_S$ can be rendered forward invariant for \eqref{cont aff sys}.
\begin{Assumption}\label{Assum feas}
For all $x\in \partial S_S$, there exists a control input $u\in \mathcal U$ such that the following condition holds:
\begin{align}\label{eq: safety cond}
L_fh_S(x)+L_gh_S(x)u\leq 0.
\end{align}
\end{Assumption}
{Similar assumptions have been used in literature, either explicitly (see e.g. \cite{romdlony2016stabilization}) or implicitly (see e.g. \cite{ames2017control}).}
\noindent Below we give some examples where Assumption \ref{Assum feas} holds.
\begin{Example}\label{example: single int}
Consider a robot modelled under single-integrator dynamics
\begin{align}\label{eq: sing int}
\dot x= u,
\end{align}
where $x,u\in \mathbb R^n$. It can be easily shown that \eqref{eq: safety cond} for \eqref{eq: sing int} for any set $S_{s}$ if $\mathcal U$ is of the form $\mathcal U = \{u\; |\; |u_i|\leq u_{M_i}\}$ for some $u_{M_i}>0$, or $\mathcal U = \{u\;|\; \|u\|\leq u_M\}$, for some $u_M>0$. This is true since $f = 0$ and $g = I$ in \eqref{eq: sing int}, and hence $L_fh_s+L_gh_su = \frac{\partial h_s}{\partial x}u$. Due to the structure of $\mathcal U$, {one can choose $ u(x) = -c\frac{1}{\|\frac{\partial h_s}{\partial x}\|}\Big(\frac{\partial h_s}{\partial x}\Big)^T$}, with $c\leq \min\{u_{M_i}\}$ or $c\leq u_M$, so that \eqref{eq: safety cond} holds, for any $x$ such that $\frac{\partial h_s}{\partial x}\neq 0$. When $\frac{\partial h_s}{\partial x} = 0$, \eqref{eq: safety cond} is satisfied trivially for $u = 0$.
\end{Example}
\begin{Example}\label{example: full act LTI}
Consider the fully-actuated nonlinear system
\begin{align}\label{eq: fully act LTI}
\dot x= f(x)+ g(x)u,
\end{align}
where $x,u\in \mathbb R^n$, $f: \mathbb R^n\rightarrow\mathbb R^n$ $g:\mathbb R^n\rightarrow \mathbb R^{n\times n}$ with $\textrm{rank} (g(x)) = n$ for all $x\in \partial S_{s}$. Assume that there are no control input constraints, i.e., $\mathcal U = \mathbb R^n$. For any $x\in \partial S_{s}$, since $g(x)$ is full-rank, choose $u(x) = -g(x)^{-1}\Big(f(x)+\Big(\frac{\partial h_s}{\partial x}\Big)^T\Big)$, so that $L_fh_s+L_gh_su = -\left\|\frac{\partial h_s}{\partial x}\right\|^2\leq 0$. Hence, under the aforementioned conditions, \eqref{eq: fully act LTI} satisfies \eqref{eq: safety cond}.
\end{Example}
In this work, we use the conditions of zeroing CBF (ZCBF) to ensure safety or forward invariance of the safe set $S_s$. The ZCBF is defined by the authors in \cite{ames2017control} as following.
\begin{Definition}
\textbf{ZCBF}: For the dynamical system \eqref{cont aff sys}, a continuously differentiable function $B:\mathbb R^n\rightarrow \mathbb R$ is called a ZCBF for the set $S_S$ if $B(x)<0$ for $x\in \textnormal{int}(S_S)$, $B(x) = 0$ for $x\in \partial S_S$, and there exists $\alpha\in \mathcal K$, such that
\begin{align}\label{eq: ZCBF}
\inf_{u\in \mathcal U}\{L_fB(x)+L_gB(x)u\}\leq \alpha(-B(x)),
\end{align}
for all $x\in S_s$.
\end{Definition}
\noindent It is easy to see that if $h_s$ is a ZCBF for \eqref{cont aff sys}, then it also satisfies \eqref{eq: safety cond}. One special case of \eqref{eq: ZCBF} is
\begin{align}\label{eq: ZCBF spec}
\inf_{u\in \mathcal U}\{L_fB(x)+L_gB(x)u\}\leq -\rho B(x),
\end{align}
with $\rho\geq 0$, as discussed in \cite[Remark 6]{ames2017control}.
\subsection{{Fixed}-time Convergence}
Inspired from \cite{lopez2019conditions}, we define a class of CLF with a user-defined fixed-time convergence guarantees to a compact set $S \subset \mathbb R^n$:
\begin{Definition}\label{def: CLF FxT}
\textbf{FxT CLF-$S$}: A continuously differentiable function $V:\mathbb R^n\rightarrow\mathbb R$ is called FxT CLF-$S$ for \eqref{cont aff sys} with parameters $\alpha_1,\alpha_2>0, \gamma_1>1,0<\gamma_2<1$, if $V(x) >0$ for all $x\notin S$, $V(x) = 0$ for all $x\in \partial S$, and the following holds:
\begin{align}\label{fxts h ineq}
\inf_{u\in \mathcal U}\{L_fV+L_gVu\}\leq -\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2},
\end{align}
for all $x\in \mathbb R^n \setminus S$, where the time of convergence $T$ satisfies $ T\leq \frac{1}{\alpha_1(\gamma_1-1)} + \frac{1}{\alpha_2(1-\gamma_2)}\leq T_{ud}$, where $T_{ud}>0$ is a user-defined fixed time.
\end{Definition}
\noindent Definition \ref{def: CLF FxT} defines a CLF that guarantees convergence of the solutions to the set $S$ within user-defined time $T_{ud}$. The traditional notions of CLF \cite{li2018formally} and exponential CLF \cite{ames2014rapidly} require $\alpha_1 = \alpha_2 = 0$, and $\gamma_1 = 1, \alpha_2 = 0$, in \eqref{fxts h ineq}, guaranteeing only asymptotic and exponential convergence, respectively. Based on \cite{lopez2019conditions}, following sufficient conditions for existence of a control input $u$ that renders the closed-loop trajectories reach the set $S_g$ in the {fixed} time $T_{ud}$ are presented in \cite{garg2019control}.
\begin{Theorem}[\cite{garg2019control}]\label{FxTS reach set S}
If there exist constants $\alpha_1, \alpha_2>0$, $\gamma_1>1$ and $0<\gamma_2<1$, satisfying
\begin{align}\label{T a1 a2 set S}
\frac{1}{\alpha_1(\gamma_1-1)} + \frac{1}{\alpha_2(1-\gamma_2)}\leq T_{ud},
\end{align}
such that $h_g$ is FxT CLF-$S_g$ with parameters $\alpha_1,\alpha_2,\gamma_1,\gamma_2$ for all $x\notin \textrm{int}(S_g)$, then there exists $u(t)\in \mathcal U$, such that the closed-loop trajectories of \eqref{cont aff sys} reach the set $S_g$ within {fixed} time $T_{ud}$ for all initial conditions $x(0)\notin S_g$.
\end{Theorem}
Note that the condition \eqref{fxts h ineq} needs to hold for all $x\notin \textrm{int}(S)$. Another way to formulate this condition is to require that
\begin{align}\label{fxts h ineq new}
\inf_{u\in \mathcal U}\{L_fh_g(x) + L_gh_g(x)u\}\leq & -\alpha_1\max\{0,h_g(x)\}^{\gamma_1}\nonumber\\
& -\alpha_2\max\{0,h_g(x)\}^{\gamma_2},
\end{align}
holds for all $x$. The condition \eqref{fxts h ineq new} guarantees that the closed-loop trajectories reach the set $S_g$ within {fixed} time $T_{ud}$ and stay there for all future times, as shown below.
\begin{Corollary}\label{Set S reach and inv}
If there exist constants $\alpha_1, \alpha_2>0$, $\gamma_1>1$ and $0<\gamma_2<1$ satisfying \eqref{T a1 a2 set S}, such that \eqref{fxts h ineq new} holds
for all $x$, then {there exists control input $u\in \mathcal U$ such that the closed-loop trajectories of \eqref{cont aff sys} reach the set $S_g$ within fixed time $T_{ud}<\infty$ for all initial conditions $x(0)\in \mathbb R^n$, and stay there for all future times.}
\end{Corollary}
\noindent As pointed out in \cite{li2018formally}, QPs can be solved very efficiently and can be used for real-time implementation. We present a QP-based formulation to compute the control input that satisfies the conditions of Theorem \ref{FxTS reach set S}.
\begin{Theorem}\label{QP P1 Theorem}
Let the solution to the following QP
\begin{subequations}\label{QP u P1}
\begin{align}
\min_{v} \;& \frac{1}{2}\|v\|^2 \\
\textrm{s.t.} \; L_fh_g(x) + L_gh_g(x)v & \leq -\alpha_1\max\{0,h_g(x)\}^{\gamma_1}\\
& -\alpha_2\max\{0,h_g(x)\}^{\gamma_2}\\
A_uv & \leq b_u,
\end{align}
\end{subequations}
with $\alpha_1 = \frac{2}{T_{ud}(\gamma_1-1)}, \alpha_2 = \frac{2}{T_{ud}(1-\gamma_2)}$,
be denoted as $v^*(\cdot)$, where $\gamma_1>1$ and $0<\gamma_2<1$. If the control input is defined as $u(\cdot) = v^*(\cdot)$, then \eqref{T a1 a2 set S} and \eqref{fxts h ineq new} hold.
\end{Theorem}
\begin{proof}
Trivial.
\end{proof}
\begin{Remark}
In contrast to \cite{li2018formally,srinivasan2018control}, QP \eqref{QP u P1} has two advantages. First, in the aforementioned work, the time of convergence $T$ depends upon the initial conditions $x(0)$, and grows larger as the distance of $x(0)$ increases from the set $S_g$; while in the above formulation \eqref{QP u P1}, it is independent of $x(0)$. Second, the previous works only concern about reaching the set $S_g$, while the formulation \eqref{QP u P1} additionally renders the set $S_g$ forward-invariant.
\end{Remark}
\section{Main Results}\label{sec: main results}
In this section, we present the main results of the paper. We first present a new FxTS result in Theorem \ref{Th: FxTS new}, which allows a positive term to appear in the time derivative of the Lyapunov candidate. Then, we design a QP, show its feasibility, show that the solution of the QP is a continuous function of $x$, and that it solves Problem \ref{P reach S} under certain conditions. In what follows, unless specified otherwise, all the results are presented under Assumptions \ref{assum: sets Ss Sg}-\ref{Assum feas}.
\subsection{New FxTS result}
Previously, conditions of the form \eqref{eq: dot V FxTS old } are considered by various authors (see for example \cite{polyakov2012nonlinear},\cite{parsegov2012nonlinear}), where the time derivative of the Lyapunov candidate is bounded by two negative terms. We relax this condition by allowing a positive term to appear in the upper bound of the time derivative, as shown in the result below.
\begin{Theorem}\label{Th: FxTS new}
Let $V:\mathbb R^n\rightarrow \mathbb R$ be a continuously differentiable, positive definite, proper function, satisfying
\begin{align}\label{eq: dot V new ineq}
\inf_{u\in \mathcal U}\{L_fV+L_gVu\} \leq -\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V,
\end{align}
with $\alpha_1, \alpha_2>0$, $\delta_1\in \mathbb R$, $\gamma_1 = 1+\frac{1}{\mu}$, $\gamma_2 = 1-\frac{1}{\mu}$ for some $\mu>1$, along the trajectories of \eqref{cont aff sys}. Then, there exists $u\in \mathcal U$, and a neighborhood $D$ of the origin such that for all $x(0)\in D$, the closed-trajectories of \eqref{cont aff sys} reach the origin in a fixed time $T$, where{\small
\begin{align}
D & = \begin{cases}\left\{x\; |\; V(x)\leq k^\mu\left(\frac{\delta_1-\sqrt{\delta_1^2-4\alpha_1\alpha_2}}{2\alpha_1}\right)^\mu\right\}; & \delta_1\geq 2\sqrt{\alpha_1\alpha_2}, \\
\; \mathbb R^n; & \delta_1<2\sqrt{\alpha_1\alpha_2},
\end{cases},\\
T & \leq \begin{cases}\frac{\mu}{\alpha_1(b-a)}\left(\log\left(\frac{b-ka}{a(1-k)}\right)-\log\left(\frac{b}{a}\right)\right); & \delta_1\geq 2\sqrt{\alpha_1\alpha_2}, \\
\frac{\mu}{\alpha_1k_1}\left(\frac{\pi}{2}-\tan^{-1}k_2\right); & 0\leq \delta_1<2\sqrt{\alpha_1\alpha_2},\label{new FxTS T est}\\
\frac{\mu\pi}{2\sqrt{\alpha_1\alpha_2}};& \delta_1\leq 0,
\end{cases},
\end{align}}\normalsize
where $0<k<1$, $a< b$ are the solutions of $\gamma(s) = \alpha_1s^2-\delta_1z+\alpha_2 = 0$, $k_1 = \sqrt{\frac{4\alpha_1\alpha_2-\delta_1^2}{4\alpha_1^2}}$ and $k_2 = -\frac{\delta_1}{\sqrt{4\alpha_1\alpha_2-\delta_1^2}}$.
\end{Theorem}
\begin{proof}
{Note that for $\delta_1\leq 0$, one can recover the right-hand side of \eqref{eq: dot V FxTS old } from \eqref{eq: dot V new ineq}, and the result follows from Theorem \ref{FxTS TH}. So, in rest of the proof we focus on the case when $\delta_1>0$.} First we show that there exists $D\subseteq \mathbb R^n$ containing the origin, such that for all $x\in D\setminus\{0\}$, $\dot V<0$. Consider the right-hand side of \eqref{eq: dot V new ineq}. For $\dot V<0$, it is needed that
\begin{align*}
\delta_1V< \alpha_1V^{\gamma_1}+\alpha_2V^{\gamma_2} \iff & \delta_1<\alpha_1V^{a_1-1}+\alpha_2V^{\gamma_2-1}\\
\iff & \delta_1<\alpha_1V^\frac{1}{\mu}+\alpha_2V^{-\frac{1}{\mu}}\\
\iff & \delta_1<\min_{V\geq 0}\alpha_1V^\frac{1}{\mu}+\alpha_2V^{-\frac{1}{\mu}}.
\end{align*}
Substitute $k = V^\frac{1}{\mu}$ to obtain $\alpha_1V^\frac{1}{\mu}+\alpha_2V^{-\frac{1}{\mu}} = \alpha_1k+\frac{\alpha_2}{k}$. The function $p:\mathbb R_+\rightarrow \mathbb R$, defined as $p(k) = \alpha_1k+\frac{\alpha_2}{k}$ is a strictly convex function, since $\frac{d^2p}{dk^2} = \frac{2\alpha_2}{k^3}>0$ for all $k>0$. Hence, the function $p$ has a unique minimizer. The derivative of $p$ reads $\frac{dp}{dk} = \alpha_1-\frac{\alpha_2}{k^2}$, which has a unique root in $\mathbb R_+$ at $k = \sqrt{\frac{\alpha_2}{\alpha_1}}$. Thus the minimum is attained for $k=\left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{\mu}{2}}$. Define $V^* = \left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{\mu}{2}}$, and $\delta = \alpha_1(V^*)^\frac{1}{\mu}+\alpha_2(V^*)^{-\frac{1}{\mu}} = 2\sqrt{\alpha_1\alpha_2}$. Thus, for $\delta_1<\delta = 2\sqrt{\alpha_1\alpha_2}$, we have that the right-hand side of \eqref{eq: dot V new ineq} is negative for all $V>0$, and hence, $\dot V<0$ for all $x\in \mathbb R^n\setminus\{0\}$.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\columnwidth,clip]{V_c3_cond2.png}
\caption{Qualitative variation of $h(V)= \alpha_1V^\frac{1}{\mu}+\alpha_2V^{-\frac{1}{\mu}}$ with $V$, for $\mu>1$. The function $h(V)$ achieves its minimum at $V = V^*$, marked by orange dashed line.}\label{fig:dot V RHS}
\end{figure}
{Now, for the case when $\delta_1\geq 2\sqrt{\alpha_1\alpha_2}$, we have that there exist $V_1\leq V^*\leq V_2$ such that $\delta_1 = \alpha_1V^{\frac{1}{\mu}}+\alpha_2 V^{-\frac{1}{\mu}}$ for both $V = V_1$ and $V = V_2$ (see Figure \ref{fig:dot V RHS}). Note that $V_1$ and $V_2$ are also solutions of $-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V = 0$, given as $V_1 \triangleq \frac{\delta_1-\sqrt{\delta_1^2-4\alpha_1\alpha_2}}{2\alpha_1}$ and $V_2\triangleq \frac{\delta_1+\sqrt{\delta_1^2-4\alpha_1\alpha_2}}{2\alpha_1}$. It can be easily verified that if $\delta_1\geq 2\sqrt{\alpha_1\alpha_2}$, then for all $V_1< V< V_2$, the expression $-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V$ evaluates to a positive value. Also, for all $V < V_1$, we have $\delta_1V<\alpha_1V^{\gamma_1}+\alpha_2V^{\gamma_2}$. Thus, for all $x\in D\setminus\{0\}$ with $D = \{x\; |\; V(x)\leq (kV_1)^\mu\}$ for some $0<k<1$, $\dot V<0$, and hence, $D$ is forward invariant.
Next, we show fixed-time convergence of the trajectories to the origin. Let $x(0)\in D$, so that $\dot V\leq 0$ for all $t\geq 0$. Thus, from \eqref{eq: dot V new ineq}, we obtain
\begin{align*}
& \frac{1}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}\frac{dV}{dt} \geq 1, \\
\implies & \int_{0}^{T}\frac{1}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}\frac{dV}{dt}dt = \int_{0}^Tdt,\\
\implies & \int_{V(x(0))}^{V(x(T))}\frac{dV}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}\geq \int_{0}^Tdt = T,\\
\implies & \int_{V_0}^{0}\frac{dV}{-\alpha_1V^{\gamma_1}-\alpha_2V^{\gamma_2}+\delta_1V}\geq T,
\end{align*}}\normalsize
where $V_0 = V(x(0))$. Denote the left-hand side of above inequality as $I$, so that we have $T\leq I$.
We consider the cases when $\delta_1<2\sqrt{\alpha_1\alpha_2}$ and $\delta_1\geq 2\sqrt{\alpha_1\alpha_2}$ separately.
First, let $\delta_1<2\sqrt{\alpha_1\alpha_2}$. Using Lemma \ref{lemma:int dot V}, we have
\begin{align}\label{eq: T bound case 1}
T\leq I \overset{\eqref{eq: I bound 1}}{\leq}\frac{\mu}{\alpha_1k_1}\left(\frac{\pi}{2}-\tan^{-1}k_2\right),
\end{align}
where $k_1 = \sqrt{\frac{4\alpha_1\alpha_2-\delta_1^2}{4\alpha_1^2}}$ and $k_2 = -\frac{\sqrt{\delta_1}}{\sqrt{4\alpha_1\alpha_2-\delta_1^2}}$.
Hence, if $\delta_1<2\sqrt{\alpha_1\alpha_2}$, we have that $\dot V<0$ for all $x(0)\neq 0$, and $V(x(t))= 0$ for all $t\geq T$, where $T\leq \frac{\mu}{\alpha_1k_1}\left(\frac{\pi}{2}-\tan^{-1}k_2\right)$. Since $V$ is proper, the origin is globally FxTS.
Now, for $\delta_1\geq 2\sqrt{\alpha_1\alpha_2}$, with $x(0)\in D = \{x\; |\; V(x)\leq (kV_1)^\mu\}$, we have that $\delta_1V_0<\alpha_1V_0^{\gamma_1}+\alpha_2V_0^{\gamma_2}$. Thus, using Lemma \ref{lemma:int dot V}, we have
\begin{align}\label{eq: T bound case 2}
T\leq I\overset{\eqref{eq: I bound 2}}{\leq} \frac{\mu}{\alpha_1(b-a)}\left(\log\left(\frac{b-ka}{a(1-k)}\right)-\log\left(\frac{b}{a}\right)\right),
\end{align}
where $a,b$ are the roots of $\gamma(z) \triangleq \alpha_1z^2-\delta_1z+\alpha_2 = 0$. {Finally, for the degenerate case of $\delta_1 = 2\sqrt{\alpha_1\alpha_2}$, we have from Lemma \ref{lemma:int dot V} that $T\leq I\leq \frac{\mu}{\sqrt{\alpha_1\alpha_2}}\Big(\frac{k}{1-k}\Big)$ for all $x\in D = \{x\; |\; V(x) \leq k^\mu\left(\frac{\alpha_2}{\alpha_1}\right)^\frac{\mu}{2}\}$. The above bound on $T$ for both the cases is independent of the initial condition $x(0)$. Thus, for all $x(0)\in D\setminus\{0\}$, the origin is FxTS.}
\end{proof}
\begin{Remark}\label{rem: new T bound}
For $\delta_1 = 0$, we have that $k_1 = \sqrt{\frac{\alpha_2}{\alpha_1}}$ and $k_2 = 0$ in \eqref{new FxTS T est}, and hence, we obtain
\begin{align}\label{eq: T bound new}
T\leq \frac{\mu\pi}{2\sqrt{\alpha_1\alpha_2}},
\end{align}
which is the same upper bound on time of convergence as given in \cite[Lemma 2]{parsegov2012nonlinear}. Also, it can be readily observed that $\frac{\mu\pi}{2\sqrt{\alpha_1\alpha_2}}<\frac{2\mu}{\sqrt{\alpha_1\alpha_2}}\leq \frac{\mu}{\alpha_1} + \frac{\mu}{\alpha_2}$, where the last inequality follows since $\alpha_1+\alpha_2\geq 2\sqrt{\alpha_1\alpha_2}$ for $\alpha_1,\alpha_2>0$. Hence, \eqref{eq: T bound new} gives a tighter upper-bound on the time of convergence as compared to \eqref{eq: T bound old}. Thus, the inequality in \eqref{T a1 a2 set S} can be replaced by
\begin{align}\label{eq: equiv a1 a2 T}
\frac{\mu\pi}{2\sqrt{\alpha_1\alpha_2}}\leq T_{ud},
\end{align}
for the case when $\gamma_1 = 1+\frac{1}{\mu}$ and $\gamma_2 = 1-\frac{1}{\mu}$ for some $\mu>1$. Equivalently, in QP \eqref{QP u P1}, the parameters $\alpha_1, \alpha_2$ can be defined as $\alpha_i = \frac{\mu\pi}{2T_{ud}} $
for $i = 1,2$, so that \eqref{eq: equiv a1 a2 T} holds.
\end{Remark}
\noindent In comparison to Theorem \ref{FxTS TH}, Theorem \ref{Th: FxTS new} allows a positive term $\delta_1V$ in the upper bound of the time derivative of the Lyapunov function. This property captures the robustness against a class of Lipschitz continuous, or vanishing, additive disturbance in the system dynamics, as shown in the following result.
\begin{Corollary}\label{cor: robust phi FxTS}
Consider the system
\begin{align}\label{eq:pert sys}
\dot x = f(x) + \psi(x),
\end{align}
where $f, \psi:\mathbb R^n\rightarrow \mathbb R^n$, $f(0) = 0$ and there exists $L>0$ such that for all $x\in \mathbb R^n$, $\|\psi(x)\|\leq L\|x\|$. Assume that the origin for the \textit{nominal} system $\dot x = f(x)$ is fixed-time stable, and that there exists a Lyapunov function $V$ satisfying conditions of Theorem \ref{FxTS TH}. Additionally, assume that there exists $k_1, k_2>0$ such that $V(x) \geq k_1\|x\|^2$ and $ \left\|\frac{\partial V}{\partial x}\right\|\leq k_2\|x\|$ for all $x\in \mathbb R^n$. Then, the origin of the perturbed system \eqref{eq:pert sys} is also FxTS.
\end{Corollary}
\begin{proof}
The time derivative of $V$ along the system trajectories of \eqref{eq:pert sys} reads
\begin{align*}
\dot V = \frac{\partial V}{\partial x}f(x) + \frac{\partial V}{\partial x}\psi(x)\leq &-aV^p-bV^q + k_2L\|x\|^2\\
\leq & -aV^p-bV^q + \frac{k_2L}{k_1}V.
\end{align*}
Hence, using Theorem \ref{Th: FxTS new}, we obtain the origin of \eqref{eq:pert sys} is FxTS { for all $x(0)\in D$, where $D$ is a neighborhood of the origin. As per the conditions of Theorem \ref{Th: FxTS new}, $D\subset \mathbb R^n$ or $D = \mathbb R^n$, depending upon the parameters $a, b, k_1, k_2$ and $L$.}
\end{proof}
Thus, this new result can also be used to guarantee FxTS of the origin in the presence of a class of vanishing disturbances.
\subsection{QP based formulation}
We are now ready to present a QP based method to find a control input that solves Problem \ref{P reach S}. { Define $z = \begin{bmatrix}v^T & \delta_1 & \delta_2\end{bmatrix}^T\in \mathbb R^{m+2}$, and consider the following optimization problem:
\begin{subequations}\label{QP gen}
\begin{align}
\min_{v, \delta_1, \delta_2} \; \frac{1}{2}z^THz & + F^Tz\\
\textrm{s.t.} \quad \quad A_uv \leq & b_u, \label{C1 cont const}\\
L_fh_g(x) + L_gh_g(x)v \leq & \delta_1h_g(x)-\alpha_1\max\{0,h_g(x)\}^{\gamma_1} \nonumber\\
& -\alpha_2\max\{0,h_g(x)\}^{\gamma_2} \label{C2 stab const}\\
L_fh_s(x) + L_gh_s(x)v \leq &-\delta_2h_s(x),\label{C3 safe const}
\end{align}
\end{subequations}
where $H = \textrm{diag}\{w_{u_1},\ldots, w_{u_m}, w_1, w_2\}$ is a diagonal matrix consisting of positive weights $w_{u_i}, w_i>0$, $F = \begin{bmatrix}\mathbf 0_m^T & q_1 & 0\end{bmatrix}$ with $q_1>0$ and $\mathbf 0_m\in \mathbb R^m$ a column vector consisting of zeros. The parameters $\alpha_1, \alpha_2, \gamma_1, \gamma_2$ are fixed, and are chosen as $\alpha_1 = \alpha_2 = \frac{\mu\pi}{2T_{ud}}$, $\gamma_1 = 1+\frac{1}{\mu}$ and $\gamma_2 = 1-\frac{1}{\mu}$ with $\mu>1$. The choice of these parameters does not affect the feasibility of the QP, as discussed below. In principle, any value of $\mu$ can be chosen as long as it is greater than 1. The linear term $F^Tz = q_1\delta_1$ in the objective function of \eqref{QP gen} penalizes the positive values of $\delta_1$ (see Theorem \ref{Th: d1 d2 P1 solve} for details on why $\delta_1$ being non-positive could be useful).} Constraint \eqref{C1 cont const} guarantees that the control input satisfies the control input constraints. Per Theorem \ref{Th: FxTS new}, the constraint \eqref{C2 stab const} guarantees convergence and the constraint \eqref{C3 safe const} ensures safety.
Parameters $\delta_1, \delta_2$ allow the upper bounds of the time derivatives of $h_s$ and $h_g$, respectively, to have a positive term when $h_s<0$ and $h_g>0$. This ensures the feasibility of the QP \eqref{QP gen} for all $x$, as demonstrated below.
\begin{Lemma}\label{lemma: QP feasibility}
There exists $(v, \delta_1, \delta_2)$ satisfying \eqref{C1 cont const}-\eqref{C3 safe const} for all $x\notin S_g$, i.e., the QP \eqref{QP gen} is feasible for all $x\notin S_g$.
\end{Lemma}
\begin{proof}
Since $x\notin S_g$, we have that $h_g(x)>0$. We consider the following two cases separately: $h_s(x) = 0$ and $h_s(x)<0$.
First, we consider the case when $h_s(x)<0$, i.e., $x\in \textrm{int}(S_s)$. Now, since $\mathcal U$ is non-empty, there exists $v = \bar v$ in $\mathcal U$ such that $\eqref{C1 cont const}$ is satisfied. Choose $\bar \delta_2 = \frac{L_fh_s(x) + L_gh_s(x)\bar v}{-h_s(x)}$, so that \eqref{C3 safe const} is satisfied with equality.
Finally, for $x\notin S_g$, we have $h_g(x)>0$. Define $\bar \delta_1 = \frac{L_fh_g(x) + L_gh_g(x)\bar v+ \alpha_1h_g(x)^{\gamma_1}+ \alpha_2h_g(x)^{\gamma_2}}{h_g(x)}$, so that \eqref{C2 stab const} holds with equality. Thus, for the case when $h_s(x)<0$, there exists $ (\bar v, \bar \delta_1, \bar \delta_2)$ such that \eqref{C1 cont const}-\eqref{C3 safe const} holds.
Next, consider the case when $h_s(x) = 0$, i.e., $x\in \partial S_s$. Per Assumption \ref{Assum feas}, we have that there exists $v = \tilde v\in \mathcal U$ such that \eqref{C3 safe const} holds. Since $h_s(x) = 0$, any value of $\delta_2$ is feasible, and hence, one can choose $\delta_2 = 0$. Hence, the choice of $(v, \delta_1, \delta_2) = (\tilde v, \bar \delta_1, 0)$ satisfies \eqref{C1 cont const}-\eqref{C3 safe const}. Thus, QP \eqref{QP gen} is always feasible.
\end{proof}
\begin{Remark}
In comparison to \cite{ames2017control}, we do not impose conditions such as $\frac{\partial h_s}{\partial x}\neq 0$ or $L_gh_s \neq 0$ in order to guarantee feasibility of the QP. { Instead, we use the slack terms corresponding to $\delta_1, \delta_2$ to guarantee that the QP is feasible.} Furthermore, in contrast to the prior work, where the feasibility is guaranteed only with the safety constraint \eqref{C3 safe const} and convergence constraint similar to \eqref{C2 stab const}, our formulation is guaranteed to have a feasible solution even with the control input constraints \eqref{C1 cont const}. Furthermore, in contrast to exponential convergence of the resulting closed-loop trajectories in \cite{ames2017control} or finite-time convergence guarantees without any control input bound in \cite{srinivasan2018control}, our proposed formulation guarantees {fixed}-time convergence in the presence of control input constraints.
\end{Remark}
For guaranteeing forward invariance of the safe set $S_S$, we use Nagumo's theorem which requires uniqueness of the system solutions. Traditionally, Lipschitz continuity of the right-hand side of \eqref{cont aff sys} is utilized in order to guarantee existence and uniqueness of the solutions of \eqref{cont aff sys}, see e.g., \cite{ames2017control,xu2015robustness,lindemann2019control}.
When the right-hand side of \eqref{cont aff sys} is only continuous, existence and uniqueness of the solutions can be established using the results in \cite[Section 3.15-3.18]{agarwal1993uniqueness} (see Theorem \ref{Th: d1 d2 P1 solve}). Next, we show that the control input $u(x)$ as a solution of the QP \eqref{QP gen} is continuous in its arguments. Define $A:\mathbb R^n \rightarrow \mathbb R^{(2m+2)\times (m+2)}$ and $b:\mathbb R^n\rightarrow \mathbb R^{(2m+2)}$ as {\footnotesize
\begin{align*}
A(x)& \coloneqq \begin{bmatrix} A_u & {{\mathbf 0_{2m}}} & {{\mathbf 0_{2m}}}\\
L_gh_G(x) & -h_G(x) & 0\\
L_gh_S(x) & 0 & h_S(x)
\end{bmatrix} \\
b(x) & \coloneqq \begin{bmatrix}b_u \\ -L_fh_G(x)-\alpha_1\max\{0,h_G(x)\}^{\gamma_1}-\alpha_2\max\{0,h_G(x)\}^{\gamma_2} \\ -L_fh_S(x)\end{bmatrix}
\end{align*}}\normalsize
where $\mathbf 0_{2m}\in \mathbb R^{2m}$ is a column vector consisting of zeros. Also, define the functions $G_i(x,z) \coloneqq A_i(x)z-b_i(x)$ where $A_i\in \mathbb R^{1\times (m+2)}$ is the $i$-th row of the matrix $A$, and $b_i\in \mathbb R$ the $i$-th element of $b$, so that the constraints \eqref{C1 cont const}-\eqref{C3 safe const} can be written as $G_i(x,z)\leq 0$ for $i\in 1, 2, \ldots, 2m+2$. Let $z^*$ and $\lambda^*\in \mathbb R_+^{2m+2}$ denote the optimal solution of \eqref{QP gen}, and the corresponding optimal Lagrange multiplier, respectively. We make the following assumption to prove our main results.
\begin{Assumption}\label{assum hs hg scs}
The strict complementary slackness holds, i.e., for each $i = 1, 2, \ldots, 2m+2$, it holds that either $\lambda_i^* > 0$ or $G_i(x,z^*) < 0$ for all $x\in \textnormal{int}(S_S)\setminus S_G$.
\end{Assumption}
Complimentary slackness, i.e., $\lambda_i^*G_i(x,z^*) = 0$, for all $i = 1, \ldots, 2m+2$, is a both necessary and sufficient condition for optimality of the solution for QPs \cite[Chapter 5]{boyd2004convex}. Note that this condition permits that for some $i$, both $\lambda_i^* = 0$ and $G_i(x,z^*) = 0$. \textit{Strict} complimentary slackness rules out this possibility, and require that for each $i$, either $\lambda_i^*$ or $G_i(x,z^*)$ is non-zero.
{
\begin{Theorem}\label{Th: QP sol cont}
Under Assumption \ref{assum hs hg scs}, the solution $z^*(\cdot)$ of \eqref{QP gen} is continuous on $ \textnormal{int}(S_S)\setminus S_G$.
\end{Theorem}
\begin{proof}
The proof is based on \cite[Theorem 2.1]{robinson1974perturbed}. Denote by $I(x)$, the indices of rows of matrix $A(x)$ corresponding to the active constraints, i.e., $j\in I(x)$ implies $A_{j}(x)z^*(x) = b_{j}(x)$, where $A_j\in \mathbb R^{1\times m+2}$ is the $j$-th row of the matrix $A$ and $b_j\in \mathbb R$ the $j$-th element of $b$. Define matrix $A_{ac}$ and $b_{ac}$ by collecting $A_j(x)$, and of $b_j$, respectively, so that $A_{ac}(x)z^*(x) = b_{ac}(x)$. Since at most one of the input constraints $u_i\leq u_{max_i}$ or $u_{min_i}\leq u_i$ can be active at any given time, the matrix $A_{ac}(x)$ has $k$ rows from $\begin{bmatrix}A_u & \mathbf 0_{2m} & \mathbf 0_{2m}\end{bmatrix}$, where $k\leq m$, which are linearly independent. Furthermore, it has $p$ rows from $\begin{bmatrix}L_gh_G & -h_G & 0\\ L_gh_S & 0 & h_S\end{bmatrix}$, where $p\leq 2$. Since $h_G, h_S\neq 0$ for $x\in \textnormal{int}(S_S)\setminus S_G$, these $k+p$ rows are linearly independent. Thus, the matrix $A_{ac}$ is full row-rank, i.e., the gradients of the active constraints $\{A_{ac_i}(x)\}$, where $A_{ac_i}(x)$ is the $i-$th row of matrix $A_{ac}(x)$, are linearly independent.
The second derivative $\nabla_{zz}L$ of the Lagrangian $L(z,x,\lambda) \coloneqq \frac{1}{2}z^THz+F^Tz + \lambda^T(A(x)z-b(x))$ with respect to $z$ is $H$, which is a positive definite matrix. Using this, and the fact that the QP \eqref{QP gen} is feasible, we have that the second-order sufficient conditions for optimality hold (see e.g. \cite[Section 2.3]{robinson1974perturbed}). Note that \cite[Theorem 2.1]{robinson1974perturbed} requires that the objective function and the functions $G_i(x,u)$ have the second derivatives jointly continuous in $(x,u)$. Since the objective function $\frac{1}{2}z^THz+F^Tz$ is independent of $x$, and the constraint functions $G_i(x,u)$ are linear in $u$, the second derivative of these functions are independent of $x$, and thus, satisfy this condition trivially. Finally, the strict complementary slackness condition is satisfied per Assumption \ref{assum hs hg scs}. Thus, all the conditions of \cite[Theorem 2.1]{robinson1974perturbed} are satisfied. Therefore, for every $x\in \textnormal{int}(S_S)\setminus S_G$, there exists an open neighborhood $\mathcal X\subset \textnormal{int}(S_S)\setminus S_G$ of $x$ such that the solution $z^*(x)$ is continuous for all $x\in \mathcal X$. Since this holds for all $x\in \textnormal{int}(S_S)\setminus S_G$, we have that the solution $z^*(x)$ is continuous for all $x\in \textnormal{int}(S_S)\setminus S_G$.
\end{proof}
Note that the above result guarantees that the control input defined as $u(x) = v^*(x)$ is continuous on $\textnormal{int}(S_S)\setminus S_G$. Theorem \ref{Th: delta a1 a2 Th 2} guarantees that the closed-loop trajectories remain bounded. Thus the solutions of the closed-loop system \eqref{cont aff sys} are well-defined, and are unique \cite[Chapter 3]{agarwal1993uniqueness}. Hence, Nagumo's theorem can be applied to guarantee safety, or forward invariance of the safe set $S_S$.
\begin{Remark}
Under Assumption \ref{assum hs hg scs}, the authors in \cite{fiacco1976sensitivity} show that the solution is continuously differentiable if the objective function and the constraints functions are twice continuously differentiable. The authors in \cite{ames2017control} assume that the functions $f,g$ and the Lie derivatives $L_fh_S, L_fh_G, L_gh_S, L_gh_G$ are locally Lipschitz continuous to show Lipschitz continuity of the solution of QP in the absence of control input constraints. Under similar assumptions, the authors in \cite{glotfelter2017nonsmooth} show that the solution of QP is guaranteed to be Lipschitz continuous (in the absence of input constraints) if the CBF constraints are inactive, i.e., the constraints are satisfied with strict inequality at the optimal solution $z^*$, which is same as Assumption \ref{assum hs hg scs}. The authors extend these results in \cite{glotfelter2018boolean} by utilizing the theory of non-smooth analysis, and strong forward invariance of sets even if the control input is not continuous. Note that we only need that the functions $f,g$ are continuous, and $h_S, h_G$ continuously differentiable in $x$.
\end{Remark}
}
\noindent Finally, we show that under some conditions, solution of \eqref{QP gen} solves Problem \ref{P reach S}.
\begin{Theorem}\label{Th: d1 d2 P1 solve}
Let Assumptions \ref{assum: sets Ss Sg}-\ref{assum hs hg scs} hold. If the solution of \eqref{QP gen}, given as $(v^*(\cdot), \delta_1^*(\cdot), \delta_2^*(\cdot))$, satisfies $\delta_1^*(x) \leq 0, \ \ \forall \;x\in S_s$, then, $u(\cdot) = v^*(\cdot)$ solves Problem \ref{P reach S} for all $x(0)\in S_s$.
\end{Theorem}
\begin{proof}
First we show convergence of the closed-loop trajectories $x(t)$ to the set $S_G$ within the user-defined time $T_{ud}$. Theorem \ref{Th: QP sol cont} guarantees that the solution of the QP \eqref{QP gen} is continuous, which implies continuity of the closed-loop system dynamics \eqref{cont aff sys} when $u(x) \coloneqq v^*(x)$. Using \cite[Theorem 3.15.1]{agarwal1993uniqueness} and choosing a Lyapunov candidate $\mathfrak{v} = \frac{1}{2}|y|^2$, it can be shown that $y\equiv 0$ is the unique solution of $\dot y = \phi(y) \coloneqq -\alpha_1|y|^{\gamma_1}-\alpha_2 |y|^{\gamma_2}$ for $y(0) = 0$. Note that $h_G(x) = 0$ for $x\in \partial S_G$ and $h_G(x)>0$ for $x\notin S_G$, i.e., the function $h_G$ is positive definite with respect to the set $S_G$. Thus, using \cite[Theorem 3.18.1]{agarwal1993uniqueness} (with $g$ in \cite[Theorem 3.18.1]{agarwal1993uniqueness} defined as $\phi$) and the fact that $\dot h_G(x)\leq \phi(h_G(x))$ for $x\notin S_G$ (since $\delta_1^*(x)\leq 0$ in \eqref{C2 stab const}), the solution of the closed-loop system \eqref{cont aff sys} exists and is unique. Since $\delta_1^*(x)\leq 0$, per Theorem \ref{Th: FxTS new}, we have that the closed-loop trajectories of \eqref{cont aff sys} with $u(x) \coloneqq v^*(x)$ reach the set $S_G$ within fixed time $T\leq \frac{\mu\pi}{2(\alpha_1 \alpha_2)^{\frac{1}{2}}} = T_{ud}$, i.e., within the user-defined time $T_{ud}$ for all $x(0)\in S_S$.
Next, we show that the closed-loop trajectories of \eqref{cont aff sys} satisfy $x(t)\in S_S$ for all $t\geq 0$ under $u(x) \coloneqq v^*(x)$. From \eqref{C3 safe const}, we have $\dot h_S(x) = L_fh_S(x) + L_gh_S(x)u \leq 0$, for all $x\in \partial S_S$. {Since the closed-loop solution of \eqref{cont aff sys} with $u(x) \coloneqq v^*(x)$ is unique, Nagumo's theorem can be applied to guarantee forward invariance of the safe set $S_S$, i.e., that $x(t)\in S_S$ for all $t\geq 0$. Therefore, the control input $u(x) \coloneqq v^*(x)$ solves Problem \ref{P reach S} for all $x(0)\in S_S$}.
\end{proof}
\noindent It is easy to see that per Theorems \ref{FxTS reach set S} and \ref{Th: d1 d2 P1 solve}, if $h_g(x)$ is a FxT CLF-$S_g$, then there exists $u\in \mathcal U$ such that the closed-loop trajectories of \eqref{cont aff sys} reach the set $S_g$ within fixed time $T$. On the basis of this, the following result can be presented.
\begin{Corollary}\label{cor: FxT CLF result}
Let Assumptions \ref{assum: sets Ss Sg}-\ref{assum hs hg scs} hold. If $h_G$ is an FxT CLF-$S_G$ for \eqref{cont aff sys} with parameters $\alpha_1, \alpha_2, \gamma_1, \gamma_2$ as defined in \eqref{QP gen}, then the QP \eqref{QP gen} with $\delta_1 = 0$ is feasible {for all $x\in \textnormal{int}(S_S)\setminus S_G$}, i.e., the solution $(v^*(x),\delta_2^*(x))$ of the QP \eqref{QP gen} after fixing $\delta_1 = 0$ exists for all $x\in \textnormal{int}(S_S)\setminus S_G$. Furthermore, if the solution $(v^*(x),\delta_2^*(x))$ is continuous for all $x\in \textnormal{int}(S_S)\setminus S_G$ with $\mathfrak{d}_2\coloneqq \sup\limits_{\textnormal{int}(S_S)\setminus S_G}|\delta_2^*(x)|<\infty$,
then the control input $u(x) \coloneqq v^*(x)$ solves Problem \ref{P reach S} for all $x(0)\in \textnormal{int}(S_S)$.
\end{Corollary}
\begin{proof}
Since $h_G$ is FxT CLF-$S_G$, there exists $v(x)\in \mathcal U$ such that \eqref{C2 stab const} is satisfied with $\delta_1 = 0$. Given that $x\in \textnormal{int}(S_S)$, i.e., $h_S(x)<0$, one can let $\delta_2(x) =\frac{L_fh_S(x) + L_gh_S(x) v(x)}{-h_S(x)}$ so that \eqref{C3 safe const} is satisfied. Thus, the QP \eqref{QP gen} is feasible with $\delta_1 = 0$ for all $x\in \textnormal{int}(S_S)$.
Next, we show that the closed-loop trajectories remain in the interior of the safe set, i.e., $x(t)\in \textnormal{int}(S_S)$ for all $0\leq t\leq T_{ud}$. For $\mathfrak{d}_2<\infty$, and for all $t\leq T_{ud}$, we have that \eqref{C3 safe const} yields:
\begin{align*}
\dot h_S(x(t)) \leq -\mathfrak{d}_2h_S(x(t)) = \psi(-h_S(x(t))),
\end{align*}
where $\psi(r) \coloneqq \mathfrak{d}_2 \ r$ is a Lipschitz continuous class-$\mathcal K$ function. Thus, using the Comparison Lemma (\cite[Lemma 3.4]{khalil2002nonlinear}) and \cite[Lemma 4.4]{khalil2002nonlinear}, we obtain that there exists a class-$\mathcal{KL}$ function $\kappa$ such that $h_S(x(t))\leq -\kappa\left(-h_S(x(0)),t\right)$. To see this, consider $\dot y = \psi(-y)$ with $y(0) = h_S(x(0))$, and define $z \coloneqq -y$ to obtain $\dot z = -\psi(z)$. Then, using \cite[Lemma 4.4]{khalil2002nonlinear}, we obtain that $z(t) = \kappa(z(0),t)$, or equivalently, $y(t) = -\kappa(-y(0),t)$. Furthermore, using the Comparison Lemma, we obtain that $h_S(x(t))\leq y(t)$, i.e., $h_S(x(t))\leq -\kappa(-h_S(x(0)),t)$. Since $\kappa\left(-h_S(x(0)),t\right)>0$ for all $t\leq T_{ud}<\infty$, it follows that $h_S(x(t))\leq -\kappa\left(-h_S(x(0)),t\right)<0$ for all $t\leq T_{ud}$, i.e., $x(t)\in \textnormal{int}(S_S)$ for all $t\leq T_{ud}$. Thus, under the assumption on continuity of the control input $u$, the same arguments as in Theorem \ref{Th: d1 d2 P1 solve} can be used to show existence and uniqueness of the solution $x(t)$ of the closed-loop system \eqref{cont aff sys} for $0\leq t\leq T_{ud}$. Thus, the control input $u(x) \coloneqq v^*(x)$ renders the set $\textnormal{int}(S_S)$ forward invariant. Finally, under the assumption that $\delta_1 = 0$, Theorem \ref{Th: d1 d2 P1 solve} ensures the convergence of the closed-loop trajectories $x(t)$ to the set $S_G$ within a fixed time $T_{ud}$; therefore, we have that $u(x) \coloneqq v^*(x)$ solves Problem \ref{P reach S} for all $x(0)\in \textnormal{int}(S_S)$.
\end{proof}
{
\begin{Remark}
Corollary \ref{cor: FxT CLF result} gives one sufficient condition under which the solution of QP \eqref{QP gen} solves Problem \ref{P reach S}. As pointed out in \cite{ames2017control}, the conflict between safety and the convergence constraint require a non-zero slack term for satisfaction of \eqref{C2 stab const}-\eqref{C3 safe const} together (see also Remark \ref{remark: d1 T U}). With this observation, one can readily conclude that if the control input constraints $U= \mathbb R^m$, or is \textit{sufficiently} large, and the goal set is in the interior of safe set, i.e., $S_g\subset\textrm{int}(S_s)$, then also, a sufficiently large value of $v$ can result into satisfaction of \eqref{C2 stab const} with $\delta_1\leq 0$. The authors in \cite{xu2015robustness} argue that in the absence of such a conflict, a larger weight $w_1$ on $\delta_1$ in \eqref{QP gen} results in solution of the QP with $\delta_1 \approx 0$.
\end{Remark}
}
Next, we list out some cases when the solution of QP \eqref{QP gen} might not solve Problem \ref{P reach S} with the specified time constraint, and from all initial condition, but it still renders the closed-loop trajectories safe, and converge to the set $S_g$ within some fixed time.
\begin{Theorem}\label{Th: delta a1 a2 Th 2}
Let Assumptions \ref{assum: sets Ss Sg}-\ref{assum hs hg scs} hold. If the solution of \eqref{QP gen}, given as $(v^*(\cdot), \delta_1^*(\cdot), \delta_2^*(\cdot))$, satisfies
\begin{align}\label{eq: delta1 a1 a2 cond 2}
\delta_1^*(x)& < 2\sqrt{\alpha_1\alpha_2}, \ \ \forall \; x\in S_s,
\end{align}
then, with $u(\cdot) = v^*(\cdot)$, the closed-loop trajectories reach the set $S_g$ within a fixed time $T_F$ for all $x(0)\in S_s$, while satisfying safety requirement, i.e., $x(t)\in S_s$ for all $t\geq 0$. If \eqref{eq: delta1 a1 a2 cond 2} does not hold, then, the result holds for all $x(0)\in D$ for some $D\subset S_s$.
\end{Theorem}
\begin{proof}
In both cases, following the proof of Theorem \ref{Th: d1 d2 P1 solve}, we have that the closed-loop trajectories satisfy $x(t)\in S_S$ for all $t\geq 0$. When \eqref{eq: delta1 a1 a2 cond 2} holds, using Theorem \ref{Th: FxTS new}, we have that there exists $T_1\leq \frac{\mu}{\alpha_1\mathrm{k}_1}\left(\frac{\pi}{2}-\tan^{-1}\mathrm{k}_2\right)$, where $\mathrm{k}_1 \coloneqq \sqrt{\frac{4\alpha_1\alpha_2-\mathfrak{d}_1}{4\alpha_1^2}}$ and $\mathrm{k}_2 \coloneqq -\frac{\mathfrak{d}_1}{\sqrt{4\alpha_1\alpha_2-\mathfrak{d}_1}}$, such that the closed-loop trajectories $x(t)$ reach the set $S_G$ within fixed time $T_1$. Also, per \eqref{eq: delta1 a1 a2 cond 2}, we have that $\mathrm{k}_1>0$ and so, $T_1<\infty$, i.e., the closed-loop trajectories $x(t)$ of \eqref{cont aff sys} under $u(x)\coloneqq v^*(x)$ reach the set $S_G$ in fixed time $T_1$.
For the case when \eqref{eq: delta1 a1 a2 cond 2} does not hold, using Theorem \ref{Th: FxTS new}, we have that there exists $D \coloneqq \{x\; | \; h_G(x)\leq k^\mu\left(\frac{\mathfrak{d}_1-\sqrt{\mathfrak{d}_1-4\alpha_1\alpha_2}}{2\alpha_1}\right)^\mu\}\cap S_S$ (which is a subset of $S_S$) and a fixed time $T_2\leq \frac{\mu}{\alpha_1(b-a)}\left(\log\left(\frac{b-ka}{a(1-k)}\right)-\log\left(\frac{b}{a}\right)\right)$, where $a, b$ are defined as in Theorem \ref{Th: FxTS new}, and $0<k<1$, such that for all $x(0)\in D$, the closed-loop trajectories reach the set $S_G$ within time $T_2$. It is easy to show that $T_2<\infty$. Hence, the closed-loop trajectories reach the set $S_G$ within a fixed time, while maintaining safety at all times.
\end{proof}
In brief, the solution of the QP \eqref{QP gen} always exists, is a continuous function of $x$, and renders the safe set $S_S$ forward invariant, i.e., guarantees safety. Furthermore, the control input is guaranteed to lead fixed-time convergence of the closed-loop trajectories to the goal set $S_G$. In the case when $\delta_1\leq 0$, the convergence is guaranteed for all $x(0)\in S_S$, and within the user-defined {fixed} time $T_{ud}$. If $\delta_1$ satisfies \eqref{eq: delta1 a1 a2 cond 2}, then also, fixed-time convergence is guaranteed for all $x(0)\in S_S$, but the time of convergence $T$ may exceed the time $T_{ud}$. Finally, if \eqref{eq: delta1 a1 a2 cond 2} does not hold, fixed-time convergence is guaranteed for all $x(0)\in D\subset S_S$, however, the time of convergence $T$ may exceed the time $T_{ud}$. Before presenting the numerical case studies, we note our final remarks.
{
\begin{Remark}\label{remark: d1 T U}
In the presence of control input bounds, it is not possible to guarantee FxTS for an arbitrary large domain of attraction. The slack term corresponding to $\delta_1$ in QP \eqref{QP gen} characterizes the trade-off between the domain of attraction and time of convergence for given control input bounds.
Intuitively, for a given control input constraint set, a larger value of $T_{ud}$ results into smaller values of $\alpha_1, \alpha_2$, which would result into satisfaction of \eqref{C2 stab const} with smaller value of $\delta_1$. Conversely, for a given $T_{ud}$ (and thus, for a given pair $\alpha_1, \alpha_2$), a larger control authority would have a similar result.
\end{Remark}
}
\begin{Remark}
Recently, the authors in \cite{cortez2019control} address the issue of sampling effects in the implementation of QP-based controllers. They consider safety constraints for discretized continuous-time systems in the presence of additive disturbances. More specifically, they discuss how to modify the CBF in the QP to guarantee forward invariance of a set
so that satisfaction of the safety constraints at the discrete time steps via robust CBF results into safety at all times. In the numerical case studies presented below, we use Euler discretization to implement the proposed method. Although a discrete implementation of the presented method without accounting for discretization errors might lead to safety violation, one can adopt a robust CLF method as discussed in \cite{cortez2019control} to guarantee that safety is preserved. A thorough analysis of safety and fixed-time convergence for discrete-time approximation of QP-based methods is out of the scope of the current paper.
\end{Remark}
\section{Numerical Case Studies}\label{sec: simulation}
We present two case studies to illustrate the efficacy of our proposed method. In the first example, we follow \cite[Section V.A]{ames2017control} to formulate the adaptive cruise control problem, where the objective is to achieve a desired speed while maintaining a safe distance from the lead vehicle (i.e., the vehicle in the front). We use QP \eqref{QP gen} to find the controller that solves the problem, and compare our results with those in \cite{ames2017control}. In the second example, we consider multi-agent motion planning problem under spatiotemporal specifications, where the objective for the agents are to visit some regions in a given time sequence, while remaining in a safe set at all times, and maintaining safe distance from each other. For both the case studies, we use Euler discretization to discretize the continuous-time dynamics, and the \textsc{Matlab} function $\texttt{quadprog}$ to solve the QP at each discrete time step.
\subsection{Adaptive Cruise Control Problem}
In this example, we consider an adaptive cruise control (ACC) problem with a following and a lead vehicle, where the objective is for the following vehicle to achieve a desired fixed speed (soft constraint). In addition, the following vehicle need to maintain a safe distance from the lead vehicle (hard constraint). Hence, when the distance of the vehicles decreases, the following vehicle reduces the speed in order to maintain safety. Considering that the two vehicles are modeled as point masses and travelling in a straight line, the system dynamics can be written as\cite{ames2017control}
\begin{align}
\dot x = f(x) + gu,
\end{align}
with
\begin{align}
f(x) = \begin{bmatrix}
-F_r(x)/M\\
a_L\\
x_2-x_1
\end{bmatrix}, \quad g=
\begin{bmatrix}
1/M \\
0 \\
0
\end{bmatrix}, \label{Ex1Dyn}
\end{align}
where $u \in (-u_{\max}, u_{\max})$ is the control input, $x = [x_1, x_2, x_3]^\mathrm{T_{ud}} = [v_f, v_l, D]^\mathrm{T_{ud}} \in {\mathbb R}^3$ is the system state with $v_f$ being the velocity of the following vehicle, $v_l$ being the velocity of the lead vehicle, and $D$ being the distance between the two vehicles.
In \eqref{Ex1Dyn}, $M$ is the mass of the following vehicle, $F_r(x) = f_0 + f_1 v_f + f_2 v_f^2$ is the drag force, and $a_L \in (-a_l \mathfrak{g}, a_l \mathfrak{g})$ is the acceleration of the lead vehicle, with $a_l$ being the fraction of the gravitational acceleration $\mathfrak{g}$.
We now define the goal and the safe sets respectively as $h_g(x) = (v_f - v_d)^2, h_s(x) = \tau_d v_f - D,$ where $v_d$ is a desired fixed velocity and $\tau_d$ is the desired time headway. We set the maximum available control effort to $u_{\max}=0.25 M\mathfrak{g}$ with $\mathfrak{g}=9.81$ $\mathrm{m/s^2}$ and $M=1650$ $\mathrm{Kg}$, the desired velocity to $v_d=22 $ $\mathrm{m/s}$, the initial velocity of the lead vehicle to $v_l(0)=10$ $\mathrm{m/s}$, initial distance to $D(0) = 150$ $\mathrm{m}$, $f_0=0.1$ $\mathrm{N}$, $f_1=5$ $\mathrm{Ns/m}$, $f_2=0.25$ $\mathrm{Ns^2/m^2}$, and $a_l = 0.3$.
We implement the QP in \eqref{QP gen} with $T_{ud}=10$ sec, and $\mu=5$ resulting in $\gamma_1=1.2$, $\gamma_2=0.8$.
\begin{figure}[!ht]
\centering
\includegraphics[ width=0.9\columnwidth,clip]{Ex1_Is_x.eps}
\caption{Tracking Performance and the safe set $h_S(x)$ for various initial follower velocities $v_f(0) \in [17,27]$ $\mathrm{m/s}$ with $T=10$ sec.}\label{ex1_1}
\end{figure}
Figures \ref{ex1_1} and \ref{ex1_2} illustrate the tracking performance of the resulting controller, where the solid lines represent the velocity of the following vehicle for different initial velocity of the following vehicle $v_f(0) \in [17, 27]$ $\mathrm{m/s}$. One can see from these figures that the desired speed is achieved when the trajectories are away from the boundaries of the safe set. As expected, close to the boundaries of the safe set $h_s(x)$, the speed of the following vehicle is reduced to maintain safety.
\begin{figure}[!t]
\centering
\includegraphics[ width=0.9\columnwidth,clip]{Ex1_Is_u.eps}
\caption{Control input for the closed-loop trajectories for various $v_f(0)$.}\label{ex1_2}
\end{figure}
As stated before, there is no guarantee for the existence of the solution of the proposed QP in \cite{ames2017control} when there is a control input constraint. For the specific problem of adaptive cruise control as in this example, the authors introduced two control barrier functions, namely optimal and conservative CBFs, based on the simplified system dynamics with no drag effect $F_r(x)$ to ensure feasibility of the solution. However, due to conservatism, the newly constructed safe sets $h_F^o(x)$ and $h_F^c(x)$ for the optimal and conservative CBFs are violated initially for large initial velocity of the following vehicle, while the actual safe set $h_s(x) = \tau_d v_f - D$ is not violated and the problem can be still feasible.
Figures \ref{ex1_0} and \ref{ex1_00} compare the tracking performance of the proposed approach and the results with optimal and conservative CBFs with $v_f(0) = 18$ $\mathrm{m/s}$.
Since we are solving the QP directly and without the aforementioned conservatism, one can see from Figure \ref{ex1_0} that our proposed control approach tracks the desired goal speed of $22$ $\mathrm{m/s}$ for a longer duration before departure from this speed for maintaining safety. Finally, Figure \ref{ex1_3} compares the control effort between the the proposed approach and the results in \cite{ames2017control}, where the proposed approach is using more available control authority. This is due to the fact that the desired goal speed is tracked for a longer duration in the proposed approach, and hence more control action is used to keep the system trajectories in the safe set as the trade-off.
\begin{figure}[!t] \vspace{-.00cm}
\centering
\includegraphics[ width=0.9\columnwidth,trim=.65cm .1cm .9cm .2cm,clip]{Compare1.jpg}
\caption{Tracking performance comparison of the proposed approach and the results in \cite{ames2017control}.}\label{ex1_0} \vspace{-.0cm}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[ width=0.9\columnwidth,trim=.6cm .1cm .9cm .2cm,clip]{Compare2.jpg}
\caption{Control inputs for Figure \ref{ex1_0}.}\label{ex1_00}
\end{figure}
\begin{figure}[!t] \vspace{-.4cm}
\centering
\includegraphics[ width=0.9\columnwidth,trim=.4cm 0 .9cm 0,clip]{Energy.eps}
\caption{Energy comparison of the proposed approach and the results in \cite{ames2017control}.}\label{ex1_3}
\end{figure}
We now examine the robustness of the proposed approach against external disturbances. To this end, we consider the system dynamics as
\begin{align}
\dot x = f(x) + gu + \phi(x),
\end{align}
where $f(x)$ and $g$ are given in \eqref{Ex1Dyn}. In addition, we consider the Lipschitz continuous disturbance $\phi(x) = \frac{d_\delta}{M} |v_f - v_d|$ where $d_\delta>0$ is a constant. In this scenario, we modify the proposed QP \eqref{QP gen} such that when the safe set $h_s(x)$ gets close to zero ($h_s(x) > -20$), the parameter $\delta_2$ is enforced to be zero.
Figures \ref{ex1_1_dist} and \ref{ex1_2_dist} present the tracking performance of the resulting controller in the presence of the external disturbance $\phi(x)$.
The safety is still achieved for all time and the following vehicle tracks the goal speed of $22$ m/s when the safe distance from the lead vehicle is possible.
\begin{figure}[!t]
\centering
\includegraphics[ width=0.9\columnwidth,trim=.55cm .0cm .9cm .2cm,clip]{Ex1_Disturb_x.eps}
\caption{Tracking Performance and the safe set $h_s(x)$ for $d_\delta\in [0,\; 100]$ (blue to red).}\label{ex1_1_dist} \vspace{-.15cm}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[ width=0.9\columnwidth,trim=.55cm 0 .9cm .2cm,clip]{Ex1_Disturb_u.eps}
\caption{Control input for Figure \ref{ex1_1_dist}.}\label{ex1_2_dist}
\end{figure}
\subsection{Two-agent motion planning with spatiotemporal specifications}\label{sec sim B}
In the second scenario, we present a multi-agent motion planning example under spatiotemporal constraints. We consider a two-agent scenario, where the robot dynamics are modeled as
\begin{align*}
\dot x_i = u_i,
\end{align*}
where $x_i,u_i\in \mathbb R^2$ for $i = 1,2$. The closed-loop trajectories for the respective agents, starting from $x_1(0)\in C_1$ and $x_2(0)\in C_3$, are required to satisfy the following STL specifications
\begin{align*}
(x_1,t) \models & G_{[0,T_4]}\phi_s\land F_{[0, T_1]}\phi_2\land F_{[T_1, T_2]}\phi_3\land F_{[T_2, T_3]}\phi_4\\
& \land F_{[T_3, T_4]}\phi_1,\\
(x_2,t) \models & G_{[0,T_4]}\phi_s\land F_{[0, T_1]}\phi_2\land F_{[T_1, T_2]}\phi_1\land F_{[T_2, T_3]}\phi_4\\
& \land F_{[T_3, T_4]}\phi_3
\end{align*}
which is explained in details below (see Figure \ref{scene:1}):
\begin{figure}[!t]
\centering
\includegraphics[ width=0.8\columnwidth,clip]{scenario_1.eps}
\caption{Problem setting for simulation example B.}\label{scene:1}
\end{figure}
\begin{itemize}
\item $x_1(t), x_2(t)\in S_s = \{x\; |\; \|x\|_1\leq 2, \|x\|_2\geq 1.5\}$ for all $t\geq 0$, i.e., the closed-loop trajectories of the two agents should stay inside the solid-blue square and outside the red-dotted circle, and maintain a minimum separation $d_m$ at all times;
\item On or before a given $T_1$ satisfying $0<T_1<\infty$, agent 1 and 2 should reach the square $C_2$;
\item On or before a given $T_2$ satisfying $T_1<T_2<\infty$, agent 1 should reach the square $C_3$ and agent 2 should reach the square $C_1$;
\item On or before a given $T_3$ satisfying $T_2<T_3<\infty$, agent 1 and 2 should reach the square $C_4$;
\item On or before a given $T_4$ satisfying $T_3<T_4<\infty$, agent 1 should reach the square $C_1$ and agent 2 should reach the square $C_3$;
\end{itemize}
\begin{figure*}[!ht]
\normalsize
\begin{subequations}\label{eq: robot STL req}
\begin{align}
(x_1,t) & \models G_{[0,T_4]} \bar \phi_s\land G_{[0, t_0]}\bar \phi_1\land F_{[0, t_0]}\bar \phi_2\land G_{(t_0,T_1]}\bar \phi_2\land F_{(t_0, T_1]}\bar \phi_3\land G_{(T_1,t_2]}\bar \phi_3\land F_{(T_1, t_2]}\bar \phi_4 \land G_{(t_2,T_2]}\bar \phi_4\land F_{(t_2, T_2]}\bar \phi_5\nonumber\\
& \land G_{(T_2,t_4]}\bar \phi_5\land F_{(T_2, t_4]}\bar \phi_6\land G_{(t_4,T_3]}\bar \phi_6\land F_{(t_4, T_3]}\bar \phi_7\land G_{(T_3,t_6]}\bar \phi_7\land F_{(T_3, t_6]}\bar \phi_8\land G_{(t_6,T_4]}\bar \phi_8\land F_{(t_6, T_4]}\bar \phi_1, \\ \nonumber\\
(x_2,t) & \models G_{[0,T_4]} \bar \phi_s\land F_{[0, t_0]}\bar \phi_5\land F_{[0, t_0]}\bar \phi_4\land G_{(t_0, T_1]}\bar \phi_4\land F_{(t_0, T_1]}\bar \phi_3\land G_{(T_1,t_2]}\bar \phi_3\land F_{(T_1, t_2]}\bar \phi_3\land G_{( t_2,T_2]}\bar \phi_3\land F_{(t_2, T_2]}\bar \phi_1 \nonumber\\
& \land G_{(T_2,t_4]}\bar \phi_1\land F_{(T_2, t_4]}\bar \phi_8\land G_{(t_4,T_3]}\bar \phi_8\land F_{(t_4, T_3]}\bar \phi_7\land G_{(T_3,t_6]}\bar \phi_7\land F_{(T_3, t_6]}\bar \phi_6\land G_{(t_6,T_4]}\bar \phi_6\land F_{(t_6, T_4]}\bar \phi_5.
\end{align}
\end{subequations}
\hrulefill
\end{figure*}
This problem is an extended version of the case study considered in \cite{lindemann2017robust,garg2019control}. Note that the sets $C_i$ are not overlapping with each other, and the corresponding functions $h_i(x)$ are not continuously differentiable.
Now, in order to be able to use QP-based formulation \eqref{QP gen}, we need to satisfy Assumption \ref{Assum feas}, i.e., find the sets $\bar S_i$ such that $\bar S_i\bigcap \bar S_{i+1}\neq\emptyset$. In order to solve this problem, we construct auxiliary sets $\bar S_i$, $i\in \{1, 2,\cdots, 8\}$ as shown in Figure \ref{scene:2}. The set $\bar S = \{x\; |\; \|x\|_2\leq 1.5\}$ and sets $\bar S_i$ are defined as follows:
\begin{figure}[!ht]
\centering
\includegraphics[ width=0.9\columnwidth,clip]{scenario_2.eps}
\caption{Construction of sets $\bar S, \bar S_1, \cdots, \bar S_8$ for simulation example 2.}\label{scene:2}
\end{figure}
\begin{itemize}
\item $\bar S_1 = \{x\; |\; \|(x-[-1.5 \; 1.5]^T)\|\leq 0.5\}$;
\item $\bar S_2 = \{x\; |\; \|(x-[0 \; 1.5]^T)\|_{P_1}\leq 1\}$;
\item $\bar S_3 = \{x\; |\; \|(x-[1.5 \; 1.5]^T)\|\leq 0.5\}$;
\item $\bar S_4 = \{x\; |\; \|(x-[1.5 \; 0]^T)\|_{P_2}\leq 1\}$;
\item $\bar S_5 = \{x\; |\; \|(x-[1.5 \; -1.5]^T)\|\leq 0.5\}$;
\item $\bar S_6 = \{x\; |\; \|(x-[0 \; -1.5]^T)\|_{P_1}\leq 1\}$;
\item $\bar S_7 = \{x\; |\; \|(x-[-1.5 \; -1.5]^T)\|\leq 0.5\}$;
\item $\bar S_8 = \{x\; |\; \|(x-[-1.5 \; 0]^T)\|_{P_2}\leq 1\}$;
\end{itemize}
where $\|z\|_{P_1} = \sqrt{\frac{z_1^2}{1.2^2}+\frac{z_2^2}{0.5^2}}$ and $\|z\|_{P_2} = \sqrt{\frac{z_1^2}{0.5^2}+\frac{z_2^2}{1.2^2}}$. Now, in order to visit square $C_2$, agent 1 can go to set $\bar S_2\setminus\bar S$ within some time $0< t_0<T_1$ and then to $\bar S_3$ before $t = T_1$. Hence, for agent 1, for time interval $[0, t_0]$, the safe set is defined by function $h_s = \max\{h_{s1}, h_{s2},h_{s3},h_{s4}\}$, where $h_{s1} = \|x\|_1-2, h_{s2} = 1.5-\|x\|, h_{s3} = d_m-\|x-x_2\|$ and $h_{s4} = \|x-[-1.5\; 1.5]^T\|-0.5$, and the goal set is defined by function $h_g = \|x-[0\; 1.5]^T\|_{P_1}-1$. For time interval $[t_0,T_1]$, the functions $h_{s4}$ and $h_g$ change, while other things remain same. With these new sets, the problem can be re-formulated for agent 1 to design a control input $u_1(t)$ such that for $x_1(0)\in \bar S_1$,
\begin{itemize}
\item For a given $t_0$ satisfying $0<t_0<T_1$, $x_1(t_0)\in \bar S_2\setminus \bar S$ and $x_2(t_0)\in \bar S_4\setminus \bar S$;
\item For a given $t_1$ satisfying $t_0<t_1\leq T_1$, $x_1(t_1)\in \bar S_3$ and $x_2(t_1)\in \bar S_3$;
\item For a given $t_2$ satisfying $T_1<t_2<T_2$, $x_1(t_2)\in \bar S_4\setminus \bar S$ and $x_2(t_2)\in \bar S_2\setminus \bar S$;
\item For a given $t_3$ satisfying $t_2<t_3\leq T_2$, $x_1(t_3)\in \bar S_5$ and $x_2(t_3)\in \bar S_1$;
\item For a given $t_4$ satisfying $T_2<t_4<T_3$, $x_1(t_4)\in \bar S_6\setminus \bar S$ and $x_2(t_4)\in \bar S_8\setminus \bar S$;
\item For a given $t_5$ satisfying $t_4<t_5\leq T_3$, $x_1(t_5)\in \bar S_7$ and $x_2(t_5)\in \bar S_7$;
\item For a given $t_6$ satisfying $T_3<t_6<T_4$, $x_1(t_6)\in \bar S_8\setminus \bar S$ and $x_2(t_6)\in \bar S_6\setminus \bar S$;
\item For a given $t_7$ satisfying $t_6<t_7\leq T_4$, $x_1(t_7)\in \bar S_1$ and $x_2(t_7)\in \bar S_5$,
\item For all $t\geq 0$, $x_1(t), x_2(t)\in S_s$ and $\|x_1(t)-x_2(t)\|\geq d_m$.
\end{itemize}
One can readily write the requirements for agent 2 in the similar manner. These requirements can be written in the form of STL formulae for the two agents as given in \eqref{eq: robot STL req}.
We can now use the formulation \eqref{QP gen} to compute the control input by solving \eqref{QP gen} sequentially, i.e., for agent 1, for $t\in [0, t_0)$, $S_g = \bar S_2$, then for $t\in [t_0, T_1)$, $S_g = \bar S_3$, and so on. We use $|u_i|\leq 10$ as the input constraints for $i = 1,2$. In order to translate the input constraint in the form of \eqref{C1 cont const}, we define $A_u = \begin{bmatrix}1& 0 \\ -1& 0\\ 0 & 1\\ 0 &-1\end{bmatrix}$ and $b_u = \begin{bmatrix}7&7&7&7\end{bmatrix}^T$, so that $u_{1x}, u_{1y}, u_{2x}, u_{2y}\in [-7, \; 7]$.
The time constraints are chosen as $T_i = 2$ for $i \in \{1,2,3,4\}$ and $t_j = 1$ for $j\in \{t_0, t_1,\cdots ,t_7\}$. We choose $\mu = 5$, so that $\gamma_1 = 1.2$ and $\gamma_2 = 0.8$. The safety distance is chosen as $d_m = 0.1$.
\begin{figure}[!ht]
\centering
\includegraphics[ width=0.9\columnwidth,clip]{ex3_agent_u1_u2_dist.eps}
\caption{Norm of the control inputs $\|u_1(t)\|, \|u_2(t)\|$ and inter-agent distance $\|x_1(t)-x_2(t)\|$ between the agents.}\label{fig: ex3 u1 u2 dist}
\end{figure}
Figure \ref{fig: ex3 u1 u2 dist} shows the control inputs for the two robots, and their inter-agent distance with time. The control input constraint $\|u_i(t)\|\leq 10$ is satisfied for both $i = 1,2$ at all times. Red-dotted line shows the minimum required inter-agent distance $d_m = 0.1$ for safety. It is clear that the control input and safety constraints are satisfied at all times.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth]{ex3_agent_traj_new_1.eps}
\caption{Closed-loop trajectories of the two robots: snapshot at $t = 1,2,3$ and $4$ sec.}\label{fig: ex3 traj 1}
\vspace{1em}
\includegraphics[width=0.45\textwidth]{ex3_agent_traj_new_2.eps}
\caption{Closed-loop trajectories of the two robots: snapshot at $t = 5,6,7$ and $8$ sec.} \label{fig: ex3 traj 2}
\end{figure}
Figures \ref{fig: ex3 traj 1} and \ref{fig: ex3 traj 2} shows the closed-loop trajectories of the two robots. The safe region $S_s\bigcap (\bigcup\limits_{i = 1}^8 \bar S_i)$ is highlighted in grey color. Purple and green lines plot the paths taken by agent 1 and 2, respectively. The agents visit all the required sets, while maintaining safe distance from each other when they meet in the sets $\bar S_3$ and $\bar S_7$. The figure shows snapshots at the instants when the agents reach the next goal set in the sequence, i.e., first snapshot at $t = 1$ is taken when agent 1 reaches the set $\bar S_2$, i.e., $x_1(t)\in \bar S_2$ and agent 2 reaches the set $\bar S_4$, i.e., $x_2(t)\in \bar S_4$. Snapshots at various time instants illustrate that the closed-loop trajectories satisfy the temporal constraints as well.
\section{Conclusions}\label{sec: conclusion}
In this paper, we presented a new condition for FxTS of the origin in terms of Lyapunov function. We illustrated the application of this new condition in establishing robustness against Lipschitz continuous disturbances. Then, we used this result to solve the problem of satisfying spatiotemporal specifications requiring the system trajectories to remain in a safe set at all times, and reach a goal set within a {fixed} time (user-defined time) in presence of control input constraints. We proposed a novel QP formulation, discussed its feasibility under the mild assumption of existence of control input that renders the safe set forward-invariant, and discussed various cases under which the solution of the QP solves the considered problem.
Two case studies are presented to illustrate applicability of the proposed method to a variety of problems.
In future, we would like to study applicability of the proposed method to large scale systems, in terms of computational power required to solve complex problems. In this work, we considered continuous-time system dynamics, without any switching in the dynamics or the system states. It would be interesting to see how the proposed method extends to a class of switched or hybrid systems, and how to formulate an efficiently solvable optimization problem based method for such class of systems under spatiotemporal specifications.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\subsection{Introduction}
Non-collinear spin textures provide a platform for the study of magnetic ordering and exchange interactions beyond conventional ferro and antiferromagnets, as well as have potential application in electronic devices and data storage \cite{Fert_skyrmion_review, Hellman_interface_magnetism_review}. Of particular recent interest are materials that support multi-\textit{q} non-collinear spin structures, where the spin structure is a superposition of multiple non-collinear orderings, denoted by ordering wavevectors, $q$, along different crystallographic directions, which can lead to topologically non-trivial spin textures, such as skyrmions \cite{Rosler_skyrmion_theory, Muhlbauer_skyrmions, Tokura_skyrmion_rev1}. Such multi-\textit{q} states are known to arise from lattice distortions that result in a non-zero Dzyaloshinskii-Moriya (DM) interaction \cite{DZYALOSHINSKY_DM, Moriya_DM, Bak_helical_theory_DM} or from frustration on triangular lattices \cite{multiq_lattice_frustration, Batista_frustration_review}. However, some materials have neither a DM interaction nor a frustrated lattice, yet still exhibit multi-\textit{q} spin textures. These include elemental Nd \cite{Nd_quad_q}, actinide monopnictides such as USb \cite{ROSSAT_Uranium_multiq, USb_multiq}, and, as very recently demonstrated, the cubic perovskite SrFeO\textsubscript{3} \cite{Ishiwata_SFO_multiQ}. The underlying source of the multi-\textit{q} behavior in these materials is unclear but is critical to understand in order to tune the properties of multi-\textit{q} spin textures. Previous theoretical studies, however, point to a third consideration--electron itinerancy--such that coupling between itinerant and localized electrons leads to multi-\textit{q} structures over single-\textit{q} structures \cite{Martin_multiq_itineracy, Ozawa_multiq_itineracy, Hayami_multiq_itineracy_2, Hayami_multiq_itineracy, Hayami_multiq_itineracy_3, Kakehashi_itinerancy_multiq}.
In order to investigate the role of itinerancy in multi-\textit{q} spin structures, we synthesized epitaxial films of SrFeO\textsubscript{3} and CaFeO\textsubscript{3} and probed their magnetic structure using magnetotransport and resonant soft x-ray magnetic diffraction. While both materials exhibit incommensurate, non-collinear helical spin structures along $\langle111\rangle$ (see Fig. \ref{XRD}(a)) \cite{takeda_SFO_magnetism, Adler_SFO_magnetism, Keimer_SFO_neutron, Mostovoy_SFO_PRL, Ishiwata_SFO_magnetism, Chakraverty_SFO_magnetism, Woodward_CFO, Kawasaki_CFO_first_transport}, SrFeO\textsubscript{3} is metallic whereas CaFeO\textsubscript{3} is insulating (CaFeO\textsubscript{3} exhibits an electronic phase transition ${\sim}$170 K above its N{\'e}el temperature) \cite{MacChesney_SFO, Matsuno_CFO_dispro}. We demonstrate that the SrFeO\textsubscript{3} film exhibits the same complex magnetic phase diagram with distinct helical structures as measured in bulk samples, while further finding that transition temperatures between the magnetic phases more closely resembles Co-doped SrFeO\textsubscript{3}, which we attribute to the moderate tensile strain induced by the substrate. Via resonant soft x-ray magnetic diffraction (RXMD), we show that the magnetic Bragg peaks of CaFeO\textsubscript{3} and SrFeO\textsubscript{3} exhibit significantly different behavior as a function of temperature. We discuss the SrFeO\textsubscript{3} results in context of the recently proposed multi-\textit{q} structures \cite{Ishiwata_SFO_multiQ}, and from this we conclude that our results are consistent with insulating CaFeO\textsubscript{3} supporting a simple multi-domain, single-\textit{q} structure. By synthesizing CaFeO\textsubscript{3}/SrFeO\textsubscript{3} superlattices with different SrFeO\textsubscript{3} layer thicknesses, we find the CaFeO\textsubscript{3} helical structure is coherent through a single unit cell-thick SrFeO\textsubscript{3} layer but is not coherent when the SrFeO\textsubscript{3} layer is increased to 6 unit cells, even though both compounds support helical magnetic structures albeit with slightly different wavevector magnitudes. This result further suggests that the differences in the helical structures of SrFeO\textsubscript{3} and CaFeO\textsubscript{3} are more significant than a simple change to the magnitude of a single \textit{q} wavevector. Our findings point to the importance of carrier itinerancy in multi-\textit{q} spin structures and provide insight into the effect of heterostructuring dissimilar helical structures.
\subsection{Film Results}
Epitaxial, (001)-oriented SrFeO\textsubscript{3} and CaFeO\textsubscript{3} films were deposited by oxygen-assisted molecular beam epitaxy at ${\sim}$650 $^\circ$C with an oxygen partial pressure of 8x$10^{-6}$ Torr (base pressure 4x$10^{-10}$ Torr). The as-grown films were subsequently annealed in the deposition chamber by heating to ${\sim}$600 $^\circ$C in oxygen plasma (200 Watts, 1x$10^{-5}$ Torr chamber pressure) and then cooled in oxygen plasma by progressively turning down the heater to zero output power over approximately one hour, followed by continued exposure to the plasma for another hour to ensure complete cooling to room temperature \cite{Rogge_PRM, Rogge_CFO_XLD}. Because ferrates are known to lose oxygen over time, prior to all measurements the films were re-annealed in oxygen plasma by the same post-growth process to mitigate oxygen deficiency.
\begin{figure}
\includegraphics{Fig1.eps}
\caption{(a) Schematic of the incommensurate, single-\textit{q} helical magnetic ordering along one of the four equivalent $\langle111\rangle$ directions. Multi-\textit{q} spin structures arise when the helical ordering exists along different $\langle111\rangle$ directions simultaneously. (b) Hard X-ray diffraction of a SrFeO\textsubscript{3} film deposited on LSAT(001) (0.5\% tensile strain) and a CaFeO\textsubscript{3} film deposited on LaAlO\textsubscript{3}(001) (0.2\% tensile strain) measured at 300 K with $E = 8,047$ eV photons (Cu $K_\alpha$). The simulated diffraction intensity of a perfect epitaxial film is shown by the dashed black line. (c) Hard X-ray reflectivity of the SrFeO\textsubscript{3} and CaFeO\textsubscript{3} films. Simulated intensity (dashed black lines) for a 14.2 nm thick (37 unit cells) SrFeO\textsubscript{3} film and a 15.8 nm thick (42 pseudocubic unit cells) CaFeO\textsubscript{3} film exhibits good agreement with measured. CaFeO\textsubscript{3} data are offset in $y$. (d) Electrical transport of the SrFeO\textsubscript{3} and CaFeO\textsubscript{3} films. Onset of helical ordering is denoted by the arrows (see text).
\label{XRD}}
\end{figure}
SrFeO\textsubscript{3} was deposited on single crystal (La$_{0.18}$Sr$_{0.82}$)(Al$_{0.59}$Ta$_{0.41}$)O$_3$ (LSAT, +0.5\% strain). CaFeO\textsubscript{3} was deposited on LaAlO$_3$ (+0.2\% strain), as were superlattices of SrFeO\textsubscript{3}/CaFeO\textsubscript{3}. X-ray diffraction of the monolithic films is shown in Fig. \ref{XRD}(b). The simulated diffraction intensity of an ideal epitaxial film exhibits good agreement with measured data. Analysis of thickness fringes from x-ray reflectivity measurements shown in Fig. \ref{XRD}(c) indicate that the SrFeO\textsubscript{3} film is 14.2 nm thick (37 unit cells) and the CaFeO\textsubscript{3} film is 15.8 nm thick (42 pseudocubic unit cells), values that approximately match those obtained from the simulated diffraction data. Electrical transport, measured with Ag paint contacts in the van der Pauw geometry, further confirms the high-quality nature of the films. As seen in Fig. \ref{XRD}(d), CaFeO\textsubscript{3} exhibits a metal-insulator transition near 270 K, and both SrFeO\textsubscript{3} and CaFeO\textsubscript{3} have a 300 K resistivity comparable to bulk samples \cite{MacChesney_SFO, Matsuno_CFO_dispro}. The electrical transport confirms that CaFeO\textsubscript{3} is insulating below its N{\'e}el temperature, whereas SrFeO\textsubscript{3} is metallic.
The electrical transport and magnetoresistance of SrFeO\textsubscript{3} is known to exhibit rich features due to its helical magnetic ordering. Electrical transport, shown in Fig. \ref{MR}(a), exhibits three anomalies as identified by the derivative of the resistivity with temperature (cooling). The first at 117 K is attributed to the onset of magnetic ordering, and a second and third anomaly occur at 110 K and 83 K, respectively. Additionally, the electrical resistivity exhibits hysteresis with temperature. These anomalies and hysteresis are consistent with previous measurements of bulk SrFeO\textsubscript{3}, although a resistivity anomaly at the onset of helical ordering was not observed previously \cite{Lebon_Keimer_SFO, Ishiwata_SFO_magnetism, Long_Co_doped_SFO, Chakraverty_SFO_magnetism}. We adopt the previously used nomenclature and label the regions as Phases I, II, and III \cite{Ishiwata_SFO_magnetism}, as shown in Fig. \ref{MR}(a). No such anomalies are observed in the CaFeO\textsubscript{3} electrical transport \cite{SI_RMD}.
\begin{figure}
\includegraphics{Fig2.eps}
\caption{(a) Electrical transport of the SrFeO\textsubscript{3} film exhibits three anomalies as indicated by the spikes in the derivative with temperature (lower panel) and are attributed to helical Phases I, II, and III. (b) Magnetoresistance measurements (H$\parallel$[001], I$\parallel$[100]) of the SrFeO\textsubscript{3} film measured in succession with increasing temperature. The data are shifted in $y$ for clarity. Double arrow scale bar denotes 5\% MR. (c) Zero-field cooled (ZFC) MR measured at various temperatures. The data are shifted in $y$ for clarity. (d) Percent hysteresis in the zero-field cooled MR as a function of temperature. The vertical, dashed line represents the transition between Phase I and II as identified by the electrical resistivity in (a).
\label{MR}}
\end{figure}
Magnetoresistance [MR $= (\rho(\textrm{H}) - \rho(\textrm{H=0}))/\rho(\textrm{H=0})$] measurements out to 9 T on SrFeO\textsubscript{3} are shown in Fig. \ref{MR}(b) and further confirm the distinct nature of the three identified magnetic phases. The transition from a negative slope (e.g., between 155 K -- 120 K) to a positive slope (e.g., 100 K) is consistent with the onset of helical ordering below 117 K \cite{Ishiwata_SFO_magnetism}. At lower temperatures, the slope of the MR changes sign as the field increases, and a small degree of hysteresis is observed. This change in slope has been attributed to a transition to a field-induced fan- or cone-like helical state (Phases IV and V) \cite{Ishiwata_SFO_magnetism}. Interestingly, this inflection point occurs at lower applied fields compared to previous measurements of both bulk and thin film samples of SrFeO\textsubscript{3}, and instead is more comparable to Co-doped SrFeO\textsubscript{3} \cite{Long_Co_doped_SFO, Chakraverty_SFO_magnetism}.
In order to further distinguish Phase I and II, zero field cooling (ZFC) MR measurements were performed. After each measurement, the sample was heated to 150 K and then cooled with no applied field. The data are shown in Fig. \ref{MR}(c) and reveal that the ZFC MR hysteresis is a strong function of temperature [hysteresis $=(\rho$\textsubscript{ZFC}(H=0) -- $\rho$\textsubscript{0T,9T,0T}(H=0))/$\rho$\textsubscript{ZFC}(H=0)]. An onset of hysteresis is not observed until ${\sim}75$ K that then increases with decreasing temperature. The ZFC MR hysteresis reaches a maximum near 40 K and vanishes below 20 K (see Fig. \ref{MR}(d)). These results are consistent with previous work that attributed the ZFC MR hysteresis to $H$-induced domain rotation, where Phase I exhibits hysteresis but Phase II, notably, does not \cite{Ishiwata_SFO_magnetism}. The onset of MR hysteresis here at ${\sim}75$ K approximately correlates with the transition from Phase II to Phase I as identified by electrical transport (83 K) in Fig. \ref{MR}(a). Additionally, the reduction in hysteresis at lower temperatures suggests that the critical field for domain rotation increases with decreasing temperature below 40 K and is consistent with previous measurements that showed a critical field of ${\sim}$15 T at 4.2 K \cite{Ishiwata_SFO_magnetism,Ishiwata_SFO_multiQ}. Similar MR measurements for CaFeO\textsubscript{3} were not possible due to its large resistivity below its N{\'e}el temperature (100 K). The main difference between these results and previous work is that the temperature range of Phase II is significantly smaller here (83 - 110 K) compared to bulk SrFeO\textsubscript{3} (56 - 110 K) \cite{Ishiwata_SFO_magnetism} and previous measurements of a SrFeO\textsubscript{3} film (46 - 104 K) \cite{Chakraverty_SFO_magnetism}. Instead, the Phase II temperature range of our SrFeO\textsubscript{3} film is comparable to Co-doped SrFeO\textsubscript{3}, approximately equivalent to 1\% Co-doping (SrFe\textsubscript{0.99}Co\textsubscript{0.01}O\textsubscript{3}) in both bulk and thin-film samples \cite{Long_Co_doped_SFO, Chakraverty_SFO_magnetism}.
To gain further insight, we probed the magnetic ordering by measuring the resonant x-ray scattering at the Fe $L$ edge along $q_{H=K=L}$ for the CaFeO\textsubscript{3} and SrFeO\textsubscript{3} films. Measurements were performed at the REIXS beamline at the Canadian Light Source. The (001)-oriented samples were mounted on a ${\sim}$55 degree wedge in order to place the (111) planes in a symmetric scattering configuration. The scattered intensity measured as a function of temperature for the CaFeO\textsubscript{3} and SrFeO\textsubscript{3} films are shown in Figs. \ref{scattering}(a) and (b), respectively. Both films exhibit scattered intensity at values of $q$ slightly suppressed compared to bulk samples: \textit{q}\textsubscript{CaFeO\textsubscript{3}}$ \sim 0.459$ \AA\textsuperscript{-1} (${\sim}1.37$ nm helical wavelength) (1\% suppressed; bulk: \textit{q}\textsubscript{CaFeO\textsubscript{3}} = 0.465 \AA\textsuperscript{-1} \cite{Woodward_CFO,Kawasaki_CFO_first_transport}); \textit{q}\textsubscript{SrFeO\textsubscript{3}}$ \sim 0.356$ \AA\textsuperscript{-1} (${\sim}1.76$ nm helical wavelength) (3\% suppressed; bulk: \textit{q}\textsubscript{SrFeO\textsubscript{3}} = 0.367 \AA\textsuperscript{-1} \cite{Keimer_SFO_neutron}). The suppressed \textit{q} vector for SrFeO\textsubscript{3} is again equivalent to approximately 1\% Co-doping \cite{Long_Co_doped_SFO}. As seen in Fig. \ref{scattering}(a), the CaFeO\textsubscript{3} peak grows uniformly in intensity with decreasing temperature. Integrating the peak area and plotting as a function of temperature in Fig. \ref{scattering}(c) reveals an onset of magnetic ordering at ${\sim}$100 K, which is slightly suppressed compared to bulk ($T_N = 120$ K \cite{CFO_Neel_temp}). Repeating for the SrFeO\textsubscript{3} film, we find an onset temperature of ${\sim}$115 K, which correlates with $T_{N3} = 117$ K as determined from the electrical transport data, and is similarly suppressed compared to bulk ($T_{N3\textrm{,bulk}} = 133$ K \cite{Keimer_SFO_neutron}).
\begin{figure*}
\includegraphics{Fig3.eps}
\caption{(a) Resonant magnetic scattering along $q_{H=K=L}$ for the CaFeO\textsubscript{3} film (E = 710.8 eV) and (b) the SrFeO\textsubscript{3} film (E = 710.6 eV). Data are offset in $y$. Inset in (b) shows the scattering geometry. (c) Total scattered intensity for the CaFeO\textsubscript{3} and SrFeO\textsubscript{3} films. (d) Magnitude of the scattering vector and (e) correlation length derived from the FWHM of the scattering peak. The FWHM and value of \textit{q} were determined by a Gaussian fit of the CaFeO\textsubscript{3} peaks; for SrFeO\textsubscript{3} the FWHM was manually extracted and \textit{q} was determined by the maximum peak intensity. (f) Gaussian fit (dashed line) of the T = 21 K scattering peak (solid line) for SrFeO\textsubscript{3} and (g) CaFeO\textsubscript{3}.
\label{scattering}}
\end{figure*}
From the magnetic scattering data, the \textit{q} vector as a function of temperature was extracted and is plotted in Fig. \ref{scattering}(d). Consistent with previous studies, the SrFeO\textsubscript{3} \textit{q} vector increases in magnitude with decreasing temperature \cite{Keimer_SFO_neutron, Chakraverty_SFO_magnetism}, although here we additionally observe a small decrease below 80 K (Phase II$\rightarrow$I transition). In contrast, CaFeO\textsubscript{3} exhibits the opposite behavior, where the \textit{q} vector decreases with decreasing temperature followed by a small increase at the lowest temperatures. Additionally, the overall change in \textit{q} for CaFeO\textsubscript{3} is approximately 2x smaller than that exhibited by SrFeO\textsubscript{3}.
Comparing the peak shape evolution with temperature between CaFeO\textsubscript{3} and SrFeO\textsubscript{3} in Figs. \ref{scattering}(a) and \ref{scattering}(b), respectively, reveals three striking contrasts. First, SrFeO\textsubscript{3} exhibits a significantly enhanced scattering intensity compared to CaFeO\textsubscript{3}. As seen in Fig. \ref{scattering}(c), the scattered intensity for SrFeO\textsubscript{3} is over 10x greater than CaFeO\textsubscript{3}, which is unexpected given the same nominal film thickness, x-ray footprint, and detector settings (because the \textit{q} vectors are different, the x-ray footprint for CaFeO\textsubscript{3} is reduced by ${\sim}$25\%, which does not sufficiently explain the 10x reduction in intensity). The second observable difference between SrFeO\textsubscript{3} and CaFeO\textsubscript{3} is whereas the CaFeO\textsubscript{3} peak is initially broad and becomes more narrow with decreasing temperature, SrFeO\textsubscript{3} exhibits a narrow peak at the onset of helical ordering that then broadens with decreasing temperature. Converting the FWHM to a correlation length, $\xi = 2\pi/$FWHM, and plotting as a function of temperature in Fig. \ref{scattering}(e) demonstrates that the SrFeO\textsubscript{3} correlation length is initially 4x greater than CaFeO\textsubscript{3} and nearly 2x greater at lower temperatures. We note that the lower bound of the correlation length for an asymmetric Bragg peak of a thin film (e.g., the 111 reflection of a (001)-oriented film) is not limited to the film's thickness.
The third major contrast between SrFeO\textsubscript{3} and CaFeO\textsubscript{3} is the significantly different peak shape evolution exhibited by SrFeO\textsubscript{3}. With decreasing temperature, the peak asymmetrically broadens, where the expansion occurs at higher values of $q$, as highlighted in Fig. \ref{scattering}(f), where a symmetric Gaussian function cannot replicate the peak at 21 K. Such broadening was not observed in previous RXMD measurements of a SrFeO\textsubscript{3} film \cite{Chakraverty_SFO_magnetism}. This broadening ceases below 80 K (see Fig. \ref{scattering}(e)), which correlates with the previously determined Phase II$\rightarrow$I transition in SrFeO\textsubscript{3} ($T_{N1}$ = 83 K) extracted from electrical resistivity and ZFC MR measurements. In contrast, the CaFeO\textsubscript{3} scattering peak at 21 K is replicated by a symmetric Gaussian curve, as shown in Fig. \ref{scattering}(g).
\subsection{Film results discussion}
The helical spin structures of our SrFeO\textsubscript{3} film exhibit features equivalent to approximately 1\% Co-doped SrFeO\textsubscript{3} \cite{Long_Co_doped_SFO, Chakraverty_SFO_magnetism}, as demonstrated by the narrowing of the Phase II temperature window, a decrease in the magnetic field strength at which the MR slope changes sign, a decrease in the magnitude of the \textit{q} vector as probed by RXMD, and a decreased N{\'e}el temperature. A decrease in the \textit{q} vector implies that the real-space length of the helix increases or, analogously, the helical angle between neighboring (111) planes, $\phi$, decreases. A previous theoretical study of the helical state in these ferrates found that the helical ordering arises due to the double exchange effect coupled with a negative charge transfer energy, $\Delta$ \cite{Mostovoy_SFO_PRL, Mostovoy_SFO_2}, where a negative $\Delta$ arises from the high formal oxidation state of Fe\textsuperscript{4+} in these ferrates \cite{Sawatzky_neg_charge_trans_1, Matsuno_CFO_dispro, Robert, Rogge_PRM}. The helical angle tracks a single parameter $\delta = (\epsilon_F - \Delta + t_{pp})/(pd\sigma)$, where $\epsilon_F$ is the Fermi level position, $t_{pp}$ is the oxygen-oxygen hopping amplitude, and $pd\sigma$ is the $\sigma$ hybridization between $p$ and $d$ orbitals. Thus a decrease in $\phi$ corresponds to a decrease in $\delta$, which can occur for an increased $\Delta$, a decreased $t_{pp}$, and/or an increased $pd\sigma$. While tensile strain would be expected to decrease $pd\sigma$, it would also be expected to increase $\Delta$ and decrease $t_{pp}$, which can account for the reduction in $\delta$ and thus a reduction in \textit{q}. In other words, we find that tensile strain acts to increase the ferromagnetic contribution to the SrFeO\textsubscript{3} spin structure, likely due to an increase in $\Delta$ and/or a decrease in $t_{pp}$.
To demonstrate that a negative $\Delta$ plays an important role in the magnetic ordering in these ferrates, we measured the resonant magnetic scattering of the SrFeO\textsubscript{3} film as a function of energy across the O $K$ edge on the Bragg condition ($q_z = 0.3548$ \AA\textsuperscript{-1}) and slightly off the Bragg condition (detector offset by 2 degrees). As seen in Fig. \ref{Oxy}, off the Bragg condition the measured intensity has features that resemble an x-ray absorption spectrum of these ferrates, as expected, where the strong prepeak feature between 527.0-529.5 eV arises from the oxygen ligand holes due to the negative charge transfer energy \cite{Abbate_SFO_XAS, Tsuyama_SFO_XPS, Rogge_PRM}. At the Bragg condition, there is a clear enhancement of the intensity within the prepeak region only. Taking the difference between the on and off Bragg conditions isolates the magnetic contribution to the scattered intensity and is shown in the lower panel of Fig. \ref{Oxy}. The observation that the magnetic contribution to the scattered intensity occurs only in the prepeak region supports the conclusion that the charge transfer energy, $\Delta$, is indeed negative and that the interaction between the O $2p$ and Fe $3d$ states is very strong. For completeness, the resonant magnetic scattering across the Fe $L$ edge for the SrFeO\textsubscript{3} and CaFeO\textsubscript{3} films are shown in the Supplemental Material \cite{SI_RMD}.
\begin{figure}
\includegraphics{Fig4.eps}
\caption{(a) Intensity measured on the magnetic Bragg condition ($I$\textsubscript{Bragg}) and off ($I$\textsubscript{off}) across the O $K$ edge for the SrFeO\textsubscript{3} film on LSAT at 21 K. The O prepeak is at 527.0-529.5 eV, and the peak at 531.5-533.5 eV is from the LSAT substrate. Lower panel shows the magnetic contribution to the scattering as determined by ($I$\textsubscript{Bragg} - $I$\textsubscript{off}).
\label{Oxy}}
\end{figure}
The resonant magnetic diffraction results for SrFeO\textsubscript{3} correlate with the magnetic phase transitions identified by electrical transport while highlighting distinct differences compared to CaFeO\textsubscript{3}. The evolution of the SrFeO\textsubscript{3} scattering peak is notably unconventional; while the CaFeO\textsubscript{3} correlation length increases slightly upon cooling, as would be expected based on simple thermal considerations, the SrFeO\textsubscript{3} correlation length decreases as a result of asymmetric peak expansion, which we attribute to the transition from Phase II to Phase I upon cooling. In contrast, the CaFeO\textsubscript{3} scattering peak remains symmetric down to at least 21 K and thus suggests that CaFeO\textsubscript{3} does not undergo additional magnetic phase transitions.
Recently, it has been proposed that SrFeO\textsubscript{3} supports multi-\textit{q} magnetic structures \cite{Ishiwata_SFO_multiQ} such that the spin structure is a superposition of multiple \textit{q} vectors along different crystallographic directions. Specifically, Phase II is proposed to be a single-domain quadruple-\textit{q} structure consisting of proper screw helical ordering with propagation vectors along the four $\langle111\rangle$ vectors of the cubic unit cell \cite{Ishiwata_SFO_multiQ}. The lack of hysteresis in the ZFC MR in Phase II is consistent with a single-domain structure. Additionally, the much larger domain size (correlation length) in Phase II SrFeO\textsubscript{3} compared to CaFeO\textsubscript{3} could arise from a single-domain structure, although we cannot rule out that the CaFeO\textsubscript{3} film may be more defective than the SrFeO\textsubscript{3} film and thus has a smaller domain size.
The source of the asymmetric change in peak shape and decreased correlation length exhibited by SrFeO\textsubscript{3} at lower temperatures is harder to disentangle with these data alone. We discuss two possible scenarios. In the first scenario, a second, symmetric scattering peak forms at slightly higher \textit{q} values, and grows in intensity with decreasing temperature. Such a scenario would be consistent with Ishiwata \textit{et al.}'s \cite{Ishiwata_SFO_multiQ} proposal of a multi-domain, double-\textit{q} structure in Phase I, where each domain includes both a proper screw ($q_1$) and a vertical cycloid ($q_2$) ordering along one of the four $\langle111\rangle$ directions, where $q_1$ and $q_2$ are along different $\langle111\rangle$ directions, and the magnitude of $q_1$ is approximately equal to the magnitude of $q_2$. In this picture, the two scattering peaks measured along the same [111] direction arise from ordering in different domains: $q_1$ of proper screw along [111] of one domain orientation and $q_2$ of vertical cycloid along [111] of another domain orientation. An additional physical implication for this scenario is that the coherence length is dramatically underestimated because the two distinct peaks were treated as one broad peak. Ishiwata \textit{et al.} also demonstrate that the scattering peak splits into 3 individual peaks upon transition to Phase I from Phase II, where the three peaks are given by ($q$, $q$, $q'$), ($q$, $q'$, $q$), ($q'$, $q$, $q$) and $q'>q$. In a second possible scenario, the asymmetric peak broadening could be due to this splitting, where the precise peak shape depends on the experimental measurement conditions, in particular how \textit{q}-space is scanned and the positions of the three peaks.
The main takeaway of both scenarios, however, is that the peak shape evolution of SrFeO\textsubscript{3} is not unexpected in context of the recent neutron diffraction measurements and is consistent with a transition from Phase II$\rightarrow$I. Thus our results, particularly the correlation length, highlight how x-ray scattering coupled with thin film effects provides another way to probe multi-\textit{q} magnetic states beyond neutron diffraction measurements. Moreover, these results reveal that the Phase II$\rightarrow$I transition is not as abrupt as indicated in previous phase diagrams, because this transition has been determined based on electrical transport measurements. Here, electrical transport identifies Phase II within 110--83 K, but as seen by the change in the scattering peak shape (as proxied by the correlation length in Fig. \ref{scattering}(e)), the spin structure of SrFeO\textsubscript{3} evolves almost constantly with temperature from 110 to 85 K in Phase II. Thus the single domain Phase II state may be stable only within a very narrow temperature range just below $T_{N2}$ (110 K). This may account for the lack of an observed decrease in total scattered intensity across the SrFeO\textsubscript{3} Phase II/I transition (single domain to multi-domain structure) because the still rapidly changing total intensity near 110 K could obscure this effect (see Fig. \ref{scattering}(c)).
These results provide important context for analyzing the CaFeO\textsubscript{3} scattering data. While previous studies have identified helical ordering in CaFeO\textsubscript{3} \cite{Woodward_CFO, Kawasaki_CFO_first_transport}, it is unknown if CaFeO\textsubscript{3} supports multi-\textit{q} spin structures, which one may expect given its similarity to SrFeO\textsubscript{3}. However, the data here demonstrate distinct differences between SrFeO\textsubscript{3} and CaFeO\textsubscript{3}. The CaFeO\textsubscript{3} peak grows uniformly with decreasing temperature whereas the SrFeO\textsubscript{3} peak asymmetrically grows and exhibits larger changes in \textit{q}. Additionally, the CaFeO\textsubscript{3} peak at low temperatures is symmetric, as seen in Fig. \ref{scattering}(g), suggesting that it does not replicate the presumed double-\textit{q} ordering seen in SrFeO\textsubscript{3} at low temperatures. Although we cannot definitively determine the precise details of the spin structure within CaFeO\textsubscript{3}, these results are consistent with a multi-domain, single-\textit{q} helical structure, where different domains have helical ordering along one of the four $\langle111\rangle$ directions. A multi-domain helical state in CaFeO\textsubscript{3} is also consistent with the significantly reduced scattering intensity of CaFeO\textsubscript{3} compared to SrFeO\textsubscript{3}--a 4x reduction would be expected given the 4 equivalent $\langle111\rangle$ directions.
A possible reason for SrFeO\textsubscript{3} hosting multi-\textit{q} spin textures but not CaFeO\textsubscript{3} is their different itinerancies. Previous theoretical studies predict that single-\textit{q} ordering can be destabilized in itinerant systems and instead multi-\textit{q} structures are preferred \cite{Martin_multiq_itineracy, Hayami_multiq_itineracy_2, Hayami_multiq_itineracy, Hayami_multiq_itineracy_3}. While a previous experimental study of the helimagnet Y\textsubscript{3}Co\textsubscript{8}Sn\textsubscript{4} has attributed itinerancy to the source of its multi-\textit{q} state, it also has a DM interaction and lattice frustration that can result in multi-\textit{q} structures \cite{Takagi_multiq_itineracy}. Here, neither lattice frustration nor the DM interaction are present in cubic SrFeO\textsubscript{3}, thus leaving electron itinerancy as the likely source of multi-\textit{q} states in SrFeO\textsubscript{3}. The fact that CaFeO\textsubscript{3} is insulating below its N{\'e}el temperature further supports the conclusion that it has a simple single-\textit{q} spin structure. Future magnetic field-dependent neutron diffraction measurements could confirm a single-\textit{q} structure in CaFeO\textsubscript{3}. Interestingly, tuning the CaFeO\textsubscript{3} metal-insulator transition temperature (e.g., through \textit{A}-site substitution (Ca\textsubscript{1-x}Sr\textsubscript{x}FeO\textsubscript{3} \cite{Takeda_CFO}) or modification of the atomic structure (i.e., octahedral rotations \cite{Antonio_CFO_strain})) below its N{\'e}el temperature could enable further studies of the role of itinerancy in multi-\textit{q} helimagnets.
\subsection{Superlattice results and discussion}
In order to probe how these different helical structures interact, we synthesized superlattices of CaFeO\textsubscript{3} and SrFeO\textsubscript{3}. Superlattices consisting of [(CaFeO\textsubscript{3})\textsubscript{20}/(SrFeO\textsubscript{3})\textsubscript{n}] x 3 for $n=$ 1, 4, and 6 unit cells were deposited on LaAlO\textsubscript{3}(001), as illustrated in Fig. \ref{superlattices_structure}(a). Non-resonant, hard X-ray reflectivity measurements, shown in Fig. \ref{superlattices_structure}(b), exhibit thickness oscillations and superlattice peaks consistent with the superstructures. The measured and simulated reflectivity for the $n=$ 1, 4, and 6 samples exhibits good agreement, and the corresponding scattering length density for the simulated data, shown in Fig. \ref{superlattices_structure}(c), confirms the superlattice structure. The top SrFeO\textsubscript{3} layer is obscured in the scattering length density for the $n=1$ and 4 superlattices due to the surface roughness (10 \AA{ }and 14 \AA, respectively). All three superlattices exhibit electrical transport similar to the monolithic CaFeO\textsubscript{3} film \cite{SI_RMD}.
\begin{figure}
\includegraphics{Fig5.eps}
\caption{(a) Schematic of the ferrate superlattices studied here. (b) X-ray reflectivity of three ferrate superlattices consisting of [(CaFeO\textsubscript{3})\textsubscript{20}/(SrFeO\textsubscript{3})\textsubscript{n}] x 3 for $n=$1, 4, and 6 unit cells (u. c.) on LaAlO\textsubscript{3}(001) measured at 300 K with $E = 8,047$ eV photons (Cu $K_\alpha$; hard X-rays). Data for $n=4$ and 6 are offset in $y$ for clarity. (c) Real part of the scattering length density corresponding to the simulated reflectivity in (b); data are shifted by 0.2 and 0.4 $r_e$/\AA\textsuperscript{3} for the $n=4$, 6 superlattices, respectively.
\label{superlattices_structure}}
\end{figure}
The resonant magnetic scattering data obtained from the three superlattices are shown in Fig. \ref{superlattices}(a) and reveal three trends. First, as seen in Fig. \ref{superlattices}(b), the onset temperature is near 115 K for all three superlattices. This is 15 K higher than that measured for the monolithic CaFeO\textsubscript{3} film and is closer to that measured for the SrFeO\textsubscript{3} film ($T_{N3} = 117$ K). The intensity of the $n=1$ superlattice trends differently with temperature compared to the other two superlattices, decreasing below 75 K and increasing again below 60 K. Second, the \textit{q} vector of all three superlattices is approximately equal to that of the monolithic CaFeO\textsubscript{3} film and the trend of the temperature dependent $q$ vector resembles that of the CaFeO\textsubscript{3} film as well.
\begin{figure}
\includegraphics{Fig6.eps}
\caption{(a) Resonant scattered intensity along $q_{H=K=L}$ for the three superlattices measured at different temperatures with $E = 711.0$ eV. The intensity at 21 K was normalized to unity because different detector settings were used among the superlattices. (b) Total scattered intensity, (c) $q$, and (d) correlation length as a function of temperature. (e) Schematic of the superlattice structure in the angled measurement geometry and depictions of the measured correlation length relative to the relevant film dimensions for the $n=1$ and 6 superlattices.
\label{superlattices}}
\end{figure}
Third, the correlation length decreases with increasing SrFeO\textsubscript{3} thickness, indicating that the SrFeO\textsubscript{3} layers disrupt the helical ordering. At the lowest temperatures, we find that the coherence length is 37 nm for the $n=1$ superlattice and decreases to 16 nm for $n=6$. For the $n=1$ superlattice, the film thickness along [111] is 41 nm, and in the simple case in which the magnetic domains are isotropic, this would indicate that the helical ordering is coherent through the entire superlattice such that the magnetic moments in the SrFeO\textsubscript{3} layers are ordered coherently with those in the neighboring CaFeO\textsubscript{3} layers. For $n=6$, the 16 nm correlation length is comparable to the thickness of the individual CaFeO\textsubscript{3} layers (${\sim}13$ nm), indicating that the helical ordering within the CaFeO\textsubscript{3} layers does not propagate through the now thicker SrFeO\textsubscript{3} layers. Although measuring down to $q = 0.35$ \AA\textsuperscript{-1}, below the SrFeO\textsubscript{3} wavevector, did not show scattered intensity (not shown), the SrFeO\textsubscript{3} layers may be too thin to detect magnetic ordering, and thus we cannot ascertain definitively if the SrFeO\textsubscript{3} layers of the $n=6$ superlattice are magnetically ordered or not. However, propagation through $n=1$ but not $n=6$ offers further evidence that the helical spin structures of SrFeO\textsubscript{3} and CaFeO\textsubscript{3} are different. Given that both compounds exhibit helical ordering with comparable wavevectors, it is surprising that only 6 SrFeO\textsubscript{3} unit cells disrupts the helical ordering if the spin texture is that of a single \textit{q} helix. The coherence of the spin structure in the $n=1$ superlattice implies that the multi-\textit{q} spin texture of SrFeO\textsubscript{3} has been converted to a single-\textit{q} texture due to proximity to CaFeO\textsubscript{3} and/or confinement effects.
In summary we have explored the role of electron itinerancy in the formation of non-collinear spin textures by studying the magnetic structures of metallic SrFeO\textsubscript{3} and insulating CaFeO\textsubscript{3}. We confirm that our SrFeO\textsubscript{3} film exhibits magnetotransport signatures consistent with its previously determined multi-\textit{q} magnetic structure, and further demonstrate that its resonant soft x-ray magnetic diffraction behavior with temperature is consistent with a multi-\textit{q} spin structure. CaFeO\textsubscript{3}, on the other hand, is found to exhibit significantly different magnetic diffraction characteristics compared to SrFeO\textsubscript{3}, which we attribute to a single-\textit{q} spin helix in CaFeO\textsubscript{3}. Additionally, by synthesizing CaFeO\textsubscript{3}/SrFeO\textsubscript{3} superlattices, we demonstrated that relatively thin layers (6 unit cells) of SrFeO\textsubscript{3} is sufficient to disrupt spin coherency through the superlattice, further supporting the conclusion that SrFeO\textsubscript{3} and CaFeO\textsubscript{3} have different helical magnetic structures. The lack of a Dzyaloshinskii-Moriya interaction and lattice frustration in cubic SrFeO\textsubscript{3}, and the presence of only a single-\textit{q} helical ordering in insulating CaFeO\textsubscript{3}, supports the conclusion that electron itinerancy plays a critical role in the formation of the multi-\textit{q} spin structures in SrFeO\textsubscript{3}. Thus, tuning electron itinerancy in other non-collinear spin structures can potentially be a path towards controlling multi-\textit{q} spin structures and their topologically non-trivial spin structures.
\begin{acknowledgments}
We thank J. A. Borchers and M. R. Fitzsimmons for useful discussions. PCR and SJM were supported by the Army Research Office, grant number W911NF-15-1-0133, and film synthesis at Drexel utilized deposition instrumentation acquired through an Army Research Office DURIP grant (W911NF-14-1-0493). RJG was supported by the Natural Sciences and Engineering Research Council of Canada. Research described in this paper was performed at the Canadian Light Source, which is supported by the Canada Foundation for Innovation, Natural Sciences and Engineering Research Council of Canada, the University of Saskatchewan, the Government of Saskatchewan, Western Economic Diversification Canada, the National Research Council Canada, and the Canadian Institutes of Health Research.
\end{acknowledgments}
|
\section{Introduction}
Determining the nanoscale positions of point emitters forms the basis of localization microscopy techniques such as single particle tracking \citep{katayama2009real, manzo2015review}, (fluorescence) photoactivated localization microscopy (f)PALM \citep{Betzig,HESS20064258}, stochastic optical reconstruction microscopy (STORM) \citep{Rust2006}, and related single molecule localization microscopy (SMLM) methods. These techniques have revolutionized biological imaging, revealing cellular processes and structures at the nanoscale \citep{SAHL2013778}. Notably, most samples of interest extend in three dimensions, necessitating three-dimensional (3D) localization microscopy \citep{von2017three}.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.335\textwidth]{main-concept.pdf}
\caption{\textbf{Optical setup and approach overview}. \textbf{a} The light emitted from a fluorescent microscopic particle is collected by the objective and focused through the tube lens into an image at the intermediate image-plane. This plane is extended using a 4f system with a phase mask placed at the Fourier plane in between the two 4f lenses. \textbf{b} The implemented phase mask (using either a Spatial Light Modulator (SLM) or fabricated fused-silica) dictates the shape of the PSF as function of the emitter's axial position. \textbf{c} After training, our CNN receives a 2D low resolution image of overlapping PSFs and outputs a 3D high-resolution volume which is translated to a list of 3D localizations. Blue empty spheres denote simulated GT positions along the surface of an ellipsoid. Red spheres denote CNN detections. The Tetrapod PSF is depicted here, however the approach is applicable to any PSF, including those optimized by the net itself (Fig. \ref{fig:psf-learning}). Scale bars are 3 \(\mu\)m.}
\label{fig:concept}
\end{figure}
\begin{figure*}[!ht]
\centering
\includegraphics{main-mp-deep.pdf}
\caption{\textbf{Comparison to MP}. \textbf{a} The trained CNN is superior to the matching pursuit approach in both detectability (Jaccard index) and in accuracy (Lateral\textbackslash Axial RMSE). Matching of points was computed with a threshold distance of 150 nm using the Hungarian algorithm \citep{kuhn1955hungarian}. \textbf{b} Example of a simulated frame of density 0.124 \(\left[\frac{\text{emitters}}{\mu m^2}\right]\) alongside 3D comparisons of the recovered positions by MP (middle) and by the CNN (right). Scale bar is 2 \(\mu\)m.}
\label{fig:comparison-mp}
\end{figure*}
In a standard microscope, the precise z position of an emitter is difficult to ascertain because the change of the point-spread function (PSF) near the focus is approximately symmetric. Furthermore, outside of this focal range (\(\approx \pm\) 350 nm for a high numerical aperture imaging system), the rapid defocusing of the PSF reduces the signal-to-noise ratio causing the localization precision to quickly degrade. One method to extend the useful z-range and explicitly encode the z position is PSF engineering \citep{pavani2009three, huang2008three,shechtman2014optimal}. Here, an additional optical element, e.g. a phase mask, is placed in the emission path of the microscope, modifying the image formed on the detector \citep{backer2014extending} (Fig. \ref{fig:concept}a); the axial position can then be recovered via image processing using a theoretical or experimentally-calibrated PSF model \citep{shechtman2014optimal, liu2013three, babcock2017analyzing, li2018real}.
In practically all applications, it is desirable to be able to localize nearby emitters simultaneously. For example, in super-resolution SMLM experiments, the number of emitters localized per frame determines the temporal resolution. In tracking applications, PSF overlap from multiple emitters often precludes localization, potentially biasing results in emitter-dense regions. The problem is that localizing overlapping emitters poses a significant algorithmic challenge even in 2D localization, and much more so in 3D. Specifically, encoding the axial position of an emitter over large axial ranges (>3 \(\mu\)m) requires the use of laterally large PSFs, e.g. the Tetrapod \citep{shechtman2014optimal, shechtman2015precise} (Fig. \ref{fig:concept} b), increasing the possibility of overlap. Consequently, while a variety of methods have been developed to cope with overlapping emitters for the in-focus, standard-PSF \citep{min2014falcon, boyd2017alternating, nehme2018deep}, the performance in high-density 3D localization situations is far from satisfactory \citep{sage2019super}.
Deep learning has proven to be adept at analyzing microscopic microscopy data \citep{rivenson2018phase, nguyen2018deep, weigert2018content, rivenson2019phasestain, liu2019deep, smith2019ultra}, especially for single-molecule localization, handling dense fields of emitters over small axial ranges (<1.5 \(\mu\)m) \citep{nehme2018deep, boyd2018deeploco, ouyang2018deep, diederich2019cellstorm, newby2018convolutional, zelger2018three, liu2018fast,hershko2019multicolor, speiser2019teaching} or sparse emitters spread over larger ranges \citep{zhang2018analyzing}. Moreover, an emerging application is to jointly design the optical system alongside the data processing algorithm, enabling end-to-end optimization of both components \citep{chakrabarti2016learning, horstmeyer2017convolutional, turpin2018light, haim2018depth, he2018learning, hershko2019multicolor, sitzmann2018end, chang2019deep, wuphasecam3d}. Here we present DeepSTORM3D, consisting of two fundamental contributions to high-density 3D localization microscopy over large axial ranges. First, we employ a convolutional neural network (CNN) for analyzing dense fields of overlapping emitters with engineered PSFs, demonstrated with the large-axial-range Tetrapod PSF \citep{shechtman2014optimal, shechtman2015precise}. Second, we design an optimal PSF for 3D localization of dense emitters over a large axial range of 4 \(\mu\)m. By incorporating a physical-simulation layer in the CNN with an adjustable phase modulation, we jointly learn the optimal PSF (encoding) and associated localization algorithm (decoding). This approach is highly flexible and easily adapted for any 3D SMLM dataset parameters, \textit{i.e.} emitter density, SNRs, and z-range. We quantify the performance of the method by simulation, and demonstrate the applicability to 3D biological samples, \textit{i.e.} mitochondria and telomere.
\begin{figure*}[!htb]
\centering
\includegraphics{main-storm3d.pdf}
\caption{\textbf{Super-resolution 3D imaging over a 4 \(\mu\)m z-range}. \textbf{a} Super-resolved image of mitochondria spanning a \(\approx\)4\(\mu\)m z-range rendered as a 2D histogram where z is encoded by color. \textbf{b} Representative experimental frame (top), and rendered frame from the 3D recovered positions by the CNN overlaid on top (bottom). \textbf{c} Diffraction limited (left), super-resolved (middle), and cross-section of the super-resolved image at \(z=1.5 \ \mu\)m (right). Scale bars are 3 \(\mu\)m.}
\label{fig:storm3d}
\end{figure*}
\section{Results}
To solve the high-density localization problem in 3D, we trained a CNN that receives a 2D image of overlapping Tetrapod PSFs spanning an axial range of 4 \(\mu\)m, and outputs a 3D grid with a voxel-size of \(27.5 \times 27.5 \times 33 \ nm^3\) (Fig. \ref{fig:concept}c). For architecture details and learning hyper-parameters see Supplementary Information sections 1.1 and 3. To compile a list of localizations, we apply simple thresholding, and local maximum finding on the output 3D grid (Supplementary Information section 3.4).
\begin{figure*}[!h]
\centering
\includegraphics{main-physical-net.pdf}
\caption{\textbf{PSF learning for high density 3D imaging}. \textbf{a} Simulated 3D emitter positions are fed to the image formation model to simulate their low resolution CCD image (Encoding). Next, this image is fed to a CNN that tries to recover the simulated emitter positions (Decoding). The difference between the simulated positions and the positions recovered by the CNN is used to jointly optimize the phase mask at the Fourier plane, and the recovery CNN parameters. \textbf{b} Simulation of the learned PSF as function of the emitter axial position (left). 3D isosurface rendering of the learned PSF (right). \textbf{c} Example frame of density 0.197 \(\left[\frac{\text{emitters}}{\mu m^2}\right]\) (top) with the same simulated emitter positions, using the Tetrapod (left) and the learned PSF (right). Jaccard index (bottom) and lateral \textbackslash axial RMSE comparison (right) between two CNNs with the same architecture, one trained to recover 3D positions from 2D images of Tetrapod PSF (black), and the second trained to recover 3D positions from 2D images of the learned PSF (orange). Scale bars are 3 \(\mu\)m.}
\label{fig:psf-learning}
\end{figure*}
\begin{figure*}[!ht]
\centering
\includegraphics{main-exp.pdf}
\caption{\textbf{Three dimensional imaging of telomeres in a single-snapshot}. \textbf{a} Schematic of imaging fixed U2OS cells with fluorescent labeled telomeres inside their nucleus. \textbf{b} Focus slice with the standard PSF inside a U2OS cell nucleus, obtained via a z-scan. The yellow rectangles mark the same emitter in all three orthogonal planes. \textbf{c} Example fit of the mean intensity in sequential axial slices used to estimate the approximate emitter axial position. \textbf{d} Experimental snapshot with the Tetrapod PSF (left), rendered image from the 3D recovered positions by the Tetrapod CNN (middle), and a 3D comparison of the recovered positions and the approximate experimental ground truth (right). \textbf{e} Experimental snapshot with the learned PSF (left), rendered image from the 3D recovered positions by the learned PSF CNN (middle), and a 3D comparison of the recovered positions and the approximate experimental ground truth (right). Scale bars are 3 \(\mu\)m.}
\label{fig:exp-demo}
\end{figure*}
We compare our method to a fit-and-subtract based Matching Pursuit (MP) approach \citep{shechtman2016multicolour} (see Supplementary Information section 4) as we are unaware of any other methods capable of localizing overlapping Tetrapod PSFs. To quantitatively compare our method with MP solely in terms of density, we simulated emitters with high signal-to-noise ratio (30K signal photons, 150 background photons per pixel) at 10 different densities ranging from 1 to 75 emitters per 13 \(\times\) 13 \(\mu m^2\) field-of-view. The results are shown in Fig. \ref{fig:comparison-mp}. As evident in both the Jaccard index (defined as \(\frac{TP}{TP + FP + FN}\), where TP, FP, FN are true positives, false positives, and false negatives \citep{sage2019super}) and the lateral/axial RMSE (Fig. \ref{fig:comparison-mp}a) the CNN achieves remarkable performance in localizing high-density Tetrapods. In the single-emitter (very low density) case, where the performance of the CNN is bounded by the discretization on the 3D grid, the RMSE of the MP localization is lower (better). This is because for a single-emitter, MP is equivalent to a continuous Maximum Likelihood Estimator (MLE) (Supplementary Information section 4), which is asymptotically optimal \citep{bickel2015mathematical}, whereas the CNN's precision is bounded by pixilation of the grid (\textit{i.e.} half voxel of 13.75 nm in xy and 16.5 nm in z). However, quickly beyond the single-emitter case, the CNN drastically outperforms MP. A similar result was obtained when compared to a leading single-emitter fitting method \citep{li2018real} applicable also for the multiple emitter case \citep{sage2019super} (see Supplementary Information section 6).
Next, we validated our method for super-resolution imaging of fluorescently labeled mitochondria in COS7 cells (Fig. \ref{fig:storm3d}). We acquired 20K diffraction limited frames of a \(50\times30 \ \mu m^2\) FOV and localized them using the CNN in about \(\approx\)10 hours, resulting in \(\approx\)360K localizations. The Tetrapod PSF was implemented using a fabricated fused-silica phase-mask (see Supplementary Information section 7.1). The estimated resolution was \(\approx\)40 nm in xy, and \(\approx\) 50 nm in z (see Supplementary Information section 7.2). To visually evaluate localization performance in a single frame (Fig. \ref{fig:storm3d}b top), we regenerated the corresponding 2D low-resolution image, and overlayed the recovered image with a uniform photons scaling on top of the experimental frame \ref{fig:storm3d}b bottom). As seen in the overlay image, the emitter PSFs (3D positions) are faithfully recovered by the CNN. Moreover, emitters with extremely low number of signal photons were ignored.
The Tetrapod is a special PSF that has been optimized for the single emitter case by Fisher Information maximization \citep{shechtman2014optimal,shechtman2015precise}. However, when considering the multiple-emitter case, an intriguing question arises: What is the optimal PSF for high density 3D localization over a large axial range? To answer this question we need to rethink the design metric; extending the Fisher Information criterion \citep{shechtman2014optimal} to account for emitter density is not-trivial, and while it is intuitive that a smaller-footprint PSF would be preferable for dense emitters, it is not clear how to mathematically balance this demand with the requirement for high localization precision per emitter.
Our PSF-design logic is based on the following: since we have already established that a CNN yields superior reconstruction for high-density 3D localization, \textit{we are interested in a PSF (\textit{encoder}) that would be optimally localized by a CNN (\textit{decoder})}. Therefore, in contrast to a sequential paradigm where the PSF and the localization algorithm are optimized separately, we adopt a co-design approach (Fig. \ref{fig:psf-learning}a). To jointly optimize the PSF and the localization CNN, we introduce a differentiable physical simulation layer, which is parametrized by a phase mask that dictates the microscope's PSF. This layer encodes 3D point sources to their respective low-resolution 2D image (see Supplementary Information section 2). This image is then fed to the localization CNN which decodes it and recovers the underlying 3D source positions. During training, the net is presented with simulated point sources at random locations and, using the difference between the CNN recovery and the simulated 3D positions, we optimize both the phase mask and the localization CNN parameters in an end-to-end fashion. The learned PSF (Fig. \ref{fig:psf-learning}b) has a small lateral footprint, which is critical for minimizing overlap at high densities. Moreover, the learned phase mask twists in a spiral trajectory causing the PSF to rapidly rotate throughout the axial range, a trait that was previously shown to be valuable for encoding depth \citep{pavani2009three}.
To quantify the improvement introduced by our new PSF, we first compare it to the Tetrapod PSF in simulations. Specifically, we train a similar reconstruction net for both the Tetrapod and the learned PSF using a matching training set composed of simulated continuous 3D positions along with their corresponding 2D low-resolution images. The learned PSF performs similar to the Tetrapod PSF for low emitter densities (Fig. \ref{fig:psf-learning}c). However, as the density goes up the learned PSF outperforms the Tetrapod PSF in both localization precision and in emitter detectability (Jaccard index) (Fig. \ref{fig:psf-learning}c). This result is not surprising, as the learned PSF has a smaller spatial foorprint, and hence it is less likely to overlap than the Tetrapod (Fig. \ref{fig:psf-learning}c).
Next, we demonstrate the superiority of the new PSF experimentally by imaging fluorescently labeled telomeres (TRF1-DsRed) in Fixed U2OS cells. The cell contains tens of telomeres squeezed in the volume of a nucleus with \(\approx 20 \ \mu\)m diameter (Fig. \ref{fig:exp-demo}a, b). From a single snapshot focused inside the nucleus, the CNN outputs a list of 3D positions of telomeres spanning an axial range of \(\approx 3 \ \mu\)m. Using the Tetrapod PSF snapshot, the Tetrapod-trained CNN was able to recover 49 out of 62 telomeres with a single false positive, yielding a Jaccard index of 0.77 (Fig. \ref{fig:exp-demo}d). In comparison, using the learned PSF snapshot, the corresponding CNN was able to recover 57 out of the 62 telomeres with only 2 false positives, yielding a Jaccard index of 0.89 (Fig. \ref{fig:exp-demo}e). The recovered positions were compared to approximated ground-truth 3D positions (Fig. \ref{fig:exp-demo}c), obtained by axial scanning and 3D fitting (see Supplementary Information section 9).
To qualitatively compare the recovered list of localizations to the acquired snapshot, we fed this list to the physical simulation layer and generated the matching 2D low-resolution image (Fig. \ref{fig:exp-demo}d,e). As verified by the regenerated images, the 3D positions of the telomeres are faithfully recovered by the CNNs. Moreover, the misses in both snapshots were either due to local aberrations and/or an extremely low number of signal photons (see Supplementary Information section 10 for more experimental results).
\section{Discussion}
In this work we demonstrated 3D localization of dense emitters over a large axial range both numerically and experimentally. The described network architecture exhibits excellent flexibility in dealing with various experimental challenges, e.g. low signal-to-noise ratios and optical aberrations. This versatility is facilitated in three ways: (1) the net was trained solely on simulated data, thus producing sufficiently large datasets for optimization; (2) the phase mask which governs the PSF was optimized with respect to the implementation in the imaging system, \textit{i.e.} the pixels of the spatial light modular, rather than over a smaller subspace, e.g. Zernike polynomials \citep{shechtman2014optimal}; (3) the CNN localization algorithm was designed in coordination with the development of the PSF, thus the system was optimized for the desired output \citep{hershko2019multicolor} rather than a proxy.
Attaining a sufficiently large training dataset has thus far been a major limitation for most applications of CNNs. With this limitation in mind, the application of CNNs to single-molecule localization would seemingly be an ideal one, since each emitter's behavior should be approximately the same. This uniformity is broken, however, by spatially-varying background, sample density, and variable emitter size in biological samples (Supplementary Information 3.1), all of which diversify datasets and necessitate relevant training data. By implementing an accurate simulator, we have shown that it is possible to build a robust network entirely in silico, generating arbitrarily large, realistic datasets with a known ground truth to optimize the nets. This aligns with our previous work in 2D SMLM \citep{nehme2018deep}).
For super-resolution reconstructions using the Tetrapod PSF, the simulator was particularly important due to the highly variable SNR of emitters in the sample. Here, our net was able to selectively localize the emitters even in very dense regions by focusing on those with a high SNR (Fig. \ref{fig:storm3d}). To optimize a PSF while simultaneously training the net, the simulator was also essential, as it would be prohibitively time consuming to experimentally vary the PSF, while recording and analyzing images to train the net.
An intriguing aspect of our optimization approach is that the optimized PSF is found by continuously varying the pixels of an initialized mask while evaluating the output of the localization net, thus the final result represents a local minimum (Fig. \ref{fig:psf-learning}). By changing the initialization conditions, we have recognized several patterns that indicate how the optimal PSF varies with the experimental conditions: namely, density, axial range, and SNR (see Supplementary Information section 1.2). Some of the recurrent features are intuitive: for example, in dense fields of emitters with limited SNR, the optimized PSFs have a small footprint over the designed axial range, enabling high density and compacting precious signal photons into as few pixels as possible. What distinguishes the net PSFs over predetermined designs is the utilization of multiple types of depth encoding; namely, simultaneously employing astigmatism, rotation, and side lobe movement (Fig. \ref{fig:psf-learning}), all of which have been conceived of and implemented previously, but never simultaneously!
This work, therefore, triggers many possible questions and research directions regarding its capabilities and limitations. For example, how globally-optimal is the resulting PSF? Similarly, how sensitive is the resulting PSF and its performance to different loss functions, CNN architectures, initializations (e.g. with an existing phase mask), and the sampled training set of locations? Currently, it is unclear how each of these components affects the learning process although we began to partially answer them in simulations (see Supplementary Information section 1.2). Finally, the co-design approach employed here paves the way to a wide variety of interesting applications in microscopy where imaging systems have traditionally been designed separately from the processing algorithm.
\section*{Funding Information}
Google; H2020 European Research Council Horizon 2020 (802567); Israel Science Foundation (ISF) (852/17); Ollendorff Foundation; Technion-Israel Institute of Technology (Career Advancement Chairship); Zuckerman Foundation.
\section*{Acknowledgements}
The authors wish to thank the Garini lab for the U2OS cells. We also thank Jonas Ries for his help with the application of SMAP-2018 to Tetrapod PSFs. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan V GPU used for this research. We thank the staff of the Micro-Nano-Fabrication \(\&\) Printing Unit (MNF-PU) at the Technion for their assist with the phase mask fabrication. Finally, we thank Google for the research cloud units provided to accelerate this research.
\section*{Author contributions}
EN, DF, TM and YS conceived the approach. EN performed the simulations and analyzed the data with contributions from all authors. EN, RG, BF, LEW, and OA took the data. RO fabricated the physical phase mask. EN, DF, LEW, TM and YS wrote the paper with contributions from all authors.
\section*{Competing Interests}
The authors declare no competing interests.
\section*{Data availability}
Data will be made available upon reasonable request.
\section*{Code availability}
Code will be made publicly available.
\bibliographystyle{ieeetr}
|
\section{Introduction}
One of the most important properties we would like to know about programs is their \emph{resource usage}, i.e., the amount of resources
(such as time, memory and energy) required for their execution. This information is useful during development, when performance bugs
and security vulnerabilities exploiting performance issues can be avoided. It is also particularly relevant for mobile applications,
where resources are limited, and for cloud services, where resource usage is a major cost factor.
In the literature, a lot of different ``cost analysis'' problems (also called ``resource bound analysis'', etc.)
have been studied (e.g.~\cite{Wegbreit:75,Rosendahl89,ACE,Albert-et-al:TCS:2011,APROVE-JAR2017,CiaoPP-TPLP2018,CHS:pldi2015,SZV:jar2017});
several of them may be grouped under the following general definition.
The \emph{countable resource problem} asks about the maximum usage of a ``resource'' that accumulates during execution,
and which one can explicitly count, by instrumenting the program with an accumulator variable and instructions to increment it where
necessary. For example, we can estimate the \emph{execution time} of a program by counting certain ``basic steps''.
Another example is counting the number of visits to designated program locations. Realistic problems of this type include bounding
the number of calls to specific functions, perhaps to system services; the number of I/O operations; number of accesses to memory, etc.
The consumption of resources such as \emph{energy} suits our problem formulation as long as such explicit bookkeeping is possible (we have to assume that the
increments, if not constant, are given by a monotone polynomial expression).
In this paper we solve the \emph{bound analysis problem} for a particular class of programs, defined in~\cite{BJK08}.
The bound analysis problem is to find symbolic bounds on the maximal possible value of an integer variable
at the end of the program,
in terms of some integer-valued variables that appear in the initial state of a computation.
Thus, a solution to this problem might be used for any of the resource-bound analyses above.
In this work we focus on values that
grow polynomially (in the sense of being bounded by a polynomial), and our goal is to find polynomial bounds that are tight, in the sense
of being precise up to a constant factor.
The programs we study are expressed by the so-called \emph{core language}. It is
an imperative language, including bounded loops, non-deterministic branches and restricted arithmetic expressions;
the syntax is shown in Figure~\ref{fig-syntax}. The semantics is explained and motivated below, but is largely intuitive;
see also the illustrative example in Figure~\ref{fig:intro-example}.
In 2008, it was proved~\cite{BJK08} that for this language it is decidable whether a computed result is polynomially bounded or not. This makes the language
an attractive target for work on the problem of computing tight bounds. However, for the past ten years there has been no improvement on~\cite{BJK08}.
We now present an algorithm to compute, for every program in the language, and every variable in the program which
has a polynomial upper bound (in terms of input values), a tight polynomial bound on its largest attainable value (informally, ``the worst-case value'')
as a function of the input values. The bound is guaranteed to be tight up to a multiplicative constant factor but constants are left implicit (for example a bound
quadratic in $n$ will always be represented as $n^2$).
The algorithm could
be extended to compute upper and lower bounds with explicit constant factors, but choosing to ignore coefficients simplifies the algorithm considerably.
In fact, we have striven for a simple, comprehensible algorithm, and we believe that the algorithm we present is sufficiently simple that, beyond
being comprehensible, offers insight into the structure of computations in this model.
Our philosophy is that research on complete solutions to static analysis questions regarding weak languages is desirable for several reasons.
First, it is theoretically satisfying---it establishes a clear and definite result. The algorithm can be possibly employed later in more complex situations,
and we will at least have an assurance that it does its part; arguments about the value of relying on decidable problems in program analysis have recently been
given in~\cite{McMillanPadon18}, and~\cite{Kincaid:sas18} gives a methodology for incorporating them as parts in a bigger system (albeit for safety problems).
Knowing that a problem is solvable for a certain weak language gives us a point of reference for future research, and makes it meaningful to further discuss
questions of computational complexity.
As pointed out in~\cite{McMillanPadon18}, when an algorithm to prove a property has a completeness proof, it usually means that it is possible to furnish a justification
for a negative answer, which is of value to the user. In our case, if the algorithm returns a bound which is higher than what you wanted, you can obtain from it
an \emph{execution pattern} which shows how the result arose.
Finally, the quest for complete solutions drives research forward by setting challenges which invite new insights and ideas.
Next, we explain the definition of the language in more detail. We will comment
about the motivation for the definitions, in particular vis-{\`a}-vis the analysis of fuller programming languages. The main argument is that choices in
the definition of the language are driven by the idea of using it as a
\emph{conservative abstraction}.
\subsection{The core language}%
\label{sec:language}
\hspace*{5cm} \\
\begin{figure}[htb]
\[\renewcommand{\arraystretch}{1.3}
\begin{array}{rcl}
\verb+X+\in\mbox{Variable} &\;\; ::= \;\; & \texttt{X}_1 \mid\texttt{X}_2 \mid \texttt{X}_3 \mid
\ldots \mid \texttt{X}_n\\
\verb+E+\in\mbox{Expression} & ::= & \verb+X+ \mid \verb/E + E/ \mid
\verb+E * E+\\
\verb+C+\in\mbox{Command} & ::= & \verb+skip+ \mid \verb+X:=E+
\mid \verb+C+_1 \texttt{;} \verb+C+_2
\mid \texttt{loop E \{C\}}
\mid \texttt{choose}\; \texttt{C}_1 \; \texttt{or} \; \texttt{C}_2
\end{array} \renewcommand{\arraystretch}{1.0}\]
\caption{Syntax of the core language.}%
\label{fig-syntax}
\end{figure}
\paragraph{\bf Data.}
The only type of data in the core language is non-negative integers.\footnote{This could be modified to all integers, as explained later.}
In a practical setting, a program may include statements that manipulate
non-integer data that can, however, be abstracted away without losing the information critical to loop control
---hence to a complexity analysis---as loops are often controlled by integer variables.
In other cases, it is possible to preprocess a program to replace complex data values with their size (or ``norm''), which is the quantity of importance
for loop control. Methods for this process have been widely studied in conjunction with termination and cost analysis.
These considerations motivate the study of weak languages that handle integers.
\paragraph{\bf Command semantics.}
The core language is inherently non-deterministic.
The {\tt choose} command represents a non-deterministic choice, and can be used to abstract any concrete conditional command by simply ignoring the condition.
Note that what we ignore is branches within a loop body
and not branches that implement the loop control, as loops are represented by a dedicated loop command.
The command $\verb+loop E {C}+$ repeats \pgt{C} a non-deterministic number of times
bounded by the value of $\pgt{E}$, which is evaluated just before the loop is entered. Thus, as a conservative abstraction, it
may be used to model different forms of loops (for-loops, while-loops)
as long as a bound on the number of iterations, as a function of the program state on loop initiation,
can be determined and expressed in the language.
There is an ample body of research on analysing programs to find such bounds where they are not explicitly given by the programmer;
in particular, bounds can be obtained from a \emph{ranking function} for the loop~\cite{PR:04,BagnaraHZ08,ADFG:2010,Ben-AmramG13jv,Ben-AmramG17}.
Note that the arithmetic in our language is too restricted to allow for the maintenance of counters and the management of \emph{while} loops,
as there is no subtraction, no explicit constants and no tests. Thus, for realistic ``concrete'' programs which use such devices, loop-bound analysis is supposed to be
performed \emph{on the concrete program} as part of the process of abstracting it to the core language. This process is illustrated in~\cite[Sect.~2]{BAPineles:2016}.
The semantics of the loop is non-deterministic, so that the loop is allowed to
actually perform fewer iterations than indicated by the bound expression; this is useful both for modeling loops that can ``break,''
as well as for using the results of auxiliary analyses, as those usually provide just a bound, not a precise number of iterations.
An interesting observation has been made by Jones and Kristiansen~\cite{JK08}: typically, algorithms whose goal is to prove loop termination
and establish loop bounds do so by
focusing on values that \emph{decrease} (examples are the Size-Change Termination principle~\cite{leejonesbenamram01} and numerous
methods that discover \emph{ranking functions}). In contrast, by abstracting to our core language (\cite{JK08} uses a very similar
language), we focus on values that \emph{grow}. Recent work in static analysis~\cite{Giesl:toplas2016} describes an analysis system which
combines a subsystem for
loop-bound analysis (via ranking functions) with a subsystem for
growth-rate analysis, which establishes symbolic bounds on data that grow along a loop.
Our definition of the core language separates the concerns and concentrates on the problem of value growth,
for a program (or program fragment) where loop bounds are already known. Note however that the loop bound is to be given as a function
of the state in which the loop is started, and may well depend on values that are the result of previous computations, as our example program in Figure~\ref{fig:intro-example} illustrates.
\begin{figure}[htb]
\begin{Verbatim}[codes={\catcode`$=3\catcode`_=8}]
loop X$_1$ {
loop X$_2$ + X$_3$ { choose { X$_3$:= X$_1$; X$_2$:= X$_4$ } or { X$_3$:= X$_4$; X$_2$:= X$_1$ } };
X$_4$:= X$_2$ + X$_3$
};
loop X$_4$ { choose { X$_3$:= X$_1$ + X$_2$ + X$_3$ } or { X$_3$:= X$_2$; X$_2$:= X$_1$ } }
\end{Verbatim}
\caption{A core-language program. \texttt{loop} $n$ \pgt{C} means ``do \pgt{C} at most $n$ times.''}%
\label{fig:intro-example}
\end{figure}
From a computability viewpoint,
the use of bounded loops restricts the programs that can be represented to such that compute
primitive recursive functions; this is a rich enough class to cover a lot of useful algorithms and make
the analysis problem challenging. In fact, our language resembles a weak version of Meyer and Ritchie's LOOP language~\cite{MR:67},
which computes all the primitive recursive functions, and where behavioral questions like ``is the result linearly bounded'' are undecidable.
\subsection{The algorithm}
Consider the program in Figure~\ref{fig:intro-example}. Suppose that it is started with the values of the variables $\texttt{X}_1,\texttt{X}_2,\dots$~being $x_1,x_2,\dots$.
Our purpose is to bound the values of all variables at the conclusion of the program in terms of those initial values. Indeed, they are all polynomially bounded, and
our algorithm provides tight bounds. For instance, it establishes that the final value of $\texttt{X}_3$ is tightly bounded (up to a constant factor) by
$\max (x_4(x_4 + x_1^2),x_4(x_2 + x_3 + x_1^2) )$.
Actually, the algorithm produces information in a more precise form, as \emph{a disjunction of simultaneous bounds}.
This means that it generates a set of
tuples of polynomials, called \emph{multi-polynomials}. Each such tuple provides simultaneous bounds on all variables in a subset of possible executions; for example,
with the program in Figure~\ref{fig:intro-example}, one such multi-polynomial is $\tuple{x_1, x_2, x_3, x_2+x_3}$, which means that, starting with the valuation
$\texttt{X}_1\mapsto x_1,\ \texttt{X}_2\mapsto x_2,\ \texttt{X}_3\mapsto x_3, \texttt{X}_4\mapsto x_4$ the value of $\texttt{X}_4$ at the end
of some possible execution (specifically, one in which the second and the third loop commands both exit immediately) is tightly (in this case, exactly)
described by $x_2+x_3$; while the other three variable retain their initial values.
Other multi-polynomials will represent other sets of possible executions, their union covering all executions.
This \emph{disjunctive} form is important in the context of a compositional analysis. To see why, suppose that we provide, for a command with variables
$\texttt{X},\texttt{Y}$, the bounds $\tuple{x,y}$ and $\tuple{y,x}$. Then we know that the \emph{sum} of their values is always bounded by $x+y$, a result that
would have not been deduced had we given the bound $\max(x,y)$ on each of the variables. The difference may be critical
for the success of analyzing an enclosing or subsequent command.
\emph{Multivariate} bounds are often of interest, and perhaps require no justification,
but let us point out that multivariate polynomials are necessary even if we are
ultimately interested in a univariate bound, in terms of some single initial value, say $n$.
This is, again, due to the analysis being compositional. When we analyze an internal command that uses variables
$\texttt{X},\texttt{Y},\dots$ we do not know in what possible contexts the command will be executed and how the values of these variables will be related to $n$.
Some highlights of our solution are as follows.
\begin{itemize}
\item We reduce the problem of analyzing any core-language program to the problem of analyzing a single loop, whose body is
already processed, and therefore presented as a collection of abstract state-transitions. This is typical of algorithms that analyze a structured imperative
language and do so compositionally.
\item Since we are computing bounds only up to a constant factor, we work with \emph{abstract polynomials}, that have no numeric coefficients.
\item We further introduce \emph{$\tau$-polynomials}, to describe the evolution of values in a loop.
These have an additional
parameter $\tau$ (for ``time''; more precisely, number of iterations).
Introducing $\tau$-polynomials was a key step in the solution.
\item The analysis of a loop is simply a closure computation under two operations: ordinary composition, and \emph{generalization} which is the
operation that predicts the evolution of values by judiciously adding $\tau$'s to \emph{idempotent} abstract transitions.
\end{itemize}
\noindent
The remainder of this paper is structured as follows. In Section 2 we give some definitions and state our main result.
In Sections~\ref{sec:algorithm-prelim}---\ref{sec:closure-algorithm} we present our algorithm.
In Section~\ref{sec:correctness-sdl}, we give the correctness statement for our algorithm, and in Section~\ref{sec:correctness-main} we give the correctness proofs for this.
In Section~\ref{sec-complexity} we consider the computational complexity of our algorithm, and in Section~\ref{sec:extensions} we consider extensions to our algorithm and
open problems. Section~\ref{sec:rw} describes related work, and Section~\ref{sec:conclusion} concludes and discusses ideas for further work.
\section{Preliminaries}%
\label{sec:prelim}
In this section, we give some basic definitions, complete the presentation of our programming language and precisely state the main result.
\subsection{Some notation and terminology}
\paragraph{\bf The language} We remark that in our language syntax
there is no special form for a ``program unit;'' in the text we sometimes use ``program'' for the subject of our analysis,
yet syntactically it is just a command.
\paragraph{\bf Polynomials and multi-polynomials}
We work throughout this article with
multivariate polynomials in $x_1,\dots,x_n$
that have non-negative integer coefficients and no variables other than $x_1,\dots,x_n$; when we speak of a polynomial we always mean one of this kind.
Note that over the non-negative integers, such polynomials
are monotonically (weakly) increasing in all variables.
Our algorithm sometimes deals with monomials, and the reader may assume that a polynomial is always represented as a non-redundant set of monomials
(i.e., $p(x)=2x$ is never represented as $x+x$ or $2x+0x^2$).
The post-fix substitution operator $[a/b]$ may be applied to any sort of expression containing a variable $b$, to substitute $a$ instead; e.g.,
$(x^2+yx+y)[2z/y] = x^2+2zx+2z$.
When discussing a command, state-transition, or program trace, with a variable $\texttt{X}_{i}$, $x_i$ will denote, as a rule, the initial value of this variable,
and $x'_i$ its final value. Thus we distinguish the syntactic entity by the typewriter font.
The parameter $n$ always refers to the number of variables in the subject program. The set $[n]$ is
$\{1,\dots,n\}$.
For a set $S$ an $n$-tuple over $S$ is a mapping from $[n]$ to $S$.
The set of these tuples is denoted by $S^n$.
Throughout the paper, various natural liftings of operators to collections of objects are used, e.g., if $S$ is a set of integers then
$S+1$ is the set $\{s+1 \mid s\in S\}$ and $S+S$ is $\{s+t \mid s,t\in S\}$. We use such lifting with sets as well as with tuples.
If $S$ is ordered, we extend the ordering to $S^n$ by comparing tuples element-wise (this leads to a partial order, in general, e.g., with natural numbers,
$\tuple{1,3}$ and $\tuple{2,2}$ are incomparable).
\begin{defi}
A function of the form $\tuple{ {\vec p}[1], \dots, {\vec p}[n] }$, i.e., an $n$-tuple of polynomials,
is called \emph{a multi-polynomial (MP)}.
We denote by ${\texttt{\upshape MPol}}$ the set of multi-polynomials, namely ${(\mathbb{N}[\vec x])}^n$,where the number of variables $n$ is fixed by context.
\end{defi}
\begin{defi}
A \emph{polynomial transition (PT)} is a computation that transforms an initial state $\vec x = \tuple{x_1,\dots,x_n}$ to a new state
${\vec x}' = \tuple{x'_1,\dots,x'_n} = {\vec p}({\vec x})$ where $\vec p \in {\texttt{\upshape MPol}}$.
\end{defi}
The distinction between a MP and a PT is perhaps a bit philosophical: An MP is a function: a mathematical object that exists independently
of computation. A PT is a computation whose effect is described by such a function. We can say, for example, that the command
$\texttt{X}_2 \verb/:= X/_2+\texttt{X}_1$\ performs a PT\@. Its associated MP is $\tuple{x_1, x_2+x_1}$.
Various operations will be applied to
MPs, mostly obvious---in particular, composition (which corresponds to sequential application of the transitions).
Note that composition of multi-polynomials, ${\vec q} \circ {\vec p}$, is naturally defined since $\vec p$ supplies $n$ values for the $n$
variables of $\vec q$,
and we have: $({\vec p}\circ{\vec q})[i]= {\vec p}[i] \circ {\vec q}$.
We define $\mathit{Id}$ to be the identity transformation, ${\vec x}' = {\vec x}$ (in MP notation: ${\vec p}[i] = x_i$ for $i=1,\dots,n$).
\subsection{Formal semantics of the core language}%
\label{sec:semantics}
The semantics
associates with every command
\verb+C+ over variables $\texttt{X}_1,\dots,\texttt{X}_n$
a relation $\sempar{C} \subseteq \mathbb{N}^n \times \mathbb{N}^n$.
In the expression $\vec x \sempar{C} \vec y$,
vector $\vec x$ (respectively $\vec y$) is the state before (after) the
execution of \verb+C+.
The semantics of {\tt skip} is the identity. The semantics of an assignment
$\texttt{X}_i\verb+:=E+$ associates to each state $\vec x$ a new state $\vec y$ obtained by replacing
the component $x_i$ by the value of the expression \verb+E+ when evaluated over state $\vec x$. This is defined in the
natural way (details omitted), and is denoted by $\sempar{E}{\vec x}$.
Composite commands are described by the straightforward equations:
\begin{eqnarray*}
\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}{\tt C}_1 {\tt ;C}_2\mbox{$\rbrack\hspace{-0.3ex}\rbrack$} &=&
\sempar{C$_2$}\circ\sempar{C$_1$} \\
\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}\verb+choose C+_1\verb+ or C+_2\mbox{$\rbrack\hspace{-0.3ex}\rbrack$} &=&
\sempar{C$_1$}\cup\sempar{C$_2$} \\
\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}\verb+loop E {C}+\mbox{$\rbrack\hspace{-0.3ex}\rbrack$} &=&
\{ (\vec{x},\vec{y}) \mid \exists i \le \sempar{E}{\vec x} :
\vec{x} \sempar{C}^i \vec{y} \}
\end{eqnarray*}
where $\sempar{C}^i$ represents $\sempar{C}\circ\cdots\circ\sempar{C}$
(with $i$ occurrences of $\sempar{C}$); and $\sempar{C}^0 = \mathit{Id}$.
\paragraph{\bf Remarks} The following two changes may enhance the applicability
of the core language for simulating concrete programs; we include
them as ``options'' because they do not affect the validity of our proofs.
\begin{enumerate}
\item
The semantics of an assignment operation may be non-deterministic:
\verb+X := E+ assigns to {\tt X} some non-negative value \emph{bounded} by {\tt E}.
This is useful to abstract expressions which are not in the core language,
and also to use the results of size analysis of subprograms. Such an analysis
may determine invariants such as ``the value of \verb+f(X,Y)+ is at most the sum
of {\tt X} and {\tt Y}.''
The reason that this change does not affect our results is that we work
exclusively with monotone increasing functions.
This assignment semantics is called \emph{a lossy assignment} in~\cite{BK11}, due to an analogy
with lossy counter machines~\cite{Mayr03}.
\item\label{rmrk-ints}
The domain of the integer variables may be extended to $\mathbb{Z}$. In this case
the bounds that we seek are on the absolute value of the output in terms of
absolute values of the inputs. This change does not affect our conclusions
because of the facts $|xy| = |x|\cdot |y|$ and $|x+y| \le |x|+|y|$.
The semantics of the loop command may be defined either as doing nothing if the loop bound is not positive, or using the absolute value as a bound.
\end{enumerate}
\subsection{Detailed statement of the main result}%
\label{sec:goals}
The \emph{polynomial-bound analysis problem} is to find,
for any given command, which output variables
are bounded by a polynomial in the input values (which are simply the values of all variables
upon commencement of the program), and to bound these output values tightly (up to constant factors).
The problem of \emph{identifying} the polynomially-bounded variables is completely solved by~\cite{BJK08}.
We rely on that algorithm, which is polynomial-time, for doing this classification (as further explained in Section~\ref{sec:basic}).
Our main result is thus stated as follows.
\begin{thm}%
\label{thm:mainresult}
There is an algorithm which, for a command \verb+C+, over variables $\texttt{X}_1$ through $\texttt{X}_n$,
outputs a set $\mathcal{B}$ of multi-polynomials, such that the following hold, where ${\texttt{PB}}$ is the set of indices $i$ of variables
$\texttt{X}_i$ whose final value under $\sempar{C}$ is polynomially bounded.
\begin{enumerate}
\item (Bounding) There is a constant $c_{\vec p}$ associated with each ${\vec p}\in \mathcal{B}$, such that
\begin{equation*}
\myforall{\vec x, \vec y} \vec x \sempar{C} \vec y \Longrightarrow \exists {\vec p}\in\mathcal{B} \,.\forall i\in{\texttt{PB}} \,.\, y_i \le c_{\vec p}{\vec p}[i](\vec x)
\end{equation*}
\item (Tightness) For every ${\vec p}\in \mathcal{B}$ there are constants $d_{\vec p}>0$, ${\vec x}_0$ such that
for all ${\vec x}\ge {\vec x}_0$ there is a $\vec y$ such that
\[
\vec x \sempar{C} \vec y \text{ and } \forall i\in{\texttt{PB}} \,.\, y_i \ge d_{\vec p}{\vec p}[i](\vec x)
\]
\end{enumerate}
\end{thm}
\noindent
A remark regarding the decision to provide bounds only up to constant factors: based on the proof,
the algorithm could be extended to explicitly generate the coefficients $c_{\vec p}$ and $d_{\vec p}$ in the above statement; but ignoring
constant factors simplifies the algorithm significantly. In essence, all the considerations regarding constants are removed from the algorithm to the proofs.
Moreover, methodically, our goal was to show that
in the sense that we consider (i.e., ``big Theta'' bounds), the problem could be solved \emph{completely}.
We do not attempt to completely solve the problem of computing bounds with the utmost precise constants.
We assume that readers with a Computer Science background are aware of the ubiquity of ``big Theta'' bounds in the analysis of algorithms and dispense
with a discussion of justifications for using them.
\paragraph{\bf A comment on counters}
As discussed in the introduction, various \emph{countable resource problems}, such as bounding the number of program
steps, may be reduced to the bound analysis problem.
We'd like to point out that the reduction, by introducing a variable to count the events of interest, may be carried out in our core
language, despite the lack of explicit constants. It suffices to reserve a dedicated variable, say \pgt{U}, to represent a unit of the resource.
Then, we advance the counter by adding \pgt{U}, and its worst-case bound will equal the worst-case bound of the resource, times \pgt{U}.
As an example, Figure~\ref{fig:instrumented} shows how the program from Figure~\ref{fig:intro-example}
could be instrumented for computing its time complexity. We use a fresh variable $\texttt{X}_5$ as a ``unit'' and another one $\texttt{X}_6$ to count the steps.
The algorithm will deduce that the final value of $\texttt{X}_6$, $x_6'$, is tightly bounded by $(x_2+x_3+x_1^3 + x_1 x_4)x_5$. Erasing the $x_5$ factor
we obtain a tight bound on the unit-cost time.
\begin{figure}[htb]
\begin{Verbatim}[codes={\catcode`$=3\catcode`_=8}]
loop X$_1$
X$_6$ := X$_6$ + X$_5$;
loop X$_2$ + X$_3$ { X$_6$ := X$_6$ + X$_5$;
choose { X$_3$:= X$_1$; X$_2$:= X$_4$ } or { X$_3$:= X$_4$; X$_2$:= X$_1$ }
};
X$_4$:= X$_2$ + X$_3$
};
loop X$_4$ { X$_6$ := X$_6$ + X$_5$;
choose { X$_3$:= X$_1$ + X$_2$ + X$_3$ } or { X$_3$:= X$_2$; X$_2$:= X$_1$ }
}
\end{Verbatim}
\caption{An instrumented core-language program.}%
\label{fig:instrumented}
\end{figure}
\section{Analysis Algorithm: First Concepts}%
\label{sec:algorithm-prelim
\label{sec:tau-poly}
The following sections describe our analysis algorithm. Naturally, the most intricate part of the analysis concerns loops.
In fact we break the description into stages: first we reduce the problem of analyzing any program to that of analyzing
\emph{simple disjunctive loops}, defined next. Then, we approach the analysis of such loops, which is the main effort in this work.
\begin{defi}%
\label{def:sdl}
A \emph{simple disjunctive loop (SDL)} is a finite set $\mathcal{S}$ of polynomial transitions.
\end{defi}
The loop is ``disjunctive'' because its meaning is that in every iteration, any of the given transitions may be chosen.
A SDL does not specify the number of iterations; our analysis generates polynomials that depend on the number of iterations as well as the initial state. For this purpose, we now
introduce $\tau$-polynomials where $\tau$ represents the number of iterations.
\begin{defi}
$\tau$-polynomials are polynomials in $x_1,\dots,x_n$ and $\tau$.
\end{defi}
If $p$ is a $\tau$-polynomial, then $p(v_1,\dots,v_n)$ is the result of substituting each $v_i$ for the respective $x_i$;
and we also write $p(v_1,\dots,v_n,t)$ for the result of substituting $t$ for $\tau$ as well.
The set of $\tau$-polynomials in $n$ variables ($n$ known from context) is denoted ${\texttt{\upshape $\tau$Pol}}$.
Multi-polynomials and polynomial transitions are formed from $\tau$-polynomials just as previously defined and are used to represent the
effect of a variable number of iterations.
For example, the $\tau$-polynomial transition $\tuple{x'_1,x'_2} = \tuple{x_1,\ x_2+\tau x_1}$ represents the effect of repeating ($\tau$ times) the assignment
$\texttt{X}_2 \verb/:= X/_2+\texttt{X}_1$. The effect of iterating the composite command
$\texttt{X}_2 \verb/:= X/_2+\texttt{X}_1\verb/; X/_3\verb/:= X/_3+\texttt{X}_2$ is described by ${\vec x}' = \tuple{x_1,\ x_2+\tau x_1, \ x_3 + \tau x_2 + \tau^2 x_1}$
(here we already have an upper bound which is not reached precisely, but is correct up to a constant factor).
We denote the set of $\tau$-polynomial transitions by ${\texttt{\upshape $\tau$MPol}}$.
Note that $\tau$ has a special status: as it does not represent a program variable, there is no component in the multi-polynomial for giving its value.
We should note that composition ${\vec q}\circ {\vec p}$ over ${\texttt{\upshape $\tau$MPol}}$ is performed by substituting ${\vec p}[i]$ for each occurrence of $x_i$ in $\vec q$.
Occurrences of $\tau$ are unaffected (since $\tau$ is not part of the state).
\section{Reduction to Simple Disjunctive Loops}%
\label{sec:basic}
We show how to reduce the problem of analysing core-language programs to the analysis of simple disjunctive loops.
\subsection{Symbolic evaluation of straight-line code}
Straight-line code consists of atomic commands---namely assignments (or \texttt{skip}, equivalent to
$\texttt{X}_1\texttt{:=}\, \texttt{X}_1$), composed sequentially. It is obvious that symbolic evaluation of such code leads to polynomial
transitions.
\begin{exa}
Consider the following command:
\begin{center}
$\texttt{X}_2 \verb/:= X/_1\verb/; X/_4 \verb/:= X/_2\verb/ + X/_3\verb/; X/_1 \verb/:= X/_2\verb/ * X/_3$
\end{center}
This is precisely represented by the transition $\tuple{x_1,x_2,x_3,x_4}' = \tuple{ x_1x_3,\ x_1,\ x_3,\ x_1+x_3}$.
\end{exa}
\subsection{Evaluation of non-deterministic choice}
Evaluation of the command \[\texttt{choose}\; \texttt{C}_1 \; \texttt{or} \; \texttt{C}_2\] yields a set of possible
outcomes.
Hence, the result of analyzing a command will be a \emph{set} of multi-polynomial transitions. We express this in the common notation of
abstract semantics:
\[
{\sempar{C}}^{S} \in \wp({\texttt{\upshape MPol}}) \,.
\]
For uniformity, we consider $\sempar{C}^{S}$ for an atomic command to be a singleton in $\wp({\texttt{\upshape MPol}})$.
Composition is naturally extended to sets, and the semantics of a choice command is now simply set union, so we have:
\begin{eqnarray*}
\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}{\tt C}_1 {\tt ;C}_2\mbox{$\rbrack\hspace{-0.3ex}\rbrack$}^{S} &=&
\sempar{C$_2$}^{S}\circ\sempar{C$_1$}^{S} \\
\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}\verb+choose C+_1\verb+ or C+_2\mbox{$\rbrack\hspace{-0.3ex}\rbrack$}^{S} &=&
\sempar{C$_1$}^{S}\cup\sempar{C$_2$}^{S}
\end{eqnarray*}
\begin{exa}
Consider the following command:
\begin{center}
$\texttt{X}_2 \texttt{:= X}_1\texttt{; choose \{ X}_4 \texttt{:= X}_2 \texttt{ + X}_3 \texttt{ \} or \{ X}_1 \texttt{:= X}_2 \texttt{* X}_3 \texttt{ \}}$
\end{center}
This is represented by the set $\{ \tuple{ x_1, x_1, x_3, x_1+x_3}, \tuple{ x_1 x_3, x_1, x_3, x_4 } \}$.
\end{exa}
\subsection{Handling loops}%
\label{sec:solve}
The above shows that any loop-free command in our language can be \emph{precisely} represented by a finite set of MPs,
that is, its semantics $\sempar{C}$ equals the union of a finite number of PTs, which we can compute by symbolic evaluation.
Consequently, the problem of analyzing any loop command is reduced to the analysis of simple disjunctive loops, by first analyzing the loop body.
The goal of the loop analysis is to
produce a set of multi-polynomials that provide
tight worst-case bounds on the results of any number of iterations of the loop body---``tight'' in the sense of Theorem~\ref{thm:mainresult}
(see also Example~\ref{ex:solve} below).
We will refer to this, concisely, as \emph{solving} the loop.
Suppose that we have an algorithm \procSDL that takes a simple disjunctive loop and computes tight bounds for its results, with explicit dependence
on the number of iterations using the parameter $\tau$.
We use it to complete the analysis of any program by the following definition:
\begin{eqnarray*}
\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}\verb+loop E {C}+\mbox{$\rbrack\hspace{-0.3ex}\rbrack$}^{S} &=&
(\procSDL(\sempar{C}^{S}))[\sempar{E}/\tau] \ .
\end{eqnarray*}
Thus, the whole solution is constructed as an ordinary abstract interpretation, following the semantics of the language, except for what happens
inside \procSDL, which is the subtle part of the algorithm; this procedure accepts a set of multi-polynomials, describing the body of the loop,
and generates a set of $\tau$-MPs that finitely describes the result of all execution sequences of the loop, thanks to a \emph{generalization}
procedure that captures the dependence of results on the iteration count by means of the parameter $\tau$. After completing this procedure, we replace
$\tau$ with the expression $\sempar{E}$, which gives the actual loop bound as a polynomial in $x_1,\dots,x_n$.
\begin{exa}%
\label{ex:solve}
Consider the following command:
\begin{center}
$\texttt{X}_4 \texttt{:= X}_1\texttt{; loop X}_4 \texttt{ \{ X}_2 \texttt{:= X}_1\texttt{ + X}_2\texttt{; X}_3 \texttt{:= X}_2 \texttt{\}}$
\end{center}
Solving just the loop yields the set $\mathcal L = \{ \tuple{ x_1, x_2, x_3, x_4}$, $\tuple{ x_1, x_2+\tau x_1, x_2+\tau x_1, x_4 } \}$
(the first MP accounts for zero iterations, the second covers any positive number of iterations).
We can now compute the effect of the given command as:
\begin{align*}
{\mathcal L}[x_4/\tau] \circ \sempar{X$_4$ := X$_1$}^{S} &=
{\mathcal L}[x_4/\tau] \circ \{ \tuple{ x_1, x_2, x_3, x_1} \} \\ &=
\{ \tuple{ x_1, x_2, x_3, x_1}, \tuple{ x_1, x_2+x_1^2, x_2+x_1^2, x_1 } \}.
\end{align*}
We can now demonstrate that the analysis of a loop may involve approximation. Suppose that we slightly modify the code as follows:
\begin{center}
$\texttt{X}_4 \texttt{:= X}_1\texttt{; loop X}_4 \texttt{ \{ X}_3 \texttt{:= X}_2\texttt{; X}_2 \texttt{:= X}_1\texttt{ + X}_2 \texttt{\}}$
\end{center}
Our algorithm will produce the same MPs, although a careful examination shows that the precise result of the loop (when it executes $x_1>0$ iterations)
is now given by
$\tuple{ x_1, x_2+x_1^2, x_2+x_1(x_1-1), x_1 }$. This expression is not a MP (since we only consider polynomials with non-negative coefficients),
but is tightly approximated by the MP we produced, as $x_1(x_1-1) = \Theta(x_1^2)$. The precise statement of what it means to tightly approximate
all possible results of a program by a finite set of MPs is the essence of Theorem~\ref{thm:mainresult}.
\end{exa}
The next section describes the procedure \procSDL, and operates under the assumption that all variables are polynomially bounded in the loop.
However, a loop can generate exponential growth. To cover this eventuality, we first apply the algorithm of~\cite{BJK08} that identifies which
variables are polynomially bounded. If some $\texttt{X}_i$ is \emph{not} polynomially bounded we replace the $i$th component of all the loop transitions
with $x_n$ (we assume $x_n$ to be a dedicated, unmodified variable). Clearly, after this change, all variables are polynomially bounded;
moreover, variables which are genuinely polynomial are unaffected, because they cannot depend on a super-polynomial quantity (given the restricted
arithmetics in our language).
In reporting the results of the algorithm, we should display ``\texttt{super-polynomial}'' instead of any expression that includes $x_n$.
\section{Simple Disjunctive Loop Analysis Algorithm}%
\label{sec:closure-algorithm}
Intuitively, evaluating $\verb/loop E {C}/$ abstractly
consists of simulating any finite number of iterations, i.e., computing
\begin{equation}
\label{eq:closure-seq}
Q_i = \{\mathit{Id}\} \cup P \cup (P\circ P ) \cup \dots \cup {P}^{(i)}
\end{equation}
where $P = \sempar{C}^{S} \in \wp({\texttt{\upshape MPol}})$. The question now is whether the sequence~\eqref{eq:closure-seq} reaches a fixed point.
In fact, it often doesn't. However, it is quite easy to see that in the \emph{multiplicative
fragment} of the language, that is, where the addition operator is not used, such non-convergence is associated with exponential growth. Indeed, since there is no
addition, all our polynomials are monomials with a leading coefficient of 1 (\emph{monic monomials})---this is easy
to verify. It follows that if the sequence~\eqref{eq:closure-seq} does not converge, higher and higher exponents must
appear, which indicates that some variable cannot be bounded polynomially. Taking the contrapositive, we conclude
that if all variables are known to be polynomially bounded the sequence will converge. Thus we have the following easy (and not so satisfying) result:
\begin{obs}
For a SDL $P$ that does not use addition, the sequence $Q_i$ as in~\eqref{eq:closure-seq} reaches a fixed point, and the fixed point provides
tight bounds for all the polynomially-bounded variables.
\end{obs}
\begin{exa}
The following loop is in the multiplicative fragment, and has no exponential behaviour:
\[ \verb/loop X/_3\verb/ { X/_1 \verb/:= X/_2\verb/ * X/_2 \verb/; X/_2 \verb/:= X/_3 \verb/ }/ \]
The effect of an iteration is given by the multi-polynomial ${\vec p} = \tuple{x_2^2,\ x_3,\ x_3}$.
Let $P=\{{\vec p}\}$.
The accumulating effect of the loop is given by
the union of:
\begin{align*}
\{&\mathit{Id}\} \\
P &= \{ \tuple{x_2^2,\ x_3,\ x_3} \} \\
P\circ P &= \{ \tuple{ x_3^2,\ x_3,\ x_3} \} \\
P^{(3)} &= P^{(2)} \\ &\hspace{-0.5em} \dots
\end{align*}
That is, the results are completely characterized by the three MPs above.
\end{exa}
\noindent
When we have addition, we find that knowing that all variables are polynomially bounded does not imply
convergence of the sequence~\eqref{eq:closure-seq}. An example is
$\texttt{loop X}_3\texttt{ \{ X}_1 \texttt{:= X}_1\texttt{ + X}_2 \texttt{ \}}$
yielding the infinite sequence of MPs $\tuple{x_1,x_2,x_3}$, $\tuple{x_1+x_2,x_2,x_3}$, $\tuple{x_1+2x_2,x_2,x_3}$, \dots~%
Our solution employs two means. One is the introduction of $\tau$-polynomials, already presented. The other
is a kind of \emph{abstraction}---intuitively, ignoring the concrete values of (non-zero) coefficients.
Let us first define this abstraction:
\begin{defi}\label{def:abspol}
${\texttt{\upshape APol}}$, the set of abstract polynomials, consists of formal sums of distinct monomials over $x_1,\dots,x_n$,
where the coefficient of every monomial included is $1$.
We extend the definition to an abstraction of $\tau$-polynomials,
denoted ${\texttt{\upshape $\tau$APol}}$.
\end{defi}
\noindent
The meaning of abstract polynomials is given by the following rules:
\begin{enumerate}
\item\label{def:alpha}
The abstraction of a polynomial $p$, $\alpha(p)$, is obtained by modifying all (non-zero) coefficients to 1.
\item Addition and multiplication in ${\texttt{\upshape $\tau$APol}}$ are defined in a natural way so that $\alpha(p)+\alpha(q) = \alpha(p+q)$
and $\alpha(p)\cdot \alpha(q) = \alpha(p\cdot q)$ (to carry these operations out, you just go through the motions of adding or
multiplying ordinary polynomials, ignoring the coefficient values).
\item\label{def:gamma}
The \emph{canonical concretization} of an abstract polynomial, $\gamma({\vec p})$ is obtained by simply regarding it
as an ordinary polynomial.
\item These definitions extend to tuples of (abstract) polynomials in the natural way.
\item The set of abstract multi-polynomials ${\texttt{\upshape AMPol}}$ and their extension with $\tau$ (${\texttt{\upshape $\tau$AMPol}}$) are defined
as $n$-tuples over ${\texttt{\upshape APol}}$ (respectively, ${\texttt{\upshape $\tau$APol}}$). We use AMP as an abbreviation for abstract multi-polynomial.
\item Composition ${\vec p}\acirc {\vec q}$, for ${\vec p}, {\vec q} \in {\texttt{\upshape AMPol}}$ (or ${\texttt{\upshape $\tau$AMPol}}$)
is defined
as $\alpha(\gamma({\vec p})\circ \gamma({\vec q}))$; it is easy to see that one can perform the calculation without the detour through polynomials
with coefficients.
The different operator symbol (``$\acirc$'' versus ``$\circ$'') helps in disambiguating expressions.
\end{enumerate}
An abstract polynomial can be reduced by deleting dominated monomials (e.g., $x^2+x \to x^2$); which clearly
preserves its meaning when viewed in ``big Theta'' terms. We expect that the combinatorial explosion in the algorithm can be attenuated
by reducing every abstract polynomial, but in this paper, we ignore this optimization to alleviate the description and proof of the algorithm.
Another useful optimization (which we also ignore in our analysis) is to delete dominated \emph{multi-polynomials} in a set of MPs.
\paragraph{\bf Analysing a SDL} To analyse a SDL specified by a set of MPs $\mathcal{S}$, we start by computing $\alpha(\mathcal{S})$. The rest of the algorithm
computes within ${\texttt{\upshape $\tau$AMPol}}$.
We define two operations that are combined in the analysis of loops.
The first, which we call \emph{closure}, is simply the fixed point of accumulated iterations as in the multiplicative case. It is introduced by the following
two definitions.
\begin{defi}[iterated composition]%
\label{def:itcomp}
Let $\vec{t}$ be any abstract $\tau$-MP\@. We define ${\vec t}^{\acirc (k)}$, for $k\ge 0$, by:
\begin{align*}
{\vec t}^{\acirc (0)} & = \mathit{Id} \\
{\vec t}^{\acirc (k+1)} & = {\vec t}\acirc {\vec t}^{\acirc (k)} .
\end{align*}
For a set $\mathcal{T}$ of abstract $\tau$-MPs, we define, for $k\ge 0$:
\begin{align*}
\mathcal{T}^{\acirc (0)} & = \{\mathit{Id}\} \\
\mathcal{T}^{\acirc (k+1)} & = \mathcal{T}^{\acirc (k)} \cup \ \bigcup_{ \vec q \in \mathcal{T},\ \vec p \in \mathcal{T}^{\acirc (k)}} {\vec q}\acirc {\vec p} \, .
\end{align*}
\end{defi}
\noindent
Note that ${\vec t}^{\acirc (k)} = \alpha({\gamma(\vec{t})}^{(k)})$, where ${\vec p}^{(k)}$ is defined using ordinary composition.
\begin{exa}
Let $\vec p = \tuple{x_1, x_1+x_2}$; clearly its $k$th iterate (for $k\ge 1$) is ${\vec p}^{(k)} = \tuple{x_1, x_1+kx_2}$. However, with abstract polynomials,
we get ${\alpha({\vec p})}^{\acirc (k)} = \tuple{x_1, x_1+x_2}$; the growth in the coefficient is abstracted away.
Similarly, we have ${\{\alpha(\vec p)\}}^{\acirc (k)} = \{\mathit{Id}, \alpha({\vec p})\}$ for all $k\ge 1$.
\end{exa}
\begin{defi}[abstract closure]
For finite
$P\subset {\texttt{\upshape $\tau$AMPol}}$, we define:
\[
\closure{P} = \bigcup_{i=0}^{\infty} P^{\acirc (i)} \,.
\]
\end{defi}
In the correctness proof, we argue that when all variables are polynomially
bounded in a loop $\mathcal{S}$, the closure of $\alpha(\mathcal{S})$ can be computed in finite time; equivalently, it equals
$\bigcup_{i=0}^{k} {(\alpha({\mathcal{S}}))}^{\acirc (i)}$ for some $k$.
The argument for finiteness is essentially the same as in the multiplicative case.
The second operation is called \emph{generalization} and its role is to capture the behaviour of accumulator
variables, meaning variables that grow by accumulating increments in the loop, and make explicit the dependence on the
number of iterations. The identification of which additive terms in a MP should be considered as increments that accumulate
is at the heart of our problem.
\begin{exa}%
\label{ex:motivate-tau}
Consider the following loop:
\begin{center}
$\texttt{loop \ldots\ \{ X}_1\texttt{:= X}_1\texttt{ + X}_3\texttt{; X}_2\texttt{:= X}_2\texttt{ + X}_3\texttt{ + X}_4\texttt{; X}_4\texttt{:= X}_3\texttt{\}}$
\end{center}
We have omitted the loop bound in order to emphasize that we are now analyzing the \emph{growth of values as the loop progresses} rather
than its final result. The generalization construction below introduces $\tau$'s into the AMPs, obtained from the loop body, in order to express
this growth in terms of the number of iterations. Informally, consider the assignment to $\texttt{X}_1$: it adds $x_3$ (the initial value of variable $\texttt{X}_3$),
and does so on every iteration: so we have an arithmetic series with an increment of $x_3$. After $\tau$ iterations, $\texttt{X}_1$ will have grown by $\tau x_3$.
The assignment to $\texttt{X}_2$ similarly adds the values of two other variables, $\texttt{X}_3$ and $\texttt{X}_4$, but as $\texttt{X}_4$ is rewritten with $x_3$ already at the end of the first
iteration, subsequent iterations add $2x_3$ (not $x_3+x_4$). This will be expressed by introducing the term $\tau x_3$ (the coefficient 2 is abstracted away).
We return to this example, and the issue it illustrates, further below.
\end{exa}
The definition of ${\vec p}^\tau$ will is greatly simplified by concentrating on idempotent AMPs, defined next. We will later give
an example that shows that our definition does not work for AMPs which are not idempotent; while the fact that a complete solution can
be obtained by generalizing only idempotent elements comes out of a Ramsey-like property of finite monoids~\cite{Simon:TCS:90}, as will be
seen in Section~\ref{sec-proof-ub}.
\begin{defi}
$\vec p \in {\texttt{\upshape $\tau$AMPol}}$ is called \emph{idempotent} if $\vec p \acirc \vec p = \vec p$.
\end{defi}
\noindent
Note that this is composition in the abstract domain. So, for instance, $\tuple{x_1, x_2}$ is idempotent, and so is
$\tuple{x_1 + x_2, x_2}$, while $\tuple{x_1x_2, x_2}$ and $\tuple{x_1 + x_2, x_1}$ are not.
\begin{defi}
For $\vec p$ an (abstract) multi-polynomial,
we say that $x_i$ is \emph{self-dependent} in $\vec p$ if $\vec p[i]$ depends on $x_i$.
We also say that the entry $\vec p[i]$ is self-dependent; the choice of term depends on context and the meaning should be clear either way.
We call a monomial self-dependent if all the variables appearing in it are.
We denote by $\sd{\vec p}$ the set $\{i \,:\, x_i \text{ is self-dependent in } \vec p\}$.
\end{defi}
We later show that in polynomially-bounded loops, if $x_i$ is self-dependent then $\vec p[i]$ must include the monomial $x_i$. To illustrate the
significance of self-dependent monomials, let us consider an
example where
${\vec p}[1] = x_1 +\dots$ and ${\vec p}[2] = x_2 + \dots$, and further ${\vec p}[3] = x_1 x_2 + \dots$. Then the monomial $x_1 x_2$ reappears in
every iterate ${\vec p}^{\acirc (k)}$. This is because all the variables in it are self-dependent.
\begin{defi}%
\label{def:MPnotation}
We define a notational convention for $\tau$-MPs, specifically for self-dependent entries of the MP\@. Assuming that $x_i$ appears in ${\vec p}[i]$, we write:
\[ {\vec p}[i] = x_i + \tau {\vec p}[i]' + {\vec p}[i]''+ {\vec p}[i]''' \,,\]
where ${\vec p}[i]'''$ includes all the non-self-dependent monomials of ${\vec p}[i]$, while the self-dependent monomials
(other than $x_i$) are grouped into two sums:
$\tau{\vec p}[i]'$, including all monomials with a positive degree of $\tau$, and ${\vec p}[i]''$ which includes all the $\tau$-free monomials.
\end{defi}
\begin{exa}%
\label{ex:MPnotation}
Let $\vec p = \tuple{ x_1 + \tau x_2 + \tau x_3 + x_3 x_4,\ x_3,\ x_3,\ x_4 }$. Then $\sd{\vec p} = \{1,3,4\}$.
Since $x_1$ is self-dependent, we will apply the above definition to ${\vec p}[1]$, so that
${\vec p}[1]' = x_3$, ${\vec p}[1]'' = x_3 x_4$ and ${\vec p}[1]''' = \tau x_2$.
Note that a factor of $\tau$ is stripped in ${\vec p}[1]'$. Had the monomial been $\tau^2 x_3$, we would have ${\vec p}[1]' = \tau x_3$.
\end{exa}
\begin{defi}[generalization]
Let $\vec p$ be idempotent in ${\texttt{\upshape $\tau$AMPol}}$; define
${\vec p}^\tau$ by: \[
{\vec p}^\tau [i] = \begin{cases}
x_i + \tau {\vec p}[i]' + \tau{\vec p}[i]''+ {\vec p}[i]''' & \text{if $i\in \sd{\vec p}$} \\
{\vec p}[i] & \text{otherwise.}
\end{cases}
\]
\end{defi}
\noindent
Note that the arithmetic here is abstract (see examples below). Note also that in the term $\tau {\vec p}[i]'$ the $\tau$ is already present in $\vec p$,
while in $\tau {\vec p}[i]''$ it is added to existing monomials. In this definition, the monomials of ${\vec p}[i]'''$ are treated like those of $\tau {\vec p}[i]'$;
however, in certain steps of the proofs we will treat them differently, which is why the notation separates them. We next give a couple of examples, just
to illustrate the definition; we later (Page~\pageref{cog}) motivate the definition in the context of the loop analysis.
\begin{exa}%
\label{ex:tauisation:1}
Let $\vec p = \fourtuple{ x_1 + x_3}{ x_2 + x_3 + x_4}{ x_3}{ x_3 }$. Note that it corresponds to the loop body from Example~\ref{ex:motivate-tau}.
Further note that ${\vec p}\acirc {\vec p} = {\vec p}$, i.e., $\vec p$ is idempotent.
We have ${\vec p}^\tau = \fourtuple{ x_1 + \tau x_3}{ x_2 + \tau x_3+x_4}{ x_3}{ x_3 }$.
\end{exa}
\begin{exa}\label{ex:tauisation:2}
Let $\vec p = \tuple{ x_1 + \tau x_2 + \tau x_3 + \tau x_3 x_4,\ x_3,\ x_3,\ x_4 }$.
Note that ${\vec p}\acirc {\vec p} = {\vec p}$.
The self-dependent variables are all but $x_2$.
We have
${\vec p}^\tau = \tuple{ x_1 + \tau x_2 + \tau x_3 + \tau x_3 x_4,\ x_3,\ x_3,\ x_4 } = {\vec p}$.
\end{exa}
\noindent
Finally we can present the analysis of the loop command.
\begin{algo} \procSDL{}$(\mathcal{S})$ \\
Input: $\mathcal{S}$, a polynomially-bounded disjunctive simple loop. \\
Output: a set of $\tau$-MPs that tightly approximates the effect of $\tau$ iterations of loop $\mathcal{S}$.
\end{algo}
\begin{enumerate}
\item Set $T = \alpha(\mathcal{S})$.
\item Repeat the following steps until $T$ remains fixed:
\begin{enumerate}
\item Closure: Set $T$ to $\closure{T}$.
\item Generalization: For all ${\vec p}\in T$ such that ${\vec p}\acirc {\vec p} = {\vec p}$, add ${\vec p}^\tau$ to $T$.
\end{enumerate}
\end{enumerate}
\begin{exa}
Consider the following loop:
\begin{center}
$\texttt{loop \ldots\ \{ X}_1\texttt{:= X}_1\texttt{ + X}_2\texttt{; X}_2\texttt{:= X}_2\texttt{ + X}_3\texttt{; X}_4\texttt{:= X}_3\texttt{ \}}$
\end{center}
The body of the loop is evaluated symbolically and yields the multi-polynomial:
\[
\vec p = \fourtuple{ x_1+x_2}{ x_2+x_3}{ x_3}{ x_3 }
\]
Now, computing within ${\texttt{\upshape AMPol}}$,
\begin{align*}
{\alpha({\vec p})}^{\acirc (2)} &= \alpha({\vec p}) \acirc \alpha({\vec p}) = \fourtuple{ x_1+ x_2+x_3}{ x_2+ x_3}{ x_3}{ x_3 } ; \\
{\alpha({\vec p})}^{\acirc (3)} &= {\alpha({\vec p})}^{\acirc (2)} \,.
\end{align*}
Here the closure computation stops.
Since $\alpha( {\vec p}^{\acirc (2)} )$ is
idempotent, we compute
\[
{\vec q} = {({\alpha({\vec p}) }^{\acirc (2)})}^\tau = \fourtuple{ x_1+\tau x_2 + \tau x_3 }{ x_2+\tau x_3 }{ x_3}{ x_3 }
\]
and applying closure again, we obtain
\[\begin{array}{ll}
{\vec q} \acirc {\alpha({\vec p}) } &= \fourtuple{ x_1+ x_2 + x_3 + \tau x_2 + \tau x_3 }{ x_2+ x_3 +\tau x_3 }{ x_3}{ x_3 } \\
{({\vec q})}^{\acirc (2)} &= \fourtuple{ x_1+ \tau x_2 + \tau x_3 + \tau^2 x_3 }{ x_2 +\tau x_3 }{ x_3}{ x_3 } \\
\end{array}
\]
where the first one simplifies to $\vec q$ by deleting dominated terms, and the second to
$\fourtuple{ x_1+ \tau x_2 + \tau^2 x_3 }{ x_2 +\tau x_3 }{ x_3}{ x_3 }$.
The last element is idempotent but applying generalization does not generate anything new. Thus the algorithm ends.
The reader may reconsider the source code to verify that we have indeed obtained tight bounds for the loop.
\end{exa}
Note that when a program contains nested loops, innermost loops will be processed first and result in a set of abstract polynomials, so we might actually
analyze any enclosing commands entirely in the abstract domain. This means rephrasing our definition of SDL by defining the input set as a set of AMPs rather
than concrete polynomials; then the initial step above, where $\mathcal{S}$ is abstracted, is skipped. In the forthcoming sections we ignore this issue and
continue to treat the input as being a set of concrete polynomials; this is for convenience only. It is easy enough to restate the results such that the input
is understood as abstract, but known to tightly describe the effect of some concrete piece of code.
\paragraph{\bf Comments on generalization}%
\label{cog}
The precise definition of the generalization operator has been one of the key steps in the development of this
algorithm (operators which are similar---but insufficient for our purpose---appear in related work~\cite{KasaiAdachi:80,
KN04,NW06,JK08,BJK08,Giesl:toplas2016}). Let us attempt to give some intuition for its definition (not a correctness proof, yet). Say $x_i$ is
self-dependent. An important point is the
partition of the terms added to $x_i$ into $\tau {\vec p}[i]'$, ${\vec p}[i]''$ and ${\vec p}[i]'''$. To see why
the non-self-dependent monomials ${\vec p}[i]'''$ are not multiplied by $\tau$, consider Example~\ref{ex:tauisation:1}. Had we been too eager to
insert $\tau$'s and do this to non-self-dependent monomials as well, we would have got:
\[
\fourtuple{ x_1 + \tau x_3}{ x_2 + \tau x_3 + \tau x_4}{ x_3}{ x_3 } \,
\]
(differing from the correct result in the second component). This would give a sound, but loose (overestimated) upper bound: as already discussed,
when the transition represented by $\vec p$ is iterated, copies of the initial value $x_4$ do \emph{not} accumulate. Next, the distinction of
${\vec p}[i]''$ from $\tau {\vec p}[i]'$ prevents $\tau$'s from being added where they already appear; this is important because we reapply generalization
to results of previous steps while analyzing a given loop.
Finally, an important aspect particular to our algorithm (compared to the above-cited) is the application of the generalization
operator only to idempotent elements. Again, we can illustrate with an example that applying it too eagerly (to all AMPs in the closure) would be
incorrect. Consider the following AMP:\@
\[
\vec p = \tuple{ x_1 + x_2 + x_3 + x_4,\ x_3,\ x_4,\ x_4 }
\]
Generalization (incorrectly applied to $\vec p$, since it is not idempotent) would yield:
\[
\fourtuple{ x_1 + x_2 + x_3 + \tau x_4}{ x_3}{ x_4}{ x_4 }
\]
This is bad because if $\vec p$ is iterated to accumulate copies of $x_4$ in $\texttt{X}_1$ (a behaviour represented by the term $\tau x_4$), the value of $\texttt{X}_2$ will no longer be $x_3$ (but $x_4$). Hence, this AMP
describes a result that is not realizable (and when we substitute the loop bound for $\tau$, we will get a loose upper bound).
\paragraph{\bf A comment on linear bounds}
When values accumulate in a loop, a non-linear result is obtained. Stated contrapositively: linear results do not involve accumulation. Indeed,
it is not hard to verify that if for linear polynomials (only) we maintain precise numeric coefficients, our algorithm would still converge. Therefore, if we wish,
we can modify our algorithm so that \emph{for linear results we have bounds in which the coefficients are explicit and precise}.
This is not a strong result as might seem at first, since it is due to the weakness of our language: linear results can only be built up by a finite number of additions.
Still, if the algorithm is employed as a brick in a larger system, this property might possibly be useful.
\section{Correctness Statement for Simple Disjunctive Loops}%
\label{sec:correctness-sdl}
We claim that our algorithm obtains a description of the worst-case results of the program that is
precise up to constant factors. That is, we claim that the set of MPs returned provides a ``big O'' upper bound (on all executions) which is also tight;
tightness means that every MP returned is also a lower bound (up to a constant factor) on an infinite sequence of possible executions.
The main, non-trivial part of the algorithm is of course the solution of a simple disjunctive loop, procedure \procSDL.
Completing this to show correctness for an arbitrary program is not difficult.
In this section we formulate the \emph{correctness statement for Simple Disjunctive Loops}, in other words, we state the contract
for procedure \procSDL, which, when satisfied, justifies its proper functioning within the whole algorithm.
This formulation forces us to step up the level of detail; specifically, we introduce \emph{traces}, to give a more concrete semantics to loops
(compared to Section~\ref{sec:semantics}).
\begin{defi}%
\label{def:trace}
Let $\mathcal{S}$ be a set of polynomial transitions.
An \emph{(abstract) trace} over $\mathcal{S}$ is a finite
sequence $\sigma = {\vec p}_1;\dots; {\vec p}_{|\sigma|}$ of elements of $\mathcal{S}$.
Thus $|\sigma|$ denotes the \emph{length} of the trace.
The set of all traces is denoted $\tseqs{\mathcal S}$.
We write $\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}{\sigma}\mbox{$\rbrack\hspace{-0.3ex}\rbrack$}$ for the composed relation
$ {\vec p}_{|\sigma|} \circ \dots\circ {\vec p}_1$ (for the empty trace, ${\varepsilon}$, we have $\mbox{$\lbrack\hspace{-0.3ex}\lbrack$}{{\varepsilon}}\mbox{$\rbrack\hspace{-0.3ex}\rbrack$} = \mathit{Id}$).
\end{defi}
\noindent
Using the following definition we will be able to give the desired correctness statement for \procSDL.
\begin{defi}%
\label{def:unimono}
Let $p(\vec x)$ be a (concrete or abstract) $\tau$-polynomial. We write $\unimono{p}$ for the \emph{linear monomials} of $p$, namely any one of
the form $ax_i$ for a constant coefficient $a$.
We write $\nunimono{p}$ for the rest. Thus $p = \unimono{p}+\nunimono{p}$.
\end{defi}
\begin{thm}[Solution of disjunctive loop problem]%
\label{thm:sdlp}
Given a polynomially-bounded SDL
represented as a set $\mathcal{S}$ of MPs, procedure \procSDL finds
a finite set $\mathcal{B}$ of $\tau$-MPs which \emph{tightly bound} all traces over $\mathcal{S}$. More precisely,
it guarantees:
\begin{enumerate}
\item (Bounding) There is a constant $c_{\vec p} > 0$ associated with each ${\vec p}\in \mathcal{B}$, such that
\begin{equation*}
\myforall{\vec x, \vec y, \sigma} \vec x \mbox{$\lbrack\hspace{-0.3ex}\lbrack$}\sigma\mbox{$\rbrack\hspace{-0.3ex}\rbrack$} \vec y \Longrightarrow \exists {\vec p}\in\mathcal{B} \,.\, {\vec y} \le c_{\vec p}{\vec p}(\vec x, |\sigma|)
\end{equation*}
\item (Tightness) For every ${\vec p}\in \mathcal{B}$ there are constants $d_{\vec p}>0$, ${\vec x}_0$
such that for all ${\vec x}\ge {\vec x}_0$ there are a trace $\sigma$ and a state vector $\vec y$ such that
\[
\vec x \mbox{$\lbrack\hspace{-0.3ex}\lbrack$}\sigma\mbox{$\rbrack\hspace{-0.3ex}\rbrack$} \vec y \, \land \, {\vec y} \ge
\unimono{\vec p}({\vec x}, |\sigma|)+ d_{\vec p}\nunimono{\vec p}({\vec x}, |\sigma|)
\,.
\]
\end{enumerate}
\end{thm}
\noindent
Note that in the lower-bound clause (2), the linear monomials of $p$ are not multiplied by the constant $d_{\vec p}$; this
sets, in a sense, a stricter requirement for them: if the trace maps $x$ to $x^2$ then the bound
$2x^2$ is acceptable, but if it maps $x$ to $x$, the bound $2x$ is not accepted. The reader may understand this technicality by considering the
effect of iteration: it is important to distinguish the transition $x'_1 = x_1$, which can be iterated ad libitum, from the transition
$x'_1 = 2x_1$, which produces exponential growth on iteration. Distinguishing $x'_1 = x_1^2$ from $x'_1 = 2x_1^2$ is not as important.
We remark that $c_{\vec p}, d_{\vec p}$ range over real numbers. However, our data and the coefficients of polynomials remain integers, it is only such
comparisons that are performed with real numbers (specifically, to allow $d_{\vec p}$ to be smaller than one; in the upper bounds, it is possible to stick
to integers).
Note also that polynomial boundedness is ensured by our algorithm
before applying the procedure (Section~\ref{sec:solve}), so the precondition of the correctness theorem is satisfied.
\section{Correctness Proofs}%
\label{sec:correctness-main}
In this section we prove the correctness of the main, non-trivial part of our algorithm, namely the solution of a simple disjunctive loop,
showing that it satisfies the requirements set forth in the last section.
\paragraph{\textbf{Overview of the proof}}
Intuitively, what we want to prove is that the multi-polynomials we compute cover all ``behaviors'' of the loop. More precisely, in the upper-bound part of the
proof we want to cover all behaviors: upper-bounding is a universal statement. To prove that bounds are tight, we show that each such bound constitutes a
\emph{lower bound} on a certain ``worst-case behavior'': tightness is an existential statement. The main aspects of these proofs are as follows:
\begin{itemize}
\item A key notion in our proofs is that of \emph{realizability}. Intuitively, when we come up with a bound, we want to show that there are traces that achieve (realize) this bound for arbitrarily large input values.
\item In the lower-bound proof, we describe a set of traces by a \emph{pattern}. A pattern is constructed like a regular expression
with concatenation and Kleene-star.
However, they allow no nested iteration constructs, and when expanding a pattern into a concrete trace,
the starred sub-expressions have to be repeated the same number of times; for example,
the pattern ${\vec p}^*{\vec q}^*$ generates the traces $\{ {\vec p}^t{\vec q}^t,\ t\ge 0\}$.
The proof constructs a pattern for every
multi-polynomial computed, showing that it is realizable. It is interesting that such simple patterns suffice to establish tight lower bounds for all loops in our class.
\item In the upper-bound proof, we describe all traces by a finite set of \emph{well-typed regular expressions}~\cite{Bojanczyk:DLT:09}.
This elegant tool channels the power of the Factorization Forest Theorem~\cite{Simon:TCS:90}; this theorem
exposes the role of idempotent elements, which is key in our algorithm.
\item Interestingly, the lower-bound proof not only justifies the tightness of our upper bounds, it also justifies the termination of the algorithm and
the application of the Factorization Forest Theorem in the upper-bound proof, because it shows that our abstract multi-polynomials generate a finite monoid.
\end{itemize}
\noindent
Recall that traces are just sequences of polynomial transitions (Definition~\ref{def:trace}).
Next, we define \emph{concrete traces}, which represent executions on concrete data. The following definition assumes that the steps of the trace are specified by
$\tau$-free multi-polynomials; it suffices for the first part of the proof (Section~\ref{sec-proof-lb}).
\begin{defi}[Concrete traces (unweighted)]
A \emph{concrete trace} corresponding to $\sigma \in \tseqs{\mathcal S}$ is a
path of labeled arcs in the state space $\mathbb{N}^n$:
\[ \concrete{\sigma} = \trans{{\vec p}_1}{s_{0}}{s_1}\dots\trans{{\vec p}_t}{s_{t-1}}{s_t},\]
where $s_{i+1} = {\vec p}_{i+1}(s_i)$.
We write $ \tseq{{s_0}}{{\sigma}}{{s_t}}$ to indicate just the initial and final states of a trace (as a special case, the empty trace ${\varepsilon}$
corresponds to a path with no arcs: $\tseq{s_0}{{\varepsilon}}{s_0}$).
The set of all concrete traces is denoted $\concrete{\tseqs{\mathcal S}}$.
\end{defi}
\noindent
Note that the semantics of polynomial transitions never block a state. That is, given $\sigma$ and $s_0$ there always is a $\concrete{\sigma}$ starting
with $s_0$.
\paragraph{\textbf{Notations}}
\begin{itemize}
\item
Concatenation of traces $\sigma, \rho$
is written as $\sigma \concat \rho$. For concrete traces, concatenation requires the final state of $\sigma$ to be the initial state of $\rho$,
assuming both are non-empty.
\item
In the proofs, we handle the abstract polynomials computed by the algorithm as if they were concrete polynomials.
This should be understood as an ``implicit cast,'' applying the $\gamma$ conversion function (see Page~\pageref{def:gamma}).
In fact, it is useful to bear in mind that
the correctness claim---namely, Theorem~\ref{thm:sdlp}---is a claim about concrete MPs, relating their numeric values to values generated by program executions.
We mostly omit $\alpha$'s and $\gamma$'s and the reader should interpret a MP as concrete or abstract according to context. For example,
comparison of value $p \ge q$ applies to concrete polynomials, while the statement that $\vec p$ is idempotent indicates that $\vec p$ is
(or should be ``cast'' into) an abstract MP\@.
\item
We call a $\tau$-MP $\vec p$ $\tau$-closed if ${\vec p}^\tau = {\vec p}$.
\end{itemize}
\subsection{Lower-Bound Correctness for Simple Loops}%
\label{sec-proof-lb}
The key notion in proving that the upper bounds that we compute are tight---equivalently, that they provide lower bounds (up to constant factors) on the
worst-case results---is that of \emph{realizability}. Intuitively, when we come up with a bound, we want to show that there are traces which achieve (realize) this bound.
Importantly, with asymptotic analysis, a bound is not justified, in general, by showing a \emph{single} trace;
what we need is a \emph{pattern} that generates arbitrarily long traces.
Formally, we define the class of \emph{patterns} (over a given set of polynomial transitions, $\mathcal S$)
and their associated languages (sets of traces). The following statements define patterns $\pi$ along with corresponding sets of languages, $L(\pi,t)$.
The role of $t$ is related to loop counts; it tells how many times to repeat $\vec a$ in a pattern ${\vec a}^*$.
\begin{itemize}
\item The empty string ${\varepsilon}$ is also a pattern. It generates the language $L({\varepsilon},t) = \{{\varepsilon}\}$, consisting of the empty trace.
\item A single MP, ${\vec p}\in \mathcal S$, is a pattern. It generates the language $L({\vec p},t) = \{ {\vec p} \}$, consisting of a single trace.
\item A concatenation of patterns is a pattern, and $L(\pi_1 \pi_2, t) \stackrel{def}{=\;} L(\pi_1, t) L(\pi_2, t)$, where the right-hand side applies concatenation to traces. \\
We define concatenation of patterns to be associative, so ${\vec p}({\vec q}{\vec r})$
and $({\vec p}{\vec q}){\vec r}$ are the same pattern, which may be written ${\vec p}{\vec q}{\vec r}$ (this works because concatenation of traces is also associative).
\item If $\pi$ is a pattern, $\pi^*$ is a pattern. However, nested application of the star is not allowed; in other words, $\pi$ is required to be star-free.
We define $L(\pi^*, t) = {L(\pi)}^t$, where $L^0 = \{{\varepsilon}\}$ and $L^{t+1} = L^t\cdot L$. \\
Parentheses are used for syntactic disambiguation of the operand of the star, e.g., ${\vec p}{({\vec q}{\vec r})}^*$.
\end{itemize}
\noindent
We use the notation $\pi^n$ as a shorthand for a concatenation of $n$ copies of $\pi$.
The set of traces corresponding to pattern $\pi$, denoted $L(\pi)$,
is defined by $L(\pi) = \bigcup_{t\ge 0} L(\pi,t)$.
E.g., $L(a^*b^*) = \{ a^n b^n \mid n\ge 1 \}$.
We denote by $\pi(n)$ the result of substituting $n$ for all the stars in $\pi$ (obtaining a star-free pattern).
In particular, $\pi(1)$ is the result of substituting 1 for all the stars in $\pi$. Since nesting of iterative expressions is not allowed,
$|\pi(t)| = \Theta(t)$.
\begin{defi}[Realizability]%
\label{def:realized}
A polynomial transition (PT) ${\vec p}\in {\texttt{\upshape $\tau$MPol}}$ is said to be \emph{realizable} over the given set $\mathcal{S}\subset {\texttt{\upshape MPol}}$ if there is a pattern $\pi$
and a constant $0 < c \le 1$,
such that for all $t\ge 1$, for all $\sigma\in L(\pi(t))$, if $\tseq{\vec{x}}{{\sigma}}{\vec{y}}$, then%
\begin{equation} \label{eq:realization}
\vec y \ge \unimono{\vec p}({\vec x}, t) + c\nunimono{\vec p}({\vec x}, t) \,.
\end{equation}
We say that $\vec p$ is realized by $\pi;c$, or that it is $c$-realizable. For $\tau$-free MPs, we use the same definition but $\pi$ should not
include a star, so $t$ may be omitted. Thus a $\tau$-free MP is to be realized by a single abstract trace.
A set of MPs is called realizable if all its members are.
\end{defi}
\begin{exa}
We reconsider the loop in Example~\ref{ex:motivate-tau}. In Example~\ref{ex:tauisation:1} we have computed from its body the $\tau$-MP
${\vec q} = {\vec p}^\tau = \fourtuple{ x_1 + \tau x_3}{ x_2 + \tau x_3+x_4}{ x_3}{ x_3 }$. We claim that it is realizable. We use the simple pattern $\pi = {\vec p}^*$.
The final state after $t\ge 1$ iterations (a trace in $L(\pi(t))$) is
\[
{\vec y} = \fourtuple{ x_1 + t x_3 }{ x_2 + (2t-1)x_3 + x_4 }{ x_3 }{ x_3 },
\]
which is easy to verify by inspection of the program. We now verify that
\[
{\vec y} \ge \unimono{{\vec q}}({\vec x}, t) + c\nunimono{{\vec q}}({\vec x}, t),
\]
where $c = 1$.
\end{exa}
Note that realizability is a monotone property in the sense that if ${\vec p} \ge {\vec q}$ and $\vec p$ is realizable, then $\vec q$ is.
This is natural, since we are arguing that $\vec p$ is a \emph{lower bound}.
Note also that in contrast to Clause~2 of Theorem~\ref{thm:sdlp}, we pass $t$ rather than $|\sigma|$ as the $\tau$ parameter to the polynomial bound.
However, since $|\pi(t)| = \Theta(t)$, the inequality in the theorem will be satisfied, just with a different constant factor.%
\footnote{This consideration only applies to the non-linear part $\nunimono{\vec p}$, since $\unimono{\vec p}$ is $\tau$-free anyway.}
\noindent
We introduce a short form that makes inequalities such as~\eqref{eq:realization} easier to manipulate:
\begin{defi}
For any $\tau$-polynomial $p$, and real number $c \le 1$,
\[
c\mnl p({\vec x},t) \stackrel{def}{=\;} \unimono{p}({\vec x}, t) + c\nunimono{p}({\vec x}, t) \,.
\]
We also write $c\mnl {\vec p}$ for component-wise application of ``$c\mnl$'' to a multi-polynomial $\vec p$.
\end{defi}
\noindent
It is useful to be aware of properties of this operation. We leave the verification of the next lemma to the reader.
\begin{lem}%
\label{lem:mnl-properties}
Let $r\in {\texttt{\upshape $\tau$Pol}}$, ${\vec p}\in {\texttt{\upshape $\tau$MPol}}$, and $c\le 1$. Then we have
\begin{align*}
(c\mnl r ) \circ {\vec p} &\ge c\mnl (r\circ {\vec p}) \\
r \circ (c\mnl {\vec p}) &\ge c^{\deg r}\mnl (r\circ {\vec p}) \,.
\end{align*}
\end{lem}
\begin{lem}
Let ${\vec p},{\vec q} \in {\texttt{\upshape $\tau$AMPol}}$. Then ${\vec p}\circ {\vec q} \ge {\vec p}\acirc {\vec q}$.
\end{lem}
\noindent
Note that on the left-hand side of the inequality, we compose in ${\texttt{\upshape $\tau$MPol}}$, while on the right-hand side, we compose in ${\texttt{\upshape $\tau$AMPol}}$. A more
explicit expression of this claim would be:
\[\gamma({\vec p})\circ \gamma({\vec q}) \ge \gamma ( {\vec p}\acirc {\vec q} ). \]
\noindent
Now this becomes obvious, since both sides of the equality have the same monomials (they are abstractly the same), and
the coefficients on the right-hand side are all 1, while on the left-hand side they are integers.
The fact that both sides have corresponding monomials allows us to strengthen the statement
to the following:
\begin{lem}%
\label{lem:circ-vs-acirc}
Let ${\vec p},{\vec q} \in {\texttt{\upshape $\tau$AMPol}}$ and $c\le 1$. Then $c\mnl({\vec p}\circ {\vec q}) \ge c\mnl({\vec p}\acirc {\vec q})$.
\end{lem}
\noindent
Recall that we aim to prove that all PTs computed by our algorithm are realizable. We do this by a series of lemmas which shows that realizability is preserved
by the various constructions in our algorithm. Note that our algorithm produces abstract MPs; so, let us clarify that when we say that ${\vec p}\in{\texttt{\upshape AMPol}}$
is realizable, Definition~\ref{def:realized} is actually applied to $\gamma({\vec p})$.
\subsubsection{Realizability without generalization}
The realizability of PTs computed in the closure step will follow quite easily using a few simple lemmas.
\begin{lem}%
\label{lem:idppol-realizable}
$\mathit{Id}$ is realizable.
\end{lem}
\begin{proof}
$\mathit{Id}$ is realized by the empty pattern.
\end{proof}
\begin{lem}
Every member of $\mathcal{S}$ is realizable.
\end{lem}
\begin{proof}
${\vec p}\in \mathcal{S}$ is realized by the pattern $\vec p$.
\end{proof}
\begin{lem}%
\label{lem:comp-realizable}
Let $\vec p$, $\vec q$ be realizable $\tau$-MPs. Then ${\vec q}\circ {\vec p}$ is realizable.
\end{lem}
\begin{proof}
Suppose that $\vec p$ is realized by $\pi_1;c_1$ and $\vec q$ by $\pi_2;c_2$. We claim that
${\vec q}\circ {\vec p}$ is realized by $\pi_1 \pi_2 ; c'$ for some $c'$. Consider a concrete transition sequence $\widetilde \sigma$ with
$\sigma\in L((\pi_1\pi_2)(t))$, it has the form $\tseq{\vec x}{\sigma_1}{\vec y}\tseq{\:}{\sigma_2}{\vec z}$ with
$\sigma_i \in L(\pi_i(t))$.
We have:
\begin{align*}
\vec y & \ge c_1\mnl{\vec p}({\vec x}, t) \\
\vec z & \ge c_2\mnl{\vec q}({\vec y}, t),
\end{align*}
consequently, using Lemma~\ref{lem:mnl-properties}:
\[
\vec z \ge c_1^d c_2\mnl({{\vec q}\circ {\vec p}})({\vec x}, t)
\]
where $d$ is the highest degree of any component of $\vec q$.
\end{proof}
\begin{lem}\label{lem:acomp-realizable}
Let ${\vec p}, {\vec q}\in {\texttt{\upshape $\tau$AMPol}}$ be realizable. Then ${\vec q}\acirc {\vec p}$ is realizable.
\end{lem}
\begin{proof}
This follows from the previous lemma and Lemma~\ref{lem:circ-vs-acirc}.
\end{proof}
\begin{cor}
Let $T$ be a set of realizable abstract $\tau$-MPs. Then every member of $\closure{T}$ is realizable.
\end{cor}
\subsubsection{Dependence graphs and neat MPs}
We are still missing a realizability lemma for the generalization operation.
As a preparation for this proof, we
introduce the dependence graph of a PT and state an important structural property of dependency graphs associated with idempotent transitions.
This leads to the definition of the class we call {\em neat MPs}.
\begin{defi}
Let ${\vec p} \in {\texttt{\upshape $\tau$MPol}}$. Its \emph{dependency graph} $G({\vec p})$ is
a directed graph with node set $[n]$. The graph includes an arc $i\to j$ if and only if
${\vec p}[j]$ depends on $x_i$.
\end{defi}
\noindent
Intuitively, $G({\vec p})$ shows the data flow in the transition $\vec p$. It is easy to see that paths in the graph correspond to data-flow
effected by a sequence of transitions. For instance, if we have a path $i\to j \to k$ in $G({\vec p})$, then $x_i$ will appear in the expression
$({\vec p}\circ {\vec p})[k]$.
\begin{lem}\label{obs:no-doublin}
Suppose that the SDL $\mathcal{S}$ is polynomially bounded.
Then for all ${\vec p}\in {\texttt{\upshape $\tau$MPol}}$ which is realizable over $\mathcal{S}$, ${\vec p}[i]$ has no monomial divisible by $x_i$ other than $x_i$.
\end{lem}
\begin{proof} Suppose that $\vec p$ has such a monomial, then it has the form ${\mathfrak m} x_i$.
If $\mathfrak m$ is $\tau$-free, then
starting with a state in which all variables have
values greater than 1,
and repeatedly executing a trace that realizes $\vec p$,
we clearly obtain an exponential growth of the value in $\texttt{X}_i$, contradicting the assumption.
In the case that $\tau$ occurs in $\mathfrak m$, suppose that $\vec p$ is realized by $\pi;c$. Choose $t \ge \max(c^{-1}, 2)$.
Then repeating the trace $\pi(t)$ creates an exponential growth in $\texttt{X}_i$.
\end{proof}
\begin{lem}%
\label{lem:nearly-dag}
Suppose that ${\vec p}\in {\texttt{\upshape $\tau$MPol}}$ is realizable, and $\alpha({\vec p})$ is
idempotent. Assuming that the loop under analysis is polynomially bounded,
$G({\vec p})$ does not have any simple cycle longer than one arc. In other words,
it consists of a directed acyclic graph (DAG) plus some self-loops.
\end{lem}
\begin{proof}
Assume, to the contrary, that $G({\vec p})$ does have a cycle $i\to \dots \to k \to i$, where $k\ne i$. Let $r$ be the length of the cycle. Then
\begin{enumerate}
\item\label{stp1}
${\vec p}^{(r)}[i]$ depends on $x_i$; we can write ${\vec p}^{(r)}[i] = x_i + p(\vec x, \tau)$ (the occurrence of $x_i$ must be in a linear monomial
with a concrete coefficient of 1, otherwise iteration would cause exponential growth).
By the remark after Definition~\ref{def:itcomp}, $\alpha({\vec p}^{(r)})[i] = ({(\alpha({\vec p}))}^{\acirc (r)})[i]$,
which, by idempotence, equals
$\alpha({\vec p})[i]$. Reading backwards, $\alpha({\vec p}[i]) = \alpha({\vec p}^{(r)})[i] = \alpha(x_i) + p(\vec x, \tau)$, so
${\vec p}[i]$ also has the form $x_i + q(x,\tau)$ (where $q$ may differ from $p$).
\item By assumption, ${\vec p}[i]$ depends on $x_k$. Since $k\ne i$, we conclude that $q$ depends on $x_k$.
Also by the assumption of the cycle in $G({\vec p})$, ${\vec p}^{(r-1)}[k]$ depends on $x_i$.
This shows that $q\circ {\vec p}^{(r-1)}$ depends on $x_i$.
\item
Now, \[
{\vec p}^{(r)}[i] = ({\vec p}\circ {\vec p}^{(r-1)})[i] = {\vec p}[i]\circ {\vec p}^{(r-1)} =
(x_i + q)\circ {\vec p}^{(r-1)} = {\vec p}^{(r-1)}[i] + (q\circ {\vec p}^{(r-1)}) \,.
\]
We argue that the last expression has at least \emph{two} occurrences of $x_i$.
First, by the same token as (\ref{stp1}), $ {\vec p}^{(r-1)}[i]$ has an occurrence of $x_i$. Secondly,
$q\circ {\vec p}^{(r-1)}$ also has one.
\item Thus ${\vec p}^{(r)}$ generates exponential growth when iterated, after all.
\end{enumerate}
We conclude that a cycle as assumed cannot exist.
\end{proof}
The realizability lemmas are intended to be applied under the assumption of Theorem~\ref{thm:sdlp}, namely
that our loop is polynomially bounded; therefore we can rely on the properties guaranteed by Lemmas~\ref{obs:no-doublin} and~\ref{lem:nearly-dag}.
We focus on a particular idempotent realizable AMP $\vec p$.
Then $G({\vec p})$, with self-loops removed, is a DAG\@. We assume, w.l.o.g., that the variables
are indexed in an order consistent with $G({\vec p})$, so that if $x_i$ depends on $x_j$ then $j\le i$.
We shall refer to an (abstract) MP satisfying the properties in Lemma~\ref{obs:no-doublin}, Lemma~\ref{lem:nearly-dag} and with variables
so re-indexed (if necessary) as being \emph{neat}.\footnote{Readers who like linear algebra may draw some intuition about neat MPs
from thinking about triangular matrices whose diagonal elements are in $\{0,1\}$.}
\begin{defi}
${\vec p}\in {\texttt{\upshape $\tau$MPol}}$ is called \emph{neat} if (a) $G({\vec p})$, with self-loops removed, is a DAG;\@
(b) for all $i$, ${\vec p}[i]$ has no monomial divisible by $x_i$ except (possibly) $x_i$.
\end{defi}
Note that $\vec p$ is not required to be idempotent; neat MPs that are not necessarily idempotent come up in the upper-bound proof.
In the rest of this subsection we study neat idempotent AMPs,
establishing (as Corollary~\ref{cor:lb-nsd} below) a property important for the subsequent realizability lemma.
This property involves the $n$th iterate ${\vec p}^{(n)}$.
It is not hard to prove that $({\vec p}^{(n)})(s') \ge {\vec p}(s)$ if
$s'[i] \ge s[i]$ for all $i$. We skip the proof, however, since
we can establish a sharper result, for states $s$, $s'$ where $s'[i]\ge s[i]$ is only asserted for $i\in \sd{\vec p}$ (i.e., $x_i$ is self-dependent in $\vec p$).
We introduce the following notation: for ${\vec p}\in {\texttt{\upshape $\tau$AMPol}}$, $\chi_{\vec p}(i)$ is $1$ if $i\in \sd{\vec p}$ and $0$ otherwise.
\begin{lem}%
\label{lem:nsd-induction}
Let ${\vec p}\in {\texttt{\upshape $\tau$AMPol}}$ be idempotent and neat. Let $s,s'\in \mathbb{N}^n$ be any state vectors such that
for $i\in \sd{\vec p}$, $s'[i] \ge s[i]$ (while for $i\notin \sd{\vec p}$ no relation is asserted). Then for all $\ell\le n$,
\( {\vec q}_{\ell} \stackrel{def}{=\;} {(c\mnl {\vec p})}^{(\ell)} \)
has the following property: for all $i\le \ell$, for all $t\ge 0$,
\[
{\vec q}_{\ell}[i](s',t) \ge (c^{{(d+1)}^i}\mnl {\vec p})[i](s,t) + (s'[i]-s[i])\cdot \chi_{\vec p}(i),
\]
where $d$ is the maximum degree in $\vec p$.
\end{lem}
\noindent
The following fact is useful for proving the lemma; it follows from Lemma~\ref{lem:mnl-properties}.
\begin{fact}%
\label{lem:inductive-mnl}
Let ${\vec p} \in {\texttt{\upshape $\tau$MPol}}$, ${\vec q}\in {\texttt{\upshape $\tau$MPol}}$. Suppose that for all $i<\ell$, we have ${\vec q}[i] \ge c^{{(d+1)}^i} \mnl {\vec p}[i]$, were $d$
is the maximum degree in $\vec p$.
And suppose that $r$ is a $\tau$-polynomial depending only on $x_1,\dots,x_{\ell-1}$, also of degree at most $d$.
Then $(c\mnl r) \circ {\vec q} \ge c^{{(d+1)}^\ell}\mnl (r\circ {\vec p})$.
\end{fact}
\begin{proof}[Proof of Lemma~\ref{lem:nsd-induction}]
We use induction on $\ell$. Since $\vec p$ is neat we know that ${\vec p}[1] = x_1$.
Thus
\[
(c\mnl{\vec p})[1](s',t) = s'[1] = s[1] + (s'[1]-s[1])\cdot \chi_{\vec p}(1) = c^{d+1}\mnl{\vec p}[1](s,t) + (s'[1]-s[1])\cdot \chi_{\vec p}(1) .
\]
Next, for $\ell > 1$, consider ${\vec q}_{\ell} = (c\mnl {\vec p}) \circ {{\vec q}_{\ell-1}}$.
If $i\le \ell$ and $x_i$ is not self-dependent in $\vec p$, then ${\vec p}[i]$ only depends on variables $x_j$ with $j<i$, so:
\begin{align*}
{\vec q}_{\ell}[i](s',t) &= (c\mnl {\vec p}[i] \circ {{\vec q}_{\ell - 1}})(s',t) \\
& \ge (c\mnl {\vec p}[i] \circ c^{{(d+1)}^{i-1}}\mnl {\vec p})(s,t) && \text{by IH} \\
& \ge (c^{{(d+1)}^i} \mnl ({\vec p}[i] \circ {\vec p}))(s,t) && \text{by Fact~\ref{lem:inductive-mnl}}\\
& \ge (c^{{(d+1)}^i}\mnl {\vec p}[i])(s,t) && \text{by idempotence} \\
&= (c^{{(d+1)}^i}\mnl {\vec p}[i])(s,t) + (s'[i]-s[i])\cdot \chi_{\vec p}(i)
\end{align*}
where the next-to-last step uses the idempotence of $\vec p$ in ${\texttt{\upshape $\tau$AMPol}}$ and Lemma~\ref{lem:circ-vs-acirc}.
If $x_i$ is self-dependent in $\vec p$, then ${\vec p}[i] = x_i + r(\vec x, \tau)$ where $r$ only depends on variables $x_j$ with $j<i$. Now by IH,
and Fact~\ref{lem:inductive-mnl}:
\begin{align*}
{\vec q}_{\ell}[i](s',t) &= (c\mnl {\vec p}[i] \circ {{\vec q}_{\ell - 1}})(s',t) \\
&= {{\vec q}_{\ell - 1}}[i](s',t) + (c\mnl r \circ {{\vec q}_{\ell - 1}})(s',t) \ge s'[i] + (c\mnl r \circ (c^{{(d+1)}^{i-1}}\mnl {\vec p}))(s,t) \\
&= s[i] + (s'[i]-s[i])\cdot \chi_{\vec p}(i) + (c\mnl r \circ (c^{{(d+1)}^{i-1}}\mnl {\vec p}))(s,t) \\
&= (c\mnl {\vec p}[i] \circ (c^{{(d+1)}^{i-1}}\mnl {\vec p}))(s,t) + (s'[i]-s[i])\cdot \chi_{\vec p}(i) \\
&\ge c^{{(d+1)}^i} \mnl ( {\vec p}[i] \circ {\vec p})(s,t) + (s'[i]-s[i])\cdot \chi_{\vec p}(i) \\
&\ge (c^{{(d+1)}^i}\mnl {\vec p}[i])(s,t) + (s'[i]-s[i])\cdot \chi_{\vec p}(i) \,. \qedhere
\end{align*}
\end{proof}
\noindent
Letting $\ell=n$ in this lemma we obtain:
\begin{cor}%
\label{cor:lb-nsd}
Let ${\vec p}\in {\texttt{\upshape $\tau$AMPol}}$ be idempotent and neat. Let $s,s'\in \mathbb{N}^n$ be any state vectors such that
for $i\in \sd{\vec p}$, $s'[i] \ge s[i]$. Then, for all $i\le n$ and all $t$, ${(c\mnl {\vec p})}^{(n)}[i](s',t) \ge (c^{{(d+1)}^n}\mnl {\vec p})[i](s,t)+ (s'[i]-s[i])\cdot \chi_{\vec p}(i)$.
\end{cor}
\subsubsection{Realizability lemma for generalization}
This is what we have been preparing for.
\begin{lem}%
\label{lem:tau-realizable}
Let $\vec p$ be a realizable, idempotent and neat abstract $\tau$-MP\@.
Then ${\vec p}^\tau$ is realizable.
\end{lem}
Note that if $\vec p$ is $\tau$-closed, then ${\vec p}^\tau = {\vec p}$ and the statement is trivial.
But if $\vec p$ is not $\tau$-closed, then we have to construct a pattern that causes additive terms to accumulate in the self-dependent
variables in order to justify the replacement of sums $x_i + {\vec p}[i]''$ by $x_i + \tau{\vec p}[i]''$.
The interesting case is when $\vec p$ is not $\tau$-free:
then $\pi$ already has stars, however nested stars are not allowed in a pattern, so we cannot iterate $\pi$.
The solution is to use ${(\pi(1))}^*$. However, in reducing $\pi$ to
the star-free $\pi(1)$ we pull the rug under $\tau$-monomials already present in $\vec p$. Therefore we use a pattern that includes both ${(\pi(1))}^*$ and
$\pi$; actually, not $\pi$ but ${\pi}^n$, in order to use Corollary~\ref{cor:lb-nsd}. Note that realizability of $\vec p$ by $\pi$ means that
$\pi(1)$ realizes ${\vec p}[1/\tau]$ (which is $\vec p$ with all $\tau$'s erased). The following observation is useful:
\begin{obs}
If ${\vec p}\in {\texttt{\upshape $\tau$AMPol}}$ is idempotent, so is ${\vec p}[1/\tau]$; and similarly for neat.
\end{obs}
\newcommand{\ensuremath{{\vec p}[1/\tau]}}{\ensuremath{{\vec p}[1/\tau]}}
\begin{proof}[Proof of Lemma~\ref{lem:tau-realizable}]
Let $\pi ; c$ realize $\vec p$. Note that $\pi$ will include stars iff $\vec p$ includes $\tau$'s. We are handling the general case so we treat
$\vec p$ as a $\tau$-MP;\@ the arguments also apply to the case that it is $\tau$-free.
So, our assumption is that:
\begin{equation} \label{eq:ass2}
\text{if } \tseq{\vec{x}}{{\pi(t)}}{\vec{y}} \text{ then } \vec y \ge c\mnl {\vec p}(\vec x, t) \,.
\end{equation}
We will show that ${\vec p}^\tau$ is realized by the pattern ${(\pi(1))}^* \pi^n$.
Now, traces in $L({(\pi(1))}^{*} \pi^n )$ have the form $\sigma_t = {(\pi(1))}^t \pi^n(t)$.
We look at concrete traces \[\tseq{s_{0}}{\pi(1)}{s_1}\tseq{\cdots}{\pi(1)}{s_{t}}\tseq{\:}{\pi^n(t)}{s_{t+1}} \,.\]
We first prove by induction on $t$,
for $x_i$ self-dependent in $\vec p$, and $t\ge 0$,
\begin{equation} \label{eq:add-t}
s_{t}[i] \ge s_{0}[i] + t (c\mnl {\vec p}[i]''(s_0))
\end{equation}
(note that the omission of the $\tau$ parameter in ${\vec p}[i]''(s_0)$ reminds us that ${\vec p}[i]''$ is $\tau$-free).
For $t=0$,
the statement is trivial.
For the inductive case,
\smallskip
\[
\begin{alignedat}{2}
s_{t}[i] &\ge c\mnl {\vec p}[i](s_{t-1}, 1) &\quad&\ \notag\\
&\ge s_{t-1}[i] + c\mnl ( {\vec p}[i]'(s_{t-1}, 1) + {\vec p}[i]''(s_{t-1}) + {\vec p}[i]'''(s_{t-1},1)) &
& \text{by structure of $\vec p[i]$} \notag\\
&\ge s_{t-1}[i] + c\mnl {\vec p}[i]''(s_{t-1}) \notag\\
&\ge s_0[i] + (t-1)(c\mnl {\vec p}[i]''(s_0)) + c\mnl {\vec p}[i]''(s_{t-1}) & & \hspace{-1cm} \text{by IH} \notag\\
&\ge s_0[i] + (t-1) (c\mnl {\vec p}[i]''(s_0))
+ c\mnl {\vec p}[i]''(s_0) &
& \hspace{-1cm}\text{(s.d.~variables are non-decreasing)} \notag\\
&\ge s_0[i] + t {\vec p}[i]''(s_0) \notag
\end{alignedat}
\]
\noindent
Now, for $i\in \sd{\vec p}$,
\begin{alignat*}{2}
s_{t+1}[i] &\ge {(c\mnl {\vec p})}^{(n)}[i](s_{t}, t) &\quad& \\
&\ge (c^{{(d+1)}^n}\mnl {\vec p})[i](s_0,t) + (s_t[i]-s_0[i]) && {\text{by Cor.~\ref{cor:lb-nsd}}} \\
& = s_{0}[i] + c^{{(d+1)}^n}\mnl (t{\vec p}[i]' (s_{0},t) + {\vec p}[i]''(s_{0}) + {\vec p}[i]'''(s_{0}, t) ) + (s_t[i]-s_0[i]) && \text{by Def.~\ref{def:MPnotation}} \\
& \ge s_{0}[i] + c^{{(d+1)}^n}\mnl (t{\vec p}[i]' (s_{0},t) + {\vec p}[i]''(s_{0}) + {\vec p}[i]'''(s_{0}, t) ) + t {\vec p}[i]''(s_0) && \text{by~\eqref{eq:add-t}} \\
& \ge s_{0}[i] + c^{{(d+1)}^n}\mnl (t{\vec p}[i]' (s_{0},t) + t{\vec p}[i]''(s_{0}) + {\vec p}[i]'''(s_{0}, t) ) \\
& = c^{{(d+1)}^n}\mnl {\vec p}^\tau[i](s_0, t) \,.
\end{alignat*}
\noindent
For $i\notin \sd{\vec p}$,
\begin{alignat*}{2}
s_{t+1}[i] &\ge {(c\mnl {\vec p})}^{(n)}[i](s_{t}, t) &\quad& \\
&\ge (c^{{(d+1)}^n}\mnl {\vec p})[i](s_0,t) + (s_t[i]-s_0[i])\cdot \chi_{\vec p}(i) && {\text{by Cor.~\ref{cor:lb-nsd}}} \\
& = (c^{{(d+1)}^n}\mnl {\vec p})[i](s_0,t) \,.
\end{alignat*}
\noindent
We conclude that the pattern ${(\pi(1))}^*\pi^n$ realizes ${\vec p}^\tau$.
\end{proof}
\noindent
We can now complete our lower-bound correctness proof:
\begin{thm}%
\label{thm:SDL-lb}
Consider any set of MPs $\mathcal{S}$.
Then the set of $\tau$-MPs returned by \procSDL{}$(\mathcal{S})$ is realizable over $\mathcal{S}$.
\end{thm}
\begin{proof}
This follows by induction on the number of operations used to construct each resulting $\tau$-MP, being either composition steps or generalization
steps (justified, respectively, by Lemma~\ref{lem:comp-realizable} or Lemma~\ref{lem:tau-realizable}).
\end{proof}
\noindent
This has a crucial corollary:
\begin{cor}%
\label{cor:termination}
Consider any set of MPs $\mathcal{S}$, that represent a SDL in which all variables are polynomially bounded.
Then the set of $\tau$-MPs generated by \procSDL{}$(\mathcal{S})$ is finite, and therefore obtained in a finite number of steps.
\end{cor}
\begin{proof}
Since we are working within ${\texttt{\upshape $\tau$AMPol}}$, it is clear that if the set of $\tau$-MPs computed by the algorithm is infinite, unbounded exponents must
appear. Since all these MPs are realizable, this indicates that some variable cannot be bounded polynomially over all executions of the loop.
Taking the contrapositive, we conclude
that if all variables are known to be polynomially bounded, the set is finite.
\end{proof}
\subsection{Upper-Bound Correctness}%
\label{sec-proof-ub}
The upper-bound correctness establishes a correspondence between the set of AMPs computed by our algorithm and a set of concrete polynomials that
actually provide upper bounds for all loop executions (as in Theorem~\ref{thm:sdlp}, Clause~1).
Implicitly, this proof provides \emph{an algorithm to compute concrete upper bounds}, while ensuring that the abstractions of these bounds
remain within our set of AMPs. But we do not attempt to explicate the algorithm as such, since our main contribution is the algorithm already given.
Firstly, we have to provide a finer definition of ``upper bound'' in terms of \emph{weighted traces}.
The main technical part of the section defines a special class of $\tau$-MPs,
called \emph{iterative}, and shows some properties of the class. Finally, the main part of the proof uses a corollary of Simon's
\emph{Forest Factorization Theorem} to construct the desired upper bounds.
The application of this theorem is justified by Corollary~\ref{cor:termination}.
Before diving into upper bounds, we give a useful definition and some related lemmas.
\begin{defi}%
\label{def:join}
For multiivariate polynomials $p,q$ we define their \emph{join} $p \sqcup q$ as the polynomial
obtained by setting the coefficient of every monomial
to the largest of its coefficients in $p$ and in $q$.
\end{defi}
\noindent
For example: $(2x_1+x_1 x_2) \sqcup (x_1+x_2+3x_1 x_2) = 2x_1 + x_2 + 3x_1x_2 \,$.
\bigskip\noindent
It is easy to see that $\sqcup$ is the join (least upper bound) operation in a natural partial order on polynomials, which we denote by $\lepoly$.
Note that this order is ``syntactic'' and is a part of the ``semantic'' relation
$p \le q$ defined by treating them as functions over $\mathbb{N}^n$. To see that the relations do not coincide, consider $xy$ versus $x^2+y^2$.
With multi-polynomials, we apply the relation component-wise.
\begin{lem}%
\label{lem:lepoly-and-circ}
Suppose that ${\vec p} \lepoly {\vec q}$ and ${\vec r} \lepoly {\vec s}$. Then
${\vec p} \circ {\vec r} \lepoly {\vec q} \circ {\vec s}$.
\end{lem}
\begin{proof}
Every monomial $\mathfrak m$ of ${\vec p} \circ {\vec r}$ is the result of substituting a certain monomial of ${\vec r}[i]$ for every occurrence
of $x_i$ in $\vec p$ (where $x_i^d$ counts as $d$ occurrences). As the same monomials (up to coefficients)
appear in, respectively, ${\vec s}[i]$ and $\vec q$, we conclude that $\mathfrak m$ (times some constant) appears in ${\vec q}\circ{\vec s}$.
\end{proof}
Using the join operator, we thus have: ${\vec p} \circ ({\vec r} \sqcup {\vec s}) \gepoly ({\vec p} \circ {\vec r}) \sqcup ({\vec p} \circ {\vec s})$.
\begin{lem}%
\label{lem:afplusbg}
For any pair of polynomials $f,g$, and numbers $a,b\ge 0$,
\[
af + bg \le (a+b) (f \sqcup g) \,.
\]
\end{lem}
\begin{proof}
Partition the monomials of $f$ $\sqcup$ $g$ into three groups: those that appear only in $f$; they are multiplied by $a$ on the LHS
and by $(a+b)$ on the RHS\@. Those that appear only in $g$; they are multiplied by $b$ on the LHS
and by $(a+b)$ on the RHS\@.
For a monomial $\mathfrak m$ that appears in both polynomials---possibly with different coefficients, say $c$ in $f$ and $d$ in $g$,
we find $ac{\mathfrak m} + bd{\mathfrak m}$ in the LHS, and $(a+b)\max(c,d){\mathfrak m}$ in the RHS\@.
\end{proof}
\paragraph{\bf Weighted traces and upper bounds}
We extend our notion of traces by defining \emph{weighted traces}.
The need for this definition arises because, in the proof, we argue about a decomposition of an actual trace into segments, and we need
to abstract from the intermediate states within a segment. The number of actual transitions in the segment matters, however,
and will be represented by its weight.
Thus a weighted step, denoted $\wtrans{{\vec p}}{s}{s'}{w}$, indicates that state $s$ has evolved into step $s' = {\vec p}(s,w)$ in $w$ actual
transitions, for an integer $w\ge 1$.
A weighted concrete trace $\sigma$ is composed of such steps in the same way as an ordinary trace is; a weighted abstract trace specifies the bounds and
weights but leaves the states unspecified, e.g., $({\vec p}_1|w_1)({\vec p}_2|w_2)({\vec p_3}|w_3)$.
The total weight of a trace $\sigma$ is denoted by $\wgt{\sigma}$ and
calculated by adding up the weights of the steps. Thus we have
$\wgt{\sigma\rho} = \wgt{\sigma} + \wgt{\rho}$.
Note that since the MPs that label a weighted step are, in general, upper bounds obtained in previous analysis steps, the value
$s' = {\vec p}(s,w)$ represents, in general, a bound and not a state that is actually reachable; however,
we argue that due to the monotonicity of the functions computable in our language, we do not lose soundness by considering $s'$ instead of a set of values
bounded by $s'$.
We denote the set of weighted traces over $\mathcal{S}$ by $\wtseqs{\mathcal{S}}$.
Note that weights are associated with abstract transitions and all concretizations of a weighted trace have the same weight.
\begin{defi}[bound for a trace]
Let $\sigma\in \concrete{\wtseqs{\mathcal S}}$.
We say that $\sigma$ admits
a $\tau$-polynomial $p$ as an upper bound on variable $j$, or that $p$ bounds variable $j$ in $\sigma$,
if the following holds:
\begin{equation} \label{eq:description}
\tseq{\vec{x}}{\sigma}{\vec{y}} \land
{t \ge \wgt{\sigma}}
\ \Longrightarrow\
y_j\le p(\vec{x}, t)
\,.
\end{equation}
If $\vec p$ bounds \emph{all} variables in a concrete weighted trace $\tseq{\vec x}{\sigma}{\vec y}$,
we say that $\vec p$ bounds this trace. If $\vec p$ bounds all concretizations of $\sigma$, we say it bounds $\sigma$.
\end{defi}
\noindent
The following lemma follows from the monotonicity of all our polynomials.
\begin{lem}%
\label{lem:comp-ub}
If $\vec p$ bounds $\sigma$ and
$\vec q$ bounds $\rho$, then
${\vec q}\circ{\vec p}$ bounds $\sigma\rho$.
\end{lem}
\begin{proof}
Let $t_s = \wgt{\sigma}$, $t_r = \wgt{\rho}$. Consider a concrete run of $\sigma\rho$ and name the states thus:
$\tseq{\vec x}{\sigma}{\vec y}\tseq{\:}{\rho}{\vec z}$. Then
\(
z_i \le {\vec q}[i](\vec y, t_r)
\)
(by assumption) and for any $j$,
\(
y_j \le {\vec p}[j](\vec x, t_s)
\).
It clearly follows that
\(
z_i \le ({\vec q}[i] \circ \vec p)(\vec x, t_s+t_r)
\).
Note that $t_s+t_r$ is precisely $\wgt{\sigma\rho}$.
\end{proof}
\paragraph{\bf Iterative MPs and upper bounds for sequences of similar MPs}
\begin{defi}
$\vec p \in {\texttt{\upshape $\tau$MPol}}$ is called \emph{iterative} if all its entries depend only on self-dependent variables. Moreover, each ${\vec p}[i]$ depends
only on variables with indices $j\le i$.
\end{defi}
\noindent
\begin{exa}
$\tuple{x_1, x_2+x_1, x_2 x_1}$ is iterative, while $\tuple{x_1, x_1, x_3+x_2+x_1}$ is not
(note that the last MP is idempotent when abstracted to ${\texttt{\upshape AMPol}}$, which demonstrates that this property does not imply iterativity; the converse does not hold, either).
\end{exa}
We abbreviate ``iterative multi-polynomial'' to IMP\@.
We next give some technical definitions and results regarding IMPs. The upshot of this part is the computation of tight upper bounds for the end-state
of certain traces composed of IMPs.
We now attempt to give some intuition. Consider a sequence of MPs
which are idempotent in their abstract form. Thus if we compose their abstractions in ${\texttt{\upshape AMPol}}$ we just get the same AMP again. This is a nice
situation which suggests that we can extrapolate the bounds for a single step to bounds for any number of steps.
But there may be
some values that grow and we will have to account for this in our upper bounds. Consider the following MP:\@
\[ \tuple{x_1+x_2+x_3, \ x_3, \ x_3} \]
Note that it is idempotent in ${\texttt{\upshape AMPol}}$. However, in concrete computation, the increments accumulate. But hastily changing the first component
to $x_1 + \tau(x_2+x_3)$ overshoots the upper bound. The correct result is $x_1 + x_2 + x_3 + \tau(2 x_3)$. We find this result by extracting an iterative
MP from the given set of MPs. In the above example, the extracted IMP will be
$ \tuple{x_1 + 2x_3, \ x_3, \ x_3} $; we will see later why.
Moving to more precise details, we are going to consider sequences of \emph{similar} MPs, where similarity means that we get the same
MP if we ignore the coefficients as well as $\tau$ symbols, as expressed by the following set of definitions.
\begin{defi}
Let $\vec p \in {\texttt{\upshape $\tau$MPol}}$. We define $\tra{\vec p}$ to be $\alpha({\vec p}[1/\tau])$.
\end{defi}
\noindent
For example, $\tra{\tuple{x_1+\tau x_2, x_1^2 + \tau x_1^2}} = \tuple{x_1+x_2, x_1^2}$.
\begin{defi}
We call ${\vec p}$, $\vec q$ \emph{similar} if $\tra{\vec p} = \tra{\vec q}$.
\end{defi}
\begin{obs}%
\label{obs:tra-composition}
Suppose that $\tra{\vec p} = \tra{\vec q}$. Then for any $\vec t$, $\tra{{\vec p}\circ {\vec t}} = \tra{{\vec q}\circ {\vec t}} =
\tra{\vec p} \acirc \tra{\vec t}$.
\end{obs}
Whenever an idempotent MP is considered, we also assume it is realizable, which in turn allows us to assume that
its variables are indexed in topological order, as in the previous section, making it neat.
One could worry that the assumptions of topological order might contradict each other in an argument that involves
several different idempotents,
but this is not a problem since all the development below only concerns a set of similar MPs,
so they agree on the topological indexing. In such a discussion we may say, for example, that some $x_i$ is self-dependent, without specifying
which MP of the set is concerned.
\begin{defi}
Let $\vec p \in {\texttt{\upshape $\tau$MPol}}$ be such that $\tra{\vec p}$ is neat. Define its \emph{self-dependent cut} $\sdpart{\vec p}$ as follows:
for all $i$ self-dependent in $\vec p$, $\sdpart{\vec p}[i] = x_i$. For all other $i$, $\sdpart{\vec p}[i] = {\vec p}[i]$.
\end{defi}
\begin{lem}%
\label{lem:sdm-property}
Let ${\vec p}_1,\dots,{\vec p}_n$ be similar $\tau$-MPs such that $\tra{{\vec p}_1}=\tra{{\vec p}_2}=\dots$ is neat. Then for all $\ell\le n$,
\( {\vec q}_\ell \stackrel{def}{=\;} {\vec p}_1 \circ \sdpart{{\vec p}_2}\cdots \circ \sdpart{{\vec p}_\ell} \)
has the following property: for all $i$, ${\vec q}_\ell[i]$ only depends on variables $x_j$ which are either self-dependent
or else $j \le i-\ell$.
\end{lem}
\begin{proof}
We use induction on $\ell$. For $\ell=1$, the result is immediate, by virtue of the topological indexing.
Next, for $\ell > 1$, consider ${\vec q}_{\ell-1} = {\vec p}_1 \circ \cdots \circ \sdpart{{\vec p}_{\ell-1}}$.
Now ${\vec q}_\ell = {\vec q}_{\ell-1} \circ \sdpart{{\vec p}_\ell}$.
By IH, ${\vec q}_{\ell-1}[i]$ only depends on variables $x_j$ which are either self-dependent or have
$j\le i - \ell + 1$. In the case that $x_j$ is self-dependent,
then $\sdpart{{\vec p}_\ell}[j] = x_j$, so ${\vec q}_\ell[i]$ also depends on $x_j$ which agrees with the lemma. In the case that $x_j$ is not
self-dependent, $\sdpart{{\vec p}_\ell}[j]$ must depend on variables $x_k$ with $k \le j-1 \le i - \ell$.
\end{proof}
\noindent
This lemma is actually formulated for induction; what we are really interested in is the MP denoted by ${\vec q}_n$, for which we introduce
a special notation while enunciating the following corollary:
\begin{cor}
Let ${\vec p}_1,\dots,{\vec p}_n$ be similar $\tau$-MPs such that $\tra{{\vec p}_1}=\tra{{\vec p}_2}=\dots$ is neat. Then
\( \sdm{{\vec p}_1\cdots {\vec p}_n} \stackrel{def}{=\;} {\vec p}_1 \circ \sdpart{{\vec p}_2}\cdots \circ \sdpart{{\vec p}_n} \)
is iterative.
\end{cor}
\noindent
\emph{Notation}: if $S$ is a set of neat, similar $\tau$-MPs as above, we let
\[ \sdc{S} \stackrel{def}{=\;} \{ \sdm{{\vec p}_1\:\dots\:{\vec p}_{n}} \, \mid \, {\vec p}_1,\dots,{\vec p}_{n} \in S \} \, . \]
Note that the set $\sdc{S}$ consists of IMPs (by the above corollary) and it is easy to see that they are neat and similar to each other.
Thus, we have a way to extract a set of IMPs from a certain set of similar MPs. We will next move to the way in which we compute upper bounds
for sequences of similar MPs using these IMPs.
Recall that a $\tau$-MP is \emph{$\tau$-closed} if ${\vec p}^\tau = {\vec p}$. In particular, $\tau$-MPs that result directly from generalization
(i.e., ${\vec q}^\tau$ for some $\vec q$) are $\tau$-closed.
We next define a special composition operator for $\tau$-closed IMPs, and show an interesting property that it has.
Note that, using the notation of Definition~\ref{def:MPnotation} (Page~\pageref{def:MPnotation}),
if ${\vec p}$ is a $\tau$-closed IMP then its self-dependent entries have the form
$ {\vec p}[i] = x_i + \tau {\vec p}[i]' $.
\begin{defi}[$\tau$-absorbing composition]
We introduce a non-standard composition operation on similar, $\tau$-closed IMPs, denoted by the operator $\mathop{\star}$ (differing
from both ordinary composition $\circ$ and abstract composition $\acirc$).
Specifically, $({\vec q}\mathop{\star} {\vec p})[i]$ is defined as follows:
\begin{enumerate}
\item If $i\in \sd{\vec q} = \sd{\vec p}$, then
$({\vec q}\mathop{\star} {\vec p})[i] \stackrel{def}{=\;} {\vec p}[i] \sqcup \tau({\vec q}[i]' \circ {\vec p}) = x_i + \tau ({\vec p}[i]' \sqcup ({\vec q}[i]' \circ {\vec p}))$.
\item Otherwise, $({\vec q}\mathop{\star} {\vec p})[i] \stackrel{def}{=\;} ({\vec q}[i]\circ {\vec p})$.
\end{enumerate}
\end{defi}
\noindent
Observe that the difference between ${\vec q}\mathop{\star} {\vec p}$ and ${\vec q} \circ {\vec p}$ is only in the numeric coefficient of some monomials.
In particular, they have the same abstraction: $\alpha({\vec q}\mathop{\star} {\vec p}) = \alpha({\vec q} \circ {\vec p}) = \alpha({\vec q})\acirc \alpha({\vec p})$.
\begin{exa}%
\label{ex:tauproduct}
Consider the following $\tau$-polynomial:
\[
\begin{array}{lrlllll}
& {\vec p} = \langle & x_1, & x_2+\tau x_1, & x_3 + \tau x_2 + \tau^2 x_1, & x_4 \rangle \\[1\jot]
\llap{\text{Then:}} \\
& {\vec p}\mathop{\star} {\vec p} = \langle & x_1, & x_2+\tau x_1, & x_3 + \tau x_2 + \tau^2 x_1, & x_4 \rangle & = {\vec p} \\[1\jot]
\llap{\text{while:}} \\
& {\vec p}\circ {\vec p} = \langle & x_1, & x_2+ 2\tau x_1, & x_3 + 2\tau x_2 + 3\tau^2 x_1, & x_4 \rangle .
\end{array}
\]
\end{exa}
Thus, the operation
keeps (some) coefficients from growing when we compose $\tau$-MPs. We will see next that this has the desirable result that a sequence of
``$\mathop{\star}$ powers'' $\mathcal{T}^{\mathop{\star} (i)}$ (defined analogously to Definition~\ref{def:itcomp} on Page~\pageref{def:itcomp})
only includes a finite number of MPs. On the other
hand we will prove that it gives a sound upper bound for a sequence of transitions.
\begin{lem}%
\label{lem:taucomp-ub}
Let $\vec p$, $\vec q$ be $\tau$-closed, similar IMPs and let $\sigma$, $\rho$ be weighted traces such that
$\vec p$ bounds $\sigma$ and
$\vec q$ bounds $\rho$. Then
${\vec q}\mathop{\star} {\vec p}$ bounds $\sigma\rho$.
\end{lem}
\noindent
Compared to Lemma~\ref{lem:comp-ub}, this lemma poses stronger conditions, but has a stronger conclusion since
${\vec q}\mathop{\star} {\vec p}$ is, in general, lower than ${\vec q}\circ {\vec p}$.
\begin{proof}
The definition of ${\vec q}\mathop{\star} {\vec p}$ makes the statement trivial for variables that are not self-dependent.
So, let $x_i$ be self-dependent. Suppose that $\tseq{\vec x}{\sigma}{\vec y}\tseq{\:}{\rho}{\vec z}$.
Let $t_s = \wgt{\sigma}$, $t_r = \wgt{\rho}$.
By assumption,
\begin{align*}
z_i & \le {\vec q}(\vec y, t_r) & \\ &
= y_i + t_r{\vec q}[i]'(\vec y, t_r) \\ &
\le \left( x_i + t_s{\vec p}[i]'(\vec x, t_s) \right) + t_r{\vec q}[i]'(\vec y, t_r) \\ &
\le \left( x_i + t_s{\vec p}[i]'(\vec x, t_s+t_r) \right) + t_r({\vec q}[i]' \circ \vec p)(\vec x, t_s+t_r) \\ &
\le x_i + (t_s+t_r) ({\vec p}[i]' \sqcup ({\vec q}[i]' \circ \vec p))(\vec x, t_s+t_r) \tag{by Lemma~\ref{lem:afplusbg}} \\
& = ({\vec q}\mathop{\star} {\vec p})[i](t_s+t_r) \,. \tag*{\qedhere}
\end{align*}
\end{proof}
\begin{lem}[The Finite Closure Lemma]%
\label{lem:tau-finiteness}
Let $\mathcal{T}$ be a finite set of $\tau$-closed, similar
IMPs. Then the set
\[
\mathcal{T}^{\bigstar} = \bigcup_{k=1}^\infty \mathcal{T}^{\mathop{\star} (k)}
\]
is finite.
Hence, there is a number $\ell$ such that $\mathcal{T}^{\bigstar} = \mathcal{T}^{\mathop{\star} (\ell)}$.
\end{lem}
\begin{exa}
Consider $\mathcal{T} = \{{\vec p}, {\vec q}\}$ with $\vec p$ from Example~\ref{ex:tauproduct} and
\[
\begin{array}{lrlllll}
& {\vec q} = \langle & x_1 + \tau x_4, & x_2, & x_3 + \tau x_4, & x_4 \rangle .
\end{array}
\]
Then
\[
\begin{array}{lrlllll}
& {(\mathcal{T})}^{\mathop{\star} (2)} = & \mathcal{T} \mathop{\star} \mathcal{T} = & \{ {\vec p}\mathop{\star} {\vec p}, & {\vec q}\mathop{\star} {\vec p}, &{\vec q}\mathop{\star} {\vec q}, & {\vec p}\mathop{\star} {\vec q} \} = \\
&&& \{ {\vec p}, & {\vec r}, & {\vec q}, & {\vec r} \}
\end{array}
\]
Where
\[
\begin{array}{lrlllll}
& {\vec r} = \langle & x_1 + \tau x_4, & x_2 + \tau x_1, & x_3 + \tau x_2 + \tau^2 x_1 + \tau x_4, & x_4 \rangle .
\end{array}
\]
Next, ${(\mathcal{T})}^{\mathop{\star} (3)}$ consists of 8 products, which contribute only one new element to the accumulated union, namely
\[
\begin{array}{lrlllll}
& {\vec p}\mathop{\star} {\vec r} = \langle & x_1 + \tau x_4, & x_2 + \tau x_1, & x_3 + \tau x_2 + 2\tau^2 x_1 + \tau x_1 +\tau x_4, & x_4 \rangle .
\end{array}
\]
Adding ${(\mathcal{T})}^{\mathop{\star} (4)}$ to the union, we only have to check ${\vec p}\mathop{\star} {\vec p}\mathop{\star} {\vec r}$ and ${\vec q}\mathop{\star} {\vec p}\mathop{\star} {\vec r}$.
It turns out that these $\tau$-MPs have appeared already. Therefore, we have reached $\mathcal{T}^{\bigstar} $.
\end{exa}
\begin{proof}
Define the following sequence:
\begin{align*}
T_{(0)} & = \mathit{Id} \\
T_{(k+1)} & = T_{(k)} \sqcup \ ( \bigsqcup \mathcal{T} ) \circ T_{(k)} \, .
\end{align*}
We prove the following statement by complete induction on $k$:
\begin{equation} \label{eq:Tstar-sdv}
{\vec r} \in \mathcal{T}^{\bigstar},\ i\le k \ \Rightarrow {\vec r}[i] \lepoly T_{(k)} [i] \,.
\end{equation}
This implies the desired result because it shows that for any $i\le n$, the set of values assumed by ${\vec r}[i]$, when $\vec r$ ranges
over $\mathcal{T}^{\bigstar}$, is finite, because it is contained in the down-set of a single polynomial, namely $T_{(i)}[i]$,
and such a down-set is always finite. Hence, $\mathcal{T}^{\bigstar}$ is finite.
It is worthwhile to observe that since the members of $\mathcal{T}$ are similar IMPs, then,
whenever $i$ is a self-dependent variable (in any of them---and then in all), $\bigsqcup \mathcal{T}$ also has the form $x_i + \text{(function of lower variables)}$
and, consequently, so do the $T_{(k)}$.
To prove~\eqref{eq:Tstar-sdv},
fix $k$ and $i\le k$. We now employ induction on the number $t$ of compositions used to construct $\vec r$ (following the definition of $\mathcal{T}^{\bigstar}$).
The base case is $t=1$, i.e., ${\vec r} \in \mathcal{T}$ where the result is immediate.
In the general case,
\[
{\vec r} = {\vec q}\mathop{\star} {\vec f}
\]
where ${\vec q} \in \mathcal{T}$ and ${\vec f}\in \mathcal{T}^{\mathop{\star} (t-1)}$.
\emph{Case~}1: we consider ${\vec r}[i]$ when $i\in \sd{\vec q}$.
Write
\[
{\vec q}[i] = x_i + \tau {\vec q}[i]'
\]
then
\[
{\vec r}[i] = (x_i + \tau {\vec f}[i]') \sqcup \tau ( {\vec q}[i]' \circ {\vec f} )
\]
Note that ${\vec q}[i]'$ only depends on variables $x_j$ with $j < i \le k$.
By the induction hypothesis (of the induction on $k$) we have
\[
{\vec f}[j] \lepoly T_{(k-1)} [j]
\]
Consequently
\[
\tau ({\vec q}[i]' \circ {\vec f})
\lepoly \left({\vec q}\circ T_{(k-1)}\right) [i]
\lepoly T_{(k)} [i] \, .
\]
By the induction hypothesis on $t$,
\[
{\vec f}[i] = x_i + \tau {\vec f}[i]' \lepoly T_{(k)} [i]
\]
and we conclude that
\[
(x_i + \tau {\vec f}[i]') \sqcup \tau ( {\vec q}[i]' \circ {\vec f} ) \lepoly
T_{(k)} [i] \, \sqcup \, T_{(k)} [i] =
T_{(k)} [i] \, .
\]
\emph{Case~}2: consider $i\notin \sd{\vec q}$.
Then ${\vec q}[i]$ only depends on self-dependent variables lower than $i$, hence lower than $k$.
By the induction hypothesis on $k$, when $j<k$ we have
\[
{\vec f}[j] \lepoly T_{(k-1)} [j]
\]
Consequently
\[
{\vec r}[i] = {\vec q}[i] \circ {\vec f} \lepoly T_{(k)} [i] \, . \qedhere
\]
\end{proof}
We are heading towards a central part of the proof, where we construct upper bounds for sequences of MPs taken from a set $\mathcal{S}$ of
similar $\tau$-MPs. Moreover, we restrict attention to MPs $\vec p$ such that $\tra{\vec p}$ is idempotent.
Lemma~\ref{lem:tau-finiteness} is used to show that a finite set of upper bounds covers such traces of any length. However, it requires the MPs to be
$\tau$-closed IMPs. We close this gap by demonstrating that we can obtain upper bounds for such traces using $\sdm{\mathcal{S}}^\tau$
instead of $\mathcal{S}$ (the $\tau$ superscript over a set name means that we apply generalization to all members of the set).
\begin{lem}%
\label{lem:bounding-by-sdm}
Let $\mathcal{S}$ be a set of similar $\tau$-MPs such that $\tra{\vec p}$, for ${\vec p}\in \mathcal{S}$, is neat.
For any $t>n$,
consider a concrete trace
\[ {\concrete{\sigma}} = \wtrans{{\vec p}_1}{s_{0}}{s_1}{w_1}\dots\wtrans{{\vec p}_t}{s_{t-1}}{s_t}{w_t} \,.\]
Let ${\vec r} = \sdm{{\vec p}_{t}\:\dots\:{\vec p}_{t-n+1}}$.
Then $s_t \le {\vec r}(s_{t-1}, w')$ where $w' = \sum_{z=t-n+1}^t w_z$. In other words, ${\vec r}$ bounds the last step of $\sigma$ provided that its weight is
suitably increased.
\end{lem}
The point is to replace ${\vec p}_t$ by $\vec r$ as a bound on the last step, the gain being the convenient form of $\vec r$, namely an IMP\@.
\begin{proof}
We use induction on $d$ to prove the following claim for all $0 < d \le n$:
\bigskip\noindent
Let ${\vec r}_d = \sdm{{\vec p}_{t-n+d}\:\dots\:{\vec p}_{t-n+1}}$ and let $w'_{d} = \sum_{z=t-n+1}^{t-n+d} w_z$.
\bigskip\noindent
Then for $i\le d$, $s_{t-n+d}[i] \le {\vec r}_d (s_{t-n+d-1},w'_{d})$.
\bigskip\noindent
Note that this claim implies the lemma's statement (setting $d=n$). We now move to its proof.
\bigskip\noindent
\emph{Base case}: $d = 1$. Hence ${\vec r}_d = {\vec p}_{t-n+1}$. The claim only concerns $i=1$, and
$s_{t-n+1}[i] \le {\vec p}_{t-n} (s_{t-n+1},w'_{1})[i]$ by definition.
\bigskip\noindent
\emph{Inductive case}: $d > 1$. Then
\[
{\vec r}_d = {\vec p}_{t-n+d} \circ \sdpart{{\vec p}_{t-n+d-1}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} \,.
\]
Consider $s_{t-n+d}[i]$, for some $i \le d$. By assumption,
\begin{equation} \label{eq:st-l}
s_{t-n+d}[i] = {\vec p}_{t-n+d} [i] (s_{t-n+d-1},w_{t-n+d}) \,.
\end{equation}
Let us consider the entries $s_{t-n+d-1}[j]$ on which the above expression may depend;
thus $j\le i$.
Let us focus first on the case that $x_j$ is self-dependent. We observe that
\begin{equation} \label{eq:j-sd}
s_{t-n+d-1}[j] = ( \sdpart{{\vec p}_{t-n+d-1}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} ) [j] (s_{t-n+d-1},w)
\end{equation}
for any $w$, since $ \sdpart{\vec p}[j] = x_j$ for all ${\vec p}\in \mathcal{S}$
(note that both sides of the equation refer to $s_{t-n+d-1}$).
When $x_j$ is not self-dependent, we have (using the induction hypothesis, and the definition of self-dependent cut)
\begin{equation} \label{eq:j-not-sd}
\begin{aligned}
s_{t-n+d-1}[j] &\le ( {\vec p}_{t-n+d-1} \circ \sdpart{{\vec p}_{t-n+d-2}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} ) [j] (s_{t-n+d-2},w'_{d-1})
\\ &= ( \sdpart{{\vec p}_{t-n+d-1}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} ) [j] (s_{t-n+d-2},w'_{d-1})
\end{aligned}
\end{equation}
We now wish to replace $s_{t-n+d-2}$ in the last expression by $s_{t-n+d-1}$. To this end
we apply Lemma~\ref{lem:sdm-property}, establishing that the entries $s_{t-n+d-2}[e]$ that influence the last expression are either self-dependent
or have index $e\le j-d+1 \le 1$. But in the latter case $e=1$ and $x_1$ is certainly self-dependent. We conclude that
${\vec p}_{t-n+d-1}[e] \ge x_ e$, hence $s_{t-n+d-1}[e] \ge s_{t-n+d-2}[e]$. We now obtain from~\eqref{eq:j-not-sd} that
\begin{align*}
s_{t-n+d-1}[j] &\le ( \sdpart{{\vec p}_{t-n+d-1}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} ) [j] (s_{t-n+d-2},w'_{d-1}) \\
&\le ( \sdpart{{\vec p}_{t-n+d-1}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} ) [j] (s_{t-n+d-1},w'_{d-1}) \,.
\end{align*}
\noindent
We substitute this in~\eqref{eq:st-l} to obtain
\begin{align*}
s_{t-n+d}[i] &\le ( {\vec p}_{t-n+d} [i] \circ ( \sdpart{{\vec p}_{t-n+d-1}} \circ \dots \circ \sdpart{{\vec p}_{t-n+1}} ) )
(s_{t-n+d-1}, w'_{d-1} + w_{t-n+d}) \\
&= {\vec r}_d (s_{t-n+d-1},w'_{d}) \,. \qedhere
\end{align*}
\end{proof}
\begin{lem}%
\label{lem:idempotent-ub}
Let $\mathcal{S}$ be a set of similar $\tau$-MPs such that $\tra{\vec p}$, for all ${\vec p}\in \mathcal{S}$, is neat.
Then every weighted trace over $\mathcal{S}$ of length greater than $n$ has an upper bound in the set
\( \left(\left.(\sdc{\mathcal S}^\tau)\right.^{\bigstar} \circ {\mathcal S}^{(n)}\right) [n\tau / \tau] \).
\end{lem}
\begin{proof}
Write such a trace as $\sigma = \wtrans{{\vec p}_1}{s_{0}}{s_1}{w_1}\dots\wtrans{{\vec p}_t}{s_{t-1}}{s_t}{w_t}$.
Let $\sigma_i$ be the $i$th step, namely $\wtrans{{\vec p}_i}{s_{i-1}}{s_i}{w_i}$.
For each $i > n $, by the last lemma, the $i$th transition is bounded by a $\tau$-MP ${\vec r}\in \sdc{\mathcal S}$,
with modified weight. If $\vec r$ bounds a certain transition, then ${\vec r}^\tau$ certainly does.
Consequently,
by Lemmas~\ref{lem:comp-ub} and~\ref{lem:taucomp-ub},
$\sigma_1 \dots \sigma_n \sigma_{n +1}' \dots \sigma_{t}'$ (where the primed transitions use modified weights)
is bounded by a $\tau$-MP ${\vec q} \in \left.(\sdc{\mathcal S}^\tau)\right.^{\bigstar} \circ {\mathcal S}^n$.
In order to get rid of the modified weight, we note that
\begin{align*}
\wgt{\sigma_1 \dots \sigma_n \sigma_{n +1}'\dots \sigma_t'} & = \wgt{\sigma_1 \dots \sigma_n } + \sum_{i=n +1}^t \wgt{\sigma_{i}'} \\
& = \sum_{i=1}^n w_i + \sum_{i=n +1}^t \,\sum_{z=i-n+1}^{i} w_z \\
& \le n \wgt{\sigma}
\end{align*}
Thus we see that ${\vec q}[n \tau/\tau]$ bounds $\sigma$ using its original weight.
\end{proof}
Let $\mathcal{S}$ be the SDL under analysis, and $\alpha(\mathcal{S})$ map it into ${\texttt{\upshape AMPol}}$. The latter is a monoid with respect to the composition operation,
and we obtain a monoid homomorphism $\alpha: \tseqs{\mathcal{S}} \to {\texttt{\upshape AMPol}}$ from abstract traces to ${\texttt{\upshape AMPol}}$.
The \emph{Factorization Forest Theorem}
of Imre Simon~\cite{Simon:TCS:90} shows that such a homomorphism induces a useful structure on the traces, provided that the codomain is a
\emph{finite} monoid. Now, ${\texttt{\upshape AMPol}}$ is, of course, infinite; but if we assume that the loop under analysis is polynomially bounded, then,
as argued in the proof of termination (Corollary~\ref{cor:termination}), the closure $\closure{\alpha(\mathcal{S})}$ is finite. Note that it is a sub-monoid of ${\texttt{\upshape AMPol}}$.
Thus, Simon's theorem can be applied. Instead of the original formulation,
we can use a convenient corollary of Simon's theorem, stated by Boja\'nczyk~\cite{Bojanczyk:DLT:09} (actually we will not use his formulation but
a simplified one, as we do not need its full power).
In the statement of this result, we refer to $\alpha(\sigma)$, with $\sigma\in \tseqs{\mathcal{S}}$, as the \emph{type} of $\sigma$.
We consider regular expressions constructed using the operators: concatenation, union and Kleene-plus (where $E^+$ generates the union of all
languages generated by $E^i$ for $i>0$).
A regular expression $E$ over the alphabet $\mathcal{S}$ is \emph{well-typed} if for each of its sub-expressions $F$ (including $E$),
all words (traces) generated by $F$ have the same type, which is then the type of the expression.
\begin{thm}[Boja\'nczyk]%
\label{thm:regexp}
The existence of a homomorphism $\alpha: {\mathcal{S}}^* \to M$, where $M$ is a finite monoid, implies that $\mathcal{S}^*$ can be generated by
a finite union of well-typed regular expressions.
\end{thm}
Now, all we have to do is prove that the set of AMPs returned by our algorithm provides upper bounds for all the traces generated by each of these
regular expressions. Note that this theorem highlights the role of \emph{idempotence} in ${\texttt{\upshape AMPol}}$, since for an expression $E^+$ to be well-typed,
$\alpha(\sigma)$ has to be the same for all words $\sigma$ generated by $E^+$, which implies that it is an idempotent element.
\begin{thm}%
\label{thm:SDL-ub}
Let $\mathcal{S}$ be a polynomially-bounded SDL\@. Let $\mathcal{A}$ be the set of abstract $\tau$-MPs returned by Algorithm \procSDL{}. Then
there is a finite set $\mathcal{B} \subset {\texttt{\upshape $\tau$MPol}}$ such that $\alpha(\mathcal{B}) \subseteq \mathcal{A}$
and for any trace $\sigma\in \tseqs{\mathcal{S}}$
there exists
${\vec p}\in\mathcal{B}$ that bounds $\sigma$.
\end{thm}
\begin{proof}
We consider the regular expressions established by Theorem~\ref{thm:regexp} and all their sub-expressions: this is a finite set.
We construct a set of bounds ${\mathcal{B}}_E$, with $\alpha({\mathcal{B}}_E)\subseteq \mathcal{A}$,
for each such sub-expression $E$, by structural induction on the expressions. Clearly, this proves the theorem.
Importantly, the construction maintains these properties:
\begin{itemize}
\item If expression $E$ has type ${\vec a}\in {\texttt{\upshape AMPol}}$ then every ${\vec p}\in {\mathcal{B}}_E$ has
$\tra{\vec p} = {\vec a}$
(intuitively, the algorithm does not change the shape of the multi-polynomials except by adding $\tau$'s).
\item Moreover, $\vec p$ is realizable (this is not hard but requires a bit of attention since Theorem~\ref{thm:SDL-lb} does not refer to the
bounds constructed in the current proof, but only to $\mathcal{A}$).
\end{itemize}
If $E$ is a single transition $\vec p$ we set $\mathcal{B} = \{\vec p\}$.
\bigskip\noindent
If $E$ is $FG$ we compose ${\mathcal{B}}_F$ with ${\mathcal{B}}_G$. Note that $\alpha({\mathcal{B}}_G \circ{\mathcal{B}}_F) = \alpha({\mathcal{B}}_G) \acirc \alpha({\mathcal{B}}_F) \subseteq \mathcal{A}$
thanks to the closure computation. Realizability follows from Lemma~\ref{lem:comp-realizable}.
\bigskip\noindent
If $E$ is $F+G$ we unite ${\mathcal{B}}_F$ with ${\mathcal{B}}_G$. Again, the abstraction of the result is in $\mathcal{A}$.
\bigskip\noindent
It remains to consider an expression of the form $F^+$.
Consider ${\mathcal{B}}_F$. By IH, it consists of realizable $\tau$-MPs and are all similar to a single idempotent AMP,
so the requirements of Lemma~\ref{lem:idempotent-ub} are satisfied. Let $\Phi$ be the set of traces generated by $F$. Then every $\sigma\in\Phi$
has a bound in ${\mathcal{B}}_F$. For a concatenation of $i\le n$ such traces we have a bound in ${({\mathcal{B}}_F)}^{(i)}$. For a concatenation of
more than $n$ traces, consider each of these traces as a weighted transition where the weight represents the trace's length.
By Lemma~\ref{lem:idempotent-ub}, the concatenation of the traces has a bound in
\begin{equation}
\left(\left.(\sdc{\mathcal{B}_F}^\tau)\right.^{\bigstar} \circ {\mathcal{B}_F}^{(n)}\right) [n\tau / \tau] \,. \label{eq:bounds}
\end{equation}
By Lemma~\ref{lem:tau-finiteness},
we can replace $\left.(\sdc{\mathcal{B}_F}^\tau)\right.^{\bigstar}$ with
$\left.(\sdc{\mathcal{B}_F}^\tau)\right.^{\mathop{\star} (\ell)}$ for some $\ell > 0$.
Moreover, it is sound to relax the upper bound to $\left.({\mathcal{B}_F}^\tau)\right.^{\mathop{\star} (\ell)}$. Now~\eqref{eq:bounds} becomes
\[
\left(\left.({\mathcal{B}_F}^\tau)\right.^{\mathop{\star} (\ell)} \circ {\mathcal{B}_F}^{(n)}\right) [n\tau / \tau] \,.
\]
We claim that this set of upper bounds satisfies all our requirements. For realizability, all we need is the observation that substituting
$n\tau$ for $\tau$ does not affect realizability; plus Lemma~\ref{lem:comp-realizable} and the fact
${\vec q}\mathop{\star} {\vec p} \le {\vec q}\circ {\vec p}$.
Finally we look at the abstractions of these $\tau$-MPs, namely the set
\[
\alpha\left(\left(\left.({\mathcal{B}_F}^\tau)\right.^{\mathop{\star} (\ell)} \circ {\mathcal{B}_F}^{(n)}\right) [n\tau / \tau]\right) =
\left.({\alpha({\mathcal{B}_F})}^\tau)\right.^{\acirc (\ell)} \acirc {(\alpha({\mathcal{B}_F}))}^{\acirc (n)}
\,.
\]
These AMPs are included
in the result of \procSDL{}, since they they are produced by closure, generalization and closure again.
Note also that the bounds conform with the type of $F$ (when $\tau$'s are ignored).
\end{proof}
\newcommand{\cscc}[0]{\boxplus}
\section{On the Computational Complexity of our Problem}%
\label{sec-complexity}
Our main goal in this research was to establish that the problem of computing tight bounds is solvable. However, once proved solvable,
the question of its complexity arises.
For simplicity we assume that we are only dealing with programs where all variables are polynomially bounded.
We consider the complexity in terms of three parameters: $|P|$, the size of the program; $n$, the number of variables; and $d$, the highest degree reached.
We note that if the user wishes to verify a desired degree bound $d$, say check that a program is at most of cubic time complexity,
it is possible to use an abstraction that truncates exponents higher than $d$ and the complexity will be reduced accordingly.
First, we give an upper bound.%
\footnote{To avoid any confusion, we are obliged to point out that in the preliminary version (proceedings of FoSSaCS 2019) a wrong expression for the
upper bound was given.}
\begin{thm}
Our algorithm runs in time polynomial in $|P|\cdot {2^{n^{d+1}}}$.
\end{thm}
\begin{proof}
We estimate the complexity of our algorithm, based on bounding the number of different AMPs that may be encountered.
The number of monomials over $n$ variables is bounded (for $n>2$) by $n^d$ (think of a monomial as a product of at most $d$ variables.
For uniqueness list the indices in descending order. The number of descending lists of length at most $d$ is bounded by $n^d$).
Since a MP is an $n$-tuple of sets of monomials, the number of possible AMPs
is less than ${(2^{n^d})}^n = 2^{n^{d+1}}$.
Next, we consider the running time of the analysis of a loop (independent of the position of the loop
in the program, so we can later just multiply this time bound by the number of loops in the program).
In procedure \procSDL{}, a set of AMPs is maintained and repeatedly enlarged (by applying closure and generalization), until stable.
The time to perform each round of enlargement is polynomial in the size of the resulting set (more precisely its representation, but with any reasonable implementation
this does not change much) and the number of rounds is clearly bounded by the size of the final set. So we deduce that the total time is polynomial in
$2^{n^{d+1}}$.
Finally we should add the time to represent non-looping code as a set of AMPs, and other ``book-keeping'' operations, but clearly they contribute at most a polynomial in $|P|$ and the number of AMPs.
\end{proof}
Is this a satisfactory upper bound? It seems high, and is probably not tight. We know, however, that a solution to our problem must use at least exponential time
in the worst case,
because it has a potentially high \emph{output size}.
\begin{clm}
There is a command, of size polynomial in $n$ and $d$, which requires \[{(\lfloor n/(2d)\rfloor)}^{\lfloor nd/2\rfloor }\] AMPs to describe its result.
\end{clm}
\begin{proof}
We assume that $n$ is even and a multiple of $d$, so we can avoid the ``floor'' signs.
We use $m = n/2$ variables called $\texttt{X}_1,\dots,\texttt{X}_{m}$ and $m$ variables called $\texttt{Y}_1,\dots,\texttt{Y}_{m}$.
As usual let $x_i$ denote the initial value in $\texttt{X}_i$.
For each $j\le m$ we write a piece of code that computes
\[ (x_1 \lor x_2 \lor \dots \lor x_{m/d}) \ast \cdots \ast (x_{m-(m/d)+1} \lor \dots \lor x_{m}), \]
where the disjunction operator represents a non-deterministic choice. Hence we choose $d$ values and multiply them together.
It should be clear that this produces one of ${(m/d)}^d$ uncomparable monomials, and the result of analysing this command will have to
list all of them. We assign the product to $\texttt{Y}_j$. Since the non-deterministic
choices are made independently for each $j$, to describe the outcome in $\texttt{Y}_1,\dots,\texttt{Y}_{m}$ we need ${({(m/d)}^d)}^m = {(m/d)}^{md}$ uncomparable AMPs.
\end{proof}
Thus, as long as we use an explicit AMP representation for the output of our algorithm, worst-case
complexity exponential in $nd$ is unavoidable---at least for $d$ smaller than $n$.
This does not necessarily rule out the applicability of the algorithm, as some algorithms susceptible to combinatorial explosion still prove usable in static analysis
applications (consider the closure-based algorithm
for Size-Change Termination). At any rate, we intend to complement the work presented in this article by further research into improving the algorithm,
which we consider only a starting point, being the first complete algorithm for this problem. We now list some speculations about this future work.
The reader may have noticed that for the ``bad'' example we used above, the description would be very compact if we were allowed to use the $\textbf{max}$
operator and embed it in expressions (for our output bound for each $\texttt{Y}_j$ is just a product of $d$ ``max'' expressions). But this does not seem to be a panacea,
and we suspect that exponential output size---and running time, for sure---may be necessary even with \textbf{max}. We leave this as an open problem;
of course, the central problem is to identify the complexity class of the analysis problem.
We can also ask what happens if we redefine our problem so that the output size is small---a natural example is the case of a univariate bound;
or ask for an output-size dependent complexity function. In these cases
the output-size excuse for exponential complexity does not apply.
Another open problem that is raised by the above considerations is: what is the best bound on $d$ in terms of $|P|$ and $n$ alone?
\section{Algorithm Extensions and Open Problems}%
\label{sec:extensions}
In this section we list some ideas about how this research might be extended, specifically in terms of adding features to the subject language
(while keeping the completeness of the algorithm!), and whether we believe that our current approach
suffices for solving these extensions.
\subsection{``Unknown'' value}
When abstracting real-world code into a restricted language, it is common to have cases in which a value
has to be treated as ``unknown.'' It might be really unknown (determined by the environment) or the result of computations
that we cannot model in the restricted language (note that in our case, if we can \emph{bound} a computation by a polynomial expression we
are happy enough). It seems useful to extend our language by a special ``unknown'' value. Another example of its usage,
following~\cite{JK08}, is to analyze the growth of variables in loops for which no iteration bound is known;
in this situation one cannot obtain a time bound for the program, but we can
still bound computed values and may be able to draw conclusions regarding other quantities of interest, perhaps space complexity.
We simulate such a loop using the ``unknown'' value as the loop bound.
Concretely, this extension can be implemented by using a dedicated variable $x_u$ for anything unknown, and, throughout the algorithm,
replacing any expression that includes $x_u$ immediately by $x_u$. This prevents an explosion of the MP set (or even failure to reach a fixed point) because
of expressions including $x_u$.
\subsection{Resets}
\cite{B2010:DICE} extended the decidability result from~\cite{BJK08} to a language that contains the \emph{reset} command
\verb/X := 0/. This addition may seem trivial at first, however for the problem of deciding polynomial boundedness it caused the
complexity of decision to jump from PTIME to PSPACE-complete. The increased complexity arises from the need to recognize the situation
that a variable ``is definitely zero,'' subsequent to a particular execution path. In this case, a loop that has this variable as counter will
\emph{definitely} not execute. So the algorithm has to deal with tracking these 0's around. However, our algorithm
already tracks data-flow rather precisely and the algorithm is exponential anyway. We believe that
our algorithm can be extended to handle
resets without further raising the complexity of the solution.
\subsection{Flowchart programs}
In~\cite{BAPineles:2016} the results of~\cite{BJK08} (and, implicitly, also~\cite{B2010:DICE}) are extended from a structured language,
that can be analyzed in a compositional manner (as we have done), to a ``flowchart'' language, where a program is presented as a
\emph{control-flow graph}, or flowchart, of arbitrary shape, together with \emph{annotations} that convey information regarding
iteration bounds. We argued there that this program form is more general and closer to the form used by several analysis tools for
real-world programs. The results of~\cite{BJK08} were carried over to this language by transforming flowchart programs into programs
in a well-structured language LARE that is slightly more expressive than our core language. It seems that the same development should be doable
with precise polynomial-bound analysis, we only have to extend our algorithm to the language LARE\@. We have not investigated this in detail yet.
\subsection{Deterministic loops}
In~\cite{BK11}, Kristiansen and Ben-Amram looked at a variant of our core language where loops are deterministic: the semantics of
\verb/loop X {C}/ is to perform \pgt{C} \emph{precisely} as many times as the value of \pgt{X}. It was shown that the decision problem
of polynomial boundedness, addressed in~\cite{BJK08}, becomes undecidable in this case; however the undecidability proof exploits
worst-case scenarios where some variables are constant while others grow. It is conjectured that the problem is decidable if one only
asks about bounds that are either univariate, or multivariate but asymptotic in all variables. We propose the same conjecture with respect
to the problem of tight polynomial bounds.
\subsection{Increments and Decrements}
In our opinion, the feature that most strikingly marks our language as weak is the absence of increments and decrements (and explicit constants
in general; but if you have increment and reset you can generate other constants). However, anyone familiar with counter machines will realize that
including them would bring
our language very close to counter machines and hence to undecidability; still, without deterministic loops, this model falls just \emph{a little}
short of counter machines. We pose the decidability of the polynomial bound problem in such a language (and the investigation of the model from a
computability viewpoint, in general) as a challenging open problem.
\subsection{Procedures}
Due to the compositional form of our algorithm, we suppose that extension of the language with first-order, non-recursive procedures is not hard, but have not
investigated this further. A much greater challenge is to allow recursive procedures (one has to figure out what is a good way to do that since, of course,
we cannot allow unbounded recursion). Another one is to include high-order functions.
Avery, Kristiansen and Moyen outline in~\cite{AKM09} a possible approach
for promoting analyses like~\cite{BJK08} to higher-order programs, but a definite decidability result is not obtained.
\section{Related Work}%
\label{sec:rw}
Bound analysis, in the sense of finding symbolic bounds for data values, iteration bounds and related quantities,
is a classic field of program analysis~\cite{Wegbreit:75,Rosendahl89,ACE}. It is also an area of active research, with tools being
currently (or recently) developed including \tool{COSTA}~\cite{Albert-et-al:TCS:2011}, \tool{AProVE}~\cite{APROVE-JAR2017},
\tool{CiaoPP}~\cite{CiaoPP-TPLP2018}
, $C^4B$~\cite{CHS:pldi2015},
\tool{Loopus}~\cite{SZV:jar2017}---and this is just a sample of tools for imperative programs. There is also work on functional and logic programs,
term rewriting systems, recurrence relations, etc.~that we cannot attempt to survey here. In the rest of this section we point out work that is more directly
related to ours, and has even inspired it.
The LOOP language is due to Meyer and Ritchie~\cite{MR:67}, who note that it computes only primitive recursive functions, but complexity can rise very fast, even for programs with nesting-depth 2.
Subsequent work concerning similar languages~\cite{KasaiAdachi:80
KN04,NW06,JK08}
attempted to analyze such programs more precisely;
most of them proposed syntactic criteria, or analysis algorithms, that are
sufficient for ensuring that the program lies in a desired class (often, polynomial-time programs),
but are not both necessary and sufficient: thus, they do not prove
decidability (the exception is~\cite{KN04} which has a decidability result for a weak ``core'' language).
The core language we use in this paper is from~Ben-Amram et al.~\cite{BJK08}, who observed that by introducing weak bounded loops instead of concrete loop commands and
non-deterministic branching instead of ``\texttt{if},'' we have weakened the semantics just enough to obtain decidability of polynomial growth-rate.
This research was motivated by observing that all the previous algorithms, although they implicitly relax the semantics (since they do not analyze
conditionals, etc.), do not provide completeness over the core language.
Justifying the necessity of these relaxations,~\cite{BK11} showed undecidability for a language that can only do addition and definite loops (that cannot exit early).
In the vast literature on bound analysis in various forms,
there are a few other works that give a complete solution for a weak language.
\emph{Size-change programs} are considered by~\cite{CDZ-MFCS14,Zuleger-CSR15}. Size-change programs abstract away nearly everything in the program,
leaving a control-flow graph annotated with assertions about variables that decrease (or do not increase) in a transition. Thus, it does not assume structured
and explicit loops, and it cannot express information about values that increase. Both works yield tight bounds on the number of transitions until termination.
Dealing with a somewhat different problem,~\cite{Seidl-polynomial-invariants,Ouaknine-polynomial-invariants}
both check, or find, \emph{invariants} in the form of polynomial equations. We find it remarkable that they give
complete solutions for weak languages, where the weakness lies in the non-deterministic control-flow, as in our language.
If one could give a complete solution for polynomial
\emph{inequalities}, this would imply a solution to our problem as well.
The \emph{joint spectral radius} problem for semigroups of matrices is related to our work as well (though we have not discovered this until recently,
due to the work being done in an entirely different context). Specifically,~\cite{JPB:2008} gives an algorithm that can be expressed in the language of our
work as follows: given a Simple Disjunctive Loop in which all polynomials are linear, the algorithm decides if the loop is polynomially bounded and if it is,
returns the highest degree of $\tau$ in the tight polynomial upper bound (over all variables). A closer inspection shows that it can actually determine the degree
in which $\tau$ enters the bound for every variable. Thus, it solves a certain aspect of the SDL analysis problem. Their algorithm is polynomial-time and uses an
approach similar to~\cite{BJK08}.
\section{Conclusion and Further Work}%
\label{sec:conclusion}
We have solved an open problem in the area of analyzing programs in a simple language with bounded loops. For our language, it has been previously
shown that it is possible to decide whether a variable's value, number of steps in the program, etc.,~are polynomially bounded.
Now, we have an algorithm that computes tight polynomial bounds on the final values of variables in terms of initial values.
The bounds are tight up to constant factors (suitable constants are also computable).
This result improves our understanding of what is computable by, and about, programs of this
form. An interesting corollary of our algorithm is that as long as variables are \emph{polynomially bounded}, their worst-case bounds are
described tightly by (multivariate) \emph{polynomials}. This is, of course, not true for common Turing-complete languages.
Another interesting corollary of the \emph{proofs} is the definition of a simple class of patterns that suffice to realize the worst-case behaviors.
There are a number of possible directions for further work.
\begin{itemize}
\item
As discussed in Section~\ref{sec-complexity}, we have not settled the computational complexity of the problem we have solved.
\item
We propose to look for decidability results for richer (yet, obviously, sub-recursive) languages. Some possible language extensions
include deterministic loops, variable resets (cf.~\cite{B2010:DICE}), explicit constants, and (recursive) procedures.
The inclusion of explicit constants is a particularly challenging open problem.
\item
Rather than extending the language, we could extend the range of bounds that we can compute.
In light of the results in~\cite{KN04}, it seems plausible that the approach can be extended to classify the
Grzegorczyk-degree of the growth rate of variables when they are super-polynomial. There may also be
room for progress regarding precise bounds of the form $2^{poly}$.
\item
Our algorithm computes bounds on the highest possible values of variables. With our restricted arithmetic we can reduce the calculation of
a ``countable resource'' like the number of steps to bounding a variable. However,
our weak language seems useless as an abstraction for more advanced resource-analysis problems,
e.g., analysis of expected costs (for programs in which such an analysis is interesting).
So we pose the problem of designing weak languages that adequately abstract some non-trivial cases of more advanced analyses and obtain computability
for them.
\item
Finally, we hope to see the inclusion of our algorithm (or at least the approach) in a system that handles a real-life
programming language. In particular, it would be interesting to see how our method works together with techniques that discover loop bounds,
typically via ranking functions.
\end{itemize}
\bigskip
\paragraph{\bf Acknowledgment.} Amir M. Ben-Amram is grateful for the hospitality at the School of Computing, Dublin City University, where part of this work has been done. The authors also thank the referees for valuable comments.
\bibliographystyle{alpha}
\newcommand{\etalchar}[1]{$^{#1}$}
|
\section{\normalsize INTRODUCTION}\label{introduction_sec}
\par
}
\noindent
Value-at-Risk (VaR) is employed by many financial institutions as an important risk management tool. Representing the market risk as one number, VaR has become a standard risk measurement metric. However, VaR cannot measure
the expected loss for extreme (violating) returns. Expected Shortfall (ES, Artzner \emph{et al.}, 1997, 1999) calculates the average of returns on the ones being below the quantile (VaR) of its distribution, and is a more coherent measure than VaR. Thus, in recent years ES has become more widely employed for tail risk measurement and is one important change appears in the Basel Accord III (Basel Committee, 2010) which is expected to occur in the period leading up to 1st January 2019. However, there is much less existing research on modeling ES compared with VaR.
In recent two decades, the availability of high frequency data enables the calculation of various realized measures of volatility, including Realized Variance (RV): Andersen and Bollerslev (1998), Andersen \emph{et al.} (2003); and Realized Range (RR): Martens and van Dijk (2007), Christensen and Podolskij (2007), etc. Realized measures of volatility now play a key role in calculating accurate volatility estimates and
forecasts, e.g. the Realized GARCH model of Hansen, Huang and Shek (2011) and earlier work by Giot and Laurent (2004) and Clements, Galvao, and Kim (2008).
The quantile regression type model, e.g. the Conditional Autoregressive Value-at-Risk (CAViaR) model of Engle and Manganelli (2004), is a popular semi-parametric approach to forecast VaR. Gerlach, Chen and Chan (2011) generalize the CAViaR models to a fully nonlinear family. In additional, the realized measures have been employed into the quantile regression framework. \v{Z}ike\v{s} and Barun\'{i}k (2014) investigate how the conditional quantiles of future returns and volatility of financial assets vary with various realized measures. Avdulaj and Barunik (2017) explore nonlinearities in returns and propose to incorporate realized measures with the nonlinear quantile regression framework using copulas, to explain and forecast the conditional quantiles of financial returns.
However, the CAViaR type models cannot directly estimate and forecast ES. A semi-parametric model that directly estimates quantiles and expectiles, and implicitly ES, called the Conditional Autoregressive Expectile (CARE) model, is proposed by Taylor (2008). Gerlach and Chen (2016) employ the daily range into the CARE framework which is further extended into fully nonlinear family. Again, realized measures have been proved to be able to provide extra efficiency for the CARE type models (Gerlach, Walpole and Wang, 2017).
To select the appropriate expectile level, a grid search process is required for the CARE type models which is relatively computationally expensive (dependent on the model complexity and the size of the grid). As an alternative, Taylor (2017) proposes a joint ES and quantile regression framework (ES-CAViaR) which employs the Asymmetric Laplace (AL) density to build a likelihood function whose maximum likelihood estimates (MLEs)
coincide with those obtained by minimisation a joint loss function for VaR and ES. Fissler and Ziegel (2016) develop a family of joint loss functions (or ``scoring rules'') of the associated VaR and ES series that are strictly consistent for the true VaR and ES series, i.e. they are uniquely minimized by the true VaR and ES series. Under specific choices of functions in the join loss function of Fissler and Ziegel (2016), it can be shown that such loss function is exactly the same as the negative of AL log-likelihood function presented in Taylor (2017). Patton, Ziegel and Chen (2017) propose new dynamics models for VaR and ES, through adopting the generalized autoregressive score (GAS) framework (Creal, Koopman and Lucas (2013) and Harvey (2013)) and utilizing the loss functions in Fissler and Ziegel (2016).
In this paper, firstly a joint Conditional Autoregressive Expectile and Expected Shortfall (ES-CARE) framework is proposed, inspired by Engle and Manganelli (2004) and Taylor (2008, 2017). Secondly, the proposed model is extended with adding a measurement equation to incorporate realized measure to drive the tail risk dynamics (Realized-ES-CARE). Thirdly, the proposed framework is extended to nonlinear qunatile and ES autoregressive specification to model the volatility asymmetry (Realized-Threshold-ES-CARE). An adaptive Bayesian MCMC algorithm is utilised for estimation and forecasting in the proposed models. To evaluate the performance of the proposed Realized-ES-CARE and Realized-Threshold-ES-CARE models, the accuracy of the associated VaR and ES forecasts are assessed via an empirical study. Over a long forecasting period which includes 2008 GFC, results illustrate that the proposed Realized(-Threshold)-ES-CARE perform favourably, compared to Taylor's CARE and ES-CAViaR models and to a range of traditional competing models.
The paper is organized as follows: A review of the ES-CAViaR and CARE models is conducted in Section \ref{model_review_section}. Section \ref{model_section} formalizes the proposed ES-CARE, Realized-ES-CARE and Realized-Threshold-ES-CARE models. The associated likelihood and the adaptive Bayesian MCMC algorithm for parameter estimation are presented in Section \ref{beyesian_estimation_section}.
The simulation studies are discussed in Section \ref{simulation_section}. The empirical results are presented in Section \ref{data_empirical_section}. Section \ref{conclusion_section} concludes the paper and discusses future work.
{\centering
\section{\normalsize ES-CAViaR and CARE MODELS}\label{model_review_section}
\par
}
Koenker and Machado (1999) show that the quantile regression estimator is equivalent to a maximum likelihood estimator when assuming that the data are conditionally Asymmetric Laplace (AL)
with a mode at the quantile. If $r_t$ is the data on day $t$ and $Pr(r_t < Q_t | \Omega_{t-1}) = \alpha$, then the parameters in the model for $Q_t$ can be estimated using a likelihood based on:
$$ p(r_t| \Omega_{t-1}) = \frac{\alpha (1-\alpha)}{\sigma} \exp \left( -(r_t-Q_t)(\alpha - I(r_t < Q_t) \right)\,\, , $$
for $t=1,\ldots,n$ and where $\sigma$ is a nuisance parameter.
Taylor (2017) extends this result to incorporate the associated ES quantity into the likelihood expression, noting a link between $ES_t$ and a dynamic $\sigma_t$, resulting in the conditional density function:
\begin{eqnarray} \label{es_var_likelihood}
p(r_t| \Omega_{t-1}) = \frac{\alpha (1-\alpha)}{ES_t} \exp \left( -\frac{(r_t-Q_t)(\alpha - I(r_t < Q_t)}{\alpha ES_t} \right)\,\, ,
\end{eqnarray}
allowing a likelihood function to be built and maximised, given model expressions for $Q_t, \text{ES}_t$. Taylor (2017) notes that the negative logarithm of the resulting likelihood function is strictly consistent
for $Q_t, \text{ES}_t$ considered jointly, e.g. it fits into the class of strictly consistent functions for VaR\&ES jointly developed by Fissler and Zeigel (2016).
Taylor (2017) incorporates two different ES components that describe the dynamics between VaR and ES and also avoid ES estimates crossing the corresponding VaR estimates, as presented in Model (\ref{es_caviar_add_model}) (ES-CAViaR-Add: ES-CAViaR with an additive VaR to ES component) and Model (\ref{es_caviar_mult_model}) (ES-CAViaR-Mult: ES-CAViaR with an multiplicative VaR to ES component):
\noindent \textbf{ES-CAViaR-Add:}
\begin{eqnarray} \label{es_caviar_add_model}
Q_{t}&=& \beta_1+ \beta_2 |r_{t-1}| + \beta_3 Q_{t-1},\\ \nonumber
\text{ES}_t&=&Q_t-w_t, \\ \nonumber
w_t&=&
\begin{cases}
\gamma_0 + \gamma_1 (Q_{t-1} - r_{t-1}) + \gamma_2 w_{t-1} & \text{if } r_{t-1} \leq Q_{t-1},\\
w_{t-1} & \text{otherwise},
\end{cases}
\end{eqnarray}
where $\gamma_0 \ge 0, \gamma_1 \ge 0, \gamma_2 \ge 0$ are constrained in Taylor (2017), to ensure that the VaR and ES series do not cross.
\noindent \textbf{ES-CAViaR-Mult:}
\begin{eqnarray} \label{es_caviar_mult_model}
Q_{t}&=& \beta_1+ \beta_2 |r_{t-1}| + \beta_3 Q_{t-1}, \\ \nonumber
\text{ES}_t&=& w_t Q_t, \\ \nonumber
w_t&=& 1+\exp(\gamma_0),
\end{eqnarray}
where $\gamma_0$ is unconstrained.
In addition, for $\alpha=$ 1\% the $w_t$ component for the ES-CAViaR-Add model has in-sample estimates as in Figure \ref{wt_example}.
This step function behavior occurs since $r_{t-1} \leq Q_{t-1}$ only occurs for 1\% of the observations in an accurate model. This behavior, exhibiting constant differences
between VaR and ES for long periods, and large, sustained jumps in $VaR_t-ES_t$, seems non-intuitive and potentially able to be improved. In addition, the ES-CAViaR-Mult model has a simple multiplicative VaR to ES ratio component, while there is no direct econometrics interpretation of the parameter $\gamma_0$ in the framework.
\begin{figure}[htp]
\centering
\includegraphics[width= 0.5\textwidth]{wt_example.eps}
\caption{\label{wt_example} In-sample $w_t$ plots estimated with ES-CAViaR-Add model with S\&P 500.}
\end{figure}
\subsection{Expectile}
Expectile is closely related to quantile. The $\tau$ level expectile $\mu_{\tau}$, defined by Aigner, Amemiya and Poirier (1976), can be estimated through minimizing the following Asymmetric Least Squares (ALS) equation (Taylor, 2008):
\begin{equation}\label{als_equation}
\sum_{t=1}^{n} |\tau-I(r_t<\mu_{\tau})|(r_t-\mu_{\tau})^2 \, ,
\end{equation}
No distributional assumption is required to estimate $\mu_{\tau}$ here.
As discussed in Section \ref{introduction_sec}, ES is defined as $\text{ES}_{\alpha}= E(Y|Y<Q_{\alpha})$, which stands for the expected
value of $Y$, conditional on the set of $Y$ that is more extreme than the $\alpha$-level quantile of Y, denoted $Q_{\alpha}$.
Newey and Powell (1987) and Taylor (2008) show that this relationship can be formulated as:
\begin{equation}\label{expectile_es_equation}
\text{ES}_{\alpha}= \left(1+\frac{\tau}{(1-2\tau)\alpha_{\tau}} \right)\mu_{\tau} \, ,
\end{equation}
where $\mu_{\tau}=Q_{\alpha}$, e.g. $\mu_{\tau}$ occurs at the quantile level $\alpha_{\tau}$ of $Y$. Thus, $\mu_{\tau}$ can be used
to estimate the $\alpha$ level quantile $Q_{\alpha}$, and then scaled to estimate the associated ES.
Taylor (2008) proposes the CARE type models which have a similar form to CAViaR models of Engle and Manganelli (2004), where lagged returns drive the expectiles,
and employed ALS for estimation. The general Symmetric Absolute Value (SAV) form of this model is:
\noindent
\textbf{CARE-SAV}:
\begin{align*}
\mu_{t;\tau} = \beta_{1} + \beta_{2} |r_{t-1}| + \beta_{3} \mu_{t-1;\tau}
\end{align*}
\noindent
where $r_t$ is the return, $\mu_{t;\tau}$ is the $\tau$-level expectile and $r_t$ is daily return, all on day $t$. The CARE-type model produces one-step-ahead forecasts of $\mu_{t;\tau}$ (expectiles), that can be employed as VaR estimates, by an appropriate choice of $\tau$. The VaR estimates can be further scaled, using Equation (\ref{expectile_es_equation}), to produce forecasts of ES which cannot be directly calculated under the CAViaR framework.
However, the selection of appropriate expectile level $\tau$ requires a grid search, based on the violation rate (VRate, the ratio of the returns exceed the VaR estimates) or quantile loss function (Gerlach and Wang, 2016b). Specifically, for each grid value of $\tau$, the ALS estimator of the CARE equation parameters $\beta$ is found, yielding an associated VRate($\tau$). $\hat{\tau}$ is set to the grid value of $\tau$ s.t. VRate is closest to the desired $\alpha$. In the real applications, this grid search approach can be computationally expensive (dependent on the model complexity and the size of the grid), and the performance can be affected by the size and gap of the grid which is normally decided under ad-hoc approach.
\vspace{0.5cm}
{\centering
\section{\normalsize MODEL PROPOSED} \label{model_section}
}
\noindent
\subsection{ES-CARE Model}
In this paper, firstly we propose a new ES-CARE framework, inspired by Engle and Manganelli (2004) and Taylor (2008, 2017), to jointly and efficiently estimate and generate VaR \& ES forecasts.
Given ES to $\tau$ level expectile ($\alpha$ level quantile) relation as in Equation (\ref{expectile_es_equation}), we have:
\begin{equation}\label{var_es_equation}
\mu_{\tau} = Q_{\alpha} = \frac{\text{ES}_{\alpha}}{1+\frac{\tau}{(1-2\tau)\alpha_{\tau}}}
\end{equation}
Putting Equation (\ref{var_es_equation}) into the CARE model as below:
\begin{eqnarray} \label{caviar_sav}
&& \mu_{t;\tau}= \beta_1+ \beta_2 |r_{t-1}| + \beta_3 \mu_{t-1;\tau}, \nonumber
\end{eqnarray}
we have:
\begin{eqnarray} \label{caviar_sav}
&&\frac{\text{ES}_{t;\alpha}}{1+\frac{\tau}{(1-2\tau)\alpha_{\tau}}} = \beta_1+ \beta_2 |r_{t-1}| + \beta_3 \frac{\text{ES}_{t-1;\alpha}}{1+\frac{\tau}{(1-2\tau)\alpha_{\tau}}}, \nonumber
\end{eqnarray}
thus an autoregressive framework of ES can be derived as:
\begin{eqnarray} \label{caviar_sav}
&&\text{ES}_{t;\alpha}= \beta_1 \left( 1+\frac{\tau}{(1-2\tau)\alpha_{\tau}} \right) + \beta_2 \left( 1+\frac{\tau}{(1-2\tau)\alpha_{\tau}} \right) |r_{t-1}| + \beta_3 \text{ES}_{t-1;\alpha}, \nonumber
\end{eqnarray}
Therefore, the new ES-CARE model is proposed as:
\noindent
\textbf{ES-CARE}:
\begin{eqnarray} \label{es_care_model}
\mu_{t;\tau}&=& \beta_1+ \beta_2 |r_{t-1}| + \beta_3 \mu_{t-1;\tau},\\ \nonumber
\text{ES}_{t;\alpha}&=& \beta_1 \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) + \beta_2 \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) |r_{t-1}| + \beta_3 \text{ES}_{t-1;\alpha}, \nonumber
\end{eqnarray}
subscripts $\tau$ is removed from $\alpha_{\tau}$ to simplify the notation. There are 4 parameters to be estimated in total in Model (\ref{es_care_model}): $\beta_1$, $\beta_2$, $\beta_3$, and $\tau$. $\tau$ is constraint with $[0,\alpha]$ based on its definition. Although stationarity conditions have not been theoretically considered in the literature, it is logical
that a necessary condition would be $\beta_3<1$, so that $\mu_{t;\tau}$ and $\text{ES}_{t;\alpha}$ do not diverge; but this is not a sufficient condition for stationarity. There are no other constraints for $\beta_1$, $\beta_2$ and $\beta_3$.
It is worth note that the ${1+\frac{\tau}{(1-2\tau)\alpha_{\tau}}}$ factor is equivalent to the $1+\exp(\gamma_0)$ factor in ES-CAViaR-Mult framework (Model \ref{es_caviar_mult_model}). However, the ES-CARE has a simple linear $\frac{\tau}{(1-2\tau)\alpha_{\tau}}$ function which is potentially easier to be identified with higher accuracy than the Exponential function in ES-CAViaR-Mult model. The simulation study actually lends evidence on this. In addition, the estimated $\tau$ has a direct econometrics interpretation (expectile level), and can be used to demonstrate why the ES-CARE model can be more efficient than the original CARE model. More results will be provided in later sections.
The new framework has several nice properties. Compared with the CARE model, the model can simultaneously estimate VaR (expectile), ES and the expectile level $\tau$ without any grid search, resulting in significantly speed up estimation process. Further, the $\tau$ is estimated under a VaR and ES join loss function, e.g. Equation (\ref{es_caviar_like_equation}), so it is a more statistical estimation procedure compared with the existing ad-hoc grid search, which can potentially improve the VaR and ES estimation and forecasting accuracy. More evidence will be provided in the later sections on the improved $\tau$, VaR and ES results. In addition, compared with the ES-CAViaR-Add model in Taylor (2017), the ES-CARE framework has a more parsimonious and dynamic ES component, which can potentially tackle the challenges presented in Figure \ref{wt_example}.
Also, the ES and VaR (expectile) are guaranteed to be not cross with each other based on the above derivations. Later on, we will provide more empirical evidence on the improved VaR and ES forecasting performance with ES-CARE, compared with CARE and ES-CAViaR.
Lastly, the ES-CARE model employs autoregressive specifications for both Expectile and ES, which enables the development of fully nonlinear threshold expectile and ES autoregressive dynamics.
\subsection{Realized(-Threshold)-ES-CARE Models}
The Realized-GARCH (Re-GARCH) framework is proposed in Hansen, Huang and Shek (2012). Compared to the conventional GARCH model, the Re-GARCH employs a measurement equation, which captures the contemporaneous relation between unobserved volatility and a realized measure. The superiority of Re-GARCH compared to GARCH and GARCH-X is well demonstrated, e.g. see Hansen, Huang and Shek (2012), Watanabe (2012) and Gerlach and Wang (2016a).
The Realized-ES-CARE (Re-ES-CARE) framework is proposed as below, through adding a measurement equation which models the relation between expectile and a realized measure into the ES-CARE framework.
\noindent
\textbf{Re-ES-CARE:}
\begin{eqnarray} \label{re_es_care}
\mu_{t;\tau}&=& \beta_1+ \beta_2 X_{t-1} + \beta_3 \mu_{t-1;\tau},\\ \nonumber
\text{ES}_{t;\alpha}&=& \beta_1 \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) + \beta_2 \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) X_{t-1} + \beta_3 \text{ES}_{t-1;\alpha}, \\ \nonumber
X_t&=& \xi+\phi |\mu_{t;\tau}|+ \delta_1 \epsilon_t + \delta_2 (\epsilon_t^2-E(\epsilon^2)) + u_t \, , \\ \nonumber
\end{eqnarray}
where $X_t$ is a realized measure observed on day $t$, details to be discussed in Section \ref{data_empirical_section}. The measurement equation here is of a standard time series form, e.g. $E(u_t) = 0$, thus the standard setting and choice of $u_t \stackrel{\rm i.i.d.} {\sim} N(0,\sigma_{u}^2)$ is made for the measurement error.
It is important to note that neither the likelihood for the ES-CARE models nor for the Realized-ES-CARE models is a parametric likelihood or leads to a parametric MLE. The likelihood assumes a given
value for $\alpha$ during estimation, thus directly targeting a specific expectile (quantile) of the conditional return distribution, without assuming it has a specific distributional form.
Compared to the ES-CAViaR or Realized-GARCH models which have only one return-related ``error'', there are two return-related ``error'' series in the proposed Realized-ES-CARE type models: one is the $z_t = r_t -\mu_{t;\tau}$, which is assumed to follow an asymmetric Laplace distribution with time varying scale, so that likelihood can be constructed based on this AL density to jointly estimate the conditional VaR and conditional ES. However, the framework does not rely on an AL or any distribution assumption for the returns. The other one is $\epsilon_t= \frac{r_t} {\mu_{t;\tau}}$, that appears in the measurement equation and is employed to capture the well known leverage effect.
Again, if $\mu_{t;\tau}$ is a multiple of $\sqrt{h_t}$ then, we will have $E(\epsilon_t)=0$, as usual, but to keep a
zero mean asymmetry term $(\epsilon_t^2-E(\epsilon^2))$, we need to know $$ E(\epsilon^2) = E\left(\frac{r_t^2}{\mu_{t;\tau}^2}\right). $$
This second moment information is not included in Realized-ES-CARE framework. Thus, we substitute it with an empirical estimate
$E(\epsilon^2) \approx \bar{\epsilon^2}$, being the sample mean of the squared multiplicative errors. We note
that $E(\epsilon_t^2-\bar{\epsilon^2})= 0$ is preserved if $\bar{\epsilon^2}$ is an unbiased estimate. Therefore, the term
$\delta_1 \epsilon_t + \delta_2 (\epsilon_t^2-\bar{\epsilon^2})$ still generates an asymmetric response in volatility to return shocks.
Further, the sign of $\delta_1$ is expected to be opposite to that from an Realized-GARCH model, since the expectile $\mu_{t;\tau}$ is negative for
the lower quantile levels , e.g. $\alpha=1\%$, considered in the paper.
Motivated by the nonlinear quantile dynamics in Gerlach, Chen and Chan (2011), the Realized-ES-CARE framework is further extended to the threshold nonlinear specifications. In addition to the nonlinear expectile (VaR) component, a non-linear ES autoregressive component is incorporated. This is benefited from the proposed ES-CARE framework which directly incorporates an ES autoregressive component. The model is named as Realized-Threshold-ES-CARE (Re-Threshold-ES-CARE):
\textbf{Re-Threshold-ES-CARE:}
\begin{eqnarray} \label{re_es_care_t}
\mu_{t;\tau}&=&
\begin{cases}
\beta_{1}+ \beta_{2} X_{t-1} + \beta_{3} \mu_{t-1;\tau} , & z_{t-1} \leq c, \\
\beta_{4}+ \beta_{5} X_{t-1} + \beta_{6} \mu_{t-1;\tau} , & z_{t-1} > c,
\end{cases} \\ \nonumber
\text{ES}_{t;\alpha}&=&
\begin{cases}
\beta_{1} \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) + \beta_{2} \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) X_{t-1} + \beta_{3} \text{ES}_{t-1;\alpha} , & z_{t-1} \leq c, \\
\beta_{4} \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) + \beta_{5} \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) X_{t-1} + \beta_{6} \text{ES}_{t-1;\alpha} , & z_{t-1} > c,
\end{cases} \\ \nonumber
X_t&=& \xi+\phi |\mu_{t;\tau}|+ \tau_1 \epsilon_t + \tau_2 (\epsilon_t^2-E(\epsilon^2)) + u_t \, , \\ \nonumber
\end{eqnarray}
$z_{t}$ is a threshold variable and chosen to be self-exciting, e.g., $z_{t}= r_{t}$, and $c$ is the threshold value and set as 0 in our paper.
{\centering
\section{\normalsize LIKELIHOOD AND BAYESIAN ESTIMATION} \label{beyesian_estimation_section}
\par
}
\noindent
\subsection{ES-CARE Likelihood Function with AL}\label{es_caviar_likelihood_section}
Taylor (2017) extended the Koenker and Machado (1999) result to incorporate the ES in the equivalent likelihood function which is given in Equation (\ref{es_caviar_like_equation}). Note here $\mu_{t;\tau}= Q_t$ as discussed in Section \ref{model_section}.
\begin{eqnarray}\label{es_caviar_like_equation}
\ell(\mathbf{r};\mathbf{\theta})= \sum_{t=1}^{n} \left( \text{log} \frac{(\alpha-1)}{\text{ES}_t} + {\frac{(r_t- \mu_{t;\tau})(\alpha-I(r_t\leq \mu_{t;\tau}))}{\alpha \text{ES}_t}} \right).
\end{eqnarray}
\subsection{Realized(-Threshold)-ES-CARE Log Likelihood}\label{re_es_caviar_likelihood_section}
Because the Realized-ES-CARE framework has a measurement equation, with $u_t \stackrel{\rm i.i.d.} {\sim} N(0,\sigma_{u}^2)$, the full log-likelihood function for Realized-ES-CARE is the sum of the log-likelihood $\ell (\mathbf{r};\mathbf{\theta})$ for the expectile and ES equation (as in Equation \ref{es_caviar_like_equation}) and the log-likelihood $\ell (\mathbf{x}|\mathbf{r};\mathbf{\theta})$ from the measurement equation:
\begin{eqnarray}\label{re_es_care_like_equation}
&\ell(\mathbf{r},\mathbf{X};\mathbf{\theta})= \ell(\mathbf{r};\mathbf{\theta})+ \ell(\mathbf{X}|\mathbf{r};\mathbf{\theta})=\\ \nonumber
& \underbrace{\sum_{t=1}^{n} \left( \text{log} \frac{(\alpha-1)}{\text{ES}_t} + {\frac{(r_t-\mu_{t-1;\tau})(\alpha-I(r_t\leq \mu_{t-1;\tau}))}{\alpha \text{ES}_t}} \right) }_{\ell (\mathbf{r};\mathbf{\theta})}\\ \nonumber
& \underbrace{-\frac {1}{2} \sum_{t=1}^{n} \big( \text{log} (2 \pi)+ \text{log}(\sigma_{u}^2)+
u_t^2/\sigma_{u}^2 \big)}_{\ell (\mathbf{X}|\mathbf{r};\mathbf{\theta})} ,\\ \nonumber
\end{eqnarray}
where $u_t= X_t- \xi - \phi |Q_{t}| - \tau_{1} \epsilon_{t} - \tau_{2} (\epsilon_{t}^2-\bar{\epsilon_{t}^2})$, $t=1,\ldots,n$.
Further, the log-likelihood function of the nonlinear Realized-Threshold-ES-CARE is the same as the Realized-ES-CARE model, except the changing nonlinear dynamics of the expectile (VaR) and ES.
\subsection{Maximum Likelihood Estimation}
We have incorporated a three step maximum Likelihood (ML) approach for the proposed Realized(-Threshold)-ES-CARE models.
In the first step, for Realized-ES-CARE, the expectile equation parameters ($\beta_1,\beta_2,\beta_3$) are estimated separately by optimizing pseudo-likelihood for a expectile regression. For Realized-Threshold-ES-CARE, a threshold expectile regression model (Gerlach and Chen, 2016) is estimated separately to get the threshold expectile equation parameters ($\beta_1$ to $\beta_6$).
In the second step, multiple starting values for the measurement equation parameters ($\xi, \phi, \tau_{1}, \tau_{2}, \sigma_{u}$) and $\tau$ are randomly sampled: 10,000 random candidate starting vectors are
used.
Finally, the estimates for (threshold) quantile equation parameters in the first step are combined with the randomly sampled candidates in the second step. The parameter set that maximizes
the log-likelihood function (\ref{re_es_care_like_equation}) is selected as the starting values for the constrained optimization routine $fmincon$ in Matlab, to generate the final ML estimates.
\subsection{Bayesian Estimation}
Motivated by the favourable estimation results for CAViaR (Gerlach, Chen and Chan, 2011) and CARE-X models (Gerlach and Chen, 2016), compared to the associated MLEs, a Bayesian
estimator is also considered.
Given a likelihood function, and the specification of a prior distribution, Bayesian algorithms can be employed to estimate the parameters of Realized-ES-CARE and Realized-Threshold-ES-CARE models.
An adaptive MCMC method, adopted from that in Gerlach and Wang (2016a) and Chen \emph{et al.} (2017) is employed in this case. Three parameter blocks were employed in the MCMC simulation:
$\utwi{\theta_1}=(\beta_1,\beta_2,\beta_3, \phi)$ for Realized-ES-CARE and $\utwi{\theta_1}=(\beta_1,\beta_2,\beta_3,\beta_4,\beta_5,\beta_6,\phi)$ for Realized-Threshold-ES-CARE, $\utwi{\theta_2}=(\tau)$, $\utwi{\theta_3}=(\xi, \delta_{1}, \delta_{2}, \sigma_{u})$, via the motivation that parameters within the
same block are likely to be more strongly correlated in the posterior, than those between blocks, allowing faster mixing of the chain (e.g. see Damien \emph{et al.}, 2013). Priors are chosen to be uninformative over the possible stationarity and positivity regions, e.g. $\pi(\utwi{\theta})\propto I(A)$, which is a flat prior for $\utwi{\theta}$ over the region $A$.
In "burn-in" period, the "epoch" method in Chen \emph{et al.} (2017) is employed. For the initial "epoch", a Metropolis algorithm (Metropolis \emph{et al.}, 1953) employing a
mixture of 3 Gaussian proposal distributions, with a random walk mean vector, is utilised for each block of parameters. The proposal var-cov matrix of each block in each mixture element is
$C_i \Sigma$, where $C_1 =1; C_2 =100; C_3 =0.01$ (allowing both very big ($i=2$) and very small ($i=3$) jumps), with $\Sigma$ initially set to $\frac{2.38}{\sqrt{(d_i)}}I_{d_i}$. Here $d_i$ is the dimension of
block ($i$); $I_{d_i}$ is the identity matrix of dimension $d_i$. The covariance matrix is subsequently tuned, aiming towards a target acceptance rate of $23.4\%$ (if $d_i>4$, or $35\%$ if $2 \le d_i \le 4$, or $44\%$
if $d_i=1$), as standard, via the algorithm of Roberts, Gelman and Gilks (1997).
In order to enhance the convergence of the chain, at the end of 1st epoch, e.g. 20,000 iterations, the covariance matrix for each parameter block
is calculated, after discarding the first e.g. 2,000 iterations, which is used in the proposal distribution in the next epoch (of e.g. 20,000 iterations).
After each epoch, the standard deviations of each parameter chain in that epoch are calculated and are collectively compared to the standard deviations from the previous epoch. This
process is continued until the mean absolute percentage change over the standard deviations of parameters is less than a pre-specified threshold (10\% is employed in the paper).
In the empirical study, on average it takes 6-10 epochs to observe an absolute percentage change lower than 10\%; thus, the chains are run in
total for 120,000-200,000 iterations as a burn-in period, in the empirical parts of this paper. A final epoch of 12,000 iterates is run with an "independent" Metropolis-Hastings algorithm,
employing a mixture of three Gaussian proposal distributions for each block. The mean vector
for each block is set as the sample mean vector of the last epoch iterates (after discarding the first 2,000 iterates) for that block. The proposal var-cov matrix in each element is $C_i \Sigma$, where
$C_1 =1;C_2 =100;C_3 =0.01$ and $\Sigma$ is the sample covariance matrix of the last epoch iterates for that block (after discarding the first 2,000 iterates). This final epoch is employed as the
sample period, where all estimation and inference (and forecasting) is done via the posterior mean.
{\centering
\section{\normalsize Realized(-Threshold)-ES-CARE SIMULATION STUDY}\label{simulation_section}
\par
}
\noindent
Simulation studies are conducted to compare the properties and performance of the Bayesian method and MLE for Realized-Threshold-ES-CARE type models, with respect to parameter estimation
and one-step-ahead VaR and ES forecasting accuracy. To compare the bias and precision performance of the MCMC and ML methods, both the mean and Root Mean Square Error (RMSE) values are calculated over the replicated datasets.
1000 replicated return series are simulated from the following specific square root Realized(-Threshold)-GARCH model, specified as Simulation Models 1 \& 2. $n=1900$ is approximately the average in-sample (fixed) size for the empirical study across 7 indices, details as in Table \ref{var_fore_table}. To match up with the forecasting study and to find properties for the estimators in a similar situation, $n=1900$ is selected as the sample size in the simulation study.
\textbf{Simulation Model 1}
\begin{eqnarray} \label{re_garch_simu}
r_t&=& \sqrt{h_t} \varepsilon_t^{*}, \\ \nonumber
\sqrt{h_t}&=& 0.02 + 0.10 X_{t-1}+ 0.85 \sqrt{h_{t-1}}, \\ \nonumber
X_t&=& 0.1+0.9 \sqrt{h_t}-0.02 \varepsilon_t^{*} + 0.02 (\varepsilon_t^{*2}-1) +u_t, \\ \nonumber
&&\varepsilon_t^{*} \stackrel{\rm i.i.d.} {\sim} N(0,1), u_t \stackrel{\rm i.i.d.} {\sim} N(0,0.3^2).\\ \nonumber
\end{eqnarray}
\textbf{Simulation Model 2}
\begin{eqnarray} \label{re_t_garch_simu}
r_t&=& \sqrt{h_t} \varepsilon_t^{*}, \\ \nonumber
\sqrt{h_t}&=&
\begin{cases}
0.05 + 0.20 X_{t-1}+ 0.80 \sqrt{h_{t-1}}, & r_{t-1} \leq 0, \\
0.10 + 0.10 X_{t-1}+ 0.75 \sqrt{h_{t-1}}, & r_{t-1} > 0,
\end{cases} \\ \nonumber
X_t&=& 0.1+0.9 \sqrt{h_t}-0.02 \varepsilon_t^{*} + 0.02 (\varepsilon_t^{*2}-1) +u_t, \\ \nonumber
&&\varepsilon_t^{*} \stackrel{\rm i.i.d.} {\sim} N(0,1), u_t \stackrel{\rm i.i.d.} {\sim} N(0,0.3^2).\\ \nonumber
\end{eqnarray}
In order to calculate the corresponding Realized-ES-CARE true parameter values, a mapping from the Simulation Model 1 to the Realized-ES-CARE is required.
Further, given $\Phi^{-1}(\alpha)$ as the standard Gaussian inverse cdf, we have $\mu_{t;\tau}=Q_{t}=\sqrt{h_t} \Phi^{-1}(\alpha)$, then
$\sqrt{h_t} =\frac{\mu_{t;\tau}} {\Phi^{-1}(\alpha)} $. Then, with $\varepsilon_t^{*} \stackrel{\rm i.i.d.} {\sim} N(0,1)$, we have $\epsilon_{t}= \frac{r_t} {\mu_{t;\tau}} = \frac{r_t} {\sqrt{h_t} \Phi^{-1}(\alpha)} = \frac{\varepsilon_t^{*}} {\Phi^{-1}(\alpha)}$. Substituting these back into the Simulation Model 1, the corresponding Realized-ES-CARE specification can be written as:
\begin{eqnarray}
\mu_{t;\tau}&=& 0.02 \Phi^{-1}(\alpha)+ 0.10 \Phi^{-1}(\alpha) X_{t-1}+ 0.85\mu_{t-1;\tau},\\ \nonumber
\text{ES}_{t;\alpha}&=& 0.02 \Phi^{-1}(\alpha) \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) + 0.10 \Phi^{-1}(\alpha) \left( 1+\frac{\tau}{(1-2\tau)\alpha} \right) |r_{t-1}| + 0.85 \text{ES}_{t-1;\alpha}, \\ \nonumber
X_t &=& 0.1-\frac{0.9} {\Phi^{-1}(\alpha)} |\mu_{t;\tau}| -0.02\Phi^{-1}(\alpha)\epsilon_{t} + 0.02 \Phi^{-1}(\alpha)^2 (\epsilon_{t}^2- \frac{1} {\Phi^{-1}(\alpha)^2} ) + u_t, \\ \nonumber
\end{eqnarray}
allowing true parameter values to be calculated or read off. These true values are presented in Table \ref{simu_table_1}. Similarly, the parameter true values of the Realized-Threshold-ES-CARE model corresponding to Simulation Model 2 are derived similarly and presented in Table \ref{simu_table_2}.
The true value of $\tau$ parameter can be calculated as well. The true in-sample and one-step-ahead $\alpha$ level VaR and ES forecast can be exactly calculated for each data set; i.e. $\text{VaR}_{t;\alpha}=\mu_{t;\tau}=Q_{t;\alpha}= \sqrt{h_{t}}\Phi^{-1}(\alpha)$,
and $\text{ES}_{t;\alpha}= -\sqrt{h_{t}} \frac{\phi(\Phi^{-1}(\alpha))}{\alpha} $, where $\phi()$ is standard Normal pdf. Via the one-to-one relationship between VaR and ES presented in Equation (\ref{expectile_es_equation}), the true value of $\tau$ in this model can be exactly calculated: 0.001461.
VaR forecast is $\text{VaR}_{n+1}= \sqrt{h}_{n+1}\Phi^{-1}(\alpha)$, and the corresponding true ES forecast is $\text{ES}_{n+1}= -\sqrt{h}_{n+1}\frac{\phi\left(\Phi^{-1}(\alpha)\right)}{\alpha}$, where $\phi$ is the standard Gaussian pdf, which are calculated for each dataset; the averages of these, over
the 1000 datasets, are given as $\text{VaR}_{n+1}$ and $\text{ES}_{n+1}$ in the "True" column of Table \ref{simu_table_1} and \ref{simu_table_2}.
The Realized-ES-CARE and Realized-Threshold-ES-CARE models are fit to the 1000 datasets generated by Simulation Models 1 and 2 respectively, once using the adaptive MCMC method and once using the ML estimator.
Estimation results of Realized-ES-CARE are summarized in Table \ref{simu_table_1}, where boxes indicate the preferred model in terms of minimum
bias (Mean closest to Ture) and maximum precision (minimum RMSE). First, both MCMC and ML generate relatively accurate parameter estimates and VaR \& ES forecasts in this case, which proves the validity of both methods
as discussed in Section \ref{beyesian_estimation_section}. For all 9 parameters and both VaR \& ES forecasts the bias results clearly favour the MCMC estimator compared to the ML. Further, the precision is higher for the MCMC method for 7 out of 9 parameters and for both VaR and ES forecasts. It is worth noting that the proposed framework can generate very close to True $\tau$ estimates, which proves the validating of the proposed framework.
As discussed in Section \ref{model_section}, the ES-CARE model has a simple linear $\frac{\tau}{(1-2\tau)\alpha_{\tau}}$ function which is potentially easier to be estimated with higher accuracy than the Exponential function in ES-CAViaR-Mult model. As can be seen, the RMSE values for the $\tau$ are quite small for both methods and are much smaller than that of the $\gamma_0$ of ES-CAViaR-Mult (simulation results not shown here). In the measurement equation, the MCMC generates clearly better estimation results for $\xi$ and $\phi$ which are known to be the two most important parameters in the realized GARCH framework.
With respect to the Realized-Threshold-ES-CARE estimation, as in Table \ref{simu_table_2}, MCMC still demonstrates it advantageous compared to ML. Accurate parameter estimates and VaR \& ES forecasting results are produced by both adaptive MCMC and ML. However, compared with the RMSE results for VaR and ES forecasts in Table \ref{simu_table_1}, increased RMSE values are observed, which is due to the challenge of estimating a more complex framework. With respect to bias results, MCMC is favoured by 5 parameters and by both the VaR and ES tail risk forecasts. Regarding precision, MCMC produces lower RMSE for 10 out 12 parameters and both VaR and ES forecasts.
\begin{table}[!ht]
\begin{center}
\caption{\label{simu_table_1} \small Summary statistics for the two estimators of the Realized-ES-CARE model, with data generated from Simulation Model 1.}\tabcolsep=10pt
\begin{tabular}{lcccccccc} \hline
$n=1900$ & & \multicolumn{2}{c}{MCMC} & \multicolumn{2}{c}{ML} \\
Parameter &True &Mean & RMSE &Mean & RMSE \\ \hline
$\beta_1$ & -0.0465 & \fbox{-0.0674} & \fbox{0.1245}& -0.0714 & 0.1574 \\
$\beta_2$ & -0.2326 & \fbox{-0.2492}& \fbox{0.0904}& -0.2495& 0.0959\\
$\beta_3$ & 0.8500 & \fbox{0.8255} & \fbox{0.1061} & 0.8222& 0.1374 \\
$\tau$ & 0.001461 & \fbox{0.001363} & \fbox{0.000311} & 0.001348 &0.000322 \\
$\xi$ & 0.1000 & \fbox{0.1807} & \fbox{0.1693}& 0.2027 & 0.3503 \\
$\phi$ & 0.3869 & \fbox{0.3394} & \fbox{0.1278}& 0.3220& 0.2646 \\
$\delta_{1}$ & 0.0465& \fbox{0.0411}& 0.0168& 0.0410& \fbox{0.0167} \\
$\delta_{2}$ & 0.1082 & \fbox{0.0961}& \fbox{0.0289} & \fbox{0.0961}& \fbox{0.0289}\\
$\sigma_{u}$ & 0.3000 & \fbox{0.2801}& \fbox{0.0204}& 0.2797& 0.0208 \\
$\text{VaR}_{n+1}$ & -1.2523& \fbox{-1.2499}& \fbox{0.0706}& -1.2497 & 0.0747\\
$\text{ES}_{n+1}$ & -1.4349 & \fbox{-1.4203}& \fbox{0.0858}& -1.4182& 0.0896\\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small A box indicates the favored estimators, based on mean and RMSE.
\end{table}
\begin{table}[!ht]
\begin{center}
\caption{\label{simu_table_2} \small Summary statistics for the two estimators of the Realized-Threshold-ES-CARE model, with data generated from Simulation Model 2.}\tabcolsep=10pt
\begin{tabular}{lcccccccc} \hline
$n=1900$ & & \multicolumn{2}{c}{MCMC} & \multicolumn{2}{c}{ML} \\
Parameter &True &Mean & RMSE &Mean & RMSE \\ \hline
$\beta_1$ & -0.1163 & -0.1483 & \fbox{0.1573} &\fbox{-0.0992}& 0.2033 \\
$\beta_2$ & -0.4653 & -0.5096& \fbox{0.1720}& \fbox{-0.4961}& 0.1727 \\
$\beta_3$ & 0.8000 & 0.7688& \fbox{0.0776} &\fbox{0.7952} &0.0922 \\
$\beta_4$ & -0.2326 & -0.2077& \fbox{0.1521}& \fbox{-0.2255} & 0.1816\\
$\beta_5$ & -0.2326& -0.2610 & \fbox{0.1452} &\fbox{-0.2527}& 0.1476 \\
$\beta_6$ & 0.7500& \fbox{0.7460}& \fbox{0.0871}& 0.7420& 0.0974\\
$\tau$ & 0.001461 & \fbox{0.001318}& \fbox{0.000320} & 0.001296& 0.000345\\
$\xi$ & 0.1000 & \fbox{0.1416} & \fbox{0.2127} &0.1465 &0.2130\\
$\phi$ & 0.3869 & \fbox{0.3714} &\fbox{0.0937}& 0.3696 &0.0940 \\
$\delta_{1}$ & 0.0465 & \fbox{0.0455} &0.0158 &0.0454 &\fbox{0.0157} \\
$\delta_{2}$ & 0.1082& 0.1094& 0.0278& \fbox{0.1092}& \fbox{0.0277} \\
$\sigma_{u}$ & 0.3000& \fbox{0.2999} &\fbox{0.0051} &0.2994 & \fbox{0.0051}\\
$\text{VaR}_{n+1}$ & -2.3443& \fbox{-2.3405} & \fbox{0.1496} & -2.3074&0.2384\\
$\text{ES}_{n+1}$ & -2.6863& \fbox{-2.6495} &\fbox{0.1825} &-2.6068& 0.2822 \\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small A box indicates the favored estimators, based on mean and RMSE.
\end{table}
{\centering
\section{\normalsize DATA and EMPIRICAL STUDY}\label{data_empirical_section}
\par
}
\subsection{Realized Measures}
Various realized measures, including realized variance (RV) and realized range (RR) are incorporated in the proposed Realized-ES-CARE and Realized-Threshold-ES-CARE type models.
To reduce the effect of microstructure noise of realized measures, Martens and van Dijk (2007) present a scaling process which is inspired by the fact that the daily squared return and range are less affected by microstructure noise than their high frequency counterparts. Therefore, the process can be used to smooth and scale RV and RR, creating less microstructure sensitive measures.
Further, Zhang, Mykland and A\"{i}t-Sahalia (2005) propose a sub-sampling process to deal with micro-structure effects for realized variance (SSRV). The sub-sampling process is applied to RR in Gerlach and Wang (2016b).
The properties of the sub-sampled RR, compared to those of other realized measures, are assessed via simulation under three scenarios in Gerlach and Wang (2016b).
The scaled RV (ScRV), Scaled RR (ScRR), sub-sampled RV (SSRV) and sub-sampled RR (SSRV) are also employed and tested in the proposed frameworks. For example, Realized-ES-CARE-RV represents Realized-ES-CARE framework employing RV, and Realized-Threshold-ES-CARE-RR represents Realized-Threshold-ES-CARE framework employing RR.
\subsection{Data Description}
Daily and high frequency data, observed at 1-minute and 5-minute frequency within trading hours, including open, high, low and closing prices, are downloaded from
Thomson Reuters Tick History. Data are collected for 7 market indices: S\&P500, NASDAQ (both US), Hang Seng (Hong Kong), FTSE 100 (UK),
DAX (Germany), SMI (Swiss) and ASX200 (Australia). The time period is Jan 2000 to June 2016.
The daily return and the daily RV, RR, Scaled RV and Scaled RR measures (using 5 minute data) are calculated; 1-minute data are employed to produce daily sub-sampled RV and sub-sampled RR measures; $q=66$ is employed for the scaling process as suggested in , e.g. around 3 months as suggested in Martens and van Dijk (2007).
\subsection{In-sample Parameter Estimates} \label{parameter_estimates_section}
Before presenting the forecasting results, the parameter estimates from Realized-Threshold-ES-CARE models are shown for S\&P 500. Table \ref{para_est_table} presents the parameter estimates for the 1st forecasting step (using 1st in-sample data set).
First, the parameter estimates are consistent with the results in Table II of Hansen, Huang and Shek (2012), after mapping between Realized-Threshold-ES-CARE and Realized-GARCH model as discussed in Section \ref{simulation_section}. For example, the $\varphi$ estimates in the measurement equation should be in general close to 1 in the Realized-GARCH framework. Dividing 1 by $\Phi^{-1}(\alpha)$ (assuming Normal error distribution) and taking negative of it (as $|\mu_{t;\tau}|$ employed), we have 0.4299 which is close to real the $\varphi$ estimates in Table \ref{para_est_table}.
Second, we can see that the absolute values of $\beta_2$ in the $r_t \leq 0$ regime is in general larger than $\beta_6$, meaning the realized measures will contribute more to the VaR\&ES forecasting when $r_t \leq 0$. Such results are consistent with our expectation and prove the proposed threshold framework can successfully capture the volatility asymmetry.
Third, Realized-Threshold-ES-CARE-RR generates smaller $\sigma_{u}$ values than Realized-Threshold-ES-CARE-RV. This result is consistent with the findings in Martins and van Dijk (2007) and Christensen and Podolskij (2007): RR can have much lower mean squared error than RV, which might provide RR with
higher accuracy and efficiency in volatility estimation and forecasting. The $\sigma_{u} $ of Realized-Threshold-ES-CARE-SSRV is smaller than that of Realized-Threshold-ES-CARE-RV. Comparing Realized-Threshold-ES-CARE-RR and Realized-Threshold-ES-CARE-SSRR, the $\sigma_{u}$ estimates are quite close and are consistent with the findings in Gerlach and Wang (2016b). It seems that $\sigma_{u}$ estimates from the models with ScRV and ScRR are not improved compared with model employing RV and RR. The results here are in general consistent with the subsequently discussed forecasting performance.
Last, as discussed in Martins and van Dijk (2007), RR is biased as a true volatility estimator, if each day $t$ is divided into finite number of equally sized intervals.
However, the RR or SSRR in the Realized-Threshold-ES-CARE models are not required to be unbiased, because the parameters in the
model can adjust such bias: an advantage of using the Realized-GARCH framework which is inherited by the Realized(-Threshold)-ES-CARE frameworks.
\begin{table}[h]
\begin{center}
\caption{\label{para_est_table} \small The estimated Re-Threshold-ES-CARE parameters for the 1st forecasting step with S\&P500.}\tabcolsep=10pt
\tiny
\begin{tabular}{lcccccccc} \hline
Parameters&RV&RR&ScRV&ScRR&SSRV&SSRR\\
\hline
$\beta_1$&-0.2019&-0.1586&-0.2449&-0.2227&-0.2198&-0.1779\\
$\beta_2$&-0.6893&-1.0399&-0.6059&-0.7013&-0.8463&-0.9625\\
$\beta_3$&0.7358&0.7461&0.7223&0.7333&0.7042&0.7315\\
$\beta_4$&-0.0538&-0.0026&-0.0765&-0.0697&-0.0938&-0.0410\\
$\beta_5$&-0.4817&-1.0102&-0.4137&-0.6289&-0.6854&-0.9343\\
$\beta_6$&0.7430&0.6876&0.7465&0.7010&0.6848&0.6735\\
$\tau$&0.001851&0.001873&0.001888&0.002012&0.001957&0.001980\\
$\xi$ &-0.0516&0.0141&-0.1375&-0.0900&-0.0677&-0.0135\\
$\varphi$ &0.3783&0.2410&0.4496&0.3750&0.3441&0.2741\\
$\tau_1$&0.0800&0.0725&0.0855&0.1018&0.1237&0.0962\\
$\tau_2$&0.3653&0.1745&0.3764&0.2342&0.2140&0.1540\\
$\sigma_{u}$&0.2551&0.1504&0.2915&0.2112&0.2037&0.1577\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Tail Risk Forecasting}
$\alpha= 1\%$ is employed for both one day ahead Value-at-Risk (VaR) and Expected Shortfall (ES) forecasting study for the 7 indices.
A rolling window with fixed in-sample size is employed for estimation to produce each one step ahead forecast in the forecast period; the in-sample size $n$ is
given in Table \ref{var_fore_table} for each series, which differs due to non-trading days in each market. In order to see the
performance during the GFC period, the initial date of the forecast sample is chosen as the beginning of 2008. On average,
2111 one day ahead VaR and ES forecasts are generated for each return series from a range of models.
24 models are tested and compared in this section. These include the proposed Realized-ES-CARE and Realized-Threshold-ES-CARE models (estimated with adaptive MCMC) with different input measures of volatility: RV \& RR, scaled RV \& RR and sub-sampled RV \& RR.
The proposed ES-CARE, original ES-CAViaR-Add and ES-CAViaR-Mult models (estimated with adaptive MCMC) are also included in the study. We have also tested the conventional GARCH (Bollerslev, 1986), EGARCH (Nelson, 1991) and GJR-GARCH (Glosten, Jagannathan and Runkle, 1993) with Student-t distribution; the GARCH employing Hansen's skewed-t distribution (Hansen, 1994) and Realized-GARCH with Gaussian and Student-t return equation error distributions (Hansen, Huang and Shek, 2012).
Further, a filtered GARCH-t historical simulation (GARCH-t-HS) approach is also included, where a GARCH-t is fit to the in-sample data. Using all the in-sample data the standardised VaR and ES are estimated via historical simulation, from the sample of returns (e.g. $r_1,\ldots,r_n$ divided by their GARCH-estimated conditional standard deviation (i.e. $r_t/\sqrt{\hat{h_t}}$).
Then final forecasts of VaR, ES are found by multiplying the standardised VaR, ES estimates by the forecast $\sqrt{\hat{h}_{t+1}}$ from the GARCH-t model.
Finally, the Threshold-GARCH model (Li and Li, (1996) and Brooks, (2001)) incorporating Hansen's Skewed-t distribution (T-GARCH-Skew-t) is also tested. All these aforementioned models are estimated by ML, using the Econometrics toolbox in Matlab (GARCH-t, EGARCH-t, GJR-t and GARCH-t-HS) or code developed by the authors (CARE-SAV, T-GARCH-Skew-t and Realized-GARCH). The actual forecast sample sizes $m$, in each series, are given in Table \ref{var_fore_table}.
Firstly, the VaR violation rate is employed to assess the VaR forecasting accuracy. VRate is simply the proportion of returns that exceed the forecasted VaR level in the forecasting period. Models with VRate closest to nominal quantile level $\alpha=1\%$ are preferred.
Several standard quantile accuracy and independence tests are also employed: the unconditional coverage (UC) and conditional
coverage (CC) tests of Kupiec (1995) and Christoffersen (1998) respectively, as well as the dynamic quantile (DQ) test of
Engle and Manganelli (2004) and the VQR test of Gaglione et al. (2011). Finally, the standard quantile loss function is also employed to compare the models for VaR forecast accuracy. Since the standard quantile loss function is strictly consistent, e.g. the expected loss is a minimum at the true quantile series. Thus, the most accurate VaR forecasting model
should produce the minimized quantile loss function, given as:
\begin{equation}\label{q_loss}
\sum_{t=n+1}^{n+m}(\alpha-I(r_t<Q_t))(r_t-Q_t) \,\, ,
\end{equation}
where $Q_{n+1},\ldots,Q_{n+m}$ is a series of quantile forecasts at level $\alpha$ for the observations $r_{n+1},\ldots,r_{n+m}$.
\subsubsection{\normalsize Value at Risk}
Table \ref{var_fore_table} presents the VRates for each model over the 7 indices. $\alpha = 1\%$ is the target violation rate. For each time series, the models are ranked according to the deviations to the 1\% target rate. Then the average these ranks across all 7 markets is presented in the "Avg Rank" column in Table \ref{var_fore_table}, to compare the overall performance of each model. A box indicates the model with VRate closest to 1\% in each market, while bold indicates the VRate is significantly different to 1\% by the UC test.
As presented in Table \ref{var_fore_table}, overall the best ranked models are proposed ES-CARE and Realized-ES-CARE-SSRR models, followed by the GARCH-Skew-t model. For 5 series the proposed ES-CARE or Realized(-Threshold)-ES-CARE models produce the best or second best VRates. Models, including GARCH-Skew-t, T-GARCH-Skew-t and CARE, also generate quite competitive VRate results. Using quantile loss, now we compare Realized(-Threshold)-ES-CARE type models and these models in more detail with respect to economic efficiency, and provide evidence on why the proposed models are preferred in VaR forecasting.
\begin{table}[hbt!]
\begin{center}
\caption{\label{var_fore_table} \small 1\% VaR Forecasting VRate with different models on 7 indices.}\tabcolsep=10pt
\tiny
\begin{tabular}{lccccccccccc} \hline
Model&S\&P500&NASDAQ&HK&FTSE&DAX&SMI&ASX200&Avg Rank\\ \hline
GARCH-t&\bf{1.467\%}&\bf{1.895\%}&\bf{1.652\%}&\bf{1.731\%}&1.362\%&\bf{1.617\%}&\bf{1.702\%}&20.71\\
EGARCH-t&\bf{1.514\%}&\bf{1.611\%}&1.215\%&\bf{1.777\%}&1.408\%&\bf{1.712\%}&\bf{1.466\%}&19.86\\
GJR-GARCH-t &\bf{1.467\%}&\bf{1.563\%}&1.263\%&\bf{1.777\%}&1.408\%&\bf{1.759\%}&\bf{1.513\%}&19.57\\
GARCH-t-HS &1.230\%&\bf{1.563\%}&1.263\%&1.123\%&\textcolor{blue}{1.127\%}&1.284\%&0.898\%&8.29\\
GARCH-Skew-t&\textcolor{blue}{1.088\%}&\textcolor{blue}{1.042\%}&1.263\%&1.169\%&\fbox{0.939\%}&1.331\%&0.804\%&\textcolor{blue}{7.14}\\
T-GARCH-Skew-t&\fbox{0.994\%}&\fbox{0.995\%}&1.312\%&1.356\%&1.174\%&1.331\%&\fbox{1.040\%}&8.00\\
CARE &1.278\%&\bf{1.563\%}&\fbox{1.020\%}&1.310\%&1.221\%&1.284\%&1.229\%&9.86\\
Re-GARCH-RV-GG &\bf{2.130\%}&\bf{1.942\%}&\bf{2.818\%}&\bf{1.777\%}&\bf{2.300\%}&\bf{1.807\%}&\bf{1.560\%}&23.57\\
Re-GARCH-RV-tG &\bf{1.467\%}&1.326\%&\bf{1.992\%}&1.310\%&\bf{1.596\%}&\fbox{1.141\%}&1.229\%&14.29\\
ES-CAViaR-Add&\bf{1.467\%}&\bf{1.516\%}&1.215\%&1.216\%&1.268\%&1.236\%&\textcolor{blue}{0.946\%}&10.57\\
ES-CAViaR-Mult&1.278\%&1.421\%&1.166\%&1.216\%&1.315\%&1.236\%&\textcolor{blue}{0.946\%}&7.71\\
ES-CARE&1.278\%&1.421\%&1.166\%&1.169\%&1.315\%&1.236\%&\textcolor{blue}{0.946\%}&\fbox{7.00}\\
Re-ES-CARE-RV&1.278\%&\bf{1.705\%}&\bf{2.187\%}&1.169\%&1.315\%&1.331\%&0.804\%&13.86\\
Re-ES-CARE-RR&\textcolor{blue}{1.088\%}&1.374\%&1.263\%&0.889\%&1.221\%&1.427\%&0.709\%&9.29\\
Re-ES-CARE-ScRV&1.325\%&\bf{1.658\%}&1.166\%&1.169\%&1.315\%&\textcolor{blue}{1.189\%}&0.898\%&10.00\\
Re-ES-CARE-ScRR&1.278\%&\bf{1.468\%}&\fbox{1.020\%}&1.123\%&1.221\%&1.236\%&0.709\%&7.57\\
Re-ES-CARE-SSRV&1.372\%&\bf{1.468\%}&1.215\%&\textcolor{blue}{1.076\%}&1.315\%&1.284\%&0.709\%&10.71\\
Re-ES-CARE-SSRR&1.278\%&1.374\%&1.166\%&\fbox{0.935\%}&1.174\%&1.331\%&0.709\%&\fbox{7.00}\\
Re-T-ES-CARE-RV&1.183\%&\bf{1.516\%}&\bf{2.041\%}&1.123\%&1.362\%&1.284\%&0.851\%&11.71\\
Re-T-ES-CARE-RR&1.136\%&1.184\%&1.263\%&0.889\%&1.268\%&\bf{1.617\%}&0.662\%&10.86\\
Re-T-ES-CARE-ScRV&1.230\%&1.421\%&1.263\%&1.169\%&1.174\%&1.331\%&0.851\%&8.57\\
Re-T-ES-CARE-ScRR&1.278\%&\bf{1.468\%}&\textcolor{blue}{1.069\%}&1.076\%&1.268\%&1.427\%&0.709\%&9.43\\
Re-T-ES-CARE-SSRV&1.420\%&1.279\%&1.166\%&1.076\%&1.362\%&1.427\%&0.851\%&10.14\\
Re-T-ES-CARE-SSRR&1.372\%&1.137\%&1.166\%&\fbox{0.935\%}&1.315\%&\bf{1.569\%}&0.757\%&9.86\\ \hline
m (Forecasting steps) &2113&2111&2058&2138&2130&2103&2115&2114\\
n (In-sample size)&1905&1892&1890&1943&1936&1930&1871&1916\\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small Box indicates the favored models based on ESRate, blue shading indicates the 2nd ranked model, for each series and average rank. Bold indicates the violation rate is
significantly different to 1\% by the UC test. $m$ is the out-of-sample size, and $n$ is in-sample size. Re-GARCH stands for the Realized-GARCH
type models. Re-ES-CARE and Re-T-ES-CARE represent Realized-ES-CARE and Realized-T-ES-CARE type models respectively.
\end{table}
The quantile loss results are presented in Table \ref{quanitl_loss_table} for each model for each series. The average rank based on ranks of quantile loss across 7 markets is calculated and shown in the ``Avg Rank'' column. 6 of the 7 series have the lowest loss produced by one of the Realized(-Threshold)-ES-CARE type models (4 from Realized-Threshold-ES-CARE type models and 2 from Realized-ES-CARE type models). The best average ranked models are Realized-ES-CARE and Realized-Threshold-ES-CARE frameworks employing SSRV. The quantile loss values for the ES-CAViaR-Add and CARE models are on average slightly higher than that of ES-CARE model, meaning ES-CARE model produces more accurate VaR forecasting results. Further, the quantile loss values of Realized-(Threshold)-ES-CARE type models are clearly lower compared to ES-CARE model, which proves that the extra efficiency can be gained by employing the realized measures through the measurement equation.
\begin{table}[hbt!]
\begin{center}
\caption{\label{quanitl_loss_table} \small 1\% VaR Forecasting quantile loss on 7 indices.}\tabcolsep=10pt
\tiny
\begin{tabular}{lccccccccccc} \hline
Model&S\&P500&NASDAQ&HK&FTSE&DAX&SMI&ASX200&Avg Rank\\
\hline
GARCH-t&81.8&92.1&98.4&81.5&93.4&\textcolor{red}{88.0}&69.7&\textcolor{red}{19.86}\\
EGARCH-t&80.3&92.2&90.3&76.9&92.2&83.1&67.3&13.29\\
GJR-GARCH-t &77.6&89.8&92.2&77.9&\textcolor{red}{93.9}&85.7&67.9&14.43\\
GARCH-t-HS &81.8&91.5&96.9&80.3&93.9&86.3&69.5&18.71\\
GARCH-Skew-t&81.7&90.9&97.3&80.1&93.6&86.1&69.8&18.29\\
T-GARCH-Skew-t&76.1&87.3&91.3&78.4&91.5&83.8&68.3&10.71\\
CARE &\bf{84.2}&\bf{95.5}&93.0&\bf{82.7}&93.3&\bf{89.8}&\bf{77.3}&\bf{21.00}\\
Re-GARCH-RV-GG &80.0&87.3&\bf{119.0}&78.1&\bf{95.2}&83.4&66.1&15.00\\
Re-GARCH-RV-tG &77.1&\fbox{85.3}&\textcolor{red}{108.6}&77.0&91.7&82&65.4&10.29\\
ES-CAViaR-Add&\bf{84.2}&\textcolor{red}{93.5}&94.9&81.1&92.8&86.3&\textcolor{red}{71.9}&19.57\\
ES-CAViaR-Mult&\textcolor{red}{83.5}&93.3&95.3&\textcolor{red}{81.7}&93.2&85.7&71.7&19.29\\
ES-CARE&83.4&93.3&95.3&81.6&93.2&85.7&71.8&19.14\\
Re-ES-CARE-RV&76.1&91.6&106.2&76.5&91.6&81.4&\textcolor{blue}{65.2}&11.14\\
Re-ES-CARE-RR&\textcolor{blue}{73.2}&86.5&101.3&76.3&90.1&79.0&67.4&8.14\\
Re-ES-CARE-ScRV&77.3&89.8&96.5&76.4&93.9&82.9&66.3&12.86\\
Re-ES-CARE-ScRR&76.0&89.4&91.8&76.3&91.0&80.6&67.8&8.43\\
Re-ES-CARE-SSRV&74.8&87.4&91.1&75.8&90.2&\textcolor{blue}{78.7}&66.2&\fbox{5.43}\\
Re-ES-CARE-SSRR&\fbox{72.7}&\textcolor{blue}{86.2}&96.9&75.5&89.7&\fbox{78.6}&66.7&\textcolor{blue}{5.71}\\
Re-T-ES-CARE-RV&78.4&91.5&102.2&\fbox{74.8}&91.9&80.6&\fbox{65.0}&10.29\\
Re-T-ES-CARE-RR&76.6&88.5&98.8&75.5&89.5&80.0&66.2&8.14\\
Re-T-ES-CARE-ScRV&79.9&90.4&92.4&74.9&92.2&82.2&65.8&10.00\\
Re-T-ES-CARE-ScRR&78.5&90.2&\textcolor{blue}{89.3}&\textcolor{blue}{74.9}&90.1&81.6&66.6&8.29\\
Re-T-ES-CARE-SSRV&77.8&88.0&\fbox{88.2}&75.1&\textcolor{blue}{89.2}&80.0&65.9&\fbox{5.43}\\
Re-T-ES-CARE-SSRR&76.3&87.7&94.5&75.7&\fbox{89.1}&79.7&66.5&6.57\\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small Box indicates the favoured model, blue shading indicates the 2nd ranked model, bold indicates the least favoured model, red shading indicates the 2nd lowest ranked model, in each column.
\end{table}
Figure \ref{var_forecast_fig} and \ref{var_forecast_fig1} provide further evidence on how and why the proposed Realized-ES-CARE framework generates clearly lower quantile loss compared with other models, combined
with relatively accurate VRates. Specifically, the VaR violation rates for the GARCH-Skew-t, ES-CAViaR-Mult and Realized-ES-CARE-RR models are 1.088\%, 1.278\% and 1.088\% respectively, for the S\&P500 returns, e.g. the three models are very similar by that metric. However, from Table \ref{quanitl_loss_table}, the quantile loss values for the 3 models are 81.7, 83.5 and 73.2 respectively, meaning the Realized-ES-CARE-RR model is the
most accurate model having clearly the lowest quantile loss. Through close inspection of Figure \ref{var_forecast_fig1}, the GARCH-Skew-t and ES-CAViaR-Mult have VaR forecasts typically quite close together in value, driving their close quantile loss values. However, both these models generate clearly more extreme (in the negative direction) VaR forecasts
on most days in the US market, compared to the Realized-ES-CARE-RR. This means that the capital set aside by financial institutions to cover extreme losses, based on such VaR forecasts, is usually at a higher level for the GARCH-Skew-t or ES-CAViaR-Mult models, than for the Realized-ES-CARE-RR.
In other words, the Realized-ES-CARE-RR model produces VaR forecasts that are relatively close to nominal VRate and are closer to the true VaR series, as measured by the loss function. VaR forecasts of this model are also closer to the data and less extreme, implying that lower amounts of capital are needed to protect against market risk. Given the forecasting steps $m=2113$ for S\&P 500, the forecasts from Realized-ES-CARE-RR were less extreme than those from ES-CAViaR-Mult on 1396 days (66\%) in the forecast period, which clearly demonstrates the advantageous of employing the RR through the measurement equation.
This suggests a higher level of information (and cost) efficiency regarding risk levels for the Realized-ES-CARE-RR model, likely coming from the improved the model specification and increased statistical efficiency of the realized range series over squared returns, compared to the ES-CAViaR-Mult and GARCH-Skew-t models. Since the economic capital is determined by financial institutions' own model and should be
directly proportional to the VaR forecast, the Realized-ES-CARE-RR model is able to decrease the cost capital allocation and increase the profitability of
these institutions, by freeing up part of the regulatory capital from risk coverage into investment, while still providing sufficient and more than
adequate protection against violations. The more accurate and often less extreme VaR forecasts produced by Realized-ES-CARE-RR are particularly
strategically important to the decision makers in the financial sector. This extra efficiency is also often observed for the Realized(-Threshold)-ES-CARE type
models in the other markets/assets.
Further, during the periods with high volatility including GFC, when there is a persistence of extreme returns, the Realized-ES-CARE-RR VaR forecasts "recover" the fastest among the 3 models, presented through close inspection as in Figure \ref{var_forecast_fig1}, in terms of being marginally the fastest to produce forecasts that again rejoin and follow the tail of the return data. Traditional GARCH models tend to over-react to extreme events and to be subsequently very slow to recover, due to their frequently estimated very high level of persistence, as discussed in Harvey and Chakravarty (2009). Realized(-Threshold)-ES-CARE models clearly improve the performance on this aspect. Generally, the Realized(-Threshold)-ES-CARE models better describe the dynamics in the volatility, compared to the traditional GARCH model and ES-CAViaR-Mult type models, thus largely improving the responsiveness and accuracy of the risk level forecasts, especially after high volatility periods.
\begin{figure}[hbt!]
\centering
\includegraphics[width=.7\textwidth]{sp500_var.eps}
\caption{\label{var_forecast_fig} S\&P 500 VaR forecasts with GARCH-Skew-t, ES-CAViaR-Mult and Realized-ES-CARE-RR.VRates: 1.088\%, 1.278\% and 1.088\%.}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=.7\textwidth]{sp500_var_zoom_in.eps}
\caption{\label{var_forecast_fig1} S\&P 500 VaR forecasts (zoomed in) with GARCH-Skew-t, ES-CAViaR-Mult and Realized-ES-CARE-RR. VRates: 1.088\%, 1.278\% and 1.088\%. Quantile loss: 81.7, 83.5 and 73.2.}
\end{figure}
Several tests are employed to statistically assess the forecast accuracy and independence of violations from each VaR forecast model.
Table \ref{var_backtest_table} shows the number of return series (out of 7) in which each 1\% VaR forecast model is rejected for each test,
conducted at a 5\% significance level. The Realized(-Threshold)-ES-CARE type models are generally less or equally likely to be rejected by the back tests
compared to other models. The Realized-Threshold-ES-CARE-ScRV achieves the least number of rejections and is rejected once, followed by Gt-HS, T-GARCH-Skew-t, Realized-ES-CARE-RV, Realized-Threshold-ES-CARE-RV and Realized-Threshold-ES-CARE-SSRV models. The G-t and Realized-GARCH-GG are rejected in all 7 series, and the EGARCH-t and GJR-GARCH-t models are rejected in 6 series, respectively.
\begin{table}[hbt!]
\begin{center}
\caption{\label{var_backtest_table} \small Counts of 1\% VaR rejections with UC, CC, IND, DQ and VQR tests for different models on 7 indices.}\tabcolsep=10pt
\tiny
\begin{tabular}{lccccccccccc} \hline
Model&UC&CC1&DQ1&DQ4&VQR&Total\\
\hline
GARCH-t&6&6&7&5&4&\bf{7}\\
EGARCH-t&5&3&4&5&2&\textcolor{red}{6}\\
GJR-GARCH-t &5&3&6&4&3&\textcolor{red}{6}\\
GARCH-t-HS &1&1&1&2&0&\textcolor{blue}{2}\\
GARCH-Skew-t&0&0&1&3&0&3\\
T-GARCH-Skew-t&0&0&0&2&0&\textcolor{blue}{2}\\
CARE &1&1&0&4&0&4\\
Re-GARCH-RV-GG &7&7&7&7&4&\bf{7}\\
Re-GARCH-RV-tG &3&2&2&1&3&3\\
ES-CAViaR-Add&2&0&0&3&0&4\\
ES-CAViaR-Mult&0&0&0&3&0&3\\
ES-CARE&0&0&0&3&1&4\\
Re-ES-CARE-RV&2&2&2&2&2&\textcolor{blue}{2}\\
Re-ES-CARE-RR&0&1&1&2&2&3\\
Re-ES-CARE-ScRV&1&1&1&2&2&3\\
Re-ES-CARE-ScRR&1&1&2&3&0&3\\
Re-ES-CARE-SSRV&1&1&2&4&2&4\\
Re-ES-CARE-SSRR&0&1&1&2&2&3\\
Re-T-ES-CARE-RV&2&1&1&1&1&\textcolor{blue}{2}\\
Re-T-ES-CARE-RR&1&1&2&2&2&3\\
Re-T-ES-CARE-ScRV&0&0&0&1&0&\fbox{1}\\
Re-T-ES-CARE-ScRR&1&1&2&2&0&3\\
Re-T-ES-CARE-SSRV&0&1&1&2&0&\textcolor{blue}{2}\\
Re-T-ES-CARE-SSRR&1&1&2&2&1&3\\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small Box indicates the model with least number of rejections, blue shading indicates the model with 2nd least number of rejections, bold indicates the model with the highest number of rejections, red shading indicates the model 2nd highest number of rejections.
All tests are conducted at 5\% significance level.
\end{table}
\subsection{\normalsize Expected Shortfall and Expectile Level}
The same 24 models are employed to generate one step ahead forecasts of 1\% ES for all 7 series during the forecasting period. Before checking the ES forecasting results, Figure \ref{tau_comparision figure} visualizes the 2113 estimated expectile level $\tau$ parameters from ES-CARE and CARE models during the forecasting period of S\&P 500. The $\tau$ of CARE is selected based on the grid search and violation rate approach of Taylor (2008), as discussed in Section \ref{model_review_section}.
Although in general the $\tau$ values estimated by ES-CARE and CARE are close to each other, we can clearly see the ES-CARE model produces more dynamic $\tau$ values, e.g. during the 2008 GFC and the mid-2012 to 2014 period. Especially, the estimated $\tau$ from ES-CARE is much more responsive to the volatility jump. Such improved responsiveness will potentially improve the accuracy of tail risk forecasts. As discussed in Section \ref{model_section}, the selection procedure of $\tau$ with the ES-CARE model is based on a strictly consistent VaR\&ES joint loss function, unlike the grid search approach. More results will be provided in the next section to support the improved ES forecasting performance from ES-CARE compared to CARE.
In addition, during the low volatility period, e.g. mid-2014 to mid-2015, the estimated $\tau$ values from ES-CARE model are clearly smaller than those from CARE. Based on Equation (\ref{expectile_es_equation}), the ratio between ES and VaR is linearly proportional to $\tau$. Therefore, the ES to VaR ratios produced by ES-CARE model, during mid-2014 to mid-2015 low volatility period (less cases of extreme returns), are smaller than that produced by CARE. Such results are consistent with the definition of VaR and ES, and lend some evidence on why the VaR and ES forecasts generated by ES-CARE are more efficient than CARE, to be further discussed in the following section.
\begin{figure}[htp]
\centering
\includegraphics[width=.9\textwidth]{tau_comparison.eps}
\caption{\label{tau_comparision figure} For S\&P 500 forecasting, top plot visualizes S\&P 500 returns, and bottom plot visualizes $m=2113$ estimated $\tau$ parameters from CARE and ES-CARE models.}
\end{figure}
Further, as discussed in Section \ref{model_review_section}, the ES component in the ES-CAViaR-Add framework would produce the in-sample $w_{t}$ as presented in Figure \ref{wt_example}, resulting ES dynamics that could be potentially improved. Figure \ref{xt_comparision figure} visualize $m=2113$ VaR and ES forecasts differences from ES-CAViaR-Add and ES-CARE models. Clearly, more dynamic ES and VaR differences are produced by the ES-CARE which incorporates a more flexible ES regression component compared with ES-CAViaR-Add. More specifically, the difference between VaR and ES should be larger during the high volatility period, based on the definition of ES, e.g. as illustrated in the 2008 GFC period. However, taking the period of early-2009 to mid-2009 as example, apparently the ES-CARE creates more responsive ES forecasting results than ES-CAViaR-Add. In the following Section, we will quantify the improvement of VaR\& ES forecasting results from ES-CARE model compared with ES-CAViaR-Add.
\begin{figure}[htp]
\centering
\includegraphics[width=.9\textwidth]{wt_fore_comparison.eps}
\caption{\label{xt_comparision figure} For S\&P 500 forecasting, top plot visualizes S\&P 500 returns, and bottom plot visualizes $m=2113$ VaR and ES forecasts differences from ES-CAViaR-Add and ES-CARE models.}
\end{figure}
\subsubsection{\normalsize VaR\&ES Joint Loss Function}
In this section, the joint VaR\&ES loss function study is conducted to compare the models VaR and ES forecasts jointly, and to help clarify and quantify any
extra efficiency can be gained from the Realized(-Threshold)-ES-CARE ES forecasts compared to its competitors.
Fissler and Ziegel (2016) developed a family of loss functions that are a joint function of the associated VaR and ES series. This loss function family are
strictly consistent for the true VaR and ES series, i.e. they are uniquely minimized by the true VaR and ES series. The general function family form is:
\begin{eqnarray*}
S_t(r_t, VaR_t, ES_t) &=& (I_t -\alpha)G_1(VaR_t) - I_tG_1(r_t) + G_2(ES_t)\left(ES_t-VaR_t + \frac{I_t}{\alpha}(VaR_t-r_t)\right) \\
&-& H(ES_t) + a(r_t) \, ,
\end{eqnarray*}
where $I_t=1$ if $r_t<VaR_t$ and 0 otherwise for $t=1,\ldots,T$, $G_1()$ is increasing, $G_2()$ is strictly increasing and strictly convex,
$G_2 = H^{'}$ and $\lim_{x\to -\infty} G_2(x) = 0$ and $a(\cdot)$ is a real-valued integrable function.
As discussed in Taylor (2017), making the choices: $G_1(x) =0$,
$G_2(x) = -1/x$, $H(x)= -\text{log}(-x)$ and $a= 1-\text{log} (1-\alpha)$, which satisfy the required criteria, returns the
scoring function (defined $r_t$ to have zero mean):
\begin{eqnarray}\label{es_caviar_log_score}
S_t(r_t, VaR_t, ES_t) = -\text{log} \left( \frac{\alpha-1}{\text{ES}_t} \right) - {\frac{(r_t-Q_t)(\alpha-I(r_t\leq Q_t))}{\alpha \text{ES}_t}},
\end{eqnarray}
where the loss function is $S = \sum_{t-1}^T S_t$. Taylor (2017) referred expression (\ref{es_caviar_log_score}) as AL log score. Compared with the likelihood function as in Equation (\ref{es_caviar_like_equation}), Equation (\ref{es_caviar_log_score}) is exactly the negative of the AL log-likelihood, and is a strictly consistent scoring rule that is jointly minimized by the true VaR and ES series. We use this to informally and jointly assess and compare the VaR and ES forecasts from all models.
Tables \ref{veloss} shows the loss function values $S$, calculated using Equation (\ref{es_caviar_log_score}), which jointly assesses the
accuracy of each model's VaR and ES forecasting series, during the forecast period for each market. Generally, the Realized(-Threshold)-ES-CARE models are better ranked with lower loss than other models in most markets. For all 7 markets, the best ranked models are from Realized(-Threshold)-ES-CARE families, with Realized-Threshold-ES-CARE employing SSRV and SSRR achieving the best performance overall.
In addition, overall the VaR \& ES joint loss values from Realized-Threshold-ES-CARE specifications are consistently lower than that from Realized-ES-CARE specifications, under the same choice of realized measures (except RR). For example, Realized-Threshold-ES-CARE-SSRR is better ranked than Realized-ES-CARE-SSRR. This proves the validity of the threshold specifications for the VaR to ES dynamics.
Lastly, overall the ES-CARE model is better ranked than ES-CAViaR-Add, ES-CAViaR-Mult and CARE models. This again proves the validity of the proposed ES-CARE framework, even without incorporating the realized measures.
\begin{table}[hbt!]
\begin{center}
\caption{\label{veloss} \small VaR and ES joint loss function values across the markets; $\alpha=1\%$.}\tabcolsep=10pt
\tiny
\begin{tabular}{lccccccccccccc} \hline
Model&S\&P500&NASDAQ&HK&FTSE&DAX&SMI&ASX200&Avg Rank\\
\hline
GARCH-t&4795.0&5067.2&5144.4&4872.1&\textcolor{red}{5285.2}&\textcolor{red}{4987.1}&4531.9&\bf{20.57}\\
EGARCH-t&4800.9&5068.7&4985.2&4837.6&5277.3&4905.2&4503.7&17.71\\
GJR-GARCH-t &4665.5&4967.7&5009.9&4793.8&\bf{5315.8}&\bf{4993.5}&4475.2&16.29\\
GARCH-t-HS &4768.8&5031.3&5100.9&4811.8&5274.5&4884.3&4510.4&18.00\\
GARCH-Skew-t&4758.9&5008.6&5104.5&4814.6&5254.1&4882.7&4508.8&17.57\\
T-GARCH-Skew-t&4613.0&4909.4&4966.4&4800.4&5195.9&4873.3&4472.7&10.86\\
CARE &\textcolor{red}{4836.7}&\bf{5201.5}&5018.5&\bf{4890.9}&5231.8&4973.2&\bf{4793.4}&\textcolor{red}{20.29}\\
Re-GARCH-RV-GG &4706.0&4948.3&\bf{5673.1}&4768.5&5275.2&4878.0&4432.7&15.71\\
Re-GARCH-RV-tG &4590.8&\textcolor{blue}{4875.4}&5288.3&4706.4&5146.0&4778.0&4386.4&10.14\\
ES-CAViaR-Add&\bf{4844.0}&\textcolor{red}{5099.7}&5069.0&4859.6&5237.1&4941.2&\textcolor{red}{4596.9}&20.00\\
ES-CAViaR-Mult&4833.6&5068.2&5071.5&\textcolor{red}{4875.6}&5234.8&4909.0&4589.5&19.71\\
ES-CARE&4832.6&5068.1&5069.8&4874.0&5234.2&4907.7&4591.2&18.86\\
Re-ES-CARE-RV&4551.6&4983.2&\textcolor{red}{5295.9}&4680.9&5152.0&4757.2&\textcolor{red}{4380.0}&11.29\\
Re-ES-CARE-RR&\textcolor{blue}{4477.6}&4877.9&5149.3&4667.8&5094.6&4718.2&4459.6&7.86\\
Re-ES-CARE-ScRV&4599.6&4963.2&5090.2&4676.8&5210.8&4759.2&4411.0&11.71\\
Re-ES-CARE-ScRR&4548.3&4934.1&4969.9&4675.5&5120.1&\textcolor{blue}{4705.9}&4480.1&7.71\\
Re-ES-CARE-SSRV&4505.1&4897.6&4974.7&4672.1&5104.2&\fbox{4702.2}&4415.3&5.29\\
Re-ES-CARE-SSRR&\fbox{4455.9}&\fbox{4872.1}&5071.0&4656.4&5089.9&4715.3&4436.2&5.71\\
Re-T-ES-CARE-RV&4615.3&4999.2&5241.7&\textcolor{blue}{4645.1}&5140.5&4753.5&\fbox{4376.1}&10.14\\
Re-T-ES-CARE-RR&4555.9&4926.2&5082.3&4655.4&5077.7&4745.1&4420.0&8.00\\
Re-T-ES-CARE-ScRV&4678.2&4993.3&4995.2&\fbox{4644.8}&5159.3&4748.5&4410.4&9.29\\
Re-T-ES-CARE-ScRR&4626.0&4958.5&\textcolor{blue}{4923.0}&4646.2&5100.4&4727.3&4449&7.43\\
Re-T-ES-CARE-SSRV&4591.3&4920.4&\fbox{4919.9}&4652.2&\textcolor{blue}{5071.7}&4731.0&4405.0&\fbox{4.86}\\
Re-T-ES-CARE-SSRR&4542.6&4914.1&5008.8&4650.9&\fbox{5070.9}&4739.9&4408.5&\textcolor{blue}{5.00}\\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small Box indicates the favoured model, blue shading indicates the 2nd ranked model, bold indicates the least favoured model,
red shading indicates the 2nd lowest ranked model, in each column.
\end{table}
To further demonstrate the extra forecasting efficiently can be gained by employing the proposed models, Figure \ref{Fig_es_fore_zoom_in} visualizes the ES forecasts from CARE, Threshold-GARCH-Skew-t and Realized-Threshold-ES-CARE-SSRR. Specifically, the ES violation rate of these three models are: 0.284\%, 0.426\% and 0.473\% respectively,
for S\&P500. As studied in Gerlach and Chen (2016, Table 1), the nominal ES quantile levels only have very small variations across different distributions or different degrees of freedom for a Student-t or skewed-t. In general, for 1\% ES the quantile levels that ES to fall is between 0.34\% and 0.37\%. Based on the models with Student-t errors, the implied quantile level that the 1\% ES is estimated to fall at is $\approx$ 0.35\% which is used as the target ES violation rate for the semi-parametric models. If this nominal level is accurate, then the
CARE has the conservative ES violation rate, while the Threshold-GARCH-Skew-t and Realized-Threshold-ES-CARE-SSRR model are slightly anti-conservative (neither are significantly
different to 0.35\% by the UC test).
However, through closer inspection of Figure \ref{Fig_es_fore_zoom_in}, the cost efficiency gains from the Realized-Threshold-ES-CARE-SSRR model are again observed, in a
similar and even clearer pattern to that from the VaR forecasting study. The CARE model achieves a lower than nominal ES VRate by generating relatively more extreme ES forecasts
than the Realized-Threshold-ES-CARE-SSRR model's on 1654 days (78\%). In addition, the ES forecasts from Threshold-GARCH-Skew-t model are more extreme than Realized-Threshold-ES-CARE-SSRR on 1471 days (70\%).
Therefore, the Realized-Threshold-ES-CARE-SSRR can improve the forecast efficiency and lead to lower capital allocations to protect against extreme returns, compared with the CARE and Threshold-GARCH-Skew-t, while still achieving an accurate violation rate. Again, such extra efficiency is also frequently observed for the Realized(-Threshold)-ES-CARE type models in other time series.
Here, compared with the VaR forecasting, we would like to emphasize the extra efficiency produced by the Realized(-Threshold)-ES-CARE models are more prominent, compared with the original ES-CAViaR and conventional GARCH models. The results lend evidence on fact that the newly developed Realized(-Threshold)-ES-CARE frameworks can produce more accurate and efficient VaR and ES forecasts.
\begin{figure}[hbt!]
\centering
\includegraphics[width=.7\textwidth]{sp500_es_zoom_in.eps}
\caption{\label{Fig_es_fore_zoom_in} S\&P 500 ES forecasts (zoomed in) with CARE, Threshold-GARCH-Skew-t and Realized-Threshold-ES-CARE-SSRR. ESRates: 0.284\%, 0.426\% and 0.473\%. VaR\&ES join loss: 4,836.7, 4613.0, and 4542.6.}
\end{figure}
\subsubsection{\normalsize Model Confidence Set}
The model confidence set (MCS) (Hansen, Lunde and Nason, 2011) is utilized to statistically compared the tested models via the VaR and ES joint loss function.
A MCS is a set of models that is constructed such that it will contain the best model with a given level of confidence (90\% is used in our paper). The Matlab code for MCS testing was
downloaded from "www.kevinsheppard.com/MFE\_Toolbox". We adapted code to incorporate the VaR and ES joint loss function values
(Equation (\ref{es_caviar_log_score})) as the loss function during the MCS calculation. The R method which uses absolute values sum during the calculation of test statistic is employed in our paper, details as in page 465 of Hansen, Lunde and Nason (2011).
Table \ref{mcs_r} presents the 90\% MCS using the R methods. Column "Total" counts the total number of times that a model is included in the 90\% MCS across the 7 return series.
Via the R method, 4 Realized(-Threshold)-ES-CARE models, including Realized-ES-CARE-SSRV, Realized-ES-CARE-SSRR, Realized-Threshold-ES-CARE-RR and Realized-Threshold-ES-CARE-SSRR, are included in the MCS for all 7 markets. There are other 4 Realized(-Threshold)-ES-CARE models included in the MCS for 6 times, together with EGARCH-t and Threshold-GARCH-Skew-t.
The GARCH-t and GARCH-Skew-t are only included the MCS for 3 times respectively.
\begin{table}[hbt!]
\begin{center}
\caption{\label{mcs_r} \small 90\% model confidence set results summary with R method, $\alpha=1\%$.}\tabcolsep=10pt
\tiny
\begin{tabular}{lccccccccccccc} \hline
Model&S\&P500&NASDAQ&HK&FTSE&DAX&SMI&ASX200&Total\\
\hline
GARCH-t&0&1&0&1&0&1&0&\bf{3}\\
EGARCH-t&0&1&1&1&1&1&1&\textcolor{blue}{6}\\
GJR-GARCH-t &0&1&1&1&0&1&1&5\\
GARCH-t-HS &0&1&0&1&0&1&0&\bf{3}\\
GARCH-Skew-t&0&1&0&1&1&1&0&\textcolor{red}{4}\\
T-GARCH-Skew-t&0&1&1&1&1&1&1&\textcolor{blue}{6}\\
CARE &0&0&1&1&1&1&0&\textcolor{red}{4}\\
Re-GARCH-RV-GG &0&1&0&1&0&1&1&\textcolor{red}{4}\\
Re-GARCH-RV-tG &0&1&0&1&1&1&1&5\\
ES-CAViaR-Add&0&0&1&1&1&1&0&\textcolor{red}{4}\\
ES-CAViaR-Mult&0&1&1&1&1&1&0&5\\
ES-CARE&0&1&1&1&1&1&0&5\\
Re-ES-CARE-RV&0&1&0&1&1&1&1&5\\
Re-ES-CARE-RR&1&1&0&1&1&1&0&5\\
Re-ES-CARE-ScRV&0&1&0&1&0&1&1&\textcolor{red}{4}\\
Re-ES-CARE-ScRR&1&1&1&1&1&1&0&\textcolor{blue}{6}\\
Re-ES-CARE-SSRV&1&1&1&1&1&1&1&\fbox{7}\\
Re-ES-CARE-SSRR&1&1&1&1&1&1&1&\fbox{7}\\
Re-T-ES-CARE-RV&0&1&0&1&1&1&1&5\\
Re-T-ES-CARE-RR&1&1&1&1&1&1&1&\fbox{7}\\
Re-T-ES-CARE-ScRV&0&1&1&1&1&1&1&\textcolor{blue}{6}\\
Re-T-ES-CARE-ScRR&0&1&1&1&1&1&1&\textcolor{blue}{6}\\
Re-T-ES-CARE-SSRV&0&1&1&1&1&1&1&\textcolor{blue}{6}\\
Re-T-ES-CARE-SSRR&1&1&1&1&1&1&1&\fbox{7}\\
\hline
\end{tabular}
\end{center}
\emph{Note}:\small Boxes indicate the favoured model, blue shading indicates the 2nd ranked model, bold indicates the least favoured model,
red shading indicates the 2nd lowest ranked model, based on total number of included in the MCS across the 7 markets, higher is better.
\end{table}
{\centering
\section{\normalsize CONCLUSION}\label{conclusion_section}
\par
}
\noindent
In this paper, we propose a realized joint conditional autoregressive expectile and expected shortfall framework which is further extended through incorporating nonlinear specifications. Improvements in the out-of-sample forecasting of tail risk measures are observed, compared to Realized-GARCH model employing realized volatility, and
traditional GARCH and CARE models, as well as the original ES-CAViaR models. Specifically, Realized(-Threshold)-ES-CARE frameworks employing sub-sampled RV and sub-sampled RR generate the best VaR and ES forecasting results in the empirical study of 7 financial return series. With respect to the back testing of VaR forecasts, the Realized(-Threshold)-ES-CARE type models are also generally less likely to be rejected than their counterparts.
Further, the model confidence set results also apparently favour the proposed Realized(-Threshold)-ES-CARE frameworks.
In addition to being more accurate, the Realized(-Threshold)-ES-CARE models generated less extreme tail risk forecasts, regularly allowing smaller amounts of capital allocation without being anti-conservative or significantly inaccurate. Further, even without incorporating the realized measures, the ES-CARE model is still favourable compared with CARE and ES-CAViaR models under almost all the measures and tests considered.
To conclude, the Realized(-Threshold)-ES-CARE type models, especially the ones use sub-sampled RV and sub-sampled RR, should be considered for financial applications when
forecasting tail risk, and should allow financial institutions to more accurately allocate capital under the Basel Capital Accords,
to protect their investments from extreme market movements. This work could be extended by using alternative frequencies of observation for the realized measures; by extending the framework to allow multiple realized measures to appear simultaneously in the model, etc.
\clearpage
\section*{References}
\addcontentsline{toc}{section}{References}
\begin{description}
\item Aigner, D.J. ,Amemiya, T., and Poirier, D. J. (1976). On the Estimation of Production
Frontiers: Maximum Likelihood Estimation of the Parameters of a Discontinuous Density
Function. \emph{International Economic Review}, 17, 377-396.
\item Andersen, T. G. and Bollerslev, T. (1998). Answering the skeptics: Yes, standard volatility models do provide accurate
forecasts. \emph{International economic review}, 885-905.
\item Andersen, T. G., Bollerslev, T., Diebold, F. X. and Labys, P. (2003). Modeling and forecasting realized volatility.
\emph{Econometrica}, 71(2), 579-625.
\item Artzner, P., Delbaen, F., Eber, J.M., and Heath, D. (1997). Thinking coherently. \emph{Risk}, 10, 68-71.
\item Artzener, P., Delbaen, F., Eber, J.M., and Heath, D. (1999). Coherent measures of risk. \emph{Mathematical Finance}, 9, 203-228.
\item Avdulaj, K. and Barunik, J. (2017). A semiparametric nonlinear quantile regression model for financial returns. \emph{Studies in Nonlinear Dynamics \& Econometrics}, 21(1), 81-97.
\item Bollerslev, T. (1986). Generalized Autoregressive Conditional Heteroskedasticity. \emph{Journal of Econometrics}, 31, 307-327.
\item Brooks, C. (2001). A Double‐threshold GARCH Model for the French Franc/Deutschmark exchange rate. \emph{Journal of Forecasting}, 20(2), 135-143.
\item Chen, W., Peters, G., Gerlach, R. and Sisson, S. (2017). Dynamic Quantile Function Models. arXiv:1707.02587.
\item Christensen, K. and Podolskij, M. (2007). Realized range-based estimation of integrated variance. \emph{Journal of Econometrics},
141(2), 323-349.
\item Christoffersen, P. (1998). Evaluating interval forecasts. \emph{International Economic Review}, 39, 841-862.
\item Clements, M.P., Galv\~{a}o, A.B. and Kim, J.H. (2008). Quantile forecasts of daily exchange rate returns from forecasts of realized volatility. \emph{Journal of Empirical Finance}, 15(4), 729-750.
\item Creal, D., Koopman, S.J. and Lucas, A. (2013). Generalized autoregressive score models with applications. \emph{Journal of Applied Econometrics}, 28(5), 777-795.
\item Embrechts, P., Resnick, S.I. and Samorodnitsky, G. (1999). Extreme value theory as a risk management tool. \emph{North American Actuarial Journal}, 3(2), pp.30-41.
\item Engle, R. F. (1982), Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of United Kingdom
Inflations. \emph{Econometrica}, 50, 987-1007.
\item Engle, R. F. and Manganelli, S. (2004). CAViaR: Conditional Autoregressive Value at Risk
by Regression Quantiles. \emph{Journal of Business and Economic Statistics}, 22, 367-381.
\item Fissler, T. and Ziegel, J. F. (2016). Higher order elicibility and Osband's principle. \emph{Annals of Statistics}, in press.
\item Gaglianone, W. P., Lima, L. R., Linton, O. and Smith, D. R. (2011). Evaluating Value-
at-Risk models via quantile regression. \emph{Journal of Business and Economic Statistics},
29, 150-160.
\item Garman, M. B. and Klass, M. J. (1980). On the Estimation of Security Price Volatilities from historical data.
\emph{The Journal of Business}, 67-78.
\item Gelman, A., Carlin, J.B., Stern, H.S. and Rubin, D.B. (2014). \emph{Bayesian data analysis (Vol. 2)}. Boca Raton, FL: CRC press.
\item Gerlach, R, Chen, C.W.S. and Chan, N.Y. (2011). Bayesian time-varying quantile forecasting for value-at-risk in financial markets. \emph{Journal of Business \& Economic Statistics}, 29(4), 481-492.
\item Gerlach, R. and Chen, C.W.S. (2016). Bayesian Expected Shortfall Forecasting Incorporating the Intraday Range,
\emph{Journal of Financial Econometrics}, 14(1), 128-158.
\item Gerlach, R. and Wang, C. (2016a). Forecasting risk via realized GARCH, incorporating the realized range.
\emph{Quantitative Finance}, 16:4, 501-511.
\item Gerlach, R. and Wang, C. (2016b). Bayesian Semi-parametric Realized Conditional Autoregressive Expectile Models for Tail Risk Forecasting. arXiv preprint arXiv:1612.08488.
\item Gerlach, R., Walpole, D. and Wang, C. (2017). Semi-parametric Bayesian Tail Risk Forecasting Incorporating Realized Measures of Volatility,
\emph{Quantitative Finance}, 17:2, 199-215.
\item Gilli, M. and Kellezi, E. (2006). An Application of Extreme Value Theory for Measuring Financial Risk. \emph{Computational Economics}, 27, 1–23.
\item Giot, P. and Laurent, S. (2004). Modelling daily value-at-risk using realized volatility and ARCH type models. \emph{Journal of Empirical Finance}, 11(3), 379-398.
\item Glosten, L.R., Jagannathan, R. and Runkle, D.E. (1993). On the relation between the expected value and the volatility of the nominal excess return on stocks. \emph{The journal of finance}, 48(5), 1779-1801.
\item Hansen, P. R., Huang, Z. and Shek, H. H. (2012). Realized GARCH: a joint model for returns and realized measures of volatility. \emph{Journal of Applied Econometrics}, 27(6), 877-906.
\item Hansen, P.R., Lunde, A. and Nason, J.M. (2011). The model confidence set. \emph{Econometrica}, 79(2), 453-497.
\item Harvey, A.C. and Chakravarty, T. (2009). Beta-t-EGARCH. Working paper. Earlier version appeared in 2008 as a Cambridge Working paper in Economics, CWPE 0840.
\item Harvey, A.C. (2013). Dynamic Models for Volatility and Heavy Tails, Econometric Society Monograph 52, Cambridge University Press, Cambridge.
\item Koenker, R. and Machado, J.A. (1999). Goodness of fit and related inference processes for quantile regression. \emph{Journal of the american statistical association}, 94(448), 1296-1310.
\item Kupiec, P. H. (1995). Techniques for Verifying the Accuracy of Risk Measurement Models. \emph{The Journal of Derivatives}, 3, 73-84.
\item Li, C. W., and Li, W. K. (1996). On a Double-Threshold Autoregressive Heteroscedastic Time Series Model. \emph{Journal of Applied Econometrics}, 11(3), 253–274.
\item Brooks, C. (2001). A Double-Threshold GARCH Model for the French Franc/Deutschmark Exchange Rate. \emph{Journal of Forecasting}, 20, 135–143.
\item Martens, M. and van Dijk, D. (2007). Measuring volatility with the realized range. \emph{Journal of Econometrics}, 138(1), 181-207.
\item McNeil, A. J. and Frey, R. (2000). Estimation of Tail-Related Risk Measures for Heteroscedastic Financial Time Series: An Extreme Value Approach. \emph{Journal of Empirical Finance} 7, 271–300.
\item Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., and Teller, E. (1953). Equation of State Calculations by Fast
Computing Machines. \emph{J. Chem. Phys}, 21, 1087-1092.
\item Parkinson, M. (1980). The extreme value method for estimating the variance of the rate of return. \emph{Journal of Business},
53(1), 61.
\item Patton, A.J., Ziegel, J.F. and Chen, R. (2017). Dynamic semiparametric models for expected shortfall (and value-at-risk). arXiv preprint arXiv:1707.05108.
\item Roberts, G. O., Gelman, A. and Gilks, W. R. (1997). Weak convergence and optimal scaling of random walk Metropolis algorithms.
\emph{The annals of applied probability}, 7(1), 110-120.
\item Taylor, J. (2008). Estimating Value at Risk and Expected Shortfall Using Expectiles. \emph{Journal of Financial Econometrics}, 6,
231-252.
\item Taylor, J. (2017). Forecasting Value at Risk and Expected Shortfall Using a Semiparametric Approach Based on the Asymmetric Laplace Distribution. \emph{Journal of Business and Economic Statistics}, DOI:10.1080/07350015.2017.1281815.
\item Watanabe, T. (2012). Quantile Forecasts of Financial Returns Using Realized GARCH Models. \emph{Japanese Economic Review}, 63(1),
68-80.
\item Zhang, L., Mykland, P. A., and A\"{i}t-Sahalia, Y. (2005). A tale of two time scales. \emph{Journal of the American Statistical
Association}, 100(472).
\item \v{Z}ike\v{s}, F. and Barun\'{i}k, J. (2014). Semi-parametric conditional quantile models for financial returns and realized volatility. \emph{Journal of Financial Econometrics}, 14(1), 185-226.
\end{description}
\end{document} |
\section{\label{sec:level1}Introduction}
Materials containing honeycomb lattices decorated by metals with strong spin-orbit coupling are potential hosts for Kitaev quantum spin liquids (QSL)~\citep{kitaev2006anyons,PhysRevLett.105.027204}. The essential ingredient for understanding magnetism in these systems is a dominant bond-directional anisotropic exchange that leads to magnetic frustration~\citep{bond.directional.kitaev,bond.directional.kitaev2}. Recent attention to $4d$ and $5d$ transition metal-based candidates such as $\alpha$-RuCl$_3$~\citep{RuCl3} and A$_2$IrO$_3$ (A=Li, Na)~\citep{PhysRevLett.105.027204,Iridate} underscores the interest in this honeycomb structural motif, with the bond-directional anisotropy a consequence of the spin-orbit coupling derived $j_{\rm{eff}}$=1/2 ground state assigned to the transition metals. Decoration of the honeycomb lattice with rare-earth ions offers an alternative to $4d$- and $5d$-based materials, with YbCl$_3$ (isostructural with $\alpha$-RuCl$_3$) suggested as a potential Kitaev QSL candidate~\citep{xing2019rare}. Addionally, a recent theoretical treatment of spin-orbital entanglement in rare-earth honeycomb magnets by Luo and Chen~\citep{rare.kitaev}, and others~\citep{kitaev.afm,clark2019two} highlights the need to explore such systems.
A nearly ideal honeycomb lattice of rare earth ions is found in a family of compounds with general formula, R$_2$T$_3$X$_9$ ( \textit{R} is a rare-earth element, \textit{T} is a transition metal element and \textit{X} is a $p$-block element)~\citep{Y2Co3Ga9,EuIrGa_structure}. This family, typified by the Y$_2$Co$_3$Ga$_9$ structure~\citep{Y2Co3Ga9} occupies a large composition space and hosts a rich variety of electronic properties ranging from complex magnetically ordered states (Dy based compounds), mixed valence (Yb,Ce based compounds) and Kondo lattice behavior (Yb based compounds) ~\citep{DyCoAl,Ce2Pd9Sb3,Yb2Co3T9,Yb2RhIR3AlGa9,R2Co3Al9,Yb2.Konod.Lattice,U2Co3Al9,niermann2004preparation}.
The layered structure of R$_2$T$_3$X$_9$, along the [001] direction can be viewed as a stacking of two alternating types of layers~\citep{EuIrGa_structure} (Fig.~\ref{figure 1}). The \textit{R} and \textit{X} atoms form a planar layer with the composition R$_4$X$_6$, which contains a two-dimensional honeycomb-like arrangement of \textit{R} atoms. The other layer is strongly puckered with \textit{T} and \textit{X} atoms forming a hexagonal arrangement, with \textit{T}:Ga ratio of
1 : 2 (T$_6$X$_{12}$). Overall stacking is such that the resulting structure is orthorhombic with space group \textit{Cmcm} no. 63. The physical properties of these compounds depend on identity of \textit{X} and \textit{T} atoms that surround the \textit{R} atom. Depending on the ligand \textit{X} atoms, systems with the same R atom can be magnetic in some cases and non-magnetic in others~\citep{DyCoAl}. It is evident that the Al-containing compounds Yb$_2$T$_3$Al$_9$ are magnetic, whereas their Ga-containing analogues are paramagnets down to the lowest temperatures measured~\citep{DyCoAl,Yb2Co3T9,Yb2RhIR3AlGa9}. Thus, the interactions between \textit{R} atoms via the RKKY mechanism coupled with the large magnetocrystalline anisotropy due to crystal electric field (CEF) acting on \textit{R} atom could lead to complex magnetic states in this family of compounds. Indeed, multiple commensurate and incommensurate phases as well as field induced metamagnetic states have been reported~\citep{DyCoAl}. As such, this family of compounds presents an opportunity to understand the interplay between various hybridization strengths of different \textit{d}, and \textit{f} electrons and ligand atoms in a layered structure where the magnetic \textit{R} atoms form a slightly distorted honeycomb lattice.
Interestingly, Tb containing compounds, such as Tb$_2$T$_3$X$_9$, have not been explored and in this article we present our study on single crystalline {\rm Tb$_2$Ir$_3$Ga$_9$}\,\,. Magnetization measurements reveal a long-range antiferromangetic ordering at 12.5 K. The magnetic susceptibility is highly anisotropic, with strong preference for in plane magnetization. Two field-induced metamagnetic transitions separated by a plateau of M$\approx$M$_{\rm{s}}$/2, where M$_{\rm{s}}$ is the saturation moment, are observed when the field is applied along the crystallographic ${\bf a} $-axis, which is also the magnetic easy axis. A finite ferromagnetic-like response was observed along the ${\bf b} $-axis, although neither ${\bf b} $ nor ${\bf c} $-axes show field-induced metamagnetic transitions. Neutron powder diffraction supports the observed magnetic behaviors by revealing the zero field ground state to be a collinear two-sublattice antiferromagnet (AFM) with spins along the ${\bf a} $ axis. Based on observed properties, we present a phenomenological spin model to describe the magnetic behavior for {\rm Tb$_2$Ir$_3$Ga$_9$}\,\,. Notably, incorporation of a bond-directional anisotropy term with significant weight, is essential for agreement between theory and experiment, making contact with the phenomenology of transition-metal based honeycomb magnets.
This article is arranged as follows. In section II, experimental methods and processes are described. In section III, results of magnetization, transport, specific heat, X-ray magnetic circular dichroism (XMCD), neutron powder diffraction (NPD) measurements, and the electronic structure calculations are presented. In section IV, a phenomenological theory of the magnetization in Tb$_2$Ir$_3$Ga$_9$ is presented. Finally, in section V we conclude with discussion of the ground state properties of Tb$_2$Ir$_3$Ga$_9$.
\section{\label{sec:level2}Experimental description}
Single crystals of Tb$_2$Ir$_3$Ga$_9$ were synthesized using a Ga-flux method. The starting materials, Tb pieces (Alfa Aesar, 99.9$\%$), Ir powder (Alfa Aesar, 99.99$\%$), and Ga pellet (Alfa Aesar, 99.9999$\%$) in the molar ratio of 1:2:20, were placed in an alumina crucible and sealed under vacuum in a fused silica tube. The ampoule was heated to 1170$^{\circ}$ C and held for 12 hours, then cooled to 500$^{\circ}$C at a rate of 5$^{\circ}$C per hour. The excess flux was removed using a centrifuge. The resultant crystals are in the form of hexagonal platelets with an average size of few millimeters (mm) on an edge, as shown in Fig.~\ref{figure 1}c.
One large piece of {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, single-crystal was cut into smaller pieces with appropriate dimensions ($\approx0.1\times0.1\times0.1 {\rm\, mm}^3$) for single crystal x-ray diffraction measurements. A tiny piece of crystal was mounted on a glass fiber and measured on a STOE IPDS 2T. Data collection, integration and absorption correction were done by the x-area software package~\citep{stoe2005x}. The structure of {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, was solved and further refined based on the full matrix least-squares using the SHELXTL program package~\citep{sheldrick2015crystal}. An empirical absorption correction was applied to the measured data. The refinement results including the lattice parameters and atomic positions are consistent with the previous reports on a polycrystalline sample~\citep{grin1989phases}. Several pieces of crystals were pulverized and powder XRD was performed on a PANAlytial X'Pert Pro diffractometer. Magnetization measurements were carried out in a Quantum Design SQUID magnetometer using the DC magnetization method. For susceptibility, both zero-field-cooled (ZFC) and field-cooled (FC) data were measured. Isothermal magnetization data were measured by first cooling the sample in zero field to 1.8 K and then applying field to $\pm$ 7 T.
Transport measurements were performed in a Quantum Design Physical Property Measurement Systems (PPMS). A piece of crystal was polished to a rectangular shape with dimensions 1$\times$0.6$\times$0.1 mm$^3$ and then gold wire (25 micron) contacts were placed using Epotek H20E Epoxy. A four-probe contact method was used for the AC resistivity measurement with an excitation current of 3-5 mA at a frequency of 57.9 Hz. Heat capacity measurement was performed in a Quantum Design Dynacool PPMS. A time relaxation method was employed and the data were measured on heating from 1.8 K to 200 K under zero applied field.
The angular dependence of the metamagnetic transition was measured in a 65 T magnet at the Pulsed-Field Facility, National High Magnetic Field Laboratory, Los Alamos~\citep{singleton2004national}. The metamagnetic transitions were measured in two rectangular samples using the Proximity-Detector Oscillator (PDO) method. Samples were rotated from field parallel to crystallographic ${\bf a}$-axis to ${\bf c} \& {\bf b}$ axes. A more in depth discussion about the PDO technique can be found in Refs.~\citep{PDO.1,PDO.2}.
Neutron powder diffraction (NPD) measurements were
performed on the time-of-flight powder diffractometer,
POWGEN, located at the Spallation Neutron Source at Oak
Ridge National Laboratory. The data were collected with neutrons
of central wavelengths 1.5 and 2.665 \AA, covering the $Q$ spacing
range 0.48$-$12.95 and 0.3$-$5.87 \AA$^{-1}$, respectively. Several high-quality crystals were pulverized to obtain around 0.55 g powder that was loaded in a special annular vanadium container to reduce the
absorption effect from Tb and Ir and gain more diffraction intensity.
A Powgen Automatic Changer (PAC) was used to cover the temperature region of
10$-$300 K. We collected the data at 10, 50 and 300 K. All of the neutron diffraction data were analyzed
using the Rietveld refinement program suite FULLPROF~\citep{Carva1993}.
XMCD data were collected at beamline 4-ID-D of the Advanced Photon Source (APS). Two single crystals with surface normals along the crystallographic ${\bf a} $- and ${\bf b} $- directions were polished to $\approx$20 microns thickness for transmission experiments at Tb L$_3$ and Ir L$_{2,3}$ absorption edges. Crystals were mounted in a variable temperature insert of a cryogenic superconducting magnet and cooled to 1.5 K in helium vapor. Magnetic field was applied parallel to ${\bf a} $- or ${\bf b} $-axis. XMCD data were collected in helicity switching mode (fixed magnetic field direction) whereby the helicity of circularly polarized x-rays produced with phase retarding optics is modulated at 13.1 Hz, and the corresponding modulation in x-ray absorption coefficient is detected with a phase lock-in amplifier~\citep{haskel2007instrument}. XMCD measurements were done with applied magnetic field both along and opposite the wave vector of the incident x-ray beam to check for experimental artifacts of non-magnetic origin.
\section{\label{sec:level3}Results \& discussion}
\subsection*{Crystal Structure}
The crystal structure of Tb$_2$Ir$_3$Ga$_9$ is orthorhombic, crystallizing in space group \textit{Cmcm} (no. 63), and is isotypic with Y$_2$Co$_3$Ga$_9$~\citep{Y2Co3Ga9}. As shown in Fig.~\ref{figure 1}, the structure can be viewed as an alternation of two layers along the [0 0 1] direction. Layer A is strongly puckered with a hexagonal arrangement and consists of Ir and Ga in a 1:2 ratio. Layer B lies in a mirror plane with the Tb atoms arranged in a slightly distorted honeycomb structure. The Tb-Ga coordination in a unit cell is Tb$_4$Ga$_6$. Together these layers form the crystal in a stacking sequence of (AB)$_2$. The ratio of lattice parameters a/b = 1.725, close to $\sqrt{3}$, reflecting a small deviation from hexagonal symmetry. This is similar to other R$_2$T$_3$Ga$_9$ compounds with Y$_2$Co$_3$Ga$_9$ structure type~\citep{EuIrGa_structure, R2T3Al9_RhIrPd,DyCoAl}.
\begin{table*}[ht]\caption{Crystal data and structure refinement for Tb$_2$Ir$_3$Ga$_9$ at 293 K.}
\centering
\small\addtolength{\tabcolsep}{-4pt}
\begin{tabular}{cc}
\hline
Empirical Formula & Ga$_9$Ir$_3$Tb$_2$ \\
Formula Weight & 1521.92 g/mol\\
Wavelength & 0.71073 \AA \\
Crystal System and Space Group& Orthorhombic, $Cmcm$ \\
Unit Cell Dimensions& a = 12.9860(5) \AA \\
& b = 7.5325(9) \AA \\
& c = 9.4349(9) \AA \\
& $\alpha$ = $\beta$ = $\gamma$ = 90$^\circ$ \\
Volume & 922.89(2) \AA$^3$\\
Density (calculated) & 10.953 mg/m$^3$\\
Linear Absorption Coefficient& 83.867 mm$^{-1}$\\
F(000) & 2560 electrons\\
Crystal Size & 0.11$\times$0.1$\times$0.95 mm$^3$\\
$\theta$ Range & 3.126 to 31.843.\\
Index Range & -19$<$=h$<$=18, -11$<$=k$<$=11, -13$<$=l$<$=13\\
No. of Reflections & 5326\\
Independent Reflections & 857[R$_{int}$ = 0.1140]\\
Absorption Correction & empirical\\
Max. and min. Transmission & 0.0786 and 0.0176\\
Refinement Method & Full-matrix least-squares on F$^2$\\
Data / Restraints / Parameters & 857 / 0 / 42\\
Goodness of Fit& 1.151\\
R$_{final}$ indices & R$_1$ = 0.0385, wR$_2$ = 0.1006\\
R indices (all data)& R$_1$ = 0.0431, wR$_2$ = 0.1038\\
Extinction Coefficient& 0.00066(7)\\
Largest diff. Peak and Hole & 4.766 and -4.565 e.\AA$^{-3}$\\
\hline
\end{tabular}
\label{tab: single crystal xrd}
\end{table*}
The results of the single-crystal x-ray diffraction are given in Table~\ref{tab: single crystal xrd}. A crystal with a shape close to a cube (0.11$\times$ 0.1$\times$ 0.095 mm$^3$) gives the lowest R$_{int}$ (11\%) and reasonable thermal displacements during the refinement. The large residual electron peaks and holes close to the heavy elements Ir and Tb atoms are likely due to an inadequate absorption correction. Detailed descriptions of the atomic coordinates, refinement parameters and bond lengths of the atoms are given in the supplemental materials (SM)~\citep{supp}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{fig1.pdf}
\caption{Crystal structure of Tb$_2$Ir$_3$Ga$_9$.(a) Schematic of crystal structure projected along \textit{ac} plane. Two different layers are identified that make up the structure in stacking sequence (AB)$_2$. (b) Layer A projected as described in the image. The Ir-Ga coordination is Ir$_6$Ga$_{12}$.(c) As grown crystal of Tb$_2$Ir$_3$Ga$_9$ on a 1 mm grid. (d) Layer B projected as indicated in the figure. The honeycomb arrangement of Tb atoms is evident.}\label{figure 1}
\end{figure}
\subsection*{Magnetism \lowercase{and} Transport}
\begin{figure}[]
\centering
\includegraphics[width=0.4\textwidth]{fig2MT}
\caption{DC Magnetization of Tb$_2$Ir$_3$Ga$_9$.(a) \textit{M(T)} with field applied along the ${\bf a}- $ and ${\bf b}- $ axes (Inset: \textit{M(T)} along ${\bf c} $-axis). The maroon solid line represents the FC data. (b) ZFC inverse susceptibility in an applied field of 0.1 T for three distinct crystallographic axes and the polycrystalline average as described in text. Green solid line represents a fit to the Curie-Weiss form as described in the text.}\label{figure 2 MT}
\end{figure}
The temperature-and magnetic-field dependent magnetization measurements, \textit{M(T,H)}, are shown in Figs.~\ref{figure 2 MT} \&~\ref{figure 3 MH}, respectively. Magnetism in Tb$_2$Ir$_3$Ga$_9$ is highly anisotropic, as is evident from the DC magnetic susceptibility shown in Fig.~\ref{figure 2 MT}a. The susceptibility along the ${\bf a} $-axis shows antiferromagnetic ordering at 12.5 K with ZFC and FC curves showing no irreversibility up to 3 K. In contrast, the susceptibility for field parallel to the ${\bf b} $-axis shows ferromagnetism with a discontinuity around 13 K. The FC curve shows large irreversibility below the transition.
The susceptibility along the ${\bf c} $-axis (Inset of Fig.~\ref{figure 2 MT}a) is characterized by a broad maximum around 65 K followed by two discontinuities at 12.5 K and 2.5 K. The broad maximum is likely attributable to higher-lying CEF states being populated by thermal excitation. While the 12.5 K transition probably marks long-range order, we cannot rule out the possibility that this feature arises from small misalignment of the crystal with respect to the field. Some Ce and Yb based compounds with similar crystal structure show broad peak in temperature ranges 150-250 K, and those are well understood by mixed valence states along with CEF effects~\citep{Ce2T3X9.transport,CEF.Yb2Ir3Ga9,Yb2RhIR3AlGa9}. The weak feature at 2.5 K is currently not understood. The ZFC and FC data only show irreversibility at the onset of the transition. For $T<T_{\rm N}$, susceptibility along ${\bf b} $ is largest followed by ${\bf a} $ and ${\bf c} $, while for $T>T_{\rm N}$, the order is $\chi_a>\chi_b>\chi_c$.
Figure~\ref{figure 2 MT}b shows inverse susceptibility along all three axes and the average susceptibility, $\chi_{\rm avg} = \frac{(\chi_a +3 \chi_c)}{3}$, with a fit to the Curie-Weiss form, $1/\chi$ = $\frac{T-\theta_{\rm W}}{C}$, shown as a solid line. The Weiss temperature, $\theta_{\rm W}$, and effective moment, $\mu_{\rm eff}$, estimated from the Curie constant,\textit{C}, provide insights on the interactions between magnetic atoms. Different values of Weiss temperature reflect the strong anisotropy in magnetization and are similar to other compounds of the R$_2$T$_3$X$_9$ family~\citep{DyCoAl}. The value of $\theta_{\rm W}$ is largest for the field along ${\bf c} $ with a value of -148$\pm$2 K followed by -5.8$\pm$0.3 K along ${\bf b} $ and -0.5$\pm$0.1 K along ${\bf a} $ axis. In all three directions, $\theta_{\rm w}<0$, indicating antiferromagnetic interactions of varying strengths. The average effective moment estimated from the fit to $\chi_{\rm avg}$, is 10.3 $\mu_{\rm B}$/Tb, close to that expected effective moment for a free Tb$^{3+}$ ion $\approx$9.7 $\mu_{\rm B}$, while the average Weiss temperature is -70 K.
\begin{figure}[]
\centering
\includegraphics[width=0.4\textwidth]{fig3MH}
\caption{Magnetization at 1.8 K and 30 K.(a) Magnetization \textit{M(H)} with field applied along three crystallographic axes at 1.8 K. The \textit{M(H)} along ${\bf a} $ axis shows ultra-sharp metamagnetic transitions with each jump corresponding to a value of M$_{\rm s}$/2. Upper left inset: \textit{M(H)} along ${\bf c} $-axis. Lower right inset: Angle dependence of critical field, $B_{\rm c}$, of metamagnetic transitions. The angle $\theta$ is measured from ${\bf a}$ toward ${\bf b}$ axis. Red solid lines are fit to the form, $B_{\rm c}/cos\theta$. (b) \textit{M(H)} at 30 K.}\label{figure 3 MH}
\end{figure}
The anisotropic magnetism in Tb$_2$Ir$_3$Ga$_9$ is also evident in \textit{M(H)} data, shown in Fig.~\ref{figure 3 MH}a for 1.8 K, and in Fig.~\ref{figure 3 MH}b for 30 K. When the field is applied along the ${\bf a} $-axis, two field-induced sharp metamagnetic transitions are seen at fields $B_{\rm c1}$ and $B_{\rm c2}$, with each transition corresponding to a jump of M$_{\rm{s}}$/2 $\approx$8 $\mu_{\rm B}$, where M$_{\rm s}$ is the saturated magnetic moment at 1.8 K and 7 T. We note that the expected M$_{\rm s}$ for a free Tb$^{3+}$ ion, $^7$F$_6$, with $g=1.5$ is 9 $\mu_{\rm B}$/Tb, thus the jumps are very close to half of that expected for local Tb$^{3+}$ moment. At 7 T and 1.8 K, the $M$=16.3 $\mu_{\rm B}$/formula unit (f.u.). The angle dependence of the metamagnetic critical field, $B_{\rm c}$, shown in the right side inset of Fig.~\ref{figure 3 MH}(a), implies that only the ${\bf a}$-axis projection of field is important. Both transitions move to higher fields with increasing angle. Beyond 75 degrees, the transitions have moved beyond the accessible field range. The solid lines are the fit to the form $B_{\rm c}$/cos$\theta$, demonstrating that both metamagnetic transitions depend only on the component of the field along the ${\bf a} $-axis.
Steplike metamagnetic transitions have been found in certain phase-separated perovskite manganites~\citep{metamag.manganite.1,metamag.manganite.2}, and intermetallic compounds such as Nd$_5$Ge$_3$\citep{Nd5Ge3.1,Nd5Ge3.2}, Gd$_5$Ge$_4$~\citep{Gd5Ge4.1} doped CeFe$_2$~\citep{CeFe2}, and LaFe$_{12}$B$_6$~\citep{LaFe12B6}. In these compounds transitions are driven by field-induced response of the phase-separated state, where the applied field favors the ferromagnetic phase over the AFM phase. The metamagnetic transition in these inhomogeneous systems is first-order and is accompanied by a large hysteresis with remanance. In contrast, {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, is a homogeneous AFM system similar to TbNi$_2$Ge$_2$~\citep{TbNi2Ge2}, TbCo$_2$Si$_2$,\citep{DyCo2Si2.1,rare.earth.metamagnetism}, TbCoGa$_5$~\citep{TbCoGa5}, TbCo$_2$Ge$_2$~\citep{TbCo2Si2}, for example. The magnetic behavior follows from the expected Ising character of Tb$^{3+}$ found in the low site symmetry, $C_s$~\citep{XMCD.TbMn03,CEF.TbAl03}. Tb$^{3+}$ is a non-Kramers’ ion, and the ground state manifold $^7F_6$ is split into 2$J$ + 1 = 13 singlets by CEF in this $C_s$ symmetry. The lowest crystal-field level is expected to be a quasidoublet that dictates the low temperature properties of the host compound, including a magnetic moment of 9 $\mu_{\rm B}$~\citep{XMCD.TbMn03,CEF.TbAl03,zvezdinmodern} and the Ising behavior. The observed metamagnetic transitions of {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, are consistent with this scenario with the Ising axis being the crystallographic ${\bf a} $-axis.
While the \textit{M(H)} along ${\bf c} $ axis is typical of an antiferromagnet, the ${\bf b} $ axis magnetization reveals a small hysteresis loop with coercivity of $\approx$ 0.5 T indicating ferromagnetism in accordance with \textit{M(T)} data and a breakdown of the purely Ising approximation. We show in section~\ref{model} below that a Dzyaloshinskii- Moriya (DM) type interaction between Tb$^{3+}$ gives rise to this FM component. The M$_{\rm s}$ at 7 T and 1.8 K along ${\bf b} $-axis is 7.5 $\mu_{\rm B}$/f.u., and the remanant moment is 1.2 $\mu_{\rm B}$/f.u.
The temperature dependent resistivity, $\rho$, with current applied along three different crystallographic axes is shown in Fig.~\ref{figure 4 resistivity}a. The resistivity is metallic, anisotropic and shows a discontinuity at the magnetic transition. Transport anisotropy, $\rho_{\bf c}/\rho_{\bf a}$ $ \approx$2.5 at 290$\sim$K. The residual resistivity ratio, $\rho_{\rm 290\,K}/\rho_{\rm 2\,K}$ , is about 6 for all axes.
\subsection*{Heat Capacity}
\begin{figure}[]
\centering
\includegraphics[width=0.5\textwidth]{fig4CpRes}
\caption{(a) Resistivity with current applied along three crystallographic axes. The anisotropy between ${\bf a} $ and ${\bf c} $ axis resistivity at 290 K is close to 2.5. (b) Specific heat, \textit{C}, of Tb$_2$Ir$_3$Ga$_9$ and its non-mangnetic counterpart Y$_2$Ir$_3$Ga$_9$ (blue line) versus temperature, \textit{T}. (c) Magnetic specific heat, \textit{C$_{\rm mag}$}, plotted against temperature.(d) Magnetic entropy, \textit{S$_{\rm mag}$}, per Tb versus temperature. \textit{S$_{\rm mag}$} was calculated from the magnetic specific heat as described in the text. Inset: low temperature region of the entropy. The horizontal dashed line (olive) is R$ln$2, which coincides with the vertical dashed line (red) marking the T$_{\rm N}$ at the inflection point.}\label{figure 4 resistivity}
\end{figure}
The temperature dependence of the heat capacity, \textit{C$_p$}, of Tb$_2$Ir$_3$Ga$_9$ is shown in Fig.~\ref{figure 4 resistivity}b. The $\lambda$-like anomaly in the heat capacity at 12.5 K denotes the bulk, long-range magnetic ordering in Tb$_2$Ir$_3$Ga$_9$. To investigate the magnetic contribution to the overall heat capacity, a non-magnetic, isostructural compound Y$_2$Ir$_3$Ga$_9$ was synthesized in single-crystal form following a similar procedure to that described above for Tb$_2$Ir$_3$Ga$_9$. The heat capacity of Y$_2$Ir$_3$Ga$_9$ does not show any magnetic ordering and hence can be taken as a basis for the lattice component of the heat capacity of Tb$_2$Ir$_3$Ga$_9$. The $T$-axis of the Y$_2$Ir$_3$Ga$_9$ specific heat data was scaled following a method developed by Bouvier et al~\citep{Bouvier.Cp.normalization} and later followed by others~\citep{Cp.Normalize2} (note: $\theta_D$ is 276 K from the single $\beta$ fit). Here the correction factor is calculated via Eq.~\ref{heat cap norm}
\footnotesize
\begin{equation}\label{heat cap norm}
\frac{\Theta_D (Tb_2Ir_3Ga_9)}{\Theta_D (Y_2Ir_3Ga_9)}=\left[\frac{2(M_Y)^\frac{2}{3}+3(M_{Ir})^\frac{2}{3}+9(M_{Ga})^\frac{2}{3}}{2(M_{Tb})^\frac{2}{3}+3(M_{Ir})^\frac{2}{3}+9(M_{Ga})^\frac{2}{3}}\right]^\frac{1 }{3},
\end{equation}
\normalsize
where \textit{M$_x$} ($x$ = Y, Ga, Ir, Tb) is the atomic mass of each of the constituent atoms. In this way, a correction factor of 0.940 was calculated. The resulting data (Fig.~\ref{figure 4 resistivity}(b)) was then used to calculate the magnetic contribution to heat capacity, $C_{\rm mag}$, for Tb$_2$Ir$_3$Ga$_9$, where $C_{\rm mag} = C_{Tb}-C_{Y}$, which is shown in Fig.~\ref{figure 4 resistivity}c. Here, in addition to the $\lambda$-like transition, a broad feature is also evident. The magnetic contribution to the entropy is estimated by integrating, $\int{\frac{C_{\rm mag}}{T}}\, dT$. The resulting entropy is shown in Fig.~\ref{figure 4 resistivity}d. The maximum entropy we find is 17 J mol$^{-1}$Tb$^{-1}$K$^{-1}$, significantly smaller than the R\textit{ln}13 = 21.3 J mol$^{-1}$Tb$^{-1}$K$^{-1}$ expected for the $^7$F$_6$ ground state.
At the onset of antiferromagnetic order, we find $S = 5.75~{\rm Jmol^{-1}K^{-1}} = R \ln 2$ (Inset of Fig.~\ref{figure 4 resistivity}d). This can be understood via the splitting of the $^7$F$_6$ states of the Tb$^{3+}$ free ion by CEF into a ground state quasidoublet separated by a large ($>>$12 K) gap from the first excited state. While the CEF energy spectrum for {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, is not known, Tb compounds with similar site symmetry for the Tb$^{3+}$ ion, for example in TbAlO$_3$, place the ground state quasidoublet about 160 meV below the first excited state~\citep{CEF.TbAl03}. Similar behavior has been reported in TbNi$_2$Ge$_2$~\citep{Rln2}, where the Ising axis is the tetragonal ${\bf c}$-axis. Here, a broad ``Schottky''-like feature was found in the magnetic heat capacity and attributed to thermal population of one or more CEF levels above the quasidoublet. A similar explanation likely applies to {\rm Tb$_2$Ir$_3$Ga$_9$}\,\,. However, attempts to model the data above $T_{\rm N}$ to a Schottky form for a two-level system~\citep{schottky} lead to poor quality fits, probably reflecting the presence of groups of levels not adequately captured by a simple two-level expression. A better understanding of the CEF levels in {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, will be needed to model these data properly. The missing entropy in {\rm Tb$_2$Ir$_3$Ga$_9$}\,\,, TbNi$_2$Ge$_2$ and other Tb$^{3+}$ containing systems~\citep{Cp.Normalize2,TbFe2.CEF,kumar2008magnetism} likely signals the existence of additional CEF levels at energies higher than that probed here.
\subsection*{Neutron Powder Diffraction}
\begin{figure}
\centering \includegraphics[width=1.0\linewidth]{figNeutron3.pdf} \caption{
Rietveld refinement of neutron powder diffraction patterns of Tb$_{2}$Ir$_{3}$Ga$_{9}$ at (a) 300 K, and (b) 10 K. The observed
data and the fit are indicated by the open circles and solid lines, respectively. The difference curve is shown at the bottom.
The vertical bars mark the positions of Bragg peaks for the nuclear phase (up) and magnetic phase
(down). The dotted rectangle in (b) marks 7 strongest magnetic Bragg peaks. (c-d) Magnetic structure deduced from the refinement to the neutron data, showing the direction of the moments along ${\bf a} $-axis.}
\label{fig:Neutron}
\end{figure}
The neutron powder diffraction pattern at 300 K, shown in Fig.~\ref{fig:Neutron}(a), evidences no secondary phases in the Tb$_{2}$Ir$_{3}$Ga$_{9}$ sample. Rietveld analysis confirms the orthorhombic structure with space group \textit{Cmcm} (No. 63) with the refinement goodness of $\chi^2\approx$3.5, as illustrated in the Fig.~\ref{fig:Neutron}(a). Upon cooling to 10 K ($<T_{\rm N}$), the intensities of more than 10 low-$Q$ peaks (Fig.~\ref{fig:Neutron}(b)) increase significantly, indicative
of a magnetic contribution to the scattering. All magnetic reflections can be indexed on the
nuclear (chemical) unit cell with a magnetic propagation vector
\textbf{k} = (0,0,0). The SARAH representational analysis program~\citep{Wills2000}
was used to derive the symmetry-allowed magnetic structures. The symmetry allowed basis vectors for Tb sites are summarized in Table S1 of the SM~\citep{supp}.
The neutron diffraction pattern is best fit using the $\Gamma$5, irreducible representations i.e., antiferromagnetic order with moment strictly along ${\bf a} $ axis with the refinement goodness of $\chi^2\approx$7.29. Allowing a spin canting toward the ${\bf b} $ axis does not improve the refinement. The ordered moment of Tb is found to be $7.5(2) \mu_{\rm B}$/Tb. Both results are
consistent with the magnetization measurements. The magnetic structure
is displayed in Fig.~\ref{fig:Neutron}. It is worthwhile pointing out that Ir does not carry an ordered moment within the instrumental resolution. \\
\subsection*{XMCD}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{xmcd}
\caption{XMCD and XAS on Tb and Ir L-edges at 1.5 K. (a) XMCD signals of Tb L$_3$-edge at 0 and 5 T field are plotted on the left axis while the XAS (solid black line) are plotted on the right axis for field parallel to crystallographic ${\bf b} $-axis. Inset: XMCD signal at 1 and 5 T with field parallel to ${\bf a} $-axis. (b) XMCD signals of Ir L$_3$ edge at 5 T field are plotted on the left axis while the XAS (solid black line) are plotted on the right axis for field parallel to crystallographic ${\bf b} $-axis. XMCD on Ir L$_3$ edge of $\beta$-Li$_2$IrO$_3$ and Sr$_2$IrO$_4$ are also plotted for comparison purposes. Inset: XMCD signal at 5 T of Ir L$_2$ edge.}
\label{fig: Tb edge}
\end{figure}
To more definitively explore the potential for Ir magnetism when the field is applied along the ${\bf b} $-axis, we have performed
X-ray magnetic circular dichroism (XMCD) measurements at beamline 4-ID-D of the Advanced Photon Source. Data shown in the Fig.~\ref{fig: Tb edge} are averages of data sets collected with opposite applied field directions. While the Tb XMCD data at selected applied field values are consistent with the magnetometry data for both crystal orientations, no detectable XMCD signals were found at Ir L-edges. We can put an upper limit $<$ 0.01 $\mu_B$/Ir to the magnitude of any Ir magnetic moment by scaling to XMCD signals in $\beta$-Li$_2$IrO$_3$~\citep{XMCD1} (0.35 $\mu_B$/Ir) and Sr$_2$IrO$_4$~\citep{XMCD2} (0.05 $\mu_B$/Ir). This indicates that the finite ferromagnetic response along \textit{b}-axis is solely due to Tb moments either by a field induced canting of Tb spins or by a small but finite DM interaction between the nearest neighbors along the \textit{c}-direction, or a combination of the two. We will discuss these possibilities further in section~\ref{model}. The non-magnetic state of Ir is in accordance with our NPD data and the general consensus that the T atom in R$_2$T$_3$X$_9$ is magnetically inactive~\citep{DyCoAl}.
\subsection*{Electronic structure \& Magnetism}
Electronic structure calculations have been carried out within density functional theory (DFT) using the all-electron, full potential code WIEN2K~\citep{blaha2001wien} based on the augmented plane wave plus local orbital (APW + lo) basis set~\citep{sjostedt2000alternative}. The Perdew-Burke-Ernzerhof (PBE) version of the generalized gradient approximation (GGA)~\citep{PhysRevLett.77.3865} was chosen as the exchange correlation potential. Spin-orbit coupling (SOC) was introduced in a second variational procedure~\citep{singh2006planewaves}. The LDA+$U$ scheme improves over GGA or LDA in the study of systems containing correlated electrons by introducing the on-site Coulomb repulsion U applied to localized electrons (e.g., $4f$). We have performed calculations within the LDA+$U$ ensatz (using the fully localized version for the double-counting correction)~\citep{PhysRevB.49.14211} taking a reasonable $U$ value for this $f$-electron system (8 eV) comparable to the values obtained for TbN. A dense \textit{K}-mesh of 16$\times$16$\times$11, was used for the Brillouin zone sampling. An R$_{mt}$K$_{max}$ of 7 was chosen for all calculations. Muffin tin radii were 2.5 \textit{a.u.} for Ir and Tb, 2.7 \textit{a.u.} for Ga.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figDOS}
\caption{Atom-resolved density of states for {\rm Tb$_2$Ir$_3$Ga$_9$}\,\,. Left panel GGA, right panel LDA+U calculations.}
\label{fig: DOS}
\end{figure}
Using the experimental structure, DFT calculations were performed for a FM state, as well as for the collinear N\'eel AFM state proposed by neutrons (where each of the three types of bonds is AFM), and a collinear striped AFM phase. The AFM state proposed by neutrons is more stable than any other magnetic configuration tested by 16 meV/unit cell (u.c.). Once spin-orbit-coupling (SOC) is introduced, the preferred direction of the magnetization is the ${\bf a}$-axis with derived magnetocrystalline anisotropy energies MAE[100]-[010]= 0.35 meV/u.c., and MAE[100]-[001] = 0.75 meV/u.c. The obtained $4f$ spin moment is 5.86 $\mu_{\rm B}$, and the orbital moment 1.34 $\mu_{\rm B}$ (increased upon inclusion of a Coulomb $U$ to 6.07 and 1.40 $\mu_{\rm B}$, respectively). The total magnetic moment is then 7.2, and 7.5 $\mu_{\rm B}$ for GGA and LDA+$U$ calculations, respectively. Both agree with the ordered moment found by NPD and DC magnetization.
In the GGA calculation, as seen in Fig.~\ref{fig: DOS}, the minority spin channel for Tb atoms corresponding to their $f$ states, is pinned at the Fermi level (partially filled). Due to the highly localized character of the $4f$ electrons, it is unlikely that the density of states can have a finite Tb-$4f$ contribution at the Fermi level. In the LDA+U ground-state, there is no $4f$ weight at the Fermi level (all the weight is Ir-$d$ and Ga-$p$). There are instead different peaks of the $4f$-projected density of states well below and well above the Fermi level (shifted by the inclusion of a Coulomb $U$).
\section{Phenomenological Model of magnetism \lowercase{in} T\lowercase{b}$_2$I\lowercase{r}$_3$G\lowercase{a}$_9$}\label{model}
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{honey3}
\end{center}
\caption{Honeycomb layer of {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, with first-neighbor interactions $J_{1}^{(1)}$ and $J_{1}^{(2)}$ and second-neighbor interaction $J_2$.}\label{honey}
\end{figure}
A theoretical understanding of the magnetism in {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, confronts two constraints posed by the measurements discussed above. The first constraint posed is the absence of any phase transition when the field is applied along ${\bf b} $ and ${\bf c}$-axes, as confirmed by the measurements of the critical fields in applied fields of up to 60 T. Another constraint is provided by the observed scaling of the critical fields
when the external magnetic field is rotated by angle
$\theta $ away from the ${\bf a} $ axis in the ${\bf a} -{\bf b} $ plane. As shown in Fig.~\ref{figure 3 MH}, measurements find that $B_{\rm c1}(\theta )\cos \theta $
and $B_{c2}(\theta )\cos \theta $ are roughly independent of $\theta $ up to $0.35 \pi $.
This implies that only the component of the field
along ${\bf a} $ controls those phase transitions. Taken together, these two considerations suggest that the Tb moments can be approximately described as Ising spins aligned along $\pm {\bf a} $. However, a simple Ising model cannot faithfully capture all of the features discussed above, and we now build a phenomenological model in accordance with these considerations. We take the four Tb ions in the unit cell to have ``spins"
\begin{equation}
{\bf S}_i = S(\sin \theta_i \cos \phi_i ,\sin \theta_i \sin \phi_i, \cos \theta_i)
\end{equation}
with $S=6$, and ($\theta$,$\phi$)$_i$ are spherical polar coordinates. Elastic neutron measurements indicate that the zero-field state of the Tb ions has $\theta_i = \pi/2$ with
$\phi_1 \approx \phi_4 \approx 0$ and $\phi_2 \approx \phi_3 \approx \pi $.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figModel2}
\end{center}
\caption{Predicted and measured magnetization with field along (a) ${\bf a} $, and (b) ${\bf b} $ axes. Inset: the predicted variation of $B_{\rm c} \cos \theta $ where $\theta $ is the angle of the field in the $xy$ plane. Dashed lines are the experimental data.}
\label{unitcell}
\end{figure}
The Hamiltonian is given by
\begin{eqnarray}
{\cal H}&=&-\frac{1}{2}\sum_{i, j}^{1st} J_{1, ij}\, {\bf S}_i \cdot {\bf S}_j -\frac{1}{2}\sum_{i, j}^{1st} J_{1, ij}^n \, {\bf S}_i \cdot {\bf n}_{ij} \, {\bf S}_j \cdot {\bf n}_{ij} \nonumber \\
&-& \frac{1}{2}J_2 \sum_{i,j}^{2nd} {\bf S}_i \cdot {\bf S}_j
- K_2 \sum_i {S_{ia}}^2 -K_p \sum_i {S_{ic}}^2
\nonumber \\ &-&
\frac{1}{2} K_h \sum_i \Bigr\{ (S_{ia} +iS_{ib})^6 +(S_{ia}-iS_{ib})^6\Bigl\} \nonumber \\
&-&\frac{1}{2}\sum_{i, j}^{1st} {\bf D}_{ij} \cdot ({\bf S}_i \times {\bf S}_j )
- g\mu_{\rm B} {\bf B} \cdot \sum_i {\bf S}_i,
\label{ham}
\end{eqnarray}
where the field is along the $\alpha =a$, $b$, or $c$ direction and
the exchange interactions are indicated in Fig.~\ref{honey}. The factors of $1/2$ avoid double counting.
First neighbors are coupled by both isotropic $J_{1, ij}$ and directional $J_{1, ij}^n$ exchange couplings.
The latter couples spins along the direction of the bond so that
\begin{equation}
{\bf n}_{ij} = \frac{{\bf R}_j - {\bf R}_i}{\vert {\bf R}_j - {\bf R}_i\vert }
\end{equation}
is a unit vector from site $i$ to site $j$.
To account for the orthorhombic distortion of the honeycomb lattice, we further break
$J_{1, ij}$ and $J_{1, ij}^n$ into two parts: $J_1^{(1)}$ and $J_1^{n(1)}$ acts between sites 1 and 2 or 3 and 4 along the $a$ axis
while $J_1^{(2)}$ and $J_1^{n(2)}$ act between sites 1 and 3 or 2 and 4 at 60$^\circ $ from the $a$ axis.
Second neighbors 1 and 4 or 2 and 3 are coupled by $J_2$.
The easy-axis anisotropy $K_2$ aligns the spins along the $a$ axis due to the orthorhombic distortion of the lattice and the easy-plane anisotropy $K_p < 0$ keeps the spins in the $a-b$ plane. The hexagonal anisotropy $K_h > 0$ favors the spins ${\bf S}_i $ to lie along the three pseudo-hexagonal axis at $\phi_i =0$ and $\pm \pi/3$. In terms of the spin angles, this energy can be writtten $-S^6 K_h \cos 6\phi_i $ at each site.
The DM interactions ${\bf D}_{ij}$ along ${\bf c} $ act only between first neighbors (with opposite signs between sites 1 and 2 and between 3 and 1 or 2 and 4) and produces the zero-field canted moment along ${\bf b} $. This DM interaction is allowed by the alternating positions of the Ir ions around each Tb-Tb bond.
This complex model is required to ``tame" the magnetization so that no phase transition occurs when the field is applied along ${\bf b} $.
A simpler model that neglects the directional exchange has a $\chi^2$ about 16 times greater.
The best fits for this model are shown in Fig.~\ref{unitcell}. Notice that this model describes the experimental measurements in all three phases at 1.8 K. The ${\bf c}$-axis behavior, not shown, is a good fit to the data shown in Fig.~\ref{figure 3 MH}b. The fits for the model were constrained by the requirement that
$B_{cn}(\theta )\cos \theta $ ($n=1$ or 2) are approximately constant as a function of $\theta $, as measured experimentally. The scaled critical fields from the model are plotted as a function of $\theta $ in the inset of Fig.~\ref{unitcell}. The scaled fields deviate from their $\theta =0$ values only above about $0.3\pi $. We have neglected the presumably weak coupling between planes. While it is not known how the neighboring planes are magnetically configured, if they respond identically to the applied fields, then their exchange coupling will not change with field.
The fitted values of the parameters for the model are given in Table~\ref{obmodes}. The largest parameter is the easy-plane anisotropy $K_p \approx -0.88$ meV, which keeps the Tb spins in the $a-b$ plane. The weak hexagonal anisotropy $K_h$ favors the spins to lie along the three hexagonal axis.
Notice that all the exchange parameters are AFM. While the first-neighbor exchange between spins 1 and 2 or 3 and 4 contains both isotropic $J_1^{(1)}$ and
directional $J_1^{n(1)}$ contributions, the exchange between spins 1 and 3 or 2 and 4 is primarily directional with $\vert J_1^{(2)}\vert \ll \vert J_1^{n (2)}\vert $. The next nearest neighbor interaction $J_2 <$ 0 stabilizes the intermediate metamagnetic phase. With $J_2$=0, B$_{\rm c2}$ = B$_{\rm c1}$ and the intermediate phase would be absent. We note that the extracted parameters are consistent with the observed transition temperature with $J \approx$ 0.025 meV and the mean-field transition temperature ($z$/3)$J$ $S$($S$+1) = 1 meV or 11.6 K.
\begin{table}
\caption{Parameters in meV}
\begin{tabular}{|cc|cc|cc|cc|}
\hline
parameter && value &\\
\hline
$J_1^{(1)}$ && -0.016 &\\
$J_1^{(2)}$ && -0.0004 &\\
$J_1^{n (1)}$ && -0.024 &\\
$J_1^{n (2)}$ && -0.063 &\\
$J_2$ && -0.0081 &\\
\hline
$K_2$ && 0.050 &\\
$K_p$ && -0.88 &\\
$K_h$ && $1.4 \times 10^{-7}$ &\\
\hline
$D$ && 0.0065 &\\
\hline
\end{tabular}
\label{obmodes}
\end{table}
\section{Conclusion}
We have synthesized single crystals of the honeycomb lattice antiferromagnet {\rm Tb$_2$Ir$_3$Ga$_9$}\,\,. The observed magnetism is highly anisotropic with an AFM transition at T$_N\approx$12.5 K. Two step-like metamagnetic transitions were found when the magnetic field was applied along the magnetic easy ${\bf a}$-axis, reflecting the Ising nature of the Tb$^{3+}$ quasidoublet. Neutron powder diffraction revealed the direction of the magnetic moment along the ${\bf a}$-axis, in accordance with the magnetization data. A broad peak found for the ${\bf c}$-axis susceptibility is attributed due to CEF effects, as is a similar broad maximum in the magnetic specific heat above the N\'eel transition. A phenomenological model was proposed that describes all of the magnetic data well, including the angle-dependence of the metamagnetic transitions. A small but finite DM interaction between nearest neighbors in Tb-Tb planes, which acts along ${\bf c}$-axis, was found to be essential in describing the observed scaling behavior of the metamagnetic transitions. Notably, the inclusion of a bond directional anisotropy to the magnetic exchange is essential to proper modeling of the data, highlighting that {\rm Tb$_2$Ir$_3$Ga$_9$}\,\, joins the family of honeycomb magnets with such anisotropic exchange. Beyond this, the Ising behavior of the Tb moments and the honeycomb lattice arrangements of the Tb atoms make this compound a fertile ground to investigate the interplay among various magnetic interactions and crystal field effects. As such, field dependent neutron scattering experiments on single crystals are a logical next step toward understanding this honeycomb lattice system.
\acknowledgments
This work was sponsored by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division. A portion of this research used resources at Spallation Neutron Source, a DOE Office of Science User Facility operated by the Oak Ridge National Laboratory. Work performed at the National High Magnetic Field Laboratory, USA, was supported by NSF Cooperative Agreements DMR-1157490 and DMR-1644779, the State of Florida, U.S. DoE, and through the DoE Basic Energy Science Field Work Project Science in 100 T. Work at the APS was supported by the U.S. Department of Energy, Office of Science, under Contract No. DE-AC02-06CH11357. A.S.B. thanks ASU for startup funds. The authors would like to thank Dr. David Parker, ORNL, for useful discussions.
|
\section{Introduction}
\addcontentsline{toc}{section}{Introduction}
The distribution of the sum of two independent \(\chi^{2}\) distributed random variables with \( m_{1} \) and \( m_{2} \) degrees of freedom is known to be \(\chi^{2}\) with \( m_{1}+m_{2} \) degrees of freedom.
However, the case of non-independent variables is less straight forward. \cite{Gunst1973} derived the distribution of a sum of two linearly correlated \(\chi^{2}\) random variables with \( \rho <1 \). More recently, \cite{Joarder2013} expressed the density for a sum of correlated \(\chi^{2}\) with common \( m \) degrees of freedom in terms of a generalized hypergeometric function.
Here it is shown that the distribution of the sum of linearly correlated \(\chi^{2}\) variables with product moment correlation coefficient \( 0 \leq \rho \leq 1 \) can be approximated by a conveniently parametrized gamma distribution. By a similar method, the distribution of the difference can be derived as well in particular cases.
Both results have straight-forward generalizations to sums of gamma variables with common scale and to differences of identically distributed gamma variables.
\section{Distribution of sum of correlated \(\chi^{2}\) variables}
Let
\[ X_{1} \sim \chi ^{2} \left( m_{1} \right) , \]
\[ X_{2} \sim \chi ^{2} \left( m_{2} \right) , \]
\[ X=X_{1}+X_{2} .\]
We will first consider the case of \( m = m_{1} = m_{2} \) . In the independent case, with shape and scale parametrization.
\[ X \sim \Gamma \left( m, 2 \right). \]
Let us consider the case of the distribution of \( X_{1} \) multiplied by a constant. In this case, a well-known result is that
\[ cX_{1} \sim \Gamma \left( \frac{m}{2},2c \right), \]
When \( c=2 \)
\[ 2X_{1} \sim \Gamma \left( \frac{m}{2},4 \right). \]
When \( X_{1}=X_{2} \),\ \( X=2X_{1} \) . Whereas previous derivations of the distribution of linear combinations of correlated \(\chi^{2}\) variable focus on \( \rho <1 \), here a different approach is used, akin to the method used by \cite{kotz1963distribution} to derive an approximate distribution for the sum of correlated gamma variables. These two results can be seen as special cases of the distribution of the sum of two positively correlated \(\chi^{2}\) variables where \( \rho \left( X_{1},X_{2} \right) =0 \) and \( \rho \left( X_{1},X_{2} \right) =1 \), respectively.
Under this perspective, it can be inferred that the general distribution of X in the correlated case reduces to a gamma at the extremes, and therefore should be gamma or generalized gamma for different values of \( 0 \leq \rho \leq 1 \). We can now deduce some other properties of this distribution based on considerations on expected value and variance.
Clearly:\par
\[ E \left( X \right) =E \left( X_{1} \right) +E \left( X_{2} \right) =2m .\] \par
The expected value should not depend on correlation. Also, it follows from basic properties of variance that, in the correlated case: \par
\[ Var \left( X \right) =Var \left( X_{1} \right) +Var \left( X_{2} \right) +2Cov \left( X_{1},X_{2} \right) ;\]
since \( Var \left( X_{1} \right) =Var \left( X_{2} \right) =2m ,\) \par
\[ Cov \left( X_{1},X_{2} \right) = \rho \; \sqrt[]{Var \left( X_{1} \right) Var \left( X_{2} \right) }=2 \rho m ,\] \par
\[ Var \left( X \right) =4m \left( 1+ \rho \right) .\] \par
The distribution of X should therefore be gamma or generalized gamma with expected value equal to the sum of the expected values, i.e. \( 2m \), and variance equal to the one of the individual \(\chi^{2}\) scaled by a \( 2 \left( 1+ \rho \right) \) factor. There is no straight-forward way to parametrize the generalized gamma distribution so that these conditions are met, since it only reduces to a gamma when the second scale parameter is 1. However, it is very easy to achieve the result by choosing appropriate values for a simple two parameters gamma distribution. Since for the gamma distribution \(\mu=k\theta\) and \(\sigma^{2} = k\theta^{2}\), the distribution for the sum of two correlated \(\chi^{2}\) variables with common \( m \) should be approximately:\par
\[ X \sim \Gamma \left( \frac{m}{1+ \rho },2 \left( 1+ \rho \right) \right) .\] \par
The result is easily generalized to the case where \( m_{1} \neq m_{2} \) . In this case: \par
\[ E \left( X \right) =E \left( X_{1} \right) +E \left( X_{2} \right) =m_{1}+m_{2} ,\] \par
\[ Cov \left( X_{1},X_{2} \right) =2 \rho \; \sqrt[]{m_{1}m_{2}} ,\] \par
\[ Var \left( X \right) =2 \left( m_{1}+m_{2}+2 \rho \; \sqrt[]{m_{1}m_{2}} \right) ,\] \par
\[ X \sim \Gamma \left( \frac{ \left( m_{1}+m_{2} \right) }{2 \left( 1+\frac{2 \rho \; \sqrt[]{m_{1}m_{2}}}{m_{1}+m_{2}} \right) },2 \left( 1+\frac{2 \rho \; \sqrt[]{m_{1}m_{2}}}{m_{1}+m_{2}} \right) \right) .\] \par
It is easily verified that the conditions concerning mean and variance are satisfied. \par
In the case of the sum of \(N \) variables with pairwise correlation coefficients \(\rho_{i,j}: i,j= 1, ..., N \),
\[ \sum _{i=1}^{N}X_{i} \sim \Gamma \left( \frac{ \sum _{i=1}^{N}m_{i}}{u},u \right) ,\]
\[ u=2 \left( 1+\frac{2 \sum _{i \neq j}^{N} \rho _{i,j} \; \sqrt[]{m_{i}m_{j}}}{ \sum _{1=1}^{N}m_{i}} \right) .\] \par
Figure \ref{fig:myrcode1} shows empirical distribution in the case of \( m=m_{1}=m_{2}=5 \), with overlapping density function of the gamma model and the exact distribution derived by Joarder et al. Data were simulated both as random draws from a bivariate \(\chi^{2}\) and by the quantile function method, i.e. by transformation of a bivariate normal variate; the results are compared between the two methods.
Interestingly, whereas, as expected, the gamma approximation has a slightly worse fit than the exact distribution on data simulated from a bivariate \(\chi^{2}\), on the other hand it shows a better fit when non-independence has a different functional form.
\begin{figure}[H]
\begin{knitrout}
\definecolor{shadecolor}{rgb}{0.969, 0.969, 0.969}\color{fgcolor}
\includegraphics[width=\maxwidth]{myrcode1-1}
\end{knitrout}
\caption{Empirical distribution and density function for a sum of \(\chi^{2}\) variables with m = 5. Variables on the left are simulated by the quantile function method, whereas variables on the right are drawn from a bivariate \(\chi^{2}\) distribution. Solid line: empirical distribution; dashed red: gamma distribution; dotted blue: exact distribution.\label{fig:myrcode1}}
\end{figure}
Figure \ref{fig:myrcode2} shows empirical distribution and overlapping density function for some combinations of parameters and correlation coefficients; data in this case were simulated by the quantile function method.
\begin{figure}[H]
\begin{knitrout}
\definecolor{shadecolor}{rgb}{0.969, 0.969, 0.969}\color{fgcolor}
\includegraphics[width=\maxwidth]{myrcode2-1}
\end{knitrout}
\caption{Empirical distribution (solid black) with overlapping gamma density function (dashed red) for a sum of \(\chi^{2}\) variables with \(m_{1}=5\)\:and\:\(m_{2}=7 \) and for a sum of correlated gamma distributions with common \(\theta\).\label{fig:myrcode2}}
\end{figure}
Since the \(\chi^{2}\) is just a gamma distribution with shape \(k=\frac{m}{2}\) and scale \(\theta = 2\), the approach can also be extended to any sum of correlated gamma variables with common scale parameter \(\theta\). If \(X_{1}, ... ,X_{N}\) are gamma variables with shape parameters \(k_{1}, ... ,k_{N}\), pairwise correlation coefficients \(\rho_{i,j}: i,j= 1, ..., N \) and common \(\theta\) then
\[ \sum _{i=1}^{N}X_{i} \sim \Gamma \left( \frac{ \sum _{i=1}^{N}k_{i}}{u},\theta u \right) ,\] \par
with
\[ u=1+\frac{2 \sum _{i \neq j}^{N} \rho _{i,j} \; \sqrt[]{k_{i}k_{j}}}{ \sum _{1=1}^{N}k_{i}} .\] \par
Figure \ref{fig:myrcode2} also displays the distribution fit for the sum of two correlated gamma distributions.
\section{Distribution of difference of correlated \(\chi^{2}\) variables}
A similar line of reasoning allows to approximate the distribution of the difference of two correlated \(\chi^{2}\) variables, at least in the case where \( m = m_{1} = m_{2} \) . \par
Let \par
\[ X=X_{1}-X_{2} .\] \par
Now when \(X_{1}\) and \(X_{2}\) are indipendent \(X\) is Variance-Gamma distributed [\cite{klar2015note}]. In particular we see from the moment generating function of the \(\chi^{2}\) distribution that\par
\[ M_{X_{1}} \left( t \right) =M_{X_{2}} \left( t \right) = \left( 1-2t \right) ^{\frac{-m}{2}} ,\] \par
\[ M_{X} \left( t \right) =M_{X_{1}} \left( t \right) M_{X_{2}} \left( -t \right) = \left( 1-4t^{2} \right) ^{\frac{-m}{2}}= \left( \frac{\frac{1}{4}}{\frac{1}{4}-t^{2}} \right) ^{\frac{m}{2}} ,\] \par
which is the moment generating function of the Variance-Gamma distribution, under generalized hyperbolic parametrization:
\[ M_{X} \left( t \right) =e^{ \mu t} \left( \frac{ \alpha ^{2}- \beta ^{2}}{ \alpha ^{2}- \left( \beta ^{2}+t^{2} \right) } \right) ^{ \lambda } ,\] \par
with \( \mu =0, \alpha =\frac{1}{2}, \beta =0, \lambda =\frac{m}{2} \) .\par
We change to the parametrization proposed by \cite{Seneta2004} for convenience, which in this case leads us to:\par
\[ c= \mu =0, \]
\[ \sigma =\sqrt[]{\frac{2 \lambda }{ \alpha ^{2}- \beta ^{2}}}=\frac{\sqrt[]{2 \lambda }}{ \alpha }=2 \; \sqrt[]{m}, \]
\[ \nu =\frac{1}{ \lambda }=\frac{2}{m}, \]
\[ \theta = \beta =0. \]
We can use the same approach as above and look at this distribution as a special case of difference of correlated \(\chi^{2}\) variables with \( \rho =0 \) . However, the approach cannot be faithfully replicated, since in the case of \( \rho =1 \) the distribution is degenerate with zero mean and variance. Here it is only known that the distribution becomes Variance-Gamma in the special case of \( \rho =0 \) . However, we know that in the more general case the distribution must have zero mean and that its variance must tend to zero as \( \rho \) approaches 1. \par
The mean of the Variance-Gamma distribution is\par
\[ E \left( X \right) =c+ \theta .\] \par
With \( c= \theta =0 \) , \( E \left( X \right) =0 \) . \par
The variance of the Variance-Gamma distribution is:\par
\[ Var \left( X \right) = \sigma ^{2}+ \theta ^{2} \nu , \] \par
which here becomes \par
\[ Var \left( X \right) = \sigma ^{2}=4m .\] \par
From properties of variance we also know that, in the correlated case:\par
\[ Var \left( X \right) =4m \left( 1- \rho \right) , \] \par
i.e. the variance should be scaled by \( 1- \rho \) factor to account for correlation. This is accomplished by multiplying \( \sigma \) by \( \sqrt[]{1- \rho } \) .\par
The distribution of a difference between two correlated \(\chi^{2}\) variables is therefore expected to be approximately\par
\[ X \sim VarianceGamma \left( 0,2\;\sqrt[]{m \left( 1- \rho \right) },0,\frac{2}{m} \right) . \]
for \( 0 \leq \rho <1 \) . As \( \rho \) approaches 1 the variance approaches 0.
The generalization to a difference of two gamma distributions with common \(k\) and \(\theta\) is very straight-forward and is obtained by setting \(\nu = \frac{1}{k}\) and \(\sigma=\theta \; \sqrt{2k}\).
If \(X_{1}\) and \(X_{2}\) are gamma variables with common shape and scale \(k, \theta\), then
\[ X_{1}-X_{2} \sim VarianceGamma \left( 0,\theta\;\sqrt[]{2 k \; \left( 1- \rho \right) },0,\frac{1}{k} \right) . \]\par
The fit of the distribution is tested on 100.000 simulations; Empirical distribution and overlapping density function are shown in Figure \ref{fig:myrcode3} for different values of \(\rho\).\par
\begin{figure}[H]
\begin{knitrout}
\definecolor{shadecolor}{rgb}{0.969, 0.969, 0.969}\color{fgcolor}
\includegraphics[width=\maxwidth]{myrcode3-1}
\end{knitrout}
\caption{Empirical distribution (solid black) with overlapping Variance-Gamma density function (dashed red) for the difference of two \(\chi^{2}\) variables drawn from a bivariate \(\chi^{2}\) distribution with m = 5 and for the difference of two arbitrary gamma distributions. \label{fig:myrcode3}}
\end{figure}
\section{Conclusions}
In this work it is shown that the distribution of the sum of a series of correlated \(\chi^{2}\) variables can be approximated by an adequately parametrized gamma distribution. The exact density function for said distribution has been derived before; however the approximation that is proposed here is much simpler and has more general applicability, since it generalizes easily to cases where multiple variables with different degrees of freedom are summed. As shown in simulations, the distribution closely fits the data; remarkably, whereas the fit was slightly worse than the exact distribution in data simulated from squaring multivariate normal variates, it was actually better when data were simulated using the quantile function method. This is likely due to the approximation being less reliant on specific distributional assumptions, which makes the method attractive to those who work with correlated gamma variables when little is known about the functional form of non-independence. \par
In the second part of the work it is shown, in a similar fashion, that the distribution of a difference between two correlated \(\chi^{2}\) variables with common degrees of freedom is, at least approximately, a variance gamma distribution with 0 mean and \( \sigma \) equal to \( 2m \;\sqrt[]{1- \rho } \). The density function fits the data in simulations. \par
Both results are generalized to sum and difference of gamma variables with common scale paramaters. These approximations are attractively simple, which given their very good fit to the empirical distributions suggest they could have widespread applications in statistics.
\renewcommand{\refname}{\vspace*{-1em}}
\specialsection*{References}
|
\section{Conclusions}
\label{sec:conclusions}
We presented maps of the Comptonization parameter of the twelve nearby, massive galaxy clusters
constituting the X-COP sample \citep{xcop:presentation}. We used an improved version of the spectral imaging
algorithm for parametric component separation proposed in \citetalias{bourdin:SZimaging}, featuring wavelet and
curvelet decomposition.
In particular, the enhancements we introduced are: (\emph{i}) a double grey body spectral energy density to model
the emission from thermal dust;
(\emph{ii}) the cleaning of residual contamination from dust and compact point sources;
(\emph{iii}) a new wavelet coefficient-wise deconvolution via beam-weighting in the calculation of the
wavelet transform.
We also illustrated a new method to estimate the error on the reconstructed tSZ signal based on bootstrap extractions
of \textsl{Planck} noise maps.
With this method, we showed it is possible to detect the signal from diffuse filaments and small substructures located
beyond $ R_{500} $. Indeed, we detected interesting features in the outskirts of the majority of the objects in the
sample under study.
In particular, we highlighted diffuse bridges in significant excess with respect to the
background level, which connect the brightest objects of the cluster systems A2029 and RXCJ1825.
These findings are consistent with ancillary surface brightness maps in the X-ray band.
When compared to its original implementation, our new deconvolution technique provided more stability,
an enhanced localisation of the central peak and a more precise reconstruction of the signal in the outskirts.
In particular, the test case of cluster A2319 showed a reduction of contamination from outliers
at radii $ r > 2 R_{500} $, and about 50 per cent lower values of the bootstrap error in the same region.
We plan to use the tSZ maps presented in this work to isolate and mask the signal from substructures located in
cluster outskirts, which will be the topic of a forthcoming paper.
Indeed, this signal may bias-high the reconstructed radial profiles of thermal pressure in the virial region, with
non-negligible consequences on the parameters of the profile, and on estimates of cluster masses relying on the
assumption of hydrostatic equilibrium.\\
Cluster A2319, which is the one detected by \textsl{Planck} with the highest S/N, will be the subject of a dedicated
upcoming work, in which we will highlight in detail the interesting features characterising the tSZ map that
we were able to reconstruct by employing sparse representations.
\section{Data set}
\label{sec:dataset}
We applied our imaging procedure to the sample of galaxy clusters studied in the XMM Cluster Outskirts Project (X-COP).
The goal of the project is to unveil the thermodynamic and dynamical properties of the ICM in the outermost cluster regions,
combining data in the X-ray and millimetre bands \citep[see][]{tchernin:a2142,xcop:presentation,
xcop:a2319,xcop:hydromassprofiles,xcop:nonthermalpressure,xcop:thermoproperties}.
The main physical properties of the X-COP clusters are summarised in Table~\ref{tab:xcop}.
\begin{table*}
\centering
\caption{\small Basic data of X-COP galaxy clusters~\citep[taken from][]{xcop:thermoproperties}.
Galactic coordinates are taken from the NASA/IPAC Extragalactic Database (\url{https://ned.ipac.caltech.edu}).}
\begin{tabular}{cccccccc}
\toprule
\midrule
Cluster name & %
\textsl{Planck} S/N & %
$ z $ & %
$M_{500}$ ($ \times 10^{14}$$ \text{M}_\odot $) & %
$R_{500} $ (kpc) & %
$\theta_{500}$ (arcmin) & %
$l $ (deg) & %
$b $ (deg)\\
\midrule
\object{A2319} & 30.8 & 0.0557 & 7.31 & 1346 & 20.8 & 75.70 & 13.51 \\
\object{A3266} & 27.0 & 0.0589 & 8.80 & 1430 & 21.0 & 272.13 & -40.13 \\
\object{A2142} & 21.3 & 0.0909 & 8.95 & 1424 & 14.1 & 44.22 & 48.68 \\
\object{A2255} & 19.4 & 0.0809 & 5.26 & 1196 & 13.1 & 93.97 & 34.95 \\
\object{A2029} & 19.3 & 0.0766 & 8.65 & 1414 & 16.3 & 6.44 & 50.53 \\
\object{A85} & 16.9 & 0.0555 & 5.65 & 1235 & 19.2 & 115.23 & -72.03 \\
\object{A3158} & 17.2 & 0.0597 & 4.26 & 1123 & 16.3 & 265.05 & -48.93 \\
\object{A1795} & 15.0 & 0.0622 & 4.63 & 1153 & 16.1 & 33.82 & 77.18 \\
\object{A644} & 13.9 & 0.0704 & 5.66 & 1230 & 15.3 & 229.93 & 15.29 \\
\object{A1644} & 13.2 & 0.0473 & 3.48 & 1054 & 19.0 & 304.88 & 45.45 \\
RXCJ1825$^{\tiny \ddag}$ & 13.4 & 0.0650 & 4.08 & 1105 & 14.8 & 58.31 & 18.54 \\
ZW1215$^{\tiny \ddag}$ & 12.8 & 0.0766 & 7.66 & 1358 & 15.7 & 282.50 & 65.19 \\
\bottomrule
\end{tabular}
\newline\newline\newline
$^{\tiny \ddag}$\tiny{The full identifiers for these two clusters are \object{RXC J1825.3+3026} and
\object{ZwCl 1215.1+0400}, respectively.}\hfill
\label{tab:xcop}
\end{table*}
The sample consists of twelve massive objects located at low to intermediate redshift, namely in the range
$ 0.04 < z < 0.10 $, observed in X-ray by \textsl{XMM-Newton} \citep{xmm:fov,xmm:resolution}
and at millimetre wavelengths by \textsl{Planck}.
As discussed in \citet{planck:PSZ1_early}, the six bands observed with the High Frequency Instrument (HFI) centred at
100, 143, 217, 353, 545, and 857 GHz are those which provide the highest signal-to-noise ratio (S/N hereafter)
for the detection of galaxy clusters.
In particular, the clusters in the X-COP sample have been detected with S/N $ > 12 $.
Their characteristic angular size $ \theta_{500} $, that is the angle subtended by the $ R_{500} $ radius of each cluster,
is larger than 10 arcmin, so that they are
well-resolved by \textsl{Planck}. A list of the HFI beams at all frequencies is reported in Table~\ref{tab:hfibeams}.
\begin{table}
\centering
\caption{Frequencies and corresponding beams of \textsl{Planck} HFI
\citep[from][]{planck:hfibeams2015}.}
\begin{tabular}{cc}
\toprule
\midrule
Frequency (GHz) & Beam FWHM (arcmin)\\
\midrule
100 & 9.69\\
143 & 7.30\\
217 & 5.02\\
353 & 4.94\\
545 & 4.83\\
857 & 4.64\\
\bottomrule
\end{tabular}
\label{tab:hfibeams}
\end{table}
We extracted $ 256 \times 256 $ pixel maps centred on each target cluster from the raw full-sky
maps provided in the latest \textsl{Planck} data release\footnote{\textsl{Planck} data are publicly
available at: \url{https://pla.esac.esa.int}} \citep[see][for details on HFI]{planck:hfi2018}.
Specifically, we retrieved the temperature maps, the associated variance maps and the jackknife maps
\citep[see][]{planck:hfijk} at each frequency band.
In particular, the latter are used in our procedure to estimate the error, as explained
in detail in Sect.~\ref{subsec:error}.
Sky patches are obtained through a gnomonic re-projection using the \texttt{HEALPix} package
\citep{healpix:presentation}.
Each pixel is 1 arcmin in size, so the maps extend over $ \approx 4.3^{\circ} $, corresponding to
sufficiently large radii (i.e. cluster-centric distances $ r > 3R_{500} $) to allow the detection of substructures in
cluster outskirts.
The signal in the raw temperature maps is corrected at all HFI frequencies by the astrophysical offsets
listed in Table~\ref{tab:offsets}, which account for the cosmic infrared background and the tSZ background.
\begin{table}[b]
\centering
\caption{Astrophysical offsets subtracted from each raw HFI map (see text for details).}
\begin{tabular}{cc}
\toprule
\midrule
Frequency (GHz) & Offset\\
\midrule
100 & \num{1.47e-5} K$_\tup{CMB}$\\
143 & \num{2.32e-5} K$_\tup{CMB}$\\
217 & \num{7.02e-5} K$_\tup{CMB}$\\
353 & \num{4.12e-4} K$_\tup{CMB}$\\
545 & \num{3.41e-1} MJy sr$ ^{-1} $\\
857 & \num{5.84e-1} MJy sr$ ^{-1} $\\
\bottomrule
\end{tabular}
\label{tab:offsets}
\end{table}
The offset values are computed as described in \citet{planck:dust2013}, from the correlation between
HFI data at 857 GHz and data delivered by the Leiden/Argentine/Bonn (LAB) radio survey of Galactic
HI \citep{lab:4hfioffsets}.
The pixels in the sky that are used to compute this correlation are selected according to the values of both the column
density and the velocity of HI clouds. In order to remove the contamination from galaxy clusters, and to account
for calibration at high frequencies, we also added two more selection criteria. One is based on masking pixels where
the the Comptonization parameter estimated from the maps obtained with the MILCA algorithm \citep{hurier:milca}
exceed the threshold value of \num{e-5}.
The second one accounts for a correction of the CMB using SMICA maps \citep{smica:presentation}
at frequencies between 100 GHz and 353 GHz as templates.
\section{Introduction}
\label{sec:introduction}
Clusters of galaxies, which are the largest gravitationally bound structures in the universe, can be studied at
millimetre wavelengths through observations of the Sunyaev--Zel'dovich (SZ) effect \citep{sz:70,sz:72}.
Indeed, as the photons of the cosmic microwave background (CMB) radiation propagate through the hot and ionised
intra-cluster medium (ICM hereafter) of galaxy clusters, they get diffused via inverse Compton scattering by
the free electrons. The consequence of this interaction is that photon energy is re-distributed; therefore, the
observed black body spectrum of the CMB gets distorted.
The spectral distortion produced by the random thermal motion of the electrons is known as the thermal SZ (tSZ) effect.
A remarkable property is its linear dependence on both electron density and temperature, making it a direct probe
of the thermal pressure of the ICM integrated along the observer line of sight.
For this reason, the tSZ effect is extremely useful when investigating the outskirts of galaxy clusters, which host complex
physical processes associated to the continuous accretion of matter
\citep[see e.g.][for a recent review]{walker:outskirtsreview}.
In this respect, the XMM Cluster Outskirts Project \citep[X-COP, see][]{xcop:presentation}, proposes combining the
tSZ effect with X-ray data at radii $ R_{500}< r < R_{200} $\footnote{Whenever the subscript $ \Delta = 200,500 $ is used, we
refer to the mass enclosed by a spherical volume of radius $ R_{\Delta} $, whose density is $ \Delta $ times the
critical density of the universe at a given redshift $ z $, $ \rho_c(z) = 3 H(z)^2/(8 \pi G) $, being $ H(z) $ the
Hubble parameter, and $ G $ the gravitational constant},
which are barely accessible to X-ray spectroscopy because of
significant contamination by unresolved astrophysical or instrumental backgrounds.
Another important feature of the tSZ effect, which makes it unique among other observables at microwave frequencies,
is its peculiar spectral signature. Indeed, the observed change of the CMB brightness manifests as a decrement at frequencies
below the zero-frequency $ \nu_0 = \SI{217}{\GHz} $, and as an increment at higher frequencies.
In the non-relativistic approximation \citep{kompaneets:articolo}, both the zero-frequency and the overall spectral
signature do not depend on cluster physical parameters. Therefore, thanks to this characteristic, it is possible to
disentangle the tSZ signal from other astrophysical sources using multi-frequency data.
The separation of different components contributing to an astrophysical signal is an issue that has been extensively
addressed in the literature with a variety of techniques, and for diverse purposes.
Among the most used methods, there are refinements of the internal linear combination (ILC),
which was first proposed by \citet{eriksen:ilc}.
These methods assume there is no correlation between the different physical components, and they are non-parametric,
which means that they do not rely on any particular model. The contribution from each astrophysical source is estimated
through a simple linear combination of the total signal at all the available frequencies. The weighting coefficients for
each frequency band are chosen in such a way that they provide unit response with respect to the component of interest,
while minimising the variance of the reconstructed map.
The increasing amount of microwave data at high angular
resolution, which can be of a few arcmin or arcsec with currently operating instruments
\citep[see e.g.][for a recent review]{mroczkowski:szreview},
and the unprecedented frequency coverage provided by
the \textsl{Planck} satellite \citep{planck:presentation}, motivated the development of several modified ILC algorithms
tailored to the extraction of the tSZ effect.
Relevant examples in this respect are provided by: \citet{remazeilles:cilc}, \citet{hurier:milca} (who propose the
modified internal linear combination algorithm: MILCA), and \citet{hurier:milcann}.
Despite their success, ILC methods for the extraction of the tSZ effect suffer from some drawbacks
\citep[see e.g.][for a discussion]{bobin:compsep}.
For instance, results can be significantly biased if the different foreground sources are correlated, as well as if
they exhibit a spatially variable spectral energy density.
Moreover, ILC methods perform better in the case of Gaussian foreground fluctuations, which is not a realistic hypothesis
for the tSZ effect and the emission from thermal dust.
In addition, the combination of heterogeneous data sets from ground-based and space-based experiments has been explored
through the proposal of more sophisticated techniques, with the aim of exploiting the advantages coming from
observations at different angular resolutions and sensitivities.
For instance, \citet{remazeilles:nilc} propose a method which implements needlet decomposition, while in the recent work
by \citet{aghanim:pact} the authors apply suitable matched multi-filters
\citep[see e.g.][]{herranz:matchingfilterSZ,melin:SPTSZmapping,tarrio:mmf16,tarrio:mmf18}
to data from \textsl{Planck} and from the Atacama Cosmology Telescope \citep{act:presentation}.
Besides non-parametric algorithms, methods based on a prior modelling of the components and on likelihood maximisation
have also been developed \citep[see e.g.][]{feroz:bayesclusters,khatri:parametricSZ}.
In particular, parametric techniques are robust, especially for the mapping of Galactic thermal dust
\citep[see for instance][]{planck:dust2013,meisner:twodust}.
On the other hand, algorithms based on sparse representations allow one to exploit some properties of the foreground
sources, such as their morphological independence or positive normalisation.
In \citet{bourdin:SZimaging} (\citetalias{bourdin:SZimaging} hereafter) the authors propose to link sparse imaging
to a parametric estimate of foregrounds, via the introduction of a spatially-weighted likelihood function through
wavelets and curvelets.
Wavelets are basis functions characterised by a scale (or equivalently by its inverse, that is the resolution) and by a
position parameter.
Wavelet transforms provide sparse representations of signals characterised by a smooth distribution, apart from localised
discontinuities. Since the ICM seems to hold such properties, wavelet transforms have been widely used in cluster
astronomy, both for X-ray \citep{slezak:waveletxray,vikhlinin:waveletxray,starck:waveletxray} and tSZ imaging
\citep{pierpaoli:waveletsz,pires:waveletsz}.
Another relevant application for millimetre astronomy is the foreground cleaning of CMB maps proposed by
\citet{bobin:compsep,bobin:compsep16}, which relies on sparse priors to spatially separate the components.
Curvelet transforms are a higher-dimensional generalisation of wavelet transforms. Indeed, curvelet bases
are parametrised by two additional quantities, namely direction and elongation \citep{candes:curvelets}.
This property makes them particularly suitable to get sparse representations of images featuring
edges and anisotropies \citep{mallat:waveletbook}, as in the case of shocks and filamentary structures in
the outskirts of galaxy clusters.
For the reasons above, in \citetalias{bourdin:SZimaging} the authors employ wavelet decomposition in
the extraction of the CMB signal and curvelet decomposition to separate the anisotropic signal from the thermal dust
and the tSZ effect.
The algorithm is tested using mock \textsl{Planck} frequency maps from the set of high-resolution
hydrodynamical simulations of galaxy clusters presented in \citet{planelles:sims14} and \citet{rasia:sims14}.
We applied for the first time the spectral imaging algorithm by \citetalias{bourdin:SZimaging}
to the frequency maps from the latest data released by the \textsl{Planck} Collaboration in 2018.
Furthermore, we improved the algorithm by implementing an efficient removal of astrophysical contaminations,
such as the emission from thermal dust and bright point sources. Most importantly, we enhanced its stability by means
of a new deconvolution technique applied to the wavelet coefficients, proposed for the first time in this work.
We show the maps of the tSZ effect for a set of twelve massive and nearby clusters of galaxies selected for X-COP.
We focus in particular on two interesting cases of possibly interacting cluster couples, and we compare our images
to ancillary X-ray data. The significance of the signal is established through a new boostrap-based procedure for
the estimate of the tSZ error maps, which we also used to show the overall enhancement of the algorithm, compared to
its previous implementation.
The paper is structured as follows.
In Sect.~\ref{sec:dataset} we briefly introduce the cluster data set.
Section~\ref{sec:methods} illustrates the details of the algorithm we used for producing maps of the tSZ effect,
focussing on its new aspects.
In Sect.~\ref{sec:results} we present and discuss our results for the X-COP sample. In particular, we make a comparison
with the implementation of the algorithm presented in \citetalias{bourdin:SZimaging}, and we detail the cases of
cluster A2029 and RXCJ1825.
Finally, we summarise our results and outline future perspectives in Sect.~\ref{sec:conclusions}.
Throughout the paper, we assume a $ \Lambda $CDM cosmological model with $ \Omega_{\Lambda} = 0.7 $, $ \Omega_m = 0.3 $
and $ h = 0.7 $.
\section{Imaging of the tSZ effect}
\label{sec:methods}
To map the tSZ signal, we used a parametric imaging algorithm featuring wavelet and curvelet decomposition.
This procedure was originally proposed in \citetalias{bourdin:SZimaging} and tested only on mock HFI observations
of interacting clusters.
The first application to real HFI data motivated substantial adaptations and improvements of
its original implementation.
In particular, we used the spectral energy density of a double grey body to model the emission from thermal dust,
and we corrected for residual contamination from both dust and bright point sources.
Another remarkable improvement concerns with the stability of the algorithm, which we obtained
through a new deconvolution procedure proposed here for the first time.
In this Section we first provide a synthetic description of the original algorithm, then we focus
on the aforementioned enhancements.
\subsection{The spectral imaging algorithm}
\label{subsec:imaging}
The novelty of our imaging method consists of combining the parametric approach to component separation
and sparse representations.
This is achieved by finding the parameters which minimise a modified chi-square, accounting for the
wavelet transform of the residuals between the data and the model map at each frequency $ \nu $.
Referring to the $ k$-th pixel in the maps, the residuals can be simply written as
\begin{equation}
\label{eqn:residuals}
\text{res}(\nu,k;s) = D_\tup{HFI}(\nu,k)-M(\nu,k;s) \ ,
\end{equation}
where $ D_\tup{HFI} $ represent the HFI data.
The source component maps $ s $, which are the parameters we want to estimate, enter the residuals via
the model map, $ M $. The latter can be written explicitly as
\begin{equation}
\label{eqn:model}
M(\nu, k; s) = \sum\limits_{i}^{N_s} f_i(\nu) \ s_i(k) + \eta(\nu,k) \ ,
\end{equation}
that is the sum over the temperature anisotropies produced by the $ N_s $ physical sources to separate, $ s_i $,
multiplied by the corresponding spectral energy density $ f_i(\nu) $, plus the instrumental noise, $ \eta $.
The dominating astrophysical sources in the frequency range covered by \textsl{Planck} HFI,
apart from the CMB, are the tSZ effect and the Galactic thermal dust \citep[see for instance][]{planck:foregrounds}.
The spectral energy density of the tSZ effect is treated in the non-relativistic case, so that
\citep[e.g.][]{birkinshaw:szreview}
\begin{equation}
\label{eqn:ftsz}
f_\tup{tSZ}(\nu) = \frac{h_p \nu}{k_B T_\tup{CMB}}\times%
\coth \left(\frac{h_p \nu}{2 k_B T_\tup{CMB}} \right) - 4 \ ,
\end{equation}
where $ h_p $ is the Planck constant, $ k_B $ is the Boltzmann constant, and $ T_\tup{CMB} = (2.725 \pm 0.001)$ K
\citep{mather:tcmb} is the CMB temperature, which is constant at all frequencies.
The amplitude of the tSZ signal is given by the Compton $ y $-parameter \citep[see e.g.][]{rephaeli:articolo}:
\begin{equation}
\label{eqn:y}
y = \frac{\sigma_T}{m_e c^2} \int_\tup{los} p(l) \ dl \ ,
\end{equation}
where the integral is calculated along the line of sight (los), being $ \sigma_T $ the Thomson cross section,
$ m_e $ the electron mass, $ c $ the speed of light, and $ p(l) $ the thermal pressure of the ICM.
The spectral modelling of Galactic thermal dust (td subscript) is presented in detail in
Sect.~\ref{subsec:imaging_novelties},
since it represents a major change in the algorithm.
The residuals in Eqn.~\eqref{eqn:residuals} can be rewritten in terms of their wavelet transform as
\citep[see e.g.][]{mallat:waveletbook}
\begin{equation}
\label{eqn:residuals15}
\begin{split}
\text{res}_{\Psi}(\nu,k;s) &= \sum_n^{N_\tup{pix}} \bar a_{j_0,n}(\nu;s) \ \Phi_{j_0,n}(k) \ +\\
&+ \sum_{j=j_0}^{N_\tup{scales}} \sum_n^{N_\tup{pix}} a_{j,n}(\nu;s) \ \Psi_{j,n}(k) \ ,
\end{split}
\end{equation}
where $ j $ and $ n $ give the dilation and translation of the wavelet basis function $ \Psi $, respectively, being
$ 2^j $ the wavelet scale. In our case, the wavelet basis function is a B$_3 $ spline, and $ \Phi $ is the dual scaling
function of $ \Psi $ at the scale $ j_0 $. This scale corresponds to the approximation level -- also called last smooth --
which encodes signal information at the lowest resolution.
The wavelet coefficients in Eqn.~\eqref{eqn:residuals15} are given by
\begin{align}
&\bar a_{j_0,n}(\nu;s) = \sum_m^{N_\tup{pix}} \text{res}(\nu,m;s) \ \Phi^*_{j_0,n}(m) \ ; \label{eqn:coeffresiduals}\\
&a_{j,n}(\nu;s) = \sum_m^{N_\tup{pix}} \text{res}(\nu,m;s) \ \Psi^*_{j,n}(m) \ , \label{eqn:coeffresidualsdetail}
\end{align}
for the approximation and detail levels, respectively.
In order to ensure normalisation and positivity when weighting the residuals, the wavelet kernel $ \Psi $ is
split in its positive and negative parts, $ \Psi_{+} $ and $ \Psi_{-} $ (see fig.~1 of \citetalias{bourdin:SZimaging}
for reference).
This yields the minimisation of two separate chi-squares, which can be expressed as the sum across all the
frequencies and pixels ($ N_{\nu} $ and $ N_\tup{pix} $, respectively) of such weighted residuals:
\begin{equation}
\label{eqn:chi2B15}
\chi^2_{\Psi_{\pm}} = \sum_{\nu}^{N_{\nu}} \sum_{k}^{N_\tup{pix}} \
\frac{\text{res}_{\Psi_{\pm}}^2(\nu,k;s)}{\sigma_\tup{HFI}^2(\nu,k)} \ ,
\end{equation}
where $ \sigma^2_\tup{HFI} $ is the HFI variance map.
Therefore, the estimate of the component maps is given by the half-sum of the results from the two separate
minimisations:
\begin{equation}
\label{eqn:wtcoeffs}
\hat s = \frac{1}{2} \left[\underset{s}{\text{argmin}}(\chi^2_{\Psi_{+}}) -
\underset{s}{\text{argmin}}(\chi^2_{\Psi_{-}})\right] \ .
\end{equation}
To maximise the efficiency in the recovery of signal anisotropies, we chose a curvelet basis to decompose
the tSZ and dust signals \citep[see e.g.][]{candes:curvelets}, and a wavelet basis for the CMB
\citepalias[see][for more details]{bourdin:SZimaging}. The decomposition has been performed over
four scales\footnote{The maximum number of scales that can be used for
wavelet decomposition is linked to the number of pixels on each side of the image as:
$ N_\tup{scales} = \text{floor} [\log(N_\tup{pix})/\log 2]$ - 4}.
The curvelet transform has been computed following the procedure described in
\citet{starck:curvelets}, that is to say trough ridgelet transforms applied to the bidimensional wavelet bands of
the B$_3 $ spline wavelet transform.
Subsequently, we applied a soft thresholding at $ 1 \sigma $ level to the curvelet and wavelet
coefficients, in order to keep only the relevant features of the signal.
Eventually, the final maps have been obtained by means of a suitable restoration operator,
which re-combines the approximation coefficients and the detail coefficients together.
Pixels plagued by contaminants, due for instance to a local bad modelling of thermal dust, have been automatically
masked. Specifically, we discarded those pixels where chi-square exceeded a given threshold value, depending on the
wavelet scale. At the same time, we imposed a condition of regularity on the error of the wavelet coefficients, in order
to select regions characterised by a high S/N of the tSZ component.
\subsection{Improvements and new features}
\label{subsec:imaging_novelties}
The three main changes we implemented with respect to the version of the imaging
algorithm presented in \citetalias{bourdin:SZimaging} are detailed in the following subsections.
Two of them concern with a more realistic treatment of astrophysical contaminants. The third one consists of a new
deconvolution technique proposed here for the first time.
\subsubsection{Modelling of thermal dust}
Following \citet{meisner:twodust} \citep[see also][]{bourdin:pressure},
we modelled the thermal dust as a double grey body, by assuming two populations of dust grains, instead of
the idealised case of a single grey body spectrum. Indeed, the latter provides an accurate representation of the
thermal emission from Galactic dust only at frequencies higher than 353 GHz.
The spectral energy density we set for this component is therefore
\begin{equation}
\label{eqn:meisnergb}
\begin{split}
f_\tup{td}(\nu) &=
f_1 \frac{q_1}{q_2} \left(\frac{\nu}{\nu_0} \right)^{\beta_1} B(\nu;T_1) \ + \\
&+ (1-f_1) \left(\frac{\nu}{\nu_0} \right)^{\beta_2} B(\nu;T_2) \ ,
\end{split}
\end{equation}
where the dimensionless constant factors $ f_1 $ and $ q_1/q_2 $ refer to the relative contribution
from the coldest component at temperature $ T_1 $ and the hottest component at temperature $ T_2 $.
The $ \beta_1 $ and $ \beta_2 $ parameters give the slopes of the two different power laws, while
$ B(\nu; T_1) $ and $ B(\nu; T_2) $ are the corresponding Planck functions describing the black body spectra.
In order to get the best-fit parameters of the model described by Eqn.~\eqref{eqn:meisnergb},
we calculated an independent fit to the dust component through a Monte Carlo Markov Chain sampling.
To treat only the signal from the dust, we limited this fit to the pixels in the frequency maps that are located
sufficiently far from the cluster, at radial distances from the centre larger than $ 5 R_{500} $.
The only spatially-variable parameter is the temperature $ T_2 $, which is fixed a priori to the value
determined by a joint fit to \textsl{IRAS} and \textsl{Planck} data, as detailed in \citet{meisner:twodust}.
From this fit we obtain maps of the dust component at all frequencies, which we plug in the model
maps of Eqn.~\eqref{eqn:model}.
\subsubsection{Removal of contamination from thermal dust and point sources}
\label{subsubsec:adaptation}
Despite the two grey body model provides an accurate description of the thermal dust component,
our maps of the Compton $ y $-parameter turned out to be affected by contamination either from diffuse signal
from thermal dust on large spatial scales\footnote{We notice, however, a significant reduction of this
contamination when comparing maps from \textsl{Planck} 2015 data release to those from the 2018 data release},
or from compact point sources, mostly at the frequency of 857 GHz.
To solve these issues, we proceeded as described in the following.
We reconstructed the tSZ images without accounting for the contribution from the 857 GHz frequency
in the last approximation of its discrete wavelet transform, $ \bar a_{j_0,n} $ (see also Eqn.~\eqref{eqn:residuals15}).
On the other hand, we took advantage of all the HFI frequency maps to compute the detail coefficients of
the wavelet transform, $ a_{j,n} $, supposedly insensitive to image gradients below the last approximation scale,
$ j_0 $. This allowed us to suppress unphysical large-scale gradients in the approximation coefficients of
some cluster images. These gradients could be possibly related to a residual contamination of the thermal dust signal
by the large-scale variance of other astrophysical sources or of the instrumental noise.
To remove the contamination from point sources, we used the masks in the second \textsl{Planck} catalogue
of compact sources (PCCS2) \citep[see][for details]{planck:compactsources,planck:compactsources2}.
These objects are dusty galaxies and synchrotron emitting in the HFI range at high and low frequencies,
respectively.
Specifically, we combined the gnomonic projection of all the masks at the six HFI frequencies into a unique
mask, after smoothing them to the common resolution of 18 arcmin. To remove the contribution from such sources,
we multiplied this mask by the wavelet kernel used in the chi-square minimisation. We applied this procedure only to
the clusters in our sample which are significantly affected by bright point sources at far infrared frequencies,
namely A3266, A85, and ZW1215.
\subsubsection{Deconvolution procedure}
\label{subsubsec:deconvolution}
As can be seen from Eqn.~\eqref{eqn:chi2B15}, the weighted chi-square mixes the different angular resolutions
of the data. Therefore, a deconvolution must be implemented to restore the signal.
The original version of the algorithm features a Van Cittert deconvolution \citep{vancittert:deconvolution},
regularised via successive projections in the significant curvelet domain.
In other words, the resolution of the model map is adjusted to the resolution of
the data map through iterative corrections of the curvelet coefficients.
However, several tests revealed that this technique hampered the reliability of the signal reconstructed in low
signal regimes, because of the amplification of pixel-sized diverging artefacts across the iterations.
To avoid this problem, we implemented a new ``wavelet coefficient-wise'' deconvolution of the \textsl{Planck} beams.
Specifically, we deconvolved the amplitude of the single wavelet coefficients, instead of deconvolving the final image.
To this end, we took the full expression of the residuals as in Eqn.~\eqref{eqn:residuals}, and we calculated their
wavelet coefficients as follows (see also Eqn.~\eqref{eqn:coeffresidualsdetail} for a comparison):
\begin{equation}
\label{eqn:coeffresidualsdeconv}
\begin{split}
a_{j,n}(\nu;\Delta s) &= \sum_m^{N_\tup{pix}} \text{res}(\nu,m; \Delta s) \ \Psi^*_{j,n}(m) \\
&=\sum_m^{N_\tup{pix}} \lbrace D_\tup{HFI} (\nu,m) - \mathcal{B}(\nu) \otimes [H M(\nu,m;\tilde s + \Delta s)+\\
&+ (1-H) M(\nu,m;\tilde s - \Delta s)]\rbrace \ \Psi^*_{j,n}(m) \ ,
\end{split}
\end{equation}
(the same holds for the approximation level).
We rewrite the parameter as $ s = \tilde s \pm \Delta s $ in Eqn.~\eqref{eqn:coeffresidualsdeconv}, to
highlight its spatial variation, $ \Delta s $.
Instead of splitting the wavelet kernel in two components, we used the Heaviside step function, $ H $,
that coincides with the positive part of the kernel, while $ (1-H) $ coincides with its negative part.
This operation is equivalent to taking the absolute value of the wavelet function.
To match the angular resolutions of the data and the model, we convolved the model maps at each frequency with the
\textsl{Planck} beam, $ \mathcal{B}(\nu) $, which can be approximated with a Gaussian having the FWHM listed in
Table~\ref{tab:hfibeams}.
The effect of such a convolution is that of correlating the contributions from the
negative and the positive components of the kernel (see also Fig.~\ref{fig:newdeconv}).
Thus, the final estimate of the parameters takes into account the results from the single decompositions.
Besides allowing us to deconvolve spatial variations of the model maps, the weighting by the absolute value of the wavelet
kernel within a single chi-square minimisation increases the support of the kernel itself, as can be seen in
Fig.~\ref{fig:kerneldeconv}. This is equivalent to enlarging the size of the sky region where we compute the chi-square,
thus enhancing the S/N.
\begin{figure}
\centering
\subfloat[Parameter $ s $]{\includegraphics[width=0.23\textwidth]{sgl_gaussconv-eps-converted-to}
\label{fig:parameterdeconv}}
\subfloat[Wavelet kernel]{\includegraphics[width=0.23\textwidth]{bsplinemod_gaussconv-eps-converted-to}
\label{fig:kerneldeconv}}
\caption{Effect of beam-weighting of wavelet coefficients.
\textbf{Fig.~\ref{fig:parameterdeconv}.} Step-like spatial variation of $ s $ (solid black line).
The filled black area represents its convolution with the instrumental beam (dashed red line).
\textbf{Fig.~\ref{fig:kerneldeconv}.} Effective spatial variation of $ s $.
The blue and red lines represent the absolute value of the positive and negative parts of the wavelet kernel,
respectively. The envelopes of the shaded areas coincide with the convolution of the instrumental beam with
the support of the positive and negative parts of the kernel.}
\label{fig:newdeconv}
\end{figure}
\subsection{Procedure for error assessment}
\label{subsec:error}
To associate a statistical error to the estimate of the component maps, we
used a bootstrap procedure that allowed us to simulate $ N_\tup{tot} $ sets of HFI maps.
More specifically, we followed the steps detailed below.
Firstly, we generated $ N_\tup{tot} = 100 $ Monte Carlo realisations of the noise at each HFI frequency,
$ \eta_u(\nu) $, with $ u = 1,\dots, N_\tup{tot} $ referring to the $ u $-th extraction.
We chose the value of $ N_\tup{tot} $ as a trade-off between the computational time needed
to perform wavelet and curvelet transforms and the statistical significance.
We assumed the instrumental noise to be Gaussian and spatially correlated, and we
imposed the noise maps to have the same power spectrum as the jackknife maps.
Subsequently, the HFI raw data at each frequency, $ D_\tup{HFI}(\nu) $, have been denoised through a simple
wavelet-based procedure, by calculating the wavelet transform on three scales, and soft-thresholding the coefficients
at $ 1.5 \sigma $. This procedure relies on spatially-variant thresholds for each wavelet band, which have been
preliminarily inferred by computing the variance of the coefficients across wavelet transforms of the noise maps,
$ \eta_u(\nu) $.
Lastly, the mock data have been obtained as the summation of the denoised maps,
$ D_\tup{HFI,den}(\nu) $, and the $ u $-th noise realisation, that is
\begin{equation}
\label{eqn:mockhfi}
{D_\tup{HFI}}_u(\nu) = D_\tup{HFI,den}(\nu) + \eta_u(\nu) \ .
\end{equation}
We used such $ N_\tup{tot} $ synthetic data sets as input to the imaging algorithm to obtain the following
vectors of maps:
$ \bm s_\tup{tSZ} = (s_\tup{tSZ}^1, \dots , s_\tup{tSZ}^{N_\tup{tot}}) $,
$ \bm s_\tup{td} = (s_\tup{td}^1, \dots , s_\tup{td}^{N_\tup{tot}}) $,
$ \bm s_\tup{CMB} = (s_\tup{CMB}^1, \dots , s_\tup{CMB}^{N_\tup{tot}}) $.
The standard deviations of all the bootstrap sets, namely $ \text{std}(\bm s_\tup{tSZ})$,
$ \text{std}(\bm s_\tup{td})$, and $ \text{std}(\bm s_\tup{CMB})$, represent our error estimates
for each component.
In the following, we label as $ \sigma_y = \text{std}(\bm s_\tup{tSZ}) $ the error for the tSZ signal only.
The level of significance of the anisotropies (such as blobs and filaments) we find in the tSZ images of each
cluster can be assessed by computing the ratio $ y/\sigma_y $, for a given minimum value of $ \sigma_y $,
which represents an `effective' S/N.
\section{Results and discussion}
\label{sec:results}
We show in Fig.~\ref{fig:ymaps} the zoomed maps of the Compton $ y $-parameter we obtained from the application of the
procedure described in Sect.~\ref{sec:methods} to the HFI maps of the X-COP clusters.
\begin{figure*}[th]
\centering
\includegraphics[width=\textwidth]{{{unique_maps_4scales_128_log_wcsPLANCK2018}}}
\caption{Compton $ y $-parameter of X-COP clusters from our spectral imaging.
The solid and dashed cyan circles on top of the images mark the $ R_{500} $ and $ R_{200} $ radii, respectively.
The colour scale is logarithmic and contours are logarithmically spaced by a factor of $ \sqrt[4]{2} $,
starting from \num{2e-6}.
The J2000 Equatorial coordinates of the cluster centres are reported in the bottom right corner of each map.}
\label{fig:ymaps}
\end{figure*}
In the majority of cases, the algorithm highlights the presence of blobs or extended filamentary structures at
radial distances $ r \gtrsim R_{500} $ from the cluster centre.
The average minimum values of the Compton $ y $-parameter we can recover from the faintest resolved
structures are \num{2.0e-6} and \num{4.2e-6}, with a minimum significance of $ 3 \sigma_y $ and $ 5 \sigma_y $, respectively.
This represents an improvement of a factor of 2.5 in sensitivity with respect to the version of the algorithm
presented in \citetalias{bourdin:SZimaging}, where the minimum signal for a $ 3 \sigma $ detection is $ y = \num{5e-6} $.
It is important to stress that, thanks to the deconvolution, the algorithm is capable to recovering the elongated
structure of the signal in the cluster central regions (see e.g. the cases of A2142 and A1644).
A particular cluster case which is worth mentioning is that of A2319, which is known to be a complex merging system
that behaves as an outlier with respect to the other clusters in the sample \citep[see][]{xcop:a2319,
xcop:nonthermalpressure,xcop:thermoproperties}.
The tSZ map of this object shows a number of blobs located in the virial region, which may be due to either substructures
or clumpy ICM patches. The detailed study of these features and of their impact on the thermodynamic properties of this
system will be addressed in a forthcoming dedicated paper.
In the following, we first perform a comparison with the results from the version of the algorithm presented in
\citetalias{bourdin:SZimaging} for cluster A2319. Subsequently, we briefly discuss a couple of two possibly
interacting systems, namely A2029 and RXCJ1825.
\subsection{Comparison with the previous version of the algorithm}
We illustrate here the improvements introduced by the adaptation of our procedure to real cluster data, and by the
wavelet coefficient-wise deconvolution. In particular, in order to highlight the impact of removing the highest
frequency from the approximation coefficients (see Sect.~\ref{subsubsec:adaptation}), we adopted a double grey
body model for thermal dust also in the procedure by \citetalias{bourdin:SZimaging}.
To perform a consistent comparison, we applied the two versions of the algorithm to the same HFI maps.
The wavelet decomposition has been performed here over three scales, and coefficients have been thresholded at $ 1\sigma $.
Fig.~\ref{fig:a2319oldnew} shows the maps of the Compton $ y $-parameter from
the old and new versions of the algorithm (see Figs.~\ref{fig:a2319old} and~\ref{fig:a2319new}, respectively),
in the case of cluster \object{A2319}.
\begin{figure}[h]
\centering
\subfloat[Van Cittert deconvolution]{\includegraphics[width=0.33\textwidth]{{{a2319_vc_sz}}}
\label{fig:a2319old}}\\
\subfloat[New deconvolution]{
\hspace{-0.10cm}\includegraphics[width=0.335\textwidth]{{{a2319_sz}}}
\label{fig:a2319new}}\\
\subfloat[HFI at 857 GHz]{
\hspace{0.02cm}\includegraphics[width=0.33\textwidth]{{{a2319_800}}}
\label{fig:a2319857GHz}}
\caption{Region centred on cluster A2319.
\textbf{Fig.~\ref{fig:a2319old}.} Compton $ y $-parameter reconstructed with Van Cittert deconvolution.
\textbf{Fig.~\ref{fig:a2319new}.} Compton $ y $-parameter reconstructed with the
wavelet coefficient-wise deconvolution.
The colour scale is logarithmic and contours are logarithmically spaced by a factor of $ \sqrt[4]{2} $.
\textbf{Fig.~\ref{fig:a2319857GHz}.} Raw HFI map at 857 GHz.}
\label{fig:a2319oldnew}
\end{figure}
This object is particularly suitable to show the effect of removing the 857 GHz channel from the last smooth
coefficients.
Indeed, in Fig.~\ref{fig:a2319old} it is possible to see a large-scale gradient to the left-hand side of the image
at $ y \approx \num{6e-6} $. Such a contaminant is likely due to a residual
signal from thermal dust, as can be seen from the comparison with the HFI map at 857 GHz, shown in
Fig.~\ref{fig:a2319857GHz}. This residual signal is significantly removed with the new adaptations,
as demonstrated in the map in Fig.~\ref{fig:a2319new}.
Most notably, Fig.~\ref{fig:a2319oldnew} highlights how the two deconvolution techniques have a different impact on
the final maps.
Indeed, the result in Fig.~\ref{fig:a2319old} is produced with the Van Cittert deconvolution using three
iterations, and with the convergence parameter $ \alpha $ set to 0.25 (see eq.~(9) of \citetalias{bourdin:SZimaging}).
It can be seen that this map shows a number of pixel-sized outliers, which plague regions where
the signal is $ y \lesssim \num{7.5e-6} $.
Such artefacts are due to a diverging amplification of the curvelet coefficients, as a consequence of the
iterative nature of this deconvolution algorithm.
The map in Fig.~\ref{fig:a2319new} shows instead that the result from the new wavelet coefficient-wise
deconvolution is cleaner, and the mildly ellipsoidal shape of the
signal in the centre is still well recovered.
It is worth noting that both techniques are sensitive to the basis functions used to project the deconvolved signal.
To be specific, Van Cittert method is a regularisation applied to curvelet coefficients, therefore it is particularly
suitable to highlight ellipsoidal and elongated features. The wavelet coefficient-wise deconvolution, on the contrary,
is applied to wavelet coefficients, thus it provides a better localisation and detection of peaks.
This can be seen in in Fig.~\ref{fig:a2319new}, where the signal in the cluster centre is recovered more efficiently, with
respect to Fig.~\ref{fig:a2319old}.
Another important element that improves peak localisation is the larger kernel
support which enters the chi-square minimisation in the new deconvolution (see Sect.~\ref{subsubsec:deconvolution}
and Fig.~\ref{fig:newdeconv}). Indeed, this feature allows the reconstruction of
the signal with a higher S/N, since in this case the wavelet coefficients are computed and deconvolved on a
larger sky region.
To further illustrate the differences between the two deconvolutions, we show as solid black lines in
Fig.~\ref{fig:boot_oldnew} the profiles of the Compton $ y $-parameter. Each profile has been extracted as the average
vertical cut within a 10 arcmin-wide band, passing through the centre of the corresponding tSZ image in
Fig.~\ref{fig:a2319oldnew}.
The superimposed light grey lines represent the same profiles from each of the $ N_\tup{tot} = 100 $ different maps of
the Comptonization parameter obtained with the bootstrap procedure (see Sect.~\ref{subsec:error}).
It can be seen from Fig.~\ref{fig:a2319boot_new_vert} that the non-iterative nature of the new deconvolution improves
significantly the stability in low-signal regimes, allowing one to reach a minimum level of $ y $ of the order of \num{e-6}.
On the other hand, the diverging artefacts which hampered the reliability of the detection of substructures in the
outskirts with the Van Cittert deconvolution, are clearly visible in several realisations of the signal
in Fig.~\ref{fig:a2319boot_old_vert}.
Moreover, the dispersion of the bootstrap profiles at radii $ r \gtrsim 2 R_{500} $ is on average 50 per cent lower
with the new procedure, resulting into a lower error in the reconstructed signal in these regions.
\begin{figure}[t]
\centering
\subfloat[Van Cittert deconvolution] %
{\includegraphics[width=0.35\textwidth]{{{a2319_bootstrap_old_reproj_vertical_avg10_log}}}
\label{fig:a2319boot_old_vert}}\qquad
\subfloat[New deconvolution] %
{\includegraphics[width=0.35\textwidth]{{{a2319_bootstrap_new_reproj_vertical_avg10_log}}}
\label{fig:a2319boot_new_vert}}
\caption{Compton $ y $-parameter profiles of cluster A2319.
The solid black curve of Fig.~\ref{fig:a2319boot_old_vert} is superimposed as a dashed red line to the cuts in
Fig.~\ref{fig:a2319boot_new_vert} for comparison purposes (see text for details).}
\label{fig:boot_oldnew}
\end{figure}
The different characterisation of the central peak is also clear from Fig.~\ref{fig:a2319boot_new_vert}, where the
value recovered with the Van Cittert deconvolution is $ \approx 23 $ per cent lower with respect to the value obtained
with the new one based on wavelet coefficients. Such a difference is due to the joint effect of the larger
S/N provided by the absolute value of the wavelet kernel, and of the limited number of iterations (set to 3) used
in the Van Cittert deconvolution. The latter has been arbitrarily chosen as a trade-off between a reliable recovery of the
peak, and the avoidance of divergences. Therefore, it has a non-negligible impact on the final results; on the contrary,
the new deconvolution does not rely on any arbitrary parameter.
From the comparisons reported above, we can conclude that the new version of the algorithm does provide a more reliable
and stable reconstruction of the Comptonization parameter.
\subsection{The cases of A2029 and RXCJ1825}
Clusters A2029 and RXCJ1825 are two interesting targets for testing the algorithm.
Among the X-COP clusters, they have been detected by \textsl{Planck} with intermediate (19.3) and low (13.4) S/N,
respectively.
A2029 has been widely studied in X-rays \citep[see e.g.][]{lewis:a2029,clarke:a2029radiox,bourdin:Ximaging,walker:a2029};
on the contrary, RXCJ1825 has been poorly investigated in this band since its discovery \citep{ebeling:rxc1825discovery}.
Both these clusters may be interacting with two known neighbouring systems, as suggested in
\citet{planck:interactingclusters}. Nevertheless, given the low significance of the data, no further analysis on
the tSZ signal from possible connecting filaments has been explored in their work.
\subsubsection{A2029}
We show in Fig.~\ref{fig:a2029_XSZ} the contours from our map of the tSZ effect, superimposed to the X-ray
surface brightness of \object{A2029}.
We report the data from the \textsl{ROSAT} satellite, which allows the detection of the X-ray emission at larger radii
than those probed by \textsl{Chandra} or \textsl{XMM-Newton}, thanks to its low particle background
\citep{rosat:presentation,vikhlinin:rosatoutskirts}. The raw data have been denoised via the adaptive
smoothing technique~\citep{eckert:a2029}.
\begin{figure}
\centering
\subfloat[A2029 from \textsl{ROSAT}/PSPC] %
{\includegraphics[height=0.23\textheight]{a2029_x_szcontours_4scalesPLANCK2018_corrected} %
\label{fig:a2029_XSZ}} \quad
\subfloat[RXCJ1825 from \textsl{XMM-Newton}] %
{\includegraphics[height=0.23\textheight]{rxc1825_x_szcontours_4scalesPLANCK2018_corrected}
\label{fig:rxc1825_XSZ}}
\caption{X-ray and tSZ signals of clusters A2029 and RXCJ2815.
The maps show the vignetting-corrected and background-subtracted surface brightness in the X-ray energy band
0.5-2.5 keV. The superimposed contours represent the Compton $ y $-parameter from our algorithm.
The solid and dashed cyan circles on top of the images are drawn at $ R_{500} $ and $ R_{200} $, respectively.
The shaded circles in the bottom left corner of the maps represent a 5 arcmin beam.
\textbf{Fig.~\ref{fig:a2029_XSZ}.} Cluster A2029 mapped by \textsl{ROSAT}/PSPC.
The cyan cross indicates the position of the neighbour cluster A2033.
\textbf{Fig.~\ref{fig:rxc1825_XSZ}.} Cluster RXCJ1825 mapped by \textsl{XMM-Newton}.
The cyan cross marks the position of cluster CIZAJ1824, while the green arrow identifies a significant
elongation which may be associated to stripped gas from a nearby group of galaxies.}
\label{fig:XSZ}
\end{figure}
It can be seen that the tSZ signal in the central region matches fairly well with the X-ray surface brightness.
Thanks to the deconvolution, we are able to detect the tSZ emission corresponding to the neighbouring cluster
\object{A2033} with a significance of $ 8 \sigma_y $, which together with A2029 belongs to a small
supercluster \citep{einasto:superclusters}.
The X-ray and the tSZ peaks of A2033 show an offset which is, in any case, smaller than the best resolution of 5 arcmin
provided by \textsl{Planck}.
Our tSZ map clearly highlights an elongated projected structure that connects the two clusters.
As testified by the image showing $ y/\sigma_y $ in Fig.~\ref{fig:a2029_snr}, this elongated excess of signal is
significant to better than 5$ \sigma_y $.
X-ray images do also show such an elongated morphology in the peripheral regions to the north-east, pointing to A2033.
This suggested a possible ongoing merger between the two objects \citep[see e.g.][]{eckert:a2029,walker:a2029}.
However, recent analyses of the reconstructed density field through gravitational lensing,
indicate that this signal is likely due to the gas in the overlapping outskirts of the two clusters at $ R_{200} $,
rather than to a filament connecting them \citep{gonzalez:a2029interacting}.
Our tSZ imaging does represent an improvement with respect to the map shown in fig.~1 of \citet{planck:interactingclusters},
where the emission from A2033 is barely detected, as well as the signal between the two clusters.
However, we cannot favour any of the two hypothesised processes.
On the other hand, our maps may help in discriminating among possible scenarios proposed to explain the observed excess,
for instance in future works combining microwave and X-ray data to model the three-dimensional thermodynamic properties of
the ICM in each component of the system.
\begin{figure}[t]
\centering
\subfloat[A2029]{\includegraphics[height=0.23\textheight]{a2029_ysigmay}
\label{fig:a2029_snr}} \quad
\subfloat[RXCJ1825]{\includegraphics[height=0.23\textheight]{rxc1825_ysigmay}
\label{fig:rxc1825_snr}}
\caption{Maps of $ y/\sigma_y $ from a bootstrap run on clusters A2029 and RXCJ1825. The overlapped contours
represent the tSZ signal.}
\end{figure}
\subsubsection{RXCJ1825}
In Fig.~\ref{fig:rxc1825_XSZ}, we show the X-ray surface brightness of \object{RXCJ1825} from \textsl{XMM-Newton},
featuring wavelet denoising. The tSZ contours from our algorithm are superimposed to the image.
As in the case of A2029, the tSZ emission follows pretty well the X-ray surface brightness. In particular,
we notice that the peaks of the signal in the two bands coincide within 1 arcmin.
The tSZ signal shows a significant elongation towards the neighbouring cluster \object{CIZA J1824.1+3029}
(CIZAJ1824 hereafter), even if the peak corresponding to this object is not clearly visible.
A recent analysis based on the kinematics of their member galaxies, suggests that RXCJ1825 and CIZAJ1824 are in a
pre-merger state \citep{girardi:rxc1825}.
Thus, also in this case, our maps may shed a new light on the hypothesis of a possible interaction in future analyses.
Interestingly, our tSZ map shows a $ 14 \sigma_y $-significant elongation to the south-west of the cluster,
highlighted with a green arrow in Fig.~\ref{fig:rxc1825_XSZ}, which agrees with the excess in the X-ray surface brightness
detected with \textsl{XMM-Newton}.
Such an emission may be due to gas stripped from the past interaction between RXCJ1825 and a small, disrupted group of
galaxies, which have been detected at the same redshift in the optical band \citep{clavico:rxc1825}.
This scenario is also supported by the recent finding of a radio halo extending in the same direction of the X-ray
elongation \citep{botteon:rxc1825radio}.
Another feature we notice in the signal is the presence of a third structure located within $ R_{500} $,
that contributes at a level of $ y \approx \num{5e-5} $, and which is not detected in X-rays.
Nevertheless, the significance of this detection is 35 per cent lower with respect to the signal in the centre,
as demonstrated by the map in Fig.~\ref{fig:rxc1825_snr}, constructed from a bootstrap run. Therefore, its detection is
likely due to some localised irregularity in the instrumental noise in the raw HFI data.
In order to further quantify the significance of the signal reconstructed with our algorithm as a function of the
radius, we computed the profiles of the ratio $ \sigma_y/y $ that is the inverse of the effective S/N.
Specifically, we show in Fig.~\ref{fig:snr} the average of the vertical cuts passing through the centre of the
maps of $ \sigma_y/y $ within a 10 arcmin-wide band, for the clusters A2029 and RXCJ1825.
It can be seen that our algorithm provides a reconstruction of the signal with an effective S/N $ > 3 $
(marked with a dashed purple line in the plots), up to radii $ r \simeq 2R_{500}$ for both clusters.
\begin{figure}
\centering
\subfloat[A2029]{
\includegraphics[width=0.35\textwidth]{a2029_SNRCURVES}\label{fig:a2029snrcurve}}\qquad
\subfloat[RXCJ1825]{
\includegraphics[width=0.35\textwidth]{rxc1825_SNRCURVES}\label{fig:rxc1825snrcurve}}
\caption{Profiles of $ \sigma_y/y $ of clusters A2029 and RXCJ1825.
The purple dashed lines correspond to an effective S/N of 3 ($ \sigma_y/y = 1/3 $, see text for details).}
\label{fig:snr}
\end{figure}
|
\section{Introduction}
This paper studies a system of stochastic differential equations that purports to express the iron laws of dynamical equilibrium behavior in the broker call money market, a market that exists for the sake of funding stock brokers' margin loans to retail clients.
\par
We assume that the (aggregate) demand side of the market is comprised of continuous time Kelly (1956) gamblers. Kelly's seminal (1956) article takes up the problem of repeated bets on independent horse races for which the gambler has a better-quality estimate of the win probabilities than does the bookie, whose beliefs are implicit in the posted odds. The common-sense insight is that the (stationary) nature of the problem dictates that the gambler should always bet a fixed fraction of his wealth on this (favorable) opportunity; for obvious reasons, the \textit{Kelly fraction} is chosen so as to optimize the asymptotic continuously-compounded per-bet capital growth rate. Leo Breiman (1961) demonstrated the competitively superior properties of the Kelly Criterion: namely, that it asymptotically almost surely beats any ``essentially different'' stratgy by an exponential factor; and it has the shortest mean waiting time for hitting a distant wealth goal.
\par
In our problem, each horse race has been replaced by a differential tick $dt$ of the market clock, whereby the stock market index $S_t$ undergoes a fluctuation $dS_t$ that determines the gambler's profit-and-loss. Unlike betting on a horse race (where you should not bet $100\%$ of your wealth because you will eventually lose it all), it is highly advisable, given a low enough margin loan interest rate, to bet more than $100\%$ of your wealth on every little movement of the stock market. To fix ideas, let us assume that the S\&P 500 index multiplies itself at an expected (logarithmic) rate of $\nu:=9\%$ a year, with $\sigma:=15\%$ annual ($\log$-) volatility. If we were to imagine, in passing, that $\sigma$ tends to zero, then it becomes clear by continuity considerations that we should be falling all over ourselves in order to borrow money at a margin rate of, say, $5\%$.
\par
In the context of leveraged investment, Kelly's \textit{fixed fraction betting scheme} implies an ostensibly counter-intuitive trading mechanic. To illustrate, let us assume that we have $\$100$, and that we resolve to act so as to maintain a constant $2\times$ level of exposure to the S\&P 500 index. This means that we must try to always maintain a margin loan (debit) balance equal to the level of our account equity; the loan-to-value ratio must always be $50\%$. Now, assume that we wake up tomorrow and the stock market has gapped up $10\%$; our new account equity is $\underbrace{\$120}_{\text{Equity}}=\$\underbrace{220}_{\text{Assets}}-\underbrace{\$100}_{\text{Liabilities}}$. Whatever were our good reasons to lever the initial $\$100$ twice over, it seems natural that they should continue to apply to our newer, wealthier self. Note well that the \textit{market} has effectively chosen for us the new leverage ratio of $1.83:1$, with a corresponding loan-to-value ratio of $45.5\%$. Thus, our fixed-fraction betting scheme dictates that we must borrow and invest an additional $\$20$ (which is equal to the profits just earned). Although this behavior gives off the optics of a trend-following, performance-chasing, or market-timing scheme, the simple fact is that we are just going back to the well so as to carry on exploiting the opportunity to borrow at a low price. To be clear, the ``trend'' in question is the exponential growth of corporate earnings and dividends that is manifest in the high drift rate of the $\log$-price of the market index.
\par
The flip side of the coin is that the gambler must unwind this very process when the market goes down; for, suppose that on the next day, the index price gets divided by a factor of $1.1$ (for a loss of $9.1\%$), e.g. it returns full circle back to its original level. Our portfolio assets have thereby dwindled in value to $\$240\div1.1=\$218.18$ (against liabilities of $\$120$), for a net equity of $\$98.18$. Thus, the ``sideways'' motion of the asset price has caused us to underperform a buy-and-hold investor; we have been ``whipsawed'' by the cold arithmetic of buying high and selling low. On that score, our leverage ratio has just balooned to $2.22\times$, which makes for a loan-to-value ratio of $55\%$. In order to remedy this (overlevered) situation, our scheme dictates that we must liquidate $\$21.82$ of assets, e.g. the amount of money that we just lost in the fire.
\par
This simple example makes it abundantly clear just what is the fundamental trade-off that is faced by the continuous time Kelly gambler. On the one hand, we expect to earn the spread between the margin loan interest rate and the compound growth rate of the market index; on the other, we must deduct the ongoing costs of the whipsaw effect, which become more pronounced with higher levels of volatility in the underlying. The sweet spot that perfectly balances these two considerations (check with David Luenberger 1998) amounts to the magic leverage ratio $b^*:=0.5+(\nu-r_L)\div\sigma^2,$ where $r_L$ is the margin loan interest rate.
\par
If we assume that margin loans are supplied perfectly elastically (e.g. a horizontal supply curve), then the continuous time Kelly gambler has access to a permanent source of funding that allows him to beat the market asymptotically almost surely by an exponential factor (cf. with Garivaltis 2019b). This ``bucket shop''\footnote{A vivid expression that can be found in Merton (1992).} envrionment, with its unlimited supply of Saps willing to provide cash to their betters for a song, has obvious practical defects from a meta-perspective. Namely, on a long enough time horizon, the Kelly gamblers must inevitably own every single dollar of stock market capitalization.
\par
Thus, in order to get a realistic equilibrium outcome, we have decided in this paper to stand the supply curve on its head. The Saps, who are in possession of a giant pool of call money, are now assumed to supply it perfectly inelastically at the going rate (e.g. as determined by a vertical supply curve). Hapless though they are, they nonetheless manage to multiply their capital at an exponential rate; this paper assumes that all principal and interest payments are continuously reinvested in the money market.
\par
How is this natural and straightforward market structure going to shake itself out in the long run, given all the reverberatory effects of so many random vibrations in the asset price? These pages contain the answer.
\section{The Model}
We assume that the stock market index or ETF has $N$ shares outstanding, and its price per share $S_t$ evolves according to the geometric Brownian motion
\begin{equation}
dS_t:=S_t\times(\mu\,dt+\sigma\,dW_t),
\end{equation}where $S_0$ is the (given) initial price at time 0. Here, $\mu$ denotes the annual drift rate, $\sigma$ is the annual volatility, and $W_t$ is a standard Brownian motion. The log-price evolves according to (cf. Paul Wilmott 2001)
\begin{equation}
d(\log S_t)=\nu\,dt+\sigma\,dW_t,
\end{equation}where $\nu:=\mu-\sigma^2/2$ is the almost-sure asymptotic continuously-compounded capital growth rate
\begin{equation}
\nu=\lim_{t\to\infty}\frac{1}{t}\log\bigg(\frac{S_t}{S_0}\bigg)=\frac{\mathbb{E}[d(\log S_t)]}{dt}.
\end{equation}We assume that, at every instant $t$, there is a quantity $q_t$ of loan money (``broker call money'') that is supplied inelastically to the retail brokerage market. The money market charges a continuously-compounded interest rate of $r_L(t)$ per year for the duration of the differential time step $[t,t+dt]$, where $r_L(t)$ will be determined below in equilibrium. Naturally, we assume that the money market continuously reinvests all proceeds (both principal and interest), and so the size of the money market evolves according to
\begin{equation}
dq_t:=q_t\times r_L(t)\times\,dt,
\end{equation}or
\begin{equation}\label{intlaw}
d(\log q_t)=r_L(t)\times\,dt,
\end{equation}where $q_0$ is exogenously given. Thus, we have the relation
\begin{equation}\label{intquantity}
q_t=q_0\times\exp\bigg\{\int_0^tr_L(s)ds\bigg\}.
\end{equation}
\par
The demand side of the broker call money market is supposed to be constituted by a (representative) continuous time Kelly (1956) gambler, that ``bets'' the fraction $b_t\in[1,\infty)$ of his wealth on the stock market for the differential time step $[t,t+dt]$. In so doing, since $b_t\geq1$, he has borrowed the quantity $q_t:=(b_t-1)\times V_t$ from the money market; the loan must be repaid (both principal and interest) ``on call'' at time $t+dt$. Starting from a given initial value of $V_0$, the gambler's fortune $V_t$ evolves according to the stochastic differential equation
\begin{multline}
dV_t:=\underbrace{\frac{bV_t}{S_t}}_{\text{number of shares}}\times\underbrace{dS_t}_{\text{profit/loss per share}}-\underbrace{q_t\times r_L(t)\times dt}_{\text{interest paid } (=dq_t)}\\=V_t\times[\{b_t\mu+(1-b_t)r_L(t)\}dt+b_t\sigma\,dW_t].\\
\end{multline}Applying It\^{o}'s Lemma (cf. with Thomas Mikosch 1998) to the transformed process $V_t\mapsto\log V_t$, we see that the gambler's log-fortune evolves according to
\begin{equation}\label{vlaw}
\boxed{d(\log V_t)=\bigg\{b_t\mu+(1-b_t)r_L(t)-\frac{\sigma^2b_t^2}{2}\bigg\}dt+b_t\sigma\,dW_t.}
\end{equation}The gambler's expected continuously-compounded growth rate over $[t,t+dt]$ is equal to (cf. with David Luenberger 1998)
\begin{equation}\label{objective}
\boxed{\Gamma(b,r_L):=\text{Growth Rate}(b,r_L)=\frac{\mathbb{E}[d(\log V_t)]}{dt}=r_L+(\mu-r_L)b-\frac{\sigma^2}{2}b^2}.
\end{equation}The \textit{Kelly bet} (cf. with Edward O. Thorp\footnote{Who made a few such bets himself.} 2006, 2017) for the next tick of the market clock ($dt$) is, by definition, the fraction of wealth that maximizes the growth rate:
\begin{equation}\label{kellycriterion}
\boxed{b_t=b^*(r_L):=\argmax_{b\geq1}\,\,\Gamma(b,r_L)=\frac{\mu-r_L}{\sigma^2}=\frac{1}{2}+\frac{\nu-r_L}{\sigma^2}.}
\end{equation}Note well that the maximized (instantaneous expected continuously-compounded) growth rate is a stochastic process\footnote{The author is aware that this is an abuse of notation, albeit a very natural one that should cause no confusion.} $(\Gamma_t)_{t\geq0}$ that fluctuates according to the prevailing margin loan interest rate $r_L(t)$; substituting the Kelly bet (\ref{kellycriterion}) into the objective function (\ref{objective}), we get the expressions
\begin{multline}
\Gamma_t:=\max_{b\geq1}\,\,\Gamma(b,r_L(t))=\Gamma(b_t,r_L(t))\\=r_L(t)+\frac{1}{2}\bigg[\frac{\mu-r_L(t)}{\sigma}\bigg]^2=r_L(t)+\frac{\sigma^2b_t^2}{2}.\\
\end{multline}We will require the fact that the process $(\Gamma_t)_{t\geq0}$ is bounded:
\begin{equation}
\mu-\sigma^2/2\leq\Gamma_t\leq\frac{\mu^2}{2\sigma^2}.
\end{equation}The minorant $\mu-\sigma^2/2\equiv\Gamma(1,r_L(t))$ is the growth rate of an unlevered investor ($b:=1$) who just buys the market index, and holds. The majorant $\mu^2\div(2\sigma^2)$ is the Kelly growth rate that obtains when the margin loan interest rate is zero. To put it differently, we have
\begin{equation}
\frac{\partial}{\partial r_L}\bigg\{\max_{b\geq1}\,\Gamma(b,r_L(t))\bigg\}=1-\frac{\mu-r_L(t)}{\sigma^2}=1-b_t<0.
\end{equation}The instantaneous demand curve for margin loans is
\begin{equation}
q_t=\bigg(\frac{\mu}{\sigma^2}-1\bigg)V_t-\frac{V_t}{\sigma^2}\times r_L(t).
\end{equation}The corresponding instantaneous inverse demand curve is
\begin{equation}
r_L=\mu-\sigma^2\bigg(1+\frac{q_t}{V_t}\bigg)=\mu-\sigma^2\times b_t,
\end{equation}and the (price) elasticity of instantaneous demand for margin loans is
\begin{equation}
\epsilon^d(q_t):=-\frac{r_L(t)}{q_t}\times\frac{dq_t}{dr_L(t)}=\bigg(\frac{\mu}{\sigma^2}-1\bigg)\times\frac{V_t}{q_t}-1.
\end{equation}Since all $q_t$ dollars of call money are supplied inelastically by the money market, we have the vertical supply curve
\begin{equation}
\text{Quantity Supplied}:=
\begin{cases}
q_t&\text{ if }r_L(t)>0\\
[0,q_t]&\text{ if }r_L(t)=0.\\
\end{cases}
\end{equation}Intersecting supply and demand, we get the equilibrium interest rate
\begin{equation}\label{intersection}
\boxed{r_L(t)=\max\bigg(\mu-\sigma^2-\sigma^2\times\frac{q_t}{V_t},0\bigg)}=(\mu-\sigma^2[1+q_t/V_t])^+,
\end{equation}where $\mu-\sigma^2=\nu-\sigma^2/2$ is the \textit{choke price} of margin debt, and $x^+:=\max(x,0)$ denotes the positive part of the number $x$. On account of the equilibrium price (\ref{intersection}), we get the formula
\begin{equation}
\boxed{b_t=\min\bigg(1+\frac{q_t}{V_t},\frac{\mu}{\sigma^2}\bigg).}
\end{equation}Thus, our dynamical model amounts to the following three assumptions:
\begin{enumerate}[(I.)]
\item
All $q_t$ dollars of call money are supplied inelastically.
\item
The broker call money market continuously reinvests all its interest $dq_t$ and principal $q_t$.
\item
All margin loans are issued to continuous time Kelly gamblers; the loans pass through costlessly from the money market, with no additional markup from stock brokers.
\end{enumerate}
To help visualize this environment, Figure \ref{demand} plots the supply and demand curves for both $t:=0$ and $t:=10$ years later, along with the corresponding sample path $(q_t,r_L(t))_{0\leq t\leq10}$ in the price-quantity plane. The simulation (50,000 steps, $\Delta t:=1.75\text{ hours}$) used the parameter values $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=300px]{demand.png}
\caption{\sc Random vibrations of the supply and demand for margin loans over the course of a decade, as generated by the sample path $(q_t,r_L(t))_{0\leq t\leq10}$ in the price-quantity plane. The (50,000-step) simulation used the parameter values $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$.}
\label{demand}
\end{center}
\end{figure}
\begin{lemma}\label{lemma}
The size of the money market relative to Kelly gamblers' total equity has the following upper bound:
\begin{equation}\label{result}
\boxed{\frac{q_t}{V_t}\leq\frac{q_0}{V_0}\times\exp\bigg(-\frac{\sigma^2t}{2}-\sigma\int_{0}^tb_sdW_s\bigg).}
\end{equation}Thus, the margin loan interest rate is bounded from below by the expression:
\begin{equation}
r_L(t)\geq\mu-\sigma^2\bigg\{1+\frac{q_0}{V_0}\exp\bigg(-\frac{\sigma^2t}{2}-\sigma\int_{0}^tb_sdW_s\bigg)\bigg\}.
\end{equation}
\begin{proof}
First, we note the fact that $q_t\leq q_0 \times e^{(\mu-\sigma^2)t}$, e.g. the call money market can never compound its money any faster than the choke price $\mu-\sigma^2$. That is, looking at the interest rate expression (\ref{intersection}), we have the upper bound $r_L(t)\leq\mu-\sigma^2$; juxtaposing this inequality with the integral (\ref{intquantity}) yields $q_t\leq q_0\times\exp\{(\mu-\sigma^2)t\}$, as promised. On the other hand, the gambler's fortune is bounded below by the quantity
\begin{equation}\label{minorant}
V_t\geq\exp\bigg\{(\mu-\sigma^2/2)t+\sigma\int_{0}^tb_sdW_s\bigg\}.
\end{equation}For, looking at the differential equation (\ref{vlaw}), and bearing in mind that $\Gamma_t\geq\mu-\sigma^2/2$, we have
\begin{equation}\label{integrate}
d(\log V_t)=\Gamma_t\,dt+b_t\sigma\,dW_t\geq(\mu-\sigma^2/2)\,dt+b_t\sigma\,dW_t
\end{equation}
\begin{equation}
\implies\log(V_t/V_0)\geq(\mu-\sigma^2/2)t+\sigma\int_{0}^tb_sdW_s.
\end{equation}Combining the majorant $q_t\leq q_0\times e^{(\mu-\sigma^2)t}$ with the minorant (\ref{minorant}) for $V_t$, we get the stated result (\ref{result}).
\end{proof}
\end{lemma}
\par
In plain language, Lemma \ref{lemma} says that the locally expected growth rate of the gambler's fortune always exceeds the expected growth rate of the market index; on the other hand, the compound growth rate of the broker call money market (at most $\mu-\sigma^2$) is expected to be lower than that of the market index. To be sure, the actual dynamics of the relative market size $(q_t/V_t)_{t\geq0}$ is ultimately determined by the realized path $(W_t)_{t\geq0}$ of the Brownian motion that drives all uncertainty in the economy. But based on the expected difference in the exponential growth rates of $V_t$ and $q_t$, it is clear that after the elapse of many years (read: decades or centuries), the chances are high that the aggregate quantity of call money will be small in relation to the total bankrolls of continuous time Kelly gamblers.
\begin{theorem}
The size of the broker call money market relative to Kelly gamblers' total equity converges in probability\footnote{It emphatically does \textit{not} converge in mean square, as we will show below.} to zero:
\begin{equation}
\boxed{\plim_{t\to\infty}\frac{q_t}{V_t}=0.}
\end{equation}Thus, the leverage ratio of Kelly gamblers converges in probability to $1$ ($b_\infty:=\plim_{t\to\infty}b_t=1$) and the margin loan interest rate converges in probability\footnote{Since the processes $b_t$ and $r_L(t)$ are bounded ($1\leq b_t\leq\mu/\sigma^2$ and $0\leq r_L(t)\leq\mu-\sigma^2)$, they do indeed converge in mean square to $1$ and $\mu-\sigma^2$, respectively. The ratio $q_t/V_t$ has no such bounds; it may take any value in $(0,+\infty)$. } to the choke price:
\begin{equation}
\boxed{r_\infty:=\plim_{t\to\infty}r_L(t)=\mu-\sigma^2=\nu-\frac{\sigma^2}{2}.}
\end{equation}The growth rate process $\Gamma_t$ converges in probability to the buy-and-hold growth rate:
\begin{equation}
\boxed{\plim_{t\to\infty}\Gamma_t=\nu=\mu-\sigma^2/2.}
\end{equation}
\end{theorem}
\begin{proof}Since the Kelly bet $b_t=\min(1+q_t/V_t,\mu/\sigma^2)$ is a continuous function of the ratio $q_t/V_t$, and the interest rate $r_L(t)=\mu-\sigma^2b_t$ is in turn a continuous function of $b_t$, it suffices to show that $q_t/V_t$ converges in probability to 0, since probability limits are preserved by continuous transformations. There follows $\plim_{t\to\infty}\Gamma_t=\Gamma(1,\mu-\sigma^2)=\mu-\sigma^2/2.$
\par
Thus, let $\epsilon$ be any positive real number. Applying Lemma \ref{lemma}, we get the relations
\begin{equation}
1\geq \text{Prob}\bigg\{\frac{q_t}{V_t}\leq\epsilon\bigg\}\geq\text{Prob}\bigg\{\underbrace{-\frac{1}{t}\int_0^tb_sdW_s-\frac{\log(\epsilon V_0/q_0)}{\sigma t}}_{:=X_t}\leq\frac{\sigma}{2}\bigg\}\to1\text{ as }t\to\infty.\\
\end{equation}That is, note that the process $X_t:=-t^{-1}\big\{\int_0^tb_sdW_s+\log(\epsilon V_0/q_0)/\sigma\big\}$ converges to zero in mean square: we have
\begin{equation}
\lim_{t\to\infty}\mathbb{E}[X_t]=\lim_{t\to\infty}\frac{-\log(\epsilon V_0/q_0)}{\sigma t}=0
\end{equation}and, combining the It\^{o} isometry (cf. Tomas Bj\H ork 1998) with the bound $1\leq b_s\leq\mu/\sigma^2$, we get
\begin{equation}
\text{Var}[X_t]=\frac{1}{t^2}\int_0^t\mathbb{E}[b_s^2]ds\leq\frac{\mu^2}{\sigma^4t}\to0,
\end{equation}so that $\lim_{t\to\infty}\text{Var}[X_t]=0.$ Since the process $(X_t)_{t\geq0}$ converges to zero in mean square, it certainly converges to zero in probability; in particular, this means that
\begin{equation}
\lim_{t\to\infty}\text{Prob}\bigg\{X_t\leq\frac{\sigma}{2}\bigg\}=1.
\end{equation}By the squeeze theorem, then, we have obtained the desired result: for all $\epsilon>0$,
\begin{equation}
\lim_{t\to\infty}\text{Prob}\bigg\{\frac{q_t}{V_t}\leq\epsilon\bigg\}=1.
\end{equation}
\end{proof}
\begin{corollary}\label{convergence}The margin loan interest rate $r_L(t)$ converges in mean square to the choke price $r_\infty:=\mu-\sigma^2=\nu-\sigma^2/2$ and the Kelly bet $b_t$ converges in mean square to $1$. The instantaneous Kelly growth rate $\Gamma_t$ converges in mean square to the buy-and-hold growth rate $\nu=\mu-\sigma^2/2$.
\end{corollary}
\begin{proof}
It suffices to show that $b_t$ converges in mean square to 1; then, on account of the linear relationship $r_L(t)=\mu-\sigma^2b_t$, we will have $\lim_{t\to\infty}\mathbb{E}[r_L(t)]=\mu-\sigma^2$ and $\lim_{t\to\infty}\text{Var}[r_L(t)]=0$.
\par
To this end, let $\epsilon$ be any positive number, and let $R_t:=q_t/V_t$ denote the relative size of the call money market. We have
\begin{multline}
\mathbb{E}[(b_t-1)^2]=\text{Prob}\{R_t<\epsilon\}\mathbb{E}[(b_t-1)^2|R_t<\epsilon]+\text{Prob}\{R_t\geq\epsilon\}\mathbb{E}[(b_t-1)^2|R_t\geq\epsilon]\\
\leq1\times\epsilon^2+\text{Prob}\{R_t\geq\epsilon\}\times(\mu/\sigma^2-1)^2.\\
\end{multline}Using the fact that $\lim_{t\to\infty}\text{Prob}\{R_t\geq\epsilon\}=0$, we see that the following relation must obtain for every $\epsilon>0$:
\begin{equation}
\limsup_{t\to\infty}\mathbb{E}[(b_t-1)^2]\leq\epsilon^2.
\end{equation}Since the $\limsup_{t\to\infty}$ of the mean-squared error is smaller than every positive number, we get the inequalities
\begin{equation}
\liminf_{t\to\infty}\mathbb{E}[(b_t-1)^2]\leq\limsup_{t\to\infty}\mathbb{E}[(b_t-1)^2]=0\leq\liminf_{t\to\infty}\mathbb{E}[(b_t-1)^2],
\end{equation}which implies that $\lim_{t\to\infty}\mathbb{E}[(b_t-1)^2]=0$.
\par
Finally, turning our attention to the instantaneous Kelly growth rate $\Gamma_t=r_L(t)+\sigma^2b_t^2/2$, it now suffices to show that $b_t^2$ converges to 1 in mean square; then, since $r_L(t)\xrightarrow{m.s.}\mu-\sigma^2$, we will have $\Gamma_t\xrightarrow{m.s.}\mu-\sigma^2+\sigma^2/2\times1=\mu-\sigma^2/2$, as promised. Accordingly, we bound the mean-squared error
\begin{equation}
\mathbb{E}[(b_t^2-1)^2]=\mathbb{E}[(b_t-1)^2(b_t+1)^2]\leq\mathbb{E}[(b_t-1)^2]\times\bigg(\frac{\mu}{\sigma^2}+1\bigg)^2\to0,
\end{equation}which proves that $b_t^2\xrightarrow{m.s.} 1.$
\end{proof}
In plain English: as time goes on, there are some (exceedingly rare) sample paths of the experiment whereby the Kelly gambler performs very poorly in relation to the money market; the ratio $q_t/V_t$ therefore spikes and the Kelly bet hits the upper bound $b_t=\mu/\sigma^2$ under an interest rate of zero. However, these rare events make a negligible contribution to the mean-squared error $\mathbb{E}[(b_t-1)^2]$, precisely because the Kelly gambler's mantra prevents him from betting more than $\overline{b}:=\mu/\sigma^2$, even when he is offered an interest rate of zero. After many years $t$ have elapsed, the density of $b_t$ becomes concentrated around $1$, albeit with a long tail that spans the interval $[1,\mu/\sigma^2]$.
\par
Figure \ref{mcarlo} illustrates the corollary by plotting a 100-year, 50,000-step sample path of the Kelly leverage ratios and margin loan interest rates in an economy generated by the parameters $(q_0,V_0,\nu,\sigma,\mu,r_\infty):=(1,1,0.09,0.15,0.1012,0.0787)$. For context, the Figure provides Monte Carlo estimates of the expected values $\{\mathbb{E}[b_t],\mathbb{E}[r_L(t)]\}_{0\leq t\leq100}$ and standard deviations $\{\text{Std}(b_t),\text{Std}(r_L(t))\}_{0\leq t\leq100}$ that were generated from $50,000$ simulations of $50,000$ steps each ($\Delta t:=17.5\text{ hours}$). In the same vein, Figure \ref{gamma} gives a 30-year sample path of the instantaneous growth rate process $(\Gamma_t)$ for the same parameters, along with Monte Carlo estimates of the functions $t\mapsto\mathbb{E}[\Gamma_t]$ and $t\mapsto\text{Std}(\Gamma_t)$ that were computed from $40,000$ simulations ($\Delta t:=6.6\text{ hours}$).
\begin{figure}[t]
\begin{center}
\includegraphics[width=410px]{mcarlo.png}
\caption{\sc 100-year sample path of the Kelly leverage ratios $(b_t)$ and corresponding margin loan interest rates ($r_L(t)$), for the parameters $(q_0,V_0,\nu,\sigma,\mu,r_\infty):=(1,1,0.09,0.15,0.1012,0.0787).$ The means and standard deviations were estimated from $50,000$ Monte Carlo simulations of $50,000$ steps each ($\Delta t:=17.5\text{ hours}$).}
\label{mcarlo}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=325px]{gamma.png}
\caption{\sc 30-year sample path of the optimum growth rate process $(\Gamma_t)$ for the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$, along with Monte Carlo estimates ($40,000$ simulations, $\Delta t:=6.6\text{ hours}$) of the functions $t\mapsto\mathbb{E}[\Gamma_t]$ and $t\mapsto\std(\Gamma_t)$.}
\label{gamma}
\end{center}
\end{figure}
\begin{theorem}
The relative size $q_t/V_t$ of the money market is a martingale (in spite of the fact that it converges in probability to zero); the Kelly bet $b_t$ is a supermartingale (e.g. it is always expected to decrease) and the margin loan interest rate $r_L(t)$ is a submartingale (e.g. it is always expected to increase).
\end{theorem}
\begin{proof}
We apply the quotient rule of the It\^{o} calculus (cf. with Ovidiu Calin 2015) to the ratio $q_t/V_t$:
\begin{equation}
d\bigg(\frac{q_t}{V_t}\bigg)=\frac{dq_tV_t-q_tdV_t-dq_tdV_t}{V_t^2}+\frac{q_t}{V_t^3}\times(dV_t)^2.
\end{equation}According to the It\^o multiplication table (e.g. Paul Wilmott 1998), we have $dq_t\times dV_t=0$ and $(dV_t/V_t)^2=b_t^2\sigma^2\times dt$. Thus, one calculates that
\begin{multline}
d\bigg(\frac{q_t}{V_t}\bigg)=\frac{q_t}{V_t}\times\bigg[(r_L+b_t^2\sigma^2)dt-\frac{dV_t}{V_t}\bigg]\\
=\frac{q_t}{V_t}\times\bigg[b_t(\underbrace{\sigma^2b_t-\mu+r_L(t)}_{=0})dt-b_t\sigma\,dW_t\bigg]=-\sigma\frac{q_tb_t}{V_t}\,dW_t.\\
\end{multline}Thus, the stochastic process $(q_t/V_t)_{t\geq0}$ is a martingale, since it has zero drift and admits the (It\^o) integral representation
\begin{equation}\label{representation}
\boxed{\frac{q_t}{V_t}=\frac{q_0}{V_0}-\sigma\int_0^t\frac{q_sb_s}{V_s}dW_s.}
\end{equation}On account of the fact that $b_t=\min(1+q_t/V_t,\mu/\sigma^2)$ is a concave function of the martingale $q_t/V_t$, we conclude that $(b_t)_{t\geq0}$ is a supermartingale, e.g. it is always expected to decrease (cf. with Lawrence Evans 2010). Likewise, the interest rate $r_L(t)=\max(\mu-\sigma^2-\sigma^2q_t/V_t,0)$ is a submartingale, since it is a convex function of a martingale.
\end{proof}
Thus, although the chances are high that the ratio $R_t:=q_t/V_t$ is very low in the long run, it nevertheless has a constant mean $\mathbb{E}[R_t]\equiv q_0/V_0$; this happens on account of a few sample paths for which the stock market dramatically underperforms the broker call money market. The (unconditionally) expected interest rate $\mathbb{E}[r_L(t)]$ is an increasing function of time that converges to $\mu-\sigma^2$; conditional on the current state of things at time $t$, the expected margin rate $\mathbb{E}[r_L(t+\Delta t)|r_L(t)]$ at any time in the future is greater than or equal to the current observation $r_L(t)$. However, the expected increases in the interest rate (and attendant decreases in the aggregate leverage ratio) are disturbed by so many random vibrations of the stock market. The margin loan interest rate responds pro-cyclically to random noise in the financial markets; the leverage ratios of continuous time Kelly gamblers respond counter-cyclically. But the underlying signal (that is, the exponential growth of asset prices) suffices to generate a permanent uptrend in margin loan interest rates.
\begin{corollary}
The probability of the margin loan interest rate ever hitting zero (between now and kingdom come) has the following majorant:
\begin{equation}
\boxed{\text{\sc Prob}\{r_L(t)\text{\sc \,is ever }0\}\leq1-\frac{r_L(0)}{r_\infty}=1-\frac{\text{\sc Current Interest Rate}}{\text{\sc Choke Price}}}\\
\end{equation}
\end{corollary}
\begin{proof}
The condition that the margin loan interest rate $r_L(t)$ hits zero at least once over a given horizon $[0,T]$ is equivalent to the condition that the ratio $q_t/V_t$ breaches $\mu/\sigma^2-1$ at least once. Since $(q_t/V_t)_{t\geq0}$ is a positive martingale, Doob's martingale inequality obtains (cf. Lawrence Evans 2010); in our context, this inequality amounts to
\begin{equation}\label{inequality}
\text{Prob}\bigg\{\max_{0\leq t\leq T}\frac{q_t}{V_t}\geq\frac{\mu}{\sigma^2}-1\bigg\}\leq\frac{\mathbb{E}[q_T/V_T]}{\mu/\sigma^2-1}=\frac{q_0}{V_0}\times\frac{\sigma^2}{\mu-\sigma^2}=1-\frac{r_L(0)}{r_\infty},
\end{equation}where we have used the fact that $\mathbb{E}[q_T/V_T]\equiv q_0/V_0$. Taking the limit of the inequality (\ref{inequality}) as $T\to\infty$, we obtain the desired result, that
\begin{equation}
\text{Prob}\bigg\{\sup_{t\geq0}\frac{q_t}{V_t}\geq\frac{\mu}{\sigma^2}-1\bigg\}\leq1-\frac{r_L(0)}{r_\infty}.
\end{equation}
\end{proof}Thus, if the current margin loan interest rate amounts to $70\%$ of the choke price, then the chance of it ever hitting zero is at most $30\%$. If the current rate is $20\%$ of the asymptotic interest rate, then the chance of it ever reaching zero is at most $80\%$, etc. Table \ref{doobtable} illustrates the majorant for different stock market volatilities and compound-annual (logarithmic) growth rates, assuming that the money market begins on par with the gambler's fortune ($q_0/V_0:=1$). Naturally, the bound becomes tighter as the stock market parameters become more favorable (higher $\nu$, lower $\sigma$); it also tightens with the relative scarcity of loanable funds (lower $q_0/V_0$).
\begin{example}
As of this writing, the broker call money rate (as reported by Bankrate.com) is $4.25\%$. Assuming the stylized parameters $(\nu,\sigma):=(0.09,0.15)$ for the S\&P 500 index, we get a choke price of $7.9\%$. Thus, we reckon that the chance of the margin loan interest rate ever hitting zero is at most $4.25\div7.9=54\%$.
\end{example}
\begin{proposition}
We have the following bounds on the (unconditional) standard deviation of the relative market size $q_t/V_t$:
\begin{equation}
\boxed{\frac{q_0}{V_0}\times\sqrt{\exp[\sigma^2t]-1}\leq\std\bigg(\frac{q_t}{V_t}\bigg)\leq\frac{q_0}{V_0}\times\sqrt{\exp[(\mu/\sigma)^2t]-1}.}
\end{equation}In particular, $\lim_{t\to\infty}\std(q_t/V_t)=+\infty$.
\end{proposition}
\begin{proof}
For notational convenience, we let $F(t):=\mathbb{E}[(q_t/V_t)^2]$ denote the second moment of the relative size process. Recalling the (It\^o) integral representation
\begin{equation}
\frac{q_t}{V_t}=\frac{q_0}{V_0}-\sigma\int_0^t\frac{q_sb_s}{V_s}dW_s,
\end{equation}the It\^o isometry implies that
\begin{equation}
\text{Var}\bigg[\frac{q_t}{V_t}\bigg]=F(t)-(q_0/V_0)^2=\sigma^2\int_0^t\mathbb{E}\bigg[\bigg(\frac{q_sb_s}{V_s}\bigg)^2\bigg]ds,
\end{equation}which, upon differentiating, gives us
\begin{equation}
\frac{dF}{dt}=\sigma^2\mathbb{E}\bigg[\bigg(\frac{q_tb_t}{V_t}\bigg)^2\bigg].
\end{equation}Now, bearing in mind that $1\leq b_t^2\leq\mu^2/\sigma^4$,
we have the inequalities
\begin{equation}
\sigma^2F(t)\leq\frac{dF}{dt}\leq\bigg(\frac{\mu}{\sigma}\bigg)^2F(t),
\end{equation}or equivalently,
\begin{equation}\label{bothsides}
\sigma^2\leq\frac{d}{dt}\log F(t)\leq\bigg(\frac{\mu}{\sigma}\bigg)^2.
\end{equation}Integrating the inequalities (\ref{bothsides}) and simplifying, we obtain the theoretical bounds
\begin{equation}\label{root}
F(0)\times\{\exp[\sigma^2t]-1\}\leq F(t)-F(0)\leq F(0)\times\{\exp[(\mu/\sigma)^2t]-1\}.
\end{equation}Remembering that $F(0)=(q_0/V_0)^2$ and $\text{Var}[q_t/V_t]=F(t)-F(0)$, taking the square root of (\ref{root}) yields the stated result.
\end{proof}Thus, although the martingale $(q_t/V_t)_{t\geq0}$ converges in probability to zero, its standard deviation grows to infinity at a geometric rate. Figure \ref{stdev} plots these theoretical bounds, along with Monte Carlo estimates of the true standard deviation, for $t\in[0,2]$ assuming the parameters $q_0:=1, V_0:=1, \nu:=0.09, \sigma:=0.15$, and $\mu:=\nu+\sigma^2/2=0.1012$. The (deterministic) function $t\mapsto\text{Std}(q_t/V_t)$ was estimated from $100,000$ experiments of $100,000$ steps each; the corresponding step size was $\Delta t:=10.5\text{ minutes}$. Note that the population standard deviation $\sqrt{\mathbb{E}[(q_t/V_t)^2]-(q_0/V_0)^2}$ is required to be increasing on account of the fact that the process $(q_t/V_t)^2$ is a submartingale (e.g. it is a convex function of $q_t/V_t$). For the sake of visualization, Figure \ref{relative} plots a 100-year sample path of $q_t/V_t$ for the same deep parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$; the experiment consisted of $100,000\text{ steps}$, for a step size of $8.8\text{ hours}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=325px]{stdev.png}
\caption{\sc Monte Carlo estimates of $\std(q_t/V_t)$ over $t\in[0,2]$, assuming the parameters $q_0:=1, V_0:=1, \nu:=0.09, \sigma:=0.15$, and $\mu:=\nu+\sigma^2/2=0.1012$. Estimates computed from the simulation of $100,000$ sample paths of $100,000$ steps each ($\Delta t:=10.5 \text{ minutes}$).}
\label{stdev}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=325px]{relative.png}
\caption{\sc 100-year sample path of $q_t/V_t$, generated by the parameters $q_0:=1, V_0:=1, \nu:=0.09, \sigma:=0.15$, and $\mu:=\nu+\sigma^2/2=0.1012$. 100,000 steps, $\text{step size}=8.8\text{ hours}$. Crossing the blue (dashed) barrier results in a margin loan interest rate of zero.}
\label{relative}
\end{center}
\end{figure}
\begin{table}
\begin{tabular}{llllll}
\textbf{\small Ratio} & \textbf{\small Vol. }&\textbf{\small CAGR}&\textbf{\small Current/Choke}&\textbf{\small Majorant}&\textbf{Actual Prob.}\\
($q_0\div V_0$)&($\sigma$)&($\nu$)&($r_L(0)/r_\infty$)&\small of $\text{P}\{r_L\text{ is ever }0\}$& \small (Monte Carlo est.*)\\
\hline\hline\\
$1$ & 10\% & 9\% & 7.5\%/8.5\% & $\leq$11.8\% & $\underset{(0.18\%)}{9.3\%}$ \\
$1$ & 15\% & 9\% & 5.6\%/7.9\% & $\leq$28.6\% & $\underset{(0.28\%)}{26.3\%}$ \\
$1$ & 20\% & 9\% & 3\%/7\% & $\leq$57.1\% & $\underset{(0.31\%)}{55.1\%}$ \\
\hline\\
$1$ & 10\% & 8\% & 6.5\%/7.5\% & $\leq$13.3\% & $\underset{(0.2\%)}{10.8\%}$ \\
$1$ & 15\% & 8\% & 4.6\%/6.9\% & $\leq$32.8\% & $\underset{(0.29\%)}{30.1\%}$ \\
$1$ & 20\% & 8\% & 2\%/6\% & $\leq$66.7\% & $\underset{(0.3\%)}{63.8\%}$ \\
\hline\\
$1$ & 10\% & 7\% & 5.5\%/6.5\% & $\leq$15.4\% & $\underset{(0.21\%)}{12.7\%}$ \\
$1$ & 15\% & 7\% & 3.6\%/5.9\% & $\leq$38.3\% & $\underset{(0.3\%)}{36.1\%}$ \\
$1$ & 20\% & 7\% & 1\%/5\% & $\leq$80\% & $\underset{(0.26\%)}{78\%}$ \\
\hline\\
$1$ & 10\% & 6\% & 4.5\%/5.5\% & $\leq$18.2\% & $\underset{(0.23\%)}{15.2\%}$ \\
$1$ & 15\% & 6\% & 2.6\%/4.9\% & $\leq$46.2\% & $\underset{(0.31\%)}{43.6\%}$ \\
$1$ & 20\% & 6\% & 0\%/4\% & $\leq$100\% & 100\%\\
\hline\\
$1$ & 10\% & 5\% & 3.5\%/4.5\% & $\leq$22.2\% & $\underset{(0.25\%)}{18.7\%}$ \\
$1$ & 15\% & 5\% & 1.6\%/3.9\% & $\leq$58.1\% & $\underset{(0.31\%)}{56\%}$ \\
$1$ & 20\% & 5\% & 0\%/3\% & $\leq$100\% & 100\%\\
\hline
\end{tabular}
\caption{Upper bounds on the probability of the margin loan interest rate ever hitting zero over $t\in[0,+\infty)$, for different stock market volatilities and growth rates, assuming that the money market starts on par with the gambler's fortune ($q_0/V_0:=1$).\\\\
*Percentage of all simulations for which the margin loan interest rate hit the zero bound (standard errors in parentheses). $25,000$ simulations per estimate, spanning 200 years each, $25,000$ steps per simulation, $\Delta t:=2.92\text{ days}$.\\}
\label{doobtable}
\end{table}
\begin{theorem}
The Kelly gambler's realized continuously-compounded capital growth rate over $[0,T]$ (namely, $\log(V_T/V_0)/T$) converges in mean square to the stock market growth rate $\nu=\mu-\sigma^2/2$; the realized continuously-compounded growth rate of the money market over $[0,T]$ (namely, $\log(q_T/q_0)/T$) converges in mean square to the choke price $r_\infty=\nu-\sigma^2/2=\mu-\sigma^2$.
\end{theorem}
\begin{proof}
On account of the expresstion
\begin{equation}
\frac{\log(q_T/q_0)}{T}=\frac{1}{T}\int_0^Tr_L(t)dt,
\end{equation}it follows that
\begin{equation}
\mathbb{E}\bigg[\frac{\log(q_T/q_0)}{T}\bigg]=\frac{1}{T}\int_0^T\mathbb{E}[r_L(t)]dt
\end{equation}is the average value of the (deterministic) function $t\mapsto\mathbb{E}[r_L(t)]$ over the interval $[0,T]$. Since $r_L(t)$ converges in mean square to $r_\infty$, we have the relation $\lim_{t\to\infty}\mathbb{E}[r_L(t)]=r_\infty$; thus, the average value of the function $t\mapsto\mathbb{E}[r_L(t)]$ must also converge to $r_\infty$. It remains to show that
\begin{equation}
\lim_{T\to\infty}\text{Var}\bigg[\frac{\log(q_T/q_0)}{T}\bigg]=0.
\end{equation}To this end, we invoke the formula (cf. with Hoel, Port, and Stone 1972)
\begin{equation}\label{doubleintegral}
\text{Var}\bigg[\frac{\log(q_T/q_0)}{T}\bigg]=\frac{1}{T^2}\int_0^T\int_0^T\text{Cov}(r_L(s),r_L(t))dsdt.
\end{equation}The Cauchy-Schwarz inequality (e.g. check with T.T. Soong 1973) says that
\begin{equation}\label{rhs}
\text{Cov}(r_L(s),r_L(t))\leq\text{Std}(r_L(s))\times\text{Std}(r_L(t)).
\end{equation}Hence, since the right-hand-side of (\ref{rhs}) is multiplicatively separable in the variables $s$ and $t$, the double integral (\ref{doubleintegral}) is majorized by the square of the unidimensional integral $\int_0^T\text{Std}(r_L(t))dt$; this gives us the variance bound
\begin{equation}\label{ub}
\text{Var}\bigg[\frac{\log(q_T/q_0)}{T}\bigg]\leq\bigg[\frac{1}{T}\int_0^T\text{Std}(r_L(t))dt\bigg]^2\to0\text{ as }T\to\infty.
\end{equation}The right-hand-side of (\ref{ub}) converges to zero as $T\to\infty$ because it is the average value of the (deterministic) function $t\mapsto\text{Std}(r_L(t))$, which itself converges to zero on account of the fact that $r_L(t)$ converges in mean square to $r_\infty$. This proves that the realized money market growth rate $\log(q_T/q_0)/T$ converges in mean square to the choke price $r_\infty=\mu-\sigma^2$.
\par
Turning our attention to the realized compound-growth rate of the Kelly bankroll over $0\leq t\leq T$, we integrate the left-hand-side of (\ref{integrate}) and obtain the expression
\begin{equation}
\frac{\log(V_T/V_0)}{T}=\underbrace{\frac{1}{T}\int_0^T\Gamma_t\,dt}_{=:x_T}+\underbrace{\frac{\sigma}{T}\int_0^Tb_tdW_t}_{=:y_T}=:x_T+y_T.
\end{equation}Bearing in mind that $\mathbb{E}[y_T]\equiv0$, we get
\begin{equation}\label{sampleavg}
\mathbb{E}\bigg[\frac{\log(V_T/V_0)}{T}\bigg]=\mathbb{E}[x_T]=\frac{1}{T}\int_0^T\mathbb{E}[\Gamma_t]dt,
\end{equation}which is the average value of the deterministic function $t\mapsto\mathbb{E}[\Gamma_t]$ over the interval $[0,T]$. Since the stochastic process $(\Gamma_t)_{t\geq0}$ converges in mean square to $\mu-\sigma^2/2$, we of course have $\lim_{t\to\infty}\mathbb{E}[\Gamma_t]=\mu-\sigma^2/2$; accordingly, the average value (\ref{sampleavg}) of the function $t\mapsto\mathbb{E}[\Gamma_t]$ must also converge to $\mu-\sigma^2/2$ as $T\to\infty$.
\par
To complete the proof, we proceed to demonstrate that that $\lim_{T\to\infty}\text{Var}[x_T+y_T]=0.$ On account of the triangle inequality $\text{Std}(x_T+y_T)\leq\text{Std}(x_T)+\text{Std}(y_T)$, it suffices to show that $\lim_{T\to\infty}\text{Var}[x_T]=0$ and $\lim_{T\to\infty}\text{Var}[y_T]=0$. We are already familiar with the fact that $y_T$ converges to zero in mean square; \textit{mutatis mutandis}, analogous to what we just did with the interest rate $r_L(t)$, we write
\begin{multline}\label{mutatis}
\text{Var}\bigg[\frac{1}{T}\int_0^T\Gamma_tdt\bigg]=\frac{1}{T^2}\int_0^T\int_0^T\text{Cov}(\Gamma_s,\Gamma_t)dsdt\\
\leq\frac{1}{T^2}\bigg[\int_0^T\text{Std}(\Gamma_s)ds\bigg]\bigg[\int_0^T\text{Std}(\Gamma_t)dt\bigg]\\
=\bigg[\frac{1}{T}\int_0^T\text{Std}(\Gamma_t)dt\bigg]^2\to0.\\
\end{multline}The last bracketed expression in (\ref{mutatis}) converges to 0 as $T\to\infty$ because it is the average value of the function $t\mapsto\text{Std}(\Gamma_t)$ over the interval $[0,T]$, a function whose value itself converges to 0 as $t\to\infty$.
\end{proof}To illustrate the Theorem, Figure \ref{growthrate} plots the realized growth rate series
\begin{equation}
\frac{1}{t}(\log(q_t/q_0),\log(S_t/S_0),\log(V_t/V_0))
\end{equation}that obtained from a 200-year, 200,000-step simulation of the model economy generated by the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=325px]{growthrate.png}
\caption{\sc The realized continuously-compounded capital growth rates in a 200-year, 200,000-step simulation of the model economy, under the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$.}
\label{growthrate}
\end{center}
\end{figure}
\begin{theorem}[Change of Num\'eraire]\label{change}
The ratio $V_t/S_t$ of the gambler's fortune to the price of one unit of the market index (e.g. the value of the bankroll as measured in shares of the ETF) is a submartingale (always expected to increase). The total size $q_t/S_t$ of the money market, as expressed in units of this num\'eraire, is a supermartingale that converges in probability to zero. The aggregate wealth in the model $(q_t+V_t)/S_t$ (money market plus gambler's equity) is a supermartingale when expressed in shares of the market index. Consequently, the ratio
\begin{equation}
\text{Kelly Gambler's Relative Growth Factor}:=\frac{V_t/V_0}{S_t/S_0}
\end{equation}has the property that
\begin{equation}
\boxed{\mathbb{E}\bigg[\frac{V_t/V_0}{S_t/S_0}\bigg]\leq1+\frac{q_0}{V_0}\text{ for all }t.}
\end{equation}
\end{theorem}The relative growth factor $(V_t/V_0)\div(S_t/S_0)$ amounts to the ratio of the Kelly gambler's bankroll to the wealth of a buy-and-hold investor ($b\equiv1$) who started with the same initial capital. Note that, although the cumulative outperformance realized by the Kelly gambler over $[0,t]$ is always expected to increase, it fails to grow to infinity at an exponential rate (as it would under perfectly elastic supply of margin loans). Rather, the asymptotic relative growth factor is a finite, random quantity that may even turn out to be less than 1 (albeit with low probability). At the start of the model, the Kelly gambler cannot expect to ever achieve more than $1+q_0/V_0$ times the wealth of a buy-and-hold investor who started with the same amount of money. Say, if the initial interest rate is positive and the initial leverage ratio is $b_0:=2$ (the maximum allowed by U.S. Regulation-T), then we cannot expect to achieve more than double the final wealth of an equivalent buy-and-hold investor.
\begin{proof}
Applying the It\^o quotient rule to the process $(V_t/S_t)_{t\geq0}$, one calculates that
\begin{equation}\label{V/S}
d\bigg(\frac{V_t}{S_t}\bigg)=\frac{V_t}{S_t}\times\bigg\{(b_t-1)(\mu-\sigma^2-r_L(t))dt+\sigma(b_t-1)dW_t\bigg\}.
\end{equation}Thus, $(V_t/S_t)_{t\geq0}$ is a submartingale because of its positive drift, which obtains on account of the fact that $b_t>1$ and $r_L<\mu-\sigma^2$. A similar calculation shows that
\begin{equation}\label{q/S}
d\bigg(\frac{q_t}{S_t}\bigg)=\frac{q_t}{S_t}\times\bigg\{-(\mu-\sigma^2-r_L(t))dt-\sigma dW_t\bigg\},
\end{equation}whence $(q_t/S_t)_{t\geq0}$ is a supermartingale because of its negative drift rate. Combining equations (\ref{V/S}) and (\ref{q/S}), and simplifying, we obtain
\begin{equation}
d\bigg(\frac{q_t+V_t}{S_t}\bigg)=d\bigg(\frac{q_t}{S_t}\bigg)+d\bigg(\frac{V_t}{S_t}\bigg)=\frac{(b_t-1)V_t-q_t}{S_t}\bigg\{(\mu-\sigma^2-r_L(t))dt+\sigma dW_t\bigg\}.
\end{equation}Recalling that $b_t=\min(1+q_t/V_t,\mu/\sigma^2)\leq1+q_t/V_t$, we see that $(q_t+V_t)/S_t$ is a supermartingale, since its drift is $\leq0$. With these facts in hand, we have the inequalities
\begin{equation}\label{relativeperf}
\mathbb{E}\bigg[\frac{V_t}{S_t}\bigg]\leq\mathbb{E}\bigg[\frac{q_t+V_t}{S_t}\bigg]\leq\frac{q_0+V_0}{S_0},
\end{equation}where we have used the fact that $\mathbb{E}[(q_t+V_t)/S_t]$ is a decreasing function of time. Multiplying (\ref{relativeperf}) through by $S_0/V_0$, we get the promised result:
\begin{equation}
\mathbb{E}\bigg[\frac{V_t/V_0}{S_t/S_0}\bigg]\leq1+\frac{q_0}{V_0}.
\end{equation}Finally, for the sake of demonstrating that $\plim_{t\to\infty}q_t/S_t=0$, we start with the upper bound
\begin{equation}
\frac{q_t}{S_t}\leq\frac{q_0}{S_0}\times\exp\bigg(-\frac{\sigma^2t}{2}+\sigma W_t\bigg);
\end{equation}If $\epsilon$ is any positive number, then
\begin{multline}
\text{Prob}\bigg\{\frac{q_t}{S_t}\leq\epsilon\bigg\}\geq\text{Prob}\bigg\{\frac{q_0}{S_0}\exp\bigg(-\frac{\sigma^2t}{2}+\sigma W_t\bigg)\leq\epsilon\bigg\}\\
=\text{Prob}\bigg\{\frac{W_t}{\sqrt{t}}\leq\frac{\log(\epsilon S_0/q_0)}{\sigma\sqrt{t}}+\frac{\sigma\sqrt{t}}{2}\bigg\}=N\bigg(\frac{\log(\epsilon S_0/q_0)}{\sigma\sqrt{t}}+\frac{\sigma\sqrt{t}}{2}\bigg),\\
\end{multline}where $N(\bullet)$ denotes the cumulative normal distribution function. Thus, we have
\begin{equation}
1\geq\lim_{t\to\infty}\text{Prob}\bigg\{\frac{q_t}{S_t}\leq\epsilon\bigg\}\geq N(\infty)=1,
\end{equation}which is the desired result.
\end{proof}Figure \ref{numeraire} supplements Theorem \ref{change} by plotting a 100-year ($100,000$-step) sample path of the time series $V_t/S_t$, $q_t/S_t$, $(q_t+V_t)/S_t$, and $r_L(t)$ for the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.08,0.2)$. For this particular simulation ($\Delta t:=8.8\text{ hours}$), we made the stock market index less favorable than it was in our previous experiments (lower $\nu$, higher $\sigma$) so as to highlight the model's behavior when the margin loan interest rate hits zero very frequently on its way up to $r_\infty$. To help visualize the population statistics under this change of num\'eraire, Figure \ref{meanratios} provides a 300-year plot of the time functions $\mathbb{E}[V_t/S_t]\pm\std(V_t/S_t),\mathbb{E}[q_t/S_t], \text{and }\mathbb{E}[(q_t+V_t)/S_t]$ for these same parameters.
\par
Figure \ref{density} gives a density estimate (Epanechnikov kernel, $\text{bandwidth}:=0.0193)$ for the random variable $\mslim_{t\to\infty}[(V_t/V_0)\div(S_t/S_0)]$, based on $100,000$ simulations generated by the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$. For these sample paths, on average, the Kelly gambler achieved a relative growth factor of $1.87$, or $87\%$ more final wealth than the equivalent buy-and-hold investor. Note that the Kelly Criterion underperformed buy-and-hold on $2.1\%$ of all sample paths (cf. the asymptotic CDF, which is supplied in the right half of the Figure).
\begin{figure}[t]
\begin{center}
\includegraphics[width=410px]{numeraire.png}
\caption{\sc 100-year sample path of the stochastic processes $V_t/S_t$, $(q_t+V_t)/S_t$, $q_t/S_t$, and $r_L(t)$ for the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.08,0.2)$. $100,000$ steps, $\Delta t:=8.8\text{ hours.}$}
\label{numeraire}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=325px]{meanratios.png}
\caption{\sc 300-year plot of the population statistics $\mathbb{E}[V_t/S_t]\pm\std(V_t/S_t),\mathbb{E}[q_t/S_t], \text{and }\mathbb{E}[(q_t+V_t)/S_t]$ under change of num\'eraire, assuming the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.08,0.2)$.}
\label{meanratios}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=425px]{density.png}
\caption{\sc Empirical distribution of the random variable $\mslim_{t\to\infty}[(V_t/V_0)\div(S_t/S_0)]$, based on 100,000 simulations generated by the parameters $(q_0,V_0,\nu,\sigma):=(1,1,0.09,0.15)$. Epanechnikov kernel $\text{bandwidth}:=0.0193$; the Kelly Criterion underperformed buy-and-hold on 2.1\% of all sample paths.}
\label{density}
\end{center}
\end{figure}
\section{Summary and Conclusions}
This paper established the core dynamical behavior of the broker call money market, which supplies cash to stock brokers for the sake of funding margin loans to retail clients. We assumed (naturally) that the demand side of the market is comprised of continuous time Kelly gamblers, who size their bets over each tick $dt$ of the clock so as to maximize their expected continuously-compounded capital growth rate over $[t,t+dt]$. Ordinarily, under perfectly elastic supply of margin loans (cf. with Garivaltis 2019a), the Kelly gambler is able to beat the market asymptotically almost surely, and by an exponential factor to boot.
\par
To model the powerful long run feedback effects that these sophisticated investors must have on the equilibrium price of margin debt, we assumed that the production side of market amounts to a giant pool of cash that supplies itself inelastically and continuously reinvests all principal and interest. Thus, although the total size of the money market ($q_t$) grows to infinity at a geometric rate, this rate of supply expansion is lower than the asymptotic growth rate $\nu$ of the market index ($S_t$) and the expected compound growth rate of the Kelly bankroll ($V_t$). Proceeding with this intuition, we found that the relative market size $(q_t/V_t)_{t\geq0}$ is a martingale (whose variance tends to infinity) that nonetheless converges to zero in probability\footnote{But \textit{not} in mean square! }; $(q_t/S_t)_{t\geq0}$ is a supermartingale that converges to zero in probability. Consequently, the margin loan interest rate is a submartingale (always expected to increase) that converges in mean square to the choke price $r_\infty=\nu-\sigma^2/2$, where $\sigma$ is the annual $\log$-volatility of the stock market. If the relative size of the money market becomes unexpectedly large (e.g. due to bad stock market performance), then the margin loan interest rate may happen to hit zero on its way up to $r_\infty$; we found a nice rule of thumb for bounding the chances of this ever happening (from here to eternity): the probability is at most $1-(\text{Current Interest Rate}\div\text{Choke Price})$. Based on numerical solutions of the differential equations, we observed that this majorant is typically within $3\%$ of the actual value.
\par
In the same vein, we concluded that the Kelly leverage ratio $(b_t)_{t\geq0}$ is a supermartingale that converges in mean square to $1:1$; thus, the very success of the leveraged investor causes a gradual degradation of the quality of his opportunity set. This is manifest in the asymptotic distribution of the Kelly gambler's performance relative to the a buy-and-hold investor with the same starting capital, e.g. the random variable $\mslim_{t\to\infty}[(V_t/V_0)\div(S_t/S_0)]$. In fact, when all feedback effects are considered, the Kelly gambler no longer beats the market by an exponential factor; his asymptotic compound growth rate, namely $\mslim_{t\to\infty}[\log(V_t/V_0)/t]$, is equal to the stock market growth rate, $\nu$. The realized money market growth rate $\log(q_t/q_0)/t$ converges in mean square to $r_\infty$.
\par
We demonstrated that the leveraged investor's relative growth factor $(V_t/V_0)\div(S_t/S_0)$ is a submartingale (always expected to increase); however, its limiting expected value is at most $1+q_0/V_0$. Thus, if the money market starts out on par with the Kelly bankroll ($q_0/V_0:=1$), then the Kelly gambler cannot expect to achieve any more than double the final wealth of the equivalent buy-and-hold investor. Simulation studies (using the stylized parameter values $(\nu,\sigma):=(0.09,0.15)$ for the S\&P 500 index) indicate that the asymptotic relative growth factor (which is negatively-skewed) has a mean of $1.87$ and a standard deviation of $0.24$; the Kelly Criterion eventually beat the market in $97.9\%$ of all (100,000) simulations. The greatest final relative growth factor ever achieved in simulation was was $2.13$; in a select few of the experiments, the gambler blew himself up spectacularly: the empirical minimum final growth relative was $0.012$. Pray that that never happens to \textit{you}.
\textit{Northern Illinois University}
\subsection*{Acknowledgment}
I thank Dr. Omri Tal, Max Planck Institute for Mathematics in the Sciences, Leipzig, for his incisive communications about the Kelly/Thorp/Cover theory of asymptotic capital growth and competitively optimal trading. Omri's latest paper, \textit{Adaptive Bet-Hedging Revisited: Considerations of Risk and Time Horizon}, is highly recommended.
\subsection*{Disclosures}
This paper is solely the work of the author, who declares that he has no conflicts of interest; the work was funded entirely through his regular academic appointment at Northern Illinois University.
|
\section{Introduction}
\input{intro}
\section{Preliminaries}
\label{sec:prelim}
We assume the reader to be familiar with basic notions of finite automata over infinite words and Markov
chains, see, e.g., \cite{GraedelThomasWilke02,Kulkarni}.
In the following we provide a brief summary of our notation and a few facts related to linear algebra.
\subparagraph*{Finite automata.}
A \emph{B\"uchi automaton}
is a tuple $\mathcal{A} = (Q,\Sigma,\delta,Q_0,F)$
where $Q$ is the finite set of states, $Q_0 \subseteq Q$ is the set of initial
states, $\Sigma$ is the finite alphabet,
$\delta: Q \times \Sigma \to 2^Q$ is
the transition function, and $F \subseteq Q$ is the set of accepting states.
We extend the transition function to $\delta: Q \times \Sigma^* \to 2^Q$
and to $\delta: 2^Q \times \Sigma^* \to 2^Q$ in the standard way.
For $q \in Q$ we write $\mathcal{A}_q$ for the automaton obtained from~$\mathcal{A}$ by making $q$ the only initial state.
Given states $q,r\in Q$ and a finite word
$w = a_0 a_1 \cdots a_{n-1} \in \Sigma^*$,
a \emph{run} for $w$ from $q$ to $r$ is a sequence
$q_0 q_1 \cdots q_n \in Q^{n+1}$ with $q_0=q$, $q_n=r$ and
$q_{i+1} \in \delta(q_i,a_i)$ for $i \in \{0, \ldots, n-1\}$.
A \emph{run} in $\mathcal{A}$ for an infinite word
$w = a_0 a_1 a_2 \cdots \in \Sigma^{\omega}$ is an infinite sequence
$\rho = q_0 q_1 \cdots \in Q^{\omega}$ such that $q_0 \in Q_0$ and
$q_{i+1} \in \delta(q_i,a_i)$ for all $i \in \mathbb{N}$. Run $\rho$ is
called \emph{accepting} if $\inf(\rho) \cap F \ne \emptyset$ where
$\inf(\rho) \subseteq Q$ is the set of states that occur infinitely
often in~$\rho$. The \emph{language} $\mathcal{L}(\mathcal{A})$ of accepted words
consists of all infinite words $w\in \Sigma^{\omega}$ that have at
least one accepting run.
$\mathcal{A}$~is called
\emph{unambiguous} if each word $w\in \Sigma^{\omega}$ has at most one accepting run.
We use the acronym UBA for unambiguous B\"uchi automaton.
We define $|\delta| := |\{(q,r) \mid \exists\, a \in \Sigma : r \in \delta(q,a)\}|$, i.e., $|\delta| \le |Q|^2$ is the number of transitions in~$\mathcal{A}$ when allowing for multiple labels per transition.
In Appendix~\ref{app-A} we give an example that shows that the number of transitions can be quadratic in~$|Q|$, even for UBAs with a strongly connected state space.
We assume $|Q| \le |\delta|$, as states without outgoing transitions can be removed.
In this paper, $\Sigma$ may be a large set (of states in a Markov chain), so it is imperative to allow for multiple labels per transition.
We use a lookup table to check in constant time whether $r \in \delta(q,a)$ holds for given $r, q, a$.
A \emph{diamond} is given by two states $q, r \in Q$ and a finite word~$w$ such that there exist at least two distinct runs for~$w$ from $q$ to~$r$.
One can remove diamonds (see Appendix~\ref{app:prelims})
\begin{restatable}{ourlemma}{lemdiamondfree}\label{lem:diamondfree}
Given a UBA, one can compute in time $O(|\delta|^2|\Sigma|)$ a UBA of at most the same size, with the same language and without diamonds.
\end{restatable}
\noindent For the rest of the paper, we assume that UBAs do not have diamonds.
\subparagraph*{Vectors and matrices.}
We consider vectors and square matrices indexed by a finite set~$S$.
We write (column) vectors $\vec{v} \in \mathbb{R}^S$ with arrows on top,
and $\vec{v}^\top$ for the transpose (a row vector)
of~$\vec{v}$. The zero vector and the all-ones vector are denoted
by~$\vec{0}$ and $\vec{1}$, respectively.
For a set $T \subseteq S$ we write $\cvec{T} \in \{0,1\}^S$ for the characteristic vector of~$T$, i.e., $\cvec{T}_s = 1$ if $s \in T$ and $\cvec{T}_s = 0$ otherwise.
A matrix
$M \in [0,1]^{S \times S}$ is called \emph{stochastic} if
$M \vec{1} = \vec{1}$, i.e., if every row of $M$ sums to one. For a
set $U \subseteq S$ we write $\vec{v}_U \in \mathbb{R}^U$ for the
restriction of~$\vec{v}$ to~$U$. Similarly, for $T, U \subseteq S$ we
write $M_{T,U}$ for the submatrix of~$M$ obtained by deleting the rows
not indexed by~$T$ and the columns not indexed by~$U$. The (directed)
\emph{graph} of a nonnegative matrix $M \in \mathbb{R}^{S \times S}$ has
vertices $s \in S$ and edges $(s,t)$ if $M_{s,t} > 0$. We may
implicitly associate~$M$ with its graph and speak about
graph-theoretic concepts such as reachability and strongly connected
components (SCCs) in~$M$.
\subparagraph*{Markov chains.}
A (finite-state discrete-time) \emph{Markov chain} is a pair
$\mathcal{M} = (S,M)$ where $S$ is the finite set of states, and
$M \in [0,1]^{S \times S}$ is a stochastic matrix that specifies
transition probabilities. An \emph{initial distribution} is a
function $\iota : S \to [0,1]$ satisfying
$\sum_{s \in S} \iota(s) = 1$. Such a distribution induces a
probability measure~$\mathrm{Pr}^{\mathcal{M}}_\iota$ on the measurable subsets
of~$S^\omega$ in the standard way, see for instance~\cite[chapter 10.1, page 758]{baie2008}.
If $\iota$ is concentrated on a
single state~$s$, we may write $\mathrm{Pr}^{\mathcal{M}}_s$
for~$\mathrm{Pr}^{\mathcal{M}}_\iota$.
We write~$E$ for the set of edges in the graph of~$M$.
Note that $|S| \le |E| \le |S|^2$, as $M$ is stochastic.
\subparagraph*{Solving linear systems.}
Let $\kappa \in [2,3]$ be such that one can multiply two $n \times n$-matrices in time $O(n^\kappa)$ (in other literature, $\kappa$ is often denoted by~$\omega$).
We assume that arithmetic operations cost constant time.
One can choose $\kappa = 2.4$, see \cite{gall2014} for a recent result.
One can check whether an $n \times n$ matrix is invertible in time $O(n^\kappa)$~\cite{bunc1974}.
Finally, one can solve a linear system with $n$ equations using the Moore-Penrose pseudo-inverse~\cite{jame1978} in time $O(n^\kappa)$~\cite{petk2009}.
\subparagraph*{Spectral theory.}
The \emph{spectral radius} of a matrix $M \in \mathbb{R}^{S \times S}$,
denoted $\rho(M)$, is the largest absolute value of the eigenvalues
of~$M$.
By the Perron-Frobenius theorem~\cite[Theorems 2.1.1, 2.1.4]{book:bermanP94}, if $M$ is nonnegative then the spectral radius $\rho(M)$ is an eigenvalue of $M$ and there is a nonnegative eigenvector $\vec{x}$ with $M\vec{x}=\rho(M)\vec{x}$.
Such a vector~$\vec{x}$ is called \emph{dominant}.
Further, if $M$ is nonnegative and strongly connected then $\vec{x}$ is strictly positive in all components and the eigenspace associated with $\rho(M)$ is one-dimensional.
\section{Algorithms}
\label{sec:uba}
Given a Markov chain~$\mathcal{M}$, an initial distribution~$\iota$, and a B\"uchi
automaton~$\mathcal{A}$ whose alphabet is the state space of~$\mathcal{M}$, the
\emph{probabilistic model-checking problem} is to compute
$\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A}))$.
This problem is PSPACE-complete~\cite{CY95,BusRubVar04}, but solvable in polynomial time if $\mathcal{A}$ is deterministic.
For UBAs a polynomial-time algorithm was described in~\cite{16BKKKMW-CAV,baieunp1}.
In this paper we obtain a faster algorithm (recall that $E$ is the set of transitions in the Markov chain):
\begin{restatable}{theorem}{thmPMCMCUBA}
\label{thm:PMC-MC-UBA}
Given a Markov chain $\mathcal{M} = (S,M)$, an initial distribution~$\iota$, and
a UBA~$\mathcal{A} = (Q,S,\delta,Q_0,F)$,
one can compute $\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A}))$ in time
$O(|Q|^\kappa |S|^\kappa + |Q|^3|E|+|\delta|^2|E|)$.
\end{restatable}
\noindent Before we prove this theorem in Section~\ref{sub:pseudo-cuts}, we describe the algorithm from~\cite{16BKKKMW-CAV,baieunp1} and analyse the runtime of an efficient implementation.
\subsection{The Basic Linear System} \label{sub:linear-system}
Let $\mathcal{M} = (S,M)$ be a Markov chain, $\iota$ an initial distribution.
Let $B \in \mathbb{R}^{(Q\times S)\times(Q\times S)}$ be the following
matrix:
\begin{equation}\label{eq:Bmat}
B_{\langle q,s\rangle,\langle q',s'\rangle} = \left\{\begin{array}{ll}
M_{s,s'} & \textrm{if $q' \in \delta(q,s)$} \\
0 & \textrm{otherwise}
\end{array}\right.
\end{equation}
Define $\vec{z} \in \mathbb{R}^{Q\times S}$ by $\vec{z}_{\langle q,s \rangle} = \mathrm{Pr}^{\mathcal{M}}_s(\mathcal{L}(\mathcal{A}_q))$.
Then $\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A})) = \sum_{q\in Q_0}\sum_{s \in S} \iota(s) \vec{z}_{\langle q,s \rangle}$.
Lemma 4 in \cite{baieunp1} implies that $\vec{z} = B\vec{z}$.
\begin{example}
Consider the UBA~$\mathcal{A}$ from Figure~\ref{fig:automaton} and the two-state Markov chain~$\mathcal{M}$ shown on the left of Figure~\ref{fig:product}.
The weighted graph on the right of Figure~\ref{fig:product} represents the matrix~$B$, obtained from
$\mathcal{A}$ and $\mathcal{M}$ according to Equation \eqref{eq:Bmat}. It is natural
to think of $B$ as a product of $\mathcal{A}$ and $\mathcal{M}$. Notice that $B$ is
not stochastic: the sum of the entries in each row (equivalently, the
total outgoing transition weight of a graph node) is not always one.
\label{ex:product}
\end{example}
\begin{figure}
\begin{center}
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,auto,node distance=2.5cm,
semithick]
\tikzstyle{every state}=[draw=black,text=black]
\node[accepting,state] (A) {$q_0$};
\node[state] (B) [right of=A] {$q_1$};
\node[state] (C) [below of=A] {$q_2$};
\node[state] (D) [below of=B] {$q_3$};
\draw[->] (-1.3,0) -- (A);
\path (A) edge [bend left] node {$a$} (B);
\path (B) edge [loop right] node {$b$} (B);
\path (B) edge node {$b$} (D);
\path (B) edge [bend left] node {$a$} (A);
\path (C) edge [loop left] node {$b$} (C);
\path (C) edge node {$b$} (A);
\path (C) edge [bend left] node {$a$} (D);
\path (D) edge [bend left] node {$a$} (C);
\node[state] (X) [below of=C] {$a$};
\node[state] (Y) [right of=X] {$b$};
\path (X) edge [bend left] node {$\frac{1}{2}$} (Y);
\path (Y) edge [bend left] node {$\frac{1}{2}$} (X);
\path (X) edge [loop above] node {$\frac{1}{2}$} (X);
\path (Y) edge [loop above] node {$\frac{1}{2}$} (Y);
\end{tikzpicture}
\quad
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,auto,node distance=2.1cm,
semithick, inner sep=1]
\tikzstyle{every state}=[draw=black,text=black]
\node[state,inner sep=0pt,scale=0.98] (A) {$\langle q_3,b \rangle$};
\node[state,inner sep=0pt,scale=0.98] (B) [right of=A] {$\langle q_1,b \rangle$};
\node[state,inner sep=0pt,scale=0.98] (C) [right of=B] {$\langle q_0,a \rangle$};
\node[state,inner sep=0pt,scale=0.98] (D) [above of=C] {$\langle q_1,a \rangle$};
\node[state,inner sep=0pt,scale=0.98] (E) [below of=B] {$\langle q_3,a \rangle$};
\node[state,inner sep=0pt,scale=0.98] (F) [right of=E] {$\langle q_2,b \rangle$};
\node[state,inner sep=0pt,scale=0.98] (G) [right of=F] {$\langle q_0,b \rangle$};
\node[state,inner sep=0pt,scale=0.98] (H) [below of=E] {$\langle q_2,a \rangle$};
\path (B) edge node{$\frac{1}{2}$} (A);
\path (B) edge [loop above] node{$\frac{1}{2}$} (B);
\path (B) edge node{$\frac{1}{2}$} (D);
\path (B) edge node{$\frac{1}{2}$} (E);
\path (C) edge node{$\frac{1}{2}$} (B);
\path (C) edge [bend left=15] node[inner sep=1pt]{$\frac{1}{2}$} (D);
\path (D) edge [bend left=15,'] node[inner sep=1pt]{$\frac{1}{2}$} (C);
\path (D) edge node{$\frac{1}{2}$} (G);
\path (E) edge node{$\frac{1}{2}$} (F);
\path (E) edge [bend left=15] node[inner sep=1pt]{$\frac{1}{2}$} (H);
\path (F) edge node{$\frac{1}{2}$} (C);
\path (F) edge [loop below] node{$\frac{1}{2}$} (F);
\path (F) edge node{$\frac{1}{2}$} (G);
\path (F) edge node{$\frac{1}{2}$} (H);
\path (H) edge node{$\frac{1}{2}$} (A);
\path (H) edge [bend left=15,'] node[inner sep=1pt]{$\frac{1}{2}$} (E);
\begin{pgfonlayer}{background}
\filldraw [line width=4mm,join=round,green!15]
(D.north -| H.west) rectangle (H.south -| D.east);
\filldraw [line width=4mm,join=round,red!15]
(A.north -| A.west) rectangle (A.south -| A.east)
(G.north -| G.west) rectangle (G.south -| G.east);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
\caption{The UBA from Figure~\ref{fig:automaton} and the Markov chain $\mathcal{M}$ on the left, and their product, $B$, on the right. The (single) accepting recurrent SCC is shaded green, and the two other SCCs are shaded red.}
\label{fig:product}
\end{figure}
Although $\vec{z}$ is a solution the system of equations $\vec{\zeta} = B\vec{\zeta}$, this system does not uniquely identify $\vec{z}$.
Indeed, any scalar multiple of $\vec{z}$ is a solution for these equations.
To uniquely identify $\vec{z}$ by a system of linear equations, we need to analyse the SCCs of $B$.
All SCCs~$D$ satisfy $\rho(D) \leq 1$, see~\cite[Proposition~7]{baieunp1}.
An SCC $D$ of~$B$ is called \emph{recurrent} if $\rho(B_{D,D})=1$.
It is called \emph{accepting} if there is $\langle q,s \rangle \in D$ with $q \in F$.
\begin{example}
The matrix~$B$ from Figure \ref{fig:product} has three SCCs, namely the two singleton sets $\{\langle q_0,b \rangle\}$ and $\{\langle q_3,b \rangle\}$, and $D = \{\langle q_0,a \rangle,\langle q_1,a \rangle,\langle q_1,b \rangle,\langle q_2,a \rangle,\langle q_2,b \rangle,\langle q_3,a \rangle\}$. Only $D$ is recurrent;
indeed, $\vec{y} = (\vec{y}_{\langle q_0,a \rangle},\vec{y}_{\langle q_1,a \rangle},\vec{y}_{\langle q_1,b \rangle},\vec{y}_{\langle q_2,a \rangle},\vec{y}_{\langle q_2,b \rangle},\vec{y}_{\langle q_3,a\rangle})^\top = (2,1,3,1,3,2)^\top$ is a dominant eigenvector with $B_{D,D}\vec{y}=\vec{y}$.
Since $q_0$ is accepting, $D$ is accepting recurrent.
\end{example}
Denote the set of accepting recurrent SCCs by $\mathcal{D}_+$ and the set of non-accepting recurrent SCCs by $\mathcal{D}_0$.
By~\cite[Lemma~8]{baieunp1}, for $D \in \mathcal{D}_+$ we have $\vec{z}_d > 0$ for all $d \in D$, and
for $D \in \mathcal{D}_0$ we have $\vec{z}_D = \vec{0}$.
Hence, for $D \in \mathcal{D}_+$, there exists a \emph{$D$-normaliser}, i.e., a vector $\vec{\mu} \in \mathbb{R}^D$ such that $\vec{\mu}^\top \vec{z}_D = 1$.
This gives us a system of linear equations that identifies $\vec{z}$ uniquely~\cite{baieunp1}:
\begin{lemma}[Lemma 12 in \cite{baieunp1}]\label{lem:linear-system}
Let $\mathcal{D}_+$ be the set of accepting recurrent SCCs, and $\mathcal{D}_0$ the
set of non-accepting recurrent SCCs. For each $D \in \mathcal{D}_+$ let
$\vec{\mu}(D)$ be a $D$-normaliser. Then $\vec{z}$~is the unique
solution of the following linear system:
\begin{equation}
\begin{aligned}
&&\vec{\zeta} & = B \vec{\zeta} \\
\text{for all } D \in \mathcal{D}_+ :&& \quad \vec{\mu}(D)^\top \vec{\zeta}_D & = 1 \\
\text{for all } D \in \mathcal{D}_0 :&& \quad \vec{\zeta}_D & = \vec{0}
\end{aligned}
\label{eq:linear-system}
\end{equation}
\end{lemma}
Uniqueness follows from the fact that the system $\vec{\zeta} = B \vec{\zeta}$ describes the eigenspace of the dominant eigenvalue (here,~$1$) of a nonnegative strongly connected matrix (here,~$B$), and such eigenspaces are one-dimensional.
This leads to the following result:
\begin{proposition}\label{prop:mccomp}
Suppose $N$ is the runtime of an algorithm to calculate a normaliser for each accepting recurrent SCC.
Then one can compute $\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A}))$ in time $O(|Q|^\kappa|S|^\kappa)+N$.
\end{proposition}
\begin{proof}
Lemma~\ref{lem:linear-system} implies correctness of the following procedure to calculate $\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A}))$:
\begin{enumerate}
\item Set up the matrix~$B$ from Equation~\eqref{eq:Bmat}.
\item Compute the SCCs of~$B$.
\item For each SCC $C$, check whether $C$ is recurrent.
\item For each accepting recurrent SCC~$D$, compute its $D$-normaliser~$\vec{\mu}(D)$.
\item Compute~$\vec{z}$ by solving the linear system~\eqref{eq:linear-system} in Lemma~\ref{lem:linear-system}.
\item Compute $\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A})) = \sum_{s \in S}\sum_{q \in Q_0}\iota(s)\vec{z}_{q,s}$.
\end{enumerate}
One can set up $B$ in time $O(|Q|^2|S|^2)$.
Using Tarjan's algorithm one can compute the SCCs of~$B$ in time linear in the vertices and edges of~$B$, hence in $O(|Q|^2|S|^2)$~\cite{tarj1972}.
One can find those SCCs $D$ which are recurrent in time $O(|Q|^\kappa|S|^\kappa)$ by checking if $I - B_{D,D}$ is invertible.
The linear system~\eqref{eq:linear-system} has $O(|Q||S|)$ equations, and thus can be solved in time $O(|Q|^\kappa|S|^\kappa)$.
Hence the total runtime is $O(|Q|^\kappa|S|^\kappa)+N$.
\end{proof}
In Section~\ref{sub:cuts} we describe the combinatorial, \emph{cut} based, approach from~\cite{16BKKKMW-CAV,baieunp1} to calculating $D$-normalisers and analyse its complexity.
In Section~\ref{sub:pseudo-cuts} we describe a novel linear-algebra based approach, which is faster in terms of the automaton.
\subsection{Calculating \texorpdfstring{$D$}{D}-Normalisers Using Cuts}
\label{sub:cuts}
For the remainder of the paper, let $D$ be an accepting recurrent SCC.
A \emph{fibre over $s \in S$} is a subset of $D$ of the form $\alpha\times\{s\}$ for some $\alpha \subseteq Q$.
Given a fibre $f = \alpha\times\{s\}$ and a state $s' \in S$, if $M_{s,s'} > 0$ we define the fibre $f \then s'$ as follows:
\begin{equation*}
f \then s' := \{\langle q,s' \rangle \; | \; q \in \delta(\alpha,s)\}\cap D.
\end{equation*}
If $M_{s,s'} = 0$, then $f \then s'$ is undefined, and for $w \in S^*$ we define $f \then w = f$ if $w = \varepsilon$ and $f \then ws' = (f \then w)\then s'$. If $f = \{d\}$ for some $d \in D$ we may write $d \then s'$ for $f \then s'$.
We call a fibre $c$ a \emph{cut} if $c = d \then v$ for some $v \in S^*$ and $d \in D$, and $c \then w \neq \emptyset$ for all $w \in S^*$ whenever $c \then w$ is defined.
Note that if $c$ is a cut then so is $c \then w$ whenever it is defined.
Given a cut $c \subseteq D$ we call its characteristic vector $\cvec{c} \in \{0,1\}^D$ a \emph{cut vector}. In the example in Figure~\ref{fig:product}, it is easy to see that $\{\langle q_1,b\rangle\} = \langle q_0,a\rangle \then b$ is a cut.
\begin{lemma}[Lemma 10 in \cite{baieunp1}]\label{lem:cutnormalises}
There exists a cut.
Any cut vector $\vec{\mu}$ is a normaliser, i.e., $\vec{\mu}^\top \vec{z}_D = 1$.
\end{lemma}
Loosely speaking, $\vec{\mu}^\top \vec{z}_D \le 1$ follows from unambiguousness, and $\vec{\mu}^\top \vec{z}_D \not< 1$ follows from an ergodicity argument (intuitively, all states in the cut are almost surely visited infinitely often).
The following lemma is the basis for the cut computation algorithm in~\cite{16BKKKMW-CAV,baieunp1}:
\begin{lemma}[Lemma 17 in \cite{baieunp1}] \label{lem:get-larger}
Let $D \subseteq Q \times S$ be a recurrent SCC.
Let $d \in D$.
Suppose $w \in S^*$ is such that $d \then w \ni d$ is not a cut.
Then there are $v \in S^*$ and $e \ne d$ with $d \then v \supseteq \{d,e\}$ and $e \then w \ne \emptyset$. For any such $e$, $d \then w \cap e \then w = \emptyset$.
Hence $d \then v w \supseteq \{d,e\} \then w \supsetneq d \then w$.
\end{lemma}
This suggests a way of generating an increasing sequence of fibres, culminating in a cut.
We prove the following proposition:
\begin{restatable}{ourproposition}{propfindcut}\label{prop:findcut}
Let $D \subseteq Q \times S$ be a recurrent SCC.
Denote by $T$ the set of edges in~$B_{D,D}$.
One can compute a cut in time
$O(|Q|^2 |\delta| |D| + |\delta| |T|)$.
\end{restatable}
\noindent Define, for some $d = \langle q,s\rangle \in D$, its \emph{co-reachability} set $\mathit{Co}(d) \subseteq D$:
it consists of those $e \in D$ such that there exists a word $w$ with $\{d,e\} \subseteq d \then w$.
Note that $\mathit{Co}(d)$ is a fibre over~$s$. In the example of Figure~\ref{fig:product} we have that $\mathit{Co}(\langle q_0, a\rangle) = \{\langle q_0, a\rangle,\langle q_2, a\rangle\}$, with $\{\langle q_0, a\rangle, \langle q_2, a\rangle\} \in \langle q_0, a\rangle \then ba$.
The following lemma (proof in Appendix~\ref{app:cuts}) gives a bound on the time to compute~$\mathit{Co}(d)$:
\begin{restatable}{ourlemma}{lemcalccod}\label{lem:calccod}
One can compute $\mathit{Co}(d)$ in time $O(|Q| |D| + |\delta| |T|)$.
Moreover, one can compute in time $O(|Q|^2 |D| + |\delta| |T|)$ a list $(\mathit{CoPath}(d)(e))_{e \in \mathit{Co}(d)}$ such that $\mathit{CoPath}(d)(e) \in S^*$ and $\{d,e\} \subseteq d \then \mathit{CoPath}(d)(e)$ and $|\mathit{CoPath}(d)(e)| \le |Q| |D|$.
\end{restatable}
\noindent The lemma is used in the proof of Proposition~\ref{prop:findcut}:
\begin{proof}[Proof sketch of Proposition~\ref{prop:findcut}]
Starting from a singleton fibre $\{d\}$, where $d = \langle q,s\rangle \in D$ is chosen arbitrarily, we keep looking for words $v \in S^*$ that have the properties described in Lemma \ref{lem:get-larger} to generate larger fibres $d\then w$:
\begin{enumerate}
\item $w := \varepsilon$ (the empty word)
\item while $\exists\, v \in S^*$ and $\exists\, e \ne d$ such that $d \then v \supseteq \{d, e\}$ and $e \then w \neq \emptyset:$ \\
\mbox{}\hspace{4mm} $w := v w$
\item return $d \then w$.
\end{enumerate}
By \cite[Lemma~18]{16BKKKMW-CAV} the algorithm returns a cut.
In every loop iteration the fibre $d \then w$ increases, so the loop terminates after at most $|Q|$ iterations.
For efficiency we calculate $\mathit{Co}(d)$ and $\mathit{CoPath}(d)$ using Lemma~\ref{lem:calccod}, and we use dynamic programming to maintain the set, $\mathit{Survives}$, of those $e \in D$ for which $e \then w \ne \emptyset$ holds.
Whenever a prefix~$v$ is added to~$w$, we update~$\mathit{Survives}$ by processing~$v$ backwards.
This leads to the following algorithm:
\begin{enumerate}
\item Calculate $\mathit{Co}(d)$ and $\mathit{CoPath}(d)$ using Lemma \ref{lem:calccod}
\item $w := \varepsilon;$ $\mathit{Survives} := (Q \times \{s\}) \cap D$
\item while $\exists\, e \in \mathit{Co}(d) \setminus \{d\}$ such that $e \in \mathit{Survives}$: \\
\mbox{}\hspace{4mm} $v_0= s$; $v_1\ldots v_n := \mathit{CoPath}(d)(e)$ \\
\mbox{}\hspace{4mm} for $i = n, n-1, \ldots, 1$: \\
\mbox{}\hspace{10mm} $\mathit{Survives} := \{\langle p,v_{i-1}\rangle \in D \mid (\delta(p, v_{i-1}) \times \{v_i\}) \cap \mathit{Survives} \neq \emptyset\}$ \\
\mbox{}\hspace{4mm} $w := v_1\ldots v_nw$
\item return $d \then w$
\end{enumerate}
The runtime analysis is in Appendix~\ref{app:cuts}.
\end{proof}
\begin{example}\label{ex:cut}
Letting $d = \langle q_0,a \rangle$ and $e = \langle q_2,a \rangle$ we have $\mathit{Co}(d) = \{d, e\}$ with $\mathit{CoPath}(d)(d) = \varepsilon$ and $\mathit{CoPath}(d)(e) = b a a$.
Initially we have $\mathit{Survives} = Q \times \{a\}$.
In the first iteration the algorithm can only pick~$e$.
The inner loop updates $\mathit{Survives}$ first to $\{q_0, q_1, q_2, q_3\} \times \{a\}$ (i.e., to itself), then to $\{q_1, q_2\} \times \{b\}$, and finally to $\{q_0,q_3\} \times \{a\}$.
Now $(\mathit{Co}(d) \setminus d) \cap \mathit{Survives}$ is empty and the loop terminates.
The algorithm returns the cut $d \then baa = \{d,e\}$.
\end{example}
Applying Proposition~\ref{prop:findcut} to the general procedure (Proposition~\ref{prop:mccomp}) leads to the following result on the combinatorial approach:
\begin{restatable}{ourtheorem}{thmPMCMCUBAcut}
\label{thm:PMC-MC-UBA-cut}
Given a Markov chain $\mathcal{M} = (S,M)$, an initial distribution~$\iota$, and
a UBA~$\mathcal{A} = (Q,S,\delta,Q_0,F)$, one can compute $\mathrm{Pr}^{\mathcal{M}}_\iota(\mathcal{L}(\mathcal{A}))$
in time
$O(|Q|^\kappa |S|^\kappa + |Q|^3 |\delta| |S|+|\delta|^2 |E|)$.
\end{restatable}
\input{pseudo-cut}
\section{Discussion} \label{sec-discussion}
We have analysed two algorithms for computing normalisers: the cut-based one by Baier et al.~\cite{16BKKKMW-CAV,baieunp1}, and a new one, which draws from techniques by Protasov and Voynov~\cite{Protasov17} for the analysis of matrix semigroups.
The first approach is purely combinatorial, and in terms of the automaton, an efficient implementation runs in time $O(|Q|^3 |\delta| + |\delta|^2) = O(|Q|^3 |\delta|)$ (Proposition~\ref{prop:findcut}).
The second approach combines a linear-algebra component to compute~$R(s)$ with a combinatorial algorithm to compute the co-reachability set~$\mathit{Co}(d)$.
In terms of the automaton, the linear-algebra component runs in time $O(|Q|^3)$ (Lemma~\ref{lem:calcR}), while the combinatorial part runs in time $O(|\delta|^2)$, leading to an overall runtime of $O(|Q|^3 + |\delta|^2)$.
Note that for all $r \in [1,2]$, if $|\delta| = \Theta(|Q|^r)$ then the second approach is faster by at least a factor of~$|Q|$.
Although it is not the main focus of this paper, we have analysed also the model-checking problem, where a non-trivial Markov chain is part of the input.
The purely combinatorial algorithm runs in time $O(|Q|^\kappa |S|^\kappa + |Q|^3 |\delta| |S|+|\delta|^2 |E|)$, and the linear-algebra based algorithm in time $O(|Q|^\kappa |S|^\kappa + |Q|^3 |E|+|\delta|^2 |E|)$.
There are cases in which the latter is asymptotically worse, but not if $\kappa = 3$ (i.e., solving linear systems in a normal way such as Gaussian elimination) or if $|E|$ is $O(|S|)$.
It is perhaps unsurprising that a factor of $|\delta|^2$ from the computation of~$\mathit{Co}(d)$ occurs in the runtime, as it also occurs when one merely verifies the unambiguousness of the automaton, by searching the product of the automaton with itself.
Can the factor $|\delta|^2$ (which may be quartic in~$|Q|$) be avoided?
\bibliographystyle{plain}
\subsection{Calculating \texorpdfstring{$D$}{D}-Normalisers Using Linear Algebra}
\label{sub:pseudo-cuts}
Recall that $D$ is an accepting recurrent SCC.
For $t \in S$ define the matrix $\Delta(t) \in \{0,1\}^{D \times D}$ as follows:
\begin{equation*}
\Delta(t)_{\langle q, s\rangle,\langle q', s'\rangle} := \left\{\begin{array}{ll}
1 & \quad\textrm{if $s' = t$, $M_{s,t} > 0$, and $q' \in \delta(q,s)$} \\
0 & \quad\textrm{otherwise}
\end{array}\right.
\end{equation*}
Note that the graph of~$\Delta(t)$ contains exactly the edges of the graph of~$B_{D,D}$ that end in vertices in $Q \times \{t\}$.
If $M_{s,t} > 0$ holds for all pairs $(s,t)$, then the matrices $(\Delta(t))_{t \in S}$ generate a semigroup of matrices, all of which have spectral radius~$1$.
Such semigroups were recently studied by Protasov and Voynov~\cite{Protasov17}.
Specifically, Theorem~5 in~\cite{Protasov17} shows that there exists an affine subspace $\mathcal{F}$ of $\mathbb{R}^{D}$ which excludes $\vec{0}$ and is invariant under multiplication by matrices from the semigroup.
Moreover, they provide a way to compute this affine subspace efficiently.
One can show that cut vectors are orthogonal to~$\mathcal{F}$.
The key idea of our contribution is to generalise cut vectors to \emph{pseudo-cuts}, which are vectors $\vec{\mu} \in \mathbb{R}^D$ that are orthogonal to~$\mathcal{F}$.
We will show (in Lemma~\ref{lem:colocal-pseudo-cut} below) how to derive a $D$-normaliser based on a pseudo-cut that is non-zero only in components that are in a co-reachability set~$\mathit{Co}(d)$ (from Lemma~\ref{lem:calccod}).
If $M_{s,t} = 0$ holds for some $s,t$ (which will often be the case in model checking), then $\Delta(s) \Delta(t)$ is the zero matrix, which has spectral radius~$0$, not~$1$.
Therefore, the results of~\cite{Protasov17} are not directly applicable and we have to move away from matrix semigroups.
In the following we re-develop and generalise parts of the theory of~\cite{Protasov17} so that the paper is self-contained and products of $\Delta(s) \Delta(t)$ with $M_{s,t} = 0$ are not considered.
Let $w = s_1 s_2 \ldots s_n \in S^*$.
Define $\Delta(w) = \Delta(s_1) \Delta(s_2) \cdots \Delta(s_n)$.
We say $w$ is \emph{enabled} if $M_{s_i,s_{i+1}} > 0$ holds for all $i \in \{1,\ldots, n-1\}$.
If $f \subseteq D$ is a fibre over~$s$ such that $s w$ is enabled, we have $\cvec{f \then w}^\top = \cvec{f}^\top\Delta(w)$.
We overload the term \emph{fibre over $s$} to describe any vector $\vec{\mu} \in \mathbb{R}^D$ such that $\vec{\mu}_{\langle q,s'\rangle} = 0$ whenever $s' \neq s$.
We define \emph{pseudo-cuts over~$s$} to be fibres $\vec{\mu}$ over~$s$ such that $\vec{\mu}^\top\Delta(w)\vec{z} = \vec{\mu}^\top\vec{z}$ holds for all $w \in S^*$ such that $s w$ is enabled.
Let $c \subseteq Q \times \{s\}$ be a cut with $s w$ enabled.
Then $c \then w$ is a cut, and $\cvec{c}^\top \Delta(w) \vec{z} = 1 = \cvec{c}^\top \vec{z}$ holds by Lemma~\ref{lem:cutnormalises}.
It follows that cut vectors are pseudo-cuts.
\begin{example}\label{ex:cutvects}
Since $c = \{\langle q_0,a \rangle, \langle q_2,a \rangle\}$ from Example~\ref{ex:cut} is a cut, $\cvec{c}$ is a pseudo-cut over~$a$.
Pseudo-cuts do not need to be combinations of cut vectors:
although the fibre $f = \{\langle q_0,a\rangle, \langle q_1,a\rangle\}$ is not a cut, $\cvec{f}$ is a pseudo-cut over~$a$.
\end{example}
Fix some $d = \langle q,s \rangle \in D$.
Recall that $\mathit{Co}(d)$ consists of those $e \in D$ such that there exists a word $w$ with $\{d,e\} \subseteq d \then w$.
We define \emph{$\mathit{Co}(d)$-pseudo-cuts} to be pseudo-cuts $\vec{\mu}$ over~$s$ such that $\vec{\mu}_d \ne 0$ and $\vec{\mu}_e = 0$ holds for all $e \not \in \mathit{Co}(d)$.
\begin{example}
Any cut vector is a $\mathit{Co}(d)$-pseudo-cut for some $d \in D$, by definition, and so are scalar multiples of cut vectors. The vector $\cvec{f}$ in Example~\ref{ex:cutvects}, however, is not a $\mathit{Co}(d)$-pseudo-cut, since $\langle q_1,a\rangle \not \in \mathit{Co}(\langle q_0, a\rangle)$ and $\langle q_0,a\rangle \not \in \mathit{Co}(\langle q_1, a\rangle)$.
\end{example}
From a $\mathit{Co}(d)$-pseudo-cut we can easily derive a $D$-normaliser:
\begin{lemma}\label{lem:colocal-pseudo-cut}
Let $\vec{\mu} \in \mathbb{R}^D$ be a $\mathit{Co}(d)$-pseudo-cut.
Then $\frac{1}{\vec{\mu}_d}\vec{\mu}$ is a $D$-normaliser.
\end{lemma}
\begin{proof}
Let $w$ be an enabled word in~$M$ such that $d \then w$ is a cut containing~$d$.
Such a word exists (see the proof sketch of Proposition~\ref{prop:findcut}).
Since $(\cvec{d}^\top\Delta(w))^\top = \cvec{d \then w}$ is a $D$-normaliser (by Lemma~\ref{lem:cutnormalises}), it suffices to prove that $\frac{1}{\vec{\mu}_d}\vec{\mu}^\top\vec{z} = \cvec{d}^\top \Delta(w)\vec{z}$.
We can write $\vec{\mu}$ as $\sum_{d' \in \mathit{Co}(d)} \vec{\mu}_{d'} \cvec{d'}$, so $\vec{\mu}^\top \Delta(w) = \sum_{d' \in \mathit{Co}(d)} \vec{\mu}_{d'} \cvec{d'}^\top \Delta(w)$. For any $d' \in \mathit{Co}(d) \setminus \{d\}$, let $w'$ be such that $\{d, d'\} \subseteq d \then w'$. Now we see that $d' \in d \then w w'$, and since $d \then w$ is a cut so are $d \then ww'$ and $d \then ww'w$. Thus,
\begin{equation*}
\cvec{d}^\top\Delta(w)\vec{z} \ = \ \cvec{d}^\top\Delta(ww'w)\vec{z} \ \geq \ \cvec{d}^\top\Delta(w)\vec{z}+\cvec{d'}^\top\Delta(w)\vec{z},
\end{equation*}
which implies $\cvec{d'}^\top\Delta(w)\vec{z} = 0$ for every $d' \in \mathit{Co}(d) \setminus \{d\}$. This means that
\begin{equation*}
\vec{\mu}^\top\Delta(w)\vec{z} = \sum_{d' \in \mathit{Co}(d)}\vec{\mu}_{d'}\cvec{d'}^\top\Delta(w)\vec{z} = \vec{\mu}_d\cvec{d}^\top\Delta(w)\vec{z}.
\end{equation*}
Since $\vec{\mu}$ is a pseudo-cut, this implies that $\frac{1}{\vec{\mu}_d}\vec{\mu}^\top\vec{z} = \frac{1}{\vec{\mu}_d}\vec{\mu}^\top\Delta(w)\vec{z} = \cvec{d}^\top\Delta(w)\vec{z}$.
\end{proof}
By Lemma~\ref{lem:colocal-pseudo-cut}, to find a $D$-normaliser it suffices to find a $\mathit{Co}(d)$-pseudo-cut.
Fix a dominant eigenvector $\vec{y}$ of~$B_{D,D}$ so that $\vec{y}$ is strictly positive in all components.
One can compute such~$\vec{y}$ in time $O(|D|^\kappa)$.
By \cite[Lemma~8]{16BKKKMW-CAV} the vector $\vec{z}_{D}$ is also a dominant eigenvector of~$B_{D,D}$, hence $\vec{y}$ and $\vec{z}_{D}$ (the latter of which is yet unknown) are scalar multiples.
In order to compute a $\mathit{Co}(d)$-pseudo-cut, we compute a basis for the space spanned by $\Delta(w)\vec{y}$ for all enabled words~$w$.
We use a technique similar to the one employed by Tzeng in \cite{tzen1992} for checking equivalence of probabilistic automata.
To make this more efficient, we compute separate basis vectors for each $s \in S$.
Define $\Delta'(t) \in \{0,1\}^{D \times D}$ as $\Delta'(t)_{\langle q_1,s_1\rangle,\langle q_2,s_2\rangle} = 1$ if $q_1=q_2$ and $s_1 = s_2 = t$ and 0 otherwise.
Note that $\Delta(s)\Delta'(s) = \Delta(s)$ holds for all $s \in S$.
\begin{restatable}{ourlemma}{lemcalcR}\label{lem:calcR}
Suppose $\vec{y} = B_{D,D} \vec{y}$ is given.
Denote by $V(s) \subseteq \mathbb{R}^D$ the vector space spanned by the vectors $\Delta'(s)\Delta(w)\vec{y}$ for $w \in S^*$ and $s \in S$.
Let $Q_{D, t} = (Q \times \{t\}) \cap D$ and let $E(t) = \{(s,t) \mid M_{s,t} > 0\}$ be the set of edges in~$M$ that end in~$t$.
One can compute a basis $R(s)$ of $V(s)$ for all $s \in S$ in time
{$O(|Q|^2 \sum_{t \in S} |Q_{D,t}||E(t)|)$}, where for each $\vec{r} \in R(s)$ we have $\vec{r} = \Delta'(s)\Delta(w)\vec{y}$ for some enabled word $s w$.
\end{restatable}
\begin{proof}[Proof sketch]
Fix an arbitrary total order $<_S$ on $S$.
We define a total order $\ll_S$ on $S^*$ as the ``shortlex'' order but with words read from right to left.
That is, the empty word~$\varepsilon$ is the smallest element, and for $v, w \in S^*$ and $s,t \in S$, we have $v s \ll_S w t$ if (1) $|v s| < |w t|$ or (2) $|vs| = |w t|$ and $s <_S t$ or (3) $|v s| = |w t|$ and $s = t$ and $v \ll_S w$.
We use a technique similar to the one by Tzeng in~\cite{tzen1992}.
At every step in the algorithm, $\mathit{worklist}$ is a set of pairs $(s w, \Delta'(s) \Delta(w) \vec{y})$.
We write $\min_{\ll_S}(\mathit{worklist})$ to denote the pair in $\mathit{worklist}$ where $s w$ is minimal with respect to~$\mathord{\ll_S}$.
\begin{enumerate}
\item for each $s \in S$, let $R(s) := \{\Delta'(s)\vec{y}\}$ and $R(s)_\bot := \{\Delta'(s)\vec{y}\}$
\item $\mathit{worklist} := \{(s t,\Delta'(s)\Delta(t)\vec{y}) \mid M_{s,t} > 0\}$
\item while $\mathit{worklist} \neq \emptyset$: \\
\mbox{}\hspace{4mm} $(t w,\vec{u}) := \min_{\ll_S}(\mathit{worklist})$; $\mathit{worklist} := \mathit{worklist} \setminus \{(t w,\vec{u})\}$ \\
\mbox{}\hspace{4mm} Using the Gram-Schmidt process\footnote{For good numerical stability, one should use the so-called Modified Gram-Schmidt process~\cite[Chapter~19]{Higham}.}, let $\vec{u}_\bot$ be the orthogonalisation of $\vec{u}$ against $R_\bot(t)$ \\
\mbox{}\hspace{4mm} if $\vec{u}_\bot \neq \vec{0}$, i.e., if $\vec{u}$ is linearly independent of $R_\bot(t)$: \\
\mbox{}\hspace{10mm} $R(t) := R(t) \cup \{\vec{u}\}$ and $R_\bot(t) := R_\bot(t) \cup \{\vec{u}_\bot\}$ \\
\mbox{}\hspace{10mm} $\mathit{worklist} := \mathit{worklist} \cup \{(s t w, \Delta'(s)\Delta(t)\vec{u}) \mid M_{s,t} > 0\}$
\item return $R(s)$ for all $s \in S$
\end{enumerate}
At any point and for all $s \in S$, the sets $R(s)$ and $R(s)_\bot$ span the same vector space, and this space is a subspace of~$V(s)$.
The sets $R(s)$ and $R(s)_\bot$ consist of linearly independent fibres over $s$, and these fibres are possibly nonzero only in the $Q_{D,s}$-components.
Hence $|\bigcup_{s \in S} R(s)| \leq |D|$ and thus there are at most $|D|$ iterations of the while loop that increase $\mathit{worklist}$.
At every iteration where $\vec{u}$ is dependent on $R(t)_\bot$ the set $\mathit{worklist}$ decreases by one, and therefore the algorithm terminates.
In Appendix~\ref{app:pseudocuts} we prove that in the end we have that $R(s)$ spans~$V(s)$, and we analyse the runtime.
\end{proof}
\begin{example}\label{ex:spanex}
Let us return to our running example. We see that the vector $\vec{y} = (\vec{y}_{\langle q_0,a \rangle},\vec{y}_{\langle q_1,a \rangle},\vec{y}_{\langle q_1,b \rangle},\vec{y}_{\langle q_2,a \rangle},\vec{y}_{\langle q_2,b \rangle},\vec{y}_{\langle q_3,a \rangle})^\top = (2,1,3,1,3,2)^\top$ is a dominant eigenvector of $B_{D,D}$.
Fix the order $a <_S b$.
Step~1 initialises $R(a)$ to $\{\Delta'(a) \vec{y}\}$ and $R(b)$ to $\{\Delta'(b) \vec{y}\}$, where $\Delta'(a) \vec{y} = (2,1,0,1,0,2)^\top$ and $\Delta'(b) \vec{y} = (0,0,3,0,3,0)^\top$.
Step~2 computes $\Delta'(a)\Delta(a)\vec{y} = (1,2,0,2,0,1)^\top$, which is linearly independent of $\Delta'(a)\vec{y}$.
However, $\Delta'(b)\Delta(a)\vec{y} = (0,0,3,0,3,0)^\top = \Delta'(b)\vec{y}$.
Also, $\Delta'(a)\Delta(b)\vec{y} = (3,0,0,0,0,3)^\top = 2\Delta'(a)\vec{y}-\Delta'(a)\Delta(a)\vec{y}$ and $\Delta'(b)\Delta(b)\vec{y} = (0,0,3,0,3,0)^\top = \Delta'(b)\vec{y}$.
One can check that $\Delta'(a) \Delta(a a) \vec{y} = \Delta'(a) \vec{y}$ and $\Delta'(b) \Delta(a a) \vec{y} = \Delta'(b) \vec{y}$.
Hence the algorithm returns $R(a) = \{(2,1,0,1,0,2)^\top,(1,2,0,2,0,1)^\top\}$ and $R(b) = \{(0,0,3,0,3,0)^\top\}$.
\end{example}
Fix $d = \langle q,s\rangle \in D$ for the rest of the paper.
The following lemma characterises $\mathit{Co}(d)$-pseudo-cuts in a way that is efficiently computable:
\begin{lemma}\label{lem:nullspclpc}
A vector $\vec{\mu} \in \mathbb{R}^D$ with $\vec{\mu}_d = 1$ and $\vec{\mu}_e = 0$ for all $e \not \in \mathit{Co}(d)$ is a $\mathit{Co}(d)$-pseudo-cut if and only if $\vec{\mu}^\top \vec{r} = \vec{\mu}^\top \vec{y}$ holds for all $\vec{r} \in R(s)$.
\end{lemma}
For an intuition of the proof, consider the affine space, $\mathcal{F} \subseteq \mathbb{R}^D$, affinely spanned by those $\Delta'(s) \Delta(w) \vec{y}$ for which $s w$ is enabled.
This affine space was alluded to in the beginning of this subsection and is visualised as a blue straight line on the right of Figure~\ref{fig:automaton}.
The shaded plane in this figure is the vector space of pseudo-cuts over~$s$.
This space is orthogonal to~$\mathcal{F}$.
The following lemma says that $\mathcal{F}$ is affinely spanned by the points in~$R(s)$.
This strengthens the property of~$R(s)$ in Lemma~\ref{lem:calcR} where $R(s)$ was defined to span a \emph{vector} space.
\begin{lemma} \label{lem:affine}
Let $w \in S^*$ be such that $s w$ is enabled.
By the definition of~$R(s)$ there are $\gamma_{\vec{r}} \in \mathbb{R}^D$ for each $\vec{r} \in R(s)$ such that $\Delta'(s) \Delta(w) \vec{y} = \sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} \vec{r}$.
We have $\sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} = 1$.
\end{lemma}
\begin{proof}
Let $c$ be a cut containing~$d$.
Since $R(s)$ is a basis, for any $\vec{r} = \Delta'(s)\Delta(w_{\vec{r}})\vec{y} \in R(s)$ the word $s w_{\vec{r}}$ is enabled.
Therefore, $c \then w_{\vec{r}}$ is a cut and by Lemma~\ref{lem:cutnormalises} we have $\cvec{c \then w_{\vec{r}}}^\top \vec{y} = \cvec{c}^\top\vec{y}$.
Hence $\cvec{c}^\top \vec{r} = \cvec{c}^\top \Delta'(s) \Delta(w_{\vec{r}}) \vec{y} = \cvec{c}^\top \Delta(w_{\vec{r}}) \vec{y} = \cvec{c \then w_{\vec{r}}}^\top \vec{y} = \cvec{c}^\top\vec{y}$.
Moreover, we have:
\begin{align*}
\cvec{c}^\top\vec{y} & = \cvec{c}^\top\Delta(w)\vec{y} &&\text{since $s w$ is enabled and by Lemma~\ref{lem:cutnormalises}} & \\
& = \cvec{c}^\top \Delta'(s)\Delta(w)\vec{y} &&\text{since $\cvec{c}$ is a fibre over $s$} \\
& = \cvec{c}^\top \sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} \vec{r} &&\text{by the definition of $\gamma_{\vec{r}}$}\\
& = \cvec{c}^\top\vec{y} \sum_{\vec{r} \in R(s)}\gamma_{\vec{r}} &&\text{as argued above.}
\end{align*}
Therefore, $\sum_{\vec{r} \in R(s)}\gamma_{\vec{r}} = 1$.
\end{proof}
Now we can prove Lemma~\ref{lem:nullspclpc}:
\begin{proof}[Proof of Lemma~\ref{lem:nullspclpc}]
For the ``if'' direction, let $w$ be such that $sw$ is enabled, and it suffices to show that $\vec{\mu}^\top \Delta(w) \vec{y} = \vec{\mu}^\top \vec{y}$.
By Lemma~\ref{lem:affine} there are $\gamma_{\vec{r}}$ such that $\Delta'(s) \Delta(w) \vec{y} = \sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} \vec{r}$ and $\sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} = 1$.
We have:
\begin{equation*}
\vec{\mu}^\top \Delta(w) \vec{y} \ = \ \vec{\mu}^\top \Delta'(s)\Delta(w)\vec{y} \ = \ \sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} \vec{\mu}^\top \vec{r} \ = \ \sum_{\vec{r} \in R(s)} \gamma_{\vec{r}} \vec{\mu}^\top \vec{y} \ = \ \vec{\mu}^\top \vec{y}\;,
\end{equation*}
where the last equality is from Lemma~\ref{lem:affine}.
For the ``only if'' direction, suppose $\vec{\mu}$ is a $\mathit{Co}(d)$-pseudo-cut.
Let $\vec{r} = \Delta'(s)\Delta(w_{\vec{r}})\vec{y} \in R(s)$.
Then $s w_{\vec{r}}$ is enabled and $\vec{\mu}^\top \vec{r} = \vec{\mu}^\top\Delta'(s)\Delta(w_{\vec{r}})\vec{y} = \vec{\mu}^\top \Delta(w_{\vec{r}})\vec{y} = \vec{\mu}^\top \vec{y}$.
\end{proof}
\begin{example}
In Example~\ref{ex:cut} we derived that $\vec{y} = (2,1,3,1,3,2)^\top$ and $R(a) = \{(2,1,0,1,0,2)^\top,(1,2,0,2,0,1)^\top\}$.
The cut vector $\vec{\mu} = (1,0,0,1,0,0)^\top$ from Example~\ref{ex:cutvects} satisfies $\vec{\mu}^\top \vec{r} = 3 = \vec{\mu}^\top \vec{y}$ for both $\vec{r} \in R(a)$.
\end{example}
Using Lemmas \ref{lem:calccod}, \ref{lem:calcR} and~\ref{lem:nullspclpc} we obtain:
\begin{restatable}{ourproposition}{propfindcodcut}\label{prop:findcodcut}
Let $D \subseteq Q \times S$ be a recurrent SCC.
Denote by $T_D$ the set of edges of~$B_{D,D}$.
For $t \in S$, let~$E(t)$ denote the set of edges of~$M$ that end in $t$, and let $Q_{D,t} = (Q \times \{t\}) \cap D$.
Let $d = \langle q,s\rangle \in D$.
One can compute a $\mathit{Co}(d)$-pseudo-cut in time
$O(|D|^\kappa + |Q||D|+|\delta||T_D| + |Q|^2 \sum_{t \in S} |Q_{D,t}||E(t)|)$.
\end{restatable}
Now our main result follows, which we restate here:
\thmPMCMCUBA*
|
\section{Introduction}\label{sec:introduction}
In many situations fitting one global model to a given data set can be very challenging,
especially if the data contains lots of different features with strong variation
and complex interactions. Therefore, separating the data into more homogeneous subgroups
based on a set of covariates first can simplify the task and fitting a local model to
each of the resulting subgroups often leads to better results. This separation can be
done by applying a tree algorithm. While almost all algorithms proposed in the literature
follow the general idea of splitting the data such that some objective function is optimized
locally, they differ in their specific approaches to selecting a split variable and the
corresponding split point.
Some of the first tree algorithms (e.g., AID, \citealp{Morgan+Sonquist:1963};
CART, \citealp{Breiman+Friedman+Stone:1984}) rely on exhaustive search procedures to find both,
the best split point and split variable in one step by directly comparing all possible split
points in all possible split variables. However, it has been shown that this is not only
computationally expensive but also biased towards split variables with many possible split
points \citep{Doyle:1973, Kim+Loh:2001}.
Therefore, selecting a split variable in a first step and then searching for the
best split point only within this variable in a separated second step is a more
promising strategy as applied for example by the algorithms QUEST \citep{Loh+Shih:1997},
GUIDE \citep{Loh:2002}, CTree \citep{Hothorn+Hornik+Zeileis:2006}
and MOB \citep{Zeileis+Hothorn+Hornik:2008}.
For the first step of selecting a split variable they all share the same basic concept of
choosing the covariate which shows the highest association to the response variable
based on $p$-values provided by a statistical test. While QUEST
and GUIDE employ
statistical significance tests for contingency tables, CTree applies permutation tests in
a conditional inference framework and MOB uses fluctuation tests based on central limit theorems
for the parameter estimators.
All these approaches have been shown to work well for various situations, however, the relative
(dis)advantages of the testing strategies have not yet been investigated and compared in detail.
Therefore, in this paper the focus will be put on that first step of tree algorithms,
i.e., the task of selecting the best split variable in a given (sub)sample. In particular, the approach of the
GUIDE algorithm is compared to the one of the CTree algorithm and the MOB algorithm
by investigating the building blocks of their testing strategies in which they differ:
(1)~Variation of the goodness-of-fit measure for the response: using residuals or full model scores.
(2)~Dichotomization of these residuals or scores.
(3)~Categorization of possible split variables.
Apart from these three main factors further aspects such as the approximation of the null
distribution (conditional vs.\ unconditional) or the type of test statistic (maximally
selected vs.\ sum of squares) will be considered as well.
For this purpose, a unifying framework for testing strategies in unbiased model-based tree
algorithms is presented such that each of the three strategies in GUIDE, CTree, and MOB can be
obtained by a specific combination of the available building blocks. This allows to systematically
vary the building blocks and assess the power of the resulting inference procedure. Moreover,
it is investigated whether the performance of the inference impacts the performance of the
trees differently under pre- vs.\ post-pruning.
In many of the considered scenarios the choice of goodness-of-fit measure heavily influences
the performance of testing strategies. In particular, using model scores leads to
overall clearly better results than employing residuals only. Moreover, the original
values of the goodness-of-fit measure are preferred over dichotomized versions of them.
Also regarding the effects of categorizing possible split variables and the selection
of a pruning strategy clear recommendations can be given based on the presented results.
The remainder of this manuscript is structured as follows:
Section~\ref{sec:urp} reviews unbiased tree models, starting with the general algorithmic idea
(Section~\ref{sec:generic_alg}) followed by the specific algorithms CTree (Section~\ref{sec:ctree}),
MOB (\ref{sec:mob}), and GUIDE (\ref{sec:guide}), before possible pruning strategies
(\ref{sec:pruning}) are discussed.
In Section~\ref{sec:unify} the unifying framework for testing strategies in unbiased model-based
recursive partitioning algorithms is presented.
The setting for the simulation study is introduced in Section~\ref{sec:simulation}
and the results are illustrated and discussed in Section~\ref{sec:results}.
\section{Unbiased recursive partitioning}
\label{sec:urp}
\subsection{Generic algorithm}
\label{sec:generic_alg}
The basic idea of building a regression tree model is to partition the data into smaller
and more homogeneous subgroups based on a set of covariates. Various tree
algorithms have been developed, following essentially the same general structure,
employing the covariates as \emph{split variables} in the tree induction.
Starting at the root of the tree, pertaining to the full available data sample, the
algorithms proceed in the following steps:
\begin{enumerate}
\item For the current sample a (possibly simple) model is fitted by optimizing
some objective function (or loss function) that reflects the goodness of fit.
\item Among all available split variables one is selected as the split variable,
choosing the split point such that goodness of fit is maximized in the
resulting subgroups.
\item Steps 1 and 2 are repeated within each subgroup until some stopping
criterion is attained.
\end{enumerate}
The term ``model'' is used here in a very broad sense and encompasses not
only least-squares or maximum-likelihood models but also simple constant
fits such as means or average proportions.
Therefore, depending on the type of model, the employed objective function can
for example be the sum of deviations from a typical/average value, but it can
also be based on a model for a response along with potential split variables.
For instance, different types of residuals (or the signs thereof) can be employed
\citep[see e.g.,][]{Loh:2002} as well as rank sums or logrank scores
\citep[as in][]{Hothorn+Hornik+VanDeWiel:2006}.
As explained in Section~\ref{sec:introduction} the considered tree algorithms CTree,
MOB, and GUIDE first select the split variable and then, in a separate step, the
split point. For the first step of selecting a split variable they all apply statistical
tests following the same basic strategy:
\begin{enumerate}
\item To capture how the objective function changes with the observations in
the current subgroup, a disaggregated, observation-wise goodness-of-fit
measure is obtained. More formally, this is an $N \times K$ matrix
where $N \in \mathbb{N}$ is the number of observations and $K \in \mathbb{N}$
the number of goodness-of-fit measures per observation.
\item The dependency or association of this goodness-of-fit matrix with
each possible split variable $Z_j$, $j \in \{1, \dots, J\}$, is assessed
using some suitable test statistic. The corresponding $p$-values allow for
a comparison of all $J$ split variables on a standardized or unified scale and in
that way for an unbiased split variable selection.
\item The split variable corresponding to the smallest $p$-value -- and thus the
highest influence on the goodness of fit of the model -- is selected for
splitting the data into subgroups.
\end{enumerate}
As an example -- and explained in more detail below -- consider a linear regression tree.
Thus, a linear regression model is fitted in each subgroup, minimizing the
residual sum of squares as the aggregated goodness-of-fit measure. The corresponding
observation-wise goodness-of-fit measure can be given by the residuals or the scores
(gradient contributions). Analogously, the log-likelihood and corresponding score function
could be used.
While this basic approach is the same for the GUIDE, CTree, and MOB algorithms, they differ
in their strategies on how to calculate test statistics. In order to point out these specific
characteristics in Section~\ref{sec:unify} the strategies of the three tree algorithms are first
explained in more detail in Sections~\ref{sec:ctree}, \ref{sec:mob}, and \ref{sec:guide}.
\subsection{CTree}\label{sec:ctree}
The CTree algorithm \citep{Hothorn+Hornik+Zeileis:2006} is based on the idea of providing
non-parametric regression tree models in a conditional inference framework by applying
permutation tests.
To select a split variable it is tested whether there is any association between the transformed
response $h(Y)$ and each possible transformed split variable $g(Z_j)$, $j = 1, \dots, J$.
The only requirement for the function $h$ is to depend on $Y$ in a permutation-symmetric way but
this encompasses ranks, scores, indicator functions, etc.\ and can also be multidimensional.
For a numeric response the identity function $h(Y) = Y$ is a common
choice while a categorical response can be mapped to a unity vector by an indicator function
$h(Y) = (0,\ldots,1,\ldots,0)^{\top}$.
Alternatively, the function $h$ can capture location and scale of $Y$ via
$h(Y) = (Y, (Y-\bar{Y})^2)^{\top}$.
If a parametric model is fitted to the response $Y$ with some covariate(s)~$X$ employed as
\textit{regressor} variable(s),
then a model-based transformation $h(Y) = s(Y,X,\hat{\beta})$ can be used, e.g.,
the residuals in a linear model with regression coefficients $\beta$. Moreover,
$s$ can be the score function pertaining to the objective function (or loss function)
$\ell$:
$$
s(Y, X, \beta) = \frac{\partial \ell(Y, X, \beta)}{\partial \beta}
$$
The estimate of the model parameters $\hat{\beta}$ is obtained by optimizing
the sum of the objective function $\ell$, aggregated over all observations.
This framework includes many different M-type estimators as special cases, including maximum
likelihood and ordinary least squares estimation. For a $K$-dimensional parameter
$\beta$ the score function evaluated for the $i$-th observation $s(y_i, x_i, \beta)$
is also a $K$-dimensional vector, i.e., the gradient contribution of the $i$-th observation.
Thus, the $N \times K$-matrix consisting of these scores or gradient contributions for all
$i = 1, \dots, N$ observations is a natural candidate for the observation-wise goodness-of-fit
measure as described in the previous Section~\ref{sec:generic_alg}.
Similarly, different types of functions can be chosen for the influence function $g$ depending
on a possible split variable $Z$. A simple choice in case of $Z$ being a numeric variable is again
the identity function $g(Z) = Z$. For categorical variables $g$ can also map its values to the
corresponding unity vectors by an indicator function $g(Z) = (0,\ldots,1,\ldots,0)^{\top}$.
To test for independence of $h(Y)$ and $g(Z)$ CTree calculates a linear association test statistic,
following the framework of \cite{Strasser+Weber:1999}. The corresponding conditional expectation and
covariance given all permutations of the response variable can be calculated and used to
standardize the test statistic. This standardized statistic has an asymptotic normal distribution
which is in fact multivariate if either of the transformation $h(Y)$ and/or $g(Z)$ is
multivariate. The actual test is carried out by mapping this standardized statistic to
the real line either by taking the absolute maximum or using a quadratic form -- with $p$-values
being computed from the analogous transformation of the normal distribution
\citep[see][for a hands-on introduction and Appendix~\ref{app:teststat_ctree} for more details on the linear test statistic]{Hothorn+Hornik+VanDeWiel:2006}.
If both variables $Y$ and $Z$ are numeric the default independence test corresponds to a Pearson
correlation test. For one numeric and one categorical variable essentially a one-way ANOVA
(analysis of variance) is employed while for two categorical variables a $\chi^2$~test is performed.
Thus, in the general CTree framework many types of tests can be specified by selecting suitable
transformations $g$ and $h$. While originally
conceived for nonparametric models, it is easy to adapt CTree to model-based testing and
recursive partitioning by choosing a model-based $h$ transformation as argued above
\citep[see the concrete examples in][]{Zeileis+Hothorn:2013,Seibold+Zeileis+Hothorn:2016}.
\subsection{MOB}\label{sec:mob}
In contrast to CTree the MOB algorithm \citep{Zeileis+Hothorn+Hornik:2008} was explicitly
designed for a model-based goodness-of-fit measure in order to embed parametric models
into a regression tree framework. Thus, MOB is based on an objective/loss function $\ell$
and corresponding score function $s$. The original paper considered generalized linear models
(GLMs) and survival regression models but subsequently various other models have been applied as well,
including beta regression \citep{Gruen+Kosmidis+Zeileis:2012}, psychometric item response theory
models \citep{Strobl+Kopf+Zeileis:2015}, or mixed effects models \citep{Fokkema+Smits+Zeileis:2018}.
But just like CTree can be applied to parametric models, MOB conversely also encompasses
simple regression and classification trees, e.g., by choosing an intercept-only model.
For selecting a split variable MOB employs a score-based test that relies on the central limit
theorem for the parameter estimate $\hat{\beta}$. The test assesses whether the scores
-- when ordered by the potential split variable $Z$ -- fluctuate randomly around their
zero mean or differ systematically in certain subgroups. The latter would indicate a
parameter instability that could be captured by fitting separate models (optimizing $\ell$)
in the resulting subgroups. In case of a numeric split variable $Z$ both the score-based statistic
and the partitioned objective function $\ell$ are maximized over all possible splits in $Z$
(subject to certain minimal subgroup size constraints). Unlike the CTree framework, MOB
relies on classical unconditional inference. For more details see \cite{Zeileis+Hornik:2007} and
Appendix~\ref{app:teststat_mob}.
\subsection{GUIDE}\label{sec:guide}
Building on earlier work for the QUEST algorithm \citep{Loh+Shih:1997}, \cite{Loh:2002}
proposed the GUIDE algorithm blending trees with parametric regression models and
encompassing simpler classification and regression trees as special cases. Thus,
linear regression models could be fitted in the nodes of a tree as well as
constant fits such as simple mean response~\mbox{$\hat{\beta} = \bar{y} = \frac{1}{N}\sum_{i=1}^N y_i$}.
The tests for selecting the split variable are then based on the corresponding residuals,
e.g., for a simple linear regression \citep[as used in][]{Loh:2002}:
$$
r(Y,X,\hat{\beta}) = Y - \hat{\beta_0} - \hat{\beta_1} \cdot X.
$$
In subsequent work other models together with an appropriate choice of residuals
have been applied such as in
regression trees for longitudinal and multiresponse data \citep{Loh+Zheng:2013},
quantile regression models \citep{Chaudhuri+Loh:2002},
and proportional hazards modeling via Poisson regression \citep{Loh+He+Man:2015}, among others.
To construct a statistical test two additional transformations are carried out:
(1)~The residuals $r(Y,X,\hat{\beta})$ are dichotomized at zero, yielding an indicator for positive vs.\ negative residuals.
(2)~Each possible split variable $Z$ is categorized, i.e., unless $Z$ is already categorical
it is split at its quartiles into four bins.
Subsequently, a $\chi^2$~test of independence is performed for the dichotomized residuals and
each categorized/categorical split variable.
After choosing the split variable showing the highest dependency by yielding the lowest
$p$-value, the split point minimizing the overall goodness-of-fit measure is selected.
Note that the split point in numeric variables $Z$ is searched over all possible splits,
not just the four bins that were constructed for the $\chi^2$~test. More details on the
applied test statistic can be found in Appendix~\ref{app:teststat_guide}.
\subsection{Pruning}
\label{sec:pruning}
To avoid overfitting recursive partitioning algorithms need to assure that trees do not grow
too large. Apart from certain minimal subgroup size or maximal tree depth constraints this is
classically accomplished by so-called ``pruning'' approaches. As the three tree algorithms
considered here (CTree, MOB, and GUIDE) differ in their default choice of pruning approach,
we briefly discuss these here. However, as all three algorithms can in principle be combined
with any of the pruning approaches this is done only relatively briefly.
The classical CART algorithm \citep{Breiman+Friedman+Stone:1984} proposed to first grow a large
tree and then prune those splits in the tree that did not increase predictive performance in a
cross-validation. This is also know as \emph{post-pruning} (after growing the initial tree) and
more specifically cost-complexity pruning.
In the unbiased recursive partitioning literature this post-pruning approach is also used frequently
\citep[e.g., in][]{Loh+Vanichsetakul:1988,Loh+Shih:1997,Kim+Loh:2001,Loh:2002} and the $p$-values
from the association tests are only employed for selecting the split variable on a unified scale.
However, \cite{Hothorn+Hornik+Zeileis:2006} proposed to also use these $p$-values for a
so-called \emph{pre-pruning} strategy which stops growing the tree as soon as no significant association
can be found in a given subgroup. This approach is the default in CTree and also in
MOB. However, \cite{Zeileis+Hothorn+Hornik:2008} also pointed out that a natural strategy for
post-pruning in model-based partitioning is to use information criteria such as AIC (Akaike information
criterion) or BIC (Bayes information criterion), following the ideas of \cite{Su+Wang+Fan:2004}.
Clearly, for inference-based pre-pruning it is crucial that the association tests employed for
split selection work well as statistical significance test, i.e., conform with their nominal size
and have high power. In contrast, when using post-pruning (either based on cross-validation or
information criteria) it might not be as crucial that the significance test works well and has
high power.
Due to these considerations we first evaluate the significance tests underlying CTree, MOB,
and GUIDE by themselves, i.e., without growing an actual tree in combination with a pruning strategy.
Subsequently we combine the tests with a cost-complexity post-pruning approach in order to assess
whether shortcomings of the tests are mitigated by pruning.
\section{Unifying framework}\label{sec:unify}
Each of the algorithms CTree, GUIDE, and MOB can be characterized by its combination
of the type of model fits, tests, and pruning strategy employed to
grow the tree. Table~\ref{tab:combinations} provides an overview of the default
combinations.
However, as discussed above, subsequent publications have emphasized that all three
algorithms can be combined with different model fitting approaches and to some degree
different pruning strategies have been explored as well. Thus, the class of tests
employed for the unbiased splitting variable selection forms the core of each of the
algorithms: conditional inference (CTree) vs.\ score-based fluctuation tests (MOB)
vs.\ residual-based $\chi^2$ tests (GUIDE). Therefore, we consider a standard class
of model fits (namely, linear regression trees) and investigate the relative
advantages and disadvantages of the tests themselves (which are essential to pre-pruning)
as well as the combination with post-pruning. Thus, subsequently
the names CTree, MOB, and GUIDE distinguish the test-based variable and split selection
rather than the entire algorithm with all default settings.
\begin{table}[t!]
\centering
\begin{tabular}{ l l l l }
\hline\noalign{\smallskip}
& Fit & Test & Pruning \\
\noalign{\smallskip}\hline\noalign{\smallskip}
CTree & Non-parametric & Conditional inference & Pre\\
MOB & Parametric & Score-based fluctuation & Pre (or post with AIC/BIC)\\
GUIDE & Parametric & Residual-based $\chi^2$ & Post (cost-complexity pruning)\\
\noalign{\smallskip}\hline
\end{tabular}
\caption{Default combinations of fitted model type, test type and pruning strategy
for the algorithms CTree, MOB, and GUIDE.
\label{tab:combinations}}
\end{table}
\subsection{Building blocks of testing strategies}\label{sec:buildingblocks}
Even though CTree, MOB, and GUIDE differ in the specific tests they apply, their approaches for split
variable selection follow the same basic structure as explained in Section~\ref{sec:generic_alg}.
In fact, the tests can be embedded in a unifying conceptual framework that yields the
different tests by combining various building blocks.
These mostly differ in the way the model for the dependent variable on the one hand and
the splitting variables on the other are prepared or transformed:
\begin{itemize}
\item \emph{Goodness-of-fit measure:}\\
Different variations of the disaggregated, observation-wise goodness-of-fit measure of the
model for the response $Y$ and possible regressors $X$ can be considered. Either residuals
$r(Y,X,\hat{\beta})$ can be used as proposed for GUIDE or model scores $s(Y,X,\hat{\beta})$
as proposed for MOB. All three algorithms can, in principle, use both goodness-of-fit
measures though, which is probably brought out most clearly in the CTree algorithm that
explicitly allowed for different transformations $h(Y)$ in its original description already.
\item \emph{Dichotomization of residuals/scores:}\\
Rather than testing independence between the split variables and the residuals/scores
themselves, it is possible
to dichotomize residuals/scores at $0$ so that only their signs are assessed (as proposed for GUIDE).
\item \emph{Categorization of split variables:}\\
Similarly, the split variables can also be categorized (for testing only). This was proposed for
GUIDE, employing binning at the quartiles yielding four categories of approximately
equal size.
\end{itemize}
The three algorithms combine these building blocks in different ways as shown in Table~\ref{tab:buildingblocks}:
When applying CTree for unbiased model-based recursive partitioning it has been suggested to use
the model scores without dichotomization and assess their association with the untransformed split
variables using a conditional inference test. This is similar to a squared correlation test statistic.
MOB also employs the scores without
dichotomization and maximally selects a score statistic over all potential split points in the
split variable. GUIDE employs the dichotomized residuals and assesses their association with
the categorized split variable in a classical (unconditional) $\chi^2$ test.
But the building blocks could be easily re-combined to yield new types of tests. For example,
in the GUIDE approach, a one-way ANOVA can be used for assessing the association of the residuals
(without dichotomization) with the categorized split variable. Or alternatively, a multivariate
one-way ANOVA can be used for the non-dichotomized scores as opposed to the residuals etc.
Note that there are further differences in the testing strategies between the three algorithms, e.g.,
using conditional vs.\ unconditional approximations of the null distributions. However, this difference
has relatively little influence compared to the other building blocks considered in detail here.
Moreover, both similarities and relative differences between these approaches have been previously
discussed, e.g., in \cite{Hothorn+Zeileis:2008} and \cite{Zeileis+Hothorn:2013}.
\begin{table}[t!]
\centering
\begin{tabular}{ l l l l l }
\hline\noalign{\smallskip}
& Scores & Dich. & Cat. & Statistic\\
\noalign{\smallskip}\hline\noalign{\smallskip}
CTree & Model scores & -- & -- & Sum of squares\\
MOB & Model scores & -- & -- & Maximally selected\\
GUIDE & Residuals & \checkmark & \checkmark & Sum of squares\\
\noalign{\smallskip}\hline
\end{tabular}
\caption{Testing strategies of CTree, MOB, and GUIDE with the corresponding
setting of the building blocks in the unifying framework and the type of test statistic.
\label{tab:buildingblocks}}
\end{table}
\subsection{Linear model tree}\label{sec:lmtrees}
To focus on the unified testing framework, as described in the previous section, we employ
the same model fits for all three algorithms. To do so, we employ linear regression models
because it is such a basic and widely used model and linear model trees were the leading
illustrations in both the original MOB \citep{Zeileis+Hothorn+Hornik:2008} and GUIDE \citep{Loh:2002} papers.
However, the conclusions drawn from this example also hold for many other model types.
To fix notation, we consider the following models for the simulation
study in Section~\ref{sec:simulation}:
$$
Y = \beta_0 + \beta_1 \cdot X + \epsilon
$$
with response variable $Y$, regressor variable $X$, and error term $\epsilon$.
In particular, in the investigated tree models the coefficients $\beta_0$ and $\beta_1$
can depend on the possible split variables $Z_j$, $j=1,\ldots,J$, such that
$$
Y = \beta_0(Z_1, \ldots, Z_J) + \beta_1(Z_1, \ldots, Z_J) \cdot X + \epsilon.
$$
This model is fitted, as usual, by ordinary least squares (OLS) to the observations in each notation,
yielding the parameter estimates $\hat{\beta} = (\hat{\beta_0}, \hat{\beta_1})^\top$.
To keep notation simple we present the following equations for the root node with all observations
$\{(y_i,x_i)\}_{i=1,\ldots,N}$ and $N \in \mathbb{N}$. In subsequent nodes the same equations
are used but just for a smaller subgroup.
The aggregated goodness-of-fit measure (or objective or loss function) in OLS estimation
is the sum of squared residuals: $\sum_{i=1}^N \ell(y_i,x_i,\beta_0,\beta_1)$ where
$$
\ell(y_i,x_i,\beta_0,\beta_1) = r(y_i,x_i,\beta_0,\beta_1)^2
$$
is the squared residual which is defined as
$$
r(y_i,x_i,\beta_0,\beta_1) = y_i - \beta_0 - \beta_1 \cdot x_i.
$$
These residuals can also intuitively be used as the corresponding disaggregated
observation-wise goodness-of-fit measure. Another natural candidate for this is the
score function for the $i$-th observation:
$$
s(y_i, x_i, \beta_0, \beta_1) =
\frac{\partial \ell}{\partial (\beta_0, \beta_1)^{\top}}(y_i, x_i, \beta_0, \beta_1) =
-2 \cdot r(y_i,x_i, \beta_0, \beta_1) \cdot (1,x_i)^\top
$$
Thus, up to a constant scaling factor of $-2$ (that could also be omitted) the first
component of the scores is in fact the residual.
However, it is complemented by a second component
that captures the slope effect of $x_i$. Therefore, when computing the score for all $N$ observations
this yields an $N \times 2$ score matrix whose first column corresponds to the $N$ residuals.
As this is the derivative with respect to the intercept parameter~$\beta_0$, it
captures changes in the intercept. Moreover, the second column of the score matrix contains
derivatives with respect to the slope parameter~$\beta_1$ and thus captures changes in this.
Hence, tests based on the full scores (as in CTree and MOB) include residual-based tests (as
in GUIDE) as a special case. Therefore, score-based tests can in principle capture all changes
in the objective function that residual-based tests can capture -- but the reverse is not
necessarily true. In the next sections we will investigate how relevant this is in practice
and how much it depends on the concrete test statistics employed.
\section{Simulation setting and evaluation}
\label{sec:simulation}
In this simulation study two different scenarios are considered for the linear model trees
presented in Section~\ref{sec:lmtrees}. First, the underlying tree structure based on which the
data is generated is a stump, i.e., a tree with only one split (``stump'' scenario, see
Figure~\ref{fig:dgp_stump}).
By keeping the tree so simple we can focus on the testing strategy only giving focus to their
power in terms of selecting the correct split variable.
In the second scenario the true tree structure contains two splits in two different variables
yielding a tree with three terminal nodes (``tree'' scenario, see Figure~\ref{fig:dgp_tree}).
It employs the same basic structure as the first scenario but simply adds another split. This
allows to evaluate the power of the three testing strategies in a more complex setting, in
combination with using a post-pruning strategy.
\subsection{Data generating process}
\subsubsection{``Stump'' scenario}
Each generated data set consists of the response and regressor variables, one true split
variable and nine noise split variables as listed in Table~\ref{tab:variables}
together with the corresponding distributions.
\begin{table}[t!]
\centering
\begin{tabular}{l l l}
\hline\noalign{\smallskip}
Name & Notation & Specification \\
\noalign{\smallskip}\hline\noalign{\smallskip}
\textit{Variables:} & & \\
Response & $Y$ & $=\beta_0(Z_1) + \beta_1(Z_1) \cdot X + \epsilon$ \\
Regressor & $X$ & $\mathcal{U}([-1,1])$ \\
Error & $\epsilon$ & $\mathcal{N}(0,1)$ \\
True split variable & $Z_1$ & $\mathcal{U}([-1,1])$\\
Noise split variables & $Z_2$--$Z_{10}$ & $\mathcal{U}([-1,1])$ or $\mathcal{N}(0, 1)$ \\
& & (alternating) \\
\noalign{\smallskip}\hline\noalign{\smallskip}
\multicolumn{3}{l}{\textit{Parameters/functions:}} \\
Intercept & $\beta_0$ & $0$ or $\pm \delta$\\
Slope & $\beta_1$ & $1$ or $\pm \delta$\\
True split point & $\xi$ & $\in \{0, 0.2, 0.5, 0.8\}$ \\
Effect size & $\delta$ & $\in \{0, 0.1, 0.2, \ldots , 1\}$ \\
\noalign{\smallskip}\hline
\end{tabular}
\caption{\label{tab:variables}Variables included in the data generating process
as used for the ``stump'' scenario. In the ``tree'' scenario $Z_2$ is also a true split variable,
not only $Z_1$, and hence $\beta_0$ and $\beta_1$ are functions $\beta_{k-1}(Z_1,Z_2)$, $k=1,2=K$.}
\end{table}
The location parameter $\mu$ of the normally distributed response variable $Y$ depends
linearly on the regressor variable $X$.
Moreover, the intercept $\beta_0$ and/or the slope parameter $\beta_1$
can depend on the true split variable $Z_1$. More specifically, three different variations
are considered for the coefficients $\beta_0$ and $\beta_1$:
\begin{enumerate}
\item The intercept $\beta_0$ varies depending on $Z_1$ while $\beta_1$ is fixed (at $1$).
\item The slope coefficient $\beta_1$ varies depending on $Z_1$ while $\beta_0$ is fixed
(at $0$).
\item Both coefficients $\beta_0$ and $\beta_1$ vary depending on $Z_1$.
\end{enumerate}
See Figure~\ref{fig:dgp_stump} for an illustration. More precisely, the coefficient that
changes~($\beta_{k-1}$ with $k=1,2=K$) switches between two values at the split point $\xi$:
\begin{align*}
\beta_{k-1}(Z_1) = \begin{cases}
-\delta \cdot (-1)^{k-1} \quad \text{if } Z_1 \leq \xi \\
+\delta \cdot (-1)^{k-1} \quad \text{if } Z_1 > \xi
\end{cases}
\end{align*}
In that way, the type of variation is the same for $\beta_0$ and $\beta_1$, however, in
opposite directions.
\begin{figure}[p!]
\setkeys{Gin}{width=1\linewidth}
\minipage{0.45\textwidth}
\includegraphics{draft-dgp_stump1}
\endminipage
\minipage{0.45\textwidth}
\includegraphics{draft-dgp_stump2}
\endminipage
\minipage{0.45\textwidth}
\includegraphics{draft-dgp_stump3}
\endminipage
\minipage{0.45\textwidth}
\includegraphics{draft-dgp_stump4}
\endminipage
\caption{\label{fig:dgp_stump}{Top left panel: True stump structure applied in the data generating process for the ``stump'' scenario with the true model parameters $\beta_0$
and $\beta_1$, either varying or being fixed at $0$ or $1$ respectively.
Top right and bottom panels: Bivariate plot of the response $Y$ on regressor $X$
with effect size $\delta = 1$ and for three variations:
varying $\beta_0$ and fixed $\beta_1$ (top right),
fixed $\beta_0$ and varying $\beta_1$ (bottom left),
varying $\beta_0$ and varying $\beta_1$ (bottom right).}}
\end{figure}
\begin{figure}[p!]
\begin{center}
\setkeys{Gin}{width=1\linewidth}
\minipage{0.5\textwidth}
\includegraphics{draft-dgp_tree1}
\endminipage
\minipage{0.45\textwidth}
\includegraphics{draft-dgp_tree2}
\endminipage
\caption{\label{fig:dgp_tree}Left panel: True tree structure
applied in the data generating process for the ``tree'' scenario
with the true model parameters $\beta_0$ and $\beta_1$ for each terminal panel.
Right panel: Bivariate plot of the response $Y$ on regressor $X$
with effect size $\delta = 1$.}
\end{center}
\end{figure}
\subsubsection{``Tree'' scenario}
For the second simulation scenario (``tree'' scenario, see Figure~\ref{fig:dgp_tree}) the same basic setting with the same
variables as in the ``stump'' scenario is applied (see Table~\ref{tab:variables}).
However, not only $Z_1$ but also $Z_2$ is used as a true split variable following a uniform distribution
on $[-1,1]$. Therefore, an additional split
is preformed yielding a tree with three terminal nodes. The first split (in $Z_2$, at~$\xi$) induces a
change in the slope parameter $\beta_1$ while the second split (in $Z_1$, also at~$\xi$) corresponds
to a change in the intercept $\beta_0$.
Hence, contrary to the ``stump'' scenario where three variations are considered for the parameters
$\beta_0$ and $\beta_1$, only this one variation with both parameters varying is investigated for
the ``tree'' scenario. In particular, the parameters depend on the split variables in the following way:
\begin{align*}
\beta_{0}(Z_1, Z_2) = \begin{cases}
0 \qquad \text{if } Z_2 \leq \xi \\
-\delta \quad \text{ if } Z_1 \leq \xi \land Z_2 >\xi \\
+\delta \quad \text{ if } Z_1 > \xi \land Z_2 >\xi
\end{cases}
\end{align*}
\begin{align*}
\beta_{1}(Z_1, Z_2) = \beta_{1}(Z_2)= \begin{cases}
+\delta \quad \text{if } Z_2 \leq \xi \\
-\delta \quad \text{if } Z_2 > \xi
\end{cases}
\end{align*}
\subsection{Evaluation}
\label{sec:evaluation}
The testing strategies are evaluated over a stepwise increasing effect size $\delta$,
on 100 replications per step each consisting of 250 observations.
To compare the performance of the evaluated testing strategies for each step
in the ``stump'' scenario the following criteria are considered:
the $p$-values pertaining only to the true split variable $Z_1$;
and the proportion of replications for which the $p$-value of $Z_1$
is the lowest and significant at 5\% level (i.e., where $Z_1$ would be
selected for splitting in a pre-pruning approach), denoted by the
``selection probability''. In that way, the power of the
considered statistical tests can be compared as they are all applied
as significance tests answering two questions at once:
(1)~Should a split be performed at all?
(2)~If so, in which variable?
For the first question the $p$-values regarding all available split variables are compared
to a predefined level of significance $\alpha = 0.05$. Only if the smallest $p$-value is smaller
than $\alpha$ a split is performed and the split variable corresponding to this $p$-value is selected.
For the ``tree'' scenario the adjusted Rand index~(ARI) is calculated
as a measure of similarity between the true tree structure and the fitted model tree.
As explained before, the aim of this simulation study is to investigate the effects
of each particular building block of the unifying framework presented in Section~\ref{sec:unify}
rather than the whole testing strategies. Therefore, other combinations as presented in
Table~\ref{tab:buildingblocks}, hence adapted versions of GUIDE, CTREE, and MOB are evaluated as well
as their original versions.
\section{Results}
\label{sec:results}
In the following we first investigate the properties of the testing
strategies in the ``stump'' scenario, focusing on the tests' $p$-values
for the true split variable $Z_1$ and its corresponding selection
probability (i.e., the association of $Y$ and $Z_1$ being significant at 5\% level
and having the lowest $p$-value among all split variables). Section~\ref{sec:scores}
begins by highlighting the importance of using full model scores vs.\ residuals only
before Section~\ref{sec:3way} considers all building blocks (scores vs.\ residuals,
dichotomization of these, and categorization of the split variable).
Subsequently, the ``tree'' scenario is employed to investigate how the
performance of the tests affects growing the trees overall. This is
evaluated using the adjusted Rand index for trees grown by pre-pruning and
cost-complexity post-pruning (Section~\ref{sec:tree}).
\newpage
\subsection[``Stump'' scenario: Residuals vs. full model scores]{``Stump'' scenario: Residuals vs.\ full model scores}
\label{sec:scores}
A crucial difference between the testing strategies of CTree/MOB and GUIDE
is the difference between using only the residuals vs.\ the full model
scores. In the literature on structural change tests it is well-established
that residual-based tests can only capture parameter differences that
affect the conditional mean \citep[see e.g.,][ and further discussion in
Section~\ref{sec:discussion}]{Ploberger+Kraemer:1992}.
Hence we compare the CTree, MOB, and GUIDE algorithms -- all three using the
default specification as shown in Table~\ref{tab:buildingblocks} -- and
additionally consider a new GUIDE flavor, denoted GUIDE+scores, that uses
dichotomized scores rather than residuals.
\begin{figure}[t!]
\setkeys{Gin}{width=\linewidth}
\includegraphics{draft-scores_stump}
\caption{\label{fig:scores_stump}
Selection probability of the true split variable $Z_1$ for
testing strategies CTree, MOB, GUIDE, and GUIDE+scores in the ``stump''
scenario. Probabilities are estimated over increasing effect size $\delta$
with 100 replications of 250 observations per step. The true split point
in $Z_1$ is either the median 0 (top) or the 90\% quantile 0.8 (bottom)
for either varying intercept $\beta_0$ (left), varying slope $\beta_1$
(middle), or both coefficients varying (right).}
\end{figure}
In Figure~\ref{fig:scores_stump} the performances of the testing strategies are
represented by the corresponding selection probability, i.e., a significance level
is incorporated. (Appendix~\ref{app:significance} shows that the same qualitative
conclusions can be drawn when the significance level is \emph{not} included.)
For a split point at the median ($\xi=0$, top row) all testing strategies
perform similarly well as long as the intercept varies (left and right panel) with CTree and
MOB being only slightly ahead. However, for the split point at the $90$\% quantile $\xi = 0.8$
(bottom row) the performance of all of the applied strategies decreases.
While both GUIDE versions struggle to detect the correct split variable even for a high effect
size, MOB is clearly ahead with CTree leading to the second best results.
This advantage of MOB over CTree is mainly due to the abrupt shift in the
model parameters $\beta_0$ and $\beta_1$ and turns into an advantage of CTree over MOB for
a smooth transition with continuously-changing parameters (see Appendix~\ref{app:contbeta}).
However, in the scenario where only $\beta_1$ (but not the intercept $\beta_0$) is affected by the split,
the residual-based GUIDE approach has no power at all even for a true split at the median (top middle panel).
It is easily possible, though, to substantially mitigate this problem by using scores
(sensitive to changes in all parameters) rather than residuals only (sensitive to changes
in the conditional mean). The remaining difference between MOB/CTree and GUIDE+scores
is due to dichotomizing the scores at zero and due to categorizing the split variables
which are investigated in more detail in the following section.
\subsection{``Stump'' scenario: Full factorial analysis of building blocks}\label{sec:3way}
To investigate the impact of each of the building blocks separately the most general
case of the ``stump'' scenario where the intercept and slope parameter are both varying is
considered. In this evaluation all possible combinations of the building blocks
have been included. The different levels of each of the three building blocks are listed in
Table~\ref{tab:levels} where $h$ and $g$ refer to the transformation functions applied to the
response $Y$ or a split variable $Z$ respectively, both as described in Section~\ref{sec:ctree}.
In the case of categorization, the split variable $Z$ is binned at the
quartiles. This corresponds to a four-dimensional 0/1 transformation
function $g(Z)$ that indicates into which of the bins each observation falls.
Maximum selection across potential split points in a variable $Z$ also
corresponds to a multivariate 0/1 transformation function $g(Z)$. However,
in this case for each potential split point an indicator is used that is 0
before and 1 after the respective split point.
(See also Table~\ref{tab:combinations_appendix} in Appendix~\ref{app:combinations}
for a more detailed overview of all 12 combinations of the building blocks.)
\begin{table}
\centering
\begin{tabular}{ l l l }
\hline\noalign{\smallskip}
Building block & Levels & Transformation \\
\noalign{\smallskip}\hline\noalign{\smallskip}
Residuals vs.\ scores & \textit{residuals} & $h(Y) = r(Y,X,\hat{\beta})$\\
($h$ transformation) & \textit{scores} & $h(Y) = s(Y,X,\hat{\beta})$\\
\noalign{\smallskip}\hline\noalign{\smallskip}
Dichotomization & \textit{yes} & $\mathds{1}_{[0,\infty)}(h)$\\
(of $h$) & \textit{no} & $h$ without\\
& & \quad dichotomization\\
\noalign{\smallskip}\hline\noalign{\smallskip}
Categorization & \textit{cat} & $g(Z)=(0,1,0,0)^{\top}$\\
($g$ transformation) & & \quad indicating the\\
& & \quad assigned bin\\
& \textit{max} & $g(Z)=(0,\ldots,0,1,\ldots,1)^{\top}$\\
& & \quad indicating the\\
& & \quad potential split point\\
& \textit{lin} & $g(Z)=Z$\\
\noalign{\smallskip}\hline
\end{tabular}
\caption{Available levels of the building blocks.
\label{tab:levels}}
\end{table}
Based on the results displayed in Figure~\ref{fig:3way} it can be stated that dichotomizing
residuals/scores decreases the performance as it leads to higher $p$-values for
the true split variable $Z_1$ and thus to lower power in the settings considered.
For a true split at the median ($\xi=0$, left panel) this effect is almost constant
across the three types of categorization considered. However, in case
of the true split point at the $90$\% quantile $\xi = 0.8$ (right panel), categorizing split
variables increases the $p$-value of $Z_1$ even more.
Not surprisingly, maximum selection is most advantageous in this case (i.e., for a late abrupt shift)
which is harder to detect based on a linear statistic. But overall it depends on the situation
whether a linear or a maximum selection of a split variable leads to lower $p$-values.
Comparing the two panels suggests that a categorization weakens the performance of the
tests unless the true split point is close to one of the breaks from the binning
(as in the left panel).
Moreover, while the effect of both transformations (dichotomization of residuals/scores and
categorization of split variables) can be observed separately in Figure~\ref{fig:3way},
combining them increases the negative impact remarkably.
As already shown in Section~\ref{sec:scores} the use of scores vs.\ residuals has a minor
effect if there is a change in the intercept which is
supported by the small differences between the dashed lines (scores) and the solid
lines (residuals) in Figure~\ref{fig:3way}.
Note that the effect size $\delta$ in the results from Figure~\ref{fig:scores_stump} has been chosen
so that $p$-values for the non-dichotomized tests are roughly comparable: $\delta = 0.3$
for the true split point at the median $\xi = 0$ (left) vs.\ a stronger effect of $\delta = 1$
for the true split point at the 90\% quantile $\xi = 0.8$ (right).
Additional evaluations for varying effect size $\delta$ can be found in Appendix~\ref{app:increasingdelta}.
\begin{figure}[t!]
\setkeys{Gin}{width=1\linewidth}
\includegraphics{draft-3way}
\caption{\label{fig:3way} Effects of the three building blocks \textit{residuals vs.\ scores},
\textit{dichotomization}, and \textit{categorization} in the ``stump'' scenario with the intercept
and the slope parameter both varying. All possible combinations have been
evaluated in two different settings and their performances are compared based on the
mean $p$-values corresponding to the true split variable $Z_1$.
Left panel: true split point at the median $0$ and with
effect size $0.3$.
Right panel: true split point at the $90$\% quantile $0.8$ and with effect
size $1$.}
\end{figure}
\subsection[``Tree'' scenario: Pre-pruning vs. post-pruning]{``Tree'' scenario: Pre-pruning vs.\ post-pruning}
\label{sec:tree}
So far the testing strategies underlying the different tree algorithms have only been considered
as classical significance tests, i.e., in terms of power and $p$-values. However, one could argue
that for a tree this is practically not really relevant -- at least when combined with a post-pruning
strategy such as cost-complexity pruning \citep{Breiman+Friedman+Stone:1984}. In the latter case it only matters that the relevant split
variables have the lowest $p$-value among all potential split variables -- but it is irrelevant whether
this is significant or not. To investigate to which extent this is actually true we evaluate the
different tree algorithms in the more complex ``tree'' scenario (see Figure~\ref{fig:dgp_tree}):
once with significance-based pre-pruning and once with cost-complexity post-pruning.
Recall that the true split structure is composed of splits in two different variables, both at
the same split point $\xi$. First, the split in $Z_2$ changes the slope from $+\delta$ to
$-\delta$. Second, the split in $Z_1$ changes the intercept in the negative slope group from
$-\delta$ to $+\delta$. In the simulation the effect size $\delta$ is increased from 0 to 1
for different split points $\xi$ from 0 to 0.8. Here, the performance is not evaluated in terms
of test properties but only in terms of tree properties, namely the adjusted Rand index (ARI)
in comparison to the true partition of the data.
\begin{figure}[t!]
\setkeys{Gin}{width=\linewidth}
\includegraphics{draft-tree_all}
\caption{\label{fig:tree_all}
Adjusted Rand index (ARI) for the testing strategies CTree, MOB, GUIDE, and GUIDE+scores in the
``tree'' scenario, once applying pre-pruning (top row) and once post-pruning (bottom row).
The ARI is evaluated over
increasing effect size $\delta$ with 100 replications of 250 observations per step.
Each panel represents a different value of the true split point $\xi$ ($0$, $0.2$, $0.5$ and
$0.8$ from left to right, corresponding to the $50$\%, $60$\%, $75$\%, and $90$\% quantile).
}
\end{figure}
The top row of Figure~\ref{fig:tree_all} presents the results for pre-pruning which, not surprisingly,
reflect the results from previous sections. Thus, CTree and MOB perform similarly and rather well
as the underlying significance tests have good power properties. Only for late abrupt shifts the
linear test statistics in CTree clearly have lower power than the maximally-selected statistics in MOB.
Again, this picture would reverse for smooth changes instead of abrupt shifts. Compared to CTree/MOB
both GUIDE flavors clearly perform worse as the underlying tests are less powerful; notably
for the residual-based GUIDE which again has problems picking up the slope change associated with $Z_2$.
When switching from pre-pruning to post-pruning (bottom row of Figure~\ref{fig:tree_all})
it is indeed shown that many of the problems stemming from the low power of the two GUIDE
flavors are indeed mitigated. Thus, by first growing a large tree without stopping upon
non-significance and then pruning back based on predictive performance instead substantially
improves the fit of the two GUIDE flavors. However, using the residuals only in GUIDE still
performs clearly worse compared to the GUIDE+scores. The latter is essentially on par with
CTree and MOB when the true split matches one of the categorization bins (i.e., $\xi = 0$
and $\xi = 0.5$, respectively) while CTree/MOB still perform somewhat better in the other
cases ($\xi = 0.2$ and $\xi = 0.8$).
In summary, there is clear support for the conventional wisdom that the power of the testing
strategy in an unbiased tree algorithm is not so important when combined with post-pruning.
However, there are limits to this. Consequently, when tests can be made more powerful --
e.g., by using scores instead of residuals -- then this improves the corresponding tree
algorithm. Finally, the simulation also supports using pre-pruning in an unbiased tree algorithm
when the underlying testing strategy also works well as a classical significance test.
\section{Discussion}
\label{sec:discussion}
The testing strategies underlying the unbiased recursive partitioning algorithms
CTree, MOB, and GUIDE have been embedded in a common inference framework, highlighting
what the tests have in common and what sets them apart. Concerning the effects
of the corresponding building blocks for the tests, three main conclusions can be drawn:
\begin{itemize}
\item \emph{Goodness-of-fit measure:} Assessing all dimensions of a model via the full
scores is to be preferred over assessing only a subset with the residuals. The former
can substantially improve performance while leading only to minor deteriorations
when it is not necessary.
\item \emph{Dichotomization of residuals/scores:} No scenarios could be found where this
is beneficial and tests without dichotomization performed clearly better in several
scenarios.
\item \emph{Categorization of split variables:} The effects of categorization are not so
clear-cut and depend on the true data structure. If there are indeed abrupt shifts
close to the breaks from the categorization, it works well. However, for splits
close to the margins performance can deteriorate and a maximally-selected test is
preferable. Finally, a linear statistic performs better for smooth rather than
abrupt changes.
\end{itemize}
Thus, also when categorization as in GUIDE is used we would recommend to employ the non-dichotomized
scores instead of the dichotomized residuals. Note that such a test corresponds to a
multiple ANOVAs (for the score components) and is easily available in statistical software.
Moreover, in the \proglang{R} system the \pkg{coin} package provides a convenient toolbox that
encompasses multivariate transformation functions $h$ and/or $g$.
Linear models as presented in this study have been chosen as they are highly relevant in practice,
allow for simple illustrations, and theoretical insights are available for the testing strategies.
However, the results can be easily extended to a wide variety of other models where the introduced building
blocks can be applied in the same way. Specifically, it has been shown theoretically that certain
changes in the parameters do not lead to shifts in the residuals (but in other components of the
scores). \cite{Ploberger+Kraemer:1992} showed that residual-based tests can detect a change in the parameters
of a linear model only if it also causes a shift in the expected value $\mathbb{E}(Y)$.
This is not the case if changes are orthogonal to the mean regressor which in our case is $(1,0)^{\top}$.
Consequently, if only the slope $\beta_1$ but not the intercept $\beta_0$ changes, the shift
is of type $(0, \delta)^\top$ and thus orthogonal to the mean regressor. Due to this residual-based
tests as in GUIDE break down and do not have power to detect this. Note that this situation does not
have to be rare in practice: Especially for binary regressor variables (e.g., as in treatment-subgroup
investigations) it can easily occur \citep[see Figure~2 in][for an illustration]{Loh+He+Man:2015}.
Also in more general models, residual-based and score-based procedures are expected to perform
equally well if all model parameters are highly correlated. But if parameters do change orthogonally
this might again be missed when only considering residuals -- and full model scores are typically
easily available as the appropriate remedy. Note that the score function~$s$ can also
simply be seen as a transformation of the response variable $Y$ (and potentially regressors $X$)
to a different space in order to allow for a well structured analysis of dependencies. This has
been exploited in several tree-based approaches previously published in the literature,
e.g., in \cite{Hothorn+Zeileis:2017} and \cite{Schlosser+Hothorn+Stauffer:2019}. Similarly, it would
be of interest to investigate a score-based version of the extended GUIDE algorithms beyond the
linear model, e.g., in \citealp{Loh+Zheng:2013}, \citealp{Chaudhuri+Loh:2002}, and \citealp{Loh+He+Man:2015}.
\section*{Computational details}
\label{sec:comp}
The applied implementation is based on the \proglang{R} package \pkg{partykit}
(version~1.2.4) which is available
on \proglang{R}-Forge at \url{https://R-Forge.R-project.org/projects/partykit/}.
The code to reproduce the simulation study is available in the supplement for this
paper on arXiv.org E-Print Archive (\url{https://arxiv.org/}).
The functions \code{ctree} and \code{mob} provide an implementation of the
two tree algorithms in their original form. For their adapted versions additional
modifications have been applied within these functions allowing for a categorization
of possible split variables and a dichotomization of scores.
To evaluate the GUIDE algorithm a reimplementation of this algorithm has been
built using the basic framework of \code{ctree} and \code{mob}.
\section*{Acknowledgments}
Torsten Hothorn received funding from the Swiss National Science
Foundation, grant number 200021\_184603.
|
\section{Introduction}
We consider a $d$-dimensional diffusion process defined by the following stochastic differential equation,
\begin{align*}
\mathrm{d}X_{t}=b\left(X_{t},\beta\right)\mathrm{d}t+a\left(X_{t},\alpha\right)\mathrm{d}w_{t},\ X_{-\lambda}=x_{-\lambda},
\end{align*}
where $\lambda> 0$, $\left\{w_{t}\right\}_{t\ge -\lambda}$ is a standard $r$-dimensional Wiener process, $x_{-\lambda}$ is an $\mathbf{R}^{d}$-valued random variable independent of $\left\{w_{t}\right\}_{t\ge -\lambda}$, $\alpha\in\Theta_{1}$ and $\beta\in\Theta_{2}$ are unknown parameters, $\Theta_{1}\subset\mathbf{R}^{m_{1}}$ and $\Theta_{2}\subset\mathbf{R}^{m_{2}}$ are compact and convex parameter spaces, $a:\mathbf{R}^{d}\times\Theta_{1}\to\mathbf{R}^{d}\otimes\mathbf{R}^{r}$ and $b:\mathbf{R}^{d}\times\Theta_{2}\to\mathbf{R}^{d}$ are known functions.
Our concern is statistical estimation for $\alpha$ and $\beta$ from observation.
$\theta_{\star}=\left(\alpha_{\star},\beta_{\star}\right)$ denotes the true value of $\theta:=\left(\alpha,\beta\right)$.
We denote the observation as the discretised process $\left\{\overline{X}_{ih_{n},n}:i=0,\ldots,n\right\}$ with discretisation step $h_{n}>0$ such that $h_{n}\to0$ and $T_{n}:=nh_{n}\to\infty$, where the convoluted process $\left\{\overline{X}_{t,n}\right\}_{t\ge0}$ is defined as
\begin{align*}
\overline{X}_{t,n}:=\int_{t-\overline{\rho}h_{n}}^{t}V_{h_{n}}\left(t-s\right)X_{s}\mathrm{d}s=\int_{\mathbf{R}}V_{h_{n}}\left(t-s\right)X_{s}\mathrm{d}s=\left(V_{h_{n}}\ast X\right)\left(t\right),
\end{align*}
where $V_{h_{n}}$ is an $\mathbf{R}^{d}\otimes\mathbf{R}^{d}$-valued kernel function whose support is a subset of $\left[0,\overline{\rho}h_{n}\right]$, and $\overline{\rho}>0$ such that $\sup_{n}\overline{\rho}h_{n}\le\lambda$. In this paper, we specify
$V_{h_{n}}=V_{\rho,h_{n}}$ which is a parametric kernel function whose support is a subset of $\left[0,\overline{\rho}h_{n}\right]$ defined as
\begin{align*}
V_{\rho,h_{n}}^{\left(i,j\right)}\left(t\right):=\begin{cases}
\left(\rho^{\left(i\right)}h_{n}\right)^{-1}\mathbf{1}_{\left[0,\rho^{\left(i\right)}h_{n}\right]}\left(t\right) & \text{if }i=j\text{ and }\rho^{\left(i\right)}>0,\\
\delta\left(t\right) & \text{if }i=j\text{ and }\rho^{\left(i\right)}=0,\\
0 &\text{if }i\neq j,
\end{cases}
\end{align*}
$\delta\left(t\right)$ is the Dirac-delta function, $\rho=\left[\rho^{\left(1\right)},\ldots,\rho^{\left(d\right)}\right]^{T}\in\Theta_{\rho}:=\left[0,\overline{\rho}\right]^{d}$ is the smoothing parameter determining the smoothness of observation. That is to say, the observed process is defined as follows:
\begin{align*}
\overline{X}_{ih_{n},n}^{\left(\ell\right)}=\begin{cases}
\left(\rho^{\left(\ell\right)}h_{n}\right)^{-1}\int_{\left(i-\rho^{\left(\ell\right)}\right)h_{n}}^{ih_{n}}X_{s}^{\left(\ell\right)}\mathrm{d}s&\text{ if }\rho^{\left(\ell\right)}>0,\\
X_{ih_{n}}^{\left(\ell\right)}&\text{ if }\rho^{\left(\ell\right)}=0,
\end{cases}
\end{align*}
for all $\ell=1,\ldots,d$.
Let us consider both the problems that (i) $\rho$ is a known parameter, and
(ii) $\rho$ is an unknown one
and this is estimated by observation $\left\{\overline{X}_{ih_{n},n}\right\}$, and the parameter space is denoted as $\Xi:=\Theta_{\rho}\times\Theta$.
When assuming $\rho$ as a known parameter, we can find researches for parametric estimation for $\alpha$ and/or $\beta$ based on observation schemes which can be represented as special cases for some specific $\rho$. If $\rho=\mathbf{0}$, our scheme is simply equivalent to parametric inference based on discretely observed diffusion processes $\left\{X_{ih_{n}}:i=0,\ldots,n\right\}$ studied in \citet{Florens-Zmirou-1989, Yoshida-1992, Bibby-Sorensen-1995, Kessler-1997, Kessler-Sorensen-1999, Yoshida-2011, Uchida-Yoshida-2012, Uchida-Yoshida-2014} and references therein. If $\rho=\left[1,\ldots,1\right]^{T}$, we can regard the problem as parametric estimation for integrated diffusion processes discussed in \citet{Gloter-2000, Ditlevsen-Sorensen-2004, Gloter-2006, Gloter-Gobet-2008, Sorensen-2011}. Even for the case $\rho=\left[0,\ldots,0,1,\ldots,1\right]^{T}$ where some axes correspond to direct observation and the others do to integrated observation, we give consistent estimators for $\alpha$ and $\beta$ by considering the scheme of convolutionally observed diffusion processes and this is one of the contributions of our study.
What is more, our contribution is to consider the scheme where $\rho$ is unknown and succeed in representation of the microstructure noise which makes the observation smoother than the latent diffusion process itself.
As \citet{Zhang-et-al-2005} studies, the existence of microstructure noise in financial data affects realised volatilities to increase as the subsampling frequency gets higher \citep[for instance, see Figure 7.1 in][]{Ait-Sahalia-Jacod-2014}. However, realised volatilities of some biological data such as EEG decrease as subsampling frequency increases: for instance, some time series data for the 2nd participant in the dataset named Two class motor imagery (002-2014) of \citet{BNCI-2014} show clear tendency of decreasing realised volatilities as subsampling frequency increases.
Figure \ref{fig:BNCI:paths} shows
the path of the 2nd axis of the data S02E.mat \citet{BNCI-2014} for all 222 seconds (the observation frequency is 512Hz,
and hence the entire data size is 113664) and that for the first one second;
it seems to perturb like a diffusion process.
\begin{figure}[ht]
\centering
\includegraphics[width=.45\textwidth]{S02E_2nd_column_whole.png}
\includegraphics[width=.45\textwidth]{S02E_2nd_column_1st_one_second.png}
\caption{The path of the second column of S02E.mat of \citet{BNCI-2014}
for all 222 seconds (left) and the first one second (right).}
\label{fig:BNCI:paths}
\end{figure}
We define realised volatilities with subsampling as for a one dimensional observation $\left\{Y_{i}\right\}_{i=0,\ldots,n}$,
\begin{align*}
RV\left(k\right)=\sum_{1\le i\le \left[n/k\right]}\left(Y_{ik}-Y_{\left(i-1\right)k}\right)^{2},
\end{align*}
where $k=1,\ldots,100$ is the subsampling frequency parameter, and provide a plot of the realised volatilities the 2nd axis of the data S02E.mat in Figure \ref{fig:BNCI:RV}:
\begin{figure}[ht]
\centering
\includegraphics[width=.8\textwidth]{BNCI_2020_RV_2nd_axis.png}
\caption{Realised volatilities with subsampling of the 2nd axis of data S02E.mat in Two class motor imagery (002-2014) \citep{BNCI-2014}.}
\label{fig:BNCI:RV}
\end{figure}
the altitudes of the graph represented in the $y$-axis correspond to the values of the realised volatilities $RV\left(k\right)$ with subsampling at every $k$ observation represented in the $x$-axis. It is observable that the increasing subsampling frequency results in decreasing realised volatilities, which cannot be explained by the existent major microstructure noises \citep[e.g., see][]{Jacod-et-al-2009, Jacod-et-al-2010, Bibinger-et-al-2014, Koike-2016, Ogihara-2018}. To explain this phenomenon, we consider the smoother process than the latent one though ordinarily microstructure noises make the observation rougher than the latent process, because quadratic variation of a sufficiently smooth function is zero. One way to deal with smoother observation than the latent state is convolutional observation. As a concrete example, we show a convolutionally observed diffusion process and its characteristics in realised volatilities: let us consider the following 1-dimensional stochastic differential equation defining an Ornstein-Uhlenbeck (OU) process:
\begin{align*}
\mathrm{d}X_{t} = -20X_{t}\mathrm{d}t+10\mathrm{d}w_{t}, X_{-\lambda}=0,
\end{align*}
where $\lambda=10^{-2/5}$. We simulate the stochastic differential equation by Euler-Maruyama method \citep[see][]{Iacus-2008} with parameters $n=10^{7}$, $h_{n}=10^{-5}$, and $T_{n}=10^{2}$ and its convolution approximated by summation with the smoothing parameter $\rho=10$ (for details, see Section 5).
Figure \ref{fig:intro:latentAndConvoluted} shows the latent diffusion process and the convoluted observation on $\left[0,1\right]$, and we can see that the observation is indeed smoothed compared to the latent state.
\begin{figure}[ht]
\centering
\includegraphics[width=.45\textwidth]{intro_latent.png}
\includegraphics[width=.45\textwidth]{intro_convoluted.png}
\caption{The left figure is the plot of the latent diffusion process, and the right one is that of the convolutionally observed process on $\left[0,1\right]$ respectively.}
\label{fig:intro:latentAndConvoluted}
\end{figure}
In Figure \ref{fig:intro:rv}, we also give the plot of realised volatilities of the convolutionally observed process with subsampling as Figure \ref{fig:BNCI:RV}.
\begin{figure}[ht]
\centering
\includegraphics[width=.8\textwidth]{intro_rv.png}
\caption{The realised volatilities of the convolutionally observed diffusion process with subsampling.}
\label{fig:intro:rv}
\end{figure}
It is seen that the convolutional observation of a diffusion process also has the characteristics of decreasing realised volatilities as subsampling frequency increases, which can bee seen in some biological data such as \citet{BNCI-2014}. Of course, graphically comparing characteristics of simulation and real data is insufficient to verify the convolutional observation with smoothing parameter $\rho>0$ in 1-dimensional case; therefore, we propose statistical estimation method for unknown $\rho$ and hypothesis test with the null hypothesis $H_{0}:\rho=\mathbf{0}$ and the alternative one $H_{1}:\rho\neq\mathbf{0}$ from convolutional observation in Section 3. Moreover, in Section 6, we examine the real EEG data plotted in Figure \ref{fig:BNCI:RV} by the statistical hypothesis testing we propose, and see it is more appropriate to consider the data as a convolutional observation of a latent diffusion process with $\rho\neq\mathbf{0}$ rather than direct observation of the latent process, which indicates the validity to deal with the problem
of the convolutional observation scheme with unknown $\rho$.
The paper is composed of the following sections: Section 2 gives the notations and assumptions used in this paper; Section 3 discusses the estimation and test for smoothing parameter $\rho$, and the discussion provides us with the tools to examine whether we should consider the convolutional observation scheme; Section 4 proposes the quasi-likelihood functions for the parameter of diffusion processes $\alpha$ and $\beta$, and corresponding estimators with consistency; Section 5 is for the computational simulation to examine the theoretical results in the previous sections;
and Section 6 shows an application of the methods we propose in real data analysis.
\section{Notations and assumptions}
\subsection{Notations}
First of all, we set $A\left(x,\alpha\right):=a\left(x,\alpha\right)^{\otimes 2}$, $a\left(x\right):=a\left(x,\alpha_{\star}\right)$, $A\left(x\right):=A\left(x,\alpha_{\star}\right)$ and $b\left(x\right):=b\left(x,\beta_{\star}\right)$. We also give the notation for a matrix-valued function $\mathbb{G}\left(x,\alpha|\rho\right)$ such that
$\mathbb{G}^{\left(i,j\right)}\left(x,\alpha|\rho\right):=A^{\left(i,j\right)}\left(x,\alpha\right)f_{\mathbb{G}}\left(\rho^{\left(i\right)},\rho^{\left(j\right)}\right)$, where
\begin{footnotesize}
\begin{align*}
&f_{\mathbb{G}}\left(\rho^{\left(i\right)},\rho^{\left(j\right)}\right)\\
&:=\begin{cases}
1& \text{ if }\rho^{\left(i\right)}=\rho^{\left(j\right)}=0,\\
1-\frac{\rho^{\left(j\right)}}{2}& \text{ if }\rho^{\left(i\right)}=0,\rho^{\left(j\right)}\in\left(0,1\right],\\
\frac{1}{2\rho^{\left(j\right)}}& \text{ if }\rho^{\left(i\right)}=0,\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\\
1-\frac{\rho^{\left(i\right)}}{2}& \text{ if }\rho^{\left(i\right)}\in\left(0,1\right],\rho^{\left(j\right)}=0,\\
\frac{1}{2\rho^{\left(i\right)}}& \text{ if }\rho^{\left(i\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}=0,\\
\frac{-3\left(\rho^{\left(i\right)}\right)^{2}\rho^{\left(j\right)}+3\rho^{\left(i\right)}\left(\rho^{\left(j\right)}\right)^{2}+6\rho^{\left(i\right)}\rho^{\left(j\right)}-2\left(\rho^{\left(j\right)}\right)^{3}}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}& \text{ if }\rho^{\left(i\right)},\rho^{\left(j\right)}\in\left(0,1\right],\rho^{\left(i\right)}>\rho^{\left(j\right)},\\
\frac{3\left(\rho^{\left(i\right)}\right)^{2}\rho^{\left(j\right)}-3\rho^{\left(i\right)}\left(\rho^{\left(j\right)}\right)^{2}+6\rho^{\left(i\right)}\rho^{\left(j\right)}-2\left(\rho^{\left(i\right)}\right)^{3}}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}& \text{ if }\rho^{\left(i\right)},\rho^{\left(j\right)}\in\left(0,1\right],\rho^{\left(i\right)}\le\rho^{\left(j\right)},\\
\frac{3\left(\rho^{\left(j\right)}\right)^{2}+3\rho^{\left(j\right)}-\left(\rho^{\left(j\right)}\right)^{3}}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}\in\left(0,1\right],\rho^{\left(i\right)}>\rho^{\left(j\right)}+1,\\
\frac{6\rho^{\left(j\right)}-1}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)},\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(i\right)}>\rho^{\left(j\right)}+1,\\
\frac{\left(\rho^{\left(i\right)}-\rho^{\left(j\right)}\right)^3 - 3\left(\rho^{\left(i\right)}\right)^2 + 6\rho^{\left(i\right)}\rho^{\left(j\right)} + 3 \rho^{\left(i\right)}- 1-\left(\rho^{\left(j\right)}\right)^{3}}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}\in\left(0,1\right],\rho^{\left(i\right)}\le \rho^{\left(j\right)}+1,\\
\frac{\left(\rho^{\left(i\right)}-\rho^{\left(j\right)}\right)^3 - 3\left(\rho^{\left(i\right)}\right)^2 + 6\rho^{\left(i\right)}\rho^{\left(j\right)} + 3 \rho^{\left(i\right)}-3\left(\rho^{\left(j\right)}\right)^{2}+3\rho^{\left(j\right)}- 2}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)},\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}<\rho^{\left(i\right)}\le \rho^{\left(j\right)}+1,\\
\frac{-\left(\rho^{\left(i\right)}-\rho^{\left(j\right)}\right)^{3}+6\rho^{\left(i\right)}\rho^{\left(j\right)}-\left(\rho^{\left(i\right)}\right)^{3}-3\left(\rho^{\left(j\right)}\right)^{2}+3\rho^{\left(j\right)}-1}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)}\in\left(0,1\right],\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}\le \rho^{\left(i\right)}+1,\\
\frac{-\left(\rho^{\left(i\right)}-\rho^{\left(j\right)}\right)^{3}-3\left(\rho^{\left(i\right)}\right)^{2}-3\left(\rho^{\left(j\right)}\right)^{2}+6\rho^{\left(i\right)}\rho^{\left(j\right)}+3\rho^{\left(i\right)}+3\rho^{\left(j\right)}-2}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)},\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(i\right)}\le\rho^{\left(j\right)}\le \rho^{\left(i\right)}+1,\\
\frac{3\left(\rho^{\left(i\right)}\right)^{2}+3\rho^{\left(i\right)}-\left(\rho^{\left(i\right)}\right)^{3}}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)}\in\left(0,1\right],\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}> \rho^{\left(i\right)}+1,\\
\frac{6\rho^{\left(i\right)}-1}{6\rho^{\left(i\right)}\rho^{\left(j\right)}}&\text{ if }\rho^{\left(i\right)},\rho^{\left(j\right)}\in\left(1,\overline{\rho}\right],\rho^{\left(j\right)}> \rho^{\left(i\right)}+1.
\end{cases}
\end{align*}
\end{footnotesize}
The continuity of these functions is examined in Lemma \ref{LemmaFunctionD} and Lemma \ref{LemmaFunctionG}. In addition, we also give the notation used throughout this paper.
\begin{itemize}
\item For every matrix $A$, $A^{T}$ is the transpose of $A$, and $A^{\otimes 2}:=AA^{T}$.
\item For every set of matrices $A$ and $B$
of the same size,
$A\left[B\right]:=\mathrm{tr}\left(AB^{T}\right)$. Moreover, for any $m\in\mathbf{N}$, $A\in\mathbf{R}^{m}\otimes\mathbf{R}^{m}$ and $u,v\in\mathbf{R}^{m}$, $A\left[u,v\right]:=v^{T}Au$.
\item $v^{\left(\ell\right)}$ and $A^{\left(\ell_{1},\ell_{2}\right)}$ denote
the $\ell$-th element of a vector $v$ and the $\left(\ell_{1},\ell_{2}\right)$-th one of a matrix $A$, respectively.
\item For any vector $v$ and any matrix $A$, $\left|v\right|:=\sqrt{\mathrm{tr}\left(v^{T}v\right)}$ and $\left\|A\right\|:=\sqrt{\mathrm{tr}\left(A^{T}A\right)}$.
\item $\left(\Omega,P,\mathcal{F},\mathcal{F}_{t}\right)$ denotes the stochastic basis, where $\mathcal{F}_{t}:=\sigma\left(x_{-\lambda},w_{s}:s\le t\right)$.
\end{itemize}
\subsection{Assumptions}
With respect to $X_{t}$, we assume the following conditions.
\begin{itemize}
\item[{[A1]}]
\begin{itemize}
\item[(i)] For a constant $C$, for all $x_{1},x_{2}\in\mathbf{R}^{d}$,
\begin{align*}
\sup_{\alpha\in\Theta_{1}}\left\|a\left(x_{1},\alpha\right)-a\left(x_{2},\alpha\right)\right\|+
\sup_{\beta\in\Theta_{2}}\left|b\left(x_{1},\beta\right)-b\left(x_{2},\beta\right)\right|\le C\left|x_{1}-x_{2}\right|.
\end{align*}
\item[(ii)] For all $p\ge0$, $\sup_{t\ge-\lambda}\mathbf{E}_{\theta_{\star}}\left[\left|X_{t}\right|^{p}\right]<\infty$.
\item[(iii)] There exists a unique invariant measure $\nu_{0}$ on $\left(\mathbf{R}^{d},\mathcal{B}\left(\mathbf{R}^{d}\right)\right)$ and for all $p\ge1$ and $f\in L^{p}\left(\nu_{0}\right)$ with polynomial growth,
\begin{align*}
\frac{1}{T}\int_{-\lambda}^{T}f\left(X_{t}\right)\mathrm{d}t\to^{P}\int_{\mathbf{R}^{d}}f\left(x\right)\nu_{0}\left(\mathrm{d}x\right).
\end{align*}
\end{itemize}
\item[{[A2]}] There exists $C>0$ such that $a:\mathbf{R}^{d}\times \Theta_{1}\to \mathbf{R}^{d}\otimes \mathbf{R}^{r}$ and $b:\mathbf{R}^{d}\times \Theta_{2}\to \mathbf{R}^{d}$ have continuous derivatives satisfying
\begin{align*}
\sup_{\alpha\in\Theta_{1}}\left|\partial_{x}^{j}\partial_{\alpha}^{i}a\left(x,\alpha\right)\right|&\le
C\left(1+\left|x\right|\right)^{C},\ 0\le i\le 2,\ 0\le j\le 2,\\
\sup_{\beta\in\Theta_{2}}\left|\partial_{x}^{j}\partial_{\beta}^{i}b\left(x,\beta\right)\right|&\le C\left(1+\left|x\right|\right)^{C},\ 0\le i\le 2,\ 0\le j\le 2.
\end{align*}
\end{itemize}
With the invariant measure $\nu_{0}$, we define
\begin{align*}
\mathbb{V}_{1} \left(\alpha|\xi_{\star}\right)
&:=-\int_{\mathbf{R}^{d}}
\left\|\mathbb{G}\left(x,\alpha|\rho_{\star}\right)-
\mathbb{G}\left(x,\alpha_{\star}|\rho_{\star}\right)\right\|^{2}\nu_{0} \left(\mathrm{d}x\right),\\
\mathbb{V}_{2}\left(\beta|\xi_{\star}\right)
&:=-\int_{\mathbf{R}^{d}}\left|b\left(x,\beta\right)-b\left(x,\beta_{\star}\right)\right|^{2}\nu_{0}\left(\mathrm{d}x\right).
\end{align*}
For these functions, let us assume the following identifiability conditions hold.
\begin{itemize}
\item[{[A3]}]
There exist $\chi_1\left(\alpha_{\star}\right)>0$ and $\chi_1'\left(\beta_{\star}\right)>0$
such that for all $\alpha\in\Theta_{1}$ and $\beta\in\Theta_{2}$,
$\mathbb{V}_{1}\left(\alpha|\xi_{\star}\right)\le -\chi_1\left(\alpha_{\star}\right)\left|\alpha-\alpha_{\star}\right|^2$
and $\mathbb{V}_{2}\left(\beta|\xi_{\star}\right)\le -\chi_1'\left(\beta_{\star}\right)\left|\beta-\beta_{\star}\right|^2$.
\end{itemize}
\section{Estimation and test of the smoothing parameter}
In this section, we discuss the case where the smoothing parameter $\rho$ of the kernel function $V_{\rho,h_{n}}$ is unknown.
The estimation is significant for estimation of $\alpha$ and $\beta$ since we utilise the estimate for $\rho$ in quasi-likelihood functions of $\alpha$ and $\beta$. The test problem for hypotheses $H_{0}:\rho=\mathbf{0}$ and $H_{1}:\rho\neq\mathbf{0}$ is also important to examine whether our framework of convolutional observation is meaningful.
\subsection{Estimation of the smoothing parameter}
For simplicity of notation, let us consider the case $\overline{\rho}>2$; otherwise the discussion is quite parallel. We should note that for all $i=1,\ldots,d$,
\begin{align*}
\mathbb{G}^{\left(i,i\right)}\left(x|\rho\right)=\begin{cases}
A^{\left(i,i\right)}\left(x\right)&\text{ if }\rho^{\left(i\right)}=0,\\
A^{\left(i,i\right)}\left(x\right)\left(1-\frac{\rho^{\left(i\right)}}{3}\right)&\text{ if }\rho^{\left(i\right)}\in\left(0,1\right],\\
A^{\left(i,i\right)}\left(x\right)\left(\frac{1}{\rho^{\left(i\right)}}-\frac{1}{3\left(\rho^{\left(i\right)}\right)^{2}}\right)&\text{ if }\rho^{\left(i\right)}\in\left(1,\overline{\rho}\right].
\end{cases}
\end{align*}
Let us consider the estimation of $\rho^{\left(i\right)}$ with using the next statistics: the full quadratic variation
\begin{align*}
\frac{1}{nh_{n}}\sum_{k=1}^{n}\left(\overline{X}_{kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(k-1\right)h_{n},n}^{\left(i\right)}\right)^{2}\to^{P} \begin{cases}
\nu_{0}\left(A^{\left(i,i\right)}\left(\cdot\right)\right)&\text{ if }\rho_{\star}^{\left(i\right)}=0,\\
\nu_{0}\left(A^{\left(i,i\right)}\left(\cdot\right)\right)\left(1-\frac{\rho_{\star}^{\left(i\right)}}{3}\right)&\text{ if }\rho_{\star}^{\left(i\right)}\in\left(0,1\right],\\
\nu_{0}\left(A^{\left(i,i\right)}\left(\cdot\right)\right)\left(\frac{1}{\rho_{\star}^{\left(i\right)}}-\frac{1}{3\left(\rho_{\star}^{\left(i\right)}\right)^{2}}\right)&\text{ if }\rho_{\star}^{\left(i\right)}\in\left(1,\overline{\rho}\right],
\end{cases}
\end{align*}
because of Proposition \ref{propEmpQ} in Appendix A,
and the reduced quadratic variation defined as $\frac{1}{nh_{n}}\sum_{2\le 2k\le n}\left(\overline{X}_{2kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(2k-2\right)h_{n},n}^{\left(i\right)}\right)^{2}$ converges in probability as follows.
\begin{lemma}\label{LemmaReducedQ}
Under [A1], we have the convergence in probability such that
\begin{align*}
&\frac{1}{nh_{n}}\sum_{2\le 2k\le n}\left(\overline{X}_{2k,n}^{\left(i\right)}-\overline{X}_{2k-2,n}^{\left(i\right)}\right)^{2}\\
&\to^{P}\begin{cases}
\nu_{0}\left(A^{\left(i,i\right)}\left(\cdot\right)\right)&\text{ if }\rho_{\star}^{\left(i\right)}=0,\\
\nu_{0}\left(A^{\left(i,i\right)}\left(\cdot\right)\right)\left(1-\frac{\rho_{\star}^{\left(i\right)}}{6}\right)&\text{ if }\rho_{\star}^{\left(i\right)}\in\left(0,2\right],\\
\nu_{0}\left(A^{\left(i,i\right)}\left(\cdot\right)\right)\left(\frac{2}{\rho_{\star}^{\left(i\right)}}-\frac{4}{3\left(\rho_{\star}^{\left(i\right)}\right)^{2}}\right)&\text{ if }\rho_{\star}^{\left(i\right)}\in\left(2,\overline{\rho}\right].
\end{cases}
\end{align*}
\end{lemma}
Then we define the ratio of those statistics such that
\begin{align*}
R_{n}^{\left(i\right)}&:=\left(\frac{1}{nh_{n}}\sum_{k=1}^{n}\left(\overline{X}_{kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(k-1\right)h_{n},n}^{\left(i\right)}\right)^{2}\right)\left(\frac{1}{nh_{n}}\sum_{2\le 2k\le n}\left(\overline{X}_{2kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(2k-2\right)h_{n},n}^{\left(i\right)}\right)^{2}\right)^{-1}\\
&\to^{P}\begin{cases}
1 & \text{ if }\rho_{\star}^{\left(i\right)}=0,\\
\left(1-\frac{\rho_{\star}^{\left(i\right)}}{3}\right)\left(1-\frac{\rho_{\star}^{\left(i\right)}}{6}\right)^{-1} & \text{ if }\rho_{\star}^{\left(i\right)}\in\left(0,1\right],\\
\left(\frac{1}{\rho_{\star}^{\left(i\right)}}-\frac{1}{3\left(\rho_{\star}^{\left(i\right)}\right)^{2}}\right)\left(1-\frac{\rho_{\star}^{\left(i\right)}}{6}\right)^{-1} &\text{ if }\rho_{\star}^{\left(i\right)}\in\left(1,2\right],\\
\left(\frac{1}{\rho_{\star}^{\left(i\right)}}-\frac{1}{3\left(\rho_{\star}^{\left(i\right)}\right)^{2}}\right)\left(\frac{2}{\rho_{\star}^{\left(i\right)}}-\frac{4}{3\left(\rho_{\star}^{\left(i\right)}\right)^{2}}\right)^{-1}&\text{ if }\rho_{\star}^{\left(i\right)}\in\left(2,\overline{\rho}\right]
\end{cases}\\
&=\begin{cases}
1 & \text{ if }\rho^{\left(i\right)}=0,\\
\left(6-2\rho_{\star}^{\left(i\right)}\right)\left(6-\rho_{\star}^{\left(i\right)}\right)^{-1} & \text{ if }\rho_{\star}^{\left(i\right)}\in\left(0,1\right],\\
\left(6\rho_{\star}^{\left(i\right)}-2\right)\left(6\left(\rho_{\star}^{\left(i\right)}\right)^{2}-\left(\rho_{\star}^{\left(i\right)}\right)^{3}\right)^{-1} &\text{ if }\rho_{\star}^{\left(i\right)}\in\left(1,2\right],\\
\left(3\rho_{\star}^{\left(i\right)}-1\right)\left(6\rho_{\star}^{\left(i\right)}-4\right)^{-1}&\text{ if }\rho_{\star}^{\left(i\right)}\in\left(2,\overline{\rho}\right],
\end{cases}\\
&=: R\left(\rho_{\star}^{\left(i\right)}\right),
\end{align*}
where $R$ has the next property.
\begin{lemma}\label{LemmaFunctionR}
$R$ is a $\left[\left(3\overline{\rho}-1\right)\left(6\overline{\rho}-4\right)^{-1},1\right]$-valued monotonically decreasing continuous function, and has a continuous inverse $R^{-1}:\left[\left(3\overline{\rho}-1\right)\left(6\overline{\rho}-4\right)^{-1},1\right]\to\left[0,\overline{\rho}\right]$.
\end{lemma}
We define $\hat{\rho}_{n}$ such that
\begin{align*}
\hat{\rho}_{n}^{\left(i\right)}&:=\begin{cases}
0 & \text{ if }R_{n}^{\left(i\right)}>1,\\
R^{-1}\left(R_{n}^{\left(i\right)}\right) & \text{ if }R_{n}^{\left(i\right)}\in \left[\left(3\overline{\rho}-1\right)\left(6\overline{\rho}-4\right)^{-1},1\right],\\
\overline{\rho} & \text{ if } R_{n}^{\left(i\right)}<\left(3\overline{\rho}-1\right)\left(6\overline{\rho}-4\right)^{-1},
\end{cases}
\end{align*}
and then continuous mapping theorem for convergence in probability verifies the next result.
\begin{theorem}\label{thmRhoEstimate}
Under [A1], $\hat{\rho}_{n}$ has consistency, i.e., $\hat{\rho}_{n}\to^{P}\rho_{\star}$.
\end{theorem}
\begin{remark}
We can compute $y=R^{-1}\left(x\right),\ x\in\left[\left(3\overline{\rho}-1\right)\left(6\overline{\rho}-4\right)^{-1},1\right]$ by solving the following equations:
\begin{align*}
\text{(i) }&x=\left(6-2y\right)\left(6-y\right)^{-1} && \text{ if }x\in\left(4/5,1\right],\\
\text{(ii) }& x=\left(6-2y\right)\left(6y^2-y^3\right)^{-1}&&\text{ if }x\in\left(5/8,4/5\right],\\
\text{(iii) }& x=\left(3y-1\right)\left(6y-4\right)^{-1}&&\text{ if }x\in\left[\left(3\overline{\rho}-1\right)\left(6\overline{\rho}-4\right)^{-1},5/8\right].
\end{align*}
\end{remark}
\subsection{Test for smoothed observation}
For all $i=1,\ldots,d$, we consider the next hypothesis testing:
\begin{align*}
H_{0}:\rho^{\left(i\right)}=0,\ H_{1}:\rho^{\left(i\right)}>0.
\end{align*}
Let us consider the following test statistic:
\begin{align*}
\mathcal{T}_{i,n}&:=\sqrt{\frac{n}{\frac{2}{3nh_{n}^{2}}\sum_{k=1}^{n}\left(\overline{X}_{kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(k-1\right)h_{n},n}^{\left(i\right)}\right)^{4}}}\\
&\quad\times\left(\frac{1}{nh_{n}}\sum_{k=1}^{n}\left(\overline{X}_{kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(k-1\right)h_{n},n}^{\left(i\right)}\right)^{2}-\frac{1}{nh_{n}}\sum_{2\le 2k\le n}\left(\overline{X}_{2kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(2k-2\right)h_{n},n}^{\left(i\right)}\right)^{2}\right)\\
&=\sqrt{\frac{3/2}{\sum_{k=1}^{n}\left(\overline{X}_{kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(k-1\right)h_{n},n}^{\left(i\right)}\right)^{4}}}\\
&\quad\times\left(\sum_{k=1}^{n}\left(\overline{X}_{k,n}^{\left(i\right)}-\overline{X}_{k-1,n}^{\left(i\right)}\right)^{2}-\sum_{2\le 2k\le n}\left(\overline{X}_{2kh_{n},n}^{\left(i\right)}-\overline{X}_{\left(2k-2\right)h_{n},n}^{\left(i\right)}\right)^{2}\right),
\end{align*}
and we abbreviate $\mathcal{T}_{i,n}$ to $\mathcal{T}_{n}$ if $d=1$.
Under $H_{0}$, we have the next result.
\begin{theorem}\label{thmTestH0}
Under $H_{0}$ and [A1], we have the convergence in law such that
\begin{align*}
\mathcal{T}_{i,n}\to^{\mathcal{L}}N\left(0,1\right).
\end{align*}
\end{theorem}
We also obtain the result to support the consistency of the test.
\begin{theorem}\label{thmTestH1}
Under $H_{1}$ and [A1], we have the divergence in probability such that for any $c\in\mathbf{R}$,
\begin{align*}
&P\left(\mathcal{T}_{i,n}<c\right)\to 1.
\end{align*}
\end{theorem}
Hence when we set the significance level $\alpha_{\mathrm{sig}}\in\left(0,1\right)$, then we have the rejection region
\begin{align*}
&\mathcal{T}_{i,n}<\Phi^{-1}\left(\alpha_{\mathrm{sig}}\right)
\end{align*}
where
$\Phi$ is the distribution function of the standard Gaussian distribution.
Theorem \ref{thmTestH1} supports the consistency of the test.
This test is essential in terms of examining the validity to consider the scheme of convolutional observation: if $\rho=\mathbf{0}$, then the ordinary observation scheme can be applied, but if $\rho\neq\mathbf{0}$, then we have the motivation to consider the convolutional
observation scheme.
\section{Least square estimation of the diffusion and drift parameters}
Let us set the least-square quasi-likelihood functions such that
\begin{align*}
\mathbb{H}_{1,n}\left(\alpha|\rho\right)&:=-\sum_{k=1}^{n}\left\|\frac{1}{h_{n}}\left(\overline{X}_{kh_{n},n}-\overline{X}_{\left(k-1\right)h_{n},n}\right)^{\otimes2}-\mathbb{G}\left(\overline{X}_{\left(k-1\right)h_{n},n},\alpha|\rho\right)\right\|^{2},\\
\mathbb{H}_{2,n}\left(\beta|\rho\right)& :=-\sum_{k=\left[\max_{i}\rho^{\left(i\right)}\right]+2}^{n}\frac{1}{h_{n}}\left|\overline{X}_{kh_{n},n}-\overline{X}_{\left(k-1\right)h_{n},n}-h_{n}b\left(\overline{X}_{\left(k-2-\left[\max_{i}\rho^{\left(i\right)}\right]\right)h_{n},n},\beta\right)\right|^{2},
\end{align*}
and the least-square estimators $\hat{\alpha}_{n}$ and $\hat{\beta}_{n}$ satisfying
\begin{align*}
\mathbb{H}_{1,n}\left(\hat{\alpha}_{n}|\rho_{\star}\right)=\sup_{\alpha\in\Theta_{1}}\mathbb{H}_{1,n}\left(\alpha|\rho_{\star}\right),\
\mathbb{H}_{2,n}\left(\hat{\beta}_{n}|\rho_{\star}\right)=
\sup_{\beta\in\Theta_{2}}\mathbb{H}_{2,n}\left(\beta|\rho_{\star}\right).
\end{align*}
when $\rho_{\star}$ is known, and
\begin{align*}
\mathbb{H}_{1,n}\left(\hat{\alpha}_{n}|\hat{\rho}_{n}\right)=\sup_{\alpha\in\Theta_{1}}\mathbb{H}_{1,n}\left(\alpha|\hat{\rho}_{n}\right),\
\mathbb{H}_{2,n}\left(\hat{\beta}_{n}|\hat{\rho}_{n}\right)=
\sup_{\beta\in\Theta_{2}}\mathbb{H}_{2,n}\left(\beta|\hat{\rho}_{n}\right).
\end{align*}
when $\rho_{\star}$ is unknown.
\begin{theorem}\label{thmThetaEstimate}
Under [A1]-[A3], $\hat{\alpha}_{n}$ and $\hat{\beta}_{n}$ are consistent, i.e., $\hat{\alpha}_{n}\to^{P}\alpha_{\star}$ and $\hat{\beta}_{n}\to^{P}\beta_{\star}$.
\end{theorem}
\section{Simulations}
In this simulation section, we only consider the case where $\rho$ is unknown and should be estimated by data with the method proposed in Section 3.
\subsection{1-dimensional simulation} We examine the following 1-dimensional stochastic differential equation whose solution is a 1-dimensional Ornstein-Uhlenbeck (OU) process:
\begin{align*}
\mathrm{d}X_{t}=\left(\beta^{\left(1\right)}X_{t}+\beta^{\left(2\right)}\right)\mathrm{d}t+\alpha\mathrm{d}w_{t},\ X_{-\lambda}=0,
\end{align*}
$\alpha\in\Theta_{1}:=\left[0.01,10\right]$, $\beta\in\Theta_{2}:=\left[-10,-0.01\right]\times\left[-10,10\right]$, and $\lambda=10^{-7/3}$. The procedure of the simulation is as follows: in the first place we iterate an approximated OU process by Euler-Maruyama scheme \citep[for example, see][]{Iacus-2008} with simulation parameters $n_{\mathrm{sim}}=10^{5+m}$, $h_{\mathrm{sim}}=10^{-10/3-m}$, $T_{\mathrm{sim}}=10^{5/3}$ where $m\in\mathbf{N}$ is
a parameter to determine the precision of approximation; secondly, we give the approximation of convolution by summation such that
\begin{align*}
\overline{X}_{ih_{n},n}\approxeq \begin{cases}\frac{1}{\left[10^{m}\rho\right]}\sum_{k=0}^{10^{m}-1}X_{ih_{n}-kh_{\mathrm{sim}}}&\text{ if }\left[10^{m}\rho\right]\ge1,\\
X_{ih_{n}}&\text{ if } \left[10^{m}\rho\right]<0,
\end{cases}
\end{align*}
where $i=0,\ldots,n$, the sampling frequency $h_{n}=10^{-10/3}$ and $n=10^{5/3}$. In this Section 5.1, we fix the true value of $\alpha$ and $\beta$ as $\alpha_{\star}=3$ and $\beta_{\star}=\left[-2,1\right]^{T}$, and change the true value of $\rho\in\Theta_{\rho}:=\left[0,100\right]$ to see the corresponding changes of performance of estimation for $\xi$, and test for $\rho$ in comparison to estimation by an existent method called local Gaussian approximation (LGA) for parametric estimation of discretely observed diffusion processes \citep[e.g., see][]{Kessler-1997} which does not concern convolutional observation. All the numbers of iterations for different $\rho$'s are 1000.
In the first place, we see the estimation and test with small values of $\rho_{\star}$ such that $\rho_{\star}=0,0.1,0.2,\ldots,1$ to observe how the performance of statistics changes by difference in $\rho$. Table \ref{tab:sim:1dim:small:rho:estimate} summarises the results of simulation of $\hat{\rho}_{n}$ for $\rho$'s with respective empirical means and root mean square error (RMSE).
\begin{table}[ht]
\begin{tabular}{c|c|c|c|c|c}
$\rho$ & $0.0$ & $0.1$ & $0.2$ & $0.3$ & $0.4$ \\\hline
mean & $0.00990$ & $0.0971$ & $0.198$ & $0.298$ & $0.398$ \\
RMSE & $(0.0182)$ & $(0.0256)$ & $(0.0235)$ & $(0.0215)$ & $(0.0197)$\\\\
\end{tabular}
\begin{tabular}{c|c|c|c|c|c|c}
$\rho$ & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 1.0\\\hline
mean & $0.498$ & $0.598$ & $0.699$ & $0.799$ & $0.899$ & $0.999$ \\
RMSE & $(0.0180)$ & $(0.0164)$ & $(0.0150)$ & $(0.0135)$ & $(0.0123)$ & $(0.0110)$
\end{tabular}
\caption{Estimation performance of $\rho$ with small $\rho$.}
\label{tab:sim:1dim:small:rho:estimate}
\end{table}
We can see the proposed estimator $\hat{\rho}_{n}$ works well for small $\rho$. With respect to the performance of the test statistic $\mathcal{T}_{n}$ proposed in Section 3.2, Table \ref{tab:sim:1dim:small:rho:test} shows the empirical ratio of the number of iterations whose $\mathcal{T}_{n}$ is lower than some typical critical values where $\Phi$ indicates the distribution function of 1-dimensional standard Gaussian distribution as well as the maximum value of $\mathcal{T}_{n}$ in 1000 iterations.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c|c|c}
& \multicolumn{5}{c|}{empirical ratio of $\mathcal{T}_{n}$ less than ...}& \multirow{2}{*}{max. of $\mathcal{T}_{n}$}\\
& $\Phi^{-1}\left(0.10\right)$ & $\Phi^{-1}\left(0.05\right)$ & $\Phi^{-1}\left(0.025\right)$ & $\Phi^{-1}\left(0.01\right)$ & $\Phi^{-1}\left(0.001\right)$\\\hline
$\rho=0.0$ & $0.101$ & $0.053$ & $0.025$ & $0.005$ & $0.000$ & $3.060$\\\hline
$\rho=0.1$ & $0.989$ & $0.980$ & $0.966$ & $0.914$ & $0.759$ & $-0.710$\\\hline
$\rho=0.2$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-4.593$\\\hline
$\rho=0.3$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-9.341$\\\hline
$\rho=0.4$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-13.985$\\\hline
$\rho=0.5$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-19.152$\\\hline
$\rho=0.6$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-24.816$\\\hline
$\rho=0.7$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-30.848$\\\hline
$\rho=0.8$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-37.557$\\\hline
$\rho=0.9$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-44.829$\\\hline
$\rho=1.0$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $1.000$ & $-52.759$\\
\end{tabular}
\caption{Empirical ratio of test statistic $\mathcal{T}_{n}$ less than some critical values, and the maximum value of $\mathcal{T}_{n}$ in 1000 iterations}
\label{tab:sim:1dim:small:rho:test}
\end{table}
Even for $\rho=0.1$, the simulation result supports the theoretical discussion of the test with consistency. Because $\Phi\left(10^{-16}\right)=-8.222$, all the iterations with $\rho\ge0.3$ result in rejection of $H_{0}$ with substantially significance level $10^{-16}$.
Let us see the estimation for $\alpha$ and $\beta$ by our proposal method and that by the LGA in Table \ref{tab:sim:1dim:small:theta:estimate}.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c|c|c|c}
\multicolumn{2}{c|}{} & \multicolumn{3}{c|}{the proposed method} & \multicolumn{3}{c}{LGA}\\\hline
\multicolumn{2}{c|}{}& $\alpha$ & $\beta^{\left(1\right)}$ & $\beta^{\left(2\right)}$ & $\alpha$ & $\beta^{\left(1\right)}$ & $\beta^{\left(2\right)}$\\\hline
\multicolumn{2}{c|}{true value}& $3.0$ & $-2.0$ & $1.0$ & $3.0$ & $-2.0$ & $1.0$\\\hline
\multirow{2}{*}{$\rho=0.0$} & mean & $3.004$ & $-2.091$ & $1.036$ & $2.999$ & $-2.095$ & $1.037$ \\
& RMSE & $(0.0109)$ & $(0.318)$ & $(0.497)$ & $(0.00679)$ & $(0.320)$ & $(0.497)$\\\hline
\multirow{2}{*}{$\rho=0.1$} & mean & $2.999$ & $-2.091$ & $1.035$ & $2.949$ & $-2.026$ & $1.003$ \\
& RMSE & $(0.0146)$ & $(0.319)$ & $(0.496)$ & $(0.0509)$ & $(0.297)$ & $(0.480)$ \\\hline
\multirow{2}{*}{$\rho=0.2$} & mean & $2.998$ & $-2.091$ & $1.035$ & $2.898$ & $-1.955$ & $0.967$ \\
& RMSE & $(0.0142)$ & $(0.319)$ & $(0.496)$ & $(0.102)$ & $(0.290)$ & $(0.464)$ \\\hline
\multirow{2}{*}{$\rho=0.3$} & mean & $2.998$ & $-2.092$ & $1.036$ & $2.846$ & $-1.885$ & $0.932$ \\
& RMSE & $(0.0139)$ & $(0.319)$ & $(0.497)$ & $(0.155)$ & $(0.299)$ & $(0.452)$\\\hline
\multirow{2}{*}{$\rho=0.4$} & mean & $2.998$ & $-2.091$ & $1.036$ & $2.792$ & $-1.815$ & $0.897$\\
& RMSE & $(0.0135)$ & $(0.319)$ & $(0.497)$ & $(0.208)$ & $(0.324)$ & $(0.442)$\\\hline
\multirow{2}{*}{$\rho=0.5$} & mean & $2.998$ & $-2.092$ & $1.036$ & $2.738$ & $-1.744$ & $0.862$\\
& RMSE & $(0.0132)$ & $(0.319)$ & $(0.497)$ & $(0.262)$ & $(0.361)$ & $(0.436)$ \\\hline
\multirow{2}{*}{$\rho=0.6$} & mean & $2.998$ & $-2.091$ & $1.036$ & $2.683$ & $-1.674$ & $0.827$\\
& RMSE & $(0.0129)$ & $(0.319)$ & $(0.497)$ & $(0.317)$ & $(0.408)$ & $(0.434)$\\\hline
\multirow{2}{*}{$\rho=0.7$} & mean & $2.998$ & $-2.092$ & $1.036$ & $2.626$ & $-1.604$ & $0.792$\\
& RMSE & $(0.0126)$ & $(0.319)$ & $(0.497)$ & $(0.374)$ & $(0.460)$ & $(0.434)$ \\\hline
\multirow{2}{*}{$\rho=0.8$} & mean & $2.998$ & $-2.092$ & $1.036$ & $2.568$ & $-1.534$ & $0.757$\\
& RMSE & $(0.0124)$ & $(0.319)$ & $(0.496)$ & $(0.432)$ & $(0.517)$ & $(0.439)$\\\hline
\multirow{2}{*}{$\rho=0.9$} & mean & $2.998$ & $-2.092$ & $1.036$ & $2.509$ & $-1.464$ & $0.722$ \\
& RMSE & $(0.0121)$ & $(0.319)$ & $(0.497)$ & $(0.491)$ & $(0.578)$ & $(0.445)$ \\\hline
\multirow{2}{*}{$\rho=1.0$} & mean & $2.998$ & $-2.091$ & $1.036$ & $2.449$ & $-1.394$ & $0.687$\\
& RMSE & $(0.0119)$ & $(0.319)$ & $(0.497)$ & $(0.551)$ & $(0.640)$ & $(0.456)$\\
\end{tabular}
\caption{Estimation of $\theta$ by the proposed method and LGA with small $\rho$}
\label{tab:sim:1dim:small:theta:estimate}
\end{table}
Note that the biases of the estimation by LGA increase as the true value of $\rho$ gets larger, while the estimation by our proposal method is not influenced by the true value of $\rho$. This result of the simulation supports the theoretical discussion in Section 4 stating the consistency of $\hat{\theta}_{n}$, and necessity to consider the convolutional observation scheme where the LGA method does not work properly.
Secondly, we consider the estimation and test with large $\rho_{\star}$ such that $\rho_{\star}=10,15,20$ to see if our proposal methods work even for large $\rho$. We note that the maximum values of $\mathcal{T}_{n}$ for $\rho=10,15,20$ in 1000 iterations are $-55.091$, $-68.462$ and $-79.105$, and hence we can detect the smoothed observation easily. Table \ref{tab:sim:1dim:large:rho:estimate} shows the empirical means and RMSEs of $\hat{\rho}_{n}$ for $\rho=10,15,20$ and we can see that the RMSEs increase as $\rho$'s increase; it indicates the difficulty to estimate $\rho$ accurately when $\rho_{\star}$ is large.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c}
& $\rho=10$ & $\rho=15$ & $\rho=20$\\\hline
mean of $\hat{\rho}_{n}$ & $9.919$ & $14.980$ & $19.751$\\
RMSE of $\hat{\rho}_{n}$ & $(0.145)$ & $(0.240)$ & $(0.409)$\\
\end{tabular}
\caption{The performance of $\hat{\rho}_{n}$ for $\rho=10, 15, 20$ in 1000 iterations}
\label{tab:sim:1dim:large:rho:estimate}
\end{table}
Table \ref{tab:sim:1dim:large:theta:estimate} summarises the estimation for $\theta$ by means and RMSE, and tells us that although the large RMSE of $\hat{\rho}_{n}$ results in increase of RMSE of $\hat{\alpha}_{n}$, estimation by our method is substantially better than that by LGA of course.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c|c|c|c}
\multicolumn{2}{c|}{} & \multicolumn{3}{c|}{the proposed method} & \multicolumn{3}{c}{LGA}\\\hline
\multicolumn{2}{c|}{}& $\alpha$ & $\beta^{\left(1\right)}$ & $\beta^{\left(2\right)}$ & $\alpha$ & $\beta^{\left(1\right)}$ & $\beta^{\left(2\right)}$\\\hline
\multicolumn{2}{c|}{true value}& $3.0$ & $-2.0$ & $1.0$ & $3.0$ & $-2.0$ & $1.0$\\\hline
\multirow{2}{*}{$\rho=10$} & mean & $2.989$ & $-2.101$ & $1.030$ & $0.933$ & $-0.204$ & $0.0811$ \\
& RMSE & $(0.0347)$ & $(0.323)$ & $(0.496)$ & $(2.067)$ & $(1.796)$ & $(0.920)$\\\hline
\multirow{2}{*}{$\rho=15$} & mean & $2.996$ & $-2.095$ & $1.027$ & $0.765$ & $-0.138$ & $0.0473$ \\
& RMSE & $(0.0475)$ & $(0.321)$ & $(0.495)$ & $(2.235)$ & $(1.862)$ & $(0.953)$ \\\hline
\multirow{2}{*}{$\rho=20$} & mean & $2.977$ & $-2.090$ & $1.024$ & $0.664$ & $-0.104$ & $0.0302$ \\
& RMSE & $(0.0526)$ & $(0.319)$ & $(0.493)$ & $(2.336)$ & $(1.896)$ & $(0.970)$ \\
\end{tabular}
\caption{Estimation of $\theta$ by the proposed method with large $\rho$}
\label{tab:sim:1dim:large:theta:estimate}
\end{table}
\subsection{2-dimensional simulation}
We consider the following 2-dimensional stochastic differential equation whose solution is a 2-dimensional OU process:
\begin{align*}
\mathrm{d}\left[\begin{matrix}
X_{t}^{\left(1\right)}\\
X_{t}^{\left(2\right)}
\end{matrix}\right]
&=
\left(
\left[\begin{matrix}
\beta^{\left(1\right)} & \beta^{\left(2\right)}\\
\beta^{\left(4\right)} & \beta^{\left(5\right)}
\end{matrix}\right]
\left[\begin{matrix}
X_{t}^{\left(1\right)}\\
X_{t}^{\left(2\right)}
\end{matrix}\right]
+
\left[\begin{matrix}
\beta^{\left(3\right)}\\
\beta^{\left(6\right)}
\end{matrix}\right]
\right)\mathrm{d}t
+
\left[\begin{matrix}
\alpha^{\left(1\right)} & \alpha^{\left(2\right)}\\
\alpha^{\left(2\right)} & \alpha^{\left(3\right)}
\end{matrix}\right]\mathrm{d}w_{t}, X_{-\lambda}=\left[\begin{matrix}
0\\
0
\end{matrix}\right],
\end{align*}
$\lambda=10^{-7/3}$.
The simulation is conducted with the settings as follows: firstly, we iterate the OU process by Euler-Maruyama scheme with the simulation sample size $n_{\mathrm{sim}}=10^{5+m}$, $T_{\mathrm{sim}}=10^{5/3}$ and discretisation step $h_{\mathrm{sim}}=10^{-10/3-m}$, where $m=2$ is the precision parameter for approximation of convolution; in the second place, we approximate the convoluted process with summation such that
\begin{align*}
\overline{X}_{ih_{n},n}^{\left(j\right)}\approxeq \begin{cases}\frac{1}{\left[10^{m}\rho^{\left(j\right)}\right]}\sum_{k=0}^{10^{m}-1}X_{ih_{n}-kh_{\mathrm{sim}}}^{\left(j\right)}&\text{ if }\left[10^{m}\rho^{\left(j\right)}\right]\ge1,\\
X_{ih_{n}}^{\left(j\right)}&\text{ if } \left[10^{m}\rho^{\left(j\right)}\right]<0,
\end{cases}
\end{align*}
where $i=0,\ldots,n$, $j=1,2,$ the sampling scheme for inference is defined as $n=10^{5}$ and $h_{n}=10^{-10/3}$; the true value of $\rho$, $\alpha$ and $\beta$ are set as $\rho_{\star}=\left[2,4\right]^{T}$, $\alpha_{\star}=\left[2,0,3\right]^{T}$, $\beta_{\star}=\left[-2, -0.4, 0, 0.1, -3, 5\right]^{T}$; the parameter spaces are defined as $\Theta_{\rho}=\left[0,10\right]^{2}$, $\Theta_{1}=\left[1+10^{-8},10\right]\times\left[-1+10^{-8},1-10^{-8}\right]\times\left[1+10^{-8},10\right]$, and $\Theta_{2}=\left[-10,10\right]^{6}$; the total iteration number is set to 1000.
Table \ref{tab:sim:2dim:rho:estimate} summarises the estimation for $\rho$ with the method proposed in Section 3 (the inverse of $r$ is computationally obtained) with empirical means and empirical RMSEs of $\hat{\rho}_{n}$ in 1000 iterations. We can see that $\hat{\rho}_{n}$ is sufficiently precise to estimate the true value of $\rho$ indeed in this result, which is significant to estimate the other parameters $\alpha$ and $\beta$.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c}
& $\rho^{\left(1\right)}$ & $\rho^{\left(2\right)}$\\\hline
true value & $2.0$ & $4.0$ \\\hline
empirical mean & $1.988$ & $3.966$ \\\hline
empirical RMSE & $(0.0207)$ & $(0.0514)$
\end{tabular}
\caption{summary for $\rho$ estimate}
\label{tab:sim:2dim:rho:estimate}
\end{table}
We also note that the maximum values of the test statistics for smoothed observation proposed in Section 3.2 in 1000 iterations are $-17.947$ and $-33.159$ for each axis. The $p$-value for them are smaller than $10^{-16}$; therefore, we can conclude that it is possible to detect the smoothed observation with the proposed test statistic in the case $\rho^{\left(i\right)}=2.0$ if $d=2$ from this result.
With respect to the estimation for $\alpha$ and $\beta$, we compare the estimates by our proposal method with that by LGA which does not concern convolutional observation. Table \ref{tab:sim:2dim:alpha:estimate} is the summary for $\alpha$ estimate by both the methods:
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c}
& & $\alpha^{\left(1\right)}$ & $\alpha^{\left(2\right)}$ & $\alpha^{\left(3\right)}$\\\hline
& true value & $2.0$ & $0.0$ & $3.0$ \\\hline
\multirow{2}{*}{Our proposal}& mean & $1.993$ & $0.000256$ & $2.992$ \\
& RMSE & $(0.0115)$ & $(0.00739)$ & $(0.0213)$\\\hline
\multirow{2}{*}{LGA}& mean & $1.295$ & $-0.00320$ & $1.442$ \\
& RMSE & $(0.705)$ & $(0.0154)$ & $(1.558)$\\
\end{tabular}
\caption{summary for $\alpha$ estimate}
\label{tab:sim:2dim:alpha:estimate}
\end{table}
we can see that the estimation precision for $\alpha$ by our proposal outperforms those by LGA. This results support validity of our estimation method if we have convolutional observation for diffusion processes. Regarding $\beta$, the simulation result is summarised in Table \ref{tab:sim:2dim:beta:estimate}:
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c|c|c|c}
& & $\beta^{\left(1\right)}$ & $\beta^{\left(2\right)}$ & $\beta^{\left(3\right)}$ & $\beta^{\left(4\right)}$ & $\beta^{\left(5\right)}$ & $\beta^{\left(6\right)}$\\\hline
& true value & $-2.0$ & $-0.4$ & $0.0$ & $0.1$ & $-3.0$ & $5.0$ \\\hline
\multirow{2}{*}{Our proposal}& mean & $-2.137$ & $-0.408$ & $-0.0439$ & $0.0788$ & $-3.103$ & $5.091$ \\
& RMSE & $(0.362)$ & $(0.252)$ & $(0.540)$ & $(0.473)$ & $(0.399)$ & $(0.777)$\\\hline
\multirow{2}{*}{LGA}& mean & $-0.917$ & $0.340$ & $-0.326$ & $-0.696$ & $0.221$ & $1.243$\\
& RMSE & $(1.093)$ & $(0.802)$ & $(0.386)$ & $(0.804)$ & $(3.242)$& $(3.765)$\\
\end{tabular}
\caption{summary for $\beta$ estimate}
\label{tab:sim:2dim:beta:estimate}
\end{table}
though the estimation for $\beta^{\left(3\right)}$ by our method has the smaller bias in comparison to that by LGA, the RMSE of our method is larger than that of LGA; in the estimation for other parameters, our proposal method outperforms the method by LGA. We can conclude that our proposal for estimation of $\alpha$ and $\beta$ concerning convolutional observation performs better than that not considering this observation scheme.
\section{Real data analysis}
In this section, we analyse the EEG dataset named S02E.mat provided in ``2.\ Two class motor imagery (002-2014)" of \citet{BNCI-2014}. The datasets including S02E.mat are also studied by \citet{Steyrl-et-al-2016}.
\subsection{Estimation and test for the smoothing parameters} In the first place, we pick up the first 15 axes of the dataset and compute $\hat{\rho}_{n}$ and $\mathcal{T}_{n}$ proposed in Section 3.1 and 3.2 respectively. The results are shown in Table \ref{tab:rda:est_and_tst}.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c|c}
& 1st axis & 2nd axis & 3rd axis & 4th axis & 5th axis\\\hline
$\hat{\rho}_{n}$ & $0.449$ & $1.037$ & $0.894$ & $0.736$ & $0.937$ \\\hline
$\mathcal{T}_{n}$ & $-20.398$ & $-58.631$ & $-46.649$ & $-35.201$ & $-49.741$
\end{tabular}\\
\begin{tabular}{c|c|c|c|c|c}
& 6th axis & 7th axis & 8th axis & 9th axis & 10th axis\\\hline
$\hat{\rho}_{n}$ & $0.951$ & $0.971$ & $1.017$ & $0.958$ & $0.967$ \\\hline
$\mathcal{T}_{n}$ & $-51.392$ & $-52.607$ & $-55.455$ & $-51.221$ & $-51.797$
\end{tabular}\\
\begin{tabular}{c|c|c|c|c|c}
& 11th axis & 12th axis & 13th axis & 14th axis & 15th axis\\\hline
$\hat{\rho}_{n}$ & $0.949$ & $0.649$ & $0.952$ & $0.977$ & $0.932$ \\\hline
$\mathcal{T}_{n}$ & $-50.457$ & $-30.094$ & $-50.633$ & $-50.978$ & $-48.842$
\end{tabular}\\
\caption{The values of $\hat{\rho}_{n}$ and $\mathcal{T}_{n}$ for the first 15 axes of S02.mat by \citet{BNCI-2014}.}
\label{tab:rda:est_and_tst}
\end{table}
We can observe that all the 15 time series data have the smoothing parameter $\rho>0$ with statistical significance when we assume ordinary significance levels. These results motivate us to use our methods for parametric estimation proposed in Section 4 when we fit stochastic differential equations for these data.
\subsection{Parametric estimation for a diffusion process}
We fit a 1-dimensional OU process for the time series data in the 2nd column of the data file S02E.mat with 512Hz observation for 222 seconds (the plot of the path can be seen in Figure \ref{fig:BNCI:paths}), whose $\hat{\rho}_{n}=1.037$ is the largest among those for the 15 axes and it is larger than 0 with statistical significance. According to the simulation result shown in Section 5.1, this size of the smoothing parameter gives critical biases when we estimate $\alpha$ and $\beta$ with LGA method not concerning convolutional observation scheme.
The stochastic differential equation with parameters $\alpha\in\Theta_{1}:=\left[0.01,200\right]$ and $\beta\in\Theta_{2}:=\left[-100,-0.01\right]\times\left[-100,100\right]$ is defined as follows:
\begin{align*}
\mathrm{d}X_{t}=\left(\beta^{\left(1\right)}X_{t}+\beta^{\left(2\right)}\right)\mathrm{d}t+\alpha\mathrm{d}w_{t},\ X_{-\lambda}=x_{-\lambda}.
\end{align*}
We set 5 seconds as the time unit: hence $n=113664$ and $h_{n}=1/\left(5\times512\right)$. If we fit the OU process with the LGA method, i.e., we do not concern convolutional observation scheme, we obtain the fitting result such that
\begin{align*}
\mathrm{d}X_{t}=\left(\left(-17.378\right)X_{t}+\left(-1.091\right)\right)\mathrm{d}t+\left(122.892\right)\mathrm{d}w_{t},\ X_{-\lambda}=x_{-\lambda}.
\end{align*}
In the next place, we fit $\alpha$ and $\beta$ with the least square method proposed in Section 4, and then we have the next fitting result:
\begin{align*}
\mathrm{d}X_{t}=\left(\left(-2.146\right)X_{t}+\left(0.552\right)\right)\mathrm{d}t+\left(151.919\right)\mathrm{d}w_{t},\ X_{-\lambda}=x_{-\lambda}.
\end{align*}
It is worth noting that this fitting result is substantially different to that by LGA as shown above: hence these results indicate the significance to examine if the observation is convoluted with the smoothing parameter $\rho>0$ and otherwise the estimation is strongly biased.
\section{Summary}
We have discussed the convolutional observation scheme which deals with the smoothness of observation in comparison to ordinary diffusion processes. The first contribution is to propose this new observation scheme with the statistical test to confirm whether this scheme is valid in real data. The second one is to prove consistency of the estimator $\hat{\rho}_{n}$ for the smoothing parameter $\rho$, those for parameters in diffusion and drift coefficient, i.e., $\alpha$ and $\beta$, of the latent diffusion process $\left\{X_{t}\right\}$. Thirdly, we have examined the performance of those estimators and the test statistics in computational simulation, and verified these statistics work well in realistic settings. In the fourth place, we have shown a real example of observation where $\rho\neq0$ holds with statistical significance.
These contributions, especially the third one, will cultivate the motivation to study statistical approaches for convolutionally observed diffusion processes furthermore, such as estimation of kernel function $V$ appearing in the convoluted diffusion $\overline{X}_{t}:=\left(V\ast X\right)\left(t\right)$, test theory for parameters $\alpha$ and $\beta$ as likelihood-ratio-type test statistics \citep[for example, see][]{Kitagawa-Uchida-2014, Nakakita-Uchida-2019b}, large deviation inequalities for quasi-likelihood functions and associated discussion of Bayes-type estimators \citep[e.g.,][]{Yoshida-2011, Ogihara-Yoshida-2011, Clinet-Yoshida-2017, Nakakita-Uchida-2018}. By these future works, it is expected that the applicability of stochastic differential equations in real data analysis and contributions to the areas with high frequency observation of phenomena such as EEG will be enhanced.
|
\section{Introduction}
Let $G$ be a countable discrete group and $\phi: G\rightarrow G$ an endomorphism.
Two elements $\alpha,\beta\in G$ are said to be
$\phi$-{\em conjugate} or {\em twisted conjugate,} iff there exists $g \in G$ with
$\beta=g \alpha \phi(g^{-1}).$
We shall write $\{x\}_\phi$ for the $\phi$-{\em conjugacy} or
{\em twisted conjugacy} class
of the element $x\in G$.
The number of $\phi$-conjugacy classes is called the {\em Reidemeister number}
of an endomorphism $\phi$ and is denoted by $R(\phi)$.
If $\phi$ is the identity map then the $\phi$-conjugacy classes are the usual
conjugacy classes in the group $G$.
Taking a dynamical point of view, we consider the iterates of $\phi$, and we may define \cite{Fel91} a Reidemeister zeta function of $\phi$ as a power series:
\begin{align*}
R_\phi(z)&=\exp\left(\sum_{n=1}^\infty \frac{R(\phi^n)}{n}z^n\right).
\end{align*}
Whenever we mention the Reidemeister zeta function $R_\phi(z)$, we shall assume that it is well-defined and so $R(\phi^n)<\infty$ for all $n>0$.
The following problem was investigated \cite{fh}: for which groups and endomorphisms is the Reidemeister zeta function a rational function? Is this zeta function an algebraic function?
When a Reidemeister zeta function is a rational function the infinite sequence of coefficients of the corresponding power series is closely interconnected, and is given by the finite set of zeros and poles of the zeta function.
In \cite{Fel91, fh, Li, fhw, Fel00}, the rationality of the Reidemeister zeta function $R_\phi(z)$ was proven in the following cases: the group is finitely generated and an endomorphism is eventually commutative; the group is finite; the group is a direct sum of a finite group and a finitely generated free Abelian group; the group is finitely generated, nilpotent and torsion free. Recently, the rationality and functional equation for the Reidemeister zeta function were proven for endomorphisms of fundamental groups of infra-nilmanifolds \cite{DeDu}
and for endomorphisms of fundamental groups of infra-solvmanifolds of type (R) \cite{FelLee}.
In this paper we continue to study the Reidemeister zeta function.
We prove P\'olya -- Carlson dichotomy between rationality and a natural boundary for analytic behavior of Reidemeister zeta function for a large class of automorphisms of Abelian groups.
We continue to study dynamical representation theory zeta functions (see \cite{FTZ}) counting
numbers of fixed irreducible unitary representations for iterates of an endomorphism.
The rationality and functional equation for these zeta functions are proven for several classes of groups.
We find a connection between these zeta functions and the Reidemeister torsions of the corresponding mapping tori.
We establish the connection
between Reidemeister zeta function and dynamical representation theory zeta functions under restriction of endomorphism to a subgroup and to a quotient group.
Our method is to identify the Reidemeister numbers with the number of fixed points of the induced
map $\hat\phi$ (respectively, its iterations)
of an appropriate subspace of the unitary dual $\wh G$,
when $R({\varphi})<\infty$.
This method is called the twisted Burnside--Frobenius
theory (TBFT), because in the case of a finite group and
identity automorphism we arrive to the classical Burnside--Frobenius
theorem on enumerating of (usual) conjugacy classes via irreducible unitary representations.
Let us present the contents of the paper in more details.
In Section 2 the rationality and functional equation for
dynamical representation theory zeta functions are proven for
endomorphisms of finitely generated Abelian groups; endomorphisms of finitely generated torsion free nilpotent groups; endomorphisms of groups with finite $\phi$-irreducible subspaces of corresponding unitary dual spaces and for automorphisms of crystallographic groups with diagonal holonomy $Z_2$.
For a periodic automorphism of a group we have proved a product formula for dynamical representation theory zeta functions which implies that these zeta functions are radicals of rational functions.
In Section 3 we investigate the rationality of these zeta functions and the connections between Reidemeister zeta function and dynamical representation theory zeta functions under restriction of endomorphism to a subgroup and to a quotient group.
We also prove the Gauss congruences for the Reidemeister numbers of iterations of endomorphism for a
group with polycyclic quotient group.
In Section 4 we presents results in support of a P\'olya -- Carlson dichotomy between rationality and a natural boundary for analytic behavior of Reidemeister zeta function for a large class of automorphisms of Abelian groups.
\noindent
\textbf{Acknowledgments.}
This work was supported by the grant Beethoven 2 of the Narodowe Centrum Nauki of Poland(NCN),
grant No. 2016/23/G/ST1/04280.
\section{Dynamical zeta functions and representations spaces}
Suppose, $\phi$ is an endomorphism of a discrete group $G$.
\rm
Denote by $\wh G$ the \emph{unitary dual} of $G$, i.e.
the space of equivalence classes of
unitary irreducible representations of $G$, equipped with the
\emph{hull-kernel} topology, denote by
$\wh G_f$ the subspace of the unitary dual formed by
irreducible finite-dimensional representations,
and by $\wh G_{ff}$ the subspace of $\wh G_{f}$ formed by
\emph{finite} representations, i.e. representations
that factorize through a finite group.
Generally the correspondence $\widehat{\phi}:\rho\mapsto \rho\circ\phi$
does not define a dynamical system (an action of the semigroup
of positive integers) on the unitary dual $\widehat{G}$
or its finite-dimensional part $\widehat{G}_f$, or finite part
$\widehat{G}_{ff}$, because in contrast with the
authomorphism case, the representation $\rho\circ\phi$
may be reducible, so it is only possible to decompose $\rho\circ\phi$ into irreducible
components and we obtain a sort of multivalued map $\widehat{\phi}$.
Nevertheless we can consider representations $\rho$ such that
$\rho \sim \rho\circ\phi$ and proceed as follows.
\begin{dfn}\label{dfn:represreidnum}
{\rm
A \emph{representation theory Reidemeister number}
$RT(\phi)$
is defined \cite{FTZ} as the number of all $[\rho]\in \widehat{G}$ such that
$\rho \sim \rho\circ\phi$. Taking $[\rho]\in \widehat{G}_f$
(respectively $[\rho]\in \widehat{G}_{ff}$) we obtain
$RT^f(\phi)$ (respectively $RT^{ff}(\phi)$). Evidently
$RT(\phi)\ge RT^f(\phi)\ge RT^{ff}(\phi)$.
}
\end{dfn}
In analogy with the Reidemeister zeta function and other similar objects we have defined in \cite{FTZ} jointly with E.Troitsky
following dynamical representation zeta functions
\begin{align*}
RT_\phi(z)&=\exp\left(\sum_{n=1}^\infty \frac{RT(\phi^n)}{n}z^n\right),\\
RT^f_\phi(z)&=\exp\left(\sum_{n=1}^\infty \frac{RT^f(\phi^n)}{n}z^n\right),\\
RT^{ff}_\phi(z)&=\exp\left(\sum_{n=1}^\infty \frac{RT^{ff}(\phi^n)}{n}z^n\right),\\
\end{align*}
when numbers $ RT(\phi^n)$
(resp, $ RT^f(\phi^n)$ , or $ RT^{ff}(\phi^n)$) are all finite.
The importance of these numbers is justified by the
following dynamical interpretation.
In \cite{RJMP} the following
``dynamical part'' of the dual space, where $\widehat{\phi}$ and all
its iterations $\widehat{\phi}^n$ define a dynamical system,
was defined.
\begin{dfn}
{\rm
Following \cite{RJMP} a class $[\rho]$ is called
a $\wh\phi$-\textbf{f}-point,
if $\rho\sim \rho\circ\phi$ (so, these are the points
under consideration in the Definition \ref{dfn:represreidnum}).
}
\end{dfn}
\begin{dfn}
{\rm
Following \cite{RJMP} an element $[\rho]\in \wh{G}$ (respectively, in $\wh{G}_f$
or $\wh{G}_{ff}$)
is called $\phi$-\emph{irreducible} if $\rho\circ \phi^n$ is
irreducible for any $n=0,1,2,\dots$.
Denote the corresponding subspaces of $\wh{G}$ (resp., $\wh{G}_f$
or $\wh{G}_{ff}$)
by $\wh{G}^\phi$ (resp., $\wh{G}^\phi_f$ or $\wh{G}^\phi_{ff}$).
}
\end{dfn}
\begin{lem}[Lemma 2.4 in \cite{RJMP}]\label{lem:keylem}
Suppose the representations $\rho$ and $\rho\circ\phi^n$ are
equivalent for some $n\geq 1$. Then $[\rho]\in \wh G^\phi$.
\end{lem}
\begin{cor}[Corollary 2.5 in \cite{RJMP}]\label{cor:periodicanddyn}
Generally, there is no dynamical system defined by $\widehat{\phi}$
on $\wh G$ (resp., $\wh G_f$, or $\wh{G}_{ff}$).
We have only the well-defined notion of a
$\wh\phi^n$-\textbf{f}-point.
A well-defined dynamical system exists on
$\wh G^\phi$ (resp., $\wh G^\phi_f$, or $\wh{G}^\phi_{ff}$).
Its $n$-periodic points are exactly $\wh\phi^n$-\textbf{f}-points.
\end{cor}
We refer to \cite{RJMP} for proofs and
details.
Once we have identified
the coefficients of representation theory zeta functions
with the
numbers of periodic points of a dynamical system,
the standard argument with the
M\"obius inversion formula
(see e.g. \cite[p.~104]{Fel00}, \cite{RJMP})
gives the following statement.
\begin{teo} (Theorem 2.7 of \cite{FTZ})\label{teo:congrue_rep_reide}
Suppose , $RT(\phi^n)<\infty$ for any $n$.
Then we have the following Gauss congruences
for representation theory Reidemeister numbers:
$$
\sum_{d\mid n} \mu(d)\cdot RT(\phi^{n/d}) \equiv 0 \mod n
$$
for any $n$.
A similar statement is true for $RT^f(\phi^n)$ and
$RT^{ff}(\phi^n)$.
\end{teo}
Here the above \emph{M\"obius function} is defined as
$$
\mu(d) =
\left\{
\begin{array}{ll}
1 & {\rm if}\ d=1, \\
(-1)^k & {\rm if}\ d\ {\rm is\ a\ product\ of}\ k\ {\rm distinct\ primes,}\\
0 & {\rm if}\ d\ {\rm is\ not\ square-free.}
\end{array}
\right.
$$
\begin{dfn}\label{dfn:tbft}
Following \cite{RJMP} we say that TBFT (resp., TBF$T_f$, TBF$T_{ff}$) takes place for
an endomorphism $\phi:G\to G$ and its iterations, if
$R(\phi^n)<\infty$ and $R(\phi^n)$ coincides with the
number of $\wh\phi^n$-\textbf{f}-points in $\wh G$
(resp., in $\wh G_f$, $\wh G_{ff}$)
for all $n\in \mathbb{N}$.
Similarly, one can give a definition for a single
endomorphism (without iterations).
\end{dfn}
The following statement follows from the
definitions.
\begin{prop}(Proposition 2.8 of \cite{FTZ}) \label{prop:tbftimplcoin}
Suppose, $\phi:G\to G$ is an endomorphism and $R(\phi)<\infty$.
If TBFT $($resp., TBFT$_f)$ is true for $G$ and $\phi$,
then $R(\phi)=RT(\phi)$ $($resp., $R(\phi)=RT^f(\phi)=RT^{ff}(\phi))$.
If the suppositions hold for $\phi^n$, for any $n$, then
$R_\phi(z)=RT_\phi(z)$ $($resp.,
$R_\phi(z)=RT^f_\phi(z)=RT^{ff}_\phi(z))$.
\end{prop}
Denote by $ AM^f(\phi^n)$ the number of \emph{isolated} $n$-periodic
points(i.e. \emph{isolated} $\wh\phi^n$-\textbf{f}-points) of the dynamical system $(\wh{\phi})^n$ on
$\wh G^\phi_f$.
If these numbers are finite for all powers of $\phi$, the corresponding
Artin--Masur representation zeta function is defined as
$$
AM^f_\phi(z) =\exp\left(\sum_{n=1}^\infty \frac{AM^f(\phi^n)}{n}z^n\right).
$$
Let $Z(\phi)$ be one of the numbers $RT(\phi)$, $ RT^f(\phi)$, $ RT^{ff}(\phi)$, $ AM^f(\phi)$ . Let
$$
Z_\phi(z)=\exp\left(\sum_{n=1}^\infty \frac{Z(\phi^n)}{n}z^n\right)
$$
be one of the zeta
functions $AM^f_\phi(z)$, $RT_\phi(z)$, $RT^f_\phi(z)$, $RT^{ff}_\phi(z).$
\begin{teo}\label{period}
Let $\phi$ be a periodic automorphism of least period $m$ of a group $G$ . Then the
zeta function $Z_\phi(z)$ is equal to
$$
Z_\phi(z)=\prod_{d\mid m}\sqrt[d]{(1-z^d)^{-P(d)}},
$$
where the product is taken over all divisors $d$ of the period $m$, and $P(d)$ is the integer
$$ P(d) = \sum_{d_1\mid d} \mu(d_1)Z({\phi}^{d\mid d_1}) . $$
\end{teo}
\begin{proof}
Since $\phi^m = id $, then $(\wh{\phi})^m =id$ as well and $ Z(\phi^j)=Z(\phi^{m+j})$ for every $j$. If $(k,m)=1$, there exist positive integers $t$
and $q$ such that $kt=mq+1$. So $ (\phi^k)^t=\phi^{kt}= \phi^{mq+1}=\phi^{mq}\phi=(\phi^m)^{q}\phi = \phi$.
Consequently, $ Z(\phi^k)=Z(\phi) $.
The same argument shows that $Z(\phi^d)=Z(\phi^{di})$ if $(i,m/d)=1$ where $d$ divisor $m$
Using these series of equal numbers we obtain by direct calculation
\begin{eqnarray*}
Z_\phi(z) & = & \exp\left(\sum_{i=1}^\infty \frac{Z(\phi^i)}{i} z^i \right)
= \exp\left(\sum_{d\mid m} \sum _{i=1}^\infty \frac{P(d)}{d}\cdot\frac{{z^d}^i}{i}\right)\\
&=&\exp\left(\sum_{d\mid m}\frac{P(d)}{d}\cdot \log (1-z^d)\right)
= \prod_{d\mid m}\sqrt[d]{(1-z^d)^{-P(d)}}
\end{eqnarray*}
where the integers $P(d)$ are calculated recursively by the formula
$$
P(d)= Z(\phi^d) - \sum_{d_1\mid d; d_1\not=d} P(d_1).
$$
Moreover, if the last formula is rewritten in the form
$
Z(\phi^d)=\sum_{d_1\mid d} P(d_1)
$
and one uses the M\"obius inversion law for real function in number theory, then
$$
P(d)=\sum_{d_1\mid d}\mu(d_1)\cdot Z(\phi^{d/d_1}),
$$
where $\mu(d_1)$ is the M\"obius function in number theory. The theorem is proved.
\end{proof}
\begin{cor}
If in Theorem \ref{period} the period $m$ is a prime number, then
$$
Z_\phi(z) = \frac{1}{(1-z)^{Z(\phi)}}\cdot \sqrt[m]{(1-z^m)^{Z(\phi) - Z(\phi^m)}}.
$$
\end{cor}
\begin{teo}\label{finite}
Let $\phi:G\rightarrow G$ be an endomorphism
of group $G$. Suppose that subspaces $\wh{G}^\phi$, $\wh{G}_f^\phi$, and $\wh{G}_{ff}^\phi$ are finite.
Then zeta function $Z_\phi(z)$ is a rational function satisfying a functional equation
$$
Z_\phi\left(\frac{1}{z}\right)
=
(-1)^a z^b Z_\phi(z).
$$
In particular we have
\begin{equation}
Z_\phi(z) = \prod_{[\gamma]} \frac{1}{1-z^{\#[\gamma]}},
\end{equation}
where the product is taken over the periodic orbits of the dynamical system $(\wh{\phi})^n$
in $\wh{G}^\phi$, resp $\wh{G}_f^\phi$, or $\wh{G}_{ff}^\phi$.
In the functional equation the numbers $a$ and $b$ are respectively
the number of periodic $\wh{\phi}$-orbits of elements
of $\wh{G}^\phi$, resp $\wh{G}_f^\phi$, or $\wh{G}_{ff}^\phi$ and the number of periodic elements
of $\wh{G}^\phi$, resp $\wh{G}_f^\phi$, or $\wh{G}_{ff}^\phi$.
\end{teo}
\begin{proof}
We shall call an element of $\wh{G}^\phi$, resp $\wh{G}_f^\phi$, or $\wh{G}_{ff}^\phi$ periodic if it is fixed by some
iteration of $\wh{\phi}$.
A periodic element $\gamma$ is fixed by $\wh{\phi}^n$ iff $n$ is divisible by
the cardinality the orbit of $\gamma$.
We therefore have
\begin{eqnarray*}
Z(\phi^n)
=
\sum_{\gamma \ periodic \atop \#[\gamma]\mid n} 1
=
\sum_{[\gamma]\ such \ that, \atop \#[\gamma]\mid n} \#[\gamma].
\end{eqnarray*}
From this follows
\begin{eqnarray*}
Z_\phi(z)
=
\exp\left(\sum_{n=1}^\infty \frac{Z(\phi^n)}{n} z^n\right)
=
\exp\left(\sum_{[\gamma]}
\sum_{n=1\atop \#[\gamma]\mid n}^\infty
\frac{\#[\gamma]}{n} z^n\right) \\
=
\prod_{[\gamma]}
\exp\left(\sum_{n=1}^\infty
\frac{\#[\gamma]}{\#[\gamma]n} z^{\#[\gamma]n}\right)
=
\prod_{[\gamma]}
\exp\left(\sum_{n=1}^\infty
\frac{1}{n} z^{\#[\gamma]n}\right) \\
=
\prod_{[\gamma]}
\exp \left( - \log \left( 1-z^{\#[\gamma]}\right)\right)
=
\prod_{[\gamma]} \frac{1}{1-z^{\#[\gamma]}}.
\end{eqnarray*}
Moreover
\begin{equation*}
\begin{aligned}
Z_\phi\left(\frac{1}{z}\right) {}&= \prod_{[\gamma]}\frac{1}{1-z^{-\#[\gamma]}}
= \prod_{[\gamma]}\frac{z^{\#[\gamma]}}{z^{\#[\gamma]}-1}
= \prod_{[\gamma]}\frac{-z^{\#[\gamma]}}{1-z^{\#[\gamma]}}
\\&= \prod_{[\gamma]} -z^{\#[\gamma]}Z_{\phi}(z)
= (-1)^{\#\{[\gamma]\}}z^{\sum\#[\gamma]}Z_{\phi}(z).
\end{aligned}
\end{equation*}
\end{proof}
\subsection{Endomorphisms of finitely generated Abelian groups}
For a finitely generated Abelian group $G$ we define the finite subgroup $G^{finite}$ to be the subgroup of
torsion elements of $G$. We denote the quotient $G^\infty:=G/G^{finite}$.
The group $G^\infty$ is torsion free.
Since the image of any torsion element by a homomorphism must be a torsion
element, the function $\phi:G\to G$ induces maps
$$
\phi^{finite}:G^{finite}\longrightarrow G^{finite},\;\;\;\;
\phi^\infty:G^\infty\longrightarrow G^\infty.
$$
If $G$ is abelian, then $\wh{G}=\wh{G}_f=\wh{G}^{\varphi}=\wh{G}^{\varphi}_f$ \cite{RJMP}.
The Lefschetz zeta function of a discrete dynamical system $\hat{\phi}$ equals:
$$
L_{\hat{\phi}}(z) := \exp\left(\sum_{n=1}^\infty \frac{L(\hat{\phi}^n)}{n} z^n \right),
$$
where
$$
L(\hat{\phi}^n) := \sum_{k=0}^{\dim X} (-1)^k \operatorname{Tr}\Big[\hat{\phi_{*k}}^n:H_k(\hat{G};Q)\to H_k(\hat{G};Q)\Big]
$$
is the Lefschetz number of $\hat{\phi}^n$.
The Lefschetz zeta function $ L_{\hat{\phi}}(z)$ is a rational function of $z$ and
is given by the formula:
$$
L_{\hat{\phi}}(z) = \prod_{k=0}^{\dim X}
\det\big(I-\hat{\phi_{*k}}\cdot z\big)^{(-1)^{k+1}}.
$$
\begin{teo}\label{fingenabelian}
Let $\phi:G\to G$ be an endomorphism of a finitely
generated Abelian group.
Then we have
\begin{equation}
Z(\phi^n) = \mid L(\hat{\phi}^{n}) \mid,
\end{equation}
where $L(\hat{\phi}^{n})$ is the Lefschetz number
of $\hat{\phi}$ thought of as a self-map of the topological space $\hat{G}$.
{}From this it follows that zeta functon $Z_\phi(z)$ is a rational function and is equal to:
\begin{equation}
Z_\phi(z) = L_{\hat{\phi}}(\sigma z)^{(-1)^r},
\end{equation}
where $\sigma=(-1)^p$ where $p$ is the number of real eingevalues
$\lambda\in \operatorname{Spectr}( \phi^\infty)$ such that $\lambda<-1$
and $r$ is the number of real eingevalues
$\lambda\in \operatorname{Spectr} (\phi^\infty)$
such that $\mid\lambda\mid > 1$.
If $G$ is a finite abelian group then this reduces to
$$ Z(\phi^n)=L(\hat{\phi}^n) \; {\rm and} \; Z_\phi(z)=L_{\hat{\phi}}(z).$$
\end{teo}
\begin{proof}
If $G$ is finite abelian then $\hat{G}$ is a discrete
finite set, so the number of fixed points is equal to
the Lefschetz number. This finishes the proof in the
case that $G$ is finite.
If $G$ is a finitely generated free Abelian group then
the dual of $G$
is a torus whose dimension is equal to the rank of $G$.
The dual of any finitely generated
discrete Abelian group is the direct sum of a torus and a
finite group.
If $G$ a finitely
generated Abelian group it is only necessary to check that
the number of fixed points of $\hat{\phi}^n$ is equal to the
absolute value of its Lefschetz number.
We assume without loss of generality that $n=1$.
We are assuming that $Z(\phi)$ is finite, so
the fixed points of $\hat{\phi}$ form a discrete set.
We therefore have
$$
L(\hat{\phi})
=
\sum_{x\in\operatorname{Fix}\hat{\phi}} \operatorname{ind}(\hat{\phi},x).
$$
Since $\phi$ is a group endomorphism, the zero element $0\in\hat{G}$ is always fixed.
Let $x$ be any fixed point of $\hat{\phi} $.
We then have a commutative diagram
$$
\begin{array}{ccccc}
g & \hat{G} & \stackrel{\hat{\phi}}{\longrightarrow} & \hat{G} & g \\
\updownarrow & \updownarrow & & \updownarrow & \updownarrow \\
g + x & \hat{G} & \stackrel{\hat{\phi}}{\longrightarrow} & \hat{G} & g + x
\end{array}
$$
in which the vertical functions are translations on $\hat{G}$ by $x$.
Since the vertical maps map $0$ to $x$, we deduce that
$$ \operatorname{ind}(\hat{\phi},x) = \operatorname{ind}(\hat{\phi},0) $$
and so all fixed points have the same index.
It is now sufficient to show that $\operatorname{ind}(\hat{\phi},0)=\pm 1$.
This follows because the map on the torus
$$ \hat{\phi}:\hat{G}_0\to\hat{G}_0 $$
lifts to a linear map of the universal cover, which is in this case the
Lie algebra of $\hat{G}$. The index is then the sign of the determinant of the identity map minus this lifted map.
This determinant cannot be zero, because $1-\hat{\phi}$ must have finite
kernel by our assumption that the $Z(\phi)$ is
finite
(if $\det(1-\hat{\phi})=0$ then the kernel of $1-\hat{\phi}$
is a positive dimensional subgroup of $\hat{G}$, and therefore
infinite). So we have $Z({\varphi}^n) = \mid L(\hat{\phi}^{n}) \mid =(-1)^{r+pn}L(\hat{\phi}^{n})$ for all $n$ (see also \cite{Fel00}).
Then the zeta function $$ Z_\phi(z) = L_{\hat{\phi}}(\sigma z)^{(-1)^r}$$ is rational function as well.
\end{proof}
\subsubsection{ Functional equation }
To write down a functional equation for the Reidemeister type zeta functions, we recall the following functional equation for the Lefschetz zeta function:
\begin{lem}[{\cite[Proposition~8]{fri1}}, see also \cite{del}] \label{Fried}
{Let $M$ be a closed orientable manifold of dimension $m$ and let $f:M\to M$ be a continuous map of degree $d$. Then
$$
L_{f}\left(\frac{\alpha}{dz}\right)=\epsilon\,(-\alpha dz)^{(-1)^m\chi(M)}\,L_{f}(\alpha z)^{(-1)^m}
$$
where $\alpha=\pm1$ and $\epsilon\in{\mathbb C}$ is a non-zero constant such that if $|d|=1$ then $\epsilon=\pm1$.}
\end{lem}
We obtain:
\begin{teo}[{Functional Equation}]\label{FE}
Let $\phi:G\to G$ be an endomorphism of a finitely
generated Abelian group of the rank $\geq 1$.
Then the zeta function $Z_{\phi}(z)$, whenever it is defined, has the following functional equation:
\begin{equation*}
Z_{\phi}\left(\frac{1}{dz}\right)
= Z_{\phi}(z)^{(-1)^m}\epsilon^{(-1)^{r}}
\end{equation*}
where $d$ is a degree $\hat\phi$, $m= \dim \hat G$, $\epsilon$ is a constant in ${\mathbb C}^\times$, $\sigma=(-1)^r$, $p$ is the number of real eigenvalues of $\phi^\infty $ which are $>1$ and $r$ is the number of real eingevalues
$\lambda\in \operatorname{Spectr} (\phi^\infty)$
such that $\mid\lambda\mid > 1$.
If $|d|=1$ then $\epsilon=\pm1$.
\end{teo}
\begin{proof}
We have $ Z_{\phi}(z) = L_{\hat\phi}(\sigma z)^{(-1)^{r}}$. By Lemma~\ref{Fried}
\begin{align*}
Z_{\phi}\left(\frac{1}{dz}\right) = L_{\hat\phi}\left(\frac{\sigma}{dz}\right)^{(-1)^{r}}
=\left(\epsilon(-\sigma dz)^{(-1)^m\chi(\hat G)}L_{\hat\phi}(\sigma z)^{(-1)^m}\right)^{(-1)^{r}}= \\
= Z_{\phi}(z)^{(-1)^m}\epsilon^{(-1)^{r}}(-\sigma dz)^{(-1)^{m+r}\chi(\hat G)}.
\end{align*}
On the other hand $\chi(\hat G)=0$ because the dual $\hat G$ of any finitely generated
discrete Abelian group of the rank $\geq 1$ is the direct sum of a torus of $\dim \geq 1$ and a
finite group, i.e. $\hat G$ is a union of finitely many tori. This finishes our proof.
\end{proof}
\subsection{Endomorphisms of nilpotent groups and crystallographic groups}
\begin{teo} \label{teo:nil}
Let $\phi:G\to G$ be an endomorphism of a finitely generated torsion free nilpotent group $G$ or
let $\phi$ be an automorphism of crystallographic group $G$ with diagonal holonomy $Z_2$.
Then the zeta function $RT^f_\phi(z)=RT^{ff}_\phi(z))$ is rational function.
\end{teo}
\begin{proof}
Any finitely generated torsion free nilpotent group is a supersolvable, hence, polycyclic group.
Any crystallographic group with diagonal holonomy $Z_2$ is
a polycyclic-by-finite group. In \cite{RJMP, crelle} twisted Burnside-Frobenius theorem($TBFT_f$ and $TBFT_{ff}$) was proven for endomorphisms of polycyclic groups and for automorphisms of polycyclic-by-finite groups. This theorem implies equality of Reidemeister zeta function $R_\phi(z)$ and zeta function $RT^f_\phi(z)=RT^{ff}_\phi(z))$. In \cite{Fel00} the rationality of the Reidemeister zeta function $R_\phi(z)$
was proven for endomorphisms of a finitely generated torsion free nilpotent groups and in
\cite{DekTerBus} the rationality of $R_\phi(z)$ was proven for automorphisms of crystallographic groups with diagonal holonomy $Z_2$. This completes the proof.
\end{proof}
\section{Connection with Reidemeister Torsion}
Like the Euler characteristic, the Reidemeister torsion is algebraically defined.
Roughly speaking, the Euler characteristic is a
graded version of the dimension, extending
the dimension from a single vector space to a complex
of vector spaces.
In a similar way, the Reidemeister torsion
is a graded version of the absolute value of the determinant
of an isomorphism of vector spaces.
Let $d^i:C^i\rightarrow C^{i+1}$ be a cochain complex $C^*$
of finite dimensional vector spaces over ${\mathbb C}$ with
$C^i=0$ for $i<0$ and large $i$.
If the cohomology $H^i=0$ for all $i$ we say that
$C^*$ is {\it acyclic}.
If one is given positive densities $\Delta_i$ on $C^i$
then the Reidemeister torsion $\tau(C^*,\Delta_i)\in(0,\infty)$
for acyclic $C^*$ is defined as follows:
\begin{dfn}
Consider a chain contraction $\delta^i:C^i\rightarrow C^{i-1}$,
ie. a linear map such that $d\circ\delta + \delta\circ d = id$.
Then $d+\delta$ determines a map
$ (d+\delta)_+ : C^+:=\oplus C^{2i}
\rightarrow C^- :=\oplus C^{2i+1}$
and a map
$ (d+\delta)_- : C^- \rightarrow C^+ $.
Since the map
$(d+\delta)^2 = id + \delta^2$ is unipotent,
$(d+\delta)_+$ must be an isomorphism.
One defines $\tau(C^*,\Delta_i):= \mid \det(d+\delta)_+\mid$
(see \cite{fri2}).
\end{dfn}
Reidemeister torsion is defined in the following geometric setting.
Suppose $K$ is a finite complex and $E$ is a flat, finite dimensional,
complex vector bundle with base $K$.
We recall that a flat vector bundle over $K$ is essentially the
same thing as a representation of $\pi_1(K)$ when $K$ is
connected.
If $p\in K$ is a basepoint then one may move the fibre at $p$
in a locally constant way around a loop in $K$. This
defines an action of $\pi_1(K)$ on the fibre $E_p$ of $E$
above $p$. We call this action the holonomy representation
$\rho:\pi\to GL(E_p)$.
Conversely, given a representation $\rho:\pi\to GL(V)$
of $\pi$ on a finite dimensional complex vector space $V$,
one may define a bundle $E=E_\rho=(\tilde{K}\times V) / \pi$.
Here $\tilde{K}$ is the universal cover of $K$, and
$\pi$ acts on $\tilde{K}$ by covering tranformations and on $V$
by $\rho$.
The holonomy of $E_\rho$ is $\rho$, so the two constructions
give an equivalence of flat bundles and representations of $\pi$.
If $K$ is not connected then it is simpler to work with
flat bundles. One then defines the holonomy as a
representation of the direct sum of $\pi_1$ of the
components of $K$. In this way, the equivalence of
flat bundles and representations is recovered.
Suppose now that one has on each fibre of $E$ a positive density
which is locally constant on $K$.
In terms of $\rho_E$ this assumption just means
$\mid\det\rho_E\mid=1$.
Let $V$ denote the fibre of $E$.
Then the cochain complex $C^i(K;E)$ with coefficients in $E$
can be identified with the direct sum of copies
of $V$ associated to each $i$-cell $\sigma$ of $K$.
The identification is achieved by choosing a basepoint in each
component of $K$ and a basepoint from each $i$-cell.
By choosing a flat density on $E$ we obtain a
preferred density $\Delta_
i$ on $C^i(K,E)$. One defines the
R-torsion of $(K,E)$ to be
$\tau(K;E)=\tau(C^*(K;E),\Delta_i)\in(0,\infty)$.
\subsection[The Reidemeister zeta function and Reidemeister torsion]{The Reidemeister type zeta functions and the Reidemeister torsion of the mapping Torus.}
Let $f:X\rightarrow X$ be a homeomorphism of
a compact polyhedron $X$.
Let $T_f := (X\times I)/(x,0)\sim(f(x),1)$ be the
mapping torus of $f$.
We shall consider the bundle $p:T_f\rightarrow S^1$
over the circle $S^1$.
We assume here that $E$ is a flat, complex vector bundle with
finite dimensional fibre and base $S^1$. We form its pullback $p^*E$
over $T_f$.
Note that the vector spaces $H^i(p^{-1}(b),c)$ with
$b\in S^1$ form a flat vector bundle over $S^1$,
which we denote $H^i F$. The integral lattice in
$H^i(p^{-1}(b),{\mathbb R})$ determines a flat density by
the condition
that the covolume of the lattice is $1$.
We suppose that the bundle $E\otimes H^i F$ is acyclic for all
$i$. Under these conditions D. Fried \cite{fri2} has shown that the bundle
$p^* E$ is acyclic, and we have
\begin{equation}
\tau(T_f;p^* E) = \prod_i
\tau(S^1;E\otimes H^i F)^{(-1)^i}.
\end{equation}
Let $g$ be the preferred generator of the group
$\pi_1 (S^1)$ and let $A=\rho(g)$ where
$\rho:\pi_1 (S^1)\rightarrow GL(V)$.
Then the holonomy around $g$ of the bundle $E\otimes H^i F$
is $A\otimes f^*_i$.
Since $\tau(E)=\mid\det(I-A)\mid$ it follows from (16)
that
\begin{equation}
\tau(T_f;p^* E) = \prod_i \mid\det(I-A\otimes f^*_i)\mid^{(-1)^i}.
\end{equation}
We now consider the special case in which $E$ is one-dimensional,
so $A$ is just a complex scalar $\lambda$ of modulus one.
Then in terms of the rational function $L_f(z)$ we have \cite{fri2}:
\begin{equation}
\tau(T_f;p^* E) = \prod_i \mid\det(I-\lambda .f^*_i)\mid^{(-1)^i}
= \mid L_f(\lambda)\mid^{-1}.
\end{equation}
From this formula and Theorem \ref{fingenabelian}
we have
\begin{teo}
Let $\phi:G\to G$ be an automorphism of a finitely generated abelian group $G$. If $G$ is infinite then one has
$$
\tau\left(T_{\hat{\phi}};p^*E\right)
=
\mid L_{\hat{\phi}}(\lambda) \mid^{-1}
=
\mid Z_{\phi}(\sigma\lambda) \mid^{(-1)^{r+1}},
$$
and if $G$ is finite one has
$$
\tau\left(T_{\hat{\phi}};p^*E\right)
=
\mid L_{\hat{\phi}}(\lambda) \mid^{-1}
=
\mid Z_{\phi}(\lambda) \mid^{-1}.
$$
where $\lambda$ is the holonomy of the one-dimensional
flat complex bundle $E$ over $S^1$, $r$ and $\sigma$ are the constants described in Theorem \ref{fingenabelian} .
\end{teo}
\subsection{Examples}
\begin{ex}\label{Kazhdan}
Let $\Gamma$ be a locally compact group. The following statements
are equivalent(see \cite{BHV}):
i) $ \Gamma$ has Kazhdan's Property (T);
(ii) $1_ {\Gamma}$ is isolated in $\wh \Gamma$;
(iii) every finite dimensional irreducible unitary representation of $\Gamma$ is isolated in $\wh \Gamma$;
(iv) some finite dimensional irreducible unitary representation of $\Gamma$ is isolated in $\wh\Gamma$.
This implies immediately that for an endomorphism of a locally compact group $\Gamma$ with Kazhdan's Property (T) the following zeta functions coincide: $RT^f_\phi(z)= AM^f_\phi(z)$.
\end{ex}
Now let us present some examples of Theorem \ref{finite}
for discrete groups with extreme properties.
Suppose, an infinite discrete group
$G$ has a finite number of conjugacy classes.
Such examples can be found in \cite{crelle}.
\begin{ex}\label{ex:osingroup}
For the Osin group (see \cite{Osin}) there is only trivial(1-dimensional)
finite-dimensional representation.
Indeed, the Osin group is an infinite finitely generated group $G$ with exactly two conjugacy classes.
All nontrivial elements of this group $G$ are conjugate. So, the group $G$
is simple, i.e. $G$ has no nontrivial normal subgroup.
This implies that group $G$ is not residually finite
(by definition of residually finite group). Hence,
it is not linear (by Mal'cev theorem)
and has no finite-dimensional irreducible unitary
representations with trivial kernel. Hence, by simplicity of $G$, it has no
finite-dimensional
irreducible unitary representation with nontrivial kernel, except for the
trivial one.
Let us remark that the Osin group is non-amenable, contains the free
group in two generators $F_2$,
and has exponential growth.
Let $\phi:G\rightarrow G$ be any endomorphism of Osin group $G$.
Thus, we have the following:
$ RT^f(\phi^n)=RT^{ff}(\phi^n)=1$ for all $n$. This implies that for any endomorphism of Osin group $G$
zeta functions $$ RT^{f}_\phi(z)=RT^{ff}_\phi(z)=\frac{1}{1-z} $$ are rational.
\end{ex}
\begin{ex}\label{ex;ivanovgroup}
For large enough prime numbers $p$,
the first examples of finitely generated infinite periodic groups
with exactly $p$ conjugacy classes were constructed
by Ivanov as limits of hyperbolic groups.
The Ivanov group $G$ is an infinite periodic
2-generator group, in contrast to the Osin group, which is torsion free.
The Ivanov group $G$ is also a simple group see \cite{crelle}.
The discussion can be completed
in the same way as in the case of the Osin group.
\end{ex}
\begin{ex}
G.~Higman, B.~H.~Neumann, and H.~Neumann
proved that any locally infinite countable group $G$
can be embedded into a countable group $G^*$ in which all
elements except the unit element are conjugate to each other
(see \cite{serrtrees}).
The discussion above related to the Osin group remains valid for $G^*$
groups.
\end{ex}
\section{Reduction to subgroups and quotient groups}
\subsection{Reduction to subgroups}
The following lemma is useful for calculating Reidemeister numbers and zeta functions.
It will also be used in the proofs of the theorems of this chapter.
\begin{lem}\label{lem:subRei}
Let $\phi:G\to G$ be any endomorphism
of any group $G$, and let $H$ be a subgroup
of $G$ with the properties
$$
\phi(H) \subset H
$$
$$
\forall x\in G \; \exists n\in {\mathbb N} \hbox{ such that } \phi^n(x)\in H.
$$
Then
$$
R(\phi) = R(\phi_H),
$$
where $\phi_H:H\to H$ is the restriction of $\phi$ to $H$.
If all the numbers $R(\phi^n)$ are finite then
$$
R_\phi(z) = R_{\phi_H}(z).
$$
\end{lem}
\begin{proof}
Let $x\in G$. Then there is an $n$ such that $\phi^n(x)\in H$.
It is known that $x$ is $\phi$-conjugate
to $\phi^n(x)$ (see Lemma 7 \cite{fh}).
This means that the $\phi$-conjugacy class $\{x\}_\phi$
of $x$ has non-empty intersection with $H$.
Now suppose that $x,y\in H$ are $\phi$-conjugate,
ie. there is a $g\in G$ such that
$$gx=y\phi(g).$$
We shall show that $x$ and $y$ are $\phi_H$-conjugate,
ie. we can find a $g\in H$ with the above property.
First let $n$ be large enough that $\phi^n(g)\in H$.
Then applying $\phi^n$ to the above equation we obtain
$$ \phi^n(g) \phi^n(x) = \phi^n(y) \phi^{n+1}(g). $$
This shows that $\phi^n(x)$ and $\phi^n(y)$ are $\phi_H$-conjugate.
On the other hand, one knows by Lemma 7 that $x$ and $\phi^n(x)$ are
$\phi_H$-conjugate, and $y$ and $\phi^n(y)$ are $\phi_H$ conjugate,
so $x$ and $y$ must be $\phi_H$-conjugate.
We have shown that the intersection with $H$ of a
$\phi$-conjugacy class in $G$ is a $\phi_H$-conjugacy class
in $H$.
We therefore have a map
$$
\begin{array}{cccc}
Rest : & {\mathcal R}(\phi) & \to & {\mathcal R}(\phi_H)\\
& \{x\}_\phi & \mapsto & \{x\}_\phi \cap H
\end{array}
$$
This clearly has the two-sided inverse
$$
\{x\}_{\phi_H} \mapsto \{x\}_\phi.
$$
Therefore $Rest$ is a bijection and $R(\phi)=R(\phi_H)$.
\end{proof}
\medskip
Let $Z(\phi)$ be one of the numbers
$RT(\phi), RT^f(\phi),
RT^{ff}(\phi)$. We shal write $ \mathcal Z(\phi)$ for one of the corresponding sets $\mathcal{RT}(\phi), \mathcal{RT}^f(\phi),
\mathcal{RT}^{ff}(\phi)$ of equivalence classes of irreducible representations.
\begin{lem}\label{lem:subrep}
Let $\phi:G\to G$ be any endomorphism
of Abelian-by-finite group $G$, and let $H$ be a subgroup
of $G$ with the properties
$$
\phi(H) \subset H
$$
$$
\forall x\in G \; \exists n\in {\mathbb N} \hbox{ such that } \phi^n(x)\in H.
$$
Then
$$
Z(\phi) = Z(\phi_H),
$$
where $\phi_H:H\to H$ is the restriction of $\phi$ to $H$.
If all the numbers $Z(\phi^n)$ are finite then
$$
Z_{\phi}(z)= Z_{\phi_H}(z).
$$
\end{lem}
\begin{proof}
All irreducible representations of Abelian-by-finite group $G$ are finite dimensional.
Let $\rho : G\rightarrow U(V)$ be irreducible representation, and suppose that there is a
matrix $M \in U(V)$ with
$$\rho\circ\phi= M\cdot\rho\cdot M^{-1},$$ i.e $ \rho \in\mathcal Z(\phi).$
Suppose that $\rho_H$, the restriction of $\rho$ to $H$, is a reducible representation i.e.
there is a decomposition $ V=V_1\oplus V_2$ into $H$-modules. We shall derive a contradiction.
We can find $g\in G$ such that $\rho(g)V_1 \not\in V_1$. However for sufficiently large $n$ we have
$\phi^n(g)\in H$. This shows that $\rho(H)M^nV_1 \not\in M^nV_1$. However since $H$ is $\phi$-invariant
and $V_1$ is $H$-invariant, $M^nV_1=V_1$. Therefore $\rho(H)M^nV_1 \not\in V_1$, which gives us a
contradiction. Consequently $\rho$ must be an irreducible representation of $H$ on $M^nV$. However
$M^nV=V$, so the representation $\rho_H$ is irreducible. Clearly the class of $\rho_H$ is the same as the
class of $\rho_H\circ\phi_H$, i.e $ \rho_H \in\mathcal Z(\phi_H).$ We thus have a map
$$ Rest: \mathcal Z(\phi)\rightarrow \mathcal Z(\phi_H), \\
\rho \rightarrow \rho_H .$$
Now let $ \rho_H \in\mathcal Z(\phi_H)$ be given. Then there is a matrix $M$ such that
$$\rho\circ\phi_H= M\cdot\rho\cdot M^{-1}.$$
If $M'$ is any other such matrix then $ M'\cdot M^{-1}$ commutes with $ \rho_H(x)$ for all $x$.
It follows that for $g\in \phi^{-n}(H)$ the element
$$ M^{-n}\cdot\rho(\phi^n(g))\cdot M^{n}$$ is independent of the chosen $M$, and depends only on $\rho, g$
and $n$. Now suppose that $\phi^n(g)=h_1\in H$ and $\phi^m(g)=h_2\in H, m > n$.
Then $\phi^{m-n}(h_1)=h_2$, which implies
$$M^{m-n}\cdot\rho(h_1)\cdot M^{n-m}=\rho(h_2),$$
and therefore
$$M^{-n}\cdot\rho(\phi^n(g))\cdot M^{n}=M^{-m}\cdot\rho(\phi^m(g))\cdot M^{m}.$$
The above expression is thus independent of $M$ and $n$, and depends only on $\rho$
and $g$. We may therefore define for $g\in G$
$$\bar\rho(g)=M^{-n}\cdot\rho(\phi^n(g))\cdot M^{n}$$
where $n$ is large enough that $\phi^n(g)\in H$. One can easily checks that $\bar\rho$ is a
representation of $G$. Since $\rho_H$ is irreducible it follows that $\bar\rho$ is irreducible.
One sees immediately that the class of $\bar\rho$ is the same as the
class of $\bar\rho\circ\phi$, i.e $ \rho \in\mathcal Z(\phi).$
Finally we have $$Rest(\bar\rho)=\rho$$ and since any other extension $\tilde\rho$ of $\rho$ to $G$
such that $\tilde\rho \in \mathcal Z(\phi)$
must satisfy $$\tilde\rho(g)=M^{-n}\cdot\rho(\phi^n(g))\cdot M^{n},$$ we have
$$\overline{Rest(\rho)}=\rho.$$
This shows that Rest is a bijection, so $\mathcal Z(\phi)= \mathcal Z(\phi_H).$
\end{proof}
\begin{cor}
Let $H=\phi^n(G)$. Suppose that all the numbers $R(\phi^k)$ and $Z(\phi^k)$, $k\in N$ are finite. Then $$R(\phi) = R(\phi_H), Z(\phi) = Z(\phi_H),
R_\phi(z) = R_{\phi_H}(z), Z_{\phi}(z)=Z_{\phi_H}(z)$$.
\end{cor}
\begin{teo}\label{teo:main1}
Let $\phi:G\to G$ be any endomorphism
of Abelian-by-finite group $G$, and let $H$ be a subgroup
of $G$ with the properties
$$
\phi(H) \subset H
$$
$$
\forall x\in G \; \exists n\in {\mathbb N} \hbox{ such that } \phi^n(x)\in H.
$$
Suppose that all the numbers $R(\phi^k)$ and $Z(\phi^k)$, $k\in N$ are finite.
If one of the following conditions is satisfied:
\begin{itemize}
\item [1.] $H$ is a finitely generated abelian group, or
\item [2.] $H$ is a finite group, or
\item [3.] $H$ is a crystallographic group
with diagonal holonomy ${\mathbb Z}_2$ and $\phi_H$ is an automorphism,
\end{itemize}
then
$$
R_\phi(z) = R_{\phi_H}(z)=Z_{\phi_H}(z)=Z_{\phi}(z)
$$
and these zeta functions are rational functions.
\end{teo}
\begin{proof}
TBFT (resp., TBF$T_f$, TBF$T_{ff}$) for
an endomorphism $\phi:G\to G$ and its iterations were proven in \cite{fh} for finitely generated
abelian groups and for finite groups.
Any crystallographic group
with diagonal holonomy ${\mathbb Z}_2$ is
a polycyclic-by-finite group and it has only finite dimensional irreducible representations. In \cite{RJMP, crelle} twisted Burnside-Frobenius theorem($TBFT_f$ and $TBFT_{ff}$) was proven for automorphisms of polycyclic-by-finite groups.
These results implie equality of the Reidemeister zeta function $R_{\phi_H}(z)$ and the zeta function $Z_{\phi_H}(z)$.
Hence from Lemma \ref{lem:subRei} and Lemma \ref{lem:subrep} it follows
that
$$
R_\phi(z) = R_{\phi_H}(z)=Z_{\phi_H}(z)=Z_{\phi}(z).
$$
In \cite{fh} the rationality of the Reidemeister zeta function $R_\phi(z)$
was proven for endomorphisms of finitely generated abelian groups and for finite groups and in
\cite{DekTerBus} the rationality of $R_\phi(z)$ was proven for automorphisms of almost-crystallographic groups with diagonal holonomy $Z_2$. This completes the proof.
\end{proof}
\medskip
\subsection{ Reduction to Injective Endomorphisms on quotient groups}
Let $G$ be a group and $\phi : G\rightarrow G$ an endomorphism.
We shall call an element $x\in G $ nilpotent if there is
an $n\in {\mathbb N} $ such that $\phi^n(x)= e$.
Let $N$ be the set
of all nilpotent elements of $G$.
Let $Z(\phi)$ be one of the numbers
$ RT^f(\phi),
RT^{ff}(\phi)$ and $ \mathcal Z(\phi)$ one of the corresponding sets of equivalence classes of irreducible representations.
\begin{teo}
The set $N$ is a normal subgroup of $G$. We have $ \phi(N) \subset N$ and
$\phi^{-1}(N)=N$.
Thus $\phi$ induces an endomorphism $[\phi/N]$ of the quotient group $G/N$ given by $[\phi/N](xN):=\phi(x)N$.
The endomorphism $[\phi/N]: G/N \rightarrow G/N$ is injective,
and we have
$$
R(\phi) = R([\phi/N]),\,\, \,\,Z(\phi) = Z([\phi/N]).
$$
Let the numbers $R(\phi^n)$ and $Z(\phi^n)$ be all finite.
Then
$$
R_\phi(z) = R_{[\phi/N]}(z), \,\, \,\, Z_{\phi}(z)=Z_{[\phi/N]}(z).
$$
If the quotient group $G/N$ is polycyclic then one has the following Gauss congruences
for Reidemeister numbers:
$$
\sum_{d\mid n} \mu(d)\cdot R(\phi^{n/d}) \equiv 0 \mod n
$$
for any $n$.
If one of the following conditions is satisfied:
\begin{itemize}
\item [1.] The quotient group $G/N$ is a finitely generated abelian group, or
\item [2.] $G/N$ is a finite group, or
\item [3.] $G/N$ is a finitely generated torsion free nilpotent group, or
\item [4.] $G/N$ is a crystallographic group
with diagonal holonomy ${\mathbb Z}_2$ and $[\phi/N]$ is an automorphism,
\end{itemize}
then
$$
R_\phi(z) = R_{[\phi/N]}(z)=Z_{[\phi/N]}(z)=Z_{\phi}(z)
$$
and these zeta functions are rational functions.
\end{teo}
\begin{proof}
(i) Let $x\in N, g\in G$. Then for some $n\in {\mathbb N}$ we have $\phi^n(x)= e $.
Therefore $\phi^n(gxg^{-1})=\phi^n(gg^{-1})= e$. This shows that $gxg^{-1}\in N$
so $N$ is a normal subgroup of $G$.
(ii) Let $x\in N$ and choose $n$ such that $\phi^n(x)=e$. Then $\phi^{n-1}(\phi(x))=e$ so $\phi(x)\in N$. Therefore $ \phi(N) \subset N$.
(iii) If $\phi(x)\in N$ then there is an $n$ such that $\phi^n(\phi(x))=e$. Therefore $\phi^{n+1}(x)=e$ so $x\in N$. This show that
$\phi^{-1}(N)\subset N$. The converse inclusion follows from (ii).
(iv) We shal write $ {\mathcal R}(\phi)$ for the set of $\phi$-conjugacy classes
of elements of $G$. We shall now show that the map $x\rightarrow xN$
induces a bijection $ {\mathcal R}(\phi)\rightarrow {\mathcal R}([\phi/N])$.
Suppose $x,y\in G$ are $\phi$-conjugate. Then there is a $g\in G$ with
$gx=y\phi(g).$ Projecting to the quotient group $G/N$ we have
$gnxN=yN\phi(g)N$, so $gNxN=yN[\phi/N](gN)$. This means that $xN$ and $yN$
are $[\phi/N]$-conjugate in $G/N$. Conversely suppose that $xN$ and $yN$
are $[\phi/N]$-conjugate in $G/N$. Then there is a $gN \in G/N$
such that $gNxN=yN[\phi/N](gN)$. In other words $gx\phi(g)^{-1}y^{-1})=e.$
Therefore $\phi^n(g)\phi^n(x)=\phi^n(y)\phi^n(\phi(g)).$
This shows that $\phi^n(x)$ and $\phi^n(y)$ are $\phi$-conjugate.
However $x$ and $\phi^n(x)$ are $\phi$-conjugate as are
$y$ and $\phi^n(y)$. Therefore $x$ and $y$ are $\phi$-conjugate.
(v) We have shown that $x$ and $y$ are $\phi$-conjugate iff $xN$ and $yN$
are $[\phi/N]$-conjugate. From this it follows that $x\rightarrow xN$
induces a bijection from $ {\mathcal R}(\phi)$ to $ {\mathcal R}([\phi/N])$.
Therefore $R(\phi)=R([\phi/N])$.
(vi) We shall now show that $Z(\phi)=Z([\phi/N])$. Let $\rho \in \mathcal Z(\phi)$
and let $M$ be a transformation for which
$$ \rho\circ\phi= M\cdot\rho\cdot M^{-1}$$
if $x\in N$ then there is an $n\in \mathbb N$ with $\phi^n(x)=e$. Thus $ N$ is contained in the kernel of $\rho$ and there is a representation $[\rho/N]$ of $G/N$ given by
$$[\rho/N](gN):= \rho(g).$$
Since $[\rho/N]$ satisfies identity
$$ [\rho/N]\circ[\phi/N]= M\cdot[\rho/N]\cdot M^{-1},$$
we have $[\rho/N]\in \mathcal Z([\phi/N])$.
(vii) Conversely if $\rho \in \mathcal Z([\phi/N])$ we may construct a $\bar\rho \in \mathcal Z(\phi)$ by
$$ \bar\rho(x):= \rho(xN).$$
It is clear that
$ \overline{[\rho/N]}=\rho$ and $ \bar\rho/N = \rho$.
In \cite{RJMP, crelle} the twisted Burnside-Frobenius theorem($TBFT_f$ and $TBFT_{ff}$) was proven for endomorphisms of polycyclic groups and for automorphisms of polycyclic-by-finite groups.
Then from (v) and (vi) it follows that
$$R(\phi^n)=R([\phi/N]^n)=Z([\phi/N]^n)= Z(\phi^n).$$
Gauss congruences now follow from Corollary \ref{cor:periodicanddyn} and the
general theory of congruences for periodic points
(cf. \cite{Smale,Z}).
More precisely, let $P_d$ be the number of
periodic points of least period $d$
of the dynamical system of Corollary \ref{cor:periodicanddyn}.
Then $R(\phi^n)= Z(\phi^n)=\sum\limits_{d|n} P_d$.
By the
M\"obius inversion formula,
$$
\sum\limits_{d|n} \mu(d) R({\varphi}^{n/d})=P_n \equiv 0
\mod n,
$$
since number $P_n$ is always divisible by $n$,
because $P_n$ is exactly $n$ times the number of
orbits of cardinality $n$.
Twisted Burnside-Frobenius theorem($TBFT_f$ and $TBFT_{ff}$) implies also equality of Reidemeister zeta function $R_\phi(z)$ and zeta function $RT^f_\phi(z)=RT^{ff}_\phi(z))$. In \cite{Fel00} the rationality of the Reidemeister zeta function $R_\phi(z)$
was proven for endomorphisms of finitely generated abelian groups and for endomorphisms of finitely generated torsion free nilpotent groups, and in
\cite{DekTerBus} the rationality of $R_\phi(z)$ was proven for automorphisms of crystallographic groups with diagonal holonomy $Z_2$. This completes the proof.
\end{proof}
\section{P\'olya -- Carlson dichotomy for Reidemeister zeta function}
In this section we present results in support of
a P{\'o}lya--Carlson dichotomy between rationality and a
natural boundary for the analytic behaviour of the Reidemeister zeta function
of Abelian group endomorphisms.
Let $\phi: G \rightarrow G$ be a endomorphism of a countable Abelian group $G$
that is a subgroup of~$\mathbb{Q}^d$, where $d\geqslant 1$.
Let $R=\mathbb{Z}[t]$ be a polynomial ring. Then the Abelian group $G$ naturally carries the structure of a $R$-module over the ring $R=\mathbb{Z}[t]$ where multiplication by~$t$
corresponds to application of the endomorphism:
$ tg=\phi(g)$ and extending this in a natural way to polynomials.
That is, for $g\in G$ and $ f=\sum_{n\in \mathbb{Z} }c_n t^n \in R=\mathbb{Z}[t]$ set
$$ fg= \sum_{n\in \mathbb{Z} }c_n t^ng=\sum_{n\in \mathbb{Z} }c_n \phi^n(g),$$
where all but finitely many $c_n\in \mathbb{Z}$ are zero.
This is a standard procedure for the
study of dual automorphisms of compact Abelian
groups, see Schmidt~\cite{Sch} for an overview.
Let us now consider the Pontryagin dual group $\wh G$ and dual endomorphism
$\widehat{\phi}:\rho\mapsto \rho\circ\phi$ on the $\wh G$.
We shall require the following statement:
\begin{lem}\cite{fh}\label{dual}
Let $\phi:G\to G$ be an endomorphism of an Abelian group $G$.
Then the kernel $\ker\left[\hat{\phi}:\hat{G}\to\hat{G}\right]$
is canonically isomorphic to the
Pontryagin dual of ${\rm Coker\ }\phi$.
\end{lem}
\begin{proof}
We construct the isomorphism explicitly.
Let $\chi$ be in the dual of ${\rm Coker\ }(\phi:G\to G)$.
In that case $\chi$ is a homomorphism
$$ \chi : G / {\rm Im\ }(\phi) \longrightarrow U(1).$$
There is therefore an induced map
$$ \overline{\chi} : G \longrightarrow U(1) $$
which is trivial on ${\rm Im\ }(\phi)$.
This means that $\overline{\chi}\circ\phi$ is trivial, or
in other words $\hat{\phi}(\overline{\chi})$ is the identity element of $\hat{G}$.
We therefore have $\overline{\chi}\in\ker(\hat{\phi})$.
If on the other hand we begin with $\overline{\chi}\in\ker(\hat{\phi})$,
then it follows that $\chi$ is trivial on ${\rm Im\ }\phi$, and so
$\overline{\chi}$ induces a homomorphism
$$ \chi : G / {\rm Im\ }(\phi) \longrightarrow U(1)$$
and $\chi$ is then in the dual of ${\rm Coker\ }\phi$.
The correspondence $\chi\leftrightarrow\overline{\chi}$ is
clearly a bijection.
\end{proof}
The following results are also needed
\begin{lem}\cite{Mi}\label{miles1}
Let $L \subset N$ be $R$-modules and $g\in R$.\\
Then
\\
(1) $$ \Bigl |\frac{N}{gN}\Bigr | = \Bigl |\frac{N/L}{g(N/L)} \Bigr |\Bigl |\frac{L}{L \cap gN} \Bigr |$$\\
(2) If $N/L$ is finite and the map $x\to gx$ is a monomorphism of $N$ then
$$\Bigl |\frac{N}{gN}\Bigr |=\Bigl |\frac{L}{gL}\Bigr |.$$
\end{lem}
Suppose that $G$ as an $R$- module satisfies the following conditions:
(1) the set of associated primes $Ass(G)$ is finite and consists entirely of non-zero
principal ideals of $R$,
(2) the map $g \rightarrow (t^j -1)g$ is a monomorphism of $G$ for all $j\in \mathbb{N}$\\
(equivalently, $t^j - 1 \notin \mathfrak{p}$ for all $ \mathfrak{p}\in Ass(G)$ and all $j\in
\mathbb{N}$),
(3) for each $\mathfrak{p}\in Ass(G)$, $ m(\mathfrak{p})= \dim_{\mathbb{K}(\mathfrak{p})}G_{\mathfrak{p}} < \infty$,
where $\mathbb{K}(\mathfrak{p})$ denotes the field of fractions of $R/\mathfrak{p}$ and $G_{\mathfrak{p}} = G\otimes_R\mathbb{K}(\mathfrak{p})$ is the localization of the module $G$ at $\mathfrak{p}$.
\begin{lem}\cite{Mi}\label{miles2}
Let N be an $R$-module for which $Ass(N)$ consists of finitely many non-trivial
principal ideals and suppose $ m(\mathfrak{p})= \dim_{\mathbb{K}(\mathfrak{p})}N_{\mathfrak{p}} < \infty$. If $g\in R$ is such that the map $x \to gx$ is a monomorphism of $N$, then
$N/gN$ is finite.
\end{lem}
If the Pontryagin dual endomorphism $\widehat{\phi}$ is an ergodic finite entropy
epimorphism of the compact connected Abelian group $\wh G$
of finite dimension $d\geq 1$ then the endomorphism $\phi: G \rightarrow G$ satisfies the conditions (1) - (3) above. Such dual groups $\wh G$ are called solenoids(see \cite{Mi, Sch}).
For the dual endomorphism~$\widehat{\phi}:\wh G\rightarrow \wh G$,
we use the following closed periodic point counting formula taken
from~\cite[Th.~1.1]{Mi} and \cite[Pr.~14]{BMW}. Let ~$F_{\widehat{\phi}}(j)=|Fix(\widehat\phi^j)|$
denotes the number of points fixed by the
endomorphism~$\widehat{\phi}^j$. Some obvious conditions such as ergodicity and finite entropy are
necessary to ensure that $F_{\widehat{\phi}}(j)$ is finite for all $j\in \mathbb{N}.$ Let $\mathcal{P}(\mathbb{K})$ denote the
set of finite places of the algebraic number field~$\mathbb{K}$. The places
of the field $\mathbb{K}$ are the equivalence classes of absolute
values on $\mathbb{K}$. When $char(\mathbb{K})=0$, the infinite places are the archimedean ones.
All other places are said to be finite. Given a finite place of $\mathbb{K}$, there
corresponds a unique discrete valuation $v$ whose precise value group is $\mathbb{Z}$.
The corresponding normalised absolute value $|\cdot|_v= |\mathcal{R}_v|^{-v(\cdot)}$,
where $\mathcal{R}_v$ is the residue class field of $v$.
For any set
of places~$S$, we
write~$|x|_S=\prod_{v\in S}|x|_v$.
\begin{prop}\cite[Th.~1.1]{Mi}, \cite[Pr.~14]{BMW}\label{main_formula1}
If~$\widehat{\phi}:\wh G\rightarrow \wh G$ is an ergodic finite entropy automorphism of a finite
dimensional compact connected Abelian group $\wh G$ , then there exist
algebraic number fields~$\mathbb{K}_1,\dots,\mathbb{K}_n$, sets
of finite places~$P_i\subset \mathcal{P}(\mathbb{K}_i)$ and
elements~$\xi_i\in \mathbb{K}_i$, no one of which is a root of
unity for~$i=1,\dots,n$, such that for any $j\in \mathbb{N}$.
\begin{equation}\label{periodic}
F_{\widehat{\phi}}(j)=\prod_{i=1}^{n}\prod_{v\in P_i}
|\xi_i^j-1|_{v}^{-1} = \prod_{i=1}^{n}|\xi_i^j-1|_{P_i}^{-1}.
\end{equation}
\end{prop}
\begin{proof}
We outline the major steps in the proof.
Under assumptions of the proposition the number of the periodic points $F_{\widehat{\phi}}(j)$ is finite for all $j\in \mathbb{N}$. Considering abelian group $G$ as $Z[t]$-module and using a straightforward duality argument in Lemma \ref{dual}( or
in \cite[Lemma 7.2]{LSW}) we have
\begin{eqnarray*}
F_{\widehat{\phi}}(j)=|Fix(\widehat\phi^j) |= | \mathop{\rm Ker}\nolimits(\widehat\phi^j-\operatorname{Id}_{\widehat G}) | =
|\widehat{{\rm Coker\ }(\phi-\operatorname{Id}_G)}| =\\
= |{\rm Coker\ }(\phi^j - \operatorname{Id}_G) | = |G/(\phi^j-1)G| = |G/(t^j-1)G|.
\end{eqnarray*}
The multiplicative set $ U = \bigcap_{\mathfrak{p}\in Ass(G)} R-{\mathfrak{p}}$
has $U\cap ann(a) = \emptyset$ for all non-zero $a\in G$, so the natural map $G\to U^{-1}G$
is a monomorphism. Identifying localizations of $R$ with subrings of $\mathbb{Q}(t)$, the
domain $ \mathfrak{R}=U^{-1}R = \bigcap_{\mathfrak{p}\in Ass(G)} R_\mathfrak{p}$ is a finite intersection
of discrete valuation rings and is therefore a principal ideal domain \cite{Matsumura}.
The assumptions of finite entropy and finite topological dimension force $U^{-1}G $ to be a Noeterian
$ \mathfrak{R}$ - module.Hence, there is a prime filtration
$$\{0\} = G_0 \subset G_1\subset \cdot \cdot\cdot \subset G_n = U^{-1} G $$
in which $ G_i/G_{i-1} \cong \mathfrak{R}/\mathfrak{q_i}$
for non-trivial primes $\mathfrak{q_i} \subset \mathfrak{R}, 1\leq i\leq n$.
Moreover, $\mathfrak{p_i}=\mathfrak{q_i}\cap R\in Ass(G)$ for all $1\leq i\leq n$.
Identifying $G$ with its image in $U^{-1}G$ and intersecting the chain above with $G$
gives a chain
$$\{0\} = L_0 \subset L_1\subset \cdot \cdot \cdot\subset L_n = G .$$
Considering this chain of $R$-modules, for each $1\leq i\leq n$ there is an induced inclusion
$$
\frac{L_i}{L_{i-1}} \hookrightarrow \frac{G_i}{G_{i-1}} \cong \frac{\mathfrak{R}}{\mathfrak{q_i}}\cong
\mathbb{K(\mathfrak{p_i})}= K_i
$$
and $N_i = L_i/L_{i-1}$ may be considered as a fractional ideal of $E_i = R/\mathfrak{p_i}$.
Using Lemma \ref{miles1}(1),
$$ \Bigl|\frac{L_i}{(t^j - 1)L_i}\Bigr | = \Bigl|\frac{N_i}{(t^j - 1)N_i} \Bigr |\Bigl|\frac{L_{i-1}}{L_{i-1} \cap (t^j-1)L_i} \Bigr |,$$
where $1\leq i\leq n$. Let $y\in L_i$, let $\eta$ denote the image of $y$ in $N_i$ and let $\xi_i$
denote the image of $t$ in $E_i$. If $(t^j-1)y\in L_{i-1}$ then $(\xi_i^j-1)\eta = 0$.
The ergodicity assumption implies $ t^j-1\notin \mathfrak{p_i}$ so $(\xi_i^j-1)\neq 0$.
Therefore, $\eta=0$ and $y\in L_{i-1}$. It follows that $L_{i-1} \cap (t^j-1)L_i = (t^j-1)L_{i-1}$
and hence,
$$ \Bigl| \frac{L_i}{(t^j-1)L_i}\Bigr | = \Bigl|\frac{N_i}{(t^j-1)N_i}\Bigr |\Bigl|\frac{L_{i-1}}{ (t^j-1)L_{i-1}} \Bigr |,$$
Successively applying this formula to each of the modules $L_i,1\leq i\leq n$, gives,
$$ | G/(t^j-1)G| = \prod_{i=1}^{n}|N_i/(t^j-1)N_i|$$
Consider now an individual term $|N_i/(t^j-1)N_i|$. Since $E_i$ is a finitely generated domain,
\cite[Th.~1.1]{Mi}\cite[Th. 4.14]{Eisenbud} shows that the integral closure $D_i$
of $E_i$ in $K_i$ is a finitely generated Dedekind domain. Therefore, $D_i$ is finitely generated as an $E_i$-module. We may consider $I_i = D_i\otimes_{E_i}N_i$ as a
fractional ideal of $D_i$. Lemma \ref{miles2} and Lemma \ref{miles1}(2) imply that
$|N_i/(\xi_i^j-1)N_i|= |I_i/(\xi_i^j-1)I_i|$ (see \cite{Mi}).
By considering $I_i/(\xi_i^j-1)I_i$ as a $D_i$-module, finding a composition series for this module and successively localizing at each of its associated primes to obtain multiplicities, it follows that
$$|I_i/(\xi_i^j-1)I_i| = \prod_{\mathfrak{m}\in Ass(I_i/(\xi_i^j-1)I_i)}q_{\mathfrak{m}}^{\delta_\mathfrak{m}(\xi_i,I_i)},$$ where $q_{\mathfrak{m}}= |D_i/\mathfrak{m}|$ and $\delta_\mathfrak{m}(\xi_i,I_i) = \dim_{D_i/\mathfrak{m}}(I_i/(\xi_i^j-1)I_i)_\mathfrak{m}$. Let
$$P_i=\{\mathfrak{m}\in Spec(D_i) : I_{\mathfrak{m}} \neq K_i\}.$$
It follows that the product above may be taken over all $\mathfrak{m}\in P_i$ to yield the same result. Each localization $(D_i)_{\mathfrak{m}}$ is a distinct valuation ring of $K_i$ and $P_i$ may be identified with a set of finite places of the global field $K_i$.
Hence, since $\delta_\mathfrak{m}(\xi_i,D_i) = v_\mathfrak{m}(\xi_i^j-1)$, finally we have
$$ |I_i/(\xi_i^j-1)I_i| = \prod_{\mathfrak{m}\in P_i}q_{\mathfrak{m}}^{\delta_\mathfrak{m}(\xi_i,D_i)}= \prod_{\mathfrak{m}\in P_i}q_{\mathfrak{m}}^{v_\mathfrak{m}(\xi_i^j-1)} = \prod_{\mathfrak{m}\in P_i}|\xi_i^j-1 |_{\mathfrak{m}}^{-1},$$
where $|\cdot|_{\mathfrak{m}}$ is the normalised absolute value arising from $D_{\mathfrak{m}}$. This concludes the proof.
\end{proof}
\begin{rk}
It is useful to note
that~$\mathbb{K}_i=\mathbb{Q}(\xi_i)$,~$i=1,\dots,n$. Applying the Artin product formula \cite{Weil}
to~(\ref{periodic}) gives
\begin{equation}\label{main_formula2}
F_{\widehat{\phi}}(j)=
\prod_{i=1}^{n}
|\xi_i^j-1|_{P_i^\infty\cup S_i},
\end{equation}
where~$P_i^\infty$ denotes the set of infinite places
of~$\mathbb{K}_i$ and~$S_i=\mathcal{P}(\mathbb{K}_i)\setminus
P_i$. It is also worth noting that~\cite[Rmk.~1]{Mi}
implies that~$|\xi_i|_v=1$ for all~$v\in P_i$,~$i=1,\dots,n$,
as~$\phi$ is an automorphism.
\end{rk}
The following results are needed to have
more ready access to the theory of linear recurrence sequences.
Relevant background on the connection between linear recurrence sequences
and the rationality may be found in the monograph of Everest, van der Poorten, Shparlinski and Ward \cite{EPSW}.
\begin{lem}(cf. \cite{BMW})\label{generating}
Let~$R(z)=\sum_{n= 1}^{\infty}R(\phi^n)z^n$.
If~$R_\phi(z)$ is rational then~$R(z)$ is rational.
If~$R_\phi(z)$ has analytic continuation beyond
its circle of convergence, then so too does~$R(z)$.
In particular, the existence of a natural boundary
at the circle of convergence for~$R(z)$ implies the
existence of a natural boundary for~$R_\phi(z)$.
\end{lem}
\begin{proof}
This follows from the fact that~$R(z)= z\cdot R'_\phi(z)/R_\phi(z)$.
\end{proof}
One of the important links between the arithmetic
properties of the coefficients of a complex power
series and its analytic behaviour is given
by the P{\'o}lya--Carlson theorem~\cite{Car},~\cite{Po}, ~\cite{Segal}.
\medskip{\noindent P\'{o}lya--Carlson Theorem.}
{\it A power series with integer coefficients
and radius of convergence~$1$ is either rational or has the
unit circle as a natural boundary.}\medskip
For the proof of the main theorem of this section we
use the following key result of Bell, Miles and Ward .
\begin{lem}[Lemma 17 in \cite{BMW}]\label{derivative}
Let~$S$ be a finite list of places of algebraic number fields
and, for each~$v\in S$, let~$\xi_v$ be a non-unit root in the
appropriate number field such that~$|\xi_v|_v=1$. Then the
function
\[
F(z)=\sum_{n= 1}^{\infty}f(n)z^n,
\]
where~$f(n)=\prod_{v\in S}|\xi_v^n-1|_v$ for~$n\ge1$, has the unit circle as
a natural boundary.
\end{lem}
The main results of this section
are the following counting formulas for the Reidemeister numbers and a P\'olya--Carlson dichotomy between rationality and a natural boundary for the analytic behaviour of the Reidemeister zeta function.
We follow the method of the proof of Bell, Miles and Ward in ~\cite[Theorem 15]{BMW} for the Artin--Masur zeta function of compact abelian groups automorphisms.
\begin{teo}\label{Main}
Let $\phi: G \rightarrow G$ be an automorphism of a countable Abelian group $G$
that is a subgroup of~$\mathbb{Q}^d$, where $d\geqslant 1$.
Suppose that the group $G$ as $R = \mathbb{Z}[t]$- module satisfies the following conditions:
(1) the set of associated primes $Ass(G)$ is finite and consists entirely of non-zero
principal ideals of the polynomial ring $R = \mathbb{Z}[t]$,
(2) the map $g \rightarrow (t^j -1)g$ is a monomorphism of $G$ for all $j\in \mathbb{N}$\\
(equivalently, $t^j - 1 \notin \mathfrak{p}$ for all $ \mathfrak{p}\in Ass(G)$ and all $j\in
\mathbb{N}$),
(3) for each $\mathfrak{p}\in Ass(G)$, $ m(\mathfrak{p})= \dim_{\mathbb{K}(\mathfrak{p})}G_{\mathfrak{p}} < \infty$.
Then there exist
algebraic number fields~$\mathbb{K}_1,\dots,\mathbb{K}_n$, sets
of finite places~$P_i\subset \mathcal{P}(\mathbb{K}_i)$,
$S_i=\mathcal{P}(\mathbb{K}_i)\setminus
P_i$,
and
elements~$\xi_i\in \mathbb{K}_i$, no one of which is a root of
unity for~$i=1,\dots,n$, such that
\begin{equation}\label{Reidemeister1}
R(\phi^j) = \prod_{i=1}^{n}\prod_{v\in P_i}
|\xi_i^j-1|_{v}^{-1} = \prod_{i=1}^{n}|\xi_i^j-1|_{P_i}^{-1} =
\prod_{i=1}^{n}
|\xi_i^j-1|_{P_i^\infty\cup S_i}
\end{equation}
for all $j\in \mathbb{N}$.
Suppose that the last
product in~\eqref{Reidemeister1} only involves
finitely many places and that~$|\xi_i|_v\neq 1$ for all~$v$
in the set of infinite places $P_i^\infty$ of $\mathbb{K}_i$ and all $i=1,\dots,n$. \\
Then the Reidemeister zeta function $R_{\phi}(z)$ is
either rational function or has a natural boundary at its circle of
convergence, and the latter occurs if and only if~$|\xi_i|_v=1$
for some~$v\in S_i$,~$1\le i\le n$.
\end{teo}
\begin{proof}
\medskip
The Reidemeister number of an endomorphism $\phi$ of
an Abelian group $G$ coincides with the cardinality of the quotient group $ {\rm Coker\ }(\phi-\operatorname{Id}_G)=G/{\rm Im}(\phi-\operatorname{Id}_G)$
(or ${\rm Coker\ }(\operatorname{Id}_G -\phi)=G/{\rm Im}(\operatorname{Id}_G-\phi)$).
A straightforward duality argument using Lemma \ref{dual} shows that
\begin{equation}\label{Pont}
R(\phi)= |{\rm Coker\ }(\phi-\operatorname{Id}_G) | = |\widehat{{\rm Coker\ }(\phi-\operatorname{Id}_G)}| = |\mathop{\rm Ker}\nolimits(\widehat\phi-\operatorname{Id}_{\widehat G}) | = |Fix(\widehat\phi) |.
\end{equation}
If the endomorphism $\phi: G \rightarrow G$ satisfies the conditions (1) - (3),
then the Pontryagin dual endomorphism $\widehat{\phi}$ is an ergodic finite entropy
epimorphism of the compact connected Abelian group $\wh G$
of the finite dimension $d\geq 1$ i.e. the Pontryagin dual group $\wh G$ is a solenoid(see \cite{Mi, Sch}). Hence the Reidemeister numbers $R(\phi^j)$ and the number of periodic points of the dual map $F_{\widehat{\phi}}(j)$ are finite for all $j\in \mathbb{N}$ . By (\ref{periodic}), (\ref{main_formula2}) and (\ref{Pont}) we have
\begin{equation}\label{RP}
R(\phi^j)= F_{\widehat{\phi}}(j)= \prod_{i=1}^{n}\prod_{v\in P_i}
|\xi_i^j-1|_{v}^{-1} = \prod_{i=1}^{n}|\xi_i^j-1|_{P_i}^{-1} =
\prod_{i=1}^{n}
|\xi_i^j-1|_{P_i^\infty\cup S_i}.
\end{equation}
Let~$S_i^*=\{v\in S_i:|\xi_i|_v\neq 1\}$, $S_i^{**}=\{v\in S_i:|\xi_i|_v> 1\}$ and let
$$
f(j)=\prod_{i=1}^n|\xi_i^j-1|_{S_i\setminus S_i^*}, \,\,\,\,\,\,
g(j)=\prod_{i=1}^n|\xi_i^j-1|_{P_i^\infty\cup S_i^*}.
$$
So,~$R(\phi^j)=f(j)g(j)$ by~(\ref{Reidemeister1}).
By the ultrametric property
$$
g(j)=\prod_{i=1}^n
|\xi_i|^j_{S_i^{**}}
\cdot|\xi_i^j-1|_{P_i^\infty}.
$$
We can
expand the product over infinite places using an appropriate
symmetric polynomial to obtain an expression of the form
\begin{equation}\label{dominant}
g(j)=\sum_{I\in\mathcal{I}}d_I w_I^j,
\end{equation}
where~$\mathcal{I}$ is a finite indexing set,~$d_I\in\{-1,1\}$
and~$w_I\in\mathbb{C}$.
Furthermore, by~\eqref{dominant},
\[
R_\phi(z)
=
\exp\left(\sum_{I\in\mathcal{I}}d_I
\sum_{j= 1}^{\infty}
\frac{f(j)(w_Iz)^j}{j}\right).
\]
If~$S_i\setminus S^*_i=\varnothing$ for all~$i=1,\dots,n$,
then~$f(j)\equiv 1$, and it follows immediately
that the Reidemeister zeta function $R_\phi(z)$ is rational function.
Now suppose that~$S_i\setminus S^*_i\neq\varnothing$ for
some~$i$. As noted in
Lemma \ref{generating}, we need only exhibit
a natural boundary at the circle of convergence for
\[
\sum_{I\in\mathcal{I}}d_I \sum_{j= 1}^{\infty} f(j)(w_Iz)^j
\]
to exhibit one for~$R_\phi(z)$.
Moreover,~$\limsup_{j\rightarrow\infty}f(j)^{1/j}=1$, so for
each~$I\in\mathcal{I}$, the series
\[
\sum_{j= 1}^{\infty} f(j)(w_Iz)^j
\]
has radius of convergence~$|w_I|^{-1}$.
Since~$|\xi_i|_v\neq 1$ for all~$v\in
P_i^\infty$,~$i=1,\dots,n$, there is a dominant term~$w_J$ in
the expansion~\eqref{dominant}, for which
\[
|w_J|
=
\prod_{i=1}^{n}
|\xi_i|_{S^{**}_i}
\prod_{v\in P_i^\infty}
\max\{|\xi_i|_v,1\}
=
\prod_{i=1}^{n}
\prod_{v\in P_i^\infty\cup \mathcal{P}(\mathbb{K}_j)}
\max\{|\xi_i|_v,1\},
\]
and~$|w_J|>|w_I|$ for all~$I\neq J$ (note that~$\log|w_J|$ is
the topological entropy, as given by~\cite{LW}).
Since~$|w_J|^{-1}<|w_I|^{-1}$ for all~$I\neq J$, this means
that it suffices to show that the circle of
convergence~$|z|=|w_J|^{-1}$ is a natural boundary for~$\sum_{j= 1}^{\infty} f(j)(w_Iz)^j$. But this is the case precisely when~$\sum_{j= 1}^{\infty} f(j)z^j$ has the unit circle as a natural boundary, and this
has already been dealt with by
Lemma~\ref{derivative}.
\end{proof}
\subsection{Examples}
To give an example of irrational Reidemeister zeta function
let us consider an endomorphism $\phi: g \rightarrow 2g$ on the module
${\mathbb Z}[\frac{1}{3}]$ which is an infinitely generated abelian group.
We follow the method and the calculations of Everest, Stangoe and Ward in Lemma 4.1 in \cite{ESW}
for the Artin--Masur zeta function of the dual compact
abelian group endomorphism $\widehat{\phi}$ .
\begin{lem}(cf. Lemma 4.1 of \cite{ESW} )
The Reidemeister zeta function $R_{\phi}(z)$
has natural boundary $\vert z\vert=\frac{1}{2}$.
\end{lem}
\begin{proof}
The dual compact abelian group $\widehat{{\mathbb Z}[\frac{1}{3}]}$ is a one dimensional solenoid.
By (\ref{periodic}), (\ref{main_formula2}) and (\ref{RP}) the Reidemeister numbers of iterations of $\phi$ and the number of periodic points of the dual map $\widehat{\phi}$ are
$R(\phi^j) = |Fix({\widehat{\phi}} ^j )|=F_{\widehat{\phi}}(j)= |2^j-1|\cdot |2^j-1|_3 .$
Let $\xi(z)=\sum_{n=1}^{\infty}\frac{z^n}{n}
\vert2^n-1\vert\cdot
\vert2^n-1\vert_3 $ so the Reidemeister zeta function
$R_{\phi}(z)=\exp(\xi(z))$.
Now
\begin{eqnarray*}
\xi(z)&=&\sum_{n=0}^{\infty}
\frac{z^{2n+1}}{2n+1}
(2^{2n+1}-1)+\sum_{n=1}^{\infty}
\frac{z^{2n}}{2n}(2^{2n}-1)\vert2^{2n}-1\vert_3\\
&=&\log\left(\frac{1-z}{1-2z}\right)
-{\frac{1}{2}}\log\left(
\frac{1-z^2}{1-4z^2}\right)+
\sum_{n=1}^{\infty}
\frac{z^{2n}}{2n}(2^{2n}-1)\vert2^{2n}-1\vert_3.
\end{eqnarray*}
Notice that
\begin{eqnarray*}
\vert2^n-1\vert_3=\vert
(3-1)^n-1\vert_3=\vert3^n-n3^{n-1}+\dots+(-1)^{n-1}3n+(-1)^n-1\vert_3=\\
= \begin{cases}
\frac{1}{3}\vert n\vert_3&\mbox{if }n\mbox{ is even,}\\
1&\mbox{if }n\mbox{ is odd.}
\end{cases}
\end{eqnarray*}
In
particular,
\begin{equation}\label{exp}
\vert4^n-1\vert_3=\vert2^{2n}-1\vert_3=\textstyle\frac{1}{3}\vert2n\vert_3=\frac{1}{3}\vert
n\vert_3.
\end{equation}
Write $\frac{1}{6}\xi_1(z)$ for the last term in an
expression for $\xi(z)$ above, so by~\eqref{exp}
\begin{eqnarray*}
\xi_1(z)&=&3
\sum_{n=1}^{\infty}
\frac{z^{2n}}{n}(4^{n}-1)\vert4^{n}-1\vert_3 =
\sum_{n=1}^{\infty}
\frac{z^{2n}}{n}(4^n-1)\vert n\vert_3.
\end{eqnarray*}
Following Lemma 4.1 in \cite{ESW} we shall show that $\xi_1(z)$ has infinitely many
logarithmic singularities on the circle $|z|=\frac{1}{2}$,
each of which corresponds to a zero of the Reidemeister zeta function $R_{\phi}(z)$.
Write $3^a\divides\!\divides n$ to mean that $3^a{\mathchoice{\mathrel{\bigm|}}{\mathrel{\bigm|}}{\mathrel{|}}{\mathrel{|}}} n$ but
$3^{a+1}\mathrel{\kern-3pt\not\!\kern3.5pt\bigm|} n$. Notice that $3^a\divides\!\divides n$ if and only if
$\vert n\vert_3=3^{-a}$. Then $\xi_1$ may be split up according to
the size of $\vert n\vert_3$ as
\begin{eqnarray*}
\xi_1(z)&=&\sum_{j=0}^{\infty}\frac{1}{3^j}\sum_{3^j\Vert n}
\frac{z^{2n}}{n}(4^n-1)=\sum_{j=0}^{\infty}\frac{1}{3^j}\eta_j^{(4)}(z),
\end{eqnarray*}
where
$
\eta_j^{(a)}(z)=\sum_{3^j\divides\!\divides n}
\frac{z^{2n}}{n}(a^n-1).
$
Then
\begin{eqnarray*}
\eta_0^{(a)}(z)&=&\sum_{3^0\divides\!\divides n}\frac{z^{2n}}{n}(a^n-1)=\sum_{n=1}^{\infty}\frac{z^{2n}}{n}(a^n-1)
-\sum_{n=1}^{\infty}\frac{z^{6n}}{3n}(a^{3n}-1)\\
&=&\log\left(\frac{1-z^2}{1-az^2}\right)-
\frac{1}{3}\log\left(\frac{1-z^6}{1-a^3z^6}\right){\!\!},
\end{eqnarray*}
\begin{equation*}
\eta_1^{(4)}(z)=\sum_{3^1\divides\!\divides n}\frac{z^{2n}}{n}(4^n-1)
=\sum_{3^0\divides\!\divides n}\frac{z^{6n}}{3n}(4^{3n}-1)
=\textstyle\frac{1}{3}\eta_0^{(4^3)}(z^3),
\end{equation*}
\begin{equation*}
\eta_2^{(4)}(z)=\textstyle\frac{1}{9}\eta_0^{(4^9)}(z^9),
\end{equation*}
and so on.
Thus
\begin{equation*}
\xi_1(z)=\log\left(\frac{1-z^2}{1-(2z)^2}\right)+
2\sum_{j=1}^{\infty}\frac{1}{9^j}\log\left(\frac{1-(2z)^{2\times3^j}}{1-z^{2\times3^j}}
\right){\!\!},
\end{equation*}
so for the Reidemeister zeta function we have
\begin{equation*}
\left\vert R_{\phi}(z)\right\vert=
\left\vert\frac{1-z}{1-2z}\right\vert\cdot
\left\vert\frac{1-(2z)^{2\vphantom{3^j}}}{1-z^2}\right\vert^{1/2}\!\!\!\!\cdot
\left\vert\frac{1-z^{2\vphantom{3^j}}}{1-(2z)^{2}}\right\vert^{1/6}\!\!\!\!\cdot
\prod_{j=1}^{\infty}\left\vert
\frac{1-(2z)^{2\times3^j}}{1-z^{2\times3^j}}\right\vert^{1/3\times9^j}.
\end{equation*}
It follows that the series defining the Reidemeister zeta function $R_{\phi}(z)$ has a zero at all
points of the form $\frac{1}{2}e^{2\pi ij/3^r}$, $r\ge1$ so $\vert
z\vert=\frac{1}{2}$ is a natural boundary for the Reidemeister zeta function $R_{\phi}(z)$.
\end{proof}
|
\section{Task Overview}
The Word-in-Context (WiC) \cite{wic2019naacl} task aims to evaluate the ability of word embedding models to accurately represent context-sensitive words. In particular, it focuses on polysemous words which have been hard to represent as embeddings due to the meaning conflation deficiency \cite{CamachoCollados2018FromWT}. The task's objective is to detect if target words occurring in a pair of sentences carry the same meaning.
Recently, contextual word embeddings from ELMo \cite{peters2018deep} or BERT \cite{bert_naacl} have emerged as the successors to traditional embeddings. With this development, word embeddings have become context-sensitive by design and thus more suitable for representing polysemous words. However, as shown by the experiments of \cite{wic2019naacl}, they are still insufficient by themselves to reliably detect meaning shifts.
In this work, we propose a system designed for the larger task of Word Sense Disambiguation (WSD), where words are matched with specific senses, that can detect meaning shifts without being trained explicitly to do so. Our WSD system uses contextual word embeddings to produce sense embeddings, and has full-coverage of all senses present in WordNet 3.0 \cite{Fellbaum2000WordNetA}. In \citet{lmmsacl2019} we provide more details about this WSD system, called LMMS (Language Modelling Makes Sense), and demonstrate that it's currently state-of-the-art for WSD. For this challenge, we employ LMMS in two straightforward approaches: checking if the disambiguated senses are equal, and training a classifier based on the embedding similarities. Both approaches perform competitively, with the latter taking the second position in the challenge ranking, and the former trailing close behind even though it's tested directly on the challenge, forgoing the training and development sets.
\section{System Description}
LMMS has two useful properties: 1) uses contextual word embeddings to produce sense embeddings, and 2) covers a large set of over 117K senses from WordNet 3.0. The first property allows for comparing precomputed sense embeddings against contextual word embeddings generated at test-time (using the same language model). The second property makes the comparisons more meaningful by having a large selection of senses at disposal for comparison.
\subsection{Sense Embeddings}
Given the meaning conflation deficiency issue with traditional word embeddings, several works have focused on adapting Neural Language Models (NLMs) to produce word embeddings that are more sense-specific. In this work, we start producing sense embeddings from the approach used by recent works in contextual word embeddings, particularly context2vec \cite{Melamud2016context2vecLG} and ELMo \cite{peters2018deep}, and introduce some improvements towards full-coverage and more accurate representations.
\subsubsection{Using Supervision}
Our set of full-coverage WordNet sense embeddings is bootstrapped from the SemCor corpus \cite{Miller1994UsingAS}. Sentences containing sense-annotated tokens (or spans) are processed by a NLM in order to obtain contextual embeddings for those tokens. After collecting all sense-labeled contextual embeddings, each sense embedding ($\vec{v}_s$) is determined by averaging its corresponding contextual embeddings. Formally, given $n$ contextual embeddings $\vec{c}$ for some sense $s$:
$$\vec{v}_{s} = \frac{1}{n}\sum_{i=1}^{n}\vec{c}_i$$
In this work, we used BERT as our NLM. For replicability, these are the relevant details: 1024 embedding dimensions, 340M parameters, cased. Embeddings result from the sum of top 4 layers ([-1, -4]). Moreover, since BERT uses WordPiece tokenization that doesn't always map to token-level annotations, we use the average of subtoken embeddings as the token-level embedding.
\subsubsection{Extending Supervision}
Despite its age, SemCor is still the largest sense-annotated corpus. The lack of larger sets of sense annotations is a major limitation of supervised approaches for WSD \cite{Le2018ADD}. We address this issue by taking advantage of the semantic relations in WordNet to extend the annotated signal to other senses.
Missing sense embeddings are inferred (i.e. imputed) from the aggregation of sense embeddings at different levels of abstraction from WordNet's ontology. Thus, a synset embedding corresponds to the average of all of its sense embeddings, a hypernym embedding corresponds to the average of all of its synset embeddings, and a lexname embedding corresponds to the average of a larger set of synset embeddings. All lower abstraction representations are created before next-level abstractions to ensure that higher abstractions make use of lower-level generalizations. More formally, given all missing senses in WordNet $\hat{s} \in {W}$, their synset-specific sense embeddings $S_{\hat{s}}$, hypernym-specific synset embeddings $H_{\hat{s}}$, and lexname-specific synset embeddings $L_{\hat{s}}$, the procedure has the following stages:
$$
\begin{matrix}
(1) & if |S_{\hat{s}}| > 0 , & \vec{v}_{\hat{s}} = \frac{1}{|S_{\hat{s}}|}\sum\vec{v}_{s} , \forall \vec{v}_{s} \in S_{\hat{s}} \\\\
(2) & if |H_{\hat{s}}| > 0 , & \vec{v}_{\hat{s}} = \frac{1}{|H_{\hat{s}}|}\sum\vec{v}_{syn} , \forall \vec{v}_{syn} \in H_{\hat{s}} \\\\
(3) & if |L_{\hat{s}}| > 0 , & \vec{v}_{\hat{s}} = \frac{1}{|L_{\hat{s}}|}\sum\vec{v}_{syn} , \forall \vec{v}_{syn} \in L_{\hat{s}}
\end{matrix}
$$
\subsubsection{Leveraging Glosses}
There's a long tradition of using glosses for WSD, perhaps starting with the popular work of \citet{Lesk1986AutomaticSD}. As a sequence of words, the information contained in glosses can be easily represented in semantic spaces through approaches used for generating sentence embeddings. While there are many methods for generating sentence embeddings, it's been shown that a simple weighted average of word embeddings performs well \cite{Arora2017ASB}.
Our contextual embeddings are produced from NLMs that employ attention mechanisms, assigning more importance to some tokens over others. As such, these embeddings already come `pre-weighted' and we embed glosses simply as the average of all of their contextual embeddings (without preprocessing). We've found that introducing synset lemmas alongside the words in the gloss helps induce better contextualized embeddings (specially when glosses are short). Finally, we make our dictionary embeddings ($\vec{v}_d$) sense-specific, rather than synset-specific, by repeating the lemma that's specific to the sense alongside all of the synset's lemmas and gloss words. The result is a sense-level embedding that is represented in the same space as the embeddings we described in the previous section, and can be trivially combined through concatenation (previously $L_2$ normalized).
Given that both representations are based on the same NLM, we can make predictions for contextual embeddings of target words $w$ (again, using the same NLM) at test-time by simply duplicating those embeddings, aligning contextual features against sense and dictionary features when computing cosine similarity. Thus, we have sense embeddings $\vec{v}_s$, to be matched against duplicated contextual embeddings $\vec{c}_w$, represented as follows:
$$
\vec{v}_s = \begin{bmatrix}
||\vec{v}_s||_2\\
||\vec{v}_d||_2
\end{bmatrix},
\vec{c}_w = \begin{bmatrix}
||\vec{c}_w||_2\\
||\vec{c}_w||_2
\end{bmatrix}
$$
\subsection{Sense Disambiguation}
Having produced our set of full-coverage sense embeddings, we perform WSD using a simple Nearest-Neighbors ($k$-NN) approach, similarly to \citet{Melamud2016context2vecLG} and \citet{peters2018deep}. We match the contextual word embedding of a target word against the sense embeddings that share the word's lemma (see Figure \ref{fig:knn}). Matching is performed using cosine similarity (with duplicated features on the contextual embedding for alignment, as explained in 2.1.3), and the top match is used as the disambiguated sense.
\begin{figure}[htb]
\centering
\includegraphics[width=0.35\textwidth]{lmms_nn.pdf}
\caption{Illustration of our $k$-NN approach for WSD, which relies on full-coverage sense embeddings represented in the same space as contextualized embeddings.}
\label{fig:knn}
\end{figure}
\subsection{Binary Classification}
The WiC task calls for a binary judgement on whether the meaning of a target word occurring in a pair of sentences is the same or not. As such, our most immediate solution is to perform WSD and base our decision on the resulting senses. This approach performs competitively, but we've still found it worthwhile to use WiC's data to train a classifier based on the strengths of similarities between contextual and sense embeddings. In this section we explore the details of both approaches.
\begin{figure*}[b]
\centering
\includegraphics[width=\textwidth]{sims.pdf}
\caption{Components and interactions involved in our approaches. The sim$_n$ labels correspond to cosine similarities between the related embeddings. Sense embeddings obtained from 1-NN matches of contextual embeddings.}
\label{fig:sims}
\end{figure*}
\subsubsection{Sense Comparison}
Our first approach is a straightforward comparison of the disambiguated senses assigned to the target word in each sentence. Considering the example in Figure \ref{fig:sims}, this approach simply requires checking if the sense $cook_{v}^2$ assigned to `makes' in the first sentence equals the sense $produce_{v}^2$ assigned to the same word in the second sentence.
\subsubsection{Classifying Similarities}
The WSD procedure we describe in this paper represents sense embeddings in the same space as contextual word embeddings. Our second approach exploits this property by considering the similarities (including between different embedding types) that can be seen in Figure \ref{fig:sims}. In this approach, we take advantage of WiC's training set to learn a Logistic Regression Binary Classifier based on different sets of similarities. The choice of Logistic Regression is due to its explainability and lightweight training, besides competitive performance. We use sklearn's implementation (v0.20.1), with default parameters.
\section{Results}
The best system we submitted during the evaluation period of the challenge was a Logistic Regression classifier trained on two similarity features (sim$_1$ and sim$_2$, or contextual and sense-level). We obtained slightly better results with a classifier trained on all four similarities shown in Figure \ref{fig:sims}, but were unable to submit that system due to the limit of a maximum of three submissions during evaluation. Interestingly, the simple approach described in 2.3.1 achieved a competitive performance of 66.3 accuracy, without being trained or fine-tuned on WiC's data. Performance of best entries and baselines can be seen on Table \ref{tab:rank}.
\begin{table}[]
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Submission} & \textbf{Acc.} \\ \hline
\begin{tabular}[c]{@{}c@{}}SuperGlue\\\cite{DBLP:journals/corr/abs-1905-00537}\end{tabular} & 68.36 \\ \hline
\begin{tabular}[c]{@{}c@{}}LMMS\\(Ours)\end{tabular} & 67.71 \\ \hline
\begin{tabular}[c]{@{}c@{}}Ensemble\\\cite{GariWiC}\end{tabular} & 66.71 \\ \hline
\begin{tabular}[c]{@{}c@{}}ELMo-weighted\\\cite{AnsellWiC}\end{tabular} & 61.21 \\ \hline\hline
BERT-large & 65.5 \\ \hline
Context2vec & 59.3 \\ \hline
ELMo-3 & 56.5 \\ \hline
Random & 50.0 \\ \hline
\end{tabular}
\caption{Challenge results at the end of the evaluation period. Bottom results correspond to baselines.}
\label{tab:rank}
\end{table}
\section{Analysis}
In this section we provide additional insights regarding our best approach. In Table \ref{tab:models}, we show how task performance varies with the similarities considered.
\begin{table}[h]
\centering
\begin{tabular}{|c|r|c|c|}
\hline
\textbf{Model} & \multicolumn{1}{c|}{\textbf{sim$_n$}} & \textbf{Dev} & \textbf{Test} \\ \hline
M0 & \multicolumn{1}{c|}{N/A} & 68.18 & 66.29 \\ \hline
M1 & 1 & 67.08 & 64.64 \\ \hline
M2 & 2 & 66.93 & 66.21 \\ \hline
M3 & 1, 2 & 68.50 & 67.71 \\ \hline
M4 & 1, 2, 3, 4 & 69.12 & \textbf{68.07} \\ \hline
\end{tabular}
\caption{Accuracy of our different models. M0 wasn't trained on WiC data, the other models were trained on different sets of similarites. We submitted M3, but achieved slightly improved results with M4.}
\label{tab:models}
\end{table}
We determined that our best system (M4, using four features) obtains a precision of 0.65, recall of 0.82, and F1 of 0.73 on the development set, showing a relatively high proportion of false positives (21.6\% vs. 9.25\% of false negatives). This skewness can also be seen in the probability distribution chart at Figure \ref{fig:prob}. Additionally, we also present a ROC curve for this system at Figure \ref{fig:curve} for a more detailed analysis of the system's performance.
\section{Conclusion and Future Work}
We've found that the WiC task can be adequately solved by systems trained for the larger task of WSD, specially if they're based on contextual embeddings, and when compared to the reported baselines. Still, we've found that the WiC dataset can be useful to learn a classifier that builds on top of the WSD system for improved performance on WiC's task of detecting shifts in meaning. In future work, we believe this improved ability to detect shifts in meaning can also assist WSD, particularly in generating semi-supervised datasets. We share our code and data at \href{https://github.com/danlou/lmms}{github.com/danlou/lmms}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.45\textwidth]{prob.pdf}
\caption{Distribution of Prediction Probabilities across labels, as evaluated by our best model on the development set.}
\label{fig:prob}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.45\textwidth]{curve.pdf}
\caption{ROC curve for results of our best model on the development set.}
\label{fig:curve}
\end{figure}
\section*{Acknowledgements}
This work is financed by National Funds through the Portuguese funding agency, FCT - Funda\c{c}\~{a}o para a Ci\^{e}ncia e a Tecnologia within project: UID/EEA/50014/2019.
|
\section{Introduction}
Polarized Galactic microwave emission poses a challenge to the search for
primordial $B$-mode polarization in the Cosmic Microwave Background (CMB). The
$B$-mode signal could provide direct constraints on the energy scale of
inflation, but the Milky Way foregrounds may outshine it at all frequencies, everywhere on the sky \citep{2016A&A...586A.133P,2016arXiv161002743A, 2016MNRAS.462.2063H, 2017MNRAS.469.2821T, 2018JCAP...04..023R,2019arXiv190210541H}.
We have a limited understanding of this foreground, and we must learn more to ensure the reliability of future $B$-mode measurements.
The foreground emission involves the turbulent interplay of gas, dust, and magnetic fields in the Galaxy's interstellar medium (ISM). The magnetic fields organize the flow and control the precession of grains that give rise to the polarized dust signal. Although Planck's 353 GHz polarization channel has given us a first look, several features of the dust polarization remain without physical explanations.
For example, the amplitude of dust polarization $B$-mode power is approximately half of $E$-mode power, $A_{BB}/A_{EE} = 0.53 \pm 0.01$, when fit on a large portion of the sky (for $f_{\rm sky}^{\rm eff}=0.52$--$0.71$).
Smaller patches also show the same mean value $ A_{BB}/A_{EE} = 0.51$, with small patch-to-patch dispersion $\sigma_{BB/EE} = 0.18$ \citep{2016A&A...586A.133P,2018arXiv180104945P}. This observation defied pre-Planck expectations. Random polarization orientations, or coherent orientations overlaying random polarization intensity fluctuations, both yield equal amounts of $E$ and $B$ \citep{2001PhRvD..64j3001Z,2014PhRvL.113s1303K}.
We have some understanding of dust physics and its relationship to polarization modes.
The amplitude and orientation of the dust signal is set by the integrated
column density and magnetic field orientation.
For $E$ to have more power than $B$
qualitatively means that density fluctuations (structures in the ISM density
field) must prefer orientations parallel or perpendicular to the local magnetic
field \citep{2018arXiv180711940R}.
This picture is borne out by measurements of the magnetic field orientation in individual, bright, filamentary structures in the Planck 353 GHz data \citep{2016A&A...586A.141P}. This is further validated by the observations that linear structures in neutral hydrogen emission, highlighted by a Rolling Hough Transformation, also correlate with the magnetic field direction indicated by Planck dust polarization \citep{2014ApJ...789...82C,2015PhRvL.115x1302C}.
We do not know if such filamentary structures are the dominant contribution to the polarization foreground. There is certainly evidence for filamentary structure in data \citep{2016A&A...586A.141P} as well as in simulations of the interstellar medium \citep{2005A&A...436..585D,2013A&A...556A.153H}, but there is no clear consensus on their origin or evolution \citep{2018arXiv181010014M,2019A&A...621A...5O}. The purpose of this paper is to explore what polarization power spectra are possible for filaments, and what the observed power spectra can tell us about their physical properties.
Other aspects of the dust polarization also need physical explanations.
Both $E$-mode and $B$-mode spectra follow power laws ($C_\ell \propto \ell^\alpha$), with approximately the same slope, $\alpha_{BB} = -2.42 \pm 0.02$ and $\alpha_{EE} = -2.45 \pm 0.03$. \citep{2016A&A...586A.133P,2018arXiv180104945P}
There is a positive correlation between dust intensity and $E$-mode
polarization \citep[noted by][]{2017ApJ...839...91C}, with correlation
coefficient $r_{TE}=0.357 \pm 0.003$ \citep{2018arXiv180104945P} and
significant scatter depending on the sky area but little evidence for scale
dependence. Perhaps more intriguing is a parity-violating, positive $TB$
correlation \citep{2018arXiv180104945P}.
Finally, the amplitude of dust polarization power correlates to intensity in
patches, roughly as $\langle I\rangle_{\rm patch}^{1.9}$, for both $E$ and
$B$ \citep{2016A&A...586A.133P}.
A few works have already tried to address these observations.
\citet{2017ApJ...839...91C} examined the dust polarization power spectra of slow, fast, and Alfv\'en MHD waves in terms of two parameters: the ratio of gas to magnetic pressure, and the anisotropy of the MHD modes around the background field direction.
They found two regions of parameter space that can account for the $E$ to $B$ ratio and positive $TE$ correlation but judged that these scenarios are unlikely due to the uniformity of the polarization power spectrum across the sky, and instead suggested that Planck may be seeing large scale displacements that are driving the turbulent ISM, rather than the turbulence itself.
On the other hand, \citet{2017MNRAS.472L..10K} argued with a similar analysis that the observed $E/B$ power ratio can be realized in an MHD model, so long at the turbulent flow is sub-Alfv\'enic. \citet{2018MNRAS.478..530K} extended this analysis to examine the $TE$ correlation and synchrotron emission.
Other works approach the problem using MHD simulations. For the most part, the ISM is filled with trans- and super-sonic flows, which are non-linear \citep[e.g.][]{2004ARA&A..42..211E,2010ApJ...708.1204B}.
Both \citet{2018PhRvL.121b1104K} and \citet{2019arXiv190107079K} made MHD simulations of the ISM, and modeled the dust polarization
signals. Both works find slopes and power ratios that are
reasonably close to the observed values, but the slopes
are especially sensitive to the masking procedure.
What MHD simulations do not provide is a straightforward and direct way to understand why these polarization properties arise.
Here we seek to gain physical intuition with very simple models of polarized filaments. We compute their temperature and polarization power spectra using a method akin to the cosmological halo model \citep[e.g.][]{2000MNRAS.318..203S,2002PhR...372....1C}. However, instead of spherical halos, we use magnetized, prolate-spheroidal filaments as the basic ingredients, and integrate over their population.
We organize this paper so that, in Section~\ref{sec:method}, we describe our formalism for characterizing the filament signal and for computing the power spectra. In Section~\ref{sec:results}, we show the power spectra and discuss how the parameters of the filament population affect them. In Section~\ref{sec:discussion}, we conclude and discuss the implications and possible future directions.
An appendix describes how the distributions of filament and magnetic field orientations in three dimensions appear when projected onto the plane of the sky.
\section{Method}\label{sec:method}
We define a projected filament profile, $f(\bm{x})$, upon which we paint the temperature (i.e. intensity) and polarization signals. Thus the temperature profile is:
\begin{equation}
T(\bm{x}) = T_0 f(\bm{x}),
\end{equation}
for sky position $\bm{x}$.
We model the polarization with an overall polarization fraction and polarization direction. In terms of the Stokes parameters, the polarization for a filament is
\begin{eqnarray}
X(\bm{x}) &=& (Q+iU)(\bm{x}) \\ \nonumber
&=& f_{\rm pol} \exp(2i \psi_{\rm pol}) T_0 f(\bm{x}).
\end{eqnarray}
In the HEALPix polarization convention \citep{2005ApJ...622..759G}, the $+x$-axis points south and the polarization angle $\psi_{\rm pol}$ increases east of south.
Because we will integrate over angles in our computation of the power spectrum (and because $E$ and $B$ fields are coordinate independent) we can analyze a filament that has its long axis aligned (in projection) with the $x$-axis without loss of generality.
For simplicity, we assume that the intrinsic, microphysical contribution to $f_{\rm pol}$ is common to all filaments, although we will account for geometrical and projection effects in this work.
If the long axis of that filament were aligned with the local magnetic field, the precession of the dust grains would cause the polarization angle to be $\psi_{\rm pol} = 90^\circ$, perpendicular to the filament axis.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{figs/filamentTQEB.pdf}
\end{center}
\caption{Stokes parameters and scalar polarization quantities for idealized filaments as we alter the aspect ratio. The magnetic field (thin arrow) is vertical, so the polarization direction is horizontal (thick line), making $Q<0$ and $U=0$. In sky convention, north is to the top and east is to the left.\label{fig:filament} The filaments are 2, 5, and 20 times longer than they are wide (axis ratio $\epsilon = 0.5, 0.2,0.05$). Scalars $E$ and $B$ are on the same color scale, which has half the range of the $Q$ scale. The $T$ scale differs from $Q$ by an arbitrary polarization fraction.}
\end{figure}
Working in the flat sky approximation, the Fourier components of the scalar polarization modes are:
\begin{equation}
(E+iB)(\bm{\ell}) = \exp(-2i \phi_{\bm{\ell}}) X(\bm{\ell}).
\end{equation}
Fig.~\ref{fig:filament} shows the Stokes $T,Q,U$ and scalar $E,B$ quantities on the sky for sample, north--south filaments with $\psi_{\rm pol}=90^\circ$, so the magnetic field is parallel to the filament direction and the polarization is perpendicular. (Our choice of coordinates implies that Stokes $U$ is zero in these cases.) When the magnetic field aligns with the filament direction, \citet{2018arXiv180711940R} pointed out that the real-space kernels for the $E/B$ signals show immediately that the $E$-type polarization is positive along the filament, regardless of its orientation. Since the temperature signal is also strong there, such filaments naturally yields a strong and positive $TE$ cross-correlation, as observed in the Planck data. The same work showed that the $B$ signal is concentrated at the ends of the filament, so filaments with long and thin aspect ratios will have less $B$ power relative to $E$ power than more squat ones.
\begin{figure}
\includegraphics[width=\columnwidth]{figs/filament_anglerotation.pdf}
\caption{Like Fig.~\ref{fig:filament}, but showing the polarization quantities as we alter the magnetic field direction (and hence the polarization angle), for a fixed filament orientation. The $Q$ and $U$ Stokes parameters follow HEALPix polarization convention for a north--south filament, but the $E$ and $B$ fields are coordinate independent and appropriate for any orientation. Note that filaments with aligned magnetic fields ($\psi_{\rm pol} = 90^\circ$) have zero $TB$ correlation. Filaments with relative polarization angles $90^\circ < \psi_{\rm pol} < 180^\circ$ have positive $TB$ correlations, as depicted, while those with $0^\circ < \psi_{\rm pol} < 90^\circ$ have negative $TB$ correlations (not shown).}\label{fig:pol_angle}
\end{figure}
By parity symmetry, the $TB$ and $EB$ cross-correlations are zero when the polarization is perpendicular to the filament (i.e.\ $\psi_{\rm pol} =90^\circ$). In Fig.~\ref{fig:pol_angle} we show how the $E$ and $B$ patterns transform into each other (and change sign) as $\psi_{\rm pol}$ varies away from $90^\circ$. In these cases, the $TB$ and $EB$ correlation can be non-zero for individual filaments, but so long as the average $\langle \psi_{\rm pol} \rangle =90^\circ$, there will be no overall cross-correlation for the whole population.
\subsection{Projection on the sky}
We next discuss the projection of a three-dimensional filament onto the plane of the sky.
Many important quantities depend on the angle to the line of sight of (1) the long axis of the filament ($\theta_L$) and (2) the magnetic field vector ($\theta_H$).\footnote{Elsewhere in the ISM literature, the angles are often given with reference to the plane of the sky, e.g. $\gamma_H = 90^\circ - \theta_H$ and so on.} Another important quantity is the the plane-of-sky projection of the angle between these vectors ($\psi_{LH}$), which controls the polarization angle and the amounts of $E$/$B$ polarization present. We depict these angles in Fig.~\ref{fig:anglefig}. If the magnetic field direction aligns somewhat with the filament direction, as is the case in strong-field MHD, all these angles will be correlated.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figs/geometry.pdf}
\end{center}
\caption{Geometry of the filament direction and the magnetic field. The long axis of the filament $\mathbf{L}$ points in the $x$--$z$ plane at an angle $\theta_L$ from the line of sight. The magnetic field $\mathbf{H}$ has an angle from the line of sight of $\theta_H$. The angular misalignment between the field and the filament is $\theta_{LH}$. The front view gives the projection of the misalignment onto the plane of the sky, $\psi_{LH}$.}
\label{fig:anglefig}
\end{figure}
We assume that on average, the filaments align with the local magnetic field.
In the appendix, we use simple geometry to compute the distribution of
the magnetic field projection angle $\theta_H$ and relative orientation angle $\psi_{LH}$ as a function of $\theta_L$.
We base the distribution on the assumption of a Gaussian distribution for the angle ($\theta_{LH}$) between the filament and the magnetic field in three dimensions, characterized by the dispersion ${\rm RMS}(\theta_{LH})$.
The field angle $\theta_H$ is correlated with $\psi_{LH}$, so our numerical procedure yields the tabulated joint distribution,
\begin{equation}
p(\psi_{LH},\theta_H | \theta_L).
\end{equation}
This distribution centers on aligned filaments ($\psi_{LH} = 0^\circ,\theta_H = \theta_L$), and the distribution for $\psi_{LH}$ broadens for filaments nearly along the line of sight. Its precise form is not vital for this discussion and is plotted in the appendix in Fig.~\ref{fig:psiLH_thetaH_dist}.
On the other hand, the probability distribution for the line-of-sight angle of randomly oriented filaments is determined purely by geometry,
\begin{equation}
p(\theta_L) = \sin \theta_L,
\end{equation}
for $\theta_L \in [0,180^\circ]$.
These quantities relate immediately to the polarization. Although the dust polarization fraction depends on the microphysical details of the emission, it has a geometric dependence like $f_{\rm pol} \propto \sin^2 \theta_H$ \citep{2000ApJ...544..830F}. Meanwhile, the polarization angle for a filament projected along the $x$-axis is $\psi_{\rm pol} = \psi_{LH} + 90^\circ$.
\begin{figure}
\includegraphics[width=\columnwidth]{figs/pol_vs_orientation.pdf}
\caption{Total intensity, polarization fraction, and polarization amplitude dependence on the filament orientation. Filament oriented along the line of sight have $\theta_L = 0^\circ$, while filaments in the plane of the sky have $\theta_L = 90^\circ$. The filament is modeled with an axis ratio $\epsilon = 0.25$, and the filament direction ($\theta_L$) and the magnetic field direction ($\theta_H$) are either perfectly aligned or stochastically misaligned in three-dimensions. }\label{fig:pol_vs_thetaL}
\end{figure}
We model the filament as a prolate spheroid, and label the major axis as $L_a$ and the minor axis as $L_b$. The axis ratio is thus $\epsilon = L_b/L_a < 1$. The column density (and therefore the surface brightness and ultimately the observed temperature perturbation) is proportional to the density and the line of sight distance through the filament, and so (approximately)
\begin{eqnarray}
T_0 &\propto& \rho_0 \left( L_a^2 \cos^2 \theta_L + L_b^2 \sin^2 \theta_L \right)^{1/2} \\ \nonumber
&\propto& \rho_0 L_a \left( \cos^2 \theta_L + \epsilon^2 \sin^2 \theta_L \right)^{1/2}.
\label{eq:column_density}
\end{eqnarray}
where $\rho_0$ is a characteristic density for the filament.
So all else being equal, a filament that lies along the line of sight will have the greatest column density and the brightest temperature signal. On the other hand, $f_{\rm pol} \propto \sin^2 \theta_H$, so if the magnetic field lies along the line-of-sight, there is no polarization. The polarization fraction is maximum when the magnetic field is perpendicular to the line of sight.
Fig.~\ref{fig:pol_vs_thetaL} relates the column density and polarization fraction effects of the line-of-sight angle. It also shows that since the polarized amplitude depends on the product of these two, the filaments with the brightest polarization are inclined, but not perpendicular, to the line of sight. The polarization maximum depends on axis ratio through its impact on the column density. Nearly round filaments have the polarization maximum when oriented near 90$^\circ$ to the line of sight, while in the limit of thin filaments ($\epsilon \rightarrow 0$) the polarization maximum orientation approaches $\theta_L = 45^\circ$ for perfect magnetic field alignment. If there is significant misalignment of the magnetic field and filament directions, the situation can become more complicated, depending on the particular combination of axis ratio and misalignment dispersion. In such cases, filaments along the line of sight can have significant polarization. Still, the typical line-of-sight orientation angle for maximum polarization, averaging over the magnetic field directions, is around $\theta_L = 45^\circ$.
We compute the filament's projected angular sizes along its two axes as if it were a cylinder. These depend its distance $R$ and are:
\begin{eqnarray}
\Theta_a &=& \left( L_a^2 \sin^2 \theta_L + L_b^2 \cos^2 \theta_L \right)^{1/2} / R \\ \nonumber
&=& \left( \sin^2 \theta_L + \epsilon^2 \cos^2 \theta_L \right)^{1/2} L_a / R \\ \nonumber
\Theta_b &=& L_b/R = \epsilon L_a/R.
\end{eqnarray}
Thus the projected axis ratio is
\begin{equation}
\epsilon_\Theta = \Theta_b/\Theta_a = \frac{\epsilon}{\left( \sin^2 \theta_L + \epsilon^2 \cos^2 \theta_L \right)^{1/2}}
\end{equation}
which goes to unity for filaments along the line of sight, and to the true value ($\epsilon$) for filaments perpendicular to the line of sight.
\subsection{Filaments in Fourier space}
For several terms in our power spectrum calculation, we need the Fourier transform of the projected filament profile:
\begin{equation}
f(\bm{\ell}) = \int d^2x\ f(\bm{x}) \exp(-i \bm{\ell} \cdot \bm{x}).
\end{equation}
Rather than project rays through a 3-dimensional model to obtain the filament profile, we make a simplifying assumption for computational efficiency. From the size and orientation of a filament, we take the angular dimensions and compute under the assumption that the profile is a distortion from an axisymmetric function $g$:
\begin{eqnarray}
f(x,y) =& g(x/\Theta_a , y/\Theta_b)\ & \\ \nonumber =& g(x^*, y^*) \quad & = \ g(r)
\end{eqnarray}
where $\Theta_a,\Theta_b$ are the semi-major and semi-minor axis of the elliptical distortion. As we stated, by convention and without loss of generality, we orient the long axis of the filament along the $x$-axis.
Then the transform of $f$ is simply related to the transform of $g$:
\begin{eqnarray}
f(\bm{\ell}) &=& \Theta_a \Theta_b \int d^2x^*\ g(\bm{x}^*) \exp(-i (\Theta_a\ell_x x^* + \Theta_b \ell_y y^*)) \nonumber \\
&=& \Theta_a \Theta_b\, g( \ell^* )
\end{eqnarray}
where
$ \ell^*(\bm{\ell}) = (\Theta_a^2 \ell_x^2 + \Theta_b^2\ell_y^2)^{1/2}$
and
\begin{eqnarray}
g(\ell^*) = \int d^2x^* \ g(\bm{x}^*) \exp(i \bm{\ell}^* \cdot \bm{x}^*) \\ \nonumber
= 2\pi \int dr \ r\ g(r) J_0(\ell^* r).
\end{eqnarray}
The input profile $g(\mathbf{x}^*)$ is real and even, and so the Fourier transform is too. The power spectra we find are not very sensitive to the profile that we use. In this work we have used an exponential for the basic filament profile ($g(r) = \exp(-r)$), but we have checked our best-fitting power spectrum model with a Gaussian profile ($g(r) = \exp(-r^2/2)$) and a Plummer profile ($g(r) = (1+r^2)^{-5/2}$), and find the same results.
\subsection{Parameters and one-filament term}
For a parameterized set of filament properties, $$\alpha = ( L_a, L_b, \psi_{LH}, \theta_L, \theta_H,R,\dots),$$ we can write the number density distribution $n(\alpha)$, so that the average number of filaments in a realization of the sky is
\begin{equation}
\langle N \rangle = \int d\Omega\ d\alpha\ n(\alpha)
\end{equation}
where the integral is over
\begin{equation} d\alpha = dL_a dL_b d\psi_{LH} d\theta_H d\theta_L dR. \end{equation}
Expressed another way, $n(\alpha) = \langle N \rangle p(\alpha)$, where the normalized probability distribution of the filament population is
\begin{equation} p(\alpha) = p(L_a,L_b)p(\psi_{LH},\theta_H | \theta_L) p(\theta_L)p(R) \end{equation}
This integral over the population is at least six dimensional. For a screen at a distance $R$, it is five dimensional integral. Since the angular power spectrum for foregrounds is a power law, if we can reproduce it on a single screen, putting that screen at different distances will maintain the same power spectrum. If we further fix the physical aspect ratio of the filaments, it is a four dimensional integral, over $L_a, \psi_{LH}, \theta_L, \theta_H$. (The projected aspect ratio will still vary with the line-of-sight angle $\theta_L$.)
The power spectrum contributions from filaments correlated with themselves are:
\begin{eqnarray}
C_\ell^{TT} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ | T(\bm{\ell},\alpha)|^2, \\ \nonumber
C_\ell^{EE} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ | E(\bm{\ell},\alpha)|^2, \\ \nonumber
C_\ell^{BB} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ | B(\bm{\ell},\alpha)|^2, \\ \nonumber
C_\ell^{TE} &=& \frac{1}{2\pi} \int d\phi_\ell \int d\alpha\ n(\alpha) \ T(\bm{\ell},\alpha) E(\bm{\ell},\alpha)^*.
\end{eqnarray}
Similar expressions hold for the other cross correlations, but these vanish if the orientations of the filaments are random. These power spectra computations are directly analogous to the 1-halo term in the cosmological halo model \citep{2000MNRAS.318..203S}.
\section{Results}\label{sec:results}
There are clear relationships between the physical properties of the filaments and the temperature and polarization power spectra that they produce. The slopes of the power spectra are determined primarily by the size distribution of filaments, with other effects responsible for the smaller differences between the components. The ratio of $BB/EE$ power is determined mostly by the aspect ratio of the filaments and somewhat by the misalignment of the filament directions to the background magnetic field. These same factors also determine the cross correlation $r^{TE}_\ell$,
but here misalignment is much more important. They also affect the $TE/EE$ power ratio, but this quantity is more directly affected by the overall polarization fraction.
\subsection{Power spectrum shape}
We can relate the slope of a power law spectrum to scaling relations for parameters in the filament profiles. This allows us to place constraints on the distribution of filament sizes and the scaling of other parameters.
For a generic parameter $\alpha_0$, if the filament's contribution to the power spectrum scales as
\begin{equation}
C_\ell \propto \int d\alpha_0\ n(\alpha_0)\times \alpha_0^q \,F( \alpha_0^r\, \ell) \label{eqn:shape_scaling}
\end{equation}
for any function $F$, and furthermore if the weighting distribution for the parameter is a power law, $n(\alpha_0) \propto \alpha_0^p$, then we can rescale the integration with a straightforward substitution, $u=\alpha_0 \ell^{1/r}$:
\begin{eqnarray}
C_\ell &\propto& \ell^{-(p+q+1)/r} \times \\ && \nonumber \quad \int d(\alpha_0 \ell^{1/r})\ (\alpha_0 \ell^{1/r})^p \times (\alpha_0 \ell^{1/r})^q \,F( ( \alpha_0\, \ell^{1/r})^r) \\ \nonumber
&\propto& \ell^{-(p+q+1)/r} \times \int du\ u^p u^q \,F( u^r).
\end{eqnarray}
The integral no longer has any multipole dependence and evaluates to some constant value, whatever the details of $F$. Thus we are left with a powerlaw power spectrum with
\begin{equation}
C_\ell \propto \ell^{-(p+q+1)/r}
\end{equation}
This argument holds not just for filaments, but for any signal with a power-law power spectrum that is built from a set of objects that are similarly related to each other, so long as they are weighted by powerlaw scalings and distributions. So if we observe a powerlaw spectrum with $C_\ell \propto \ell^s$, it implies that the parameter distribution's index is $p = -rs - q - 1$, regardless of the objects' profiles.
We walk through this scaling argument for a simple (and unrealistic) case---with plane-of-sky filaments with identical surface brightnesses ($T_0$ is the same for all filaments) and a constant projected-axis-ratio ($\Theta_b = \epsilon \Theta_a$)---and analyze the distribution for $\Theta_a$, the angular size of filaments. For the filament Fourier transform, we have $f(\bm{\ell}) \propto \Theta_a^2 g(\bm{\ell}^*)$ with $\bm{\ell}^* \propto \Theta_a$. The power spectrum contribution is proportional to $f^2$, so comparing the scaling for angular size parameter $\Theta_a$ to equation~\ref{eqn:shape_scaling}, we find $q=4$ and $r=1$.
In the polarization case, to reproduce $C_\ell \propto \ell^{-2.4}$ (meaning $s=-2.4$), the number density distribution of such objects on the sky must approximately scale like $n(\Theta_a) \propto \Theta_a^p$ where $p = 2.4 - 4 - 1 = -2.6$. Indeed this yields the desired power spectrum slope when calculated in our model.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{figs/power_spectrum.pdf}
\end{center}
\caption{Power spectra slopes of temperature and polarization are set chiefly by the distribution of filament lengths, which has been chosen here so that the slopes match the dotted line, $C_\ell \propto \ell^{-2.4}$. The ratios between the power spectra are set by the overall polarization fraction, the aspect ratio of filaments, and the misalignment between the filaments and their local magnetic field. When maximum and minimum sizes truncate the distribution of the filaments, the power falls below the targeted slope (semi-transparent colors).}
\label{fig:slope}
\end{figure}
In a more realistic case, with three-dimensional filaments, we can make a similar argument to deduce the distribution of filament lengths. Surface brightness depends on column density, which is proportional to length (after integrating out any distribution of axis ratios---compare equation~\ref{eq:column_density}---and assuming a density normalization independent of length). The solid angle scales like length squared. After squaring those three powers during the computation of the power spectrum, the overall scaling is $q=6$. The multipole $\ell$ scaling should also go like length, so $r = 1$, the same as the plane-of-sky case above. So with no other dependence on length, we should have distribution of lengths $n(L_a) \propto (L_a)^p$ where $p = 2.4 - 6 - 1 = -4.6$.\footnote{If the column density normalization depends on length, this procedure yields a net distribution that is a product of the size distribution and the density squared distribution (both as a function of length). Such a case could arise, for example, if small filaments are collapsed versions of large ones and have higher density.}
We have verified that this distribution produces the proper slope in Fig.~\ref{fig:slope}. All the temperature and polarization spectra have the specified slope in common. The complications of the modeling of the three dimensional orientation are not important to the slope, only the weighting and distribution of filament size.
Other than the slope, there are not clear features in the Planck-measured spectra. We note that features in the distribution of filament sizes would break the powerlaw behavior of the resulting spectra. For example, if we impose a maximum filament size (semi-transparent lines in Fig.~\ref{fig:slope}), it causes the low-$\ell$ behavior of $C_\ell$ to deviate: at scales much larger than the filament, the temperature spectrum adopts the flat, white, Poisson spectrum of point sources. The $EE$ and $BB$ spectra flatten the same way, and on scales large compared to the filaments, the aspect ratios of the filaments become unimportant and the amount of power in $EE$ and $BB$ equalize. The $TE$ cross-correlation falls off at large scales, possibly because the positive and negative contributions to $E$ are being averaged over. If, on the other hand, we impose a minimum filament size, it causes the high-$\ell$ behavior to deviate: the spectrum will drop off with increasing $\ell$, where there is no more contribution to the power.
In yet more realistic cases, we can make the slopes of all the temperature and polarization spectra differ. For example, this can happen if there is another effect that changes the size scaling of the polarization relative to the temperature. For example, to make polarization slope shallower than the temperature slope, one could make smaller objects more polarized than large objects, or if the aspect ratio of filaments changes as a function of size.
Considering the Planck data, it is not immediately clear what conclusion to draw. \citet{2018arXiv180104945P} quote $EE, BB, TE$ slopes for the dust foreground, but not the $TT$ slope. Using our own tools, we have computed the $TT$ power spectrum based on the Planck data, and find a $TT$ slope that is about $-2.6$, somewhat steeper than the polarization spectra at $-2.4$ to $-2.5$. Like \citet{2018arXiv180104945P}, for the mask we used the LR71 polarization mask supplemented with a point source mask (based on intensity maps), resulting in a mask with effective $f_{\rm sky} \simeq 0.6$.
We can approximately reproduce a $-2.6$ temperature slope and $-2.4$ polarization slope with $n(L_a) \propto L_a^{-4.4}$ and $f_{\rm pol} \propto L_a^{-0.08}$. This argues that in the unmasked region, the polarization is higher in smaller filaments.
However, this conclusion may not be correct because it depends sensitively on the mask. We reasoned that small filaments, oriented along the line of sight, might look like a point source and be included in the masked area. These end-on filaments could also have low polarization (note Fig.~\ref{fig:pol_vs_thetaL}), and excluding them might not have much effect on the polarization results. Thus for comparison, we recomputed the $TT$ spectrum with different masks. When we use only the polarization LR71 mask without removing the additional point sources from the intensity map, we get a shallower $TT$ spectrum with slope $-2.5$. When we use a mask that keeps the same large scale features but does not mask any point sources (Planck's publicly available GAL70 mask) we find a $TT$ slope of $-2.1$, notably shallower than the polarization spectra. The polarization spectra change somewhat between these masks, but the changes in the polarization slopes are small compared to the change in the TT slopes.
Some of the masked sources are extragalactic, so this slope with all point sources unmasked is probably too shallow to describe the ISM component, but can serve as a bound. The upshot is that we are not certain whether the spectrum for all filaments is steeper or shallower in temperature than polarization, and so it is difficult to draw conclusions on the size dependence of the polarization fraction.
Another feature of the Planck data is the differing slopes in $E$ and $B$. We can reproduce this feature by varying the aspect ratio as a function of filament size. For the LR71 mask in \citet{2018arXiv180104945P}, the slopes for $(BB, EE, TE)$ are roughly $(-2.5, -2.4, -2.5)$ and we found a $TT$ slope of $-2.6$. So $BB$ is steeper than $EE$, which should happen if smaller filaments are proportionally thinner than longer ones. Modifying the aspect ratio in this way also affects the $TT$ slope, breaking the simple relation that we saw earlier in this section. By trial and error, we found that this set of slopes are approximately reproduced with the following parameter dependence: $\epsilon \propto L_a^{0.1}$, $n(L_a) \propto L_a^{-4.45}$, and $f_{\rm pol} \propto L_a^{-0.1}$. Here we are simply exploring what is possible, but the relationship between the measured slopes and these filament parameters should be made more systematic and quantitative.
In light of these complications and uncertainties, in what follows we keep a common slope of $-2.4$ for all the temperature and polarization components while we explore their other parameter dependences.
\subsection{BB/EE power ratio}
The aspect ratio of the filaments is the major factor determining the ratio of $B$-mode power to $E$-mode power. In Fig.~\ref{fig:EEBBrat}, we plot the power ratio against the aspect ratio for varying degrees of filament--magnetic field misalignment. To reproduce the Planck-observed ratio of $\sim 0.5$, filaments need to have an aspect ratio $\epsilon$ slightly less than 0.26, so filaments must be slightly less than four times longer than they are wide. If the model deviates too much from this ratio, the required magnetic field misalignment is made so large that the model has trouble fitting the $TE$ correlation.
It is difficult to compare this result quantitatively to the aspect ratios of observed filaments from the literature without making a detailed accounting of the filament selection function. Projection effects will also tend to lower observed aspect ratios. The stacked filaments in Fig.~7 of \citet{2016A&A...586A.141P} appear to have axis ratios not so far from what we are finding here. The filaments identified by the Rolling Hough transformation in \citet{2014ApJ...789...82C} on HI maps tend to be longer and thinner than this.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{figs/BErat_vs_eps.pdf}
\end{center}
\caption{Ratios of $BB$ to $EE$ power as a function of the physical filament axis ratio. Long and thin filaments ($\epsilon$ small) have less $B$-mode power than $E$-mode power. Short and squat filaments (aspect ratio $\epsilon$ close to unity) have $B$ power close to the $E$ power. An aspect ratio of about $\epsilon = 0.26$ can reproduce the Planck-observed ratio of about one half, but this can be traded off against a slight dependence with the dispersion in the misalignment angle $\theta_{LH}$ between the filament direction and the magnetic field direction in three dimensions.}
\label{fig:EEBBrat}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{figs/rTE_vs_eps.pdf}
\end{center}
\caption{Correlation coefficient of $TE$ power as a function of the axis ratio. Thinner filaments (small $\epsilon$) have stronger $TE$ correlations for various magnetic field--filament misalignments. Since the $B$-to-$E$ power ratio requires $\epsilon \approx 0.26$ (Figure \ref{fig:EEBBrat}), the $TE$ correlation is diagnostic of the misalignment necessary to produce the Planck-observed $r^{TE} \approx 0.35$, which needs ${\rm RMS}(\theta_{LH}) \approx 50^\circ$.}\label{fig:rTE}
\end{figure}
\subsection{TE cross-correlation}
In the context of the filament model, we find that the level of the $TE$ cross correlation implies that filaments cannot be precisely aligned to their local magnetic field direction. The correlation coefficient is defined as
\begin{equation}
r_\ell^{TE} = C_\ell^{TE}/\sqrt{C_\ell^{EE} C_\ell^{TT}},
\end{equation}
and perfect alignment of the filaments and the fields causes far too much $TE$ correlation compared to the Planck observations.
The Planck dust data show $r_\ell^{TE} \approx 0.35$ with little scale dependence \citep{2018arXiv180104945P}. Fig.~\ref{fig:rTE} shows that to match this, the field misalignment angle $\theta_{LH}$ must have an RMS dispersion of nearly $50^\circ$, while maintaining the axis ratio $\epsilon \approx 0.26$ needed to reproduce the $BB/EE$ power ratio. If the misalignment dispersion is independent of filament size, as in our modeling, it causes no scale dependence: $r^{TE}_\ell$ is constant.
{Projection effects cause the distribution of the projected angle $\psi_{LH}$ to have a positive kurtosis (see appendix, Fig.~\ref{fig:psiLH_marg}), and so we can describe its dispersion in a few ways. For the ${\rm RMS}(\theta_{LH}) = 50^\circ$ case, 68 percent of the probability is bounded by $|\psi_{LH}| < 45^\circ$. Alternatively, $[{\rm Var}(\psi_{LH})]^{1/2} = 48^\circ$.} For comparison, \citet{2016A&A...586A.141P} fit a Gaussian a with $19^\circ$ dispersion (1$\sigma$) to the projected field--projected filament histogram of relative orientations for the filaments they found. Again this comparison is not direct because of selection effects. Their Hessian-based selection of filaments would disfavor filaments with small projected aspect ratios (close to the line of sight), and such filaments can have the largest differences in the projected orientation.
The overall level of $C_\ell^{TE}$ (and the polarization spectra) depends on the polarization fraction. The relative power ratio has a dependence like
\begin{equation}
C_\ell^{TE}/C_\ell^{EE} \propto \langle f_{\rm pol} \rangle/\langle f_{\rm pol}^2 \rangle,
\end{equation}
while for the cross correlation it is
\begin{equation}
r_\ell^{TE} \propto \langle f_{\rm pol} \rangle/\langle f_{\rm pol}^2 \rangle^{1/2}.
\end{equation}
Thus a purely multiplicative rescaling of the polarization fraction affects the ratios of the power in $TT/TE/EE$ but not the correlation coefficient $r^{TE}$.
To reproduce the Planck-measure ratio $C_\ell^{TE}/C_\ell^{EE} \sim 2.7$ (in our case that already fits $C_\ell^{BB}/C_\ell^{EE}$ and $r_\ell^{TE}$) requires $f_{\rm pol} = 0.15\, \sin^2 \theta_H$.
We have only modeled the polarization fraction amplitude and the geometric dependence on the magnetic field orientation, but in addition, the polarization fraction depends on grain geometry and small-scale turbulence \citep{2000ApJ...544..830F}, and filaments need not in reality have all the same intrinsic polarization fraction.
Our other tests have shown that the $TE$ correlations differ in their sensitive to intrinsic dispersion in the polarization fraction. For example, the correlation $r^{TE}$ is not very sensitive to the maximum polarization fraction, but the power ratio is very sensitive to it: decreasing the maximum polarization fraction decreases $C_\ell^{TE}$ but decreases the denominator $C_\ell^{EE}$ more, and so raises the ratio.
\subsection{Parity violation: TB and EB}
One surprising finding in the \citet{2018arXiv180104945P} dust spectra is a non-zero $TB$ correlation, with $r_\ell^{TB} \approx 0.05$. Like $r^{TE}$, the observed $r^{TB}$ correlation has little scale dependence (up to multipoles of several hundred).
Because non-zero $TB$ and $EB$ are parity-violating correlations, our model cannot reproduce them for randomly-oriented filaments. To get a positive $TB$ correlation we would need to favor polarization angles in the range $\psi_{\rm pol} \in [90^\circ,180^\circ]$ relative to the filament direction (compare Fig.~\ref{fig:pol_angle}). Equivalently, this corresponds to projected field angles in the range $\psi_{LH} \in [0^\circ,90^\circ]$. Such an effect may be due to some large scale feature in the Galaxy's magnetic field or differential gas flow \citep[e.g.][]{2018arXiv180104945P,2019A&A...621A..97B}.
We can determine how far away from random this correlation is by artificially weighting the distribution of the projected misalignment angles, favoring the $\psi_{LH} > 0$ portion of the distribution of $p(\psi_{LH},\theta_H | \theta_L)$ over the $\psi_{LH} < 0$ portion, while keeping the same functional form. We find that we can approximately reproduce the Planck measured $TB$ correlation by giving the preferred $\psi_{LH}$ directions about 55 percent of the total weight, rather than the 50 percent than comes naturally from randomly oriented filaments. Similar to the $r^{TE}$ correlation (also set by field--filament misalignment), this effect is not scale dependent, and so $r^{TB}$ is constant to high $\ell$ in this model.
Our modeling comes from the 1-halo term only, and shows that the $TB$ correlation can be explained if filaments orientations in projection are slightly twisted counterclockwise from the projected local magnetic field. Our model does not address the structure of that underlying field, but we may speculate that some differential, shearing hydrodynamic forcing could preferentially twist the filaments, according to our point of view, from the global mean field direction of the Milky way.
\citet{2019A&A...621A..97B}, argue that the observed $TE$ and $TB$ correlations may be features of the large scale structure of the Galactic magnetic field. They show that a helical component can create such correlations, but in their modeling, the correlation show a strong scale dependence, with $r^{TE,TB}_\ell$ falling substantially already by $\ell = 22$. The Planck spectra have much flatter $r^{TE,TB}_\ell$ correlations, consistent with the filament modeling here.
Planck did not detect an $EB$ correlation, but since $E$ and $B$ both have a factor of the polarization fraction, we would naturally expect this correlation to be smaller. It may be there, hidden in the noise. In the presence of a positive $TB$ correlation, in the context of the filament modeling, we would expect a positive $EB$ correlation too (including at high-$\ell$), and it should be a target for future experiments. Both $TB$ and $EB$ dust correlations can potentially interfere with sky-calibration of the polarization angles of CMB-instruments \citep{2016MNRAS.457.1796A} or with CMB lensing reconstruction \citep[e.g.][]{2012JCAP...12..017F,2018JCAP...04..018C}.
\section{Conclusions}\label{sec:discussion}
We do not know how much of the dusty microwave polarization foreground is due to filamentary structure, but if it is a substantial portion, we can discern details of the filament population from the foreground power spectra. We showed that the slopes of the power spectra relate to the distribution of lengths. We showed that the $BB$/$EE$ power ratio relates to the filament axis ratio. We showed that the $TE$ cross-correlation relates to the axis ratio and the RMS misalignment of filaments to the magnetic field. We showed that $TB$ correlations could be caused by a slight preference for one handedness in the misalignment between the magnetic field and the filament orientation.
Despite its relative success in reproducing the features of the dust polarization power spectrum, this formalism lacks some essential features for modeling the real sky. Foremost, this formalism includes only the one-halo term in the power spectra. This is obviously an approximation, for the Planck data have shown that the Galaxy's projected magnetic field has coherent, large-scale features, and the HI-identified filaments are clearly correlated with it and with starlight polarization measurements \citep{2015PhRvL.115x1302C}.
On the other hand, the transition from one-halo-dominated to two-halo-dominated scales often leaves a mark on the power spectrum. Since in the dust polarization spectra there are not clear features, like a break in the slope, we may speculate that the two-halo component may not be necessary to describe the main properties of the power spectra.
Inclusion of a proper two-halo formalism is complicated by the correlated direction dependence of the filaments. We may be able to import some of the techniques developed to describe galaxy intrinsic alignments \citep[e.g][]{2010MNRAS.402.2127S}, since the mathematical description of the problem is similar.
We have not tried to systematically probe the parameter degeneracies or place proper uncertainties on any of the parameters of this filament model. We can do this straightforwardly by interfacing the model with a Monte Carlo Markov Chain and developing a likelihood based on the Planck dust spectra. We plan to pursue this in further work.
By looking at observations and simulations of the filamentary ISM, we could attempt to verify some of the population statistics for filaments. For example, we could compare the length distribution of actual or simulated filaments to that implied by the slope of the power spectra.
Because this filament model is non-Gaussian, we may be able to use it to design novel diagnostics to probe for residual foregrounds in surveys that aim for the primordial $B$-modes \citep[in the spirit of][]{2014PhRvL.113s1303K,2016JCAP...09..034R,2018MNRAS.479.5577P,2019arXiv190104515C}. Similarly, we could use this model to compute the four-point contributions to polarized CMB lensing estimators. This could help place constraints on potential foreground contamination. Such statistics may be sensitive to the internal density structure of the filaments is a way that the power spectrum is not.
Due to its flexibility, its ability to model the Planck dust polarization data, its ease of computation, and its straightforward interpretation, this filament model may become a useful tool in the study of CMB polarization foregrounds.
\section*{Acknowledgments}
KMH, AR, and DCC acknowledge support from the NASA ATP program under grant NNX17AF87G. KMH acknowledges support from the NSF AAG program under grant 1815887. AR acknowledges support from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No 725456, CMBSPEC). We thank J.~Colin Hill, Susan Clark, Marc Kamionkowski, and Carlos Hervias-Caimapo for useful conversations. We thank Francois Boulanger for providing access to masks used in the \citet{2018arXiv180104945P} analysis.
|
\section{Introduction}
The role of multiple parton interactions in proton-proton collisions is believed to become increasingly important when one explores the energy frontier in particle physics. As such, the relevance in LHC observables of two simultaneous hard scatterings, usually referred to as Double Parton Scatterings (DPS), has attracted much attention in the last decade with significant theory advances related to perturbative QCD~\cite{Blok:2011bu,Diehl:2011yj,Diehl:2011tt,Gaunt:2011xd,Manohar:2012pe,Gaunt:2012dd,Blok:2013bpa,Diehl:2014vaa,Diehl:2015bca,Rinaldi:2016jvu,Buffing:2017mqm,Diehl:2017kgu,Vladimirov:2017ksc,Diehl:2018wfy,Gaunt:2018eix}. Since DPS are higher-twist effects in total cross sections compared to the conventional single parton scatterings (SPS), quantitative studies of DPS remain challenging though not impossible both on the theoretical and experimental sides. These are particularly interesting since they provide us with means to study parton correlations inside the proton (see {\it e.g.}\ \cite{Rinaldi:2016jvu,Rinaldi:2018bsf,Rinaldi:2018slz}).
Among the possible hard probes of DPS at high-energy hadron colliders, the associated production of quarkonia (see~\cite{Lansberg:2019adr} for an exhaustive review) provides unique opportunities to measure DPS in gluon-induced reactions thus to study gluon-gluon correlations in the proton. Numerous measurements of quarkonium associated processes have been performed at the Tevatron and the LHC. They can mainly be categorised as di-quarkonium production ($J/\psi+J/\psi$~\cite{Aaij:2011yc,Abazov:2014qba,Khachatryan:2014iia,Aaboud:2016fzt,Aaij:2016bqq}, $J/\psi+\Upsilon$~\cite{Abazov:2015fbl}, $\Upsilon+\Upsilon$~\cite{Khachatryan:2016ydm}), associated production with a vector boson ($J/\psi+W^{\pm}$~\cite{Aad:2014rua}, $J/\psi+Z$~\cite{Aad:2014kba}) or with another heavy quark ($J/\psi+$open charm~\cite{Aaij:2012dz}, $\Upsilon+$open charm~\cite{Aaij:2015wpa}). All these measurements cover different kinematical regions with different momentum transfers in the hard scattering. Their theoretical analysis is highly non-trivial, which has triggered many theoretical studies in the recent years~\cite{Li:2009ug,Qiao:2009kg,Ko:2010xy,Kom:2011bd,Berezhnoy:2011xy,Lansberg:2013qka,Li:2013csa,Lansberg:2014swa,Sun:2014gca,Lansberg:2015lva,He:2015qya,Baranov:2015cle,Shao:2016wor,Lansberg:2016rcx,Lansberg:2016muq,Likhoded:2016zmk,Borschensky:2016nkv,Lansberg:2017chq,Lansberg:2017dzg,Cisek:2017ikn,Gridin:2019nhc}. Very recently, the first calculation of triple-$J/\psi$ production showed that it can help us probe both DPS and triple parton scatterings (TPS)~\cite{Shao:2019qob}.
In this context, we focus in this paper on the di-$J/\psi$ case with the aim to improve the existing perturbative QCD calculations for the SPS. To do so, we consider higher-order corrections in both the strong coupling constant, $\alpha_S$, and the heavy-quark velocity, $v$. First, we study the impact of a gauge-invariant and infrared-safe subset of loop-induced (LI) contributions. Our analysis follows the lines of a similar study for $J/\psi+\Upsilon$ production~\cite{Shao:2016wor}. Such contributions appear at next-to-next-to-leading order (NNLO) in $\alpha_S$ but could be enhanced at large rapidity differences and high invariant masses of the $J/\psi$ pair because of the presence of topologies with double $t$-channel gluon exchanges between both charm-anticharm quark lines. Second, we perform a comprehensive survey of the impact of the colour-octet (CO) contributions in three kinematical domains covered by the existing LHC measurements~\cite{Khachatryan:2014iia,Aaboud:2016fzt,Aaij:2016bqq} considering the various existing fits of the non-perturbative CO long-distance matrix elements (LDMEs).
In order to disentangle DPS from SPS in observables where two particles are observed, one usually relies on the analysis of specific kinematical dependences which are believed to be drastically different in both samples. Common choices of variables are the azimuthal and the rapidity separations between both observed particles, $\Delta \phi$ and $\Delta y$.
The DPS contribution, coming from two {\it a priori} independent parton scatterings, is expected to be flatter than the SPS one in both distributions.
For double-$J/\psi$ studies, the analysis of the $\Delta y(J/\psi,J/\psi)$ distributions should be preferred compared to that of $\Delta \phi(J/\psi,J/\psi)$ since the $\Delta y(J/\psi,J/\psi)$ distribution of the SPS yield is much less affected by possible non-perturbative intrinsic $k_T$ of the colliding gluons~\cite{Kom:2011bd} than the $\Delta \phi(J/\psi,J/\psi)$ one, which can become as flat as the DPS ones in some cases. In general, one expects the DPS fraction to be the largest at large $|\Delta y(J/\psi,J/\psi)|$. A precise determination of the DPS yield therefore requires a good knowledge of the SPS in this region. Both the LI and CO topologies with $t$-channel-gluon exchanges could result into a flat $\frac{d\sigma}{d\Delta y}$ like in the $J/\psi+\Upsilon$ case~\cite{Shao:2016wor}.
Assuming $\alpha_S\sim v^2$, the colour-singlet (CS) LI contribution should be of the same magnitude as the leading order (LO) CO contribution (yet both smaller that the bulk of the CS yield in the absence of the possible kinematical enhancement which we are after here). According to the NRQCD velocity scaling rules~\cite{Bodwin:1994jh}, the former one is indeed $\mathcal{O}(\alpha_S^6v^3)$ while the latter one is $\mathcal{O}(\alpha_S^4v^7)$. This justifies why we consider both of them in this study.
This article is organised as follows. In section \ref{sec:lhcmeasure}, we first quickly review the existing LHC measurements used in our comparisons~\footnote{We do not consider the D0 measurement~\cite{Abazov:2014qba} at the Tevatron since no corrected distribution was released which could be used for a direct data-theory comparison.}. Then, we discuss our theory framework in section~\ref{sec:theory}. Section \ref{sec:lipart} gathers our discussion of the impact of the inclusion LI CS contribution and section \ref{sec:copart} comprises a comprehensive analysis of complete LO CO contribution. The appendix \ref{app:moreplots} collects additional plots relevant for further theory-data comparisons.
\section{LHC measurements and kinematical variables\label{sec:lhcmeasure}}
\subsection{kinematical variables\label{app:variables}}
We start by introducing the kinematical variables relevant for di-quarkonium production. On the experimental side, the second LHCb analysis~\cite{Aaij:2016bqq} bears on the largest set of the kinematical variables whose distribution is used for comparisons between the experimental measurements and the theoretical calculations. Since some of these variables may not be very common, we summarise the description of their names or labels in Table~\ref{tab:variables}. In particular, the transverse momentum asymmetry is defined as
\begin{eqnarray}
A_{T}(J/\psi,J/\psi)\equiv \left|\frac{P_T(J/\psi_1)-P_T(J/\psi_2)}{P_T(J/\psi_1)+P_T(J/\psi_2)}\right|,
\end{eqnarray}
where $J/\psi_1$ and $J/\psi_2$ are respectively denoted as the first and second hardest $J/\psi$ with ordered in the transverse momentum.
\begin{table}[ht]
\begin{center}
\begin{tabular}{c||c}
\hline\hline
$P_T(J/\psi+J/\psi)$ & $y(J/\psi+J/\psi)$\\\hline
Transverse momentum of the pair & Rapidity of the pair \\
\hline\hline
$P_T(J/\psi)$ & $y(J/\psi)$ \\\hline
Transverse momentum of a randomly chosen $J/\psi$ & Rapidity of a randomly chosen $J/\psi$\\\hline\hline
$\Delta \phi(J/\psi,J/\psi)$ & $\Delta y(J/\psi,J/\psi)$\\\hline
Azimuthal angle difference in the transverse plane & Rapidity separation \\\hline\hline
$M(J/\psi+J/\psi)$ & $A_{T}(J/\psi,J/\psi)$ \\\hline
Invariant mass of the pair & Transverse momentum asymmetry \\\hline
\hline
\end{tabular}
\caption{Summary of the kinematical variables.}\label{tab:variables}
\end{center}
\end{table}
\subsection{Available data sets}
Four LHC studies of double prompt $J/\psi$ production have so far been performed~\cite{Aaij:2011yc,Khachatryan:2014iia,Aaboud:2016fzt,Aaij:2016bqq}. LHCb performed two measurements in the same kinematical region, one at $\sqrt{s}=7$ TeV and another at $\sqrt{s}=13$ TeV; we will focus on the latter which is more precise~\cite{Aaij:2016bqq}. The various kinematical cuts used in the ATLAS, CMS and LHCb analyses are summarised in Table.~\ref{tab:cuts} along with the corresponding centre-of-mass energy $\sqrt{s}$ . It is useful to note that due to the different trigger and acceptance constraints on the ATLAS, CMS and LHCb data taking, the 3 samples cover complementary domains in $P_T$ and $y$. In particular, ATLAS~\cite{Aaboud:2016fzt} imposes the largest $P_T(J/\psi)$ cut (as large as 8.5 GeV), while LHCb~\cite{Aaij:2016bqq} does not impose any lower $P_T$ cut on the observed $J/\psi$. As such, LHCb events are mostly located at low $P_T(J/\psi)$. CMS~\cite{Khachatryan:2014iia} imposes varying cuts from $P_T(J/\psi)>4.5$~GeV to $P_T(J/\psi)>6.5$~GeV depending on the rapidity. Moreover, LHCb can only detect forward particles whereas ATLAS/CMS have a generally larger rapidity coverage but in the central-rapidity region. In section \ref{sec:copart}, we will discuss how these kinematical coverages can be relevant to determine the proper CO LDMEs.
\begin{table}[ht]
\begin{center}\small
\begin{tabular}{c|c|c}
Experiment & $\sqrt{s}$ [TeV] & Kinematical cuts\\\hline\hline
\multirow{3}{*}{CMS~\cite{Khachatryan:2014iia}} & \multirow{3}{*}{$7$} & $P_T(J/\psi)>$6.5 GeV when $|y(J/\psi)|<1.2$;\\
& & $P_T(J/\psi)>6.5-\frac{200}{23}(|y(J/\psi)|-1.2)$ GeV when $1.2<|y(J/\psi)|<1.43$;\\
& & $P_T(J/\psi)>4.5$ GeV when $1.43<|y(J/\psi)|<2.2$ \\\hline
\multirow{3}{*}{ATLAS~\cite{Aaboud:2016fzt}} & \multirow{3}{*}{$8$} & $P_T(\mu)>$2.5 GeV, $|\eta(\mu)|<2.3$;\\
& & One $J/\psi$ has two muons with $P_T(\mu)>$ 4 GeV; \\
& & $P_T(J/\psi)>$ 8.5 GeV, $|y(J/\psi)|<2.1$\\\hline
LHCb~\cite{Aaij:2016bqq} & $13$ & $P_T(J/\psi)<14$ GeV, $2.0<y(J/\psi)<4.5$ \\
\hline
\end{tabular}
\caption{Summary of kinematical cuts of the double-$J/\psi$ measurements by the LHC experiments which we will consider here.}\label{tab:cuts}
\end{center}
\end{table}
\section{Theory framework}
\label{sec:theory}
In this section, we briefly address some specificities of our theoretical computations, which however remain very standard.
\subsection{Intrinsic initial-$k_T$ smearing\label{sec:ktsmear}}
An important effect for an accurate description of double-$J/\psi$ hadroproduction is known to be the smearing of the kinematics arising from the intrinsic $k_T$ of the gluons~\cite{Sridhar:1998rt}. It is in principle a non-perturbative effect which cannot properly be accounted for by the collinear factorisation. In fact, double-$J/\psi$ production can provide new insights in the transverse dynamics of the gluons as it was shown~\cite{Lansberg:2017dzg} using the transverse-momentum dependent (TMD) factorisation. Clearly, a collinear computation is not meant to encapsulate such effects. As a makeshift, we simply rely on an empirical procedure to deal with them which we believe to be sufficient for our phenomenological purpose. In particular, the whole $k_T$ smearing is assumed to be factorised out by
\begin{eqnarray}
\frac{d\sigma}{d\Phi_{\langle k_T \rangle}}=\int_{0}^{+\infty}{dk_T^2 \frac{\pi}{8\langle k_T\rangle^2}e^{-\frac{\pi}{8}\frac{k_T^2}{\langle k_T\rangle^2}}\frac{d\sigma}{d\Phi}},
\end{eqnarray}
where the phase-space mapping $\Phi\rightarrow \Phi_{\langle k_T \rangle}$ is determined by boosting the whole event according to the generated transverse-momentum imbalance $|\overrightarrow{k_T}|=k_T$ with a uniform distribution of the azimuthal angle in the transverse plane. Other forms are of course possible.
In the present study, we assume $\langle k_T \rangle$ to be the same for all three experimental coverages and fix its value to be $3.0$ GeV. The distributions with other $\langle k_T \rangle$ values are also not shown but can easily be obtained with the help of {\sc\small HELAC-Onia}~\cite{Shao:2012iz,Shao:2015vga}. In fact, the NLO$^\star$ distributions with $\langle k_T \rangle=0.5$ GeV and $2.0$ GeV can be found in a theory-data comparison made by LHCb~\cite{Aaij:2016bqq}. The $k_T$-smearing effect is only visible for the $P_T(J/\psi+J/\psi)$, $\Delta \phi(J/\psi,J/\psi)$ and $A_T(J/\psi,J/\psi)$ distributions.
\subsection{Parameters entering our calculations\label{sec:setup}}
We now quickly describe our set-up for the present calculations before discussing the numerical results. We have fixed the charm quark mass to be 1.5 GeV and only the light $u$, $d$, $s$ (anti)quarks and the gluons are allowed in the initial states. In order to be compatible with our previous NLO$^\star$ calculations, we have used the NLO parton-distribution functions (PDFs) CTEQ6M~\cite{Pumplin:2002vw} for the calculations in the ATLAS and CMS acceptances and NNPDF3.0~\cite{Ball:2014uwa} for those in the LHCb acceptance. We have explicitly checked that the PDF dependence is less than $20\%$ and is thus a minor source of uncertainty compared to the (dominant) scale uncertainty which we discuss below. The missing higher-order terms in $\alpha_S$ are estimated in the usual way by independently varying the factorisation and renormalisation scales as $(\mu_F,\mu_R)=(\zeta_1\mu_0,\zeta_2\mu_0)$, with $\zeta_{1,2}=\frac{1}{2},1,2$, where the central scale $\mu_0$ is chosen to be $\mu_0=\sqrt{\left(P_T(J/\psi)\right)^2+ (4 m_c)^2}$, like in Refs.~\cite{Lansberg:2013qka,Lansberg:2014swa}. The CS LDME is estimated via
$\langle\mathcal{O}^{H_{Q\bar{Q}}}(^3S_1^{[1]})\rangle=2N_c\frac{3}{4\pi}\left|R^{H_{Q\bar{Q}}}(0)\right|^2$, where the wave function at the origin $R^{H_{Q\bar{Q}}}(0)$ can be determined by solving the Schr{\"o}dinger equation with a given QCD potential. We will use the numerical values $\left|R^{J/\psi}(0)\right|^2=0.8$ GeV$^3$ and $\left|R^{\psi(2S)}(0)\right|^2=0.5$ GeV$^3$ derived in Ref.~\cite{Eichten:1995ch} using the QCD-motivated Buchm{\"u}ller-and-Tye potential~\cite{Buchmuller:1980su}. For the CS SPS yield, the feed-down contribution from the $\psi(2S)$ decays is as large as the direct double $J/\psi$ production. In practice, we take it to be equal to 2. It is thus mandatory to take it into account.
\section{Colour-singlet contributions: partial loop-induced corrections\label{sec:lipart}}
In principle, considering the square of a one-loop amplitude by itself should give divergent results from both the infrared and ultraviolet regions. Such one-loop amplitudes squared are part of the NNLO contributions, at $\mathcal{O}(\alpha_S^6)$ in the case of double $J/\psi$ hadroproduction. The cancellation of the aforementioned infrared divergences would be achieved as usual by considering two-loop, one-loop single-real-emission and double-real-emission amplitudes. Such a computation is obviously beyond the scope of this study -- it is not even available for single $J/\psi$.
However, a subset of such one-loop diagrams, restricted to the sole topologies with two separate charm-quark lines forming each a $J/\psi$, happens to be free of any divergence and is, in addition, gauge invariant. Correspondingly, the possible double-real emissions which could develop infrared divergences do not contribute when one of the external gluon becomes soft. This is akin to the absence of any infrared divergences at $P_T \to 0$ for $gg \to J/\psi g$. Such a subset is in fact that of the LI contribution to $pp\rightarrow J/\psi+\Upsilon$~considered in Ref.~\cite{Shao:2016wor}
The square of the amplitude from these one-loop diagrams is what we refer here to as the (partial) LI corrections. Their computation is included in the {\sc\small HELAC-Onia} code~\cite{Shao:2012iz,Shao:2015vga} and is thus available to everybody. In fact, another gauge-invariant $\mathcal{O}(\alpha_S^6)$ part, namely from $pp\rightarrow J/\psi+J/\psi+c\bar{c}$, is known~\cite{Lansberg:2014swa}. It turns out to be small and can safely be ignored for our purposes. However, we wish to point out that the process $pp\rightarrow J/\psi+J/\psi+c\bar{c}$ has its own interest as it can be a potential probe of the TPS at the LHC.
Let us add that we do not expect any specific kinematical enhancement of other NNLO topologies, in particular that of the double-real-gluon emission in view of the results of $pp\rightarrow J/\psi+\Upsilon$~\cite{Shao:2016wor}. This is partly explained by the vanishing of these contributions when one gluon becomes soft, precisely where one can minimise the off-shellness of the other particles involved in the scattering and thus where these contributions could have been the largest.
\begin{figure}[hbt!]
\centering
\foreach \page in {3,1}{
\includegraphics[page=\numexpr \page\relax, width=.48\textwidth,draft=false]{./Figures/CMS/sigma_CMSarXiv14060484_CSM_LI.pdf}
}
\caption{Rapidity gap $|\Delta y(J/\psi, J/\psi)|$ (left) and invariant mass $M(J/\psi+J/\psi)$ (right) distributions for di-$J/\psi$ production in CSM via SPS within CMS $\sqrt{s}=7$ TeV acceptance~\cite{Khachatryan:2014iia}.}\label{fig:cmsli}
\end{figure}
\begin{figure}[hbt!]
\centering
\foreach \page in {4,1}{
\includegraphics[page=\numexpr \page\relax, width=.48\textwidth,draft=false]{./Figures/ATLAS/sigma_ATLASCONF2016047_CSM_LI.pdf}
}
\caption{Rapidity gap $|\Delta y(J/\psi, J/\psi)|$ (left) and invariant mass $M(J/\psi+J/\psi)$ (right) distributions for di-$J/\psi$ production in CSM via SPS within ATLAS $\sqrt{s}=8$ TeV acceptance~\cite{Aaboud:2016fzt}.}\label{fig:atlasli}
\end{figure}
\begin{figure}[hbt!]
\centering
\foreach \page in {8,2}{
\includegraphics[page=\numexpr \page\relax, width=.48\textwidth,draft=false]{./Figures/LHCb/sigma_LHCbarXiv161207451_CSM_LI.pdf}
}
\caption{Rapidity gap $|\Delta y(J/\psi, J/\psi)|$ (left) and invariant mass $M(J/\psi+J/\psi)$ (right) distributions for di-$J/\psi$ production in CSM via SPS within LHCb $\sqrt{s}=13$ TeV acceptance~\cite{Aaij:2016bqq}.}\label{fig:lhcbli}
\end{figure}
The cross-sections differential in the absolute rapidity difference between the $J/\psi$ pair $|\Delta y(J/\psi,J/\psi)|$ are shown in the left panels of Figures~\ref{fig:cmsli},~\ref{fig:atlasli} and \ref{fig:lhcbli} and are compared to the CMS, ATLAS and LHCb data. The NLO$^\star$ CS calculations are displayed by the red hatched bands in the figures. The partial LI contributions are represented by the green bands. As expected, the (partial) LI is significant at large $|\Delta y(J/\psi,J/\psi)|$ region but negligible at small and intermediate $|\Delta y|$. An order of magnitude enhancement to the CS cross section is expected when $|\Delta y(J/\psi,J/\psi)|\geq 3.0$. Nonetheless, despite the very large theoretical uncertainties from the scale variations, a discrepancy between the CS SPS and the experimental data is clearly visible at large $|\Delta y(J/\psi,J/\psi)|$, that is exactly where the DPS is expected to be important.
The invariant mass of the meson pair is also closely related to the rapidity gap $|\Delta y(J/\psi,J/\psi)|$ (see the discussion in Ref.~\cite{Lansberg:2014swa}). Large $M(J/\psi+J/\psi)$ bins are usually populated by large $|\Delta y(J/\psi,J/\psi)|$ events. Similar enhancements from the LI contributions can be seen in the tail of the invariant-mass distributions of Figures~\ref{fig:cmsli} and \ref{fig:atlasli}. The measurements by CMS and ATLAS are consistent with the SPS CS alone at low invariant masses and depart from the SPS CS bands (NLO$^\star$ and NLO$^\star$+LI) at large $M(J/\psi+J/\psi)$ values. In contrast, for the LHCb acceptance, the LI part is negligible compared to the NLO$^\star$ contributions due to the limited $M(J/\psi+J/\psi)$ range, below 14 GeV.
We have collected additional data-theory-comparison plots between the SPS CS yields and the LHC measurements for other observables in the appendix \ref{app:LImoreplots}. The data are compatible with the CS theoretical predictions but the LI contributions are found to be negligible for all the other distributions.
\section{Comprehensive assessment of the colour-octet contributions\label{sec:copart}}
The whole LO CO contributions to di-$\psi$ hadroproduction at the LHC up to $\mathcal{O}(v^7)$ in NRQCD have recently been computed~by He and Kniehl~\cite{He:2015qya}. Their study however bears on a single CO LDME set from an out-of-date LO single $J/\psi$ hadroproduction fit~\cite{Braaten:1999qk} which was made with the early Tevatron data. Yet, their calculation seems to indicate that the CO contributions might be relevant at large $|\Delta y(J/\psi,J/\psi)|$ and large $M(J/\psi+J/\psi)$ due to similar $t$-channel gluon exchange diagrams than for the CS LI contributions. The aforementioned remaining discrepancy between this full SPS LO NRQCD calculation and the CMS data at large $|\Delta y(J/\psi,J/\psi)|$ was then attributed to unknown missing higher-order QCD corrections to the CO contributions.
We however note that we do not anticipate any such so-called ``giant" $K$ factors in this region. Currently, no complete NLO CO calculation exists. Since it is important to deal with a complete set of CO channels in order to guarantee the large cancellation between $S$-wave and $P$-wave contributions involved the NLO LDME fits of hadroproduction data, we consider that to rely on a LO --but complete-- perturbative calculation and then to estimate the size of the missing higher-order corrections via the scale uncertainty is probably the most reasonable procedure to adopt.
An alternative approach to investigate the presence of possible ``giant" $K$ factors from new fragmentation topologies --if some are indeed relevant-- without performing a full computation is that recently proposed by one of us in Ref.~\cite{Shao:2018adj}. It has been proved useful for the single $J/\psi$ hadroproduction case. The method is in principle general and applicable for the double $J/\psi$ hadroproduction as well, although a new infrared divergence in double $P$-wave channels emerges~\cite{He:2018hwb}. We leave it for future studies since it may not apply to the whole phase space which we wish to consider here. Finally, we note that a similar enhancement from $t$-channel gluon exchange was expected for di-$\chi_c$ production but its feed down was also found to be insignificant in the di-$\psi$ yield~\cite{Cisek:2017ikn}.
\subsection{Status and issues with the colour-octet transitions}
Although the possibility for CO transitions is a robust prediction from NRQCD, their actual impact in the phenomenology has been the subject of debates for decades. The most glaring observations for the necessity of their presence are twofold. First, CO provide a natural solution for the infra-red divergence issue in $P$-wave production. Second, the LO $v^2$ NRQCD calculation involving only CS transitions still underestimates --even after including NLO QCD corrections-- the yields of single $J/\psi$ and $\psi(2S)$ hadroproduction at large $P_T$ at the Tevatron and the LHC.
However, NRQCD computations even including CO contributions are unable to coherently describe --{\it i.e.} with the same CO LDMEs-- the world data for $pp$, $ep$, $\gamma p$, $\gamma \gamma$ and $e^+e^-$ collisions. For a recent review, we guide the reader to Ref.~\cite{Lansberg:2019adr}.
The CO LDMEs are predicted to be universal non-perturbative objects by NRQCD, which should yield predictions compatible with all the data. The current status of their extractions is very confusing as their numerical values and their uncertainties are very disparate. The results of the fits of different groups disagree with each others. As long as the situation is not clarified, we believe that it is necessary to comprehensively consider these analyses instead of drawing conclusions based on a single CO LDME set as it is often done in the analysis of associated production of quarkonium (see Ref.~\cite{Lansberg:2019adr} for some examples).
As such, we will use different LDME sets of which we briefly review the status and the possible limitations. As we said above, the available CO LDMEs for prompt $J/\psi$ production are extracted from fits. According to the QCD accuracy of the short-distance coefficients (SDCs), we will categorise them in the 4 groups shown in Table~\ref{tab:coldme}. Namely,
\begin{enumerate}
\item three fits are based on LO SDCs~\cite{Braaten:1999qk,Kramer:2001hh,Sharma:2012dy},
\item four fits based on NLO SDCs~\cite{Butenschoen:2011yh,Gong:2012ug,Shao:2014yta,Han:2014jya},
\item one fit based on a low-$P_T$ leading-logarithm (LL) resummed SDC ~\cite{Sun:2012vc},
\item one fit using a SDC using leading-power (LP) fragmentation matched to NLO SDC~\cite{Bodwin:2014gia}.
\end{enumerate}
All of them have shortcomings and/or limitations. We enumerate them below:
\begin{enumerate}
\item First of all, we wish to emphasise that the LO fits are out-of-date and should be viewed as a pure tuning of the normalisation of the single $J/\psi$ data. Since all of the LO fits are mainly performed with the help of intermediate and large $P_T$ hadroproduction data, where the ``giant" $K$ factors from NLO QCD corrections emerge, it is very hard to imagine that these values will give correct predictions for independent observables, like the double-$J/\psi$ hadroproduction in our case, for which $K$ factors would be different. We will therefore use them here for a pure illustrative purpose.
\item The LL fit in Ref.~\cite{Sun:2012vc} concentrates on the $P_T(J/\psi)< m_c$ region. The authors performed a small-$P_T$ resummation but without considering the contribution from the CS channel which is however known to saturate the data in this region~\cite{Brodsky:2009cf,Feng:2015cba}. The values of these LDMEs have never been used for the single $J/\psi$ production at intermediate and large $P_T$ regions. They are included in our discussion like the LO fits in order to be exhaustive.
\item The NLO fit in Ref.~\cite{Butenschoen:2011yh} used the world data before 2011 without subtracting the feed-down contributions. The fit seems to yield a good agreement with the $P_T(J/\psi)<30$ GeV $J/\psi$ yields data at different colliders but for $\gamma \gamma$ and $e^+e^-$ collisions. However, it overshoots the $P_T>30$ GeV yields and fails to reproduce the polarisations of $J/\psi$, the energy-fraction distribution of the $J/\psi$ in jets~\cite{Bain:2017wvk} and the yields of $\eta_c$ (by using heavy-quark spin symmetry). In addition, the SPS $P_T$-differential cross section of $J/\psi+\gamma$~\cite{Li:2014ava} turns out to be negative at NLO with these values of CO LDMEs.
\item The NLO fit by Gong et al.~\cite{Gong:2012ug} focus on the $P_T(J/\psi)>7$ GeV data. The feed-down contributions are subtracted. This LDME set is however not compatible with the yields ({\it e.g.}\ $pp$, $\gamma p$ and $e^+e^-$) when $P_T(J/\psi)<7$ GeV, the polarisation of forward $J/\psi$~\cite{Aaij:2013nlm} and the $\eta_c$ production. In addition, it yields to --unphysical-- negative cross sections in $pp\rightarrow J/\psi+\gamma$. In principle, this set is only applicable to $J/\psi$ production with $P_T(J/\psi)>7$ GeV, {\it i.e.}\ only to the ATLAS fiducial region for our forthcoming discussion of double $J/\psi$ production.
\item The two sets denoted sets 7 and 8 in Table~\ref{tab:coldme} are two extreme cases of the PKU fit~\cite{Shao:2014yta,Han:2014jya} after including the constraints from LHC $\eta_c$ data~\cite{Shao:2014yta,Han:2014jya}. They supersede the fits including the $P_T(J/\psi)>7$ GeV hadroproduction data described in Refs.~\cite{Ma:2010yw,Chao:2012iv}. These LDME sets cannot reproduce the CDF polarisation measurement~\cite{Abulencia:2007us} --like all the other sets in fact-- and are not applicable to $P_T(J/\psi)<7$ GeV. Both sets should only be used to di-$\psi$ production in the ATLAS fiducial region.
\item The NLO+LP fit of Ref.~\cite{Bodwin:2014gia} --as well as its update~\cite{Bodwin:2015iua}-- has been presented by its authors as the only fit able to reproduce the $J/\psi$ data (both yields and polarisations) above $10$ GeV after including the LP fragmentation contributions on top of the NLO calculations. However, it does not yield the correct $\eta_c$ cross section in the same $P_T$ region under the heavy-quark spin symmetry. As what concerns predictions for double $J/\psi$ production, it is marginally applicable only in the ATLAS fiducial region with $P_T(J/\psi)>8.5$ GeV instead of 10 GeV.
\end{enumerate}
Since we aim at a comprehensive analysis, we have considered all of the $9$ sets listed in Table~\ref{tab:coldme} to show how strongly the CO contributions depend on the available CO LDMEs. We should however recall during the discussion what we believe to be the region of applicability in $P_T(J/\psi)$ for the NLO(+LP) fits.
\begin{table*}[htpb]
\begin{center} \small
\begin{tabular}{c|c|c|c|c|c}
\cline{1-5}
& \multicolumn{3}{c}{LO fits} & \multicolumn{1}{|c|}{LL fit} & \\\cline{1-5}
& Set 1 & Set 2 & Set 3 & Set 4 & \\
\cline{1-5}
$\langle \mathcal{O}^{J/\psi}(^3S_1^{[1]})\rangle$ [GeV$^3$] & $1.2$ GeV$^3$ & $1.4$ & $1.16$ & $1.16$ & \\
$\langle \mathcal{O}^{J/\psi}(^3S_1^{[8]})\rangle$ [GeV$^3$]& $1.3\cdot 10^{-3}$ & $3.9\cdot 10^{-3}$ & $1.2\cdot 10^{-2}$ & $-9.3\cdot 10^{-3}$ & \\
$\langle \mathcal{O}^{J/\psi}(^1S_0^{[8]})\rangle$ [GeV$^3$]& $1.8\cdot 10^{-2}$ & $0$ & $0$ & $0.14$ & \\
$\langle \mathcal{O}^{J/\psi}(^3P_0^{[8]})\rangle$ [GeV$^5$]& $3.5\cdot 10^{-2}$ & $4.4\cdot 10^{-2}$ & $2.9\cdot 10^{-2}$ & $-3.9\cdot 10^{-2}$ & \\
$\langle \mathcal{O}^{\psi(2S)}(^3S_1^{[1]})\rangle$ [GeV$^3$]& $0.76$ & $0.67$ & $0.76$ & $0$ & \\
$\langle \mathcal{O}^{\psi(2S)}(^3S_1^{[8]})\rangle$ [GeV$^3$]& $3.3\cdot 10^{-3}$ & $3.7\cdot 10^{-3}$ & $5\cdot 10^{-3}$ & $0$ & \\
$\langle \mathcal{O}^{\psi(2S)}(^1S_0^{[8]})\rangle$ [GeV$^3$]& $8.0\cdot 10^{-3}$ & $0$ & $0$ & $0$ & \\
$\langle \mathcal{O}^{\psi(2S)}(^3P_0^{[8]})\rangle$ [GeV$^5$]& $1.6\cdot 10^{-2}$ & $5.0\cdot 10^{-3}$ & $1.2\cdot 10^{-2}$ & $0$ & \\
$\langle \mathcal{O}^{\chi_{c0}}(^3S_1^{[8]})\rangle$ [GeV$^3$]& $1.9\cdot 10^{-3}$ & $1.9\cdot 10^{-3}$ & $3.1\cdot 10^{-3}$ & $0$ & \\
$\langle \mathcal{O}^{\chi_{c0}}(^3P_0^{[1]})\rangle$ [GeV$^5$]& $0.11$ & $9.1\cdot 10^{-2}$ & $0.11$ & $0$ & \\
\hline
& \multicolumn{4}{c}{NLO fits} & \multicolumn{1}{|c}{NLO+LP fit} \\\hline
& Set 5 & Set 6 & Set 7 & Set 8 & Set 9 \\
\hline
$\langle \mathcal{O}^{J/\psi}(^3S_1^{[1]})\rangle$ [GeV$^3$]& $1.32$ GeV$^3$ & $1.16$ & $1.16$ & $1.16$ & $1.16$ \\
$\langle \mathcal{O}^{J/\psi}(^3S_1^{[8]})\rangle$ [GeV$^3$]& $2.2\cdot 10^{-3}$ & $-4.6\cdot 10^{-3}$ & $1.1\cdot 10^{-2}$ & $9.0\cdot 10^{-3}$ & $1.1\cdot 10^{-2}$ \\
$\langle \mathcal{O}^{J/\psi}(^1S_0^{[8]})\rangle$ [GeV$^3$]& $5.0\cdot 10^{-2}$ & $9.7\cdot 10^{-2}$ & $0$ & $1.5\cdot 10^{-2}$ & $9.9\cdot 10^{-2}$ \\
$\langle \mathcal{O}^{J/\psi}(^3P_0^{[8]})\rangle$ [GeV$^5$]& $-1.6\cdot 10^{-2}$ & $-2.1\cdot 10^{-2}$& $4.2\cdot 10^{-2}$ & $3.4\cdot 10^{-2}$ & $1.1\cdot 10^{-2}$ \\
$\langle \mathcal{O}^{\psi(2S)}(^3S_1^{[1]})\rangle$ [GeV$^3$]& $0$ & $0.76$ & $0.76$ & $0.76$ & $0$ \\
$\langle \mathcal{O}^{\psi(2S)}(^3S_1^{[8]})\rangle$ [GeV$^3$]& $0$ & $3.4\cdot 10^{-3}$ & $6.1\cdot 10^{-3}$ & $1.2\cdot 10^{-3}$ & $0$ \\
$\langle \mathcal{O}^{\psi(2S)}(^1S_0^{[8]})\rangle$ [GeV$^3$]& $0$ & $-1.2\cdot 10^{-4}$ & $0$ & $2.0\cdot 10^{-2}$ & $0$ \\
$\langle \mathcal{O}^{\psi(2S)}(^3P_0^{[8]})\rangle$ [GeV$^5$]& $0$ & $9.5\cdot 10^{-3}$ & $2.2\cdot 10^{-2}$ & $0$ & $0$\\
$\langle \mathcal{O}^{\chi_{c0}}(^3S_1^{[8]})\rangle$ [GeV$^3$]& $0$ & $2.2\cdot 10^{-3}$ & $2.2\cdot 10^{-3}$ & $2.2\cdot 10^{-3}$ & $0$\\
$\langle \mathcal{O}^{\chi_{c0}}(^3P_0^{[1]})\rangle$ [GeV$^5$]& $0$ & $0.11$ & $0.11$ & $0.11$ & $0$\\\hline
$P_T(J/\psi)$ region & $<30$ GeV & $>7$ GeV & $>7$ GeV & $>7$ GeV & $>10$ GeV \\
\hline
\end{tabular}
\caption{Summary of LDMEs we used from various fits [Set 1: Sharma et al.~\cite{Sharma:2012dy}; Set 2: Braaten et al.~\cite{Braaten:1999qk}; Set 3: Kr{\"a}mer~\cite{Kramer:2001hh}; Set 4: Sun et al.~\cite{Sun:2012vc}; Set 5: Butensch{\"o}n et al.~\cite{Butenschoen:2011yh}; Set 6 : Gong et al.~\cite{Gong:2012ug}; Set 7: Shao et al.~\cite{Shao:2014yta}: Set 8: Han et al.~\cite{Han:2014jya}: Set 9: Bodwin et al.~\cite{Bodwin:2014gia}].}
\label{tab:coldme}
\end{center}
\end{table*}
\subsection{Numerical results}
In this section, we will present our numerical results with the LO CO channels summed to the pure NLO$^\star$ CS channel $^3S_1^{[1]}+^3S_1^{[1]}$. Although the CS LDMEs $\langle\mathcal{O}^{J/\psi}(^3S_1^{[1]})\rangle$ and $\langle\mathcal{O}^{\psi(2S)}(^3S_1^{[1]})\rangle$ vary from set to set in Table \ref{tab:coldme}, we will fix these values for the NLO$^\star$ CS channel $^3S_1^{[1]}+^3S_1^{[1]}$ to those used in section ~\ref{sec:setup}. The uncertainty from these LDMEs is systematically subdominant compared to the scale uncertainty. All the feed-down contributions are properly taken into account as well.
\subsubsection{LHCb data at $\sqrt{s}=13$ TeV}
We start our discussion with the LHCb acceptance~\cite{Aaij:2016bqq}, where the $P_T(J/\psi)$ can be as low as zero. We have compared the CSM NLO$^\star$+COM LO SPS (the green bands) with the data in Figure~\ref{fig:colhcbdy} for the $\Delta y(J/\psi,J/\psi)$ distribution and in Figure~\ref{fig:colhcbdm} for the invariant mass of the pair $M(J/\psi+J/\psi)$ distribution. Like we have found for the CS LI contributions, the CO contributions are not relevant in the invariant mass distribution of LHCb. They start to be slightly visible in the tail of the $\Delta y(J/\psi,J/\psi)$ distribution.
\begin{figure}[hbt!]
\centering
\foreach \page/\ipage in {2/1}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax, scale=.32,draft=false,draft=false,trim = 0mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dy_sigma_LHCbarXiv161207451_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {9/2,8/3}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax, scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dy_sigma_LHCbarXiv161207451_CO.pdf}}\hspace*{-.2cm}
}
\foreach \page/\ipage in {4/4}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax, scale=.32,draft=false,trim = 0mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dy_sigma_LHCbarXiv161207451_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {1/5,3/6}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax, scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dy_sigma_LHCbarXiv161207451_CO.pdf}}\hspace*{-.2cm}
}
\foreach \page/\ipage in {6/7}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax, scale=.32,draft=false]{./Figures/LHCb/dy_sigma_LHCbarXiv161207451_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {7/8,5/9}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax, scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dy_sigma_LHCbarXiv161207451_CO.pdf}\hspace*{-.2cm}}
}
\caption{$\Delta y(J/\psi,J/\psi)$ distributions in NLO$^\star$ CS and LO CO via SPS within LHCb $\sqrt{s}=13$ TeV acceptance~\cite{Aaij:2016bqq}.}\label{fig:colhcbdy}
\end{figure}
\begin{figure}[hbt!]
\centering
\foreach \page/\ipage in {2/1}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/LHCb/dM_sigma_LHCbarXiv161207451_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {9/2,8/3}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dM_sigma_LHCbarXiv161207451_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {4/4}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/LHCb/dM_sigma_LHCbarXiv161207451_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {1/5,3/6}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dM_sigma_LHCbarXiv161207451_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {6/7}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/LHCb/dM_sigma_LHCbarXiv161207451_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {7/8,5/9}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/LHCb/dM_sigma_LHCbarXiv161207451_CO.pdf}\hspace*{-.2cm}}
}
\caption{$M(J/\psi+J/\psi)$ distributions in NLO$^\star$ CS and LO CO via SPS within LHCb $\sqrt{s}=13$ TeV acceptance~\cite{Aaij:2016bqq}.}\label{fig:colhcbdm}
\end{figure}
This observation however very much depends on the set of CO LDMEs used. In particular, the only plausible set, {\it i.e.}\ set 5, in the small $P_T(J/\psi)$ region does not yield any significant contribution to the cross section. It also seems clear that none of the sets can fully account for the discrepancy between SPS and LHCb data in the last bins of $\frac{d\sigma}{d\Delta y}$. Additional plots for the comparisons between CS NLO$^\star$+CO LO SPS and data can be found in appendix~\ref{app:comoreplots}. The impact of the CO contributions on these additional distributions is in general minor.
\subsubsection{CMS data at $\sqrt{s}=7$ TeV}
The events analysed by CMS have larger $P_T(J/\psi)$ above $4.5$ GeV to $6.5$ GeV depending on the rapidity. In this region, the only applicable NLO fit is still set 5 taken from Ref.~\cite{Butenschoen:2011yh}. As opposed to the conclusion made in Ref.~\cite{He:2015qya}, the CO SPS contribution is either much suppressed compared to the CS SPS contributions or much smaller than the experimental data as shown in Figure~\ref{fig:cocmsdy} and Figure~\ref{fig:cocmsdm}. Given that the LO fits (like that used in Ref.~\cite{He:2015qya} ({\it i.e.}\ set 2)) are not plausible any more and that the only applicable fit is the NLO fit given by set 5, we draw the conclusion that our extraction of DPS in Ref.~\cite{Lansberg:2014swa} --made by neglecting the CO contributions-- is still sound, which actually has been shown to be consistent with the ATLAS measurement thanks to a completely different method to disentangle the DPS from the SPS contributions and in a different kinematical region.
\begin{figure}[hbt!]
\centering
\foreach \page/\ipage in {2/1}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/CMS/dy_sigma_CMSarXiv14060484_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {9/2,8/3}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,trim = 21.8mm 0mm 0mm 0mm,clip,draft=false]{./Figures/CMS/dy_sigma_CMSarXiv14060484_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {4/4}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/CMS/dy_sigma_CMSarXiv14060484_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {1/5,3/6}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,trim = 21.8mm 0mm 0mm 0mm,clip,draft=false]{./Figures/CMS/dy_sigma_CMSarXiv14060484_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {6/7}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/CMS/dy_sigma_CMSarXiv14060484_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {7/8,5/9}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,trim = 21.8mm 0mm 0mm 0mm,clip,draft=false]{./Figures/CMS/dy_sigma_CMSarXiv14060484_CO.pdf}\hspace*{-.2cm}}
}
\caption{$\Delta y(J/\psi,J/\psi)$ distributions in NLO$^\star$ CS and LO CO via SPS within CMS $\sqrt{s}=7$ TeV acceptance~\cite{Khachatryan:2014iia}.}\label{fig:cocmsdy}
\end{figure}
\begin{figure}[hbt!]
\centering
\foreach \page/\ipage in {2/1}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/CMS/dM_sigma_CMSarXiv14060484_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {9/2,8/3}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/CMS/dM_sigma_CMSarXiv14060484_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {4/4}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/CMS/dM_sigma_CMSarXiv14060484_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {1/5,3/6}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/CMS/dM_sigma_CMSarXiv14060484_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {6/7}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/CMS/dM_sigma_CMSarXiv14060484_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {7/8,5/9}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/CMS/dM_sigma_CMSarXiv14060484_CO.pdf}\hspace*{-.2cm}}
}
\caption{$M(J/\psi,J/\psi)$ distributions in NLO$^\star$ CS and LO CO via SPS within CMS $\sqrt{s}=7$ TeV acceptance~\cite{Khachatryan:2014iia}.}\label{fig:cocmsdm}
\end{figure}
\subsubsection{ATLAS data at $\sqrt{s}=8$ TeV}
The transverse momentum cut on single $J/\psi$ is largest in the ATLAS data sample with selected events satisfying $P_T(J/\psi)>8.5$ GeV. This leaves the LDME sets 5-8 as possible good fits. ATLAS used a 2D $(|\Delta y(J/\psi,J/\psi)|,\Delta \phi(J/\psi,J/\psi))$ data-driven template fit to separate SPS and DPS events~\cite{Aaboud:2016fzt}. The control region used to determine the normalisation of DPS is $(|\Delta y(J/\psi,J/\psi)|\geq 1.8,\Delta \phi(J/\psi,J/\psi)\leq \frac{\pi}{2})$. The requirement of $\Delta \phi(J/\psi,J/\psi)\leq \frac{\pi}{2}$ will significantly reduce the CO fraction at large $|\Delta y(J/\psi,J/\psi)|$. The $t$-channel gluon exchange diagrams mainly make the two $J/\psi$ recoiling against each other. It thus populates the region where $\Delta \phi(J/\psi,J/\psi)\rightarrow \pi$. The simultaneous cuts on $|\Delta y(J/\psi,J/\psi)|$ and $\Delta \phi(J/\psi,J/\psi)$ ensure that the DPS extraction in Ref.~\cite{Aaboud:2016fzt} is reliable but for the low statistics in the control region. From Figure~\ref{fig:coatlasdy} and Figure~\ref{fig:coatlasdm}, one sees that the CO yields predicted with the set 7 \& 8 agree reasonably well with the data at large $M(J/\psi,J/\psi)$ and $|\Delta y(J/\psi,J/\psi)|$ with a slight overestimation in the middle of the corresponding distributions. The sets 5 \& 6 however do not agree with the data. Strong conclusions about the relevance of CO transitions in these regions would thus probably be premature in the absence of a complete NLO study and the disparate values of the existing CO LDMEs.
\begin{figure}[hbt!]
\centering
\foreach \page/\ipage in {2/1}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/ATLAS/dy_sigma_ATLASCONF2016047_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {9/2,8/3}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/ATLAS/dy_sigma_ATLASCONF2016047_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {4/4}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/ATLAS/dy_sigma_ATLASCONF2016047_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {1/5,3/6}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/ATLAS/dy_sigma_ATLASCONF2016047_CO.pdf}\hspace*{-.2cm}}
}
\foreach \page/\ipage in {6/7}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false]{./Figures/ATLAS/dy_sigma_ATLASCONF2016047_CO.pdf}}
}
\hspace*{-.35cm}
\foreach \page/\ipage in {7/8,5/9}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/ATLAS/dy_sigma_ATLASCONF2016047_CO.pdf}\hspace*{-.2cm}}
}
\caption{$\Delta y(J/\psi,J/\psi)$ distributions in NLO$^\star$ CS and LO CO via SPS within ATLAS $\sqrt{s}=8$ TeV acceptance~\cite{Aaboud:2016fzt}.}\label{fig:coatlasdy}
\end{figure}
\begin{figure}[hbt!]
\centering
\foreach \page/\ipage in {2/1}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,,draft=false]{./Figures/ATLAS/dM_sigma_ATLASCONF2016047_CO.pdf}}}
\hspace*{-.35cm}
\foreach \page/\ipage in {9/2,8/3}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/ATLAS/dM_sigma_ATLASCONF2016047_CO.pdf}\hspace*{-.2cm}}}
\foreach \page/\ipage in {4/4}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,,draft=false]{./Figures/ATLAS/dM_sigma_ATLASCONF2016047_CO.pdf}}}
\hspace*{-.35cm}
\foreach \page/\ipage in {1/5,3/6}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/ATLAS/dM_sigma_ATLASCONF2016047_CO.pdf}\hspace*{-.2cm}}}
\foreach \page/\ipage in {6/7}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,,draft=false]{./Figures/ATLAS/dM_sigma_ATLASCONF2016047_CO.pdf}}}
\hspace*{-.35cm}
\foreach \page/\ipage in {7/8,5/9}{
\subfloat[Set \ipage]{\includegraphics[page=\numexpr \page\relax,scale=.32,,draft=false,trim = 21.8mm 0mm 0mm 0mm,clip]{./Figures/ATLAS/dM_sigma_ATLASCONF2016047_CO.pdf}\hspace*{-.2cm}}}
\caption{$M(J/\psi,J/\psi)$ distributions in NLO$^\star$ CS and LO CO via SPS within ATLAS $\sqrt{s}=8$ TeV acceptance~\cite{Aaboud:2016fzt}.}\label{fig:coatlasdm}
\end{figure}
\section{Conclusions \label{sec:conclusion}}
We have examined two SPS production mechanisms for di-$J/\psi$ production at the LHC, which can be relevant in the control region used to determine the DPS. These are the partial LI CS contributions at $\mathcal{O}(\alpha_S^6)$ and the LO CO contributions at $\mathcal{O}(\alpha_S^4)$. We have also extensively compared our new SPS calculations with the existing LHC data. Our study indeed shows that the LI corrections can enhance the NLO$^\star$ SPS cross section at large $|\Delta y(J/\psi,J/\psi)|$ and large invariant mass $M(J/\psi+J/\psi)$. However, they are not sufficient to explain the discrepancy between SPS theoretical results and the LHC data in these regions. The inclusion of the DPS in the predictions is still crucial to account for the measurements.
On the other hand, the relevance of the CO contributions in the SPS yield strongly depends on the considered LDME set, thus with a very low predictive power --given the current status of understanding of the COM. It is in any case confined to the large $|\Delta y(J/\psi,J/\psi)|$ region. We anyhow conclude that the CO contributions can only be important when compared to the ATLAS data but that the ATLAS DPS extraction via a 2D data-driven fit is very likely free of any bias due to a possibly underestimated CO contribution in their control region. Such a conclusion is backed up by studies~\cite{Yamanaka:2018blj,CEM:forthcoming} made in the colour-evaporation model which offers a complementary framework to study the impact of CO transitions.
\begin{acknowledgments}
The work of JPL, HSS, YJZ is supported in part by CNRS via the LIA FCPPL. JPL is supported in part by the TMD@NLO IN2P3 project. The work of HSS is supported by the ILP Labex (ANR-11-IDEX-0004-02, ANR-10-LABX-63). YJZ is supported by the National Natural Science Foundation of China (Grants No. 11722539). NY was supported
by the JSPS Postdoctoral Fellowships for Research Abroad.
\end{acknowledgments}
|
\section{Introduction} \label{sec:intro}
Intermediate Polars (IPs) are a subgroup of Cataclysmic Variables (CVs) in which a moderately magnetic (10$^{5-7}$\,G) white dwarf (WD) accretes matter via Roche-lobe overflow from a companion, usually a low-mass star on or near the main sequence. An accretion disk is usually formed and its inner part is truncated by the magnetic field, which channels the gas in accretion columns towards the magnetic poles of the WD. Strong shocks are formed in the columns, heating up the gas to a high temperature defined in the first order by the depth of gravitational potential well of the WD ($k$T\,$\propto$\,M$_{WD}$/R$_{WD}$). The shocked gas subsequently settles onto the WD surface as it is cooled by radiating X-rays and cyclotron emission. Thus the X-ray emission is intrinsically multi-temperature in nature.
Information on the accretion geometry and the white dwarf mass is encoded in the intrinsic emission and its subsequent interaction with its surroundings. The observed radiation in IPs has a significant contribution from X-ray photons that have interacted with the environment, including the stellar surface of the WD itself, the accretion column, the accretion disk, and any circumstellar medium that might exist.
Part of the interacting X-rays suffers photoelectric absorption inducing a deficit in the observed X-rays which is more pronounced for softer photons. Another part is reflected and results in excess in the continuum around 10--30\,keV as a Compton hump and in fluorescent emission lines, notably the Fe\,K line at 6.4\,keV.
The description of both components, absorption and reflection, is necessary for an in-depth understanding of the accretion geometry and characteristics of the intrinsic X-rays, which can also lead to the determination of the WD mass. This task requires wide-bandpass spectroscopy covering both soft and hard X-rays -- especially because the partial-covering absorption, expected in IPs due to their complex environment, can have a similar effect on the spectral shape around the Fe K edge as reflection.
The advent of imaging hard X-ray observations using NuSTAR has allowed the detailed investigation of X-ray spectra of IPs, including both absorption and reflection \citep{2015ApJ...807L..30M}. While other recent papers have concentrated on deriving the white dwarf mass using hard X-ray spectroscopy \citep{2018MNRAS.476..554S}, including the effects of tall shocks and small inner radius of the disk \citep{2019MNRAS.482.3622S}, concerns remain that such investigations can be compromised if inadequate attention is paid to the complex interplay of absorption, reflection, and the intrinsic spectral shape. Here we present our detailed analysis of wide band X-ray spectra of two IPs, TV\,Col and V2731\,Oph, to test the potential degeneracy in spectral analysis results. We also follow \citet{2015ApJ...807L..30M} in analyzing the hard X-ray spin modulation as a method to break the degeneracy.
\section{The targets}
According to GAIA DR2 parallax, TV\,Col is at a distance of 513$^{+4}_{-5}$\,pc
\citep{2016A&A...595A...1G,2018arXiv180409365G}. The system has been investigated with many X-ray observatories since its first detection with the Ariel V satellite \citep[2A 0526-328;][]{1978MNRAS.182..489C}.
EXOSAT observations ($\sim$\,2--10\,keV) revealed it to be an absorbed hard X-ray emitter with flux modulated at 1,911\,s, interpreted as the white dwarf spin period \citep{1987Ap&SS.130..261S}. Based on the energy-dependent pulse profile from EXOSAT data, \citet{1989MNRAS.237..853N} argued for the presence of a complex absorption structure that could be explained by a partial-covering absorption.
The spin period was refined to 1,909.7$\pm$2.5 s by \citet{2004AJ....127..489R} from a nearly sinusoidal modulation by using RXTE, ROSAT, and ASCA data,
which also showed the 5.5\,h orbital modulation.
The authors also claimed the presence of a strong attenuation of the X-rays and signals of partial-covering absorption from hardness ratio variations, and spin modulation decreasing with energy. Broadband RXTE spectra (PCA and HXTE; $\sim$\,3--100\,keV) confirmed that TV\,Col is a hard, thermal, X-ray source presenting a clear Fe\,K$\alpha$ complex \citep{2005A&A...435..191S}.
A single temperature thermal fit of the high energy BAT/Swift (15--195\,keV) data resulted in a temperature of $k$T\,=\,21.6$\pm$2.4\,keV \citep{2009A&A...496..121B}. An analysis of the Suzaku data in the 3--50\,keV range, with the inclusion of a partial-covering absorber, indicated a shock temperature $k$T\,=\,45.7$^{+16.6}_{-9.1}$\,keV in a plasma with a subsolar abundance \citep[0.49\,Z$_{\sun}$;][]{2010A&A...520A..25Y}.
At a distance of 2300$^{+330}_{-270}$\,pc \citep[][]{2016A&A...595A...1G,2018arXiv180409365G}, V2731\,Oph is the most hard X-ray luminous IP known \citep[log L$_{(X;14-195\,keV)}$\,$>$\,33.9 for $d$\,$>$\,1\,kpc;][]{2014MNRAS.442.2580P}.
First investigated in X-rays by \citet{2008A&A...481..149D} using XMM-{\it Newton} and INTEGRAL observations, V2731\,Oph was revealed to have a complex X-ray emission. The investigation in a broadband energy (0.2-100\,keV) showed two optically thin plasma components, one cold ($k$T\,$\sim$\,0.2\,keV) and another hot ($k$T\,$\sim$\,60\,keV), with an additional contribution of a hot blackbody ($k$T\,$\sim$\,90\,eV). The intrinsic X-rays suffer the effects of a complex (local) distribution of cold intervening material plus a warm absorber, the latter being suggested by an OVII absorption edge at 0.74\,keV. As with TV\,Col, the pulse modulation for V2731\,Oph depends on the energy. In the case of V2731\,Oph, the X-rays are spin modulated (128.02$\pm$0.02\,s),
while UV emission is unmodulated at the 3$\sigma$ level \citep{2008A&A...481..149D}.
More recently, both TV\,Col and V2731\,Oph were observed with NuSTAR and investigated in the 3-50\,keV band by \citet{2016ApJ...826..160H} in the context of the X-ray emission of the Galactic center. We revisit those observations for a more in-depth investigation of the X-ray reflection and absorption in these systems, and consequently the characterization of IPs in general. We include in the investigation the Swift/XRT observation carried out during the NuSTAR observation of TV\,Col, and we show that the inclusion of soft X-rays is crucial to disentangle the effects of absorption and reflection in IP systems. For the same reason, we discuss the constraints imposed by previous XMM-{\it Newton}, Suzaku, and Swift observations of V2731\,Oph. Also, contrary to the previous study with NuSTAR data \citep{2016ApJ...826..160H}, we fit the spectra using the cooling flow model that takes into account the ionized Fe\,K lines.
\section{Observations and data reduction}
\label{sct:obs}
Details of the X-ray observations used in this investigation are given in Table \ref{tbl:obs}.
TV\,Col was observed by the NuSTAR satellite on 2014-05-11 for 49.7 ks (ObsID 30001020002). The system was also simultaneously observed by the Neil Gehrels Swift Observatory (hereafter Swift) for $\sim$\,1.9\,ks on 2014-05-12 (ObsID 00080734001). For the first time, we explored the corresponding NuSTAR FPMA/FPMB and Swift/XRT datasets to carry out consistent wide-band spectroscopy from 0.35 to 75 keV. We also investigate five other Swift/XRT observations in order to obtain clues on the spectral variations of the source.
We investigated V2731\,Oph using a NuSTAR observation carried out for $\sim$\,50\,ks on 2014-05-14 (ObsID 30001019002), for which there is no contemporaneous X-ray observation with another satellite. However, we used archival observations, namely XMM-{\it Newton}/EPIC (2005-08-29, by 13.5\,ks; ObsID 0302100201), Suzaku (2009-02-16, by 32.9\,ks; ObsID 403026010), and the longest Swift (2007-02-23, by 12.8\,ks; ObsID 00035086002) observations in order to evaluate the source in soft X-rays and discuss implications to the description of the system.
The data reduction followed standard procedures with the appropriated tools (with HEASOFT-v6.22 for the NuSTAR, Swift, and Suzaku data, and SAS-v16.1.0 for the XMM-{\it Newton} data), with the latest calibration files that were available in 2017 November.
We use different extraction regions for the NuSTAR modules, FPMA and FPMB, based on their individual images, to accommodate the relative astrometric offset between them.
All spectra from both TV\,Col and V2731\,Oph were binned such that each energy bin contains at least 25 counts, except for the Swift/XRT data of TV\,Col. For those data a 5 counts limit was adopted
as enforcing 25 counts per bin, making Gaussian errors a
reasonable approximation, would have meant too few spectral bins
for the spectrum to be useful.
Spectral fits were done with Xspec v12.9.1m by using the $\chi$-squared as fit and test statistics, but applying the C statistic for the fit in all cases in which the Swift/XRT data of TV\,Col were investigated
-- as it is appropriate to low count per energy bin, following the deviation from Gaussian to Poisson data.
\begin{table}
\begin{center}
\caption{Journal of X-ray observations. \label{tbl:obs}}
\begin{tabular}{lcccc}
\tableline\tableline
\tableline
& ObsID & Date & Exposure \\
\tableline
TV\,Col \\
\tableline
Swift & 00035282001 & 2007-03-30 & 7\,ks \\
& 00035282002 & 2007-04-23 & 0.8\,ks \\
& 00037150002 & 2007-12-21 & 19.4\,ks \\
& 00037150003 & 2007-12-27 & 4.6\,ks \\
& 00037150005 & 2008-01-14 & 5.3\,ks \\
& 00080734001 & 2014-05-12 & 1.9\,ks \\
NuSTAR & 30001020002 & 2014-05-11 & 49.7\,ks \\
\tableline
V2731\,Oph\\
\tableline
XMM-{\it Newton}& 0302100201 & 2005-08-29 & 13.5\,ks\\
Swift & 00035086002 & 2007-02-23 & 12.8\,ks\\
Suzaku & 403026010 & 2009-02-16 & 32.9\,ks\\
NuSTAR & 30001019002 & 2014-05-14 & 50\,ks \\
\tableline
\end{tabular}
\end{center}
Notes: The covered energy range was 0.3-10\,keV for the XMM-{\it Newton}, Swift, and Suzaku satelites; for NuSTAR, it was 3--60\,keV for TV\,Col and 3--75\,keV (or 20--75\,keV; see text) for V2731\,Oph.\\
\end{table}
\begin{table*}
\begin{center}
\caption{Spectral models. \label{tbl:models}}
\begin{tabular}{llllllll}
\tableline\tableline
\tableline
Target & Model & Description \\
\tableline
TV\,Col & M1 & \textsc{constant*phabs*[(apec} or \textsc{mkcflow) + gauss]} \\
& M2 & \textsc{constant*phabs*[reflect*(apec} or \textsc{mkcflow) + gauss]} \\
& M3 & \textsc{constant*phabs*pwab*[(apec} or \textsc{mkcflow) + gauss]}\\
& M4 & \textsc{constant*phabs*pwab*[reflect*(apec} or \textsc{mkcflow) + gauss]} \\
& M5 & \textsc{constant*phabs* pwabs*[reflect*(apec} or \textsc{mkcflow) + (apec} or \textsc{bbody) + gauss]} \\
\tableline
V2731\,Oph & M6 & \textsc{constant*phabs*pwab*edge*(apec+mkcflow+bbody+gaussian)} \\
& M7 & \textsc{constant*phabs*pwab*edge*(apec+reflect*mkcflow+bbody+gaussian)} \\
& M8 & \textsc{constant*phabs*pwab*edge*(vapec+vmcflow+gaussian)}\\
& M9 & \textsc{constant*phabs*pwab*edge*(vapec+vmcflow+bbody+gaussian+gaussian+gaussian)}\\
\tableline
\end{tabular}
\end{center}
\end{table*}
\section{X-ray spectroscopy}
Throughout this work, we use \textsc{xspec} and models available within it to carry out modeling of XMM-{\it Newton}, Swift, Suzaku, and NuSTAR X-ray spectra. Two {\it optically thin} thermal models were used -- a single temperature \textsc{apec} model and a multi-temperature cooling flow model with \textsc{mkcflow}, also based on \textsc{apec} (with switch parameter equal to 2), which is model M1. In addition, an optically thick blackbody model, \textsc{bbody}, was included as necessary. The abundance table applied in both optically thin thermal models was that of \citet{2009ARA&A..47..481A}. The redshift required by construction in the \textsc{mkcflow} model cannot be zero. It was assumed to be 8.5867$\times$10$^{-8}$ for TV\,Col
and 5.367$\times$10$^{-7}$ for V2731\,Oph,
as estimated from the GAIA DR2 distances \citep[513$^{+4}_{-5}$\,pc and 2300$^{+330}_{-270}$\,pc, respectively;][]{2016A&A...595A...1G,2018arXiv180409365G} and standard cosmological values of \textsc{xspec}. The low temperature of the \textsc{mkcflow} was fixed to the minimum value allowed by the model ($k$T\,=\,80.8\,eV).
The interaction of the radiation with matter was accounted by the following models.
For the photoelectric absorption by cold matter, we use both the
single absorber \textsc{phabs} and the complex absorber \textsc{pwab} \citep{1998MNRAS.298..737D} components. The usual partial covering
absorber models (either the \textsc{pcfabs} multiplicative model or the
\textsc{partcov} convolution model combined with any absorption model)
represent two lines of sights, one with and one without an intervening
absorber. However, the absorption in the pre-shock accretion columns of
magnetic CVs requires a model with a continuous distribution of
covering fraction as a function of column height. \citet{1998MNRAS.298..737D}
developed \textsc{pwab} to represent this situation in which the covering
fraction is assumed to be a power-law function of the column ($\beta$ index,
from a minimum (N$_{H,min}$) to a maximum (N$_{H,max}$) equivalent hydrogen columns).
An \textsc{edge} component was used to account for a warm absorber, only necessary in the case of V2731\,Oph, associated with the OVII absorption edge at 0.74\,keV.
The \textsc{reflect} model was applied when investigating reflection of X-rays in the systems -- in all cases keeping the (unconstrained) inclination angle ($i$) of the reflecting surface equal to the default value in the model, such that cos($i$) is equal to 0.45.
The \textsc{gauss} model was used to describe the 6.4\,keV fluorescence iron line, with the line energy (E$_{c}$) and width ($\sigma$) fixed to 6.4\,keV and 0.01\,keV, respectively.
An energy-independent multiplicative factor, the \textsc{constant} model, was included to account for the cross-calibration uncertainties of the different instruments (namely, the impact on normalization in spectral modeling) when two or more spectra were used in simultaneous fits.
\textsc{constant} also helps with source flux variability, since the observations are not all simultaneous.
All parameters cited above as being fixed were not constrained when they were allowed to vary freely during the fit.
We notice that the statistical significance for the NuSTAR and Swift simultaneous fit of the TV\,Col spectra is biased to the NuSTAR data, and thus the visual check was crucial to choose between the models.
For TV\,Col, we started with the simplest model based on single absorption for the two optically thin thermal models cited above (\textsc{apec} and \textsc{mkcflow}; M1),
then adding the reflection model (M2),
replacing reflection model by the partial-covering model (M3),
and finally using both partial-covering and reflection models (M4).
In all cases, a Gaussian line at 6.4 keV was added to describe the Fe\,K$\alpha$ fluorescence feature.
Some Swift spectra of TV\,Col presented an excess in soft X-rays which was accounted for as an additional thermal component (M5). Thus, the four initial models are a subgroup of M5 as described in Table \ref{tbl:models}.
For V2731\,Oph, the starting point was the result of \citet{2008A&A...481..149D} from XMM-{\it Newton} and INTEGRAL data. Finally, the model applied to the system was: \textsc{constant*phabs*pwab*edge*(mkcflow+apec+bbody+ gaussian)}.
Table \ref{tbl:models} summarizes the models.
When using the \textsc{reflect} component we extended the energy range over which the model is calculated to 100\,keV, because photons with energies above the instrumental coverage can be Compton down-scattered to the energy range which is covered by the instrument (command \textsc{energies extend high 100.0}, in \textsc{xspec}). The output spectrum was limited to 80\,keV to speed up the calculation of the spectrum (command \textsc{xset reflect\_max\_e 80.0} in \textsc{xspec}).
\begin{figure*}
\centerline{
\includegraphics[angle=-90,scale=1]{tvcol_swiftnustar_spct_resid4100.eps}
}
\caption{TV\,Col: NuSTAR/FPMA and Swift/XRT spectra fitted with models M1 (cyan), M2 (blue), M3 (green), and M4$_{\rm m}$ (black; see text for futher details). The lower panel
shows the residual with respect to M4$_{\rm m}$.} \label{fig:tvspct}
\end{figure*}
\subsection{TV\,Col}
TV\,Col is among the brightest hard X-ray IPs \citep{2014MNRAS.442.2580P}. Its X-rays are due to optically-thin, thermal plasma, as supported by the strong Fe\,K line complex.
Because its Galactic latitude of -30.6 deg and distance of 513$^{+4}_{-5}$\,pc, most of the interstellar medium (ISM) in its direction is between us and TV Col. However, the system occupies a relatively unobscured part of the Galaxy with an integrated HI column density estimate of only $\sim$\,2.39$\times$10$^{20}$\,cm$^{-2}$ \citep{2005A&A...440..775K}.
Thus, the X-rays from TV\,Col are only slightly affected by the ISM and any absorption we measure is due to local contributions.
These circumstances make it an ideal target for a detailed study of intrinsic absorbers in the system and therefore of reflection effects. We base our spectral analysis on these two premises and proceed from the simplest model to more complex ones, until satisfactory fits were obtained.
\subsubsection{Broad band spectroscopy from simultaneous Swift and NuSTAR observations}
\label{sct:TVCol_joiningfit}
We first investigate TV\,Col by simultaneously fitting the contemporaneous NuSTAR (FPMA\,+\,FPMB; 3--60\,keV) and Swift/XRT spectra (0.35--10\,keV; Sect. \ref{sct:obs}) that together covered the wide energy range from 0.35 to 60 keV. The iron line complex (via NuSTAR) and the soft part of the spectrum (via XRT) were always included, which played a key role in disentangling the models.
A thermal model affected by a single absorption component (M1) with either \textsc{apec} and \textsc{mkcflow} does not describe the data. It fails in the description of the continuum, especially at low energies (see Fig. \ref{fig:tvspct}). However, this simple model immediately revealed the presence of a complex absorption pattern in the system. This model underestimates the flux for E\,$<$\,1.5\,keV appreciably, showing that the absorber with a high column density that is required to explain the spectral shape at higher energies absorbs too many soft photons.
Similar results were obtained when including the reflection component (M2).
The fit was substantially improved when the reflection component was replaced by the partial-covering model (M3), but still yielded an unacceptable description for energies up to $\sim$\,6\,keV, below which the model oscillates between underestimating and overestimating the flux. Finally, the reflection was added to M3, by using the M4 model. This model significantly improved the fit over the whole spectrum, as described below.
\begin{table}
\begin{center}
\caption{TV\,Col: the best-fit spectral parameters from the simultaneous fit of the contemporaneous NuSTAR (FPMA \& FPMB) and Swift/XRT spectra from best fit M4$_{\rm m}$ model (as in Table \ref{tbl:models}). \label{tbl:spct_tvcol}}
\begin{tabular}{llllllll}
\tableline\tableline
\tableline
Component & Parameter & Value \\
\tableline
\textsc{phabs} & N$_{H}$ (10$^{22}$\,cm$^{-2}$) &0.0239$^{(*)}$\\
\tableline
\textsc{pwab} & N$_{H,min}$ (10$^{22}$\,cm$^{-2}$) & 10$^{-7}$\,$^{(*)}$\\
& N$_{H,max}$ (10$^{22}$\,cm$^{-2}$) & 10.5$^{+1.7}_{-0.8}$\\
& $\beta$ &-0.41$^{+0.03}_{-0.04}$\\
\tableline
\textsc{reflect} & rel$_{refl}$ & 0.88$^{+0.13}_{-0.13}$\\
& cos($i$) & 0.45\,$^{(*)}$\\
\tableline
\textsc{mkcflow} & kT$_{low}$ (keV) & 0.0808\,$^{(*)}$ \\
& kT$_{max}$ (keV) & 31.0$^{+1.1}_{-1.1}$\\
& Z ($\times$Z$_{\odot}$) & 0.49$^{+0.04}_{-0.04}$\\
& redshift & 1.197$\times$10$^{-7}$\,$^{(*)}$ \\
& switch & 2\\
\tableline
$\chi^{2}_{\nu}$/d.o.f. & & 0.95/1007 \\
\tableline
\end{tabular}
\end{center}
Notes: (*) fixed parameter (see text).\\
\end{table}
Contrary to the first three models, M4 describes the continuum from 0.35 to 60\,keV (see Fig. \ref{fig:tvspct}).
The distinction between the model M4 using \textsc{apec} (M4$_{\rm a}$) or \textsc{mkcflow} (M4$_{\rm m}$) comes from the description of the Fe K complex and the surrounding continuum. While the spectral fit from \textsc{apec} substantially underestimates the flux from the ionized Fe\,XXV line, the \textsc{mkcflow} model describes both it and the Fe\,XXVI lines and results in a better description of the continuum at $\sim$\,4--8\,keV -- M4$_{\rm m}$, the assumed model for TV\,Col in this work, resulting in $\chi^2_{\nu}$ of 0.95 for 1007 degrees of freedom.
We present the best-fit parameters for M4$_{\rm m}$ in Table \ref{tbl:spct_tvcol}. The equivalent hydrogen column from the \textsc{phabs} model is constrained to be less than 3.0$\times$10$^{20}$\,cm$^{-2}$ such that any intrinsic absorption would be mitigated by the effect of the dominant local partial-covering absorber. Thus, although the values of all parameters are still mutually consistent at 1$\sigma$, we assume a conservative upper limit by fixing the column value to the total Galactic column in the direction of TV\,Col \citep[2.39$\times$10$^{20}$\,cm$^{-2}$;][]{2005A&A...440..775K}.
This procedure improves the determination of the absorption derived from the partial-covering absorption \textsc{pwab} model, resulting in N$_{H,max}$\,=\,1.05$^{+0.17}_{-0.08}$$\times$10$^{23}$\,cm$^{-2}$ and $\beta$\,=\,$-$0.41$^{+0.03}_{-0.04}$. The primary source of X-rays is a high-temperature plasma with an abundance of 0.49$\pm$0.04 times the solar values, cooling down from a maximum temperature $k$T of 31.0$\pm$1.1\,keV. If we take this as the measurement of the gravitational potential at the surface of the WD, then its mass is 0.735$\pm$0.015 M$_\odot$ \citep[following][]{1973PThPh..49.1184A}. Our subsequent analysis justifies this assumption, as we discuss in Section \ref{sct:tandrefl}.
We find that the reflection of the primary X-rays makes a non-negligible contribution to the observed X-rays. According to the best-fit model (M4$_{\rm m}$), the reflection amplitude is constrained to be 0.88$^{+0.13}_{-0.13}$.
The reflection accounts for about 16\% out of a total unabsorbed flux of $\sim$\,1.37$\times$10$^{-10}$ erg\,s$^{-1}$\,cm$^{-2}$ at 0.3--75\,keV, increasing to 30\% out of a total of $\sim$\,5.3$\times$10$^{-11}$ erg\,s$^{-1}$\,cm$^{-2}$ at 10--50\,keV (and 7.6\% out of $\sim$\,7.9$\times$10$^{-11}$ erg\,s$^{-1}$\,cm$^{-2}$ at 0.3--10\,keV). TV\,Col was observed with a luminosity of 4.3$\times$10$^{33}$\,erg\,s$^{-1}$ at 0.3--75\,keV on 2014-05-11/12, and a bolometric X-ray luminosity of (4.6$\pm$0.3)$\times$10$^{33}$\,erg\,s$^{-1}$.
The NuSTAR data suggest a pronounced fluorescent iron line at 6.4\,keV with an equivalent width (EW) estimated in 179$^{+19}_{-61}$ eV. As we show below, this value is significantly higher than that derived for V2731\,Oph with the same instrument, even though V2731\,Oph is significantly more absorbed than TV\,Col. Thus, the fluorescent line at least for TV\,Col seems to have the contribution not only from the absorber but also from the reflector.
\subsubsection{Spectral evolution from Swift/XRT observations}
We investigated individually five other archival Swift/XRT observations of TV\,Col in order to evaluate its spectral evolution (Table \ref{tbl:obs}). Thermal models resulted in unconstrained values for the plasma temperature and abundance while they are consistent with those obtained from the simultaneous NuSTAR and Swift observations. Subsequently, we fitted the spectra by assuming the model M4$_{\rm m}$, keeping fixed the plasma temperature and abundance values of the \textsc{mkcflow} and the parameters of the \textsc{reflect} component to those determined previously from the NuSTAR/Swift observations of 2014-05-11/12, presented in Table \ref{tbl:spct_tvcol}.
We also adopted the same fixed parameter values as done previously for the M4$_{\rm m}$ model. The best-fit spectral parameters are summarized in Table \ref{tbl:tvcolspctparam}.
There is an excess emission below $\sim$\,1\,keV that indicates the presence of an additional thermal component for two observations (0003715002 and 0003715005), and which is suspected for two others (00035282001 and 0003715003). Its description suggests that it is more likely due to a cool optically thin plasma ($k$T\,$\sim$\,0.15\,keV) but a blackbody emission ($k$T\,$\sim$\,0.12\,keV) cannot be ruled out.
No strong change in spectral shape of TV\,Col is observed in the first five Swift observations listed in Table \ref{tbl:obs} but they still exhibit variability in the absorbers and also in the unabsorbed flux at 0.3-10\,keV. Remarkably, during the fifth observation, the values increased by 50--100\%.
\begin{table*}
\begin{center}
\caption{TV\,Col: best-fit spectral parameters from Swift/XRT observations. \label{tbl:tvcolspctparam}}
\begin{tabular}{cccccccc}
\tableline\tableline
\tableline
& \multicolumn{2}{c}{\textsc{pwab}} & \textsc{apec} & \textsc{bbody} & & \\
\cline{2-3}
Swift/XRT & N$_{H,max}$ & $\beta$ & $k$T & $k$T & F$_{unabs,0.3-10keV}$ & $\chi^{2}_{\nu}$/d.o.f. \\
(ObsID) & (10$^{22}$\,cm$^{-2}$)& & (keV) & (keV) & ($\times$10$^{-11}$\,erg\,cm$^{-2}$\,s$^{-1}$)& \\
\tableline
00035282001 & 13.8$^{+3.4}_{-2.1}$ & -0.41$^{+0.03}_{-0.03}$ & ... & ... & 6.5 & 1.04/398 \\
& 12.7$^{+2.6}_{-2.7}$ & -0.38$^{+0.04}_{-0.01}$ & 0.14$^{+0.06}_{-0.04}$ & ... & 6.7 & 1.04/396 \\
& 12.7$^{+3.0}_{-3.2}$ & -0.38$^{+0.09}_{-0.02}$ & ... & 0.11$^{+0.06}_{-0.05}$ & 6.6 & 1.04/396 \\
00035282002 & 13.6$^{+8.0}_{-5.1}$ & -0.40$^{+0.09}_{-0.06}$ & ... & ... & 6.9 & 0.88/65 \\
00037150002 & 17.2$^{+3.9}_{-1.4}$ & -0.37$^{+0.02}_{-0.03}$ & 0.18$^{+0.01}_{-0.02}$ & ... & 8.6 & 1.12/689 \\
& 15.9$^{+2.1}_{-0.2}$ & -0.35$^{+0.01}_{-0.02}$ & ... & 0.12$^{+0.01}_{-0.01}$ & 8.5 & 1.13/689 \\
00037150003 & 13.4$^{+3.5}_{-2.9}$ & -0.35$^{+0.06}_{-0.04}$ & 0.18$^{+0.05}_{-0.06}$ & ... & 7.8 & 1.01/297 \\
& 10.1$^{+4.2}_{-2.4}$ & -0.23$^{+0.20}_{-0.11}$ & ... & 0.15$^{+0.02}_{-0.03}$ & 7.9 & 1.01/297 \\
00037150005 & 25.3$^{+17.6}_{-5.4}$ & -0.43$^{+0.02}_{-0.04}$ & 0.15$^{+0.03}_{-0.02}$ & ... & 11.2 & 1.00/389 \\
& 31.7$^{+13.6}_{-9.2}$ & -0.45$^{+0.03}_{-0.03}$ & ... & 0.09$^{+0.02}_{-0.02}$ & 11.6 & 1.03/389 \\
\tableline
\end{tabular}
\end{center}
Notes: from M4$_{\rm m}$ or M5 (see Table \ref{tbl:models}).\\
\end{table*}
\subsection{V2731\,Oph}
It is clear from the investigation of TV\,Col that a simultaneous or at least contemporaneous broadband coverage is essential in describing the effects of both absorption and reflection of X-rays and hence in characterizing the system. Even though V2731\,Oph has been observed with high signal-to-noise ratio in soft and hard X-rays, the coverage in both energy bands with different satelites was not contemporaneous.
We started by comparing the previous XMM-{\it Newton} observation (carried out on 2005-08-29) already reported by \citet{2008A&A...481..149D} with the Suzaku \citep[2009-02-16;][]{2010A&A...520A..25Y} and the longest Swift (2007-02-23) archival observations of V2731\,Oph (Table \ref{tbl:obs}).
Those observations were investigated separately and, subsequently, via a joint fit with the 20-75 keV NuSTAR (2014-05-14) spectra.
For the purpose of obtaining the based fit of the XMM-{\it Newton}
data below 10 keV, we fit them jointly with the NuSTAR spectra
restricted to 20-75 keV range, allowing us to utilize the
parameters that best describe the high energy shape of the
spectrum. In this procedure, we ignored the NuSTAR data below
20 keV, because of our suspicion (later confirmed) that there
is more spectral variability at low energies.
We then proceeded with an investigation of the NuSTAR observations of the system separately using the whole available spectral range (3-75\,keV), and then with the Swift and Suzaku spectra, also separately, in all cases fixing the values of the cooling flow component to those derived from the simultaneous fit of the XMM-{\it Newton} and the 20-75 keV NuSTAR spectra. Figure \ref{fig:v2731ophspct} shows the spectra.
\subsubsection{Clues about the spectral evolution}
\label{sect:V2731_spctevol}
As reported by \citet{2008A&A...481..149D}, an XMM-{\it Newton} observation revealed a complex X-ray spectrum of V2731\,Oph.
A good fit required multiple thermal plasma components, with both a cold and another hot optically thin plasmas, and an additional blackbody emission. Once produced, the X-rays are affected by a complex distribution of cold matter and also suffers absorption from a warm material, as indicated by the OVII absorption edge at 0.74\,keV. The authors showed that the spectral distribution is well explained using the model \textsc{wabs*pcfabs*edge(mekal+bbody+mekal+gaussian)}.
We first applied the model used by \citet{2008A&A...481..149D} to carry out a simultaneous fit of the three EPIC XMM-{\it Newton} spectra (pn, MOS1, and MOS2) and reproduced their results. However, in the following we adopt the (updated) \textsc{phabs} model instead of \textsc{wabs}, and the (more appropriate) \textsc{pwab} model in the place of \textsc{pcfabs}. Also, we replace the \textsc{apec} describing the hottest thermal component by a \textsc{mkcflow} model, in line with the case of our other target, TV\,Col, and as it is expected for IPs. Thus, the final model is M6: \textsc{constant*phabs*pwab*edge*(mkcflow+apec+bbody+ gaussian)}.
The model M6 describes the XMM-{\it Newton} data well, resulting in $\chi^{2}_{\nu}$\,=\,1.05 for 834 degrees of freedom, but the hottest component of the \textsc{mkcflow} component converged to the hard limit of $k$T\,=\,79.9\,keV. The coldest component, as for the case of TV\,Col, was not constrained and therefore fixed to $k$T\,=\,80.8\,eV. We added to the fit the NuSTAR spectra limited to the 20--75\,keV energy range. This inclusion improves the statistic to $\chi^{2}_{\nu}$\,=\,1.01, for 962 degrees of freedom, while resulting in well-constrained values (Table \ref{tbl:v2731spctparam}). Starting with the plasma parameters, the temperature derived from the \textsc{mkcflow} model peaks at $k$T\,=\,47.6$^{+4.2}_{-4.5}$\,keV.
A cold ($k$T\,=\,0.17$^{+0.01}_{-0.02}$\,keV), optically thin plasma component is necessary to account for a substantial part of the $\sim$\,0.5--1.5\,keV range which is not described by the \textsc{mkcflow} model even taking its low temperature as a free parameter during the fit. A fit without the cold plasma, but with the low temperature left to vary free, only achieves $\chi^{2}_{\nu}$\,=\,1.15.
On the other hand, we confirm that a blackbody component (with $k$T\,=\,0.094$^{+0.002}_{-0.002}$\,keV) is a good description for the excess emission in soft X-rays, not explained by the optically thin plasma emission. However, this is unphysical for a white dwarf, as it is argued at the end of this section. In terms of absorption, the modeling with \textsc{pwab} resulted in a complex structure with N$_{H,max}$\,=\,156$^{+270}_{-42}$$\times$10$^{22}$\,cm$^{-2}$ and $\beta$\,=\,-0.68$^{+0.01}_{-0.01}$. Although poorly constrained, the lower limit to N$_{H,max}$ is still high: at least $\sim$\,10$^{24}$\,cm$^{-2}$. The N$_{H,min}$ parameter, unconstrained, was fixed to 10$^{15}$\,cm$^{-2}$ -- the minimum value allowed by the model. The \textsc{phabs} resulted in the equivalent to N$_{H}$\,=\,(3.3$\pm$0.2)$\times$10$^{21}$\,cm$^{-2}$.
Such a value is higher than the total Galactic HI column density in the direction of V2731\,Oph of about 1.6$\times 10^{21}$ cm$^{-2}$ estimated by \citet{2005A&A...440..775K} but slightly lower than the value of 3.9$\times 10^{21}$ cm$^{-2}$ as derived from the 3D Dust Mapping with Pan-STARRS 1\footnote{http://argonaut.skymaps.info/} \citep{2018MNRAS.478..651G}.
The abundance was tied to the \textsc{apec} and \textsc{mkcflow} models, resulting in Z\,=\,0.30$\pm$0.09\,Z$_{\odot}$. This value is consistent with the two values reported by \citet{2008A&A...481..149D}, of 0.33$^{+0.37}_{-0.19}$\,Z$_{\odot}$ from the 0.3-10\,keV XMM-{\it Newton} EPIC data alone and 0.40$^{+0.07}_{-0.08}$Z$_{\odot}$ when the 20-100 keV INTEGRAL/ISGRI spectrum was also included in the analysis. The energy of the \textsc{edge} component was fixed at 0.74\,keV \citep[corresponding to the OVII absorption edge reported by][]{2008A&A...481..149D}.
\begin{figure*}
\centerline{
\includegraphics[angle=-90,scale=1]{v2731_spct.eps}
}
\caption{V2731\,Oph: XMM/pn, Swift/XRT, NuSTAR/FPMA, Suzaku/XIS0, and Suzaku/XIS1 spectra and the corresponding descriptions from M6. The inset presents the unfolded XMM/pn and NuSTAR/FPMA data and modeling around the Fe\,K complex. For Suzaku/XIS0 and XIS1, see zoom in Fig. \ref{fig:v2731Suzaku}. \label{fig:v2731ophspct}}
\end{figure*}
The aim of this analysis is to compare how the spectral energy distribution has evolved when comparing observations carried out at different epochs. In this exercise we explore the longest archival Swift observations and the Suzaku observation of V2731\,Oph, verifying the consistency of the spectra with the model used to describe the 0.3--10\,keV XMM-{\it Newton} plus the 20--75\,keV NuSTAR spectra.
We first investigate the Swift/XRT spectrum (ObsID 00035086002).
The model and values derived from the joint XMM-{\it Newton} and NuSTAR analysis, keeping only the normalization as free parameters, results in an unacceptable description of the Swift spectrum ($\chi^{2}_{\nu}$\,=\,3.18).
The statistic is slightly improved when allowing the parameter of the \textsc{phabs} to vary ($\chi^{2}_{\nu}$\,=\,2.54).
No improvement is obtained by assuming the N$_{H,max}$ parameter of the \textsc{pwab} free during the fit, while it is not constrained, but a significant improvement is obtained allowing the $\beta$ to vary ($\chi^{2}_{\nu}$\,=\,1.59) -- keeping the lower limit for the absorption still fixed.
The statistic improves to $\chi^{2}_{\nu}$\,=\,1.14 when the absorption depth of the \textsc{edge} component is also a free parameter.
Allowing the plasma temperature of the \textsc{apec} component to also vary improves the fit description ($\chi^{2}_{\nu}$\,=\,1.06).
On the other hand, there is no improvement when allowing the temperature
of the \textsc{bbody} component and the high temperature parameter of the \textsc{mkcflow} model to vary.
As their values are consistent with those derived from the XMM-{\it Newton} plus NuSTAR analysis, we kept them fixed to the corresponding values. The best-fit spectral parameters are summarized in Table \ref{tbl:v2731spctparam}. The analysis revealed a significant change in the partial-covering absorption and in the temperature of the cold plasma component, while the spectral energy distribution of hard X-rays remained essentially unchanged.
\begin{table*}
\begin{center}
\caption{V2731\,Oph: best-fit spectral parameters. \label{tbl:v2731spctparam}}
\begin{tabular}{ccccccccccc}
\tableline\tableline
\tableline
&\textsc{phabs}& \multicolumn{2}{c}{\textsc{pwab}$^{1}$} & \textsc{edge}$^2$ & \textsc{apec} & \textsc{bbody} & \textsc{mkcflow}$^{3}$ & & \\
\cline{3-4}
&N$_{H}$ & N$_{H,max}$ & $\beta$ & abs. depth & $k$T & $k$T & $k$T$_{max}$ &$Z$$^4$ & $\chi^{2}_{\nu}$/d.o.f. \\
\cline{2-3} \cline{6-8}
& \multicolumn{2}{c}{(10$^{22}$\,cm$^{-2}$)} && ($@$0.74\,keV) & \multicolumn{3}{c}{(keV)} & ($Z_{\odot}$) & \\
\tableline
\tableline
\tableline
\multicolumn{5}{l}{Without \textsc{reflect} from M6 and its variants:}\\
\tableline
\multicolumn{8}{l}{Model: \textsc{constant*phabs*pwab*edge*(apec + mkcflow + bbody + gaussian)}} \\
XMM\,+\,NuSTAR$^5$ & 0.33$^{+0.02}_{-0.02}$ & 156$^{+270}_{-42}$ & -0.68$^{+0.01}_{-0.01}$ & 1.86$^{+0.08}_{-0.09}$ & 0.17$^{+0.01}_{-0.02}$ & 0.094$^{+0.002}_{-0.002}$ & 47.6$^{+4.2}_{-4.5}$ & 0.30$^{+0.09}_{-0.09}$ & 1.01/962 \\
Swift & 0.29$^{+0.03}_{-0.03}$ & 156$^6$ & -0.75$^{+0.04}_{-0.04}$ & 1.46$^{+0.22}_{-0.24}$ & 0.64$^{+0.05}_{-0.06}$ & 0.094$^2$ & 47.6$^6$ & 0.30$^6$ & 1.02/150 \\
\tableline
\multicolumn{8}{l}{Model: \textsc{constant*pwab*edge*(apec + mkcflow + bbody + gaussian)}} \\
Suzaku (XIS0+3) & ... & 156$^6$ & -0.78$^{+0.01}_{-0.01}$ & 0.72$^{+0.07}_{-0.07}$ & 0.25$^{+0.01}_{-0.01}$ & 0.21$^{+0.01}_{-0.01}$ & 47.6$^6$ & 0.30$^6$ & 1.06/1663 \\
Suzaku (XIS0+1+3) & ... & 156$^6$ & -0.80$^{+0.01}_{-0.01}$ & 0.75$^{+0.04}_{-0.04}$ & 0.28$^{+0.01}_{-0.01}$ & 0.18$^{+0.01}_{-0.01}$ & 47.6$^6$ & 0.30$^6$ & 1.18/2574 \\
\tableline
\multicolumn{8}{l}{Model: \textsc{constant*pwab*(mkcflow + gaussian)}} \\
NuSTAR (3-75\,keV) & ... & 753$^{+60}_{-62}$ & -0.68$^{+0.01}_{-0.01}$ & ... & ... & ... & 47.6$^6$ & 0.30$^6$ & 1.00/782\\
NuSTAR (3-75\,keV) & ... & 766$^{+131}_{-88}$ & -0.68$^{+0.01}_{-0.01}$ & ... & ... & ... & 46.7$^{+3.6}_{-3.7}$ & 0.30$^6$ & 1.02/781\\
\tableline
\tableline
\tableline
\multicolumn{5}{l}{With \textsc{reflect} from M7 and its variants:}\\
\tableline
\multicolumn{8}{l}{Model: \textsc{constant*phabs*pwab*edge*(apec + reflect*mkcflow + bbody + gaussian)}} \\
XMM\,+\,NuSTAR$^5$ & 0.36$^{+0.02}_{-0.02}$ & 150$^{+487}_{-62}$ & -0.76$^{+0.01}_{-0.01}$ & 1.83$^{+0.09}_{-0.09}$ & 0.17$^{+0.01}_{-0.02}$ & 0.093$^{+0.002}_{-0.002}$ & 54.5$^{+6.4}_{-4.3}$ & 0.23$^{+0.07}_{-0.05}$ & 0.98/962\\
XMM\,+\,NuSTAR$^5$ & 0.36$^{+0.02}_{-0.02}$ & 121$^{+135}_{-35}$ & -0.76$^{+0.01}_{-0.01}$ & 1.83$^{+0.09}_{-0.09}$ & 0.17$^{+0.01}_{-0.02}$ & 0.092$^{+0.002}_{-0.002}$ & 55.1$^{+6.2}_{-4.4}$ & 0.30$^6$ & 0.98/963\\
Swift & 0.59$^{+0.09}_{-0.11}$ & 121$^6$ & -0.93$^{+0.03}_{-0.03}$ & 1.28$^{+0.24}_{-0.27}$ & 0.22$^{+0.02}_{-0.01}$ & 0.04$^{+0.01}_{-0.01}$ & 55.1$^6$ & 0.30$^6$ & 0.97/149
\\
\tableline
\multicolumn{8}{l}{Model: \textsc{constant*pwab*edge*(apec + reflect*mkcflow + bbody + gaussian)}} \\
Suzaku (XIS0+3)& ... & 121$^6$ & -0.88$^{+0.01}_{-0.01}$ & 0.74$^{+0.06}_{-0.07}$ & 0.25$^{+0.01}_{-0.01}$ & 0.20$^{+0.01}_{-0.01}$ & 55.1$^6$ & 0.30$^6$ & 1.05/1663 \\
Suzaku (XIS0+1+3) & ... & 121$^6$ & -0.90$^{+0.01}_{-0.01}$ & 0.74$^{+0.04}_{-0.04}$ & 0.30$^{+0.01}_{-0.01}$ & 0.18$^{+0.01}_{-0.01}$ & 55.1$^6$ & 0.30$^6$ & 1.18/2574 \\
\tableline
\multicolumn{8}{l}{Model: \textsc{constant*pwab*(reflect*mkcflow + gaussian)}} \\
NuSTAR (3-75\,keV) & ... & 848$^{+174}_{-167}$ & -0.82$^{+0.01}_{-0.01}$ & ... & ... & ... & 55.1$^6$ & 0.30$^6$ & 0.96/782\\
NuSTAR (3-75\,keV) & ... & 895$^{+206}_{-198}$ & -0.82$^{+0.01}_{-0.01}$ & ... & ... & ... & 55.8$^{+5.3}_{-3.8}$ & 0.30$^6$ & 0.98/781\\
\tableline
\end{tabular}
\end{center}
Notes:
$^1$ N$_{H,min}$\,=\,10$^{15}$\,cm$^{-2}$;
$^2$ threshold energy equal to 0.74\,keV;
$^3$ $k$T$_{low}$\,=\,80.8\,eV;
$^4$ $Z$\, tied to the \textsc{apec}, \textsc{mkcflow}, and, when present, \textsc{reflect} components, for all the elements;
$^5$ using only the 20-75\,keV NuSTAR spectra (FPMA and FPMB);
$^6$ fixed to the values of the joining XMM\,+\,NuSTAR analysis.
For the \textsc{reflect} component, when applied, the reflection scaling factor is equal to 1 and cos($i$)\,=\,0.45.
The lack of coverage for E\,$<$\,3\,keV from NuSTAR data does not allow a reliable determination of the flux at 0.3-10\,keV.
For the Suzaku analysis, see text.
\\
\end{table*}
We extended the same methodology to the Suzaku XIS0, XIS1, and XIS3 spectra. Allowing only the normalization to vary during the fit resulted in $\chi^{2}_{\nu}$\,=\,6.30. As for the XRT spectrum, the description is improved when allowing the absorption column of the \textsc{phabs} ($\chi^{2}_{\nu}$\,=\,3.70), the $\beta$ parameter of the \textsc{pwab} model ($\chi^{2}_{\nu}$\,=\,2.34), the absorption depth of the \textsc{edge} component ($\chi^{2}_{\nu}$\,=\,1.54), and the plasma temperature of the \textsc{apec} component ($\chi^{2}_{\nu}$\,=\,1.42) to vary. The column derived from the \textsc{phabs} model is virtually equal to zero (N$_{H}$\,$\sim$\,10$^{16}$\,cm$^{-2}$) and unphysical, and this component was removed without impacting the fit. The lower limit of the absorption in the \textsc{phabs} converges to the minimum value when assumed as a free parameter. In contrast with the Swift spectrum, the fit for the Suzaku spectra is statistically improved when the temperature of the \textsc{bbody} component is free during the fit ($k$T\,$\sim$\,0.18\,keV; $\chi^{2}_{\nu}$\,=\,1.22). Table \ref{tbl:v2731spctparam} includes the best-fit parameters of the Suzaku observation.
\begin{figure*}
\centerline{
\includegraphics[angle=-90,scale=1]{v2731_spctSuzaku_withresid.eps}
}
\caption{V2731\,Oph: Suzaku/XIS0 and XIS3 fitted with model M9 (green and blue, respectively) and Suzaku/XIS1 (cyan) with models M6 (red) and M9 (black). See text for futher details.}
\label{fig:v2731Suzaku}
\end{figure*}
However, there are four significant issues with the fit to the Suzaku data.
First, the model underestimates the flux around 0.5\,keV and also around 0.8 keV (see Fig. \ref{fig:v2731Suzaku}).
Second, the column density for the \textsc{phabs} component is unrealistically low in comparison with the total Galactic HI column density in this direction ($\sim$\,1.6$\times 10^{21}$ cm$^{-2}$ according to \citet{2005A&A...440..775K} and 3.9$\times 10^{21}$ cm$^{-2}$ from the 3D Dust Mapping with Pan-STARRS 1, by \citet{2018MNRAS.478..651G}). Given the large distance to V2731 Oph (at least 2 kpc), we expect that a substantial fraction of this column is between V2731 Oph and the solar system. The fact that N$_H$ for the \textsc{phabs} component was negligible when
fitting the Suzaku data, while it was higher than the Galactic value in
the XMM-Newton and Swift data, points to a possible problem with the fit.
Third, the blackbody temperature, already worryingly high in the XMM-{\it Newton} fit, is so high to be clearly unphysical for a white dwarf \citep{1987MNRAS.226..725W}; we will return to this point in Section \ref{sct:v2731complex}.
Fourth, this model does not fit the XIS1 data, which show significant residuals below $\sim$0.8\,keV, especially around 0.5\,keV. While some cross-calibration issues among the XIS units of Suzaku are commonly seen, what we see in V2731\,Oph is far beyond residuals seen in other Suzaku observations. One likely explanation is that the model is wrong but nevertheless works well enough, phenomenologically, with XMM-{\it Newton}, Swift/XRT and Suzaku XIS0+XIS3 data (Table \ref{tbl:v2731spctparam}). In fact, the best-fit spectral parameters from the XIS0+XIS3 data with M6 and M7-like models are similar to those obtained from the XIS0+XIS1+XIS3 data.
Because the Suzaku/XIS1, using an advanced backside-illuminated (BI) CCD chip, has a relatively high effective area below 0.8\,keV and a superior spectral resolution at these energies than those of the other cameras used in this work, some inadequacies of these models becomes readily visible in the XIS1 data. Thus, we explore in the next section an alternative spectral description to V2731\,Oph.
\begin{table*}
\begin{center}
\caption{V2731\,Oph: best-fit spectral parameters by using \textsc{vapec} and \textsc{vmcflow}. \label{tbl:v2731highnitrogen}}
\begin{tabular}{llcccccc}
\tableline
& & M8 & M8 & M9\\
& & XMM\,+\,NuSTAR$^1$ & Suzaku & Suzaku \\
& & & (XIS0+1+3) & (XIS0+1+3) \\
\tableline
\textsc{phabs} \\
&N$_{H}$ (10$^{22}$\,cm$^{-2}$) & 0.14$^{+0.02}_{-0.02}$ & $<$\,0.004 & 0.29$^{+0.02}_{-0.03}$\\%2
\tableline
\textsc{pwab}$^2$ \\
& N$_{H,max}$ (10$^{22}$\,cm$^{-2}$) & 612$^{+183}_{-155}$ & 374 & 68$^{+15}_{-20}$\\%4
& $\beta$ & -0.63$^{+0.01}_{-0.01}$ & -0.79$^{+0.01}_{-0.01}$ & -0.77$^{+0.02}_{-0.01}$\\%5
\tableline
\textsc{edge}$^3$ \\
& Abs. depth ($@$0.74\,keV) & 1.49$^{+0.08}_{-0.08}$ & 0.52$^{+0.06}_{-0.05}$ & 0.97$^{+0.08}_{-0.06}$\\%7
\tableline
\textsc{vapec} \\
& $k$T (keV) & 0.23$^{+0.01}_{-0.01}$ & 0.50$^{+0.01}_{-0.02}$ & 0.49$^{+0.02}_{-0.02}$\\%8
\tableline
\textsc{vmcflow}$^4$ \\
& $k$T$_{max}$ (keV) & 43.4$^{+5.0}_{-4.0}$ & 43.4$^6$ & 47.6$^7$\\
\tableline
& $Z_{N}$ ($Z_{N,\odot}$)$^5$ & 5.8$^{+1.0}_{-0.7}$ & 16.9$^{+2.3}_{-2.1}$ & 69$^{+16}_{-10}$\\%11
& $Z_{O}$ ($Z_{O,\odot}$)$^5$ & 0.5$^{+0.1}_{-0.1}$ & 1.3$^{+0.2}_{-0.1}$ & 4.6$^{+0.8}_{-0.5}$\\%12
& $Z_{F;Ni}$ ($Z_{F;Ni;\odot}$)$^5$ & 0.37$^{+0.04}_{-0.04}$ & 0.19$^{+0.01}_{-0.02}$ & 0.6$^{+0.1}_{-0.1}$\\%20
\tableline
\textsc{bbody} \\
& $k$T (eV) & ... & ... & 39$^{+5}_{-4}$\\%43
\tableline
\textsc{gaussian}\\
& Line energy (keV) & ... & ... & 0.58$^{+0.01}_{-0.01}$\\%48
& Line energy (keV) & ... & ... & 0.94$^{+0.01}_{-0.01}$\\%45
\tableline
& $\chi^{2}_{\nu}$/d.o.f. & 1.08/962 & 1.19/2571 & 1.12/2565\\
\tableline
\tableline
\end{tabular}
\end{center}
Notes: $^1$ Using only the 20-75\,keV NuSTAR spectra (FPMA and FPMB);
$^2$ N$_{H,min}$\,=\,10$^{15}$\,cm$^{-2}$;
$^3$ threshold energy equal to 0.74\,keV;
$^4$ $k$T$_{low}$\,=\,80.8\,eV;
$^5$ $Z$\, tied to the \textsc{vapec} and \textsc{vmcflow} components, and fixed to 0.3\,Z$_{\odot}$ for all other elemental abundance;
$^6$ fixed to the value of the joining XMM\,+\,NuSTAR analysis from M8; $^7$ fixed to the value of the joining XMM\,+\,NuSTAR analysis from M6 (in Table \ref{tbl:v2731spctparam}).
\end{table*}
\subsubsection{Alternative spectral description}
The Suzaku/XIS1 spectrum indicates a more pronounced excess around 0.5\,keV than in other spectra, suggesting the presence of an intense N{\small VII}\,L$\alpha$ line (see red line from M6, in Fig. \ref{fig:v2731Suzaku}). In order to check for an overabundance of N, we replace \textsc{apec} by \textsc{vapec} and \textsc{mkcflow} by \textsc{vmcflow} and apply the model without reflection to the Suzaku XIS0, XIS1, and XIS3 spectra, simultaneously -- as the variants of the first models that allow the user to set the abundance of the main elements individually. The abundance for the \textsc{vapec} and \textsc{vmcflow} were tied together, allowing Z$_{N}$ to vary freely during the fit but keeping all other elemental abundances equal to 0.30\,Z$_{\sun}$. The fit improves significantly with a better description of the soft tail ($\chi^2_{red}$ from 1.50 to 1.25; yet unnaceptable and not presented). It indicates Z$_{N}$\,=\,3.2($\pm$0.3)\,Z$_{\odot}$ while the normalization of the \textsc{bbody} component tends to zero and therefore that such a component would not be necessary anymore.
The \textsc{phabs} N$_{H}$ value of 7.4$^{+1.0}_{-0.9}$$\times$10$^{20}$\,cm$^{-2}$ is consistent with the ISM value while the complex absorption via \textsc{pwab} indicates a N$_{H,max}$ of at least 10$^{24}$\,cm$^{-2}$ with $\beta$ equal to -0.90$\pm$0.01. Another difference with respect to the values inferred from the other spectra is the substantially lower absorption depth associated with the Oxygen edge, of 0.21$\pm$0.05.
We then assumed the O abundance as a free parameter during the fit, and also the Fe which was tied to the Ni abundance. The fit resulted in Z$_{N}$\,=\,16.9$^{+2.3}_{-2.1}$\,Z$_{\odot}$, Z$_{O}$\,=\,1.3$^{+0.2}_{-0.1}$\,Z$_{\odot}$, and Z$_{Fe}$\,(=\,Z$_{Ni}$)\,=\,0.19$\pm$0.02\,Z$_{\odot}$, with $\chi^2_{red}$\,=\,1.19.
In this scenario, the absorption from the \textsc{phabs} is less than 10$^{20}$\,cm$^{-2}$, with the \textsc{pwab} component indicating a N$_{H,max}$ of at least 10$^{24}$\,cm$^{-2}$ and $\beta$ equal to -0.79$\pm$0.01.
The temperature from the \textsc{vbapec} component is $k$T\,=\,0.50$^{+0.1}_{-0.2}$\,keV ($k$T$_{max}$ in \textsc{vmcflow} was kept fixed to 43.4\,keV as it was derived from the XMM+NuSTAR for the same model).
The improvement with this model comes from a better description of the continuum around 2-5\,keV but it degrades the description of the ionized iron lines around 6.7\,keV and also fails to explain the X-rays in the 0.55-0.7\,keV and 1-1.5\,keV energy ranges. The addition of another thermal plasma component did not improve the fit.
Again, as in the analysis presented in Section \ref{sect:V2731_spctevol}, the \textsc{phabs} results in a unrealistically low column density. The results are summarized in Table \ref{tbl:v2731highnitrogen}.
The model has the OVII edge at 0.74\,keV, and some of the oxygen ions that absorbed such photons would emit the OVII He-like triplet. We then added a Gaussian line to the variable abundance model and its energy centroid converged to 0.58\,keV, the expected value of the OVII triplet. It results in $\chi^{2}_{\nu}$\,=1.15 but this time not only with low column density from the \textsc{phabs} component but understimating the spectra below 0.6\,keV. We them added the \textsc{bbody} component, resulting in a better description. The continuum below 0.6\,keV is well described, the column density from \textsc{phabs} converges to $\sim$3.8$\times$10$^{21}$\,cm$^{-2}$, and the temperature of the \textsc{bbody}, contrary to the model presented in Section \ref{sect:V2731_spctevol}, has a low temperature of $\sim$\,28\,eV but an unrealistic normalization of 2.5. The $\chi^{2}_{\nu}$\, is 1.16.
Because of the excess observed around 0.90-0.94\,keV, which is consistent with the position of the NeIX triplet, we add a new Gaussian component. The centroid energies of the Gaussian lines converge to 0.58$\pm$0.01\,keV and 0.94$\pm$0.01\,keV in good agreement with what is expected for the OVII and NeIX He-like triplets, respectively.
For the blackbody component, we derive $k$T\,=\,39$^{+5}_{-4}$\,eV and normalization of (4.3$^{+4.0}_{-2.4}$)$\times$10$^{-2}$.
The column from the \textsc{phabs} component is 2.9$^{+0.2}_{-0.3}$$\times$10$^{21}$\,cm$^{-2}$, consistent with the value expected from the ISM in the light of sight of V2731\,Oph. The fit also results in N$_{H,max}$\,=\, 68$^{+15}_{-20}$$\times$10$^{22}$\,cm$^{-2}$ and $\beta$\,=\,-0.77$^{+0.02}_{-0.01}$ for \textsc{pwab}, and $k$T\,=\,0.49$\pm$0.02\,keV for \textsc{vapec}. For the \textsc{vmcflow} component, we kept $k$T$_{max}$\,=\,47.6\,keV.
Also the fit, with $\chi^2_{red}$\,=\,1.12, resulted in
Z$_{N}$\,=\,69$^{+16}_{-10}$\,Z$_{\odot}$, Z$_{O}$\,=\,4.6$^{+0.8}_{-0.5}$\,Z$_{\odot}$,
and Z$_{Fe}$\,(=\,Z$_{Ni}$)\,=\,0.6$^{+0.1}_{-0.1}$\,Z$_{\odot}$.
The best fit spectral parameters are presented in Table \ref{tbl:v2731highnitrogen}. In Fig. \ref{fig:v2731Suzaku} we present the fits from M6 and M9 for Suzaku/XIS1 (lines red and black, respectively) where are also presented the XIS0 and XIS3 spectra, and the corresponding results from M9.
We applied the above models to the XMM-{\it Newton} EPIC spectra. The fit with variable Z$_N$ was found to be statistically worse than the previous description using model M6 (Table \ref{tbl:v2731spctparam}), with $\chi^{2}_{\nu}$ of 1.14. When both Z$_{N}$ and Z$_{O}$ are allowed to be free, without the O absorption edge, results in difficulties to explain the emission from 0.7 to 0.8\,keV. Finally, we applied M8 (see Table \ref{tbl:models}) to the combined XMM-{\it Newton} and 20--75\,keV NuSTAR data, and obtained a $\chi^{2}_{\nu}$ of 1.08 (compared to 1.01 for M6) with similar parameter values when exploring only the XMM-{\it Newton} spectra, particularly $k$T$_{max}$, which shows no statistically significant changes. The results are shown in Table \ref{tbl:v2731highnitrogen}. The addition of a \textsc{bbody} and the two \textsc{gaussian} line as in M9 do not improve the fit. With a low temperature ($k$T\,$\sim$\,15\,eV), the \textsc{bbody} component converges to a normalization of only $\sim$\,10$^{-19}$ meaning that such an addition is not supported from the data. A \textsc{gaussian} component at 0.94\,keV, as applied to the Suzaku data, is not constrained by the fit and fixing its parameters to the values derived from the Suzaku spectra does not improve the fit. Also, although still in agreement with the XMM-{\it Newton} data, the \textsc{gaussian} line around 0.58\,keV results in no significant improvement in the fit, still with $\chi^{2}_{\nu}$ of 1.08. The values derived from M8 (Table \ref{tbl:v2731highnitrogen}) are mutually consistent with those when applying M9 to the 0.3--10\,keV XMM-{\it Newton} plus the 20-75\,keV NuSTAR spectra althought these fits are not shown.
Thus, even though we have not
found a completely satisfactory fit to the soft X-ray spectrum, our idea to combine XMM-{\it Newton} and 20--75\,keV NuSTAR data appears to be valid. First, the spectral energy distribution for hard X-rays in both set of data are mutually consistent -- and confirmed from the whole spectral range covered by NuSTAR (Section \ref{sct:V2731Oph_Nustar}).
The difficulties with either model in describing the Suzaku data with all applied models, and the XMM-{\it Newton} spectra with the alternative M8 and M9 models, are limited to the soft end of the spectrum ($<$1\,keV).
We notice that the elemental abundances derived independently from the Suzaku and XMM-{\it Newton} spectra do not converge, both indicate that anomalies are present at least for the N, O, and Fe.
Second, the absorbed and unabsorbed 20-75\,keV flux differ by less than 5\%, showing that the absorption such as detected in the system, even though it is very high, has no significant rule in this spectral region.
In addition, the investigation revealed that the luminosity of the source at 20--75\,keV during the XMM-{\it Newton} observation is $\sim$\,35\% less than that observed during the NuSTAR observation, with the unabsorbed flux varying from
$\sim$\,2.5$\times$10$^{-11}$\,erg\,cm$^{-2}$\,s$^{-1}$ to $\sim$\,3.9$\times$10$^{-11}$\,erg\,cm$^{-2}$\,s$^{-1}$ (over 9 years).
The estimated unabsorbed 20--75\,keV flux from both Swift and Suzaku datasets is $\sim$\,2.1$\times$10$^{-11}$\,erg\,cm$^{-2}$\,s$^{-1}$ (over 2 years).
Since the soft X-ray complexity is limited to energies below 0.8\,keV, we will proceed by returning to the spectral model of \citet{2008A&A...481..149D} in the analysis of hard X-ray data, and comment on the soft X-ray complexity in Section \ref{sct:v2731complex}.
\subsubsection{Spectroscopy using the full NuSTAR data}
\label{sct:V2731Oph_Nustar}
The variability of V2731\,Oph in X-rays prevents the use of non-contemporaneous campaigns to carry out a simultaneous spectral fit.
Therefore, we instead explored the full spectral range covered by NuSTAR (3--75\,keV) initially using models M6. As expected, the \textsc{phabs} (that goes virtually to zero; N$_{H}$\,=\,10$^{13}$\,cm$^{-2}$), \textsc{bbody}, \textsc{apec}, and \textsc{edge} components do not contribute to the fit and they were removed from the model.
A fit that keeps the N$_{H,max}$ of the \textsc{pwab} component fixed to 156$\times$10$^{22}$\,cm$^{-2}$ and the $k$T$_{hot}$ of the \textsc{mkcflow} to 47.6\,keV, as derived from the analysis of the XMM-{\it Newton} and 20--50\,keV NuSTAR spectra, yields $\chi^{2}_{\nu}$\,=\,1.48.
Freezing the thermal component to 47.6\,keV and allowing the N$_{H,max}$ to vary during the fit results in a good description of the spectra with N$_{H,max}$\,=\,753$^{+60}_{-62}$$\times$10$^{22}$\,cm$^{-2}$ and $\beta$\,=\,-0.68$^{+0.01}_{-0.01}$, for $\chi^{2}_{\nu}$\,=\,1.00 (Table \ref{tbl:v2731spctparam}). The abundance was kept fixed to 0.30\,$Z_{\odot}$. Similar values were obtained when allowing the high temperature of the \textsc{mkcflow} varying during the fit.
kT$_{max}$ of 47.6$^{+4.2}_{-4.5}$\,keV derived from the joint NuSTAR-XMM-{\it Newton} fit suggests a WD mass of 0.92$\pm$0.04 M$_\odot$, if the shock temperature reflects the gravitational potential just above the white dwarf surface \citep[following][]{1973PThPh..49.1184A}.
\subsubsection{Investigating the hypothesis of reflection}
We move back to the spectral investigation of V2731\,Oph to verify how the occurrence of X-ray reflection in the system would affect the results. Contrary to the case of TV\,Col, it was not possible to constrain the reflection - if it is present - for V2731\,Oph. Adding the \textsc{reflect} component and allowing the reflection scaling factor to vary during the fits resulted, systematically, in values greater than 3, which are unrealistic. We froze this value to 1 while keeping cos$i$\,=\,0.45 and see how this conservative assumption would impact the spectral results when reflection is present. In doing this, we followed the same procedure as in the case without \textsc{reflect}: we investigate the EPIC XMM-{\it Newton} and the 20--75\,keV NuSTAR (FPMA\,+\,FPMB) simultaneously, then individually the Swift/XRT, the Suzaku (XIS0+XIS3 and XIS0+XIS1+XIS3), and the 3--75\,keV NuSTAR spectra.
The results are summarized in Table \ref{tbl:v2731spctparam}.
$k$T$_{max}$ of 55.1$^{+6.2}_{-4.4}$\,keV derived from the joint NuSTAR-XMM-{\it Newton} fit (with the abundance fixed to 0.3\,Z$_\odot$) suggests a WD mass of 0.99$^{+0.05}_{-0.03}$ M$_\odot$, if the shock temperature reflects the gravitational potential just above the white dwarf surface.
The inclusion of reflection does not change at the 1\,$\sigma$ level the high temperature in the \textsc{mkcflow} model. But this result can be masked by effects of the photoelectric absorption, that is not well constrained from the NuSTAR data alone. A conclusive investigation of reflection in V2731\,Oph demands high signal-to-noise spectroscopy covering at least the 0.3--75\,keV energy range.
High-resolution spectroscopy would play a crucial role in the description of the lines in the soft X-rays and therefore of the continuum and, finally, of the intervening absorber. As for now, our results suggest that 35\% of the total flux at 10-50\,keV derived for V2731\,Oph from the XMM-{\it Newton} fit can be due to reflection -- similar to the fraction inferred for TV\,Col, in which the presence of reflection is clear.
We estimate from the XMM-{\it Newton} (pn; on 2005/08/29) observation an EW of 145$\pm$40\,eV for the fluorescent iron line at 6.4\,keV. The NuSTAR data have a poorer spectral resolution but suggest a lower value on 2014/05/14 of about 69\,eV even in a situation in which X-rays seems to be more absorbed than they were during the XMM-{\it Newton} observation (Figure \ref{fig:v2731ophspct}, inset). Perhaps the low equivalent width suggests that the 6.4\,keV line can be accounted only by the absorbers with only a small contribution, if any, from reflection.
\subsubsection{X-ray luminosity and mass accretion rate}
V2731\,Oph is a luminous IP in the X-rays, with a high mass accretion rate. Here we concentrate on results from the XMM-{\it Newton} and NuSTAR analyses. From the cooling flow model, we infer directly from its normalization parameter value that the mass accretion rate during the XMM-{\it Newton} observation (2005-08-29) was 5.1$^{+2.7}_{-0.7}$$\times$10$^{-9}$ M$_{\odot}$\,yr$^{-1}$, with a bolometric luminosity of 4.2$\times$10$^{35}$($d$/2300 pc)$^{2}$\,erg\,s$^{-1}$, while the rate during the NuSTAR observation (2014-05-14) was (8.4$\pm$0.2)$\times$10$^{-9}$ M$_{\odot}$\,yr$^{-1}$. Because NuSTAR does not cover energies below 3\,keV, where the system has a significant contribution from additional components which are therefore not constrained from those data, we cannot infer the X-ray bolometric luminosity securely from the NuSTAR data. To carry out a more realistic comparison, we adopt the luminosity in the 3--75\,keV band. The values are 4.4$\times$10$^{34}$($d$/2300 pc)$^{2}$\,erg\,s$^{-1}$ and 7.8$\times$10$^{34}$($d$/2300 pc)$^{2}$\,erg\,s$^{-1}$ for the XMM-{\it Newton} and NuSTAR observations, respectively.
We estimate that the 3-75\,keV luminosity is about 10\% less than the bolometric X-ray luminosity.
\section{Timing properties from the NuSTAR observations}
The NuSTAR observations allowed us to investigate the spin variability of both TV\,Col and V2731\,Oph in hard X-rays.
We proceeded from background subtracted light curves in three energy bands: 3--6 keV, 6--10 keV, and 10--30 keV.
We chose 60 s bins for TV Col, given its long spin period (resulting
in 32 bins per cycle), and 10 s bins for V2731 Oph, given its short
period (13 bins per cycle).
Since the NuSTAR count rates of these IPs are modest, we neglected deadtime correction, which should be small. We also neglected barycentric correction to simplify the analysis procedure, since we intended to limit our analysis to the times of the NuSTAR observations. We combined the FPMA and FPMB light curves in the three bands and searched for periodic modulations using Fourier power spectrum and the normalization according to \citet{1982ApJ...263..835S}.
There is no Swift data at the time of the NuSTAR observation of
V2731 Oph, while that contemporaneous with the NuSTAR observation
of TV Col is too short to provide conclusive results.
As timing analysis of all
archival X-ray data is beyond the scope of our work, given that extensive body of
work already exists in print, we focus on the NuSTAR observations.
\begin{figure}
\centerline{
\includegraphics[angle=-0,scale=0.53]{tvcol-scargle-new.eps}
}
\caption{TV\,Col: periodogram from NuSTAR data in three X-ray energy bands; 3-6\,keV (bottom), 6-10\,keV (middle) and 10-30 keV (top).
The set of panels show the power spectra over the range of periods
we investigated (longer than 120\,s). The blue dashed line is drawn
at the known spin period of TV\,Col, 1909.7\,s, while the red dashed
line indicates the prominent peak at 2976\,s. The right set of panels
show the power spectra around 1909.7\,s in more detail.
\label{fig:tvcol_scargle}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=-0,scale=0.53]{tvcol-fold-new.eps}}
\caption{TV\,Col: folded light curves (on the prominent period of 2976 s, left, and on the known spin period of 1909.7 s, right) from NuSTAR data in three X-ray energy bands; 3-6\,keV (bottom), 6-10\,keV (middle) and 10-30 keV (top). \label{fig:tvcol_lc}}
\end{figure}
\subsection{TV\,Col}
The known spin modulation in TV\,Col, refined by \citet{2004AJ....127..489R} as 1,909.7$\pm$2.5\,s, was not detected in the NuSTAR data. Figure \ref{fig:tvcol_scargle} shows the power spectra up to 8 mHz from the three energy ranges cited above, where the blue vertical dashed lines mark the position of the expected spin frequency. The red lines in the figure mark the highest peak below 10\,keV, at $\sim$\,2,976\,s, while it is the second in power in the 10--30\,keV energy range after the one at very low frequency ($\sim$\,96,000\,s). The peak at 2,976\,s is statistically significant if white (i.e., frequency independent) noise is assumed. However, no previous reports of a $\sim$3,000 s period exist, so we consider this to be a noise peak.
\begin{figure}
\centerline{
\includegraphics[angle=-0,scale=0.53]{v2731scargle-new.eps}}
\caption{V2731\,Oph: periodogram from NuSTAR data in three X-ray energy bands; 3-6\,keV (bottom), 6-10\,keV (middle) and 10-30 keV (top).
The set of panels show the power spectra over the range of periods
we investigated (longer than 20\,s). The right set of panels show
the power spectra around 128\,s in more detail (note the different
vertical scales).
\label{fig:v2731_scargle}}
\end{figure}
\begin{figure}
\centerline{
\includegraphics[angle=-0,scale=0.53]{v2731fold-new.eps}
}
\caption{V2731\,Oph: light curves folded on the 127.98\,s spin
period from NuSTAR data in three X-ray energy bands; 3-6\,keV (bottom), 6-10\,keV (middle) and 10-30\,keV (top). \label{fig:v2731_lc}}
\end{figure}
We show folded light curves with the periods of 1,909\,s and 2,976\,s in the left and right panels of Fig. \ref{fig:tvcol_lc}, respectively.
As anticipated from the power spectrum, no modulation is associated with the spin period.
Using amplitude/average convention, the relative modulation amplitudes for the 2,976 s peak are
$\sim$\,7.5\% (3-6 keV), $\sim$\,6.6\% (6--10 keV), and 8.4\% (10--30 keV), and they can be taken as
conservative upper limits of the spin (1,909.7\,s) signal.
In comparison, \citet{2004AJ....127..489R} found spin amplitudes of 14\% (2-5 keV), 8.5\% (5-10 keV), and 6\% (10-20 keV). Such modulations would have been detectable in the NuSTAR data below 10 keV, if not above. Thus, we conclude that the X-ray spin amplitude of TV Col is variable and that we caught it in a low spin modulation state.
We note that optical observations of TV\,Col have long indicated
variable spin modulation amplitude \citep[see, e.g.,][]{1985A&A...143..313B,1988MNRAS.233..759B}.
\subsection{V2731\,Oph}
The 128\,s spin period of V2731 Oph \citep{2008A&A...481..149D} and the first harmonic are
clearly dominant in the periodogram presented in Fig. \ref{fig:v2731_scargle}.
The left panels show the periodogram over the entire frequency range (0-50 mHz,
from the light curves with time bin of 10\,s) in the three X-ray energy bands under investigation. All the three left panels have the same Y scale.
The right panels show the expanded view around the fundamental frequency (with variable Y scales), where the peak corresponding to the 128\,s period is seen and its two 1-cycle-per-spacecaft-orbit aliases.
The spin period for V2731\,Oph is estimated to be 127.981$\pm$0.022\,s during the NuSTAR observation, consistent with the previous determination in the optical
\citep[e.g., 127.999909(49)\,s;][]{2005MNRAS.361..141G}
and X-rays \citep[e.g., 128.02$\pm$0.02\,s;][]{2008A&A...481..149D}.
The folded light curves are not quite sinusoidal, although
mostly so in the 3--6 keV range (Fig. \ref{fig:v2731_lc}). The red dashed lines in the upper and middle panels are copies of the 3--6 keV spin profile, showing that the main peak remained the same, but a secondary
peak is very pronounced in the 10--30 keV range.
\section{Discussion}
\subsection{Soft X-ray complexity of V2731\,Oph}
\label{sct:v2731complex}
The fact that V2731\,Oph has a complex X-ray emission was already reported by \citet{2008A&A...481..149D}. Our analysis show that their model cannot fit the Suzaku XIS1 spectrum. Moreover, a blackbody temperature of 0.21 keV and 0.18 keV, as derived from Suzaku XIS0+XIS3 and XIS0+XIS1+XIS3 spectra, respectively, is clearly unphysical for WD accretion. Considering the local Eddington limit, the highest effective temperature for a black body contribution would be $\sim$140\,eV for a 1.4\,M$_{\odot}$ WD \citep{1987MNRAS.226..725W}, which would imply a shock temperature $kT_{max}$ of 230\,keV. Even the 94\,eV temperature, as derived from XMM-{\it Newton} data and consistent with Swift observations, demand a WD mass of at least 1.2\,M$_{\odot}$, with a high specific accretion rate of about 200\,g\,cm$^{2}$\,s$^{-1}$ and $k$T$_{max}$\,$>$\,100\,keV, and therefore inconsistent with our spectral analysis.
Based on the Suzaku XIS1 spectrum, we have proposed an alternative spectral model that does not include a hot blackbody. For other instruments, but still acceptable for the Suzaku XIS0 and XIS3 spectra,
this model results in a poorer fit than was obtained with the model with a blackbody. One plausible explanation is the limitation of the {\tt pwab} model. While this model is the most sophisticated in terms of reproducing the likely distribution of the number of lines of sight as a function of N$_H$ \citep{1998MNRAS.298..737D}, it assumes that the absorber has the solar abundance and is not ionized. If the emitting plasma is overabundant in nitrogen, so, too, should the absorbing plasma, which is not currently taken into account. Moreover, the presence of the OVII edge means the latter assumption is invalid. We currently model this using an edge model in addition to {\tt pwab}, but we believe this is not the correct model for this situation. In reality, the absorption below 0.78 keV should be less than what {\tt pwab} predicts, because the complex absorber is ionized and is mostly transparent at these energies. We therefore believe that the current inadequecies of the model is the likely explanation for the imperfect fit. It is likely that high-resolution, high signal-to-noise X-ray spectroscopy combined with an improved grid of models would be needed to see if the alternative spectral description is correct. In the mean time, the abundance of N should be investigated using optical and UV spectroscopy.
Nitrogen overabundance has been seen in a number of CVs, both magnetic
and non-magnetic \citep[see, e.g.,][]{2003ApJ...594..443G,2012ApJS..199....7F}, and indicates the presence of
CNO-processed matter in the system. This, in turn, can be either due
to prior nuclear evolution of the donor \citep{2002MNRAS.337.1105S} or pollution by previous nova eruptions
\citep{1998MNRAS.301..699M}.
For the main purposes of this work, the exercise we carried out by using three sets of X-ray observations for V2731\,Oph shows that: (i) while there is no significant evidence of change in the spectral distribution of the hottest thermal component, its luminosity is variable and its X-ray photons suffer the effect of a variable absorbing column, (ii) the hottest thermal component is in agreement with a cooling flow, stratified multi-thermal model which is expected for IP systems, (iii) soft X-ray spectrum is variable in flux and in spectral distribution.
Given these findings, we believe our results are robust
regarding the complex absorber, reflection amplitude, and the
maximum temperature of the post-shock region, despite our
imperfect understanding of the soft ($<$1\,keV) spectrum of
V2731 Oph.
\subsection{Clues from the plasma temperature and reflection amplitude}
\label{sct:tandrefl}
While the maximum temperature ($k$T$_{max}$) derived from the \textsc{mkcflow} model traces the gravitational potential well at the accretion column shock, the reflection amplitude ($\Omega_{r}$) depends on the shock height ($h$) with respect to the stellar surface. Thus, these two spectral signatures allow us to put constraints on the white dwarf mass and on the accretion geometry of the system.
For a point-like emission region placed at height $h$ above the white dwarf, with $h$ expressed as a fraction of the stellar radius, the stellar surface covers the solid angle \citep[see Figure 2 of][]{2018PASJ...70..109T}:
\begin{equation}
\Omega\,=\,2\pi (1 - \sqrt{1-1/(1+h)^2})
\label{eq:omega}
\end{equation}
The reflection amplitude is defined as the solid angle covered by the reflector as seen from the emitter in such a way that it is 1.0 when the solid angle subtended is 2$\pi$ (half the sky). This condition corresponds to $h$\,=\,0 in Eq. \ref{eq:omega}, and therefore $\Omega_{r}$\,$\equiv$\,$\Omega$/2$\pi$. Thus, a reflection amplitude of 0.88$\pm$0.13 as derived from the spectral fit of TV\,Col (see Table \ref{tbl:spct_tvcol}) requires $h$ of order 0.7\% of the white dwarf radius. Such a small shock height implies that $k$T$_{max}$ traces the gravitational potential well virtually at the WD surface and therefore the white dwarf mass determined from our spectral fit (0.735$\pm$0.015 M$_\odot$) does not require additional corrections.
On the other hand, as first derived by \citet{1973PThPh..49.1184A}, the geometrical parameter $h$ is physically determined by the
post-shock cooling time ($t_{c}$), that defines the remaining time for the accreted material to settle onto the white dwarf. Higher density ($\rho$) in the accretion column accelerates the cooling ($t_{c}$\,$\propto$\,1/$\rho$), resulting in smaller $h$. Thus, $h$ is a function of white dwarf mass and specific (or per unit area) accretion rate (g\,s$^{-1}$\,cm$^{-2}$).
Following \citet{1973PThPh..49.1184A}, the low $h$ derived from the reflection in TV\,Col ($\sim$\,0.7\%) is due to a
high specific accretion rate, of order 15\,g\,s$^{-1}$\,cm$^{-2}$.
Comparing the specific accretion rate with the mass accretion rate derived from the \textsc{mkcflow} model, of (3.8$\pm$0.2)$\times$10$^{-10}$ M$_{\odot}$\,yr$^{-1}$, we estimate that the fractional area over which
accretion happens ($f$) is of about 0.02\% for TV Col: a relatively high specific accretion rate, and a small accretion spot.
The best estimate for the spot size in IPs is that for
XY Ari, for which \citet{1997MNRAS.291...71H} used the X-ray
eclipses to arrive at an estimate of $<$\,0.002. If reliable
determination of reflection amplitude becomes routine, there
is a potential to repeat the kind of analysis we have performed
for TV Col to estimate the spot sizes of IPs. This might allow
future researchers to perform a statistical study against other
system parameters.
On the other hand, we did not detect an unambiguous sign of reflection in the NuSTAR observation of V2731\,Oph. Thus we are unable to derive the specific accretion rate or the shock height in V2731\,Oph from its X-ray spectrum.
\subsection{Additional constraints from the photometry}
X-ray spin modulation of IPs is usually due to complex absorption \citep{1989MNRAS.237..853N}, which is not expected to produce significant spin modulation above 10\,keV. Therefore, when spin modulation is seen above 10\,keV, a geometrical explanation is required. A likely scenario is that non-negligible shock height ($h$\,$>$\,0.1) results in both poles being visible over a range of viewing geometry \citep{2001A&A...377..499D}.
The lack of detectable spin modulation in the NuSTAR data on TV\,Col can be explained as follows. First, we do not expect geometrical spin modulation because the shock in TV\,Col is very close to the stellar surface ($h \sim$0.7\%). Secondly, the maximum energy at which we can detect X-ray spin modulation due to complex absorption) depends on N$_{H,max}$. Our spectral fit shows N$_{H,max}$ of only $\sim$\,10$^{23}$\,cm$^{-2}$ during the NuSTAR observation, so the effects of complex absorption are mostly confined to photon energy below $\sim$3\,keV. The complex absorption appears to be variable from epoch to epoch according to our analysis of the Swift data, which also extends to lower energies: however, these Swift data do not have sufficient spin phase coverage and our timing analysis was inconclusive. We expect that X-ray observations which showed spin modulations above 3 keV were obtained when TV\,Col was in a higher N$_{H,max}$ state.
In V2731\,Oph, N$_{H,max}$ is high, so the low energy ($<$\,10 keV) spin modulation may well have a contribution due to complex absorption. Nevertheless, the clear detection of the spin period above 10\,keV (Fig. \ref{fig:v2731_scargle}) implies a strong geometrical component to the X-ray spin modulation in this IP. This indicates a non-negligible shock height for V2731\,Oph. We thus expect a low reflection amplitude, which can explain the absence of evidence of reflection in the X-ray spectrum of the system. In that sense, V2731\,Oph appears similar to V709\,Cas \citep{2015ApJ...807L..30M}.
At first sight, it might seem surprising that a system as luminous V2731\,Oph should have a non-negligible shock height $h$. We suggest two factors which may explain the large $h$. First, recall that $h$ is determined, not by the total accretion rate, but by the specific accretion rate. We can explain the non-negligible $h$ by invoking a large $f$. A major factor controlling $f$ is the pinching by the dipole magnetic field (B\,$\propto$\,$r^3$), the degree of which is determined by the magnetospheric radius. If we assume approximate spin equilibrium, V2731\,Oph (P$_{spin}$\,=\,128\,s) must have a much smaller magnetosphere than TV\,Col (P$_{spin}$\,=\,1,911\,s), so dipole pinching is much less effective. Secondly, V2731\,Oph may well have a massive WD. The free-fall velocity is higher, which reduces the post-shock density, which reduces the X-ray cooling efficiency. Since the post-shock temperature is also high, this increases $t_c$, and hence $h$.
\subsection{Our targets and the standard evolutionary scenario}
The mass transfer in CVs with orbital period less than $\sim$10 hrs
are driven by angular momentum loss, and is expected to be of order
3-5$\times$10$^{-9}$ M$_{\odot}$\,year$^{-1}$ for a 5.5 hr period CV,
as appropriate for TV\,Col
\citep{2011ApJS..194...28K}. Both the normalization of the \textsc{mkcflow} model
and the bolometric X-ray luminosity of TV\,Col suggests a lower
accretion rate, unless a large amount of transferred material is
not accreted and is lost, perhaps propellered out of the system, therefore reducing the expected accretion rate and consequent X-ray emission. The observed X-ray luminosity is typical of IPs \citep{2018A&A...619A..62S}.
This standard evolutionary scenario does not apply to V2731\,Oph,
and so we have no quantitative prediction for the mass transfer
rate in this system. The BAT band luminosity of V2731\,Oph is already
log\,L$_{X}$(erg\,s$^{-1}$)\,=\,34.6 \citep{2018A&A...619A..62S}, the highest of all IPs,
and the uncertain evolutionary driver must be capable of sustaining
a high mass transfer rate.
The overabundance of N, if confirmed, would place V2731\,Oph among
the subset of CVs known to be overabundant in N
\citep[see, e.g.,][]{2003ApJ...594..443G}. This would imply that the donor matter has been
CNO-processed: the secondary of V2731\,Oph may be evolved in the
normal (single star evolution) sense, or it may indicate that this is
a post thermal-timescale mass-transfer (TTMT) system \citep{2002MNRAS.337.1105S}.
\subsection{White dwarf masses}
\label{sct:wdmass}
We have shown that both a complex absorber and reflection are necessary to fit the wide-band X-ray spectrum of TV\,Col. The combined NuSTAR+Swift data allowed us to constrain both, as well as the maximum temperature of the X-ray emitting plasma.
The high reflection amplitude supporting a small shock height $h$ and the long spin period indicate that the inner radius of the truncated disk is large. Taking the maximum plasma temperature as a measure of the gravitational potential just above the WD surface, we determine the WD mass to be 0.735$\pm$0.015 M$_\odot$. Our value is somewhat smaller than, but consistent with, the previous NuSTAR measurement by \citet{2016ApJ...826..160H}.
On the other hand, we were unable to constrain the reflection amplitude in V2731\,Oph. This has several effects on the mass measurement. First, our derived values for $kT_{max}$ depends on whether we include reflection (with amplitude fixed to 1.0) or not. Also, our values, particularly the case without reflection (0.92$\pm$0.04 M$_\odot$) is significantly lower than what \citet{2016ApJ...826..160H} derived; the reason for this is unclear. Moreover, we were unable to constrain the shock height, but the high energy ($>$10 keV) spin modulation suggests it is not small. Because of this, and because V2731\,Oph is a fast rotator that probably requires correction for small inner radius of the disk, we consider our numbers to be lower limits for the true WD mass in this system.
\section{Summary}
We have combined analyses of broad-band X-ray spectroscopy and photometry, as in \citep{2015ApJ...807L..30M}, to study two IPs, TV\,Col and V2731\,Oph, already analyzed by \citet{2016ApJ...826..160H}. We have gained valuable insights into the geometry and physics of accretion in these systems.
\begin{enumerate}
\item TV\,Col:
\begin{itemize}
\item Complex absorption, such as described by the {\tt pwab} model is required to obtain a good spectral fit.
\item Reflection is also clearly detected, with an amplitude of 0.88$\pm$0.13, indicating a small shock height caused by high specific accretion rate.
\item We do not detect the spin modulation in the NuSTAR data,
showing that the amplitude of the spin modulation can vary over different epochs, as is also seen in at optical wavelengths.
\item The white dwarf mass is determined to be 0.735$\pm$0.015 M$_\odot$.
\end{itemize}
\item V2731\,Oph:
\begin{itemize}
\item The X-ray spectrum is affected by a complex and variable local absorption, with a much higher N$_H$ than for TV\,Col.
\item The spectral model, which includes a hot blackbody, used to fit the XMM-{\it Newton} data does not work for Suzaku XIS1 data. We have tentatively proposed an alternative model without such a blackbody, but with an overabundance of nitrogen.
\item The NuSTAR data did not allow us to constrain the reflection amplitude. The detection of high energy ($>$10\,keV) spin modulation suggests a tall shock.
\item A combination of clues points to a high mass white dwarf in V2731\,Oph.
\end{itemize}
\end{enumerate}
\acknowledgments
We thank the anonymous referee for numerous helpful suggestions.
R.L.O. was partially supported by the Brazilian agency CNPq (Universal Grants 459553/2014-3, PQ 302037/2015-2, and PDE 200289/2017-9).
This work has made use of data from the European Space Agency (ESA) mission
{\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by the {\it Gaia}
Data Processing and Analysis Consortium (DPAC,
\url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for the DPAC
has been provided by national institutions, in particular the institutions
participating in the {\it Gaia} Multilateral Agreement.
This research has made use of the NuSTAR Data Analysis Software (NuSTARDAS) jointly developed by the ASI Science Data Center (ASDC, Italy) and the California Institute of Technology (Caltech, USA). This research has made use of the XRT Data Analysis Software (XRTDAS) developed under the responsibility of the ASI Science Data Center (ASDC), Italy.
\vspace{5mm}
\facilities{NuSTAR, Swift, Suzaku, XMM-{\it Newton}.}
|
\section{Introduction}
One of the most curious and unexpected effects related to the wave nature of quantum particles is the so called Quantum Reflection (QR). It consists in the reflection of a quantum particle that moves under the action of a potential that decreases monotonically in the direction of the particle motion, even without the existence of any turning point. Classically, the particle would suffer a force pushing it in the forward direction. Rather than, increasing its velocity, the particle is reflected by such a decreasing potential. The probability of occurrence of QR increases as the wave nature of the particles becomes more pronounced\cite{DeBroglie}, facilitating QR of particles with low masses at low energies. As a result there is a non-zero probability of an atom at low energies, attracted by a wall due to dispersive forces, to be reflected before reaching the wall. It is worth mentioning that this intriguing phenomenon is not restricted to quantum mechanics. Indeed, it is a general feature of wave propagation in inhomogeneous media and it may occur for mechanical waves or electromagnetic waves in dielectrics and transmission lines\cite{Brekhovskikh-2012}.
The first experiments on QR were performed with helium and hydrogen atoms reflected by liquid helium surfaces \cite{Nayak-1983, Berkhout-1989, Doyle-1991, Yu-1993}. Due to the low mass of atomic specimens in these experiments, QR regime was reached with relatively high energies ($1-10 neV$). Heavier the particle, lower its energy should be in order to reach the QR regime. In references \cite{Pasquini-2004, Pasquini-2006, Marchant-2016} deep QR regimes have been reached for sodium and rubidium atoms prepared in Bose-Einstein states with normal incidence and energies of the order of $10^{-4}neV$. Lower energy regimes can be reached with oblique incidence of the incoming particles \cite{Druzhinina-2003, Zhao-2008}. Since the pioneering paper by Shimizu \cite{Shimizu}, in which an ingenious setup based on QR was developed to investigate the power laws of the non-retarded and retarded dispersive interactions between an atom and a wall, many different approaches have been put forward to probe dispersive forces via QR\cite{Oberst-2005, Zhao-2008, Zhao-2010, Zhao-2013, Barnea-2017}. In this context, developing alternative mechanisms to control and/or tuning the probability of a beam of atoms to be reflected by a wall may open new possibilities for designing new atomic mirrors \cite{Cote-2003, Segev-1997} or even atomic traps \cite{Hammes-2002, Crepin-2017, Jurisch-2008}. Quantum reflection plays an important role in many other areas of physics, which include atom optics \cite{Cronin-RMP-2009, Deutschmann-1993, Landragin-1997, Shimizu-Fujita-2002, Kohno-2003, Savalli-2002} and, more recently, in high precision measurements of the short-range regime of gravitational forces \cite{Perez-2015, Dufour-2014}.
Quantum reflection strongly depends on the Casimir-Polder interaction between the incident particles and the reflecting wall. This interaction may change substantially if one changes the material properties of the wall and the type of incident particles. In this context 2D materials, such as graphene, emerged as good candidates control QR due to their remarkable electromagnetic and mechanical properties. Indeed, it has been shown that interactions mediated by vacuum fluctuations in graphene are highly tunable by varying the chemical-potential \cite{Cysne-2016, Bordag-2016, Henkel-2018}, external magnetic field \cite{Cysne-Kort-Kamp_2014, Macdonald_PRL}, strain\cite{Nichols-2016}, and by stacking of many graphene sheets \cite{Khusnutdinov-2016, Khusnutdinov-2018}. Regarding QR by graphene sheets, some theoretical works do exist~\cite{Judd-2011}, but only very few of them explore the important application of tuning this effect using external parameters, such as strain~\cite{Nichols-2016}. Taking advantage of the remarkable magneto-optical control of the Casimir-Polder interaction between atoms and graphene sheets \cite{Cysne-Kort-Kamp_2014}, we put forward an alternative, realistic method for tuning QR of cold atoms by a graphene sheet by applying a perpendicular magnetic field. We demonstrate that the quantum reflection probability for a given energy as a function of the external magnetic field shows discontinuities, a direct consequence of the structure of the Landau levels of the electronic spectrum of the graphene sheet.
This paper is organized as follows: in the next section we present the methodology to be used in the computation of QR probabilities. However, instead of using power law expressions for the dispersive forces which are valid only in the two opposite regimes of short distance (van der Waals regime) and large distance regime (retarded regime) we use the complete Casimir-Polder potential computed by means of Lifshitz formula \cite{Lifshitz-1956, Milonni}, valid for all distance regimes. In Section III we present our results and show that graphene is, indeed, a very good platform to be used as reflecting material and permits high tunability of QR. We discuss QR for three atoms with different masses (He, Na, Rb) and show that depending on the atom mass its QR probability will be significant for different atom-graphene distance regimes. Section IV is left for conclusions and final remarks.
\section{Methodology}
We consider a system constituted of a beam of atoms moving towards a graphene sheet with normal incidence, with an applied magnetic field, perpendicular to the sheet as shown in Fig.~\ref{Figure1}.
\begin{figure}[h]
\centering
\includegraphics[scale = 0.7]{Boneco.jpg}
\caption{Atom being reflected by the attractive Casimir-Polder force exerted by a graphene sheet before the atom reaches the sheet. Graphene is under the action of an external magnetic field perpendicular to the sheet.}
\label{Figure1}
\end{figure}
The common procedure to calculate the QR probability $R$ with normal incidence of the atoms towards the wall under consideration is to solve a one-dimensional Schr\"odinger equation with the corresponding potential $U(z)$ ($z$ being the distance from the atom to the wall) and appropriate boundary conditions. It is usual in the literature \cite{Friedrich-2002, Segev-1997, Oberst-2005,Arnecke-2006} to consider the van der Waals potential $U_{vdw}(z) = - C_3/z^3$, valid for short distances, or the (asymptotic) retarded potential ${U_{ret}(z) = -C_4/z^4}$, valid for large distances, $C_3$ and $C_4$ being positive constants, or even a simple phenomenological interpolating potential of the form $U_{int}(z) = -C_4/[z^3(z + \ell)]$, where $\ell$ is a parameter with dimensions of length which depends on the incident atom \cite{Judd-2011, Voronin-2005}. A detailed discussion on the comparison of this kind of phenomenological potential and the exact one given by Lifshitz formula can be found in Ref.\cite{Bezerra-2008}. In the present work we do not restrict ourselves to the approximate expressions for the interacting potential between the atom and the wall. Rather, we compute the QR probability by using the full Casimir-Polder potential valid for all distance regimes, which constitutes an important methodological progress with respect to the vast majority of existing theoretical works on QR. In the present case, this potential is the Casimir-Polder potential between a neutral but polarizable atom and a graphene sheet with a magnetic field acting perpendicularly on the sheet at low temperatures, which is given by \cite{Cysne-Kort-Kamp_2014}
\begin{eqnarray}
U(z)&=&\frac{\hbar}{\epsilon_0 c^2}\int_0^{\infty}\frac{d\xi}{2\pi} \xi^2 \,\alpha(i\xi)\int\frac{d^2\textbf{k}}{(2\pi)^2}\frac{e^{-2\kappa z}}{2\kappa} \cr\cr
&& \hspace{-40pt}\times \Bigg[ r^{s,s}(\textbf{k},i\xi, B)- \Bigg(1+\frac{2c^2k^2}{\xi^2}\Bigg)r^{p,p}(\textbf{k},i\xi, B)\Bigg]\, ,
\label{u}
\end{eqnarray}
where ,
$\kappa=\sqrt{\xi^2/c^2 + k^2}$, $\alpha(i\xi)$ is the electric polarizability of the atom, and $r^{s,s}(\textbf{k},i\xi, B)$,
$r^{p,p}(\textbf{k},i\xi, B)$ are the diagonal reflection coefficients associated to graphene ($s$ and $p$ mean, as usual, the transverse electric and transverse magnetic polarizations, respectively). The reflection coefficients of graphene in presence of magnetic field, the model for atomic polarizabilities, and a discussion of the profile of function $U(z)$ can be found in the Appendix.
For an atom of mass $m$ and energy $E$ under the action of a potential $U(z)$, the Schr\"odinger equation reads
\begin{eqnarray}
\frac{\partial^2 \psi (z)}{\partial z^2}+\frac{p^2(z)}{\hbar^2}\psi(z)=0, \label{SE}
\end{eqnarray}
where
\begin{equation}\label{p}
p(z) = \sqrt{2m [E-U(z)]}\, .
\end{equation}
Since the WKB solutions are good approximations when the atom is far from the graphene sheet (compared to the length scale associated with CP interaction, namely, $c/\xi_l$ [see the table \ref{table1} in the appendix]), it is convenient to try a solution of the form \cite{Berry-1972},
\begin{eqnarray}
\psi(z)=\frac{c_+(z)}{|\sqrt{p(z)}|}e^{i\phi(z)}+\frac{c_-(z)}{\sqrt{|p(z)|}}e^{-i\phi(z)}\, , \label{WKB}
\end{eqnarray}
with $\phi(z)$ given by
\begin{eqnarray}
\phi(z)=\int^{z}_{z_0}\frac{p(z')}{\hbar}dz' \, . \label{WKBphi}
\end{eqnarray}
Note that there is no approximation in writing previous equations, since $c_+(z)$ and $c_-(z)$ still need to be determined. However, the previous ansatz is a very convenient one, since it transforms the second order Schr\"odinger equation into a set of two coupled first order differential equations for $c_+(z)$ and $c_-(z)$. In fact, substituting Eq(s) (\ref{WKB}) and (\ref{WKBphi}) into Eq.(\ref{SE}) it is straightforward to show that
\begin{eqnarray}\label{DE1}
\frac{\partial c_{+}(z)}{\partial z} &=& e^{- 2 i \phi (z)}\frac{c_-(z)}{2p(z)}\frac{\partial p(z)}{\partial z}\\\cr
\frac{\partial c_-(z)}{\partial z} &=& e^{+2 i \phi (z)}\frac{c_+(z)}{2p(z)}\frac{\partial p(z)}{\partial z}\, .
\label{DE2}
\end{eqnarray}
Reagarding boundary conditions, it is reasonable to impose that $c_+(0) = 0$ and $c_-(0) = 1 $, which means that any atom that reaches the graphene sheet will not be reflected, but rather adsorbed to it. By definition, the quantum reflection probability is \cite{Dufour-2013},
\begin{eqnarray}
R=\lim_{z\rightarrow \infty}\Big| \frac{c_+(z)}{c_-(z)}\Big|^2\, . \label{QRP}
\end{eqnarray}
The previous set of coupled first order differential equations for $c_+(z)$ and $c_-(z)$ will be solved numerically. Information about the efficiency of a certain potential $U(z)$ to give rise to QR can be extracted from the function \cite{Dufour-2013,Dufour-2-2013},
\begin{eqnarray}
Q(z)=\frac{\hbar^2}{2p^2(z)}\Bigg[ \frac{\phi '''(z)}{\phi'(z)}-\frac{3}{2}\Bigg(\frac{\phi'' (z)}{\phi'(z)}\Bigg)^2\Bigg]\, , \label{QF}
\end{eqnarray}
with $p(z)$ and $\phi(z)$ given by (\ref{p}) and (\ref{WKBphi}). It can be shown that the highest probabilities of occurrence of QR correspond to the regions of highest values of $Q(z)$ \cite{Dufour-2013, Dufour-2-2013, Dufour-2015-tesis}. Moreover, for a given energy, this function exhibits a peak. Let us denote by $z_m$ the position of this peak, that is, the distance between the atom and the graphene sheet where $Q(z)$ is maximum. In order to solve numerically the set of coupled differential equations (\ref{DE1}) and (\ref{DE2}), we choose a point close to graphene sheet ($z_i$) and a point far from graphene ($z_f$) such that $z_i \ll z_m \ll z_f$. The differential equations are solved in the region between $z_i$ and $z_f$ which contains the peak of $Q(z)$. The limit in Eq. (\ref{QRP}) is numerically calculated by taking values of coefficients $c_+(z)$ and $c_-(z)$ at point $z_f$. Parameters $z_f$ and $z_i$ are convergence parameters. We have numerically established that a good convergence of the results occurs whenever changes in $z_i$ and $z_f$ do not affect any more the value of the QR probability $R$. A detailed discussion of this method can be found in reference \cite{Dufour-2015-tesis}.
As already mentioned, in most calculations of QR, an interpolation between the non-retarded van der Waals potential ($\approx - C_3/z^3$) and the retarded potential ($\approx - C_4/z^4$) regimes is used, allowing for a semi-analytical solution of the coupled differential equations for the WKB coefficients $c_+(z)$ and $c_-(z)$. In the present work we solve the coupled equations (\ref{DE1}) and (\ref{DE2}) by using the complete Casimir-Polder potential whose dependence with distance is quite involved, so that a full numerical procedure is demanded. The consideration of the whole potential is crucial for obtaining reliable results on the control of QR on graphene with the aid of an external magnetic field, and constitutes an important methodological advance with respect to the traditional theoretical approach to QR. In the next section we present the main results of this work obtained by numerical calculations. As we shall see, the complex behavior of the complete Casimir-Polder potential with the external magnetic field has non trivial consequences on the QR of different atomic specimens. It is worth mentioning that the Zeeman coupling in graphene can be safely neglected due to the more pronounced effect of Landau quantization of electronic motion, related to the relativistic nature of the Dirac spectrum \cite{Gusynin-2005-1, Gusynin-2005-2}. In addition, we have verified that the Zeeman effect upon the atoms only leads to negligible corrections in the computation of the dispersive energy interaction between graphene and the atoms, for all species considered, as previously shown in Ref.~\cite{Cysne-Kort-Kamp_2014}.
\section{Results and discussions}
In order to investigate the consequences of the application of a magnetic field perpendicular to graphene sheet in the probability of an atom to be reflected by this sheet, we consider three atomic species of experimental relevance: He, Na and Rb. These atomic species have already been used in QR experiments\cite{Nayak-1983, Berkhout-1989, Doyle-1991, Yu-1993, Pasquini-2004, Pasquini-2006, Marchant-2016}. The atomic polarizabilities of these atoms will be used in the Casimir-Polder potential [Eq. (\ref{u})] and are described by the single Lorentz-oscillator model, whose fitted parameters are well known from the literature \cite{Khusnutdinov-2016} [see Appendix]. Once we would like our results to be valid regardless the retard/non-retard regime, we numerically evaluate the Casimir-Polder potential in the distance range $z_i< z< z_f$ at given values of the magnetic field [see Appendix]. In order to solve numerically the system of coupled differential equations (\ref{DE1}) and (\ref{DE2}), we use a standard procedure of interpolation by a polynomial function of successive points of the CP potential and, finally, we may obtain the QR reflection probability $R$ using Eq. (\ref{QRP}).
\begin{figure}[h]
\vspace{0.1in}
\centering
\includegraphics[scale = 0.55]{Figura_2.png}
\caption{Quantum reflection probability $R$ as a function of the energy of the incident particle for (a) He, (b) Na and (c) Rb at given values of the external magnetic field.
We considered the chemical potential of the graphene sheet fixed at $\mu_c=0.115$eV.}
\label{Figure2}
\end{figure}
In Fig. \ref{Figure2} we plot the QR probabilities $R$ of He, Na and Rb as functions of their incident energies for four values of the magnetic field ($B=0,2,7,14$T). The chemical potential of graphene sheet is set to $\mu_c=0.115$eV. Note that, in all cases, $R\rightarrow 1$ as $E \rightarrow 0$, which is a direct consequence of the fact that the de Broglie wavelength associated to the atom increases as $E$ decreases, i.e. the wave-like nature of the particles becomes more important the lower their energy is. By the same token, $R\rightarrow 0$ as $E \rightarrow \infty$, as expected, since the de Broglie wavelength decreases as $E$ increases. The limit where $\lambda \rightarrow 0$ is analogous to the geometrical optics limit in wave optics and hence the particle behavior of the atom must show up. In between these two regimes, there is an intermediate one in which QR is substantially influenced by the external magnetic field. This is one of the main results of our paper, showing that a magnetic field applied to a graphene sheet may be used as an efficiently way of controlling and tuning QR by an external agent. Moreover, this influence depends on the atom under consideration. For He atom [Fig. \ref{Figure2}, panel (a)] the QR probability at a given energy decreases when the magnetic field is applied ($B=2, 7, 14$T curves). On other hand, for Rb atom [Fig. \ref{Figure2}, panel (c)] the QR probability is enhanced in presence of magnetic field ($B=2,7,14$T). For Na [panel (b) of Fig. \ref{Figure2}], there is a non-monotonic behaviour with B, i. e. QR decreases for $B=2$T and increases for $B=7, 14$T when compared to the case of $B=0$ T.
In order to understand these different behaviors of the QR probability for different atoms, it is necessary to analyze the interplay between the tunability of CP energy with the magnetic field at different distance regimes and the function $Q(z)$. The position of the peak of $Q(z)$ indicates the most probable region of space where QR can occur. To this end, we show in the Figs. \ref{Figure3}, \ref{Figure4}, and \ref{Figure5} the dependence of QR probability [panel (a)], $Q(z)$ function [panel (b)] and the relative change of CP energy with magnetic field [panel (c)] for He, Na and Rb atoms, respectively. In the case of He atom, we show in panel (a) of Fig. \ref{Figure3} the quantum reflection probability as a function of the applied magnetic field for three different values of the chemical potential of the graphene sheet. The incident energy of the atom was fixed at a value that leads to optimal tunability in a magnetic field, as it can be seen in Fig. \ref{Figure2} ($E=10^1$neV for the case of He in Fig. \ref{Figure3}, panels (a) and (b)). The QR probability as a function of the magnetic field for He atom shows a general decrescent behavior, consistent with the discussion of the previous paragraph. In addition to this decrescent trend, there exist discontinuities that are related to discontinuities in the CP energy whenever a Landau level of the spectrum of graphene (which varies with B) crosses the chemical potential \cite{Cysne-Kort-Kamp_2014}. The effect of varying the chemical potential is just a shift in values of the magnetic field where these discontinuities take place. In panel (b) of Fig. \ref{Figure3}, we show the function $Q(z)$ for chemical potential $\mu_c = 0.115$ eV, incident energy of He atom $E=10^1$ neV and four values of the external magnetic field. The maximum of $Q(z)$ function occurs at distances from the graphene sheet where the CP energy is enhanced by the external magnetic field [As it can be seen from the panel (c) of Fig. \ref{Figure3}]. This enhancement of CP energy with the magnetic field in the region of space where QR takes place is related to the decrescent behavior of the QR probability with the magnetic field for the He atom discussed in the previous paragraph. In the case of Rb, in the panel (a) of the Fig. \ref{Figure5}, we show the QR probability as a function of the external magnetic field for three values of the chemical potential. The incident energy of the Rb atom was set to correspond to the optimal tunability in a magnetic field on Fig. \ref{Figure2} ,{\it i. e.}, $E=10^{-5}$neV for panels (a) and (b) in Fig. \ref{Figure5}. Again, the QR probability as a function of the magnetic field shows discontinuities associated with the crossing of Landau levels through the chemical potential in the electronic spectrum of the graphene sheet. However, here QR probability has a crescent behavior. In the panel (b) of the figure, it is possible to see that the maximum of the function $Q(z)$ occurs in the region of space where the CP energy decreases with the presence of the magnetic field. This explains the different behaviors of QR probability with magnetic field for different atoms. In the case of Rb atom, QR takes place in a region where the CP energy decreases with the magnetic field, so that the QR probability is enhanced. Furthermore, in between the discontinuities at QR probability for Rb [panel (a) Fig. \ref{Figure5}], there exist plateaus. This behavior is a direct consequence of the plateaus in the CP energy as a function of the magnetic field at long-distance regimes (where QR takes place), which occur due the prevalence of low-frequency modes in Lifshitz formula \cite{Cysne-Kort-Kamp_2014}. Finally, in the case of Na (shown in Fig. \ref{Figure4}), the QR takes place at intermediate distances, i. e., in the region between the short distance regime, where CP energy increases with the magnetic field, and the long-distance regime, where CP energy decreases with the magnetic field. This explains the non-monotonic dependence of QR probability with the magnetic field for Na atom discussed in the previous paragraph and shown in the panel (b) of Fig.\ref{Figure2}.
\begin{figure}[h]
\vspace{0.1in}
\centering
\includegraphics[scale = 0.55]{Figura_3_He.png}
\caption{For He atom, Panel (a): QR probability as a function of the external magnetic field for chemical potentials of the graphene sheet $\mu_c=0.105$ eV (Red Dashed), $\mu_c=0.115$ eV (Blue Solid), $\mu_c=0.125$ eV (Green Dot-Dashed). Panel (b): Function $Q(z)$ of Eq. (\ref{QF}) for chemical potential of the graphene sheet $\mu_c=0.115$ eV and external magnetic field intensities $B=0, 2, 7, 14$ T. For all data at the panels (a) and (b) we set the energy of the incident particle $E=10^1$ neV. Panel (c): Relative variation of CP energy with the magnetic field as a function of distance ($U^B(z)/U^{B=0}(z)$) for the chemical potential of the graphene sheet $\mu_c=0.115$eV external magnetic field intensities $B=0, 2, 7, 14$ T.}
\label{Figure3}
\end{figure}
\begin{figure}[h]
\vspace{0.1in}
\centering
\includegraphics[scale = 0.55]{Figura_4_Na.png}
\caption{For Na atom, Panel (a): QR probability as a function of the external magnetic field for chemical potentials of the graphene sheet $\mu_c=0.105$ eV (Red Dashed), $\mu_c=0.115$ eV (Blue Solid), $\mu_c=0.125$ eV (Green Dot-Dashed). Panel (b): Function $Q(z)$ of Eq. (\ref{QF}) for chemical potential of the graphene sheet $\mu_c=0.115$ eV and external magnetic field intensities $B=0, 2, 7, 14$ T. For all data at the panels (a) and (b) we set the energy of the incident particle $E=10^{-3}$ neV. Panel (c): Relative variation of CP energy with the magnetic field as a function of distance ($U^B(z)/U^{B=0}(z)$) for the chemical potential of the graphene sheet $\mu_c=0.115$eV external magnetic field intensities $B=0, 2, 7, 14$ T.}
\label{Figure4}
\end{figure}
\begin{figure}[h]
\vspace{0.1in}
\centering
\includegraphics[scale = 0.55]{Figura_5_Rb.png}
\caption{For Rb atom, Panel (a): QR probability as a function of the external magnetic field for chemical potentials of the graphene sheet $\mu_c=0.105$ eV (Red Dashed), $\mu_c=0.115$ eV (Blue Solid), $\mu_c=0.125$ eV (Green Dot-Dashed). Panel (b): Function $Q(z)$ of Eq. (\ref{QF}) for chemical potential of the graphene sheet $\mu_c=0.115$ eV and external magnetic field intensities $B=0, 2, 7, 14$ T. For all data at the panels (a) and (b) we set the energy of the incident particle $E=10^{-5}$ neV. Panel (c): Relative variation of CP energy with the magnetic field as a function of distance ($U^B(z)/U^{B=0}(z)$) for the chemical potential of the graphene sheet $\mu_c=0.115$eV and external magnetic field intensities $B=0, 2, 7, 14$ T.}
\label{Figure5}
\end{figure}
All results presented in this paper have been derived by the Casimir-Polder interaction computed from Lifshitz formula at zero temperature [Eq. (\ref{u})]. The detailed study of the effect of finite temperature in quantum reflection is beyond the scope of the present work, but we shall briefly comment on the thermal effects on our results. It is well known that thermal effects in dispersive forces, which have been investigated in {\it e.g. } Ref. ~\cite{Bhuman-Schell}, change the long-distance regime of the Casimir-Polder interaction. This distance regime is characterized by the thermal wavelength, which is proportional to $(k_B T)^{-1}$ and thermal corrections may influence the quantum reflection quantitatively \cite{Bezerra-2008}. These corrections should be more pronounced for heavy atoms, for which quantum reflection is dominated by long-distance regimes of the Casimir-Polder potential (see Appendix). Moreover, finite temperature effects cause a broadening in the Fermi-Dirac distribution which, in turn, makes the discontinuities in the Casimir-Polder potential as a function of the magnetic field smear out. However, thermal effects do not prevent QR to be tuned by an magnetic field, since the Casimir-Polder potential at finite temperature is still substantially changed by a magnetic field \cite{Cysne-Kort-Kamp_2014}. Therefore, we expect that even at high temperatures, quantum reflection by a graphene sheet may still be tuned by a magnetic field, although the discontinuities present in panels (a) of Figs. \ref{Figure3},\ref{Figure4}, and \ref{Figure5} will disappear. The discontinuities in Figs \ref{Figure3}, \ref{Figure4} and \ref{Figure5} must be observable for small temperatures (thermal energy $k_B T$) compared to the energy spacing between two successive Landau levels. These energy spacings become larger for high magnetic fields. This fact can be used in order to reach the regime where discontinuities in quantum reflection probability should be observable.
\section{Conclusions}
In summary, we have demonstrated that an external magnetic field may be used to tune Quantum Reflection (QR) in a graphene sheet. We have calculated the attractive Casimir-Polder potential using Lifshitz formula, valid at any distance between the atom and the graphene sheet. We have considered three atomic species of experimental relevance He, Na, Rb, for which we calculate the QR probability using a full numerical approach. We identify three distinct different behaviors in the QR probability, depending on the magnitude of the applied magnetic field, which can be explained in terms of the different characteristic distance regimes of the Casimir-Polder potential. We also conclude that the effects of the magnetic field on QR are more pronounced for lighter atomic species at the short distance regime of the Casimir-Polder, while for heavier atoms the magnetic effects tend to be more intense in the long-distance regime. However, in all cases and for all the investigated atoms, QR in graphene can be efficiency tuned by applying an external magnetic field. We show that the QR probability exhibits discontinuities as a result of the quantization of the electronic spectrum of graphene. These discontinuities persist at low temperatures and for high values of the magnetic field. Altogether our findings not only allow for an alternative way to control quantum reflection at the nanoscale, but also open the door for the design of novel reflective optical elements, such as Fresnel mirrors~\cite{judd2010}, which can be used for tunable reflective focusing of matter waves.
\section*{Acknowledgements}
The authors thank the Brazilian Agencies CNPq, CAPES, and FAPERJ for their financial support.
|
\section{Introduction}
Vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) communications, also known as V2X communications together, have become a hot research topic especially since the emergence of the fifth-generation (5G) communication systems \cite{MBoban_VTMag}. With advent of autonomous driverless vehicles, transmission units mounted on vehicles will need to exchange massive amount of information signals including speed, traffic condition, direction, location, traffic incidents etc. at the frequency of ten times or even more every second through high-speed wireless links. For example, when a car pulls an emergency breaking due to an unexpected emergency situation, the signal should be transmitted to surrounding vehicles to make them aware of potential hazard buildup in the vicinity. With the deployment of 5G wireless communication technology, it is no surprise that the vehicles can communicate with other communication devices, such as mobile phones and smart computers or even facilities like building, traffic lights and road and so on, which will enhance the proliferation and security of autonomous driving in the future \cite{6g_vision, what_will_5g_be}.
The V2X communication systems feature several unique characteristics as opposed to conventional cellular communications including high mobility, rapid change of direction as well as location, and stringent quality-of-service (QoS) requirements. Furthermore, road safety concerns impose very strict requirements on ultra-low latency and high reliability in V2X communications. However, due to high mobility, the V2X communication channel state information (CSI) becomes outdated quite rapidly. Thus faster optimization techniques are of paramount interest for reliable V2X communications.
Furthermore, in order to improve the reliability of V2X communications, proper interference management as well as resource allocation strategies must be in place for the V2V and V2I links. Traditional interference and resource management schemes for V2X interference-limited scenarios operate in iterative manner, which has high computational complexity \cite{d2d_v2v, d2d_v2x}. While these conventional approaches provide a good understanding of the problem domain and reveals valuable insights of the V2X systems, these are unsuitable for most practical V2X systems requiring ultra-reliable low-latency communications (URLLC) \cite{what_will_5g_be, 6g_vision}.
Recently, machine learning approaches have gained momentum in the wireless communications domain due to their inherent capability of efficiently dealing with large-scale problems. Another reason for the boom of machine learning across a wide range of application domains is that machine learning approaches can combine learning process with existing field technologies. Through appropriate training process, the knowledge can generally decide on a particular hypothesis class. Wireless channel estimation and resource allocation problems are potential examples where machine learning are increasingly exploited \cite{jiang2016machine}. Of particular interest are deep learning techniques enabled by deep neural networks (DNN) due to their reduced computation time. Once trained properly, DNN can provide real-time resource allocation solutions, which is very crucial for V2X communications \cite{zhang2019deep}.
In \cite{drl_v2v_conf, drl_v2x_jrnl}, the authors have extended the works in \cite{d2d_v2x} by solving the same resource allocation problem using deep reinforcement learning (DRL) technique. In particular, the authors in \cite{drl_v2v_conf, drl_v2x_jrnl} developed a decentralized resource allocation technique for V2V communications based of DRL technique. However, with cloud radio access network (C-RAN) being an inherent part of 5G, it is generally expected that radio resource management strategies will be implemented at the centralized cloud \cite{what_will_5g_be}. The main advantage of C-RAN is that with access to enormous information in the cloud, radio resources can be allocated more efficiently.
In this paper, we propose a centralized resource allocation strategy using deep learning technique for V2X communications, as opposed to the decentralized approaches in \cite{drl_v2v_conf, drl_v2x_jrnl}. In particular, we consider a sum rate maximization problem under individual power constraints for each V2I and V2V communication link using supervised learning technique. The main contributions in this paper are listed below:
\begin{itemize}
\item a) We first develop an iterative power allocation algorithm for the proposed V2X communications system.
\item b) We then propose a DNN based power allocation scheme for the V2V and the V2I links. We apply the mini-batch gradient descent (MBGD) algorithm by identifying the suitable batch size and learning rate through cross validation.
\item c) We have validated the proposed learning algorithm through extensive simulations that demonstrate the suitability of the algorithm particularly for V2X communications.
\end{itemize}
\begin{figure}[ht!]
\includegraphics[width=\linewidth]{v2x_sys_mod.eps}
\caption{The proposed V2V and V2I communication system.} \label{sys_mod}
\end{figure}
\section{System Model}\label{sec_vision}
Let us consider a high-mobility vehicular communication system (cf. Fig.~\ref{sys_mod}) in which $K$ pairs of vehicular nodes share high-capacity V2I communication links with $M$ cellular users (CUs) in a 6G communication system. For the ease of exposition, we consider single-antenna communication nodes. We also assume that all vehicles may accommodate both CUs and VUs simultaneously and are capable of radio transmissions from separate antennas. The CUs communicate with the roadside infrastructure which may be a cellular base station (BS), traffic lamppost, building or any other fixed structure. We also assume that the vehicles travel at a constant speed along the roads and data is transmitted at each time slot using the same carrier frequency.
The channel power gain between the $m$th CU and the BS is defined as \cite{d2d_v2x}
\begin{align}
h_{m}^{\rm c} = G_{m}^{\rm c}\beta_{m}^{\rm c}Ad_{m}^{{\rm c}^{-\gamma}}, \label{ch_gain}
\end{align}
where $G_{m}^{\rm c}$ is assumed to be exponentially distributed fast fading power gain, $\beta_{m}^{\rm c}$ is the log normal shadow fading component, $A$ is the constant pathloss, $d_{m}^{\rm c}$ is the distance between the $m$th CU and the BS, and $\gamma$ is the pathloss exponent. Similarly, we can define the channel $h_{k}^{\rm v}$ between the $k$th V2V transmitter and the BS as well as the $k$th V2V link $g_{k}^{\rm v}$. Thus the signal received at the BS from the $m$th CU can be expressed as
\begin{multline}
y_m^{\rm c} = \underbrace{h_{m}^{\rm c}s_m^{\rm c}}_{\text{desired signal}} + \underbrace{\sum_{n=1, n \ne m}^{M} h_{n}^{\rm c} s_n^{\rm c}}_{\text{interference from the CUs}} + \!\!\! \underbrace{\sum_{k=1}^{K} h_{k}^{\rm v}s_k^{\rm v}}_{\text{interference from the VUs}}\\ + \underbrace{n_m}_{\text{noise}}, \label{rx_sig}
\end{multline}
where $s_i^{\rm x} \sim \mathcal{CN}(0,P_i^{\rm x}),~ x \in \{\rm c, v\},$ is the signal transmitted by the $i$th transmitter and $n_m \sim \mathcal{CN}(0,\sigma^2)$ is the additive white Gaussian noise component at the BS.
Thus the received signal to noise plus interference ratio (SINR) at the BS for the $m$th CU can be expressed as
\begin{equation}
\gamma_m^{\rm c} = \frac{P_m^{\rm c}|h_{m}^{\rm c}|^2}{\sum_{n=1,n\ne m}^M P_n^{\rm c}|h_{n}^{\rm c}|^2 + \sum_{k=1}^K P_k^{\rm v}|h_{k}^{\rm v}|^2 + \sigma^2}. \label{sinr_c}
\end{equation}
Similarly, the received SINR in the $k$th V2V link is given by
\begin{equation}
\gamma_k^{\rm v} = \frac{P_k^{\rm v}|g_{k}^{\rm v}|^2}{\sum_{l=1,l\ne k}^K P_l^{\rm v}|g_{k}^{\rm v}|^2 + \sum_{m=1}^M P_m^{\rm c}|h_{m,k}|^2 + \sigma^2}, \label{sinr_v}
\end{equation}
where $h_{m,k}$ is the interfering channel from V2I user $m$ to V2V receiver $k$.
\subsection{Problem Formulation}
Our aim is to maximize the overall system throughput of the V2X links by optimally allocating transmit power among all the users. Accordingly, we formulate the following optimization problem:
\begin{subequations}\label{prob_msr0}
\begin{align}
\max_{\{P_m^{\rm c}\},\{P_k^{\rm v}\}} ~&~ \sum_{m=1}^M \alpha_m \log\left(1 + \gamma_m^{\rm c} \right) + \sum_{k=1}^K \alpha_k \log\left(1 + \gamma_k^{\rm v} \right) \label{prob_msr0_o}\\
{\rm s.t.} ~&~ 0 \le P_m^{\rm c} \le P_{\rm max}^{\rm c}, \forall m, \label{prob_msr0_c1}\\
~&~ 0 \le P_k^{\rm v} \le P_{\rm max}^{\rm v}, \forall k, \label{prob_msr0_c2}\\
~&~ \alpha_m, ~ \alpha_k \ge 0, \forall m, \forall k,\label{prob_msr0_c3}
\end{align}
\end{subequations}
where $P_{\rm max}^{\rm c}$ and $P_{\rm max}^{\rm v}$ are the maximum transmit power budget of the CUs and the VUs, respectively, $\{\alpha_m\}, ~ \{\alpha_k\}$ are the weights, which determine the priority of the corresponding V2I and V2V links. Note that the problem \eqref{prob_msr0} is non-convex and hence the exactly optimal solution is non-trivial mainly due to the $\log(\cdot)$ in the objective. In the following, we will first develop an acceptable solution to the problem following traditional alternating approaches, and then following machine learning techniques.
\section{The Weighted MMSE Algorithm}\label{sec_wmmse}
In this section, we propose an alternative approach of solving problem \eqref{prob_msr0}. Instead of solving problem \eqref{prob_msr0} directly, we solve the equivalent WMMSE minimization problem, inspired by \cite{wmmse}. The equivalence is guaranteed by the well-known MMSE-SINR equality as proved in \cite{palomar_majorization}. Note that here we downscale the WMMSE algorithm originally proposed for MIMO systems in \cite{wmmse} to the equivalent SISO channels for convenience.
Since neural networks work only with real numbers, the absolute value terms in \eqref{sinr_c} and \eqref{sinr_v} representing channel power gains tend to be beneficial in the neural network based design in Section~\ref{sec_dnn}. However, in the conventional WMMSE algorithm, real channel power gains are also convenient for mathematical operations. Inspired by \cite{wmmse}, we assume that $v_i^{\rm x},~ x \in \{\rm c, v\},$ is the amplifier gain used to transmit signal by the $i$th transmitter, while $u_i^{\rm x}$ is the receiver amplifier gain used to estimate the desired signal. Accordingly, the estimated real symbol $\hat{s}_m^{\rm c}$ is given by \cite{kay}
\begin{multline}
\hat{s}_m^{\rm c} = u_m^{\rm c} \left(|h_{m}^{\rm c}|v_m^{\rm c}s_m^{\rm c} + \sum_{n\ne m}^M|h_{n}^{\rm c}|v_n^{\rm c} s_n^{\rm c} \right.\\
\left.+ \sum_{k = 1}^K|h_{k}^{\rm v}|v_k^{\rm v} s_k^{\rm v} + n_m \right), \quad m = 1, \cdots, M. \label{s_hat}
\end{multline}
Thus the MSE of estimating $s_m^{\rm c}$ is given by \cite{wmmse, jrnl_mur1}
\begin{multline}
\varepsilon_m^{\rm c} = {\rm E}\left(\hat{s}_m^{\rm c} - s_m^{\rm c}\right)^2\\
= \left|u_m^{\rm c} h_{m}^{\rm c}v_m^{\rm c} - 1\right|^2 + \sum_{n\ne m}^M\left|u_m^{\rm c}h_{n}^{\rm c}v_n^{\rm c} \right|^2\\
+ \sum_{k = 1}^K\left|u_m^{\rm c}h_{k}^{\rm v}v_k^{\rm v} \right|^2 + \sigma^2 |u_m^{\rm c}|^2, \quad m = 1, \cdots, M, \label{mse1}
\end{multline}
where ${\rm E}(\cdot)$ indicates the statistical expectation operation. Similarly, the MSE of estimating $s_k^{\rm v}$ is given by
\begin{multline}
\varepsilon_m^{\rm v} = \left|u_k^{\rm v} h_{k}^{\rm v}v_k^{\rm v} - 1\right|^2 + \sum_{l\ne k}^K\left|u_k^{\rm v}h_{l}^{\rm v}v_l^{\rm v} \right|^2\\
+ \sum_{m = 1}^M\left|u_k^{\rm v}h_{m}^{\rm c}v_m^{\rm c} \right|^2 + \sigma^2 |u_k^{\rm v}|^2, \quad k = 1, \cdots, K. \label{mse1_v}
\end{multline}
Thus following the MMSE-SINR equality $\varepsilon_m = \frac{1}{1 + \gamma_m^{\rm c}}$ derived in \cite{palomar_majorization}, the weighted sum rate (WSR) maximization problem \eqref{prob_msr0} can be equivalently expressed as the following WMMSE minimization problem
\begin{subequations}\label{prob_mmse1}
\begin{align}
\min_{\{u_i^{\rm x}\},\{v_i^{\rm x}\},\{w_i^{\rm x}\}} ~&~ \sum_{x \in \{\rm c, v\}}\sum_i \left(w_i^{\rm x}\varepsilon_i^{\rm x} - \log(w_i^{\rm x}) \right) \label{prob_mmse1_o}\\
{\rm s.t.} ~&~ 0 \le v_m^{\rm c} \le \sqrt{P_{\rm max}^{\rm c}}, \forall m, \label{prob_mmse1_c1}\\
~&~ 0 \le v_k^{\rm v} \le \sqrt{P_{\rm max}^{\rm v}}, \forall k, \label{prob_mmse1_c2}\\
~&~ \alpha_m, ~ \alpha_k \ge 0, \forall m, \forall k,\label{prob_mmse1_c3}
\end{align}
\end{subequations}
where $w_i^{\rm x}$ is a positive weight factor.
The WMMSE minimization problem \eqref{prob_mmse1} can be solved using block coordinate descent (BCD) method \cite{boyd}. In each phase, the BCD method optimizes one set of variables while fixing the rest. Checking the first optimality condition, one can find the optimal weight $w_i^{\rm x} = \frac{1}{\varepsilon_i^{\rm x}}$ in closed form and the optimal receiver gain as the well-known Wiener filter \cite{icspcs10}:
\begin{align}
u_i^{\rm x} = \frac{h_{i}^{\rm x}v_i^{\rm x}}{\sum_{m=1}^M|h_{m}^{\rm c}|^2|v_m^{\rm c}|^2 + \sum_{k=1}^K|h_{k}^{\rm v}|^2|v_k^{\rm v}|^2 + \sigma^2}.\label{rx_gain}
\end{align}
In order to obtain the optimal transmitter gain $v_i^{\rm x}, ~ x \in \{\rm c, v\}$, we can decouple problem \eqref{prob_mmse1} for each transmitter, fixing the other variables, as
\begin{subequations}\label{prob_mmse2}
\begin{align}
\min_{\{v_i^{\rm x}\}} ~&~ \sum_i \left(w_i^{\rm x}\varepsilon_i^{\rm x} - \log(w_i^{\rm x}) \right) \label{prob_mmse2_o}\\
{\rm s.t.} ~&~ 0 \le v_i^{\rm x} \le \sqrt{P_{\rm max}^{\rm x}}. \label{prob_mmse2_c1}
\end{align}
\end{subequations}
Applying the Lagrangian multiplier approach \cite{boyd}, problem \eqref{prob_mmse2} can be solved for $v_i^{\rm x}$. Considering the first-order optimality condition from the Lagrangian approach, we obtain the optimal $v_i^{\rm x}$ as
\begin{align}
v_i^{\rm x} = \frac{\alpha_i^{\rm x}w_i^{\rm x}u_i^{\rm x}|h_i^{\rm x}|}{\sum_{x \in \{\rm c, v\}}\sum_i\alpha_i^{\rm x}w_i^{\rm x}|u_i^{\rm x}|^2|h_{i}^{\rm x}|^2 + \mu_i^{\rm x}\sigma^2}.\label{tx_gain}
\end{align}
Here $\mu_i^{\rm x} \ge 0$ is the Lagrange multiplier, which should chosen such that the complementary slackness condition on the power constraint \eqref{prob_mmse2_c1} is satisfied. The overall WMMSE procedure of solving problem \eqref{prob_mmse1} is summarized in Algorithm~\ref{alg_conv}.
It has been proven in \cite{d2d_v2x} that the WMMSE algorithm eventually converges to a stationary point.
\begin{algorithm} [ht]
\caption{WMMSE algorithm for solving problem \eqref{prob_mmse1}}\label{alg_conv}
\begin{algorithmic}[1]
\STATE Initialize $v_i^{\rm x}(0), ~ x \in \{\rm c, v\}$ such that $0 \le v_i^{\rm x} \le \sqrt{P_{\rm max}^{\rm x}}$.
\STATE Compute $u_i^{\rm x}(0) = \frac{h_{i}^{\rm x}v_i^{\rm x}(0)}{\sum_{x \in \{\rm c, v\}}\sum_i|h_{i}^{\rm x}|^2|v_i^{\rm x}(0)|^2 + \sigma^2}$.
\STATE Compute $w_i^{\rm x}(0) = \left[1 - u_i^{\rm x}(0)|h_{i}^{\rm x}||v_i^{\rm x}(0)|\right]^{-1}$.
\STATE Set $n:=0$.
\STATE \textbf{repeat}
\STATE $v_i^{\rm x}(n+1) \leftarrow \frac{\alpha_i^{\rm x}w_i^{\rm x}(n)u_i^{\rm x}(n)|h_i^{\rm x}|}{\sum_{x \in \{\rm c, v\}}\sum_i\alpha_i^{\rm x}w_i^{\rm x}(n)|u_i^{\rm x}(n)|^2|h_{i}^{\rm x}|^2}$.
\STATE $u_i^{\rm x}(n+1) \leftarrow \frac{h_{i}^{\rm x}v_i^{\rm x}(n+1)}{\sum_{x \in \{\rm c, v\}}\sum_i|h_{i}^{\rm x}|^2|v_i^{\rm x}(n+1)|^2 + \sigma^2}$.
\STATE $w_i^{\rm x}(n+1) \leftarrow \left[1 - u_i^{\rm x}(n+1)|h_{i}^{\rm x}||v_i^{\rm x}(n+1)|\right]^{-1}$.
\STATE $n:= n + 1$.
\STATE \textbf{until} convergence.
\STATE \textbf{Output:} Optimal power profile $P_i^{\rm x} = v_i^{\rm x}(n))^2$.
\end{algorithmic}
\end{algorithm}
\section{Proposed Machine Learning Approach}\label{sec_dnn}
It has been shown in \cite{drl_v2v_conf, drl_v2x_jrnl} that a multi-layer neural network (MLNN) can provide a very good approximation of the WMMSE algorithm to leverage the computational efficiency of the DNN. Hence in the following, we propose a deep learning based power allocation scheme for the WSR maximization problem \eqref{prob_msr0}. The proposed DNN algorithm offers multi-fold benefits compared to the WMMSE approach.
The DNN algorithm operates by continuous mapping of values from the iterations of the WMMSE algorithm. In other words, with the target output from the WMMSE algorithm, the DNN needs to learn and approximate the unknown relationship between the input and the output. Consequently, this `black box' is transformed to a nonlinear mapping that can perform like the WMMSE approach.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\linewidth]{dnn.eps}
\caption{The proposed deep neural network for approximating the WMMSE power control problem. Input: channel power gains, Output: optimal power.}\label{fig_dnn}
\end{figure}
\subsection{Defining the Neural Network}
For the machine learning based power allocation scheme, we consider a supervised learning approach which approximates the iterative WMMSE algorithm using a fully connected neural network. The DNN consists of one input layer, multiple hidden layers and one output layer as shown in Fig.~\ref{fig_dnn}. While increasing the number of hidden layers may help in decreasing the number of hidden neurons at each hidden layer thus improving the computation efficiency, the optimal number of hidden layers is intractable. In our simulations (in Section~\ref{sec_sim}) we consider three hidden layers resulting in good performance efficiency. The channel power gains of the V2I and the V2V links are the inputs to the DNN and the optimal power allocation is the output of the DNN. Note that we do not consider any bias inputs to the neurons at any layer. We apply rectified liner unit (ReLU) as the activation function at the hidden layers, while the activation function at the output layer is specifically tailored to enforce the power constraints in problem \eqref{prob_msr0}. In particular, the hidden layer activation function is defined as
\begin{align}
y_{\text{hidden}} = \max(0,x_{\text{hidden}}), \label{act_hidden}
\end{align}
and that at the output layer is defined as
\begin{align}
y_{\text{out}} = \min\left(\max(0,x_{\text{out}}),P_{\max}^{\rm x}\right). \label{act_out}
\end{align}
\subsection{Training the DNN}
For training the DNN, we first generate a large set of channel realizations $\left\{h_{m,{\rm B}}\right\}$ and $\left\{g_{k,{\rm B}}\right\}$ following certain channel distributions to reflect the V2I and V2V channels as defined in \eqref{ch_gain}. Then we generate the corresponding power allocation for each training sample using the DNN in Fig.~\ref{fig_dnn}. The training data sets are used to optimize the weights of the neural network such that the MSE of the target power allocation obtained from the WMMSE algorithm and the current DNN output is minimized. Accordingly, we define the cost function as the MSE between DNN output and the target output as
\begin{align}
MSE = \frac{1}{M + K}\sum_{i = 1}^{M + K}(\text{out}_i - \text{target}_i)^2.
\end{align}
Since training the DNN with a very large data samples is highly time-demanding, we use the so called mini-batch gradient descent (MBGD) algorithm for training the proposed DNN. An efficient implementation of the MBGD algorithm which divides the learning rate (or, gradient) by a running average of magnitudes of recent gradients. This approach is referred to as RMSprop \cite{rmsprop}. In RMSprop, the running average of the squared gradients is updated as
\begin{multline}
E\left\{\left(\nabla_w(t)\right)^2\right\} = 0.9*E\left\{\left(\nabla_w(t-1)\right)^2\right\}\\ + 0.1*\left(\nabla_w(t)\right)^2, \label{rms_prop}
\end{multline}
and the weights ($w$) at iteration $t$ is updated as \cite{adaptive_sg}
\begin{multline}
w(t+1) = w(t) - \frac{\eta}{\sqrt{E\left\{\left(\nabla_w(t)\right)^2\right\} + \epsilon}}\nabla_w(t), \label{rms_prop}
\end{multline}
where $E$ indicates arithmetic average, $\nabla_w(t) = \frac{\partial MSE}{\partial w(t)}$ is the gradient of the learning objective (in our case, the MSE) and $\epsilon$ is a smoothing term to avoid division by zero (usually on the order of $1e-8$) \cite{grad_desc}. Note that the running average at time $t$ in \eqref{rms_prop} depends only on the most recent average of the squared gradients (at time $t-1$) and the current gradient. Interestingly, the previous average squared-gradient carries more weight ($90\%$) than the current gradient ($10\%$). It has been shown in \cite{adaptive_sg} that dividing the gradient by the square root of the running average makes the learning work much better. In order to normalize the variance of each neuron's output, we then divide the weights of each neuron by the square root of is number of inputs. Similar tactic has also been applied in \cite{lto_conf}.
\subsection{Testing the DNN}
Once we learn the optimal weights of the neural network from the training stage, the next task is to validate the performance of the neural network based approach with testing data set. Again we generate a reasonably large number of random channel realizations following the same distribution as the did for training. Each set of channels is then applied to the input of the \textit{trained} neural network for an optimal power allocation at the output. The sum rate is then averaged over the number of test data sets.
\section{Numerical Simulations}\label{sec_sim}
In this section, we perform numerical simulations to demonstrate the effectiveness of the proposed machine learning based power allocation scheme for wireless V2X communications. Throughout this section, we compare the performance of the proposed approach against the WMMSE based iterative benchmark scheme, originally proposed for a MIMO system in \cite{wmmse}. Towards this end, we first simulate the equivalent WMMSE scheme developed in Section~\ref{sec_wmmse}, following Algorithm~\ref{alg_conv}.
The V2X network consists of one cellular base station, $M = 8$ CUs and $K = 10$ V2V transmitters. We construct a fully connected DNN for the system with one input layer, three hidden layers and one output layer. The input layer consists of $N+ K + N \times K$ neurons, the three hidden layers consist of $50, 22, 20$ neurons, respectively, and the output layer has $N+K$ neurons to produce the power profile. The V2X channels are generated following the model in \cite{d2d_v2x} and assumed to be exponentially distributed. In all cases, the noise variance is assumed to be $0.1$.
It is important to choose a suitable batch-size as well as the learning rate for the MBGD algorithm since both parameters can affect the network performance and efficiency. Unfortunately, there is no generic mathematical model for defining suitable values for these two parameters. Therefore, we choose these parameters by cross-validation during the training phase.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{bat_size.eps}
\caption{Effect of batch size on learning time.}\label{bat_size}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{bat_epoch.eps}
\caption{Effect of batch size on training epoch.}\label{bat_epoch}
\end{figure}
The effects of `batch size' on learning time and the required number of epochs have been demonstrated in Figs.~\ref{bat_size} and \ref{bat_epoch}, respectively. In both pictures, the fluctuation of batch size = $100$ is much more significant than other two curves, while batch size 1000 experiences the smallest fluctuation. The results in Fig.~\ref{bat_size} show the training speed with batch size $100$ requiring the least amount of time to reach the convergence, while batch size $1000$ costs the most time. On the other hand, Fig.~\ref{bat_epoch} demonstrates that the batch size $1000$ requires the least number of epochs to converge, which is in contradiction with the results in Fig.~\ref{bat_size}. We must make our choice trading off the opposing tendencies of the two graphs. While the different batch size results in Fig.~\ref{bat_size} meet after a while, those in Fig.~\ref{bat_epoch} do not. Although the graphs in Fig.~\ref{bat_epoch} are different from each other, we must stop at some point, just like the results in Fig.~\ref{bat_size}. If the epoch extends to a very large value, the difference at the beginning may become hard to be distinguished, especially for batch size $500$ and batch size $1000$.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{lrate_all.eps}
\caption{Impact of learning rate on training MSE.}\label{lrate_all}
\end{figure}
The effect of learning rate on the training MSE has been shown in Fig.~\ref{lrate_all}. The results demonstrate that the cost function converges faster for larger learning rates at the low end, however, if the learning rate is significantly high (e.g., $0.03$), the MSE performance may not show a similar tendency and start worsening instead. Although intuitively a larger learning rate means faster converging speed, the learning rate cannot be increased arbitrarily as indicated by the results in Fig.~\ref{lrate_all}.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{test_error.eps}
\caption{Performance on the test data set. The data set is exclusively generated after the training phase.}\label{test_error}
\end{figure}
After parameter selection by cross-validation and improvement of DNN, we evaluate the performance of this trained network. In Fig.~\ref{test_error}, we compare the MSE performance of the test data set not exclusively used in training the DNN against the training data set. The blue curve is the MSE during training the DNN while the red one denotes to the testing error. From Fig.~\ref{test_error}, it is obvious that:
a) no `over fitting' problem occurs, and the gap between training error and testing error is negligible,
b) the direction of convergence on the target data set is as desirable since there is no big fluctuation,
c. the performance of the DNN is determined by its testing error and the performance in this case is acceptable since the error is very low.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{v2x_cdf.eps}
\caption{The CDF of the achievable sum rates for the proposed DNN approach and the WMMSE-based baseline scheme.}\label{v2x_cdf}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{v2x_pdf.eps}
\caption{The probability density function (PDF) of the proposed scheme against the WMMSE scheme.}\label{v2x_pdf}
\end{figure}
Next, we evaluate the sum rate performance of the proposed DNN-based power allocation scheme against the WMMSE scheme proposed in \cite{wmmse} originally for a conventional interference system. Fig.~\ref{v2x_cdf} shows the cumulative distribution function (CDF) of the two schemes for $10^5$ random test samples. Since we adopted the supervised learning technique, the DNN approach can not outperform the WMMSE algorithm. However, it can be easily observed from Fig.~\ref{v2x_cdf} that the proposed scheme has comparable sum rate performance against the WMMSE approach.
Finally, we plot the probability density functions (PDFs) of the sum rate for both algorithms in Fig.~\ref{v2x_pdf}. The result of the trained network and the WMMSE algorithm is shown by their PDFs. From Fig.~\ref{v2x_pdf}, it is obvious that the DNN approach follows almost identical distribution to the WMMSE approach. These results further validates the effectiveness of the proposed machine learning approach.
\section{Conclusion}\label{sec_con}
We have considered the resource allocation problem for wireless V2X communications and shown that deep learning techniques bring enormous prospects for V2X communications where we need fast processing of optimization algorithms. Although training the DNN is a time-consuming procedure, the trained network can, however, offer a much faster solution. We have applied supervised learning approach to develop a DNN based power allocation scheme taking the classical iterative WMMSE based solution as the base line. While the supervised learning can not offer a better solution than the target one, our results demonstrate that the proposed DNN based solution can achieve very close performance to the WMMSE approach. A particular aspect of the proposed solution is that the trained DNN offers a real-time solution which is crucial for V2X communications. However, developing an unsupervised learning based power allocation scheme could be an interesting future work.
\label{sec4}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}\footnotesize{
|
\section{Introduction}
The distribution over the eigenstates of a quantum system weakly coupled to a thermal reservoir is formed by the coupling-induced interlevel transitions. In thermal equilibrium, the transitions are balanced, the rates of transitions $m\to n$ and $n\to m$ weighted with the state populations are equal for any states $m,n$ \cite{Einstein1917}.
This detailed balance condition does not hold away from thermal equilibrium. When the distribution over the states of a nonequilibrium system is stationary, it means that the overall probability to come to a state from all other states is equal to the overall probability to leave this state. In this case, if the distribution over the states is steep, the transitions to a given state from remote states may become important. This is despite the fact that the rates of such transitions, which are determined by the overlapping of the corresponding wave functions, are exponentially small.
One can think of the interstate transitions due to the coupling to a reservoir as a random walk over the states. Where the transitions between remote states are important, the walk becomes nonlocal, in the state space.
An important group of nonequilibrium system where the locality of the walk over the states can be explored is provided by periodically driven systems. The interest in such system has surged in recent years in various areas of physics and in various contexts, cf. \cite{Peano2006,Martin2017,Weinberg2017,Desbuquois2017,Schmidt2018a,Seetharam2018a,Lohse2018} and references therein. Periodically driven systems have well-defined quantum states, the Floquet or quasienergy states ~\cite{shirley1965,zeldovich1967,ritus1967}, which form a complete set. It has been appreciated \cite{kohn2001} that, when a periodically driven system is brought in contact with a thermal reservoir, there is no detailed balance, and the resulting stationary distribution over the Floquet states is complicated.
A remarkable and important exception to the lack of detailed balance in Floquet systems is provided by a nonlinear oscillator driven close to its eigenfrequency or parametrically modulated close to twice the eigenfrequency \cite{Drummond1980,Kryuchkyan1996}. For the basic model of such an oscillator, that includes the quartic nonlinearity of the potential and a simple dissipation mechanism, the oscillator has detailed balance, if the temperature of the thermal reservoir is zero. The detailed balance breaks down for nonzero temperature \cite{Dykman1979,Dykman1988a,Marthaler2006}. However, the very occurrence of it for $T=0$ is fascinating. Since it happens for two major models, one may ask whether this is a generic feature of periodically driven dissipative oscillators.
In this paper, we study the rates of transitions between the Floquet states and the stationary distribution over these states for a nonlinear oscillator driven close to triple its eigenfrequency $\omega_0$. The quantum coherent behavior of such an oscillator displays peculiar features such as the nontrivial geometric phase and the oscillations of the Floquet wave functions in the classically inaccessible region \cite{Guo2013a,Zhang2017}. An observation of period tripling in the quantum regime with a coupled-modes superconducting cavity has been already reported \cite{Svensson2017a}.
Here we explore the dissipative dynamics of this system and the associated quantum fluctuations. It turns out that the analytical tools developed for the previously explored oscillator models \cite{Dykman1988a,Marthaler2006,Peano2014} do not apply in the present case. Moreover, even for $T=0$ the oscillator does not have detailed balance. This suggests that the detailed balance for resonant and parametric driving could be an artifact.
Another important feature of the oscillators driven close to the eigenfrequency or modulated close to twice the eigenfrequency is that their stationary distribution over the Floquet states is formed locally, i.e., by transitions between a few close states. This is the case both at zero temperature and above a certain very low temperature set by $\hbar$ weighted with a combination of the nonlinearity and driving parameters \cite{Guo2013,Peano2014}. This again poses the question of whether the locality is a generic property of the $T=0$- stationary distribution of a driven oscillator. Answering this question for period tripling is one of the central topics of the present paper.
Classically, an oscillator that displays period tripling has four stable states: three states of period-three vibrations, with the same amplitude but the phases that differ by $2\pi/3$, see Fig.~\ref{fig:quasienergy_surface}~(a), and also the quiet state with no vibrations excited. Quantum fluctuations lead to switching between these states. For the driving at frequency close to $\omega_0$ and $2\omega_0$ it was found that, unless the dissipation rate is exponentially small, the switching occurs via transitions over the quasienergy barrier even for $T=0$. The mechanism was called quantum activation \cite{Marthaler2006}. The period-3 vibrations are different in several respects. One has to find how the oscillator switches between the period-3 states and between these states and the quiet state. In particular, the oscillator can possibly switch directly between the period-3 states or first go to the quiet state and then switch from this state.
In Sec.~II below, for the oscillator that displays resonant period tripling, we provide the Hamiltonian in the rotating wave approximation (RWA) and discuss the Floquet eigenstates and the Wannier-type states, which are relevant where the wells of the RWA Hamiltonian have many bound states. In Sec.~III we develop the WKB theory of the rates of dissipation-induced transitions between the intrawell states and demonstrate the breaking of the detailed balance even for $T=0$. In Sec.~IV we study the distribution over the states using an eikonal-type approximation. In this approximation the distribution is formed by transitions between nearby states. Sec.~V is one of the central parts of the paper. Here we analyze the breaking of the local approximation for $T=0$. We also show that, already for small Planck numbers of the oscillator, the locality is restored. The results are compared with numerical analysis based on calculating the transitions rates using the Wannier-type wave functions. In Sec.~VI we study switching between different wells of the RWA Hamiltonian in the regime of weak damping and also in the vicinity of the bifurcation points where the period-3 states of the oscillator emerge. Sec.~VII contains conclusions. The Appendices describe technical details of the calculations.
\section{The model and the rotating wave approximation}
\label{sec:model}
Period tripling generically occurs in strongly driven nonlinear systems. However, for weakly damped oscillators it may emerge already for a relatively weak drive provided the frequency of the driving force $\omega_F$ is close to triple the oscillator eigenfrequency $\omega_0$ \cite{Nayfeh2004}. Several aspects of the quantum formulation of the problem in the absence of dissipation were discussed previously \cite{Zhang2017}. To set the scene, here we briefly summarize and generalize the formulation.
A simple Hamiltonian of the oscillator that displays resonant period tripling is
\begin{align}
\label{eq:hamiltonian}
H&=\frac{1}{2}(p^2 + \omega_0^2 q^2) +\frac{1}{4} \gamma q^4 -\frac{1}{3}F_0q^3\cos\omega_F t,
\end{align}
where $p$ and $q$ are the momentum and coordinate of the oscillator. The conditions of resonant driving and comparatively weak nonlinearity are
\[|\delta\omega|, |\gamma|\langle q^2\rangle \ll \omega_0, \quad \delta\omega = \frac{1}{3}\omega_F - \omega_0.\]
Without loss of generality, we set $\gamma>0$ and $F_0>0$. The analysis of \cite{Zhang2017} referred to the case $\delta\omega>0$. An alternative scaling that allows one to consider the case $\delta\omega=0$ is considered in Appendix~\ref{sec:fixed_point_analysis}.
To study the oscillator dynamics near its eigenfrequency, we switch to the rotating frame at one third of the driving frequency via the unitary transformation $U(t)=\exp(-i\omega_Ft a^\dagger a/3)$, with $a=[(\omega_F q/3)+ip](3/2\hbar\omega_F)^{1/2}$. We then introduce slowly varying in time coordinate $Q$ and momentum $P$ of the oscillator in the rotating frame,
\begin{align}
\label{eq:new_variables}
&U^\dagger(t )q U(t) = C(Q\cos\phi +P\sin\phi), \\
& U^\dagger (t) p U(t) =- C\frac{\omega_F}{3}(Q\sin\phi -P\cos\phi),\qquad \phi = \omega_Ft/3,\nonumber
\end{align}
where $C=(8\omega_F|\delta\omega|/9\gamma)^{1/2}$. The operators $Q,P$ satisfy the commutation relation
\begin{align}
\label{eq:commutation}
[Q,P]=i\lambda, \quad \lambda= \frac{27\gamma\hbar}{8\omega_F^2|\delta\omega|}.
\end{align}
Parameter $\lambda$ plays the role of the Planck constant in the rotating frame. It is dimensionless, and in what follows we use the WKB approximation, which applies for $\lambda \ll 1$. In terms of the raising and lowering operators, $Q=(\lambda/2)^{1/2}(a+a^\dagger)$ and $P=-i (\lambda/2)^{1/2}(a-a^\dagger)$.
\begin{figure}[t]
\includegraphics[width = 4.1 cm]{./{period_tripling_v2}.pdf} \hfill
\includegraphics[width = 4.1 cm]{./{quasienergy_surface_v2}.pdf} \hfill
\includegraphics[width = 4.2 cm]{./orbits_v4.pdf} \hfill
\includegraphics[width = 4.1 cm]{./quasienergy_spectrum_v4.pdf}
\caption{(a) The oscillator coordinate $q(t)$ in the three period-3 vibrational states with the phases differing by $2\pi/3$ and the drive $F(t)=F_0\cos\omega_Ft$.
(b) The classical RWA Hamiltonian $g(Q,P)$, Eq.~(\ref{eq:Hamiltonian_RWA}), as a function of the coordinate and momentum in the rotating frame for the scaled amplitude of the driving field $f=1$ and $\delta\omega>0$. (c) The classical orbits of the Hamiltonian $g$. The local maximum and the three local minima are marked by the squares, and the three saddle points are marked by the filled circles. (d) Evolution of the scaled quasienergies (calculated modulo $\hbar\omega_F/3$) with the varying $f$. The scaled Planck constant is $\lambda = 0.04$. The solid lines show the eigenvalues $g_n^{(k)}$ for $k=0$ (red), 1 (blue), and 2 (green). The dashed lines show the value of the function $g(Q,P)$ at the local minima (green) and the saddle (red); at its extremum at the origin $g(0,0)=1/4$.}
\label{fig:quasienergy_surface}
\end{figure}
In the rotating wave approximation (RWA), the motion in the rotating frame is described by the canonically transformed Hamiltonian $U^\dagger H U-i\hbar U^\dagger\dot U = (\hbar|\delta\omega|/\lambda)g(Q,-i\lambda\partial_Q)$, where
\begin{align}
\label{eq:Hamiltonian_RWA}
&g(Q,P) = \frac{1}{4}[Q^2+P^2 -{\rm sgn}(\delta\omega)]^2 -\frac{1}{3} f(Q^3 - 3PQP) \nonumber\\
&f=F_0/(8\omega_F \gamma |\delta\omega|)^{1/2}.
\end{align}
Here, $g(Q,P)$ as a function of the operators $Q,P$ is the scaled dimensionless Hamiltonian of the oscillator in the rotating frame calculated in the RWA. It depends on a single parameter, the scaled field amplitude $f$, and does not depend on time.
The Hamiltonian $g(Q,P) $ as a function of the classical coordinate and momentum is plotted in Fig.~\ref{fig:quasienergy_surface}~(b). In the case shown in the figure, $g(Q,P)$ has three local minima at a nonzero $Q^2+P^2$. For the frequency detuning $\delta\omega>0$ the onset of the minima has no threshold in the scaled field amplitude $f$, whereas for $\delta\omega<0$ the minima emerge for $f>2$, see Appendix~\ref{sec:fixed_point_analysis}. In the classical limit and in the presence of weak dissipation, these minima correspond to the three period-3 vibrational states in the laboratory frame, cf. Fig.~\ref{fig:quasienergy_surface}~(a).
The Hamiltonian $g(Q,P)$ also has a local maximum or minimum at $Q=P=0$ depending on whether $\delta\omega >0$ or $\delta\omega<0$. It corresponds to a classically stable quiet state with no vibrations. The dissipative classical dynamics is outlined in Appendix~\ref{sec:Langevin}
\subsection{The quasienergy spectrum}
\label{subsec:quasienergy_spectrum}
The Schr\"odinger equation in the rotating frame in terms of the scaled RWA Hamiltonian $g$ reads
\begin{align}
i\lambda \frac{\partial}{\partial \tau}|\phi\rangle = g(Q, -i\lambda\partial_Q) |\phi\rangle.
\end{align}
We have introduced here the dimensionless time $\tau = |\delta \omega| t$.
The Hamiltonian $g$ is invariant with respect to a rotation in the $Q,P$ phase plane by $2\pi/3$. Such rotation is described by the operator $\hat N_3 = \exp(-2\pi ia^+a /3)$, which commutes with $g$.
Therefore the eigenfunctions $\phi_n^{(k)}$ of $g$ are also eigenfunctions of $\hat N_3$,
\begin{align}
\label{eq:N_3}
\hat N_3\phi_n^{(k)} = \exp(-2\pi i k/3) \phi_n^{(k)}.
\end{align}
The superscript $k=0,1,2$ here determines the eigenvalue of the rotation operator $\hat N_3$. The subscript $n$ is the other number that enumerates the eigenstates of the Hamiltonian $g(Q,-i\lambda\partial_Q)$; this Hamiltonian has an infinite countable set of eigenfunctions and eigenvalues.
With the account taken of the interrelation between the oscillator Hamiltonian in the laboratory frame (\ref{eq:hamiltonian}) and $g$, the eigenvalues $g_n^{(k)}$ of $g$ determine the quasienergies $\varepsilon_n^{(k)}$ of the driven oscillator in the RWA \cite{Zhang2017},
\[\varepsilon_n^{(k)} = \frac{\hbar |\delta\omega|}{\lambda}g_n^{(k)} +\frac{2\pi k}{3}\hbar\omega_F.\]
As seen from this expression, the eigenvalues $g_n^{(k)}$ give the the scaled quaisenergies calculated in the reduced Brillouin zone $0\leq \varepsilon < \hbar\omega_F/3$. In what follows we call these eigenvalues scaled RWA energies, or just RWA energies, for brevity.
To calculate the eigenvalues and eigenfunctions of the scaled RWA Hamiltonian $g$, it is convenient to rewrite $g$ in terms of the ladder operators $a$ and $a^\dagger$,
\begin{align}
\label{eq:g_in_ladder_ops}
g& = \lambda \Bigl[-{\rm sgn}(\delta\omega)( a^\dagger a+1/2) + \lambda a^\dagger a(a^\dagger a +1) \nonumber\\
&- \frac{f}{3}\sqrt{2\lambda}(a^{\dagger 3}+a^3)\Bigr] + \frac{1}{4}(1+\lambda^2).
\end{align}
Several low-lying eigenvalues of $g$ as functions of the scaled driving amplitude $f$ are shown in Fig.~\ref{fig:quasienergy_surface}(d). It is clear from Eq.~(\ref{eq:N_3}) and can be also directly seen from Eq.~(\ref{eq:g_in_ladder_ops}) that the eigenvalues $g_n^{(k)}$ form groups with different values of $k=0,1,2$. The eigenvalues with different $k$ can cross with the varying $f$, as they correspond to different eigenvalues of $\hat N_3$. The eigenvalues with the same $k$ anticross.
An important qualitative feature of the evolution of the spectrum is that, with the increasing $f$, the eigenvalues with different $k$ cluster into triples. This can be understood from the form of the function $g(Q,P)$ in Fig.~\ref{fig:quasienergy_surface}~(b) and (c). For nonzero $f$ and $\delta\omega>0$, this function has three minima separated by three saddle points. The clustering occurs once the values of $g_n^{(k)}$ become smaller than the saddle-point value $g_s$ of $g(Q,P)$, which is shown as the red dashed line in Fig.~\ref{fig:quasienergy_surface}. The states with $g_n^{(k)}<g_s $ correspond to linear combinations of the intrawell states of $g(Q,P)$ that have the same RWA energy in the neglect of interwell tunneling. The tunneling between the three wells leads to the level splitting within a triple of the eigenstates. It sharply falls off as the states go deeper into the wells, see Sec.~\ref{subsec:intrawell}.
As the driving strength increases, the wells of $g(Q,P)$ become deeper whereas the saddle points get closer to the local maximum of $g(Q,P)$ at $Q=P=0$. Thus, the number of intrawell states increases, whereas the number of the states localized near $Q=P=0$ decreases.
\subsection{Intrawell states}
\label{subsec:intrawell}
Of central interest for this paper is the regime of a relatively strong drive where the three wells of the Hamiltonian function $g(Q,P)$ are well-separated in phase space compared to the typical quantum scale of the phase-space area $\lambda$. One can then think of the states localized mostly inside the wells. The number of such states is $\propto 1/\lambda$. We will be interested in the case where this number is large.
We will denote the intrawell states by $\Ket{n}_\nu\equiv |\psi_\nu(Q; n)\rangle$, where $\nu=0,1,2$ enumerates the wells. In what follows we choose the $\nu=0$ well to have a minimum on the $P=0$-axis, see Fig.~\ref{fig:quasienergy_surface}. The wells with $\nu=1,2$ are symmetrically located with respect to the $P=0$ axis, with their minima at the same $Q$. We use $({}Q_\nu,{}P_\nu)$ to denote the positions of the minima of the wells. These points are vortices of an equilateral triangle on the $(Q,P)$ plane, see Fig.~\ref{fig:quasienergy_surface}. We denote the minimal value of $g$ as $g_{\min}\equiv g({}Q_\nu,{}P_\nu)$.
The wave functions $\psi_\nu(Q;n)$ are approximate eigenstates of the Hamiltonian $g$. Near the bottom of a well $\nu$, functions $\psi_\nu(Q;n)$ have the form of the eigenfunctions of the harmonic oscillator in their central part. The Hamiltonian of this oscillator is obtained by expanding $g(Q,P)$ about $({}Q_\nu,{}P_\nu)$ to the second order in $Q-{}Q_\nu, P - {}P_\nu$. The lowest state was described earlier \cite{Zhang2017}. The large-$n$ states can be obtained in the WKB approximation,
\begin{align}
\label{eq:WKB_nu}
&\psi_\nu(Q;n)=C _n^{(\nu)}(\partial_P g)^{-1/2}\exp[iS^{(\nu)}(Q,g_n)/\lambda],\nonumber\\
&\partial_Q S^{(\nu)} = P^{(\nu)}(Q,g_n)
\end{align}
Here, $P^{(\nu)}(Q,g)$ is the solution of the equation $g(Q,P)=g$ that corresponds to the Hamiltonian orbit circling the $\nu$th minimum of $g(Q,P)$ with the classical RWA energy $g$. Parameter $C^{(\nu)}_n$ is the normalization constant. For large $n$, one finds $g_n$ by calculating $S_\nu(Q,g)$ and applying the standard Bohr-Sommerfeld quantization condition \cite{landau1977}. By symmetry the values of $g_n$ are the same for all wells $\nu=0,1,2$.
The intrawell wave functions of different wells overlap, which is the tunneling effect. However, for small $\lambda$ the overlapping is exponentially small. As a result we have, to the leading order in the overlapping,
\begin{align}
\label{eq:g_matrix}
_\nu\!\Bra{n} g \Ket{n}_\nu = g_n,\qquad _\nu\!\Bra{n} g \Ket{n}_{\nu\pm 1} = J_{\nu\pm}(g_n).
\end{align}
The hopping integral $J_{\nu\pm}(g_n)$ for the lowest intrawell states was found in Ref.~\onlinecite{Zhang2017}; for higher states it is discussed in Sec.~\ref{subsec:onset_quantum_activation}. The notion of intrawell states is meaningful if $|J_{\nu\pm}(g_n)|\ll |g_n-g_{n\pm 1}|$. Since the symmetry operator $\hat N_3$ can be thought of as rotating all wells together, $\nu\to \nu+1, \forall \nu$, we see that $J_{\nu\pm}$ is independent of $\nu$ and $J_{\nu+}\equiv J_{0+}=J_{\nu-}^*\equiv J_{0-}^*$.
To the lowest order in the overlap integrals, the eigenfunctions $\phi_\nu^{(k)}$ and eigenvalues $g_n^{(k)}$ ($k=0,1,2$) of the RWA Hamiltonian $g$ can be found in the same way as the Bloch wave functions and the energy levels in the tight-binding approximation in solid-state physics \cite{Ziman1979},
\begin{align}
\label{eq:nth_wave_functions}
&\phi_n^{(k)}(Q) \approx 3^{-1/2}\sum_\nu \psi_\nu(Q;n) e^{-2\pi i \nu k/3},\nonumber\\
&g_n^{(k)} - g_n \approx 2{\rm Re}[ J_{0+}(g_n)\exp(-2\pi i k/3)].
\end{align}
From Eqs.~(\ref{eq:WKB_nu}) - (\ref{eq:nth_wave_functions}), the approximately calculated Bloch-type wave functions $\phi_n^{(k)}(Q)$ have maxima inside the wells of $g(Q,P)$ and are small outside the wells. It is clear from qualitative arguments that this is also true for the wave functions $\phi_n^{(k)}(Q)$ calculated exactly, for example, by numerically diagonalizing the Hamiltonian (\ref{eq:g_in_ladder_ops}). It is important that, since the operator $g(Q,-i\lambda\partial_Q)$ is real, its eigenfunctions can be also chosen to be real.
Using the exact Bloch-type wave functions $\phi_n^{(k)}(Q)$, one can construct Wannier-type wave functions
\begin{align}
\label{eq:Wannier_defined}
\psi_{\nu; {\rm W}}(Q;n)= 3^{-1/2}\sum_k \phi_n^{(k)}(Q)\exp(2\pi i\nu k/3).
\end{align}
Since the exact wave functions $\phi_n^{(k)}(Q)$ are orthogonal and normalized, the Wannier functions (\ref{eq:Wannier_defined}) are also orthogonal and normalized, in contrast to $\psi_\nu(Q;n)$. However, one can make functions $\psi_\nu(Q;n)$ and $\psi_{\nu; {\rm W}}(Q;n)$ with the same $\nu$ and $n$ very close to each other inside the $\nu$th well, simultaneously for all $\nu$. This requires choosing the exact functions $\phi_n^{(k)}(Q)$ to be real and properly choosing the signs of all $\phi_n^{(k)}(Q)$ with a given $n$. As Eq.~(\ref{eq:nth_wave_functions}) suggests, inside the well $\nu=0$, where only $\psi_0(Q;n)$ is large (in the absolute value), if we want it to be close to $\psi_{0; {\rm W}}(Q;n)$, all exact wave functions $\phi_n^{(k)}(Q) $ should be close to each other. This determines the choice of the otherwise arbitrary signs of the numerically calculated $\phi_n^{(k)}(Q) $ with different $k$.
\section{Dissipation-induced intrawell transitions}
\label{sec:intrawell_transitions}
Coupling of the oscillator to a thermal reservoir leads to dissipation and to transitions between the intrawell states $|n\rangle_\nu\equiv \Ket{\psi_\nu(Q;n}$ of the Hamiltonian $g(Q,-i\lambda\partial_Q)$. We consider the parameter range where the dissipation-induced widths of the intrawell levels $g_n$, which are determined by the dissipation rate, are much larger than the tunnel splitting $\sim |J_{0\pm}(g_n)|$. In this case the statistical distribution over the intrawell states of the oscillator is formed much faster then the distribution over different wells. By symmetry arguments, the intrawell transition rates are the same for the different wells of $g(Q,P)$. For concreteness, we will consider transitions between the states in the well $\nu= 0$ and use a simplified notation $|n\rangle \equiv |n\rangle _{\nu=0}$.
As mentioned above, we enumerate the intrawell states in such a way that $g_n>g_m$ if $n>m$. Even at zero temperature, where dissipation corresponds to emission of excitations into the thermal reservoir and thus the oscillator makes transitions only from its Fock states with higher (true) energy to the Fock states with lower energy, in terms of the intrawell states $\Ket{n}$ the transitions occur both toward larger and lower $n$. Formally, this is a consequence of the states $|n\rangle$ being linear combinations of the Fock states. The transitions between the states $|n\rangle$ are sketched in Fig.~\ref{fig:schematic_hoppings}.
\begin{figure}[h]
\includegraphics[width = 7.5 cm]{./{schematic_hoppings_v2}.jpg}
\caption{The cross-section of the Hamiltonian function $g(Q,P)$ by the plane $P=0$ and the sketch of dissipation-induced transitions between the intrawell states. Even at zero temperature there occur transitions both up and down the $g$-axis. The direction of more probable transitions determines which Floquet state is predominantly occupied in the presence of dissipation. In particular, inside the well of $g(Q,P)$, the transitions toward the minimum of the well (blue arrow) have higher rates than away from the minimum. For the oscillator driven at $\approx 3\omega_0$, the transition rates do not satisfy the detailed balance condition even for $T=0$, in contrast to the oscillator driven at $\approx \omega_0$ or $\approx 2\omega_0$.}
\label{fig:schematic_hoppings}
\end{figure}
The rate of a transition $|n\rangle \to |m\rangle$ is determined by the overlapping of the wave functions $|n\rangle,|m\rangle$. Therefore it falls off exponentially with the ``distance'' $|n-m|$. However, as we show in the next section, because the stationary state populations $\rho_n$ also fall off exponentially with increasing $n$, transitions from the remote small-$n$ states close to the bottom of the well can make a dominating contribution to the populations $\rho_n$ of large-$n$ states. This means strong non-locality of the distribution $\rho_n$ in the state-number space. Such effect is fairly general for systems far from equilibrium, whereas we are not aware of systems in thermal equilibrium where it would be known, unless there are selection rules that make the transition rates very different for different groups of states.
\subsection{Balance equation for the state populations}
\label{subsec:balance_equation}
To describe dissipation, we assume that the periodically driven nonlinear oscillator is weakly coupled to a thermal reservoir and that the coupling is linear in the oscillator dynamical variables. Then dissipation comes from the oscillator transitions between neighboring Fock states with emission or absorption of excitations in the reservoir. We further assume that the spectral density of these excitations weighted with the interaction is nearly constant around the oscillator eigenfrequency within a band with width given by the detuning $|\delta \omega|$ and the nonlinear shift of the oscillator eigenfrequency $|\gamma| \langle q^2\rangle/\omega_0$. Then the oscillator dynamics in slow time $\tau = |\delta\omega|t$ is Markovian and is described by a quantum kinetic equation of the Landau-Lindblad form (cf. \cite{Landau1927,Walls2008}),
\begin{align}
\label{eq:master_general}
\frac{d}{d\tau} \rho =& (i/\lambda)[\rho, g] - \kappa{\cal D}[a]\rho,\nonumber\\
{\cal D}[a]\rho =& (\bar n+1)(a^\dagger a\rho - 2a\rho a^\dagger + \rho a^\dagger a)\nonumber\\
&+\bar n (a a^\dagger\rho -2 a^\dagger \rho a + \rho a a^\dagger).
\end{align}
Here, $\kappa$ is the dimensionless decay rate of the oscillator amplitude; it is related to the decay rate of the amplitude in the unscaled time (the coefficient of viscous friction, in the classical description) $\Gamma$ as $\kappa=\Gamma/\|\delta\omega|$; $\bar n$ is the Planck number, $\bar n = [\exp(\hbar\omega_0/k_B T)-1]^{-1}$.
Since the overlap integrals of the wave functions in different wells of $g(Q,P)$ are exponentially small, one can disregard mixing of the states in different wells by the relaxation superoperator ${\cal D}$. For $\kappa \gg |J_{0\pm}|$, dissipation-induced intrawell transitions occur much faster than interwell transitions. For the dimensionless time $\tau$ small compared to the long time of the interwell relaxation, the interwell transitions can be disregarded; they are discussed below in Sec.~\ref{sec:tunneling}.
If the distance between the intrawell levels $g_{n+1} -g_n$ is large compared to the level broadening $\propto \lambda\kappa$, the off-diagonal matrix elements of the density matrix $\langle n|\rho|n'\rangle$ ($n\neq n'$) are small compared to the state populations $\rho_n \equiv \langle n|\rho|n\rangle$ (we recall that $|n\rangle$ refers to the intrawell states in the well $\nu=0$; for concreteness, we consider the matrix elements of $\rho$ for the states in this well). If we disregard the off-diagonal matrix elements, we obtain a balance equation for the state populations $\rho_n$,
\begin{align}
\label{eq:balance_equation}
\frac{d}{d\tau} \rho_n = -\sum_{n'} W_{nn'} \rho_n + \sum_{n'} W_{n'n} \rho_{n'},
\end{align}
where $W_{nn'}$ are the rates of the dissipation-induced transition $|n\rangle \to |n'\rangle$. From Eq.~(\ref{eq:master_general}),
\begin{align}
\label{eq:hopping_rates}
W_{nn'} = 2\kappa \left[ (\bar n +1)\left|\langle n'|a|n\rangle\right|^2 + \bar n \left|\langle n|a|n'\rangle \right|^2\right]
\end{align}
We note that in a nonequilibrium system the ratio of the transition rates is not given by the Einstein relation for equilibrium systems, $W_{nn'}/W_{n'n} \neq (\bar n + 1)/\bar n$~\cite{Kubo1957}.
\subsection{Transition rates in the semiclassical limit}
\label{sec:matrix_elements}
The rates of the dissipation-induced intrawell transitions $W_{n\, n\pm m}$ are determined by the matrix elements
\[\langle n+m|\hat a|n\rangle \equiv a_m(g_n).\]
For small $\lambda$ and for not too large $|m|\ll 1/\lambda$
such matrix elements are given by the Fourier components of the vibrations of the corresponding classical dynamical variable with the scaled RWA energy $g_n$~\cite{landau1977},
\begin{align}
\label{eq:matrix_elements}
a_m(g) =\frac{\omega(g)}{2\pi}\int_{-\pi/\omega(g)}^{\pi/\omega(g)}d\tau \exp[-im\omega(g)\tau]a(\tau; g).
\end{align}
Here, $\omega(g)$ is the cyclic frequency of the periodic motion along the intrawell classical trajectory $g(Q,P)=g$, which is described by the Hamiltonian equations
\begin{align}
\label{eq:classical_eom}
\frac{d}{d\tau} Q = \partial_P g, \qquad \frac{d}{d\tau} P = -\partial_Q g.
\end{align}
The function $a(\tau;g)$ is calculated as a function of time on this trajectory and is expressed in terms of the coordinate $Q$ and momentum $P$ on the trajectory as $a(\tau;g) = (2\lambda)^{-1/2}[Q(\tau;g)+iP(\tau;g)]$. Equations (\ref{eq:hopping_rates}) - (\ref{eq:classical_eom}) allow one to calculate the hopping rates numerically in the WKB approximation.
The WKB ansatz (\ref{eq:matrix_elements}) requires $n, n+m\gg 1$. For $n, n+m$ close to 1, the wave functions $\Ket{n}, \Ket{n+m}$ are close to the wave functions of a harmonic oscillator (see below) and the calculation of the matrix elements of $Q$ and $P$ is straightforward.
In terms of the matrix elements (\ref{eq:matrix_elements}), the transition rates become
\begin{align}
\label{eq:semiclassical_hopping_rates}
W_{n \,n+m} = 2\kappa [(\bar n +1)|a_m(g_n)|^2 + \bar n |a_{-m}(g_n)|^2].
\end{align}
Here, we have neglected corrections $\sim \lambda$ and, for large $n$, set $a_{-m}(g_{n+m})\approx a_{-m}(g_n)$, in which case $W_{n\,n+m} \approx W_{n-m \,n}$.
The matrix elements $a_{m}(g_n)$ slowly vary with $n$ for large $n$, but rapidly change with $m$. The structure of the solution of the balance equation depends on the asymptotic behavior of $a_m$ for large $|m|$, such that $1\ll |m|\ll1/\lambda$. To find the large-$|m|$-behavior of $a_m(g)$, one can shift the integration contour in Eq.~(\ref{eq:matrix_elements}) into complex time. For positive and negative $m$ the contour is shifted into the lower and upper halfplane of the $\tau$-plane, respectively. The locations of the singularities of the function $a(\tau;g)$ in the complex time plane determine the exponential decay rate of the matrix element $a_m$.
The details of this calculation are in Appendix~\ref{sec:contour_integral}. For large negative $m$, we have
\begin{align}
\label{eq:a_m<0}
|a_{m<0}(g_n)|\approx &\frac{(3/2)^{1/6}\Gamma(1/3)}{2\pi \sqrt{\lambda}}\left[ \frac{\omega^2(g_n)}{|m|f}\right]^{1/3} \nonumber \\
&\times \exp[-|m|\omega(g_n)\tau_\infty(g_n)].
\end{align}
For large positive $m$, we have
\begin{align}
\label{eq:a_m>0}
|a_{m>0}(g_n)|\approx &\frac{(2/3)^{1/6}\Gamma(2/3)}{2\pi\sqrt{\lambda}}\left[\frac{f\omega(g_n)}{m^2}\right]^{1/3} \nonumber \\
&\times \exp[-m\omega(g_n)\tau_\infty(g_n)].
\end{align}
The parameter $\tau_\infty$ is the smallest distance along the imaginary time $\tau$ to the singularity of $Q(\tau;g)$ and $P(\tau;g)$. It is defined by Eq.~(\ref{eq:imaginary_time}) and is the same in the upper and the lower half-plane of the complex $\tau$ plane. It is shown in Fig.~\ref{fig:tau_infty}.
Since the distance to the singularities is the same in the both half-planes, the matrix elements $|a_m(g_n)|^2$, and thus the transition rates, fall off with $|m|$ exponentially with the same exponent proportional to $\tau_\infty>0$ for $m>0$ and $m<0$. This is qualitatively different from the rates of transitions between the Floquet states of an oscillator driven by a force with frequency close to the eigenfrequency or a parametrically modulated oscillator at frequency close to twice the oscillator eigenfrequency \cite{Marthaler2006,Guo2013,Peano2014}. In the both latter cases, for $T=0$, the rates of transitions away from the stable state fall off exponentially faster with the interstate distance than the rates of transitions toward the stable state.
\begin{figure}[ht]
\includegraphics[width = 7.5 cm]{./{tao_infinity_v3}.pdf}
\caption{The product $\omega(g)\tau_\infty(g)$ as a function of the distance to the bottom of the well $\Delta g{} = (g-g_{\min})/(g_s-g_{\min})$; here, $\tau_\infty(g)$ is the minimal imaginary time to go to infinity along the Hamiltonian trajectory (\ref{eq:classical_eom}), whereas $\omega(g)$ is the classical frequency of intrawell vibrations with a given $g$. The curves with different colors refer to the scaled drive amplitude $f$ = 0.1 (red), 0.5 (black), and 2 (green); $\delta\omega>0$. The insets shows $\omega(g)$ and $\tau_\infty(g)$ separately for the same values of $f$. Note that the values $g_{\rm min}$ and $g_s$ of $g(Q,P)$ at the minimum and the saddle point depend on $f$. }
\label{fig:tau_infty}
\end{figure}
We show in Fig.~\ref{fig:tau_infty} the decrement $\omega(g)\tau_\infty(g)$ of the fall-off rate of the matrix elements Eqs.~(\ref{eq:a_m<0}) and (\ref{eq:a_m>0}) with $|m|$. The decrement monotonically decreases with the increasing $g$, and interestingly, for a given $\Delta g{}$, it depends only weakly on the scaled drive amplitude $f$. For $g$ close to $g_{\rm min}$, it diverges logarithmically
\begin{align}
\label{eq:tao_infty_diverge}
\omega(g)\tau_\infty \approx -\frac{1}{2}\ln \Delta g{}, \qquad \Delta g{}\equiv \frac{g-g_{\rm min}}{g_s-g_{\rm min}}\ll 1,
\end{align}
due to the divergence of $\tau_\infty(g)$ seen in the inset of Fig.~\ref{fig:tau_infty}, and weakly depends on $f$, for a given $\Delta g$. This is because classical vibrations near the minima of $g(Q,P)$ become almost harmonic. Therefore, higher-order Fourier components of the classical motion decay fast. For $g$ close to the saddle point value $g_s$, the period of the classical orbit diverges logarithmically and the motion becomes extremely anharmonic: very slow near the saddle points, but fast away from them. As a result, $\omega(g)\tau_\infty(g)$ approaches zero logarithmically for $g$ close to $g_s$.
Another distinctive and important feature of period tripling is that the singularities of $a(\tau;g)$ in the complex plane $\tau$ are branching points rather than poles, as for resonant driving or parametric modulation. The branching points at $ i\tau_\infty$ and $-i\tau_\infty$ are different. Respectively, the prefactors in Eqs.~(\ref{eq:a_m<0}) and (\ref{eq:a_m>0}) show a power-law dependence on $m$, but with different exponents.
The parameter used to obtain the large-$|m|$ asymptotic behavior of the coefficients $a_m(g)$ is $|m|\omega(g)\tau_\infty\gg 1$. For $f\lesssim 1$ we have $\tau_\infty \sim 1$, see Fig.~\ref{fig:tau_infty} ($\tau_\infty \gg 1$ for $g\to g_{\min}$). Then Eqs.~(\ref{eq:a_m<0}) and (\ref{eq:a_m>0}) apply provided $|m|\omega(g) \gg 1$. In this case for $\bar n =0$ the ratio of the rates of transitions away and toward $g_{\min}$ is $W_{n\,n+m}/W_{n\,n-m} \propto [f^2/m\omega(g)]^{2/3} \ll 1$ ($m>0$). This shows that, in the stationary regime, the system is mostly localized near the minimum of the well of $g(Q,P)$: it is much more probable to make a transition toward this minimum than away from it. From Eq.~(\ref{eq:Hamilton_eqns}), for $f\gg 1$ the frequency $\omega(g)$ scales with $f$ as $f^2$, and therefore still $W_{n\,n+m}/W_{n\,n-m}\ll 1$ for $m>0$ as long as $m\omega(g)\tau_\infty(g)\gg 1$.
\subsection{Breakdown of the detailed balance}
\label{subsec:detailed balance}
A periodically driven oscillator is far from thermal equilibrium. Generally, it should not have detailed balance. The absence of detailed balance was shown for a resonantly driven oscillator in the classical limit \cite{Dykman1979}. However, as mentioned in the Introduction, in the deeply quantum regime of zero temperature, a resonantly driven Duffing oscillator with the relaxation described by the master equation (\ref{eq:master_general}) has detailed balance in the RWA \cite{Drummond1980}. Moreover, for the same type of relaxation, detailed balance for zero temperature was found also for a Duffing oscillator parametrically modulated close to twice its eigenfrequency \cite{Kryuchkyan1996}. The physical reason for the detailed balance, which exists only for $T=0$ and only for this relaxation mechanism is not known (it breaks down in the presence of dephasing \cite{Marthaler2006}). It is interesting and, in a way, illuminating, to find out whether detailed balance persists for $T=0$ and relaxation described by Eq.~(\ref{eq:master_general}) in the case of an oscillator driven close to triple the eigenfrequency.
The condition of detailed balance is the relation between the rates of interstate transitions that shows that the ratio of the rates of direct transitions back and forth between the states is equal to the ratio of the rates of transitions via an intermediate state,
\[ \frac{W_{nn'}}{ W_{n'n}} = \frac{W_{nn''} }{W_{n'n''}} \frac{W_{n''n'}}{ W_{n''n}} .\]
For $T=0$ we have $W_{nn'} = 2\kappa |a_{n'-n}(g_n)|^2$. As seen from Eqs.~(\ref{eq:a_m<0}) and (\ref{eq:a_m>0}), for period tripling there is no detailed balance even for $T=0$. This shows that the detailed balance for modulation at frequencies $\approx \omega_0$ and $\approx 2\omega_0$ for $T=0$ is non-generic.
\section{Quasistationary distribution over intrawell states}
\label{sec:distribution}
Over the dimensionless relaxation time $\sim \kappa^{-1}$, transitions between the intrawell states form a quasistationary distribution over these states. This distribution can be found from the stationary solution of the balance equation (\ref{eq:balance_equation}). An important property of the transition rates is
\[W_{n+m\,n} >W_{n\,n+m}, \quad m>0,\]
that is, the system is more likely to go from a higher-lying to a lower-lying state than in the opposite direction. This property is seen from Eqs.~(\ref{eq:a_m<0}) and (\ref{eq:a_m>0}) for large $m$ and from the numerical evaluation of the semiclassical rates $W_{nn'}$ based on Eq.~(\ref{eq:matrix_elements}) for $|n-n'|\sim 1$. It is clear then that the stationary distribution will be maximal near the bottom of the well of $g(Q,P)$. However, since the probabilities of transitions up along the $g_n$ axis are nonzero, the stationary distribution over the levels $g_n$ will have a nonzero width.
We are interested in the semiclassical parameter range $\lambda\ll 1$, where the number of intrawell bound states is large.
The analysis of the distribution should be done differently in different ranges of the values of the scaled RWA energy $g$. We note that it does not apply in a narrow range of $g$ close to the saddle value $g_s$, where the interwell transitions have to be taken into account.
\subsection{The vicinity of the minima of the wells of $g(Q,P)$}
\label{sec:harmonic_approximation}
The distribution over intrawell states can be found analytically near the minima of the RWA Hamiltonian function $g(Q,P)$. In the rotating frame, the vibrations of $Q,P$ about a minimum are almost harmonic and, respectively, the levels $g_n$ are almost equidistant. To find the distribution for the well $\nu=0$, we expand $g(Q,P)$ about the $\nu=0$-minimum,
\begin{align}
\label{eq:g_expanded}
&g \approx g_{\rm min} + \frac{1}{2}g_{PP} P^2 + \frac{1}{2}g_{QQ} (Q-Q_0)^2, \nonumber \\
& g_{PP} = 3fQ_0, \qquad g_{QQ} = fQ_0 +2{\rm sgn}(\delta\omega), \nonumber\\
&Q_0 = \frac{1}{2}[f+\sqrt{f^2+4{\rm sgn}(\delta\omega)}].
\end{align}
Here $(Q_0,P_0=0)$ is the position of the minimum and $g_{\min}$ is the value of $g(Q,P)$ at the minimum, $ g_{\min}=fQ_0^2(3f-4Q_0)/12$.
It is convenient to change from $P$ and $Q-Q_0$ to the raising and lowering operator $b^\dagger$ and $b$ using a squeezing transformation,
\begin{align}
\label{eq:g_harmonic}
& Q - Q_0 + iP = (2\lambda)^{1/2}(b \cosh\phi_* - b^\dagger \sinh\phi_*), \nonumber \\
& g \approx g_{\rm min} + \lambda \omega_{\rm min}(b^\dagger b + 1/2),\quad \omega_{\rm min} = \sqrt{g_{PP}g_{QQ}},
\end{align}
where $\omega_{\rm min}$ is the vibration frequency at the bottom of the well of $g(Q,P)$ and the squeezing parameter $\phi_*$ is given by the equation $\tanh\phi_* = (|g_{QQ}|^{1/2}-|g_{PP}|^{1/2})/(|g_{QQ}|^{1/2}+|g_{PP}|^{1/2})$. The intrawell states $\Ket{n}$ near the bottom of the well are well approximated by the eigenstates of the operator $b^\dagger b$.
From Eq.~(\ref{eq:g_harmonic}), the lowering operator of the oscillator $a= (2\lambda)^{-1/2}(Q+iP)$ is a linear combination of the operators $b$ and $b^\dagger$. Therefore the transitions rates $W_{n\,n+m}$, Eq.~(\ref{eq:hopping_rates}) are nonzero for $m=\pm 1$, and this is the case even for $\bar n=0$, where the oscillator makes dissipative transitions only to a lower Fock state. The rates $W_{n\,n+m}$ are easy to find from Eq.~(\ref{eq:g_harmonic}), taking into account that $\Bra{n}b\Ket{n+m} = (n+1)^{1/2}\delta_{m,1}$. Substituting the rates into the balance equation (\ref{eq:balance_equation}), we find that the stationary solution of this equation near the minimum of the well is
\begin{align}
\label{eq:Boltzmann}
&\rho_n \propto \exp(-n\lambda \omega_{\rm min}/T_{\rm eff}), \nonumber \\
&\lambda \omega_{\rm min}/T_{\rm eff} = \log[(\bar n_{\rm eff} + 1)/\bar n_{\rm eff}].
\end{align}
where $n_{\rm eff}$ is the effective Planck number,
\begin{align}
\label{eq:effective_nbar}
\bar n_{\rm eff} = \bar n + (2\bar n +1) \sinh^2\phi_*.
\end{align}
Equation (\ref{eq:Boltzmann}) has the form of the Boltzmann distribution. As seen from Eq.~(\ref{eq:effective_nbar}), the effective temperature of the distribution over the intrawell states is nonzero even where the effective temperature of the thermal reservoir is equal to zero. This is the quantum heating effect \cite{Dykman1988a,Marthaler2006}. For a resonantly driven oscillator it was observed experimentally in Ref.~\onlinecite{Ong2013}.
\begin{figure}
\includegraphics[width = 5.5 cm]{./{harmonic_R_v4}.pdf}
\caption{The scaled inverse temperature $\lambda \omega_{\rm min}/T_{\rm eff}$ in the harmonic approximation as a function of the scaled drive amplitude $f$ for $\delta\omega>0$. The red, black, blue, and green lines are given by Eqs.~(\ref{eq:Boltzmann}) and (\ref{eq:effective_nbar}) with the equilibrium Planck number $\bar n =0, 0.01,0.1, 1$, respectively. The dots show the result of the eikonal approximation, Eq.~(\ref{eq:equation_for_xi}), for $g$ close to $g_{\rm min}$ ($\Delta g = 0.0001$). We define the effective temperature in this case from the expression $\lambda \omega_{\rm min}/T_{\rm eff}=[R'(g) \omega(g)]_{g\to g_{\min}} $. The difference between the dots and the curves is pronounced for $\bar n=0$ and small $|f-1/\sqrt{2}|$, where $T_{\rm eff} \to \infty$ in the harmonic approximation that gives Eqs.~(\ref{eq:Boltzmann}) and (\ref{eq:effective_nbar}). This difference decreases with the decreasing $\Delta g$. }
\label{fig:harmonic_R'}
\end{figure}
The dependence of $\bar n_{\rm eff}$ on $f$ is different for different signs of the detuning of the drive frequency $\delta\omega= (\omega_F/3)-\omega_0$. For $\delta\omega<0$, as mentioned previously, $g(Q,P)$ has a minimum at $Q=P=0$, whereas the wells at nonzero $Q^2+P^2$ appear only for $f>2$. In this case $\bar n_{\rm eff}$ monotonically decreases as $f$ increases. Near the threshold $f = 2$, $\bar n_{\rm eff}$ diverges as $(f-2)^{-1/4}$. For large $f\gg 1$, $\bar n_{\rm eff}$ asymptotically approaches $ (2\bar n + 1)/\sqrt{3}-1/2$.
For $\delta\omega>0$ function $g(Q,P)$ has the form of a Mexican hat, with the top at $Q=P=0$ and the rim at $Q^2+P^2 = 1$ for $f=0$. The wells $\nu=0,1,2$ emerge for $f>0$ with no threshold, as a modulation of the depth of the rim. The effective Planck number $\bar n_{\rm eff}$ is non-monotonic as a function of the scaled drive amplitude $f$. It diverges as $f^{-1/2}$ for $f\to 0$. As $f$ increases, $\bar n_{\rm eff}$ reaches a minimum at $f=1/\sqrt{2}$, where $\bar n_{\rm eff}=\bar n$. This corresponds to the disappearance of the squeezing, $\sinh\phi_* = 0 $ for $f=1/\sqrt{2}$. For $f>1/\sqrt{2}$, $\bar n_{\rm eff}$ increases with the increasing $f$ and approaches the same asymptotic value as in the case $\delta\omega<0$.
We show in Fig.~\ref{fig:harmonic_R'} the scaled effective inverse temperature $\lambda \omega_{\rm min}/T_{\rm eff}$ as a function of the scaled drive amplitude $f$ for $\delta\omega>0$. For $\bar n = 0$, the inverse temperature diverges at $f=1/\sqrt{2}$ as $\left|\log(f-1/\sqrt{2})^2\right|$.
For $f=1/\sqrt{2}$ and $\bar n=0$ the analysis of the distribution over the states $\Ket{n}$ has to go beyond the harmonic approximation (\ref{eq:g_harmonic}) and take into account the terms of higher order in $P$ and $Q-Q_0$ in Eq.~(\ref{eq:g_expanded}). By perturbation theory in the parameter $\lambda$, the cubic terms in $P$ and $Q-Q_0$ lead to a finite transition rate $W_{02}\propto \lambda$, and more generally, $W_{n\,n+2}\propto \lambda$ for $n\sim 1$. The finite rates of the upward in $g$ transitions result in a nonzero population of the excited states; the distribution over the states falls off as a power series in $\lambda$.
\subsection{The eikonal approximation}
\label{subsec:eikonal}
To find the stationary solution of the balance equation (\ref{eq:balance_equation}) away from the minimum of $g(Q,P)$ it is convenient to seek this solution in the eikonal form,
\begin{align}
\label{eq:eikonal}
\rho_n = \exp[-R(g_n)/\lambda].
\end{align}
As it stands, Eq.~(\ref{eq:eikonal}) just uses the fact that $\rho_n\geq 0$. It applies near the minimum of the well of $g(Q,P)$ as well, as seen from Eq.~(\ref{eq:Boltzmann}), but is not limited to this range of $g$.
We will assume, and then check a posteriori, that $R(g)$ is a smooth function of $g$. This is an analog of the eikonal approximation in optics or the WKB approximation in quantum mechanics. In this approximation, for $|n'-n|\ll 1/\lambda$ we have
\begin{align}
\label{eq:local_approximation}
R(g_{n'}) &\approx R(g_n) + (g_{n'} - g_n) R'(g_n) \nonumber \\
& \approx R(g_n) + \lambda \omega(g_n)R'(g_n) (n'-n),
\end{align}
where $R'(g)\equiv dR/dg$; we have used here the semiclassical expression for the quasienergy level spacing in terms of the classical vibration frequency $\omega(g)$.
The approximation (\ref{eq:local_approximation}) was used to find the stationary distribution of resonantly and parametrically driven quantum oscillators \cite{Dykman1988a,Marthaler2006} from the balance equation for the state populations (\ref{eq:balance_equation}). As we show, it allows one to solve the balance equation for the considered here period tripling problem in a broad parameter range, but it can break down for $\bar n =0$. This is qualitatively different from the cases considered earlier.
We emphasize that Eq.~(\ref{eq:local_approximation}) assumes that the {\em logarithm} of the distribution $\rho_n$ smoothly depends on $n$, whereas the distribution itself does not have to be a smooth function of $n$ and can vary by a factor ${\cal O}(1)$ when $n$ is incremented by 1. This is similar to the WKB approximation, where the wave function is oscillating fast, but the wavelength changes only a little on the distance given by the wavelength itself.
From Eqs.~(\ref{eq:eikonal}) and (\ref{eq:local_approximation}), we see that $\rho_{n+m}/\rho_n = \exp[-m\omega(g_n)R'(g_n)]$ for $|m|\ll 1/\lambda$. This means that $\omega(g_n)R'(g_n)$ can be interpreted as the $g_n$--dependent inverse temperature of the Boltzmann-like distribution over the RWA energy levels $g_n$.
Equations (\ref{eq:balance_equation}) and (\ref{eq:local_approximation}) reduce the problem of finding the distribution over the excited intrawell states ($n\gg 1$) to an algebraic equation
\begin{align}
\sum_m W_{n+m\,n} (\xi_n^m-1 )=0, \quad \xi_n = e^{-R'(g_n)\omega(g_n)}.
\label{eq:equation_for_xi}
\end{align}
Here the hopping rates $W_{nn'}$ are given by the semiclassical expression (\ref{eq:semiclassical_hopping_rates}). As indicated above, for $|n-n'|\sim 1$ they can be calculated numerically from Eqs.~(\ref{eq:classical_eom}), whereas for large $|n-n'|$ we have the explicit expressions (\ref{eq:semiclassical_hopping_rates}) - (\ref{eq:a_m>0}). In writing Eq.~(\ref{eq:equation_for_xi}) we used $W_{n \,n+m}=W_{n-m \, n}$ for large $n\gg 1$ and $|m|\ll n$.
A crucial feature of Eq.~(\ref{eq:equation_for_xi}) is that it is {\it local} in the level number $n$. It assumes that the stationary population $\rho_n$ is formed by the transitions from a few states $n+m$ surrounding a given state $n$, with $|m|\ll n, 1/\lambda$. Formally, it requires that the sum over $m$ in Eq.~(\ref{eq:equation_for_xi}) converges. If this is the case, Eq.~(\ref{eq:equation_for_xi}) defines a smooth function $R'(g)$ which depends on the single parameter, the scaled amplitude of the driving field $f$. However, the applicability of this approximation is not known a priori. We note that Eq.~(\ref{eq:equation_for_xi}) does not contain the scaled Planck constant $\lambda$; in particular, $R'$ is independent of $\lambda$.
\subsubsection{The vicinity of the minima of the Hamiltonian function}
Typtically, the WKB approximation applies for highly excited states that correspond to large $n$ in Eq.~(\ref{eq:equation_for_xi}). In the considered case the situation is different. Because near the minima of the quasienergy surface the dynamics in the rotating frame can be mapped onto the dynamics of an auxiliary weakly nonlinear oscillator, cf. Eq.~(\ref{eq:g_harmonic}), the semiclassical approximation works for $g$ approaching $g_{\rm min}$. One can show that for $g-g_{\min} \ll 1$, the semiclassical transition rates $W_{n+m\,n}$ in Eq.~(\ref{eq:semiclassical_hopping_rates}) scale as $W_{n+m\,n}\propto (g-g_{\rm min})^{|m|}$. If one sets to zero the rates $W_{n+m\,n}$ with $|m|\geq 2$, Eq.~(\ref{eq:equation_for_xi}) gives the same result for the distribution as in Eq.~(\ref{eq:Boltzmann}), with $R'(g_{\rm min}) \omega(g_{\rm min})\to \lambda \omega_{\rm min}/T_{\rm eff}$ for $g\to g_{\min}$. Clearly, the relation $W_{n\,n+m} = W_{n-m\,n}$ does not apply for small $n\sim 1$, but the range of small $n$ corresponds to a very narrow range of $g$ for small $\lambda$.
In Fig.~\ref{fig:harmonic_R'} we show the result of solving Eq.~(\ref{eq:equation_for_xi}) for $g$ close to $g_{\rm min}$. It matches the solution (\ref{eq:Boltzmann}) obtained in the harmonic approximation well in a broad parameter range. However, it also shows that the transitions with $|m|>1$ lead to an occupation of the excited intrawell states for $\bar n=0$ even for the value of the driving amplitude $f$ where, in the harmonic approximation, such states remain unoccupied. This shows that $T_{\rm eff}$ becomes nonzero if the transitions with $|m|>1$ are taken into account.
\subsubsection{Classical limit}
For large Planck numbers $\bar n$, Eq.~(\ref{eq:equation_for_xi}) greatly simplifies, since $\omega(g) R'(g) \ll 1$. One can then expand $\exp[-m\omega(g)R']$ in a series to second order in $R'$. Then, from Eqs.~(\ref{eq:semiclassical_hopping_rates}) and (\ref{eq:equation_for_xi}),
\begin{align}
\label{eq:small_R_prime}
R' (g) &= \frac{2\sum_m |a_{-m}(g)|^2 m}{\omega(g)(2\bar n +1) \sum_m |a_{-m}(g)|^2 m^2}
\end{align}
Using that the matrix elements $a_m(g)$ are the Fourier components of the dynamical variable $a(\tau;g)\propto Q(\tau;g)+iP(\tau;g)$ calculated for a classical orbit (\ref{eq:classical_eom}) with a given $g$, the sums in Eq.~(\ref{eq:small_R_prime}) can be written as integrals over the area enclosed by this orbit in the phase space~\cite{goldstein2001}. Then Eq.~(\ref{eq:small_R_prime}) can be rewritten as,
\begin{align}
\label{eq:small_R_prime_simplified}
&R'(g)= \frac{2 M(g)}{(2\bar n +1)N(g)},\quad M(g) = \iint dQdP, \nonumber \\
& \,\, N(g) = \frac{1}{2} \iint {\bm\nabla}^2 g(Q,P) dQdP.
\end{align}
As indicated above, the integrals here are taken over the interior of the region $g(Q,P)=g$ within a well of $g(Q,P)$ and ${\bm\nabla}^2 g(Q,P) \equiv \partial^2_Q g + \partial^2_P g = 4(P^2+Q^2)-2\,{\rm sgn}(\delta\omega)$.
In the high temperature range $k_B T \gg \hbar \omega_0$, $R'$ is inversely proportional to $T$. The distribution over intrawell states goes over into the classical (but non-Boltzmann) distribution formed as a result of thermal-noise-induced diffusion over the RWA-energy.
Equation~(\ref{eq:small_R_prime}) applies for an arbitrary temperature near the threshold of the period tripling, but not too close to the threshold, so that the condition $\omega(g_{\min})\gg \kappa$ still holds. Being near the threshold means $f\ll 1$ for $\delta\omega>0$ and $f-2\ll 1$ for $\delta\omega<0$, to zeroth order in $\kappa$. Near the threshold, the vibration frequency $\omega (g) \ll 1$ for the whole range of the intrawell values of $g$. Indeed, near the minimum of $g(Q,P)$ we have $\omega_{\rm min} \approx(6f)^{1/2} \ll 1$ for $\delta\omega>0$ and $ \omega_{\rm min} \approx 2\sqrt{3}(f-2)^{1/4} \ll 1$ for $\delta\omega<0$. For larger $g$, $\omega(g)$ monotonically decreases and approaches zero as $g$ approaches $g_s$.
We compare in Appendix~\ref{sec:classical_compare} the result of the classical limit given by Eq.~(\ref{eq:small_R_prime}) with the semiclassical calculation based on Eq.~(\ref{eq:equation_for_xi}). The classical results match well the semiclassical ones already for moderately small $\bar n \sim 1$
\section{The breakdown of the local approximation}
\label{sec:non_locality}
The eikonal approximation relies on the stationary distribution being formed locally by transitions between a few nearest states. The number of the states involved should be much smaller than $1/\lambda$, otherwise the expansion of the logarithm of the probability distribution (\ref{eq:local_approximation}) does not apply. As a consequence, the polynomial equation for $R'$ (\ref{eq:equation_for_xi}), which is a direct analog of the equation for the derivative of the logarithm of the wave function in the WKB approximation, does not apply either. It turns out that, for $\bar n =0$, the locality, and thus the eikonal approximation, can break down for the oscillator that displays period tripling. In this section, we study the condition of such a breakdown and show how to go beyond the local approximation. We focus on the case of a positive detuning $\delta\omega>0$.
\subsection{The locality condition and its breakdown for zero temperature}
The stationarity of the intrawell state populations $\rho_n$ means that the probability flux out of a state $n$ (the rate of leaving the state) and the probability flux into the state are equal. We can write the balance equation (\ref{eq:balance_equation}) in terms of the relative incoming and outgoing fluxes as
\begin{align}
\label{eq:fluxes_introduced}
&\frac{d}{d\tau} \log\rho_n = \sum_m\left[{\cal J}^{\rm in}_m(g_n)-{\cal J}^{\rm out}_m(g_n)\right], \nonumber\\
&{\cal J}^{\rm in}_m(g_n)=W_{n+m\,n}\frac{\rho_{n+m}}{\rho_n}, \quad {\cal J}^{\rm out}_m(g_n)=W_{n\,n+m}.
\end{align}
The exponential decay of the rate $W_{n\,n+m}\propto \exp[-2|m|\omega(g_n)\tau_\infty(g_n)]$ for large $|m|$ guarantees that the total relative outgoing flux $\sum_m{\cal J}^{\rm out}_m$ is always finite. Since the stationary state populations $\rho_n$ decrease with the increasing $n$, that is, $\rho_{n+m}/\rho_n <1$ for $m>0$, the relative incoming flux from the states with larger RWA energy $\sum_{m>0}{\cal J}^{\rm in}_m(g_n)$ is also finite. Moreover, it is local, only a few nearest states contribute to this flux.
In contrast, but for the same reason, the relative flux ${\cal J}^{\rm in}_{m<0}$ into a given state from the states with lower RWA energy can be non-local. If the increase of the population $\rho_{n-|m|}$ with the increasing $|m|$ is faster than the decrease of the transition rate $W_{n-|m|\,n}$, the incoming relative flux ${\cal J}^{\rm in}_{m<0}$ increases with the increasing distance between the states $|m|$.
If we use the eikonal approximation for the state populations, Eq.~(\ref{eq:local_approximation}), we find that, for large $|m|$ (but $|m|\ll 1/\lambda$) and for a given $n$, ${\cal J}^{\rm in}_{m<0}\propto \exp[-|m|\omega (2\tau_\infty-R')]$. Thus, the locality condition for a given $n$ is
\begin{align}
\label{eq:locality_condition}
\lambda \ll \omega(g_n)[2\tau_\infty(g_n)-R'(g_n)].
\end{align}
The right-hand side of this inequality gives the fall-off rate of the relative influx with the distance to the lower-lying state, in the eikonal approximation. The left-hand side is proportional to the reciprocal number of states in the well of $g(Q,P)$, and thus puts a limit on the number of the lower-lying states. When the condition (\ref{eq:locality_condition}) is not met, the eikonal approximation (\ref{eq:local_approximation}) does not allow one to find the stationary probability distribution over the intrawell states. One needs to use the full balance equation (\ref{eq:balance_equation}) (we remind that, in fact, we are discussing the quasi-stationary distribution, as we have so far disregarded interwell transitions). Also, one should not use Eq.~(\ref{eq:semiclassical_hopping_rates}) to calculate the transition rates, but rather the full WKB wavefunctions or the Wannier-type functions to calculate the relevant matrix elements.
We start with analyzing the locality breakdown in the semiclassical limit $\lambda \rightarrow 0$. The function $\omega(g) [2\tau_\infty(g)-R'(g)]$ calculated in the ``local'' approximation (\ref{eq:equation_for_xi}) is shown in Fig.~\ref{fig:breakdown_locality}. For relatively large (purple line) or small (red line) values of the scaled driving field $f$, function $2\tau_\infty(g)-R'(g)$ remains positive for all intrawell values of $g$, from $g = g_{\rm min}$ to $g= g_s$. Thus, the locality holds. However, for $f\sim 1$, function $2\tau_\infty(g)-R'(g)$ becomes equal to zero at some $g=g_{\rm NL}$ as seen from the data shown by the blue and green lines in Fig.~\ref{fig:breakdown_locality}. For $g>g_{\rm NL}$, Eq.~(\ref{eq:equation_for_xi}) does not have a nontrivial solution for $R'$, signaling the breakdown of the locality.
\begin{figure}[ht]
\includegraphics[width =5.5 cm]{./{breakdown_locality_v4}.pdf} \hfill
\caption{The rate $\omega(g) [2\tau_\infty(g)-R'(g)]$ of the fall-off of the relative probability flux ${\cal J}^{\rm in}_{-|m|}(g)$ [in Eq.~(\ref{eq:fluxes_introduced})] into the intrawell state with quasienergy $g$ from lower-lying states, for different $\Delta g = (g-g_{\min})/(g_s-g_{\min})$. The lines refer to the scaled driving amplitude $f$ = 0.1 (red), 0.25 (blue), $0.5$ (green), and 2 (purple). As the RWA energy $g$ approaches $g_{\rm min}$, one can show that $\omega(g) [2\tau_\infty(g)-R'(g)]$ diverges $\propto |\ln(g-g_{\rm min})|$.
}
\label{fig:breakdown_locality}
\end{figure}
We denote the minimum value of $g$ where the locality breaks down by $g_{\rm NL}$. In Fig.~\ref{fig:gNL} we show the dependence of $g_{\rm NL}$ on the scaled driving amplitude $f$. As seen from the figure, $g_{\rm NL}$ first emerges with the increasing $f$ at the saddle point value $g_s$ of $g(Q,P)$ for $f\approx 0.2$. It then moves deeper into the well of $g(Q,P)$. The distance $\Delta g_{\rm NL}$ from $g_{\rm NL}$ to the bottom of the well scaled by the depth of the well $g_s-g_{\min}$ is minimal for $f \approx 0.5$. As $f$ further increases, $g_{\rm NL}$ goes back to $g_s$. For $f\gtrsim 1.4$ the nonlocality disappears.
\begin{figure}[h]
\includegraphics[width =5.5 cm]{./{gNL_v2}.pdf}
\caption{The minimal scaled RWA energy $\Delta g_{\rm NL}=(g_{\rm NL}-g_{\min})/(g_s-g_{\min})$ where the locality breaks down as a function of the scaled driving amplitude $f$ for $\delta\omega>0.$}
\label{fig:gNL}
\end{figure}
To find accurately the location of $g_{\rm NL}$ from Eq.~(\ref{eq:equation_for_xi}), it is important to take into account the large-$|m|$ contribution to the influx ${\cal J}^{\rm in}_{m<0}$. This is because for $g$ close to $g_{\rm NL}$, ${\cal J}^{\rm in}_{m<0}(g)$ falls off with $|m|$ rather slowly, with the fall-off exponent $\omega(g) [2\tau_\infty(g)-R'(g)]$ close to zero as can be seen in Fig.~\ref{fig:breakdown_locality}. The large-$|m|$ contribution can be calculated using the asymptotic expression for the transition rates, Eqs.~(\ref{eq:hopping_rates}) and (\ref{eq:a_m>0}).
In the region of nonlocality we have $R'(g)>2\tau_\infty(g)$. The stationary probability distribution in this region is found numerically and is discussed in Sec.~\ref{subsec:numerical_Wannier}.
\subsection{Effect of nonzero temperature}
The very existence of the non-locality in the transitions among intrawell states depends sensitively on the interrelation between the temperature and the scaled Planck constant $\lambda$. In the limit $\lambda \rightarrow 0$ the locality is formally preserved at any nonzero temperature. This is seen from the expression (\ref{eq:semiclassical_hopping_rates}) for the transition rates $W_{nn'}$. At nonzero temperatures, the second term in $W_{nn'}$, which comes from the absorption of energy from the thermal reservoir, becomes non-zero. Although this term is small for low temperatures, it can significantly change the distribution. We now provide analytical arguments to show that this is the case. The numerical results are given in the next section.
The relative influx into the state with a given RWA energy $g_n$ can be written as a sum of the contribution from nearby states [small $|m|$ in Eq.~(\ref{eq:fluxes_introduced})] and that from the states that are further away (large $|m|$). We will use that, where the locality holds, $\rho_{n+m}/\rho_n = \exp[-mR'(g_n)\omega(g_n)]$ and consider the ``most dangerous'' situation where $R'\approx 2\tau_\infty$ and therefore the contribution of the remote states is significant. Using Eqs.~(\ref{eq:a_m<0}) and (\ref{eq:a_m>0}) we find
\begin{align}
\label{eq:influx_sum}
&\sum_{m<0} {\cal J}^{\rm in}_{m} \approx (\bar n+1) C_1 \Gamma(-1/3)[\omega(2\tau_\infty-R')]^{1/3} \nonumber\\
&+ \bar n C_2 \Gamma(1/3)[\omega(2\tau_\infty-R')]^{-1/3} + C_0.
\end{align}
Here $C_1,C_2$ are constants independent of $R'$, whereas the parameter $C_0$ comes from the small-$|m|$ terms in Eq.~(\ref{eq:fluxes_introduced}) and depends on $2\tau_\infty-R'$ in a nonsingular way.
As seen from Eq.~(\ref{eq:influx_sum}), for $R'(g) \to 2\tau_\infty(g)$ the part of the relative influx, which is proportional to $\bar n$, diverges. The other terms in Eq.~(\ref{eq:fluxes_introduced}) remain finite. Therefore, for any nonzero $\bar n$, the``local'' balance equation (\ref{eq:equation_for_xi}) has a nontrivial stationary solution for $R'(g)$ with $R'(g) < 2\tau_\infty(g)$, as the term $\propto \bar n$ can compensate other terms. In other words, the locality is preserved in the limit $\lambda \rightarrow 0$ but $\bar n \neq 0$. However, as we have seen, the locality can break down for $\bar n=0$.
Where $R'(g_n)$ approaches $2\tau_\infty(g_n)$, the number of states that contribute to the influx into the state $n$ increases $\propto [2\tau_\infty(g_n)-R'(g_n)]^{-1}$. For $\bar n\to 0$ this number diverges. Physically, the total number of states in a well of $g(Q,P)$ is always finite. It is $\propto 1/\lambda$. Therefore for a nonzero $\lambda$, the transition from the non-locality at $\bar n=0$ to the locality at $\bar n \neq 0$ is smeared out. The ``non-locality'' means that essentially all states with $g_m<g_n$ contribute to the stationary population $\rho_n$, whereas the ``locality'' means that the number of such states is much smaller. For a small nonzero $\bar n$, the number of the contributing states is given by the condition that the term $\propto \bar n [2\tau_\infty(g)-R'(g)]^{-1/3}$ in Eq.~(\ref{eq:influx_sum}) is of the same order as other terms in the balance equation. The number of the contributing states $\sim [2\tau_\infty(g)-R'(g)]^{-1}$ is thus $\propto \bar n^{-3}$. Therefore the locality holds where $\bar n^3 \gtrsim \lambda$.
We note that there is a profound difference between the non-locality for period tripling discussed here and that for an oscillator driven near its eigenfrequency~\cite{Guo2013,Peano2014} or parametrically modulated near its second overtone~\cite{Marthaler2006}. There, the locality holds for $\bar n =0$. The non-locality emerges only in a narrow range of nonzero $\bar n$ due to the asymmetry of the decay of the semiclassical matrix elements $a_{m>0}$ and $a_{m<0}$ with $|m|$. In contrast, in the period-tripling case, the semiclassical matrix elements $a_{m>0}$ and $a_{m<0}$ decay with $|m|$ with the same exponents, albeit with different prefactors. The non-locality then occurs at zero temperature.
Another qualitative distinction is that, for period tripling, the logarithm of the probability distribution $\log\rho_n=-R(g_n)/\lambda$ remains a smooth function of $g_n$. The function $R(g)$ can still be expanded in a series, it is just that keeping the first term in this expansion is not sufficient for finding the probability distribution where the locality breaks down.
\subsection{Numerical analysis using the Wannier functions}
\label{subsec:numerical_Wannier}
To confirm the breakdown of the local approximation, we numerically find the Wannier-type intrawell wave functions defined by Eqs.~(\ref{eq:nth_wave_functions}) and (\ref{eq:Wannier_defined}), then calculate the transition rates using the matrix elements of the operators $a,a^\dagger$ on these wave functions, and then solve the balance equation~(\ref{eq:balance_equation}). This method does not require the local approximation, but it significantly relies on the scaled Planck constant $\lambda$ being nonzero. An alternative approach is to calculate the matrix elements using the Landau method \cite{landau1977} for states with significantly different quantum numbers. The Landau method was used \cite{Guo2013} in the problem of an oscillator driven close to its eigenfrequency, where the singularities of the classical trajectories (\ref{eq:classical_eom}) were simple poles periodically located on the complex-time plane. In the present case the trajectories have multiple branching points, which significantly complicates the application of the method.
We compare in Fig.~\ref{fig:compare_Wannier} the inverse temperature $R'(g)$ obtained using the Wannier-type functions with that obtained from Eq.~(\ref{eq:equation_for_xi}).
For a broad range of parameters, the inverse temperatures $R'(g_n)$ obtained using the two methods coincide. When $2\tau_\infty$ becomes close to $R'(g_n)$, the result obtained from Eq.~(\ref{eq:equation_for_xi}) deviates from the Wannier-functions method, indicating the breakdown of the local approximation that underlies Eq.~(\ref{eq:equation_for_xi}).
\begin{figure}[ht]
\includegraphics[width=5cm]{./{compare_Wannier_v2}.pdf}
\caption{The scaled inverse temperature $R'(g)$ of the intrawell probability distribution as a function of the RWA energy $g$, $\Delta g = (g-g_s)/(g_s-g_{\min})$.
The scaled driving amplitude is $f=0.5$. The dots refer to $R'$ obtained by solving the balance equation~(\ref{eq:balance_equation}) using the Wannier functions, $R'(g_n) = -\lambda \log (\rho_{n+1}/\rho_{n-1})/(g_{n+1}-g_{n-1})$. The scaled Planck constant is $\lambda=0.004$. There are approximately 50 levels in each well. The solid lines refer to $R'$ obtained from Eq.~(\ref{eq:equation_for_xi}), which does not contain $\lambda$. The black dashed line shows $2\tau_{\infty}(g)$. At $\bar n = 0$, $R'(g)$ obtained using Eq.~(\ref{eq:equation_for_xi}) (red solid line) becomes equal to $2\tau_\infty(g)$ for $g = g_{\rm NL}$. For larger $g$, where $R'(g)>2\tau_\infty(g)$, the local approximation breaks down, Eq.~(\ref{eq:equation_for_xi}) no longer applies, and this is why the red line terminates at $g_{\rm NL}$. The difference between the dots and the solid lines of the same color is due to $\lambda$ being nonzero and is most significant when $R'$ becomes close to $2\tau_\infty$
}
\label{fig:compare_Wannier}
\end{figure}
To visualize the nonlocality, we show in Fig.~\ref{fig:influx} the normalized probability flux
\begin{align}
\label{eq:normalized_flux}
\overline{\rho_{n'} W_{n'n}}=\rho_{n'} W_{n'n}/\rm {max}_{n'} (\rho_{n'} W_{n'n})
\end{align}
from the state $\Ket{n'}$ to the state $\Ket{n}$ inside a well of $g(Q,P)$ in the stationary regime. The normalization factor is the maximal over $n'$ value of the flux $\rho_{n'} W_{n'n}$. The flux is obtained from Eq.~(\ref{eq:balance_equation}) using the transition rates calculated with the Wannier-type wave functions (\ref{eq:Wannier_defined}). At zero temperature (left panels), the fall-off of the flux (\ref{eq:normalized_flux}) with the increasing interstate distance $n-n'$ becomes strongly non-exponential as the RWA energy of the state into which the transition occurs $g_n$ approaches $g_{\rm NL}$. For relatively large values of $g_n$, the flux $\rho_{n'} W_{n'n}$ becomes nonmonotonic in $n-n'$. This is a manifestation of the nonlocality: a significant portion of the probability flux into a high-lying state $\Ket{n}$ comes from the states much deeper in the well, because the increase of their population with $n-n'$ is faster than the fall-off of the transition rate $W_{n'n}$. Moreover, the flux from the states that are further away can be larger than from at least some states closer to a given state $n$. At finite temperatures (right panels), such behavior goes away and $\rho_{n'} W_{n'n}$ falls off exponentially as $|n-n'|$ increases.
\begin{figure}[ht]
\includegraphics[width=4.2cm]{./{influx_matrix_a}.jpg}\hfill
\includegraphics[width=4.2cm]{./{influx_matrix_b}.jpg} \\
\includegraphics[width=4cm]{./{influx_cut_a}.pdf} \hfill
\includegraphics[width=4cm]{./{influx_cut_b}.pdf}
\caption{Top panels: The normalized stationary probability flux $\Ket{n'}\to \Ket{n}$ given by the function $\overline{\rho_{n'} W_{n'n}}$, Eq.~(\ref{eq:normalized_flux}). $\Delta g_n = (g_n-g_s)/(g_s-g_{\rm min})$. The left and right panels refer to $\bar n = 0$ and $\bar n = 0.05$, respectively. The scaled drive amplitude and the dimensionless Planck constant are $f=0.5$ and $\lambda = 0.004$ in the both panels, so that the number of levels in a well of $g(Q,P)$ is $\approx 50$. Bottom panels: the normalized flux $\overline{\rho_{n'} W_{n'n}}$ as a function of $n'$ for the three values of $n$ that correspond to the vertical red, black and blue cuts in the corresponding top panels. The dots are the cross-sections of the flux shown in the upper panels along the cuts, and the color coding of the dots corresponds to the color coding of the cuts. The left lower panel shows that, for $\bar n=0$, the decay of $\overline{\rho_{n'} W_{n'n}}$ with the increasing interstate distance slows down for larger $g_n$ and even becomes non-monotonic (see the blue dots). The right lower panel shows that,
as the temperature increases, the decay of the flux $\overline{\rho_{n'} W_{n'n}}$ becomes exponential in $|g_n-g_{n'}|$, indicating that the locality has been recovered.}
\label{fig:influx}
\end{figure}
\section{Escape from the period-three states of a dissipative oscillator}
\label{sec:tunneling}
In the discussions in Secs.~\ref{sec:distribution} and~\ref{sec:non_locality}, we have disregarded tunneling between the states localized in different wells of $g(Q,P)$. As explained in Sec.~\ref{subsec:intrawell}, the tunneling causes splitting of the intrawell levels into triplets with the level spacing $\sim |J_{0\pm}|$, see Eq.~(\ref{eq:nth_wave_functions}). If the scaled relaxation rate $\kappa$ is smaller than $|J_{0\pm}|/\lambda$, dissipation leads to transitions within the triplets and between different triplets, and ultimately there is formed a stationary distribution over the broadened tunnel-split Floquet states of the oscillator. However, for small $\lambda$ the level splitting $\sim |J_{0\pm}(g_n)|$ is exponentially small. Therefore such picture is relevant provided $\kappa$ is also exponentially small.
In this section we focus on the parameter regime where the tunnel splitting is small compared to the dissipative level broadening, $ |J_{0\pm}|\ll \lambda\kappa $. In this case coherent tunneling is replaced by an interwell hopping (switching) with a rate $\propto |J_{0\pm}|^2/\lambda^2\kappa$, an analog of the quantum diffusion in solids \cite{Kagan1992}. Since the intrawell states in the rotating frame correspond to the vibrations at frequency $\omega_F/3$ with different phases, such interwell switching corresponds to phase-flip transitions in the laboratory frame.
If one disregarded the population of the excited intrawell states, the switching would occur in the lowest state. The dissipation-induced transitions between the intrawell states significantly complicate the picture. The distribution over the intrawell states is formed fast, over the dimensionless time $\kappa^{-1} \ll \lambda^2\kappa/|J_{0\pm}|^2$. The populations of the excited intrawell states $\rho_n$ fall off exponentially with the increasing $g_n$, whereas the tunneling matrix elements $|J_{0\pm}(g_n)|$ exponentially increase with $g_n$. The rate of the interwell hopping is then determined by some optimal $g_n$ where $\rho_n|J_{0\pm}(g_n)|^2$ has a sharp maximum. Such competition of the exponential factors is similar to that in systems in thermal equilibrium, where the intrawell distribution is of the Boltzmann form \cite{Larkin1985}.
As we show below, in our case the optimal value of the RWA energy for interwell switching corresponds to the height of the barrier between the wells of $g(Q,P)$, i.e., the saddle-point value $g_s$. In other words, escape from a well of $g(Q,P)$ occurs via an over-barrier transition. This is an analog of thermal activation in systems in thermal equilibrium and is called quantum activation, since it is due to quantum fluctuations and occurs even for $T=0$. For quantum oscillators driven close to their eigenfrequency or parametrically modulated close to twice the eigenfrequency, quantum activation was found earlier \cite{Dykman1988a,Marthaler2006}. These systems have detailed balance for $T=0$ \cite{Drummond1980,Kryuchkyan1996}, whereas our system does not, and therefore the nature of the escape can be expected to be different.
We will focus on the case of positive detuning, $\delta\omega>0.$ For negative detuning, $\delta\omega<0$, along with interwell tunneling there also occurs tunneling between the intrawell states and the state localized near $Q=P=0$.
\subsection{The condition for the onset of quantum activation}
\label{subsec:onset_quantum_activation}
For small $\lambda$ the tunneling matrix element $J_{0\pm}(g_n)$ can be calculated in the WKB approximation. The problem is somewhat different from the problem of tunneling of a particle in a potential well, where the momentum as a function of coordinate and energy has only two values and is either real or imaginary. In the present case, the classical momentum $P(Q,g)$ defined by the equation $g(Q,P)=g$ has 4 branches as a function of the coordinate and is complex rather than purely imaginary in the classically inaccessible region. As a result the WKB wave function is not just decaying, but also oscillating in this region. Nevertheless, extending the analysis for the lowest intrawell state \cite{Zhang2017}, one can show that
\begin{align}
\label{eq:tunnel_matrix_element}
&|J_{0\pm}(g)|\propto \omega(g) \exp[-S_{\rm tun}(g)/\lambda], \nonumber \\
&S_{\rm tun} (g) = {\rm Im} \int dQ P(Q,g).
\end{align}
Here, $S_{\rm tun}(g)$ is the imaginary part of the complex classical action calculated along the Hamiltonian trajectory (\ref{eq:classical_eom}) that goes in complex time from one well of $g(Q,P)$ to another well with the same RWA energy $g$. One has to choose the trajectory with the minimal $S_{\rm tun}$.
To find the optimal value of the RWA energy $g$ for interwell hopping, we will use a quasicontinuous approximation for the intrawell distribution and replace $\rho_n=\exp[-R(g_n)/\lambda]$ with $\rho(g) = \exp[-R(g)/\lambda]$. In this approximation the rate of interwell hopping with the RWA energy between $g$ and $g+dg$ is $\propto dg\,\exp\{-[R(g) + 2S_{\rm tun}(g)]/\lambda\}$. The optimal $g$ is determined by the extremum of the exponent.
If escape from a well of $g(Q,P)$ occurs via quantum activation, it means that the amplitude of the exponent decreases with the increasing $g$ for all values of $g$ inside the well of $g(Q,P)$,
\begin{align}
\label{eq:q_activation_condition}
\partial_g R (g) + 2\partial_g S_{\rm tun}(g) < 0\quad {\rm for }\quad g_{\min} \leq g \leq g_s.
\end{align}
\begin{figure}[t]
\includegraphics[width =5.5 cm]{./{tunneling_time_v2}.pdf}
\caption{The imaginary tunneling time $|\tau_{\rm tun}(g)|$ (solid lines) and the limiting steepness of the exponent of the intrawell distribution $\tau_\infty(g)$ (dashed lines) as functions of $g$, $\Delta g=(g-g_{\min})/(g_s- g_{\min})$. The red, black, and green curves refer to the scaled driving amplitude $f =$ 0.1, 0.5, and 2, $\delta\omega>0.$ }
\label{fig:compare_tau_tun}
\end{figure}
We found in Sec.~\ref{sec:non_locality} that, where the intrawell distribution is formed by transitions between comparatively close states (the locality condition), $R'(g_n)$ is limited by the decrement $2\tau_\infty(g_n)$ of the decay of the intrawell transition rates $W_{nn'}$ with $|n-n'|$, i.e., $0<R'(g)\leq 2\tau_\infty(g)$. This limits the first term in Eq.~(\ref{eq:q_activation_condition}).
The second term in Eq.~(\ref{eq:q_activation_condition}) is negative and is determined by the imaginary part of the time of moving along a Hamiltonian trajectory (\ref{eq:classical_eom}) through the classically inaccessible region between the wells,
\begin{align}
\label{eq:tunnel_time_general}
\partial_g S_{\rm tun} = \tau_{\rm tun}(g) = {\rm Im}\,\int dQ[\partial_Pg(Q,P)]^{-1}.
\end{align}
The derivative $\partial_Pg$ is calculated here for $P=P(Q,g)$. The details of the calculation are given in Appendix~\ref{sec:imaginary_tunneling_time}. We note that the calculation is done differently in different regions of $g$ where the trajectories in the classically allowed region are ellipse-like or horse-shoe-like.
We show in Fig.~\ref{fig:compare_tau_tun} the imaginary tunneling time $\tau_{\rm tun}$ for various values of the driving amplitude. For $g$ close to $g_{\rm min}$, $\tau_{\rm tun}$ diverges logarithmically,
\[
\tau_{\rm tun}\sim \omega_{\min}^{-1}\log(g-g_{\rm min}).
\]
For $g$ close to $g_s$, one can show that $\tau_{\rm tun}$ is linear in $g_s-g$ and is finite at $g=g_s$; see Appendix~\ref{sec:imaginary_tunneling_time}.
Also shown in Fig.~\ref{fig:compare_tau_tun} is the function $\tau_\infty(g)$, which limits $R'(g)/2$ in the region of locality. As seen from Fig.~\ref{fig:compare_tau_tun}, $|\tau_{\rm tun}|>\tau_\infty$, thus guaranteeing that the quantum-activation condition (\ref{eq:q_activation_condition}) holds true. Note that for $g$ close to $g_{\rm min}$, both $\tau_\infty$ and $\tau_{\rm tun}$ diverge logarithmically, but the prefactor of $\tau_{\rm tun}$ is twice as large as that of $\tau_\infty$ as can be seen from Eqs.~(\ref{eq:tao_infty_diverge}) and (\ref{eq:tunnel_time_general}).
In the range of the driving amplitude and the RWA energy where, for $\bar n=0$, the intrawell probability distribution is formed by transitions from remote states, we have $R'(g)>2\tau_\infty(g)$. We have found that the numerically calculated value of $R'$ in this range satisfies the condition (\ref{eq:q_activation_condition}). Therefore escape from a well of $g(Q,P)$ does occur via quantum activation.
\subsection{Quantum activation energy}
The rate of escaping from a well of $g(Q,P)$ in the rotating frame is the rate of escaping from a period-3 vibrational state in the laboratory frame. To logarithmic accuracy this rate is
\begin{align}
\label{eq:escape_rate}
W_{\rm esc}\sim \exp[-R_A/\lambda], \quad R_A = \int_{g_{\rm min}}^{g_s} dg R'(g).
\end{align}
The parameter $R_A$ is the quantum activation energy of escape. The quantum nature of the fluctuations leading to escape is clear from the fact that the escape rate displays activation dependence on $\hbar$, $\log W_{\rm esc}\propto \lambda^{-1}\propto \hbar^{-1}$. Yet, as we mentioned, the escape occurs via an overbarrier transition, as if it were thermally activated, with $\hbar$ playing the role of temperature.
We show in Fig.~\ref{fig:RA_semiclassical} the dependence of the activation energy $R_A$ on the scaled driving amplitude $f$. In contrast to the inverse effective temperature of the intrawell distribution near the minima of $g(Q,P)$, which is shown in Fig.~\ref{fig:harmonic_R'}, $R_A$ monotonically increases as $f$ increases. We present the results for $\bar n>\lambda^{1/3}$. In this case the intrawell distribution is formed by transitions $\Ket{n}\to \Ket{n'}$ with $|n-n'|\ll 1/\lambda$, and then $R_A$ is independent of $\lambda$. Already for the Planck number $\bar n\sim 1$, $R_A$ approaches the classical limit described by Eq.~(\ref{eq:small_R_prime}) and scales inversely proportional to $2\bar n +1$.
\begin{figure}[t!]
\includegraphics[width =5.5 cm]{./{RA_semiclassical}.pdf}
\caption{The semiclassical activation energy $R_A$ as a function of the scaled drive amplitude $f$. The black, blue, and green dots refer to $\bar n = 0.01, 0.1$ , and 1, respectively. The dashed line refers to the result of the classical limit obtained from Eq.~(\ref{eq:small_R_prime}). }
\label{fig:RA_semiclassical}
\end{figure}
An important comment is due here. In classical systems, in a fluctuation-induced escape from a metastable state, with overwhelming probability the system goes over the saddle point on the boundary of the attraction basin to the metastable state, cf. \cite{Freidlin1998}. In the case of the oscillator we are considering, as seen from the phase portrait in Appendix~\ref{sec:Langevin}, Fig.~\ref{fig:phase_portrait}, the classical switching occurs from a period-3 state to the zero-amplitude state, not to another period-3 state.
A quantum oscillator, however, can directly switch between the period-3 states. This is because the condition of small tunnel splitting compared to the level broadening can be violated near the saddle point, as the tunneling exponent $S_{\rm tun}$ falls off linearly with $g$ near the saddle-point value $g_s$, cf. Eq.~(\ref{eq:tunnel_time_general}) and Fig.~\ref{fig:compare_tau_tun}. If the tunnel splitting becomes comparable to the level broadening, the oscillator will tunnel between the wells once it has reached the values of $g$ close to $g_s$. The overall rate of interwell switching will still be close to that given by Eq.~(\ref{eq:escape_rate}), as this equation gives the rate of reaching $g_s$, to the logarithmic accuracy.
\subsection{Escape near the bifurcation point where the period-3 state emerges}
\label{subsec:bifurcation}
In the previous sections we considered the weak damping limit, where the intrawell level spacing is much larger than the level widths. If the level broadening is disregarded, the onset of the period-3 vibrational states has no threshold in the amplitude of the driving field for $\delta\omega>0$. In the presence of dissipation, the driving field must have a finite amplitude for the period-3 states to emerge. In terms of the theory of classical dynamical systems, the onset of period-3 vibrations is a bifurcation where the system acquires new stable states, in this case, period-3 states.
Generically, one of the dynamical variables of the system becomes slow near a bifurcation point (the bifurcation parameter values). This variable is an analog of a soft mode. It controls the overall dynamics, as other dynamical variables follow it adiabatically \cite{Guckenheimer1997}. In the case of a driven quantum oscillator this significantly simplifies the problem of fluctuations \cite{Dykman2007,*Dykman2012}. Indeed, since the oscillator dynamics in the rotating frame is Markovian, and since it is controlled by a single dynamical variable, the operator nature of this variable becomes irrelevant. The variable commutes with itself, moreover, it commutes with itself at different instants of time. The fluctuations are still quantum, their intensity is nonzero even for $T\to 0$, but otherwise the dynamics is fully classical.
Classical dynamics of an oscillator resonantly driven close to triple its eigenfrequency is well understood \cite{Nayfeh2004,jordan2007}. This dynamics is briefly summarized and the way to incorporate fluctuations consistent with the kinetic equation (\ref{eq:master_general}) is described in Appendix~\ref{sec:Langevin}. The important feature for the discussion here is that the period-3 states emerge as a result of three simultaneous saddle-node bifurcations occurring at three points $(Q_B^{(i)},P_B^{(i)})$ $(i=1,2,3)$ located at the vortices of an equilateral triangle on the $(Q,P)$-plane. At each point $(Q_B^{(i)},P_B^{(i)})$ there merge a stable and an unstable stationary state in the rotating frame, which correspond to the appropriate stable and unstable period-3 states in the laboratory frame.
If the parameter $\kappa$ is close to its bifurcational value $\kappa_B \equiv \kappa_B(f)$, one can expand the classical equations of motion near a point $(Q_B^{(i)},P_B^{(i)})$ in $\delta Q^{(i)}=Q-Q_B^{(i)},\; \delta P^{(i)}=P-P_B^{(i)}$ and rotate the variables so as to single out the slow variable $z=\delta Q^{(i)}\cos\phi_B^{(i)} + \delta P^{(i)}\sin\phi_B^{(i)}$. In the quantum Langevin equation for this variable one should keep the leading-order terms in $z, \kappa-\kappa_B$. This equation generically has the form \cite{Dykman2012}
\begin{align}
\label{eq:Langevin_soft_mode}
\frac{d}{d\tau}z = a_Bz^2 -b_B(\kappa-\kappa_B)+\xi_z(\tau),
\end{align}
where $\langle \xi_z(\tau)\xi_z(\tau')\rangle = \lambda\kappa(2\bar n+1)\delta(\tau-\tau')$ is the quantum noise, which is $\delta$-correlated in the slow time $\tau$ (such structure of the noise corresponds to the Markovian dynamics in slow time). The explicit expressions for $Q_B^{(i)}, P_B^{(i)}, \phi_B^{(i)}$, and the parameters $a_B, b_B$ are given in Appendix~\ref{sec:Langevin}. Here we note that $a_Bb_B<0$.
\begin{figure}
\includegraphics[width = 5.5cm]{./{bifurcation_diagram_v2}.pdf}
\caption{The interrelation between the amplitude and frequency of the drive where the period tripling emerges. The abscissa shows the detuning of the driving field frequency $\delta\omega$ divided by the oscillator relaxation rate in the unscaled time, $\kappa^{-1}=\delta\omega/\Gamma$. The ordinate shows the scaled squared amplitude of the driving field $\tilde f^2 = F_0^2/24\omega_0^2\gamma\Gamma$; the scaling is independent of the frequency detuning. Above the bifurcation curve the oscillator has three stable period-3 states and the stable zero-amplitude state; below this curve, only the zero-amplitude state is stable. The plot refers to $\gamma\,\delta\omega>0$.
}
\label{fig:bifurcation}
\end{figure}
For $\kappa <\kappa_B(f)$ the system (\ref{eq:Langevin_soft_mode}) has a stable and an unstable stationary state, if we disregard fluctuations. They are located at $\mp z_{\rm st}\,{\rm sgn} \,a_B$, where
\[z_{\rm st} = [b_B(\kappa-\kappa_B)/a_B]^{1/2}.\]
At $\kappa=\kappa_B(f)$ these states merge together, and for the scaled decay rate $\kappa>\kappa_B(f)$ the oscillator does not have period-3 states in the laboratory frame. Equivalently, period-3 states exist in the range of the scaled field amplitude where
\begin{align}
\label{eq:f_B_kappa}
f^2 >f_B^2(\kappa) = 2[(1+\kappa^2)^{1/2}-{\rm sgn}(\delta\omega)]
\end{align}
(see Apendix~\ref{sec:Langevin}). In Fig.~\ref{fig:bifurcation} we plot $\tilde f_B^2 = f_B^2/\kappa$; the parameter $\tilde f^2 = f^2/\kappa = F_0^2/24\omega_0\gamma\Gamma$ is advantageous as it does not depend on the frequency detuning of the drive from $3\omega_0$.
Fluctuations caused by the noise $\xi_z(t)$ in Eq.~(\ref{eq:Langevin_soft_mode}) can lead to escape from the dynamically stable state. Even though our system is far away from thermal equilibrium, the escape rate $W_{\rm esc}$ near the bifurcation point can be found in the same way as for a Brownian particle in a potential well. The dynamics of such a particle in the case where inertial effects can be disregarded (the overdamped limit) is described by the Langevin equation that has the same form as Eq.~(\ref{eq:Langevin_soft_mode}), except that the noise is thermal. Using the Kramers solution of the thermal-equilibrium problem \cite{Kramers1940} and taking into account the explicit form of the intensity of the quantum noise $\xi_z(t)$, we obtain
\begin{align}
\label{eq:escape_bif}
\ln W_{\rm esc}\approx -\frac{2}{3}\frac{|b_B (\kappa-\kappa_B)|^{3/2}}
{|a_B|^{1/2}\kappa_B\lambda(2\bar n+1)}.
\end{align}
The exponent of the escape rate scales with the distance to the bifurcation point as $|\kappa-\kappa_B|^{3/2}$ and scales with temperature as $(2\bar n+1)^{-1}$. In terms of the parameters of the driving field, near the bifurcation point we have $\kappa-\kappa_B \approx \Gamma[(\delta\omega)_B - \delta\omega]/(\delta\omega_B)^2\approx (f-f_B)/(df_B/d\kappa)$, where $(\delta\omega)_B$ is the bifurcational value of the frequency detuning and $f_B(\kappa)$ is given by Eq.~(\ref{eq:f_B_kappa}). The scaling (\ref{eq:escape_bif}) with the parameters of the driving field and with temperature holds also for a resonantly driven oscillator near a bifurcation point \cite{Dykman2007,*Dykman2012} and has been seen for such an oscillator in the experiment with Josephson junction oscillators \cite{Vijay2009,Vijay2012}.
We note that, even though the zero-amplitude state of the oscillator that displays period tripling is stable, for sufficiently strong driving quantum fluctuations lead to escape from this state into period-3 states. For a strong field, $f^2\gtrsim (1+\kappa^2)/\lambda(2\bar n+1)$, one can show that quantum fluctuations essentially wash away the zero-amplitude state. The full analysis of escape from the zero-amplitude state is beyond the scope of this paper.
\section{Summary and outlook}
This paper describes the dissipative quantum dynamics of a nonlinear oscillator resonantly driven into period tripling. The analysis is based on a minimalistic model. The scaled RWA Hamiltonian in the rotating frame has one parameter, the scaled amplitude of the drive $f$, the dissipation is described by one parameter, the scaled linear friction coefficient $\kappa$, and the only other parameter is the scaled Planck constant $\lambda\propto \hbar$. The model has allowed revealing generic features of the dissipative Floquet dynamics of a multi-state system. Of particular importance is that it revealed the features that remained hidden in the previously explored models.
For period tripling, the RWA Hamiltonian as a function of the phase-space variables $g(Q,P)$ has three symmetrically located wells. In the most interesting regime, the wells contain many bound states. The number of the states is $\propto 1/\lambda \gg 1$. Because of the three-fold symmetry, the intrawell states would be degenerate if there were no interwell tunneling. The tunneling splits the eigenvalues of $g$ into triplets.
Coupling to a thermal reservoir leads to transitions between the states. Where the transition rates ($\propto \kappa$) exceed the exponentially small tunnel splitting (divided by $\hbar$), the distribution over the states is formed in two stages. First, over the dimensionless time $\sim 1/\kappa$ there is formed a distribution over the intrawell states. Then over an exponentially longer time, there is formed a distribution over the wells.
Arguably, the most unexpected feature of the dynamics is that the intrawell distribution is formed by transitions between remote states within the well, i.e., by a nonlocal walk over the intrawell states. This is despite the corresponding transition rates being exponentially small. The effect is a consequence of a very steep decay of the distribution with the increasing RWA energy.
A broader implication of this effect is related to the fact that, even though the distribution is very steep, its logarithm is a smooth function of the number $n$ of an intrawell state [or, equivalently, of the eigenvalue $g_n$ of the RWA Hamiltonian]. A standard approach to the analysis of the distribution in this case is based on the eikonal approximation. Here, the logarithm of the distribution is expanded in the state number and the balance equation for the state populations is reduced to the equation for the derivative of the logarithm of the distribution over $g_n$, cf. \cite{Dykman1988a}. Such approach is a complete analog of the WKB approximation where the Schr\"odinger equation is reduced to the equation for the derivative of the logarithm of the wave function. It has been used to analyze numerous problems of chemical kinetics and population dynamics \cite{Kamenev2011,Assaf2017}, in which case $n$ stands for the number of species and often has several components that correspond to different species.
Usually, the WKB approximation in quantum mechanics breaks down locally near singularities of the wave front, cf.~\cite{landau1977,Berry1989}. A similar behavior was found for the distribution of the resonantly driven oscillator \cite{Guo2013,Peano2014}. In contrast, in the present case, the logarithm of the probability distribution remains smooth. However, the local approximation, that underlies the eikonal method, breaks down. Such behavior may be expected also in other types of physical systems away from thermal equilibrium as well as in the problems of chemical kinetics and population dynamics.
The other generic feature of the dynamics of period tripling is the lack of detailed balance. This contrasts the previous results \cite{Drummond1980,Kryuchkyan1996,Marthaler2006} for the oscillator driven close to its eigenfrequency or twice the eigenfrequency, where the detailed balance holds at $T=0$. At the technical level, the semiclassical dultiple singularities of the classical Hamiltonian trajectories in the complex time and the branching at the singularities. This significantly complicates the analysis, but also broadens the applicability of the results.
In common with the previously considered driven oscillators, the switching between the period-3 states occurs via transitions over the barrier of $g(Q,P)$ rather than via interwell tunneling, if the decay rate exceeds the tunnel splitting divided by $\hbar$. We have also found the scaling behavior of the switching rate with the distance to the bifurcation point where the period-3 states disappear.
We have not provided the results for the switching from the quiet state.
The rate of switching from this state is a peculiar and important problem on its own, starting with the very possibility to switch to any of the period-3 states. The switching rate can be made comparatively large in this case, while still remaining much smaller than the decay rate. A larger switching rate should grossly simplify the observation of the switching in the experiment, thus providing the means to study quantum-fluctuations induced switching where the detailed balance condition is strongly violated. Such an observation should be contrasted with the important experiment on the switching of a resonantly driven quantum oscillator in the regime where the detailed balance was effectively present \cite{Vijay2009}. A detailed analysis of the switching from the quiet state is left for the future work.
In conclusion, our results show that the dissipative dynamics of resonant period tripling provides new insights into the effects of dissipation and quantum fluctuations in Floquet systems. Experimentally, these effects can be studied with various nonlinear microwave and optical cavities and also with nanomechanical systems, where quantum regime has been already recently reached \cite{Satzinger2018,Chu2018}.
\acknowledgments
We are grateful to J. Ankerhold, S. M. Girvin and J. Gosner, who participated in the work at the early stage. We also acknowledge valuable discussions with C.~Bruder, N.~L\"orch and V.~Peano. Y.Z. was supported by the National Science Foundation (DMR-1609326). M.I.D. acknowledges partial support from the National Science Foundation (Grants No. DMR-1806473 and and CMMI-1661618).
|
\section{Introduction}\label{s:introd}
This paper concerns the relationship between two
notions of \emph{codimension-$1$ rectifiability} in the Heisenberg
group $(\mathbb{H}^n,d_{\mathbb{H}}) = (\mathbb{R}^{2n+1},\cdot,d_{\mathbb{H}})$, where
"$\cdot$" is the group product
\begin{displaymath}
(x_1,\ldots,x_{2n},t) \cdot (x_1',\ldots,x_{2n}',t') =
\left(\sum_{i=1}^{2n} x_i + x_i',t+t'+\tfrac{1}{2}\sum_{i=1}^n x_i
x_{n+i}'-x_{n+i}x_i'\right)\in \mathbb{R}^{2n}\times\mathbb{R},
\end{displaymath} and $d_{\mathbb{H}}$ is the Kor\'anyi distance
$d_{\mathbb{H}}(p,q) := \|q^{-1} \cdot p\|$ (with $\|(x,t)\| :=
\sqrt[4]{|x|^4 + 16t^{2}}$ for $(x,t)\in \mathbb{R}^{2n}\times
\mathbb{R}$). Metric notions in $\mathbb{H}^{n}$, notably Hausdorff measures, are defined using the metric $d_{\mathbb{H}}$. Metric notions in $\mathbb{R}^{n}$ are defined using the standard
Euclidean distance.
In $\mathbb{R}^{n}$, the notion of rectifiability can be defined in two
equivalent ways. For $0 < m < n$, an $\mathcal{H}^{m}$ measurable set $E
\subset \mathbb{R}^{n}$ is called \emph{$m$-rectifiable} if $\mathcal{H}^{m}$
almost all of $E$ can be covered by either \begin{enumerate}
\item countably many Lipschitz $m$-images, or
\item countably many Lipschitz $m$-graphs.
\end{enumerate}
Here, a \emph{Lipschitz $m$-image} means a Lipschitz image of
$\mathbb{R}^{m}$, while a \emph{Lipschitz $m$-graph} means a set of the
form $\{v + A(v) : v \in V\}$, where $V \subset \mathbb{R}^{n}$ is an
$m$-dimensional subspace, and $A \colon V \to V^{\perp}$ is a
Lipschitz map. The equivalence of "Lipschitz image rectifiability" and "Lipschitz graph rectifiability" is well-known. In particular, Lipschitz $m$-graphs are trivially Lipschitz $m$-images, since $v \mapsto v + A(v)$ is Lipschitz whenever $A$ is.
We then discuss the analogues of these notions in $\mathbb{H}^{n}$.
Recall first that $(\mathbb{H}^n,d_{\mathbb{H}})$ is a metric space of
Hausdorff dimension $2n+2$. A common notion of codimension-$1$
rectifiability (see \cite[Definition 4.33]{MR2836591}) is
\emph{intrinsic Lipschitz graph (iLG) rectifiability}:
an $\mathcal{H}^{2n + 1}$ measurable set $E \subset \mathbb{H}^{n}$ is called iLG rectifiable if $\mathcal{H}^{2n+1}$ almost all of $E$ can be covered by countably many
iLGs over \emph{vertical subgroups of codimension $1$}. Here, vertical subgroups refer to codimension-$1$ subspaces of $\mathbb{R}^{2n+1}$
containing the $t$-axis, cf.\ Section \ref{s:C1alpha}, while iLGs were introduced by Franchi, Serapioni, and Serra
Cassano \cite{FSS} in 2006. They are natural $\mathbb{H}^{n}$
counterparts of Lipschitz graphs in $\mathbb{R}^{n}$, see Definition
\ref{d:intrinsicGraphs}.
The notion of \emph{Lipschitz image (LI)
rectifiability} in $\mathbb{H}^n$ was first studied by
Pauls \cite{MR2048183} in 2004. An
$\mathcal{H}^{2n+1}$ measurable set $E \subset \mathbb{H}^n$ is called
LI rectifiable if $\mathcal{H}^{2n+1}$ almost all of $E$ can be
covered by Lipschitz images of closed subsets of codimension-$1$
vertical subgroups. All of these subgroups (for $n \geq 1$ fixed) are isometrically isomorphic to each other. If $n=1$, they are further isometrically isomorphic
to the \emph{parabolic plane}
\begin{displaymath} \Pi := (\mathbb{R}^{2},+,\| \cdot\| ), \quad \text{where} \quad \|(y,t)\| := \sqrt[4]{y^{4} + 16t^{2}}, \end{displaymath}
and if $n\geq
2$, they are
isometrically isomorphic to $(\mathbb{H}^{n-1}\times
\mathbb{R},\cdot_{\mathbb{H}^{n-1}\times \mathbb{R}},\|\cdot\|)$
with
\begin{equation}\label{eq:HeisxR}
((z,t),s)\cdot_{\mathbb{H}^{n-1}\times \mathbb{R}}
((z',t'),s')=\big(\big(z+z',t+t'+\tfrac{1}{2}\sum_{i=1}^{n-1}
(z_iz_{n-1+i}'-z_i'z_{n-1+i}) \big),s+s'\big)
\end{equation}
and
\begin{displaymath}
\|((z,t),s)\| = \sqrt[4]{|(z,s)|^4 + 16 t^2}.
\end{displaymath}
So, $E$ is LI rectifiable if and only if $\mathcal{H}^{2n+1}$ almost
all of $E$ can be covered by countably many Lipschitz images of
closed subsets of $\Pi$ (if $n=1$) or
$\mathbb{H}^{n-1}\times\mathbb{R}$ (if $n\geq 2$). The metric
induced by $\|\cdot\|$ in $\Pi$ is denoted $d_{\Pi}$, and in
$\mathbb{H}^{n-1}\times \mathbb{R}$ by $d_{\mathbb{H}^{n-1}\times
\mathbb{R}}$.
The connection between iLG and LI rectifiability in $\mathbb{H}^n$ is
poorly understood. It is neither known if (a) LIs of vertical subgroups are iLG
rectifiable, nor if (b) iLGs are LI rectifiable. It may appear surprising that question (b) is open: after all, to show that Lipschitz $m$-graphs in $\mathbb{R}^{n}$
are Lipschitz $m$-images, one only needed to observe that the graph map $v \mapsto v + A(v)$
is Lipschitz whenever $A$ is. In $\mathbb{H}^{n}$, this argument fails completely. We will discuss the matter further in a moment.
The purpose of this paper is to make progress in question (b). In brief, we will show intrinsic $C^{1,\alpha}$-graphs in $\mathbb{H}^{n}$ are LI rectifiable for any $\alpha > 0$. In $\mathbb{H}^{1}$, we can say something a little better. For precise statements, see Theorems \ref{mainIntro}, \ref{mainQualitative}, and \ref{mainIntroVertical}. Before formulating these new results in detail, we define our objects of study more carefully, and describe some previous work on the topic.
\begin{definition}\label{d:intrinsicGraphs} An \emph{intrinsic graph over the vertical subgroup $\mathbb{W}=\{x_1=0\}$} in $\mathbb{H}^n$ is a set of the form
\begin{equation}\label{d:ILG} S = \{w \cdot \varphi(w) : w \in \mathbb{W}\}, \end{equation}
where $\varphi \colon \mathbb{W} \to \mathbb{V} =: \{(x_1,0,\ldots,0) : x_1 \in
\mathbb{R}\}$ is an arbitrary function. The graph $S$ determines $\varphi$
uniquely. Further, $S$ is an intrinsic $L$-\textbf{Lipschitz}
graph ($L$-iLG) over $\mathbb{W}$ if it satisfies a \emph{cone condition}
of the form
\begin{displaymath} S \cap (p \cdot \mathcal{C}(\alpha)) = \{p\}, \qquad p \in S, \; 0 < \alpha < L^{-1}. \end{displaymath}
Here $\mathcal{C}(\alpha) := \{q \in \mathbb{H}^n : \|\pi_{\mathbb{W}}(p)\| \leq
\alpha \|\pi_{\mathbb{V}}(p)\|\}$, and $\pi_{\mathbb{W}} \colon \mathbb{H}^n \to \mathbb{W}$ and
$\pi_{\mathbb{V}} \colon \mathbb{H}^n \to \mathbb{V}$ are the \emph{vertical and
horizontal} projections induced by the splitting $\mathbb{H}^n = \mathbb{W} \cdot
\mathbb{V}$. If $S \subset \mathbb{H}^n$ is an ($L$-)iLG, the function $\varphi$
is called an ($L$-)\emph{intrinsic Lipschitz function}.
\end{definition} Now, viewing \eqref{d:ILG}, it is clear that an
iLG $S \subset \mathbb{H}^n$ has a "canonical" parametrisation by the
\emph{graph map} $\Phi \colon \mathbb{W} \to S$, defined by $\Phi(w) := w
\cdot \varphi(w)$. However:
\begin{itemize}
\item $\varphi \colon (\mathbb{W},d_{\mathbb{H}}) \to (\mathbb{V},d_{\mathbb{H}})$ is not always
a Lipschitz function, and \item the graph map $\Phi \colon
(\mathbb{W},d_{\mathbb{H}})\to (S,d_{\mathbb{H}})$ is "almost never" Lipschitz.
\end{itemize}
Regarding the first point, \cite[Example 3.3]{FSS} suggests that
$\varphi(0,x_2,t) = (1 + t^{1/2},0,0)$ is an intrinsic Lipschitz
function in $\mathbb{H}^1$, which is not a Lipschitz function. For
the second point, consider the constant function $\varphi(0,x_2,t)
\equiv (1,0,0)$. Then the graph map
\begin{displaymath} \Phi(0,x_2,t) = (0,x_2,t) \cdot (1,0,0) = (1,x_2,t - \tfrac{x_2}{2}) \end{displaymath}
parametrises the vertical plane $S = \mathbb{W}' =
\{x_1=1\}\subset\mathbb{H}^1$, a prototypical iLG. However, $\Phi$
is not Lipschitz on any open subset of $\mathbb{W}$, because the only
rectifiable curves on $\mathbb{W},\mathbb{W}'$ are the \emph{horizontal lines}
contained on $\mathbb{W},\mathbb{W}'$, and $\Phi$ sends the horizontal lines on
$\mathbb{W}$ to non-horizontal lines on $\mathbb{W}'$ (by \textbf{right}
translation).
In spite of these difficulties, the graph map is sometimes useful
for Lipschitz parametrising iLGs: if an iLG $S \subset \mathbb{H}^{1}$
has enough \emph{a priori} regularity, then
$\Phi$ can precomposed with something known as the
\emph{characteristic straightening map} $\Psi \colon \Pi \to \mathbb{W}$
in such a way that $\Phi \circ \Psi \colon \Pi \to S$ is locally
Lipschitz -- or even bilipschitz. The following theorem is due to
Cole and Pauls \cite{MR2247905} from 2006 (the addition of the
letters "bi" is due to Bigolin and Vittone \cite{MR2603594} from
2010):
\begin{thm}[Cole-Pauls, Bigolin-Vittone]\label{CPBV} Every non-characteristic point on a Euclidean $C^{1}$ surface $S \subset \mathbb{H}^1$
has a neighbourhood which is the bilipschitz image of an open
subset of $\Pi$. \end{thm} Both proofs reduce the problem to
intrinsic Lipschitz graphs $S = \Phi(\mathbb{W})\subset \mathbb{H}^1$, and the
Euclidean $C^{1}$-smoothness of $S$ then translates to properties
of the intrinsic Lipschitz function $\varphi \colon \mathbb{W} \to \mathbb{V}$.
The essential hypothesis is that $\varphi$ is a Euclidean
$C^{1}$-function, although it might suffice that $\varphi$ is Euclidean Lipschitz, viewed as a function $\mathbb{R}^{2} \to \mathbb{R}$.
The characteristic straightening map only plays a small (and rather implicit) role in this paper, see Lemma \ref{flagimpliesbilip}. So, we refer to the "Outline of proofs" section in \cite{MR3400438}, or the proof of \cite[Theorem 3.1]{MR2603594} for more details.
In brief, the regularity of $\varphi$ is, in Theorem \ref{CPBV}, required to control the regularity of $\Psi$,
and if $\varphi$ fails to be Lipschitz $\mathbb{R}^{2} \to \mathbb{R}$, the map $\Psi$ does not appear to be useable for the Lipschitz parametrisation problem.
In fact, Bigolin and Vittone in \cite{MR2603594} show that Theorem \ref{CPBV} can fail without the $C^{1}$-regularity assumption.
For $\tfrac{1}{2} < \alpha < 1$, they consider the intrinsic Lipschitz function
\begin{equation}\label{BVEx} \varphi(0,x_2,t) = \begin{cases} -\tfrac{t^{\alpha}}{1 - \alpha}, & \text{if } t \geq 0, \\
0, & \text{if } t < 0, \end{cases} \qquad (0,x_2,t) \in \mathbb{W}, \end{equation}
and its intrinsic graph $S = \Phi(\mathbb{W})$, which fails to be
Euclidean $C^{1}$-regular in any neighbourhood of the line
$\{(0,x_2,0) : x_2 \in \mathbb{R}\}$. They show that no Lipschitz map from
an open subset of $\Pi$ to a neighbourhood of $0 \in \Gamma$ can
have a Lipschitz inverse.
The example \eqref{BVEx} is a good prelude to the results of this
paper. The main novelties will be to
\begin{itemize}
\item[(a)] say something about the LI
rectifiability of iLGs below the critical $C^{1}$-regularity of
$\varphi$ (in particular, our results apply to the example in
\eqref{BVEx} for $\tfrac{1}{2} < \alpha < 1$),
\item[(b)] consider the problem in higher Heisenberg groups, where the technique via the
characteristic straightening map does not seem to be easily available.
\end{itemize}
Here is the first main result:
\begin{thm}\label{mainIntro} Let $\alpha > 0$, and let $S \subset \mathbb{H}^n$ be the intrinsic graph of a globally defined but compactly supported
intrinsic $C^{1,\alpha}$-function. Then $S$ has big pieces of
bilipschitz images of the parabolic plane $(\Pi,d_{\Pi})$ if
$n=1$, or of $(\mathbb{H}^{n-1}\times
\mathbb{R},d_{\mathbb{H}^{n-1}\times \mathbb{R}})$ if $n\geq 2$.
In particular, $S$ is LI rectifiable.
\end{thm}
The following corollary is easier to read:
\begin{thm}\label{mainQualitative} Let $S \subset \mathbb{H}^n$ be a $C^{1,\alpha}_{\mathbb{H}}$-surface. Then $S$ is LI rectifiable. \end{thm}
\begin{remark}
A first version of the present paper, by the third author,
contained Theorems \ref{mainIntro} and \ref{mainQualitative} in $\mathbb{H}^{1}$. After that version appeared on the arXiv, Antonelli
and Le Donne proved in \cite{antonelli2019pauls}, building on \cite{LDY}, that every $C^{\infty}$ hypersurface $S$
in $\mathbb{H}^n$, $n\geq 2$, is rectifiable by bilipschitz images
of subsets of $\mathbb{H}^{n-1}\times\mathbb{R}$.
The result of Antonelli-Le Donne follows from Theorem
\ref{mainQualitative} (or rather, the bilipschitz version stated in
Theorem \ref{surfaces}): the $C^{\infty}$ regularity of $S$,
and a result of Balogh \cite{MR2021034}, imply that
$\mathcal{H}^{2n+1}$ almost every point on $S$ has an
open neighbourhood $U$ such that $S\cap U$ is given as the level
set of a $C^{\infty}$ function $f:U \to \mathbb{R}$ with
nonvanishing \emph{horizontal gradient} $\nabla_{\mathbb{H}}f$.
Using that $f$ is Euclidean $C^{\infty}$ (or even Euclidean $C^{2}$), one concludes that $f$
satisfies condition \eqref{hHolder} in Definition
\ref{C1alpha}, possibly on a slightly smaller open set. Thus,
outside an $\mathcal{H}^{2n+1}$ null set, $S$ is locally a $C_{\mathbb{H}}^{1,\alpha}$ surface, and it
follows from Theorem \ref{surfaces} that $S$ is rectifiable by
bilipschitz images of compact subsets of $\mathbb{H}^{n-1}\times \mathbb{R}$.
\end{remark}
The notions of regularity appearing in the theorems above will be
formally introduced in Section \ref{s:C1alpha}.
In brief, a $C^{1,\alpha}_{\mathbb{H}}$-surface is an $\mathbb{H}$-regular surface whose \emph{horizontal normal} is $\alpha$-H\"older continuous (in the metric $d_{\mathbb{H}}$). Then, roughly speaking, an intrinsic $C^{1,\alpha}$-function is a function $\mathbb{W} \to \mathbb{V}$ whose intrinsic graph $\Phi(\mathbb{W})$ is a $C^{1,\alpha}_{\mathbb{H}}$-surface, but this is a little inaccurate; see Definition \ref{C1alphaGraphs} and Remark \ref{r:warning} for more precision.
The next example points out that Theorem \ref{mainQualitative} applies to the function in \eqref{BVEx}:
\begin{ex} The horizontal normal of the intrinsic graph of the function $\varphi$ from \eqref{BVEx} is
\begin{displaymath} \nu_{\mathbb{H}}(\Phi(0,x_2,t)) = \left(\frac{-1}{\sqrt{1 + |c_{\alpha}t^{2\alpha - 1}|^{2}}},\frac{c_{\alpha}t^{2\alpha - 1}}{\sqrt{1 + |c_{\alpha}t^{2\alpha - 1}|^{2}}} \right), \qquad x_2 \in \mathbb{R}, \; t > 0, \end{displaymath}
where $c_{\alpha} := \alpha/(1 - \alpha)^{2}$, and
$\nu_{\mathbb{H}}(\Phi(0,x_2,t)) \equiv (-1,0)$ for $x_2 \in \mathbb{R}$, $t \leq
0$. This follows by combining the expression for the
\emph{intrinsic gradient} of $\varphi$ on \cite[p. 166]{MR2603594}
with a known relationship between the horizontal normal of
$\Phi(\mathbb{W})$ and the intrinsic gradient of $\varphi$, see the
references around \eqref{form32}. Noting that $(2\alpha -
1)/\alpha \in (0,1]$ for $\tfrac{1}{2} < \alpha \leq 1$, we have
for arbitrary $x_2,x_2'\in \mathbb{R}$ that
\begin{displaymath} |s^{2\alpha - 1} - t^{2\alpha - 1}| \leq |s^{\alpha} - t^{\alpha}|^{(2\alpha - 1)/\alpha}
\lesssim_{\alpha} d_{\mathbb{H}}(\Phi(0,x_2,s),\Phi(0,x_2',t))^{(2\alpha
- 1)/\alpha}, \quad s,t\geq 0, \end{displaymath} so we conclude
that $\Phi(\mathbb{W})$ satisfies the assumptions of Theorem
\ref{mainQualitative}.\end{ex}
The next definition explains the rest of the terminology in Theorem \ref{mainIntro}:
\begin{definition}[BP$G$BI]\label{BPdef} Fix
$n\in \mathbb{N}$ and set $(G,d_{G})=(\Pi,d_{\Pi})$ if $n=1$, and
$(G,d_{G})=(\mathbb{H}^{n-1}\times\mathbb{R},d_{\mathbb{H}^{n-1}\times\mathbb{R}})$
if $n\geq 2$. Let $E \subset \mathbb{H}^n$ be closed and
$(2n+1)$-regular. Then, $E$ has \emph{big pieces of $G$
bilipschitz images} (BP$G$BI) if there exist constants $L \geq 1$
and $\theta > 0$ such that the following holds: for every $p \in
E$ and $0 < r \leq \operatorname{diam}_{\mathbb{H}}(E)$, there exists a compact set $K
\subset B(0,r) \subset G$ and an $L$-bilipschitz map $f \colon K
\to \mathbb{H}^n$ such that
\begin{displaymath} \mathcal{H}^{2n+1}(f(K) \cap [E \cap B(p,r)]) \geq \theta r^{2n+1}. \end{displaymath}
\end{definition}
So, up to some technical assumptions,
Theorem \ref{mainIntro} states that intrinsic $C^{1,\alpha}$-graphs in $\mathbb{H}^n$ are \emph{uniformly rectifiable} by $(G,d_{G})$,
in the spirit of David and Semmes \cite{DS1}. We next formulate a stronger result in $\mathbb{H}^{1}$. Informally, the point is that we can relax $C^{1,\alpha}$-regularity to Lipschitz regularity in the "horizontal" directions, but we still need to assume an $\epsilon$ of additional \emph{a priori} regularity in the vertical direction. To motivate the definition, we note that intrinsic $C^{1,\alpha}$-functions are locally Euclidean $(1+\alpha)/2$ H\"older
continuous along vertical lines by \cite[Proposition 4.2]{CFO2}. In $\mathbb{H}^{1}$, it turns out that this property of
\emph{extra vertical H\"older regularity} alone implies the conclusion of Theorem
\ref{mainIntro}.
\begin{thm}\label{mainIntroVertical} Let $S \subset \mathbb{H}^1$
be the intrinsic graph of a globally defined but compactly
supported intrinsic Lipschitz function with extra vertical H\"older
regularity, see Definition \ref{d:ExtraVerticalHolder}.
Then $S$ has big pieces of bilipschitz images of the parabolic
plane $(\Pi,d_{\Pi})$. In particular, $S$ is LI rectifiable.
\end{thm}
The "extra vertical H\"older regularity" is often weaker than intrinsic $C^{1,\alpha}$-regularity: for example, intrinsic Lipschitz functions of the form $\varphi(0,y,t) = \tilde{\varphi}(y)$, with $\tilde{\varphi} \colon \mathbb{R} \to \mathbb{R}$ Lipschitz, are not necessarily intrinsic $C^{1,\alpha}$, but they are very smooth along vertical lines.
Theorem \ref{mainIntroVertical} can be used to give alternative
proofs for the LI rectifiability of Euclidean $C^1$ hypersurfaces
in $\mathbb{H}^1$ (originally due to Cole-Pauls \cite{MR2247905}) and of
the $n=1$ case of Theorem \ref{mainIntro} (originally due to the
third author). Intrinsic graphs that satisfy the assumptions of
Theorem \ref{mainIntroVertical} are examples of sets on which the
$3$-dimensional \emph{Heisenberg Riesz transform} is
$L^2$-bounded, see
\cite{fssler2018riesz}. The essential property of such graphs used here is
that they can be well approximated by \emph{Lipschitz flags}
(Definition \ref{d:FlagSurface}), and that Lipschitz flags can be
bilipschitz parametrised using the characteristic straightening
map of Cole-Pauls and Bigolin-Vittone. No counterparts for these
properties are known in higher dimensions.
\medskip
We close this section with a few questions:
\begin{questions}Are all intrinsic Lipschitz graphs in $\mathbb{H}^n$
LI rectifiable? If so, do they have big pieces of Lipschitz
images of $G$, or
BP$G$BI? In the converse direction: Are (bi-)Lipschitz images of vertical subgroups in $\mathbb{H}^n$ iLG rectifiable? \end{questions}
\subsection{Bilipschitz maps between big pieces of metric spaces} As mentioned above Theorem \ref{mainIntro},
adapting the techniques of Cole-Pauls and Bigolin-Vittone seems
difficult without something close to $C^{1}$-regularity, or if
$n\geq 2$.
Instead, Theorems \ref{mainIntro} and \ref{mainIntroVertical} will follow from an application of a general result concerning metric spaces, Theorem \ref{main} below. We now formulate the abstract hypotheses of that theorem.
Let $(G,d_{G})$ and $(M,d_{M})$ be metric spaces. Assume the
following "local correspondence" between $G$ and $M$,
for constants $\alpha > 0$, $L\geq 1$, $A \geq 1$, and for some $x_{0} \in G$ and $p_{0} \in M$ fixed.
For every $x \in B_{G}(x_{0},1)$, $p \in B_{M}(p_{0},1)$ and
$n \in \mathbb{N} \cup \{0\}$ there exists a map $i^{n}_{x \to p} \colon G \to M$ with $i^{n}_{x \to p}(x) = p$ such that
\begin{equation}\label{ISO}
L^{-1}d_G(y,z)-A2^{-n(1+\alpha)}\leq d_M(i^{n}_{x \to
p}(y),i^{n}_{x \to p}(z))\leq L d_G(y,z)+ A 2^{-n(1+\alpha)},\;
y,z\in B(x,2^{-n}).
\end{equation} Moreover, if
$n \geq 0$, $x,y \in B_{G}(x_{0},1)$ and $p,q \in B_{M}(p_{0},1)$
with $d_{G}(x,y) \leq 2^{-n}$ and $i^{n}_{x \to p}(y) = q$, then
\begin{equation}\label{comp} d_{M}(i^{n}_{x \to p}(z),i^{n + 1}_{y \to q}(z)) \leq A2^{-n(1 + \alpha)} \qquad z \in B(x,2^{-n}). \end{equation}
These assumptions are reminiscent of \cite[(1.6)-(1.9)]{MR2907827}. See also \cite[Appendix 1]{CheegerColding} for related results. The assumption \eqref{ISO} postulates that the maps $i_{x \to p}^{n}$ are bilipschitz continuous at the scale $2^{-n}$, up to an error which is much smaller than $2^{-n}$. The assumption \eqref{comp} is "compatibility condition": it states that the maps $i_{x \to p}^{n}$ and $i_{y \to q}^{n + 1}$ nearly coincide at scale $2^{-n}$, again up to an error which is much smaller than $2^{-n}$. The \emph{a priori} condition "$i_{x \to p}^{n}(y) = q$" above \eqref{comp} is a technically convenient way of assuming that $p$ and $q$ are close to each other on $M$: indeed it follows from the hypotheses, including \eqref{ISO}, that
\begin{displaymath} d_{M}(p,q) = d_{M}(i_{x \to p}^{n}(x),i_{x \to p}^{n}(y)) \leq Ld_{G}(x,y) + A2^{-n(1 + \alpha)} \lesssim_{A,L} 2^{-n}. \end{displaymath}
The assumptions \eqref{ISO}-\eqref{comp} are designed to enable
the construction of bilipschitz maps from subsets of $G$ to $M$,
at least if $G$ is complete and doubling:
\begin{thm}\label{main}Assume that $(G,d_{G})$ is a complete metric space with $\operatorname{diam}_{G}(G) \geq 1$
and equipped with a nontrivial doubling measure $\mu$, and
$(M,d_{M})$ is complete. Assume that the conditions
\eqref{ISO}-\eqref{comp} hold for some $\alpha > 0$, $L\geq 1$,
and $A \geq 1$. Then, there exists a constant $\delta > 0$, a
compact set $K \subset B(x_{0},1)$ with $\mu(K) \geq \delta
\mu(B(x_{0},1))$ and a $2L$-bilipschitz embedding $F \colon K \to
M$ with $F(K) \subset B(p_{0},1)$. The constant $\delta > 0$ only
depends on the doubling constant of $(G,d_G,\mu)$, and the
constants $\alpha$, $L$, and $A$ in \eqref{ISO}-\eqref{comp}.
\end{thm}
The plan of the paper is to prove Theorem \ref{main} in Section
\ref{s:construction} and Appendix \ref{FCSPropProof}, and then
apply it to prove Theorem \ref{mainIntro} for $\mathbb{H}^n$, $n>1$, in
Section \ref{s:C1alpha}. In Section \ref{s:LipFlag} we prove
Theorem \ref{mainIntroVertical}, which yields as a corollary the
case $n=1$ of Theorem \ref{mainIntro}. Theorem
\ref{mainQualitative} is "morally" a direct corollary of Theorem
\ref{mainIntro}: by the implicit function theorem of Franchi,
Serapioni, and Serra Cassano \cite[Theorem 6.5]{FSSC}, a
$C^{1,\alpha}_{\mathbb{H}}$-surface is locally parametrisable by an
intrinsic $C^{1,\alpha}$-function over some vertical subgroup $\mathbb{W}
\subset \mathbb{H}^n$. However, the parametrisation may only be defined
on a strict subset of $\mathbb{W}$, and fail to literally satisfy the
assumptions of Theorem \ref{mainIntro}. As far as we know, there
is no extension theorem for intrinsic $C^{1,\alpha}$-functions
available. To bypass the issue, Appendix \ref{extensionAppendix}
contains a proposition saying that every point on a
$C_{\mathbb{H}}^{1,\alpha}$-surface has a neighbourhood which is
contained on the intrinsic graph of a globally defined, compactly
supported intrinsic $C^{1,\alpha/3}$-function. With this
proposition in hand, Theorem \ref{mainQualitative} is indeed a
corollary of Theorem \ref{mainIntro}.
\subsection{Notations} For $A,B > 0$, we write $A\lesssim B$ if there is a constant $C >
0$ such that $A \leq CB$. If we want to specify that the value of
$C$ is allowed to depend on an auxiliary parameter $h$, we will
write $A\lesssim_h B$.
\subsection{Acknowledgements}
K.F. and T.O. would like to thank Enrico Le Donne and S\'everine
Rigot for many fruitful discussions on the topic of the paper. We
are also grateful to Davide Vittone for tips on proving
Proposition \ref{appProp}, which was needed to reduce Theorem
\ref{mainQualitative} to Theorem \ref{mainIntro}. Finally, we
thank Damian D\k{a}browski for pointing out the reference
\cite{2017arXiv171103088A}. D.D.D. would like to thank Enrico Le
Donne and Raul Serapioni for important suggestions on the subject.
\section{Proof of the main theorem for metric spaces}\label{s:construction}
The proof of Theorem \ref{main} is inspired by the recent work of Le Donne and Young \cite{LDY} on the Carnot rectifiability of sub-Riemannian manifolds.
Before starting the proof of Theorem \ref{main} in earnest, we
need to introduce some terminology. Assume for a moment that
$n_{0} \geq 0$, and there exist families $\{\mathcal{D}_{n}\}_{n \geq
n_{0}}$ of subsets of $G$, known as \emph{cubes}, with the
following properties:
\begin{itemize}
\item[(i)\phantomsection \label{i}] Each $\mathcal{D}_{n}$ consists of a finite number of disjoint non-empty compact sets, and in particular $\operatorname{card} \mathcal{D}_{n_{0}} = 1$.
\item[(ii) \phantomsection \label{ii}] For $n > n_{0}$, each cube $Q \in \mathcal{D}_{n}$ is contained in a unique cube $\hat{Q} \in \mathcal{D}_{n - 1}$, called the \emph{parent} of $Q$. For $Q \in \mathcal{D}_{n - 1}$ fixed, write $\mathrm{ch}(Q) := \{Q \in \mathcal{D}_{n} : \hat{Q} = Q\}$.
\item[(iii) \phantomsection \label{iii}] $\operatorname{diam}_{G}(Q) < 2^{-n}$ for all $Q \in \mathcal{D}_{n}$.
\item[(iv) \phantomsection \label{iv}] There are constants $\epsilon, \tau > 0$ such that if $n \geq 0$ and $Q_{1},Q_{2} \in \mathcal{D}_{n}$ are distinct, then $d_{G}(Q_{1},Q_{2}) \geq \tau 2^{-(1 + \epsilon)n}$.
\end{itemize}
An \emph{$(\epsilon,n_{0},\tau)$-Cantor set} is any set of the form
\begin{displaymath} K := \bigcap_{n \geq n_{0}} K_{n} := \bigcap_{n \geq n_{0}} \bigcup_{Q \in \mathcal{D}_{n}} Q, \end{displaymath}
where the families $\mathcal{D}_{n}$, $n_{0} \geq 0$, satisfy properties (i)-(iv).
\begin{definition}\label{FCS} A metric measure space $(X,d,\mu)$ \emph{admits fat Cantor sets} if for all $\epsilon > 0$ and $n_{0} \geq 0$, there exist constants $\delta = \delta(n_{0}) > 0$, and $\tau = \tau(\epsilon) > 0$ such that the following holds. For all $x \in X$, there exists an $(\epsilon,n_{0},\tau)$-Cantor set $K \subset B(x,1)$ with $\mu(K) \geq \delta \mu(B(x,1))$. \end{definition}
We will only use the assumption that $(G,d_{G},\mu)$ is doubling and complete to ensure that it admits fat Cantor sets.
\begin{proposition}\label{FCSProp} Every doubling and complete metric measure space $(X,d,\mu)$ of diameter $\geq 1$ admits fat Cantor sets. In other words, for every $\epsilon > 0$ and $n_{0} \geq 0$, the constants $\delta(n_{0}) > 0$ and $\tau(\epsilon) > 0$ can be found as in Definition \ref{FCS}. They are also allowed to depend on the doubling constant of $(X,d,\mu)$.
\end{proposition}
The proof of the proposition above is essentially contained in
\cite[Section 4.1]{LDY}, but since the statement is not explicitly
given in \cite{LDY}, we repeat the details in Appendix
\ref{FCSPropProof}. What follows next is a proof of Theorem
\ref{main}, assuming Proposition \ref{FCSProp}. Fix $x_{0} \in G$
and $p_{0} \in M$. Let $\epsilon := \alpha/2$ (where $\alpha > 0$
is the parameter appearing in \eqref{ISO}-\eqref{comp}), let
$n_{0} \geq 0$ be a large integer to be determined later, and let
$K \subset B(x_{0},1)$ be an $(\epsilon,n_{0},\tau)$-Cantor set
associated to families of cubes $\{\mathcal{D}_{n}\}_{n \geq n_{0}}$, as
in \nref{i}-\nref{iv}. For each $Q \in \mathcal{D}_{n}$, $n \geq n_{0}$,
pick a \emph{centre} $c_{Q} \in Q$. Set $\mathcal{D}_{n_{0}-1} := \{G\}$
and $K_{n_{0} - 1} := G$.
The map $F \colon K \to M$ will be defined as the limit of certain intermediate maps $F_{n} \colon K_{n} \to M$ for $n \geq n_{0} - 1$. Set $F_{n_{0}-1} \equiv p_{0}$. To proceed, assume that $n \geq n_{0}$, and the map $F_{n - 1} \colon K_{n - 1} \to M$ has already been defined. Then, fix $Q_{n - 1} \in \mathcal{D}_{n - 1}$, and set
\begin{displaymath} F_{n}|_{Q} := i^{n}_{c_{Q} \to F_{n - 1}(c_{Q})}|_{Q}, \qquad Q \in \mathrm{ch}(Q_{n - 1}) \subset \mathcal{D}_{n}. \end{displaymath}
The next lemma shows, in particular, that if $x \in K$, then the sequence $(F_{n}(x))_{n \in \mathbb{N}}$ is Cauchy in $(M,d_{M})$. Hence $F(x) := \lim_{n \to \infty} F_{n}(x)$ exists by the completeness of $(M,d_{M})$.
\begin{lemma} If $n_{0} \geq 1$ is large enough (depending on $A$ and the constant $L$ in \eqref{ISO}), the following holds for all $w \in K$ and $n \geq n_{0}$:
\begin{equation}\label{form15} d_{M}(F_{n}(w),F_{n + 1}(w)) \leq A2^{-n(1 + \alpha)}. \end{equation}
\end{lemma}
\begin{proof} For $w \in K$, let $c_{n}(w)$ be the centre of $Q_{n}(w)$, where $Q_{n}(w)$ is the unique cube in $\mathcal{D}_{n}$ containing $w$. One can infer \eqref{form15} from the compatibility condition \eqref{comp} in the following way:
\begin{align*} d_{M}(F_{n}(w),F_{n + 1}(w)) = d_{M}(i^{n}_{c_{n}(w) \to F_{n - 1}(c_{n}(w))}(w),i^{n + 1}_{c_{n + 1}(w) \to F_{n}(c_{n + 1}(w))}(w)) \leq A2^{-n(\alpha + 1)}. \end{align*}
To check that the assumptions of \eqref{comp} are really in force, use the notational substitutions
\begin{displaymath} q := F_{n}(c_{n + 1}(w)), \: p := F_{n - 1}(c_{n}(w)), \: x := c_{n}(w), y := c_{n + 1}(w), \quad \text{and} \quad z := w.\end{displaymath}
Then, note that $d_{G}(x,y) \leq \operatorname{diam}_{G}(Q_{n}(x)) < 2^{-n}$ by \nref{iii}, $z \in B(x,2^{-n})$ again by \nref{iii}, and
\begin{equation}\label{form38} q = F_{n}(c_{n + 1}(w)) = i_{c_{n}(w) \to F_{n - 1}(c_{n}(w))}^{n}(c_{n + 1}(w)) = i^{n}_{x \to p}(y). \end{equation}
by definition. This shows that the assumptions of \eqref{comp} are valid, except for one small issue: is it clear that $p,q \in B(p_{0},1)$? For $n = n_{0}$, simply $p = F_{n_{0} - 1}(c_{n_{0}}(w)) = p_{0}$. Also, using \eqref{form38} and \eqref{ISO},
we find that
\begin{displaymath} d_{M}(q,p) = d_{M}(i^{n}_{x \to p}(y),i^{n}_{x \to p}(x)) \leq L d_{G}(x,y) + A2^{-n(1 + \alpha)}
\lesssim_{L} A2^{-n}. \end{displaymath}
Recalling the definitions of $p,q$, and using the estimate above
repeatedly shows that $\max\{d_{M}(p,p_{0}),d_{M}(q,p_{0})\}
\lesssim_{L} A2^{-n_{0}}$ for all $n \geq n_{0}$. In particular,
$p,q \in B(p_{0},1)$ if $n_{0} \geq 1$ is large enough, depending
on $A$ and the constant $L$ in \eqref{ISO}. The proof of the lemma
is complete.
\end{proof}
As an immediate corollary, one deduces the useful estimate
\begin{equation}\label{form4} d_{M}(F_{n}(x),F(x)) \lesssim A2^{-n(1 + \alpha)}, \qquad x \in K, \: n \geq n_{0}. \end{equation}
It remains to prove that $F$ is $2L$-bilipschitz on $K$ if the
index $n_{0} \in \mathbb{N}$ was chosen large enough, depending on the
parameters
$\alpha > 0$, $L\geq 1$ and $A \geq 1$. Fix $x,y \in K$ arbitrary with $x \neq y$, and let $Q \in \mathcal{D}_{n}$, $n \geq n_{0}$, be the smallest cube with $x,y \in Q$ (thus $x,y$ lie in distinct cubes in $\mathcal{D}_{n + 1}$). Write $c := c_{Q} \in Q$. Then, using properties \nref{iii}-\nref{iv} of the cubes $\mathcal{D}_{n}$,
\begin{equation}\label{form1} \max\{d_{G}(x,c),d_{G}(y,c)\} \leq 2^{-n} \quad \text{and} \quad \tau 2^{-(n + 1)(1 + \epsilon)} \leq d_{G}(x,y) \leq 2^{-n}. \end{equation}
Also, by definition,
\begin{displaymath} F_{n}(x) = i^{n}_{c \to F_{n - 1}(c)}(x) \quad \text{and} \quad F_{n}(y) = i^{n}_{c \to F_{n - 1}(c)}(y), \end{displaymath}
since $x,y \in Q$. We deduce from \eqref{ISO} (using \eqref{form1}) that
\begin{equation}\label{form7} d_{M}(F_{n}(x),F_{n}(y)) = d_{M}(i^{n}_{c \to F_{n - 1}(c)}(x),i^{n}_{c \to F_{n - 1}(c)}(y)) \geq
L^{-1} d_{G}(x,y) - A2^{-n(1 + \alpha)}. \end{equation} We also
have the upper bound analogous to \eqref{form7},
\begin{displaymath} d_{M}(F_{n}(x),F_{n}(y)) \leq
L d_{G}(x,y) + A2^{-n(1 + \alpha)}. \end{displaymath} Recalling
that $\epsilon = \alpha/2$ and $\tau=\tau(\varepsilon)$, the error
term $A2^{-n(1 + \alpha)}$ is smaller than $$L^{-1}\tau 2^{-(n +
1)(1 + \epsilon)}/4 \leq L^{-1}d_{G}(x,y)/4$$ for $n \geq n_{0}$,
if $n_{0} \in \mathbb{N}$ was chosen large enough, depending on $A$, $L$
and $\alpha$. Thus,
\begin{equation}\label{form2}L^{-1} \frac{3 d_{G}(x,y)}{4} \leq d_{M}(F_{n}(x),F_{n}(y)) \leq L\frac{5 d_{G}(x,y)}{4}. \end{equation}
Finally, if $n_{0} \geq 0$ is large enough, depending again on
$\alpha$, $L$ and $A$, one sees from \eqref{form4}-\eqref{form1}
that
\begin{equation}\label{form3} \max\{d_{M}(F(x),F_{n}(x)),d_{M}(F(y),F_{n}(y))\} \leq L^{-1}\frac{d_{G}(x,y)}{8}. \end{equation}
It then follows by combining \eqref{form2}, \eqref{form3}, and the triangle inequality that
\begin{displaymath} L^{-1}\frac{d_{G}(x,y)}{2} \leq d_{M}(F(x),F(y)) \leq L 2d_{G}(x,y), \end{displaymath}
as desired. The proof of Theorem \ref{main} is complete, except for the claim $F(K) \subset B(p_{0},1)$. Pick $x \in K \subset Q_{n_{0}} \subset B(c_{Q_{n_{0}}},2^{-n_{0}})$, and write, using \eqref{ISO},
\begin{displaymath} d_{M}(p_{0},F_{n_{0}}(x)) = d_{M}(i^{n_{0}}_{c_{Q_{n_{0}}} \to p_{0}}(c_{Q_{n_{0}}}),i^{n_{0}}_{c_{Q_{n_{0}}} \to p_{0}}(x)) \leq
Ld_{G}(c_{Q_{n_{0}}},x) + A2^{-n_{0}} \leq (L + A)2^{-n_{0}}.
\end{displaymath} Further, it follows from \eqref{form4} that
$d_{M}(F(x),F_{n_{0}}(x)) \lesssim A2^{-n_{0}}$. So, if $n_{0}
\geq 2$ was chosen large enough, depending on $A$ and $L$, one has
$F(x) \in B(p_{0},\tfrac{1}{2})$. Finally, since $\operatorname{diam}_{G}(K)
\leq 2^{-n_{0}} \leq \tfrac{1}{4L}$ for $n_0$ large enough
depending on $L$, and $F$ is $2L$-Lipschitz, one has $F(K) \subset
B(F(x),\tfrac{1}{2}) \subset B(p_{0},1)$. The proof of Theorem
\ref{main} is complete.
\section{Graphs and surfaces with H\"older-continuous horizontal normals}\label{s:C1alpha}
Throughout this section, we use coordinates
$(x_1,\ldots,x_{2n},t)$ on
$\mathbb{H}^n$ as defined at the beginning of Section
\ref{s:introd}. In these coordinates, a frame for the
left invariant vector fields on $\mathbb{H}^n$ is given by
\begin{equation}\label{eq:vfd}
X_i = \partial_{x_i}-\tfrac{x_{n+i}}{2}\partial_t,\quad X_{n+i} = \partial_{x_{n+i}}+\tfrac{x_{i}}{2}\partial_t,\text{ for
$i=1,\ldots,n$,}\quad
\text{and}\quad
T=\partial_t,
\end{equation} which yields the nontrivial commutator relations
\begin{displaymath}
[X_i,X_{n+i}]=T,\quad i=1,\ldots,n.
\end{displaymath}
The \emph{horizontal gradient} of a function $f:U \subset
\mathbb{H}^n \to \mathbb{R}$ is
\begin{displaymath}
\nabla_{\mathbb{H}}f=(X_1 f,\ldots,X_{2n} f).
\end{displaymath}
A \emph{vertical subgroup $\mathbb{W}$ of codimension $1$} in
$\mathbb{H}^n$ is, in the above coordinate system, a
$2n$-dimensional subspace of $\mathbb{R}^{2n+1}$ containing the
$t$-axis. Given such a subgroup $\mathbb{W}$, we denote by
$\mathbb{V}$ its Euclidean orthogonal complement. We recall that
$\pi_{\mathbb{W}} \colon \mathbb{H}^n \to \mathbb{W}$ is the \emph{vertical projection}
to $\mathbb{W}$, and $\pi_{\mathbb{V}} \colon \mathbb{H}^n \to \mathbb{V}$ is the
\emph{horizontal projection} to $\mathbb{V}$, induced by the splitting
$\mathbb{H} = \mathbb{W} \cdot \mathbb{V}$, see \cite[Proposition 2.2]{FSS}. In
particular,
\begin{displaymath}
p= \pi_{\mathbb{W}}(p) \cdot \pi_{\mathbb{V}}(p),\quad\text{for all
}p\in\mathbb{H}^n.
\end{displaymath}
\subsection{Definitions and preliminaries}\label{s:DefPrelim} A \emph{$C^{1,\alpha}_{\mathbb{H}}$-surface} is locally a non-critical level set of a $C^{1,\alpha}_{\mathbb{H}}$-function $f \colon \mathbb{H}^n \to \mathbb{R}$. Here is the precise definition:
\begin{definition}[$C^{1,\alpha}_{\mathbb{H}}$-surfaces]\label{C1alpha} Let $S \subset \mathbb{H}^n$ be an $\mathbb{H}$-regular surface in the sense of Franchi, Serapioni, and Serra Cassano \cite[Definition 6.1]{FSSC}: for every $p \in S$, there exists an open ball $B(p,r)$ and a function $f \in C^{1}_{\mathbb{H}}(B(p,r))$ such that $\nabla_{\mathbb{H}}f(p) \neq 0$, and
\begin{equation}\label{eq:Slevel} S \cap B(p,r) = \{q \in B(p,r) : f(q) = 0\}. \end{equation}
For $0 < \alpha \leq 1$, the set $S$ is called a
\emph{$C^{1,\alpha}_{\mathbb{H}}$-surface} if one can choose $f$ so that
there exists a constant $H = H_{p} \geq 1$ such that
\begin{equation}\label{hHolder} |\nabla_{\mathbb{H}}f(q_{1}) - \nabla_{\mathbb{H}}f(q_{2})|
\leq Hd_{\mathbb{H}}(q_{1},q_{2})^{\alpha}, \qquad q_{1},q_{2} \in S \cap B(p,r). \end{equation}
\end{definition}
\begin{remark}\label{normalRemark} If $S \cap B(p,r) = B(p,r) \cap \{f = 0\}$, as above, then \cite[Theorem 6.5]{FSSC} states that, after making $r > 0$ possibly a little smaller, the inward-pointing horizontal normal of $E = \{f < 0\}$ is given by the expression
\begin{displaymath} \nu_{E}(q) = -\frac{\nabla_{\mathbb{H}}f(q)}{|\nabla_{\mathbb{H}}f(q)|}, \qquad q\in S \cap B(p,r). \end{displaymath}
Clearly, if $f$ satisfies \eqref{hHolder}, then this choice of
horizontal normal is (locally) $\alpha$-H\"older continuous as a
map $(S,d_{\mathbb{H}}) \to S^{2n-1}$.
Conversely, if $S\subset \mathbb{H}^n$ is an $\mathbb{H}$-regular
surface with $\alpha$-H\"older continuous $\nu_E$, if $p$ is an
arbitrary point in $S$, and $r> 0$ is small enough, we claim that
there exists $f\in C_{\mathbb{H}}^1(B(p,r))$ so that
\eqref{eq:Slevel} and \eqref{hHolder} hold. Not every function $f$
which satisfies \eqref{eq:Slevel} necessarily fulfills
\eqref{hHolder}, cf.\ the related Remark \ref{r:warning} below. In
order to find $f$ which satisfies simultaneously the two
conditions, it is convenient to write $S$ locally as
\emph{intrinsic graph}. First, by assumption there exists $i\in
\{1,\ldots,2n\}$ so that the component $\nu_{E}^i$ does not vanish
on $S \cap B(p,r)$ for small enough $r>0$. Without loss of
generality, we may assume that $i=1$ and $\nu_{E}^i< 0$ on $S \cap
B(p,r)$. Then, the implicit function theorem of Franchi,
Serapioni, and Serra Cassano \cite[Theorem 6.5]{FSSC}, combined
with \cite[Theorem 1.2]{MR2223801}, implies that for small enough
$r>0$, the set $S \cap B(p,r)$ can be written as {intrinsic graph}
in $X_1$-direction of a function $\varphi$ whose \emph{intrinsic
gradient} $\nabla^{\varphi}\varphi$ (in the sense of Definition
\ref{d:IntrGrad}) exists and is a continuous
$\mathbb{R}^{2n-1}$-valued function. This allows to express
locally the horizontal normal $\nu_{E}$ in a convenient form,
cf.\ \eqref{form32}. Using this expression, it is easy to see that
$\alpha$-H\"older continuity of $\nu_{E}$ implies
$\alpha$-H\"older continuity of $[\nabla^{\varphi}\varphi] \circ
\pi_{\mathbb{W}}|_{S\cap B(p,r)}$ for $r$ small enough, cf.\
\eqref{eq:HolForm}. Then
\begin{displaymath}
f(x_1,\ldots,x_{2n},t):=
x_1-\varphi\left(\pi_{\mathbb{W}}(x_1,\ldots,x_{2n},t)\right)=
x_1-\varphi\left(0,x_2,\ldots,x_{2n},t+\tfrac{1}{2}x_1x_{n+1}\right)
\end{displaymath}
has the properties \eqref{eq:Slevel} and \eqref{hHolder}. To see
this, use \cite[Proposition 2.22]{MR2223801} for the expression of
$\nabla_{\mathbb{H}}f$ in terms of $\nabla^{\varphi}\varphi$, and note
that $X_1 f =1$.
\end{remark}
A case of particular interest in this paper are intrinsic graphs
$S \subset \mathbb{H}^n$ that also happen to be
$C^{1,\alpha}_{\mathbb{H}}$-surfaces. Specifically, consider the
following definition:
\begin{definition}[Intrinsic $C^{1,\alpha}$-functions and intrinsic $C^{1,\alpha}$-graphs]\label{C1alphaGraphs} Let
\begin{displaymath}
\mathbb{W} := \{(0,x_2,\ldots,x_{2n},t) : (x_2,\ldots,x_{2n},t) \in
\mathbb{R}^{2n}\}
\end{displaymath}
and $\mathbb{V} := \{ (x_1,0\ldots,0): x_1 \in \mathbb{R}\} \cong \mathbb{R}$. Let $U \subset \mathbb{W}$ be open, and let $\varphi \colon U \to \mathbb{V}$ be continuous. Write $\Phi(w) := w \cdot \varphi(w)$ for the \emph{graph map} of $\varphi$. We say that $\varphi$ is an \emph{intrinsic $C^{1,\alpha}$-function on $U$}, denoted $\varphi \in C^{1,\alpha}_{\mathbb{H}}(U)$ if
\begin{itemize}
\item[(i)] $S := \Phi(U) \subset \mathbb{H}^n$ is a
$C^{1,\alpha}_{\mathbb{H}}$-surface in the sense of Definition
\ref{C1alpha}, and \item[(ii)] the horizontal normal $\nu_{\mathbb{H}} =
(\nu_{\mathbb{H}}^{1},\ldots,\nu_{\mathbb{H}}^{2n}) \colon S \to \mathbb{R}^{2n}$ of the
subgraph
\begin{displaymath}
\{w \cdot v : v < \varphi(w)\}
\end{displaymath}
satisfies $\nu_{\mathbb{H}}^{1}(p) < 0$ for all $p \in S$.
\end{itemize}
The intrinsic graph $S$ of any intrinsic $C^{1,\alpha}$-function
is an \emph{intrinsic $C^{1,\alpha}$-graph}.
\end{definition}
\begin{remark}\label{r:warning} The conditions (i)-(ii) are $C^{1,\alpha}$-versions of the conditions appearing in \cite[Theorem 1.2(i)]{MR2223801}. Condition (ii) is not as odd as it looks: a similar hypothesis would also be required to characterise $C^{1,1}$-functions $f \colon \mathbb{R} \to \mathbb{R}$ via the properties of $\Gamma(f) := \{(x,f(x)) : x \in \mathbb{R}\}$. To see this, consider $f \colon [0,\infty) \to \mathbb{R}$ given by $f(x) = \mathrm{sgn}(x)\sqrt{|x|}$. Then $\Gamma(f)$ is a $C^{1,1}$-surface as a subset of $\mathbb{R}^{2}$, because $\Gamma(f)$ can also be written as $\Gamma(f) = \{(\operatorname{sgn}(y)y^{2},y) : y \in \mathbb{R}\}$, where $y \mapsto \operatorname{sgn}(y)y^{2} \in C^{1,1}(\mathbb{R})$. Nonetheless, $f \notin C^{1,1}(\mathbb{R})$. \end{remark}
A previous notion of \emph{intrinsic $C^{1,\alpha}$-functions} and
\emph{graphs} in $\mathbb{H}^1$ already exists, see
\cite[Definition 2.16]{CFO2} or \eqref{CFOdef} below, and it looks
different than Definition \ref{C1alphaGraphs}. The connection
needs to be clarified immediately, because a result concerning
intrinsic $C^{1,\alpha}$-functions in the sense of \cite{CFO2},
namely \cite[Proposition 4.2]{CFO2}, will be used also in this
paper. Definition 2.16 in \cite{CFO2} was stated for
$\mathbb{H}^1$, but it can be extended in an obvious way to higher
dimensional Heisenberg groups, and this version appears in
Proposition \ref{equivProp}. The precise formulation requires the
notion of \emph{intrinsic differentiability}.
\begin{definition}\label{d:IntrDiff} A function $\varphi:\mathbb{W} \to
\mathbb{V}$ is \emph{intrinsically differentiable} at the point $w_0\in
\mathbb{W}$ if there exists a map $L:\mathbb{W} \to \mathbb{V}$ whose intrinsic graph
$\{w\cdot L(w):\;w\in \mathbb{W}\}$ is a vertical subgroup and which
satisfies
\begin{displaymath}
\lim_{\|w\|\to 0} \frac{\|L(w)^{-1}\cdot
\varphi^{(p^{-1})}(w)\|}{\|w\|}=0,\quad w\in \mathbb{W}.
\end{displaymath}
Here $p=\Phi(w_0)$, and $\varphi^{(p^{-1})} \colon \mathbb{W} \to \mathbb{V}$ is
the unique function $\mathbb{W} \to \mathbb{V}$ whose intrinsic graph is $p^{-1}
\cdot \Phi(\mathbb{W})$ (see \eqref{form34} for a formula).
\end{definition}
For equivalent definitions, see \cite[Proposition
4.76]{MR3587666}. If $\varphi$ is intrinsically differentiable at
$w_0$, then there is a unique map $L:\mathbb{W} \to \mathbb{V}$ with the
properties stated in Definition \ref{d:IntrDiff}. This map is
called the \emph{intrinsic differential} of $\varphi$ at $w_0$ and
it is denoted by $d^{\varphi}\varphi(w_0)$. Moreover, there is a
unique vector $\nabla^{\varphi}\varphi(w_0)\in\mathbb{R}^{2n-1}$,
such that
\begin{displaymath}
d^{\varphi}\varphi(w_0)(w)=\langle\nabla^{\varphi}\varphi(w_0),
\pi(w) \rangle,\quad w\in \mathbb{W},
\end{displaymath}
where $\langle \cdot,\cdot\rangle$ denotes the scalar product in
$\mathbb{R}^{2n-1}$, and
\begin{displaymath}
\pi(0,x_2,\ldots,x_{2n},t)=(x_2,\ldots,x_{2n}).
\end{displaymath}
\begin{definition}\label{d:IntrGrad}
If $\varphi:\mathbb{W}\to \mathbb{V}$ is intrinsically
differentiable at $w_0 \in \mathbb{W}$, its \emph{intrinsic gradient at
$w_0$} is the vector $\nabla^{\varphi}\varphi(w_0)$. The
components of $\nabla^{\varphi}\varphi(w_0)$ are denotes as
follows:
\begin{displaymath}
\nabla^{\varphi}\varphi(w_0)=(D_2^{\varphi}\varphi(w_0),\ldots,D_{2n}^{\varphi}\varphi(w_0)).
\end{displaymath}
\end{definition}
More information about the functions $D_i^{\varphi}\varphi$,
$i=2,\ldots,2n$, will only be required once we arrive at the proof
of Proposition \ref{p:Approx}, so we postpone the detailed
discussion. At this point we just mention that the component
$D_{n+1}^{\varphi}\varphi$ will play a distinguished role as we
consider intrinsic graphs in the $X_1$ direction, and
$[X_1,X_i]=0$ for all $i=1,\ldots,2n$, except for $i=n+1$.
\begin{proposition}\label{equivProp} Let $0<\alpha \leq 1$ and $\varphi \colon \mathbb{W} \to \mathbb{V}$ be a compactly supported function between the subgroups
$\mathbb{W}$ and $\mathbb{V}$ in $\mathbb{H}^n$. Then $\varphi \in
C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ in the sense of Definition
\ref{C1alphaGraphs} if and only if $\varphi$ is intrinsically
differentiable, and the intrinsic gradient
$\nabla^{\varphi}\varphi$ satisfies
\begin{equation}\label{CFOdef} |\nabla^{\varphi^{(p^{-1})}} \varphi^{(p^{-1})}(w) - \nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(0)| \leq H\|w\|^{\alpha}, \qquad w \in \mathbb{W}, \: p \in \Phi(\mathbb{W}) \end{equation}
for a constant $H \geq 1$.
\end{proposition}
\begin{remark}\label{r:GradProp} Since \cite[Definition 2.16]{CFO2} imposes "global" H\"older continuity for $\nabla^{\varphi}\varphi$, whereas the assumptions in Definition \ref{C1alphaGraphs} are of local nature, the notions cannot be equivalent without some \emph{a priori} assumptions
-- as the compact support of $\varphi$ in Proposition
\ref{equivProp}. We also remark that condition \eqref{CFOdef}
implies that $\nabla^{\varphi}\varphi$ is continuous, as can be
deduced for instance from formula \eqref{eq:translatedGrad}
below.
\end{remark}
\begin{proof}[Proof of Proposition \ref{equivProp}] Assume that $\varphi \in C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ in the sense of Definition \ref{C1alphaGraphs}.
Then, \cite[Theorem 1.2]{MR2223801} states \textrm{in particular}
that $\varphi$ is intrinsically differentiable, the intrinsic
gradient $\nabla^{\varphi}\varphi(w)$ exists for all $w \in \mathbb{W}$, and $w
\mapsto \nabla^{\varphi}\varphi(w)$ is continuous (see also
\cite[Theorem 4.95]{MR3587666}).
Additionally, \cite[Theorem 1.2]{MR2223801} promises that the
horizontal normal $\nu_{\mathbb{H}}$ in Definition
\ref{C1alphaGraphs}(ii) has the representation
\begin{equation}\label{form32} \nu_{\mathbb{H}}(\Phi(w)) = \left(\frac{-1}{\sqrt{1 + |\nabla^{\varphi}\varphi(w)|^{2}}},\frac{\nabla^{\varphi}\varphi(w)}{\sqrt{1 + |\nabla^{\varphi}\varphi(w)|^{2}}} \right), \qquad w \in \mathbb{W}. \end{equation}
Let us now argue that $\nu_{\mathbb{H}}$, above, is (globally)
$\alpha$-H\"older continuous on $S = \Phi(\mathbb{W})$. Recall that by the
assumption that $S$ is a $C^{1,\alpha}_{\mathbb{H}}$-surface, and Remark
\ref{normalRemark}, for every $p \in S$ there exists \textbf{some}
H\"older continuous choice of a horizontal normal $\nu_{\mathbb{H}}^{p}$,
defined in a neighbourhood $S \cap U$ of $p$ (note that there are
two horizontal normals at every point in $S$). However, it is easy
to see that $\nu_{\mathbb{H}}^{p}$ must coincide on $S \cap U$ with
either $\nu_{\mathbb{H}}$ or $-\nu_{\mathbb{H}}$ whenever $S \cap U$ is
connected (that is, the sign depends only on $U$). But since $S$
is locally connected, and since $\nu_{\mathbb{H}}(p) \equiv
(-1,0,\ldots,0)$ outside the compact set $\Phi(\operatorname{spt} \varphi)
\subset S$, one infers that the horizontal normal $\nu_{\mathbb{H}}$ in
\eqref{form32} is $\alpha$-H\"older continuous on $S$.
Another remark: using again that $\operatorname{spt} \varphi$ is compact,
$\nabla^{\varphi}\varphi$ is continuous and supported in $\operatorname{spt}
\varphi$, we see that $L :=
\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})} < \infty$. From
\eqref{form32} and the $\alpha$-H\"older continuity of $\nu_{\mathbb{H}}$
on $S=\Phi(\mathbb{W})$, it can easily be deduced that
\begin{displaymath}
\ell:(S,d_{\mathbb{H}})\to \mathbb{R},\quad p\mapsto\ell(p):=
\sqrt{1+|\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))|^2}
\end{displaymath}
is $\alpha$-H\"older continuous with H\"older constant bounded in
terms of $L$ and the $\alpha$-H\"older constant of
$\nu_{\mathbb{H}}$. Then
\begin{align}\label{eq:HolForm}
\left|\ell(p)\nu_{\mathbb{H}}(p)-\ell(p')\nu_{\mathbb{H}}(p')
\right|&\leq \ell(p)
\left|\nu_{\mathbb{H}}(p)-\nu_{\mathbb{H}}(p') \right|+
\left|{\ell(p)}-{\ell(p')}\right|\lesssim_L d(p,p')^{\alpha}
\end{align}
for $p,p'\in S$. Now we are quite well equipped to check that
$\varphi$ satisfies \eqref{CFOdef}. To this end, fix $w \in \mathbb{W}$ and $p \in \Phi(\mathbb{W})$. Observe the
explicit formula
\begin{equation}\label{eq:translatedGrad} \nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(w) = \nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p \cdot w)),
\qquad w \in \mathbb{W}, \: p \in \Phi(\mathbb{W}), \end{equation} proven in Lemma \ref{invintrgrad} below. Thus, starting from the left hand
side of \eqref{CFOdef},
\begin{align}
|\nabla^{\varphi^{(p^{-1})}} \varphi^{(p^{-1})}(w) - \nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(0)|
& = |\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p \cdot w)) - \nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))| \notag \\
&= |\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(\Phi[\pi_{\mathbb{W}}(p \cdot
w)])) - \nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))| \notag
\\&\label{form33} \overset{\eqref{eq:HolForm}}{\lesssim}_L d_{\mathbb{H}}(\Phi(\pi_{\mathbb{W}}(p \cdot w)),p)^{\alpha}. \end{align}
The following formula is needed to make sense of the right hand side:
\begin{lemma}\label{l:translated} For any $p \in \mathbb{H}^n$ and $w \in \mathbb{W}$, the following relation holds:
\begin{equation}\label{form36} \Phi(\pi_{\mathbb{W}}(p \cdot w)) = p \cdot \Phi^{(p^{-1})}(w). \end{equation}
Here $\Phi^{(p^{-1})}$ is the graph map of $\varphi^{(p^{-1})}$.
\end{lemma}
\begin{proof} The function $\varphi^{(p^{-1})} \colon \mathbb{W} \to \mathbb{V}$ is explicitly given by
\begin{equation}\label{form34} \varphi^{(p^{-1})}(w) = \pi_{\mathbb{V}}(p^{-1}) \cdot \varphi(\pi_{\mathbb{W}}(p \cdot w)), \end{equation}
where $\pi_{\mathbb{V}}(x_1,\ldots,x_{2n},t) = x_1$ is the horizontal
projection induced by the splitting $\mathbb{H}^n = \mathbb{W} \cdot \mathbb{V}$, see for
instance \cite[Lemma 4.7]{CFO} (for $n=1$). The map $\pi_{\mathbb{V}}$ is
a group homomorphism $\mathbb{H}^n \to \mathbb{V}$ with $\pi_{\mathbb{V}}(w) = 0$ for all
$w = (0,x_2,\ldots,x_{2n},t) \in \mathbb{W}$. Therefore,
\begin{align*} & \Phi(\pi_{\mathbb{W}}(p \cdot w)) = p \cdot \Phi^{(p^{-1})}(w)\\
& \quad \overset{\eqref{form34}}{\Longleftrightarrow} \pi_{\mathbb{W}}(p \cdot w) \cdot \varphi(\pi_{\mathbb{W}}(p \cdot w)) = p \cdot [w \cdot \pi_{\mathbb{V}}(p^{-1}) \cdot \varphi(\pi_{\mathbb{W}}(p \cdot w))]\\
& \quad \Longleftrightarrow \pi_{\mathbb{W}}(p \cdot w) = p \cdot w \cdot \pi_{\mathbb{V}}(p^{-1}) = p \cdot w \cdot [\pi_{\mathbb{V}}(p)]^{-1} \\
& \quad \Longleftrightarrow p \cdot w = \pi_{\mathbb{W}}(p \cdot w) \cdot
\pi_{\mathbb{V}}(p) = \pi_{\mathbb{W}}(p \cdot w) \cdot \pi_{\mathbb{V}}(p \cdot w).
\end{align*} The last equation is a true true by the definition of
the projections $\pi_{\mathbb{W}}$ and $\pi_{\mathbb{V}}$, so the proof is
complete. \end{proof}
We deduce that the right hand side of \eqref{form33} equals, up to
a multiplicative constant, the term
$d_{\mathbb{H}}(\Phi^{(p^{-1})}(w),0)^{\alpha}$, which we will now
further bound from above in order to conclude the proof of the
first implication in Proposition \ref{equivProp}. To do so, we
observe that \cite[Theorem 1.2]{MR2223801} implies more than mere
intrinsic differentiability of $\varphi$: it shows that $\varphi$
is \emph{uniformly intrinsically differentiable} in the sense of
\cite[Definition 3.16]{MR2496655}. Recalling that $\varphi$ has
compact support, this implies that there exists a function
$\varepsilon=\varepsilon_{\mathrm{spt}\varphi}:(0,\infty) \to
(0,\infty)$ with $\lim_{s\to 0}\varepsilon(s)=0$ such that
\begin{displaymath}
|\varphi^{(p^{-1})}(y,t) - \langle \nabla^{\varphi}\varphi(w_0),y\rangle|
\leq\varepsilon\left(\|(y,t)\|\right) \|(y,t)\|, \qquad
p=\Phi(w_0)\in \Phi(\mathbb{W}),\;(y,t) \in
\mathrm{spt}\,\varphi^{(p^{-1})}.
\end{displaymath}
Then there is a constant $\delta>0$ such that
\begin{displaymath}
|\varphi^{(p^{-1})}(y,t)|\leq |\varphi^{(p^{-1})}(y,t) - \langle
\nabla^{\varphi}\varphi(w_0),y\rangle|+|\langle
\nabla^{\varphi}\varphi(w_0),y\rangle|\leq (1+L)\|(y,t)\|
\end{displaymath}
for all $p=\Phi(w_0)\in \Phi(\mathbb{W})$ and all $(y,t)\in
\mathrm{spt}\,\varphi^{(p^{-1})}$ with $\|(y,t)\|\leq \delta$,
where $L:=\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})}$. On the
other hand, if $\|(y,t)\|\in \mathrm{spt}\,\varphi^{(p^{-1})}$
satisfies $\|(y,t)\|> \delta$, then trivially,
\begin{displaymath}|\varphi^{(p^{-1})}(y,t)| = \|\pi_{\mathbb{V}}(\Phi(w_0)^{-1})\cdot \varphi(\pi_{\mathbb{W}}(\Phi(w_0)\cdot w)\|
\leq \tfrac{2\|\varphi\|_{L^{\infty}(\mathbb{W})}}{\delta} \|(y,t)\|. \end{displaymath}
In conclusion, there exists a constant $L'\geq 1$ such that for
all $p\in \Phi(\mathbb{W})$, it holds
\begin{equation}\label{eq:LipConclu}
\|\Phi^{(p^{-1})}(w)\|\leq \|w\|+ |\varphi^{(p^{-1})}(w)|\leq L'
\|w\|,\quad w\in \mathbb{W}.
\end{equation}
Finally, plugging formula \eqref{form36} into \eqref{form33} and
using the left-invariance of $d_{\mathbb{H}}$,
\begin{equation}\label{form35} |\nabla^{\varphi^{(p^{-1})}} \varphi^{(p^{-1})}(w) - \nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(0)|
\lesssim d_{\mathbb{H}}(\Phi^{(p^{-1})}(w),0)^{\alpha}
\overset{\eqref{eq:LipConclu}}{\lesssim} \|w\|^{\alpha}.
\end{equation}
Now \eqref{form35} proves that $\varphi$ satisfies \eqref{CFOdef}.
The converse implication stated in Proposition \ref{equivProp} is
not needed in the paper, so we only sketch the argument. Let
$\varphi$ be intrisically differentiable with intrinsic gradient satisfying
\eqref{CFOdef}. Then $\varphi$ is again uniformly intrinsically
differentiable. This is a consequence of Proposition
\ref{p:Approx},
see also \cite[Remark 2.24]{CFO2}. Therefore, according to \cite[Theorem 1.2]{MR2223801},
the intrinsic graph $S = \Phi(\mathbb{W})$ is an $\mathbb{H}$-regular surface, and the condition
(ii) in Definition \ref{C1alphaGraphs} holds.
So, it remains to check that the horizontal normal of $S$ is $\alpha$-H\"older continuous.
This follows from the expression \eqref{form32} (which is available by \cite[Theorem 1.2]{MR2223801})
using estimates similar to those above \eqref{form39}.
\end{proof}
The proof of Proposition \ref{equivProp} referred to
the following auxiliary lemma.
\begin{lemma}\label{invintrgrad}
Let $\varphi :\mathbb{W} \to \mathbb{V}$ be defined on the codimension-$1$ vertical subgroup
$\mathbb{W}\subset \mathbb{H}^n$. If $p\in \mathbb{H}^n$ and $w\in \mathbb{W}$ are such
that $\varphi$ is intrinsically differentiable at $\pi_{\mathbb{W}}(p\cdot
w)$, then
\begin{equation}\label{eq:LeftInvGrad} \nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(w) = \nabla^{\varphi}\varphi(\pi _\mathbb{W}(p\cdot w)). \end{equation}
\end{lemma}
\begin{proof}
By its very definition, a function $\psi:\mathbb{W}\to
\mathbb{V}$ is intrinsically differentiable at a point $w_0\in
\mathbb{W}$ if and only if $\psi^{(p_0^{-1})}$ for $p_0=w_0\cdot
\psi(w_0)$ is intrinsically differentiable at $0$, and in that
case
\begin{equation}\label{eq:LeftInvGrad}
\nabla^{\psi}\psi(w_0)=
\nabla^{\psi^{(p_0^{-1})}}\psi^{(p_0^{-1})}(0),
\end{equation}
see, for instance, \cite[Definition 4.71, Proposition 4.76.]{MR3587666} and \cite[top of p.192]{MR2223801}.
Now fix $\varphi$, and points $w \in \mathbb{W}$, and $p \in \mathbb{H}^n$ as in
the statement of the lemma. We first apply formula
\eqref{eq:LeftInvGrad} to $\psi:=\varphi$ and $w_0:=\pi_{\mathbb{W}}(p
\cdot w)$. Hence,
\begin{displaymath}
p_0 = w_0 \cdot \varphi(w_0) = \pi_{\mathbb{W}}(p \cdot w) \cdot
\varphi(\pi_{\mathbb{W}}(p \cdot w)) = p \cdot w\cdot \pi_{\mathbb{V}}(p^{-1})
\cdot \varphi(\pi_{\mathbb{W}}(p \cdot w))\overset{\eqref{form34}}{=} p
\cdot w\cdot \varphi^{(p^{-1})}(w)
\end{displaymath}
and \eqref{eq:LeftInvGrad} reads
\begin{equation}\label{eq:grad1}
\nabla^{\psi}\psi(\pi _\mathbb{W}(p\cdot w))= \nabla^{\varphi^{([ p \cdot
w\cdot \varphi^{(p^{-1})}(w)]^{-1})}}\varphi^{([ p \cdot w\cdot
\varphi^{(p^{-1})}(w)]^{-1})}(0).
\end{equation}
On the other hand, denoting \begin{displaymath} q_0 = w \cdot
\varphi^{(p^{-1})}(w),
\end{displaymath}
we observe that the graph of $[\varphi^{(p^{-1})}]^{(q_0^{-1})}$
is $q_0^{-1}\cdot \Gamma'$, where $\Gamma'$ is the graph of
$\varphi^{(p^{-1})}$, so
\begin{displaymath}
q_0^{-1}\cdot \Gamma'= q_0^{-1} \cdot p^{-1}\cdot
\Phi(\mathbb{W})=[p\cdot q_0]^{-1} \cdot \Phi(\mathbb{W}).
\end{displaymath}
This shows that
\begin{displaymath}
\varphi^{([p\cdot q_0]^{-1})}= [\varphi^{(p^{-1})}]^{(q_0^{-1})}
\end{displaymath}
and hence
\begin{displaymath}
\eqref{eq:grad1}= \nabla^{\varphi^{([p\cdot
q_0]^{-1})}}\varphi^{([p\cdot q_0]^{-1})}(0)=
\nabla^{[\varphi^{(p^{-1})}]^{(q_0^{-1})}}[\varphi^{(p^{-1})}]^{(q_0^{-1})}(0).
\end{displaymath}
In particular, $[\varphi^{(p^{-1})}]^{(q_0^{-1})}$ is
intrinsically differentiable at $0$. Formula
\eqref{eq:LeftInvGrad} applied to $\psi:=\varphi^{(p^{-1})}$,
$p_0=q_0$ and $w_0:=w$ yields
\begin{equation}\label{eq:grad2}
\nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(w) =
\nabla^{[\varphi^{(p^{-1})}]^{(q_0^{-1})}}[\varphi^{(p^{-1})}]^{(q_0^{-1})}(0). \end{equation}
The lemma follows by observing that the right hand sides of
\eqref{eq:grad1} and \eqref{eq:grad2} are equal.
\end{proof}
With the main definitions now in place, we repeat Theorem
\ref{mainIntro} below.
\begin{thm}\label{mainGraphs} Let $S = \Phi(\mathbb{W}) \subset \mathbb{H}^n$, where $\varphi \in C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ is compactly supported.
Then $S$ has BP$G$BI in the sense of Definition \ref{BPdef}.
\end{thm}
Before proving this, let us deduce the qualitative corollary,
Theorem \ref{mainQualitative}:
\begin{thm}\label{surfaces} Let $S \subset \mathbb{H}^n$ be a $C^{1,\alpha}_{\mathbb{H}}$-surface. Then
$\mathcal{H}^{2n+1}$ almost all of $S$ can be covered by
bilipschitz images of closed subsets of codimension-1 vertical
subgroups. In particular, $S$ is LI rectifiable. \end{thm}
\begin{proof} There are (at least) two possible approaches. One is to use the implicit function theorem \cite[Theorem 6.5]{FSSC}
to express the surface $S$ locally as the intrinsic graph of a \emph{locally defined} intrinsic $C^{1,\alpha}$-function.
This function does not, literally, satisfy the assumptions of Theorem \ref{mainGraphs}, but the proof of Theorem \ref{mainGraphs} could be localised with some effort.
The biggest difficulty in this approach is of expository nature as
localising the proof of Theorem \ref{mainGraphs} would lead to a
more cumbersome version of
Proposition \ref{p:Approx} below. So we take the following alternative route: in Appendix
\ref{extensionAppendix}, we show that every point on $S$ has a
neighbourhood which can be contained on the intrinsic graph
$\Gamma$ of a globally defined, compactly supported intrinsic
$C^{1,\alpha/3}$-function. Since $\Gamma$ is LI rectifiable by
Theorem \ref{mainGraphs}, the same is also true for $S$.
\end{proof}
It remains to prove Theorem \ref{mainGraphs}. Recall the notation
$\mathbb{V}$ and $\mathbb{W}$ from Definition \ref{C1alphaGraphs}. We will
frequently abbreviate $(0,x_2,\ldots,x_{2n},t) =:
(x_2,\ldots,x_{2n},t)=:(y,t)$ for points in $\mathbb{W}$, and continue to
use the notation
\begin{displaymath}
\langle \nabla^{\varphi}\varphi(w),y\rangle =\sum_{i=2}^{2n}
D_i^{\varphi} \varphi(w) x_i
\end{displaymath}
for an intrinsically differentiable $\varphi$. The functions
$D_i^{\varphi} \varphi$ have been introduced as the components of
the intrinsic gradient $\nabla^{\varphi}\varphi$ in Definition
\ref{d:IntrGrad}, but under additional regularity assumptions on
$\varphi$, they can also be obtained as derivatives of $\varphi$
in the direction of the vector fields
\begin{displaymath}
D^{\varphi}_j: = X_j,\quad j\in \{2,\ldots,2n\}\setminus
\{n+1\}\quad\text{and}\quad D^\varphi _{n+1}: = \partial_{x_{n+1}}
+ \varphi
\partial_t.
\end{displaymath} A first result of this type is \cite[Proposition
3.7]{MR2223801}, and more specific statements will follow shortly
in the proof of the next result, which is a key ingredient in the
proof of Theorem \ref{mainGraphs}.
\begin{proposition}\label{p:Approx} Assume that $\varphi:\mathbb{W} \to \mathbb{V}$
is intrinsically differentiable on $\mathbb{W}$ and it has a continuous
intrinsic gradient which satisfies \eqref{CFOdef} with constant
$H\geq 1$ and $0<\alpha\leq 1$. Suppose further that
$L:=\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})}<\infty$, and $p =
\Phi(w)$ for some $w \in \mathbb{W}$. Then,
\begin{equation}\label{CFOEstimate} |\varphi^{(p^{-1})}(y,t) - \langle \nabla^{\varphi}\varphi(w),y\rangle|
\lesssim \|(y,t)\|^{1 + \alpha}, \qquad (y,t) \in \mathbb{W}.
\end{equation} The implicit constant in \eqref{CFOEstimate} only
depends on $H$ and $L$.
\end{proposition}
The proposition says, in a "left-invariant" way, that $\varphi$ is
locally well-approximated by linear functions. The main corollary
is Proposition \ref{approxProp}, which quantifies how well the
intrinsic graph of $\varphi$ around $\Phi(w)$ is approximated by
the \emph{vertical tangent plane} determined by
$\nabla^{\varphi}\varphi(w)$.
\begin{proof}[Proof of Proposition \ref{p:Approx}]
For $n=1$ the proposition was established in \cite[Proposition
2.23]{CFO2}. The case $n>1$ can be proven in a simpler way without
the arguments that were used in \cite[Proposition 4.2]{CFO2}. We
include here a self-contained proof for that case. By the
definition of the intrinsic differentiability and intrinsic
gradients, we have
\begin{equation*}
|\varphi ^{(p^{-1})}(y,t) -
\langle\nabla^{\varphi}\varphi(w),y\rangle | = |\varphi
^{(p^{-1})} (y,t) - \langle \nabla ^{\varphi ^{(p^{-1})}} \varphi
^{(p^{-1})}(0) , y\rangle |
\end{equation*}
and so we just prove that
\begin{equation}\label{finale1}
|\varphi (y,t) - \langle \nabla ^{\varphi } \varphi (0), y \rangle| \lesssim \| (y,t)\|^{\alpha + 1}, \quad \mbox{ for all } (y,t) \in \mathbb{W},
\end{equation}
under the assumption that $p=0$ and $\varphi (0)=0$. Notice that
the constants $L$ and $H$ are not changed under left translations.
To explain the idea, let us first consider a $C^{1,\alpha}(\mathbb{R})$
function $h:\mathbb{R} \to \mathbb{R}$ with $h(0)=0$. Then, for $y>0$,
\begin{displaymath}
|h(y)-h'(0)y|=\left|\int_0^y h'(s)-h'(0)\,ds\right|\leq \int_0^y
|h'(s)-h'(0)|\,ds \lesssim |y|^{1+\alpha}.
\end{displaymath}
To prove \eqref{finale1}, we apply the same idea, but we integrate
along integral curves of vector fields $D_j^{\varphi}$,
$j=2,\ldots,2n$. We use the assumption $n>1$ to ensure that the
origin can be connected to any point
\begin{displaymath}
(y,t)=(0,x_2,\ldots,x_{2n},t)\in \mathbb{W} \end{displaymath} by a curve
$\gamma :I \to \mathbb{W}$ that is defined as a
concatenation $\gamma := \gamma_1 \star \cdots \star \gamma
_{2n+3} $ of the following curves:
\begin{itemize}
\item $\gamma_1$ is an integral curve of
\begin{displaymath}
D^\varphi _{n+1} = \partial_{x_{n+1}} + \varphi \partial_t
\end{displaymath}
that connects $0$ to a point of the form
\begin{displaymath}a:= (0,\ldots , 0 ,x_{n+1},0,\ldots,0, \tau
(x_{n+1})).\end{displaymath} \item $ \gamma _2 \star \dots \star
\gamma _{2n+3}$ is a concatenation of integral curves of
\begin{displaymath}
D^{\varphi}_j = X_j,\quad j\in \{2,\ldots,2n\}\setminus \{n+1\}
\end{displaymath}
with the property that
$$\mbox{length}_{\mathbb{H}^n} ( \gamma _2 \star \dots \star
\gamma _{2n+3}) \lesssim \|(y,t)\|$$ and $ \gamma _2 \star \dots
\star \gamma _{2n+3}$ connects $a$ to $(y,t)$.
\end{itemize}
A similar construction was used in \cite[Proposition
6.10]{antonelli2019pauls}. We now explain in detail how
$\gamma_1,\ldots,\gamma_{2n+3}$ are defined. First, let
$\lambda_{n+1}$ be an integral curve of $D_{n+1}^{\varphi}$, given
by
\begin{equation}\label{eq:lambda_n+1}
\lambda_{n+1}(s):= (0,\ldots 0,s,0,\ldots,\tau (s)),
\end{equation} where $s$ is the component corresponding to the
coordinate $x_{n+1}$, and $\tau: J\to \mathbb{R}$ is a solution of the
Cauchy problem
\begin{equation*}
\left\{
\begin{array}{lcl}
\tau '(s)= \varphi(0, \ldots , 0 , s, 0,\ldots,0,\tau (s)), \\
\tau (0)=0. \\
\end{array}
\right.
\end{equation*}
Such a solution exists by Peano's theorem since $\varphi$ is
continuous, and we assume that $J$ is the maximal interval of
existence for $\tau$ containing the point $0$. As moreover
$\nabla^{\varphi}\varphi$ is continuous, which follows from the
assumption \eqref{CFOdef} by Remark \ref{r:GradProp}, we find by
\cite[Lemma 4.4]{CFO2} that
\begin{equation}\label{(6.31).2:_n+1}
\begin{aligned}
(\varphi \circ \lambda _{n+1})' (s) &= D^\varphi_{n+1}\varphi (\lambda _{n+1}(s)),\quad s\in J. \\
\end{aligned}
\end{equation}
To be precise, \cite[Lemma 4.4]{CFO2} is stated in $\mathbb{H}^1$,
but since $\lambda_{n+1}$ is entirely contained in the
$x_{n+1}t$-plane, we can apply the result to
\begin{displaymath}
(x_{n+1},t)\mapsto \varphi(0,\ldots,0,x_{n+1},0,\ldots,0,t),
\end{displaymath}
interpreted as a function on a vertical subgroup in
$\mathbb{H}^1$. Moreover, using the same proof as for
\cite[(4.4)]{CFO2}, one can show that in fact $J=\mathbb{R}$, and
\begin{equation}\label{(6.31)}
|\tau (s) | \lesssim _L |s|^2, \qquad s\in \mathbb{R}.
\end{equation}
The curve $\gamma_1$ will be an appropriately parametrized
subcurve of $\lambda_{n+1}$. If $x_{n+1}>0$, we naturally define
$\gamma_1$ to be the restriction of $\lambda_{n+1}$ to the
interval $[0,x_{n+1}]$. If $x_{n+1}<0$, we have to reverse the
order of the parametrization, so we make the general definition:
$\gamma_1(s):= \lambda_{n+1}(\mathrm{sgn}(x_{n+1})s)$, $s\in
[0,|x_{n+1}|]$, if $x_{n+1}\neq 0$, and we let $\gamma_1$ be the
constant curve $\gamma_1\equiv 0$ otherwise. Note that the points
$a:=\gamma_1(|x_{n+1}|)$ and $(y,t)$ belong to
\begin{displaymath}
\mathbb{G}:=\{(z_1,\ldots,z_{2n+1})\in \mathbb{R}^{2n+1}:\,
z_1=0\quad\text{and}\quad z_{n+1}=x_{n+1}\},
\end{displaymath}
and $\mathbb{G}$ with the group law and metric induced from
$\mathbb{H}^n$ is isometrically isomorphic to $\mathbb{H}^{n-1}$
with $d_{\mathbb{H}}$. We next show that there exists a compact
interval $I=[|x_{n+1}|,b]$ such that $a$ and $(y,t)$ can be
connected by a concatenation $\gamma _2 \star \dots \star \gamma
_{2n+3}:I\to \mathbb{G}$ of integral curves of $D^\varphi _j=X_j$
, $j \in \{2,\dots, 2n \}\setminus\{n+1\}$ so that
\begin{equation}\label{equation1803}
\max_{s\in I} \|(\gamma _2 \star \dots \star \gamma _{2n+3}) (s)
\| \lesssim_L \|(y,t)\| \mbox{\,\, and \,\, length}_{\mathbb{H}^n}
( \gamma _2 \star \dots \star \gamma _{2n+3}) \lesssim
d_{\mathbb{H}}(a,(y,t)).
\end{equation}
The first inequality in \eqref{equation1803} follows from the
second one since
\begin{equation}\label{eq:a_2}
\|a\|\overset{\eqref{(6.31)}}{\lesssim_L } |x_{n+1}|\quad
\text{and}\quad |x_{n+1}|\leq\|(y,t)\|
\end{equation}
and
\begin{align*}\|(\gamma _2 \star \dots \star \gamma
_{2n+3}) (s) \| &\leq \mathrm{length}_{\mathbb{H}^n} ( \gamma _2
\star \dots \star \gamma _{2n+3})+\|a\|,\quad s\in I.
\end{align*}
Thus it suffices to find curves which satisfy the second condition
in \eqref{equation1803}. To achieve this, first concatenate curves
$\gamma_2,\ldots,\gamma_{2n-1}$ in the following way: follow the
unique integral curve of $D^\varphi _2 =X_2$ starting at $a$ for
time $x_2$, then follow the integral curve of $D^\varphi _{j}
=X_{j}$ for time $x_{j}$, etc.\ for $j=3,\ldots,n,n+2,\ldots,2n$,
until you reach $a' =(y,\tau (x_{n+1}) + \frac 1 2 \sum_{i=2}^n
x_{i}x_{n+i})$. Then connect $a'$ to $(y,t)$ by a curve $
\gamma_{2n}\star \dots \star \gamma _{2n+3}$ with
\begin{displaymath}
\mathrm{length}_{\mathbb{H}^n}(\gamma_{2n}\star \dots \star \gamma
_{2n+3}) \lesssim \left|t-\left(\tau (x_{n+1}) + \tfrac{ 1}{ 2}
\sum_{i=2}^n x_{i}x_{n+i}\right)\right|^{1/2}.
\end{displaymath}
This is possible since $[X_2,X_{n+2}]=\partial_t$. Now $\gamma _2
\star \dots \star \gamma _{2n+3}$ connects $a$ to $(y,t)$ as
desired, and
\begin{equation}\label{nuvola}
\begin{aligned}
\mbox{length}_{\mathbb{H}^n} ( \gamma_{2} \star \dots \star \gamma _{2n+3})& \lesssim |x_2|+\cdots+|x_{2n}|+
\left|t-\left(\tau (x_{n+1}) + \tfrac{ 1 }{2}
\sum_{i=2}^n x_{i}x_{n+i}\right)\right|^{1/2}\\
& \lesssim d_{\mathbb{H}}(a,(y,t)) + |\tau (x_{n+1}) -t|^{1/2}\\
& \lesssim d_{\mathbb{H}}(a,(y,t)).\\
\end{aligned}
\end{equation}
Hence we have found curves $\gamma_2,\ldots,\gamma_{2n+3}$ so that
the conditions in \eqref{equation1803} are satisfied.
Now we are ready to implement the idea explained at the beginning
of the proposition. Namely, we will write $\varphi(y,t)$ as an
integral of $(\varphi \circ \gamma)'$, where $\gamma$ is the
concatenation of the curves $\gamma_1,\ldots,\gamma_{2n+3}$. To
relate this to the intrinsic gradient, we apply again the
intrinsic differentiability of $\varphi$, and observe that
\begin{equation}\label{(6.31).2}
\begin{aligned}
(\varphi \circ \lambda _j)' (s) &= D^\varphi_j\varphi (\lambda _j(s)),\\
\end{aligned}
\end{equation}
for all $j=2,\ldots,2n$, where $\lambda_{n+1}$ is as in
\eqref{eq:lambda_n+1}, and $\lambda_j$ for $j\neq n+1$ is an
arbitrary integral curve of $D_j^{\varphi}=X_j$. For $j=n+1$, we
saw \eqref{(6.31).2} already in \eqref{(6.31).2:_n+1}. For $j\neq
n+1$, the vector field $ D^\varphi_j$ is linear and independent of
$\varphi$, and \eqref{(6.31).2} follows directly from the
intrinsic differentiability assumption by the argument given at
the beginning of the proof of \cite[Proposition 3.7]{MR2223801}.
Hence, if $\gamma$ is the piecewise $C^1$ curve given by
$$\gamma (s)=(\gamma_1 \star \cdots \star \gamma
_{2n+3}) (s)=(y(s),t(s)),\quad s\in[0,b],$$ we get that $\varphi
\circ \gamma$ is piecewise $C^1$ and
\begin{equation}\label{stima702}
\begin{aligned}
|\varphi(y,t) - \langle \nabla ^{\varphi } \varphi (0) ,y \rangle| & =
\left| \int_0^b (\varphi \circ \gamma )' (s)\, \mathrm{d} s -\int _0^b \langle \nabla ^{\varphi } \varphi (0) , \dot{y}(s)\rangle \, \mathrm{d} s \right|\\
& \overset{\eqref{(6.31).2}}{\leq} \int _0^b | \nabla ^\varphi \varphi (\gamma (s) ) - \nabla ^\varphi \varphi (0)| \, \mathrm{d} s, \\
& \lesssim _H \int _0^b \|\gamma (s)\|^\alpha \, \mathrm{d} s \\
\end{aligned}
\end{equation}
where in the last inequality we have used the assumption that
$\varphi$ satisfies \eqref{CFOdef}. In the first inequality, we
used the fact that for almost all $s\in [0,b]$, the tangent vector
$\dot{y}(s)$ exists by construction, and is of the form
$(0,\ldots,0,\pm 1,0,\ldots,0)$ with $$(\varphi \circ
\gamma)'(s)\overset{\eqref{(6.31).2}}{=}
D_j^{\varphi}\varphi(\gamma(s))=\langle
\nabla^{\varphi}\varphi(\gamma(s)),\dot{y}(s)\rangle$$ if the
$x_j$-component of $\dot{y}(s)$ is $+1$, and $$(\varphi \circ
\gamma)'(s)\overset{\eqref{(6.31).2}}{=}
-D_j^{\varphi}\varphi(\gamma(s))=\langle
\nabla^{\varphi}\varphi(\gamma(s)),\dot{y}(s)\rangle$$ if the
$x_j$-component of $\dot{y}(s)$ is $-1$. Having established
\eqref{stima702}, we will estimate from above the expression
\begin{equation*}
\int _0^b \|\gamma (s)\|^\alpha \, \mathrm{d} s = \int _0^{|x_{n+1}|} \|\gamma _1(s)\|^\alpha \, \mathrm{d} s +
\int _{|x_{n+1}|}^b \| (\gamma _2 \star \dots \star \gamma _{2n+3}) (s)\|^\alpha \,
\mathrm{d} s.
\end{equation*}
Firstly, since \eqref{(6.31)} holds, we have
\begin{equation*}
\begin{aligned}
\int _0^{|x_{n+1}|} \|\gamma _2(s)\|^\alpha \, \mathrm{d} s & \lesssim
\int _0^{|x_{n+1}|} |s |^\alpha + \sqrt{|\tau (\mathrm{sgn}(x_{n+1})s)|^\alpha } \, \mathrm{d} s \lesssim _L \| (y,t)\|^{\alpha +1} \\
\end{aligned}
\end{equation*}
and secondly, by definition of $\gamma_j$, $j \in \{2,\dots, 2n
+3\} $, we have
\begin{equation*}\begin{aligned}
|b-|x_{n+1}||& = \mbox{length}_{\mathbb{H}^n} (\gamma _2 \star
\dots \star \gamma _{2n+3})
\stackrel{\eqref{eq:a_2},\eqref{nuvola}}{ \lesssim }_L \| (y,t)\|. \\
\end{aligned}
\end{equation*}
This combined with the first condition in \eqref{equation1803}
yields
\begin{equation*}
\begin{aligned}
\int _{|x_{n+1}|}^b \| (\gamma _2 \star \dots \star \gamma _{2n+3}) (s)\|^\alpha \,ds \lesssim_L \| (y,t)\|^{\alpha +1} .
\end{aligned}
\end{equation*}
Putting together \eqref{stima702} and the last estimates, we can conclude
\begin{equation*}
\begin{aligned}
|\varphi (y,t) - \langle \nabla ^{\varphi } \varphi (0), y\rangle|
& \lesssim_H \int _0^b \|\gamma (s)\|^\alpha \, \mathrm{d} s \lesssim_{H,L} \| (y,t)\|^{\alpha +1}\\
\end{aligned}
\end{equation*}
and the proof of Proposition \ref{p:Approx} is complete.
\end{proof}
\begin{remark}\label{r:constants}
Recall from Proposition \ref{equivProp} that a compactly supported
function $\varphi \in C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ (as in Theorem
\ref{mainGraphs}) satisfies \eqref{CFOdef} for some constant $H
\geq 1$. The letter $H$ will refer to this constant for the rest
of Section \ref{s:C1alpha}. We also remark that $\varphi$ is
intrinsic Lipschitz, recall Definition \ref{d:intrinsicGraphs}.
Indeed, a $C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ function is intrinsically
differentiable with continuous intrinsic gradient, and the compact
support assumption implies that $\nabla^{\varphi}\varphi \in
L^{\infty}(\mathbb{W})$. In the case $n=1$, \cite[Lemma 2.22]{CFO2} states
that then $\varphi$ is intrinsic Lipschitz. One could adapt the
proof to higher dimensions using Proposition \ref{p:Approx}, or
alternatively refer to \eqref{eq:LipConclu} to conclude also for
$n>1$ that $\varphi$ is intrinsic Lipschitz. We denote by $L$ the
maximum of the intrinsic Lipschitz constant of $\varphi$, and the
sup-norm $\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})}$. The
compact support assumption of $\varphi \in C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$
is initially needed to ensure that $\max\{H,L\} < \infty$.
However, the constants are then left-invariant: if $p \in \mathbb{H}$,
then $\varphi^{(p^{-1})}$ is intrinsically differentiable, its
intrinsic gradient is continuous, and satisfies \eqref{CFOdef}
with the same constant $H$ (see
\cite[Lemma 2.25]{CFO2}),
$\|\nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}\|_{L^{\infty}(\mathbb{W})}\leq
L$ by Lemma \ref{invintrgrad}, and $\varphi^{(p^{-1})}$ is
intrinsic Lipschitz with constant $L$, even though the support of
$\varphi^{(p^{-1})}$ of course depends on $p$.
\end{remark}
We use Proposition \ref{p:Approx} to quantify how well the
intrinsic graph $S\subset \mathbb{H}^n$ of a compactly supported
$C^{1,\alpha}_{\mathbb{H}}$ function is approximated at a point $p\in S$
by a certain vertical plane. For $p \in S$, let $\mathbb{W}_p = W_p
\times \mathbb{R}$ be the unique vertical subgroup with the
property that $W_p$ is a $(2n-1)$ dimensional subspace of
$\mathbb{R}^{2n}$ which is perpendicular to the line spanned by
the vector $\nu_{\mathbb{H}}(p)$ using coordinates as in \eqref{form32}.
\begin{proposition}\label{approxProp} Fix $n\in \mathbb{N}$ and $0<\alpha\leq 1$. Let $\varphi \in C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ be compactly supported
on the codimension-$1$ vertical subgroup $\mathbb{W}\subset \mathbb{H}^n$, and
write $S := \Phi(\mathbb{W})$. Then, there exists a constant $A = A(H,L)
\geq 1$ such that for every $p \in S$,
\begin{equation}\label{form20} \operatorname{dist}_{\mathbb{H}}(q,S) \leq A d_{\mathbb{H}}(p,q)^{1 + \alpha}, \qquad q \in p \cdot \mathbb{W}_{p}. \end{equation}
Here $H$ and $L$ are defined as in Remark \ref{r:constants}, that
is, $\varphi$ satisfies \eqref{CFOdef} with constant $H$,
$\varphi$ is intrinsic $L$-Lipschitz, and
$\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})}\leq L$.
\end{proposition}
\begin{proof}[Proof of Proposition \ref{approxProp}] The plan is to apply estimate \eqref{CFOEstimate} from Proposition \ref{p:Approx}. Note that
\begin{displaymath}
L_{p}(y,t) := (0,y,t) \cdot
(\langle\nabla^{\varphi}\varphi(w),y\rangle,0,0)
\end{displaymath}
defines a map $\mathbb{W} \to \mathbb{W}_{p}$ since
\begin{equation}\label{eq:W_p}
W_p = \left\{\left(\sum_{i=2}^{2n} D_i^\varphi\varphi(w)x_i,x_2,\ldots,x_{2n}\right):\; (x_2,\ldots,x_{2n})\in \mathbb{R}^{2n-1}\right\}
\end{equation}
is a $(2n-1)$-plane perpendicular to
\begin{displaymath}
\begin{pmatrix}-1\\D_2^{\varphi}\varphi(w)\\\vdots\\ D_{2n}^{\varphi}\varphi(w)\end{pmatrix}
\end{displaymath}
in $\mathbb{R}^{2n}$.
In fact, $L_{p}$ is a bijection $\mathbb{W} \to \mathbb{W}_{p}$, and for all $ (y,t)=(0,x_2,\ldots,x_{2n},t) \in
\mathbb{W}$,
\begin{equation}\label{form22} \|L_{p}(y,t)\| = \left\|\left(\langle\nabla^{\varphi}\varphi(w),y\rangle,y,t -
\tfrac{1}{2}x_{n+1}\langle\nabla^{\varphi}\varphi(w),y\rangle\right)\right\|
\sim_{L} \|(y,t)\|,
\end{equation} because $|\nabla^{\varphi}\varphi(w)| \leq
\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})} \leq L$. The last
observation immediately implies the inequality ``$\lesssim_L$'' in
\eqref{form22}. It also gives the converse inequality, since
\begin{align*}
\|(y,t)\|&\lesssim |y| + |t|^{\frac{1}{2}}\lesssim |y| + \left|t -
\tfrac{1}{2}x_{n+1}\langle\nabla^{\varphi}\varphi(w),y\rangle\right|^{\frac{1}{2}}
+
\left|\tfrac{1}{2}x_{n+1}\langle\nabla^{\varphi}\varphi(w),y\rangle\right|^{\frac{1}{2}}\\&\lesssim_L
|y|+ \left|t -
\tfrac{1}{2}x_{n+1}\langle\nabla^{\varphi}\varphi(w),y\rangle\right|^{\frac{1}{2}}
\lesssim_L \|L_{p}(y,t)\|.
\end{align*}
Moreover, for arbitrary $w=(y,t)\in \mathbb{W}$, we have $d_{\mathbb{H}}(q,S)
\leq d_{\mathbb{H}}(q,p \cdot \Phi^{(p^{-1})}(y,t))$, where
$\Phi^{(p^{-1})}$ is the graph map of $\varphi^{(p^{-1})}$, simply
because $\Phi^{(p^{-1})}(\mathbb{W}) = p^{-1} \cdot S$. Therefore, by the
left-invariance of $d_{\mathbb{H}}$, one has for $q=p\cdot L_{p}(y,t)$,
\begin{align*} \operatorname{dist}_{\mathbb{H}}(q,S) \leq d_{\mathbb{H}}(p^{-1} \cdot q,\Phi^{(p^{-1})}(y,t))
& = d_{\mathbb{H}}(L_p(y,t),\Phi^{(p^{-1})}(y,t))\\
& = |\varphi^{(p^{-1})}(y,t) - \langle\nabla^{\varphi}\varphi(w),y\rangle| \stackrel{\eqref{CFOEstimate}}{\lesssim}_{H,L} \|(y,t)\|^{1 + \alpha} \notag\\
& \stackrel{\eqref{form22}}{\sim}_{H,L} \|L_{p}(y,t)\|^{1 +
\alpha} = d_{\mathbb{H}}(p,q)^{1 + \alpha}. \notag\end{align*} This
proves \eqref{form20}, and hence the proposition. \end{proof}
\subsubsection{Reduction to unit scale}\label{ss:red_unit_scale}
The rest of Section \ref{s:C1alpha} is devoted to the proof of
Theorem \ref{mainGraphs} in the case $n>1$. With the earlier
preparations in place, a proof for the case $n=1$ could be
obtained along the same lines, but some steps would require a
separate discussion. Instead, we will deduce the case $n=1$ later
from a more general result, see Theorem \ref{mainGraphsn_1}. So we
fix $n>1$ for the rest of this section, and constants will be
allowed to depend on $n$ without special mentioning.
Theorem \ref{mainGraphs} for $n>1$ is essentially a corollary of
Theorem \ref{main} applied to
\begin{equation}
\label{eq:G_M} (G,d_{G},\mu) =(\mathbb{H}^{n-1}\times
\mathbb{R},d_{\mathbb{H}^{n-1}\times \mathbb{R}},\mathcal{L}^{2n})\quad
\text{and} \quad (M,d_{M}) = (S,d_{\mathbb{H}}), \end{equation} where
$(G,d_G)$ is defined as in \eqref{eq:HeisxR} and the line below
it.
Once the hypotheses of Theorem \ref{main} have been verified -- a
task occupying the next section -- the theorem will yield the
existence of $2L'$-bilipschitz maps $f \colon K \to S \cap
B(p,1)$, $p \in S$, where $K \subset G$ with $\mathcal{H}^{2n+1}(K) \geq
\delta
>0$. The constant $\delta > 0$ will only depend on $\alpha$, the H\"older
constant $H$ in \eqref{CFOdef}, the constant $L$ that bounds the
intrinsic Lipschitz constant of $\varphi$ and the
$L^{\infty}$-norm of $\nabla^{\varphi}\varphi$, whereas the
constant $L'$ will depend only on $L$. This is saying, in particular, that the BP$G$BI condition holds at
unit scale. How about other scales? The following easy lemma shows
that property \eqref{CFOdef} improves under ``zooming in'':
\begin{lemma}\label{blowUpLemma} Let $\varphi: \mathbb{W} \to \mathbb{V}$ be
intrinsically differentiable with continuous intrinsic gradient
$\nabla^{\varphi}\varphi$ that satisfies \eqref{CFOdef} with
constants $\alpha > 0$ and $H \geq 1$. For $r>0$, let
\begin{displaymath} \varphi_{r}(w) := \tfrac{1}{r}[\varphi \circ \delta_{r}],\quad w\in \mathbb{W}. \end{displaymath}
Then, $\psi:=\varphi_{r}$ is an intrinsically differentiable
function with intrinsic graph $\delta_{1/r}(\Phi(\mathbb{W}))$, its
intrinsic gradient is continuous, and satisfies \eqref{CFOdef}
with constants $\alpha$ and $r^{\alpha}H$, that is
\begin{displaymath}
|\nabla^{\psi^{(p^{-1})}}\psi^{(p^{-1})}(w)-\nabla^{\psi^{(p^{-1})}}\psi^{(p^{-1})}(0)|\leq
r^{\alpha} H \|w\|^{\alpha},\quad w\in \mathbb{W},p\in
\delta_{1/r}(\Phi(\mathbb{W})).
\end{displaymath}
\end{lemma}
\begin{proof}
Fix $0 < r \leq 1$ and let $w$ be an arbitrary point in $\mathbb{W}$.
Then
\begin{displaymath}
\delta_{\frac{1}{r}}[w\cdot \varphi(w)] =
\delta_{\frac{1}{r}}(w)\cdot \delta_{\frac{1}{r}}(\varphi(w)) =
\delta_{\frac{1}{r}}(w)\cdot
\delta_{\frac{1}{r}}(\varphi(\delta_r(\delta_{\frac{1}{r}}(w)))),
\end{displaymath}
which shows that $\delta_{\frac{1}{r}}(\Phi(\mathbb{W}))$ is the intrinsic
graph of $\psi=\varphi_r$ as defined in the lemma. Since the
Heisenberg dilations are group isomorphisms which commute with
vertical projections, it is easy to see that $\psi$ is
intrinsically differentiable with intrinsic gradient
\begin{displaymath}
\nabla^{\psi}\psi = \nabla^{\varphi}\varphi \circ \delta_r.
\end{displaymath}
Moreover, since $\psi^{(p^{-1})} = \frac{1}{r}
\varphi^{(\delta_r(p)^{-1})} \circ \delta_r$ for $p\in
\delta_{1/r}(\Phi(\mathbb{W}))$, we have
\begin{displaymath}
\nabla^{\psi^{(p^{-1})}}\psi^{(p^{-1})}=
\nabla^{\varphi^{(\delta_r(p)^{-1})} }
\varphi^{(\delta_r(p)^{-1})} \circ \delta_r,
\end{displaymath}
which yields the remaining claims in the lemma.
\end{proof}
Returning to the proof of Theorem \ref{mainGraphs}, let $p \in S$
and, first, $0 < r \leq C$, where $C := 2\operatorname{diam}_{\mathbb{H}}(\Phi(\operatorname{spt}
\varphi))$. Using the previous lemma, and also recalling that
dilations have no effect on intrinsic Lipschitz constants,
$S_{1/r} := \delta_{1/r}(S)$ is an intrinsic graph of an intrinsic
Lipschitz function with essentially bounded intrinsic gradient
satisfying \eqref{CFOdef} with constants depending only on the
corresponding constants for $S$, and $C$.
Therefore,
by the BP$G$BI property at scale $r = 1$, to be established in the
next section, every ball $S_{1/r} \cap B(p,1)$ contains the image
of a $2L'$-bilipschitz map $g$ from a compact set $K \subset G$
with $\mathcal{H}^{2n+1}(K) \geq \delta = \delta(C) > 0$. Now, one may
simply pre- and post-compose $g$ with the natural dilations in $G$
and $\mathbb{H}^n$ to produce a $2L'$-bilipschitz map $g_{r} \colon
\delta_{r}(K) \to S \cap B(\delta_{r}(p),r)$ (note also that
$\mathcal{H}^{2n+1}(\delta_{r}(K)) = r^{2n+1}\mathcal{H}^{2n+1}(K) \geq \delta
r^{2n+1}$).
Next, consider the case $r > C$. Then, if $p \in S$ is arbitrary,
the set $S \cap B(p,r)$ satisfies
\begin{displaymath} \mathcal{H}^{2n+1}([S \cap B(p,r)] \cap \mathbb{W}) \gtrsim \mathcal{H}^{2n+1}(S \cap B(p,r)). \end{displaymath}
Thus, the restriction of $\mathrm{Id}$ to $[S \cap B(p,r)] \cap
\mathbb{W}$ (composed with an isometry $G \cong \mathbb{W}$) yields the desired
bilipschitz map in this case.
\subsection{Proof for graphs with H\"older continuous normals}\label{verification}
In this section we complete the proof of Theorem \ref{mainGraphs}.
After the reduction to unit scale in Section
\ref{ss:red_unit_scale}, it remains to verify the hypotheses of
Theorem \ref{main} for $n>1$, $(G,d_{G}):=(\mathbb{H}^{n-1}\times
\mathbb{R},d_{\mathbb{H}^{n-1}\times \mathbb{R}})$, and $(M,d_{M}) :=
(S,d_{\mathbb{H}})$, where $S = \Phi(\mathbb{W})$, as in Theorem
\ref{mainGraphs}. To be accurate, also take $x_{0} = 0 \in G$, and
fix $p_{0} \in M$ arbitrary. We start by defining the maps $i_{w
\to p} \colon G \to S$. They will not depend on the scale index $k
\geq 0$, that is, $i_{w \to p}^{k} = i_{w \to p}$ for all $k \geq
0$, and they can also be defined for all points $w \in G$, $p \in
M$ (and not only those close to $x_{0}$ and $p_{0}$).
To construct the maps
$i_{w \to p} \colon G \to S$, we will first define certain
bilipschitz maps $\Psi_p:\mathbb{W} \to \mathbb{W}_p$. We know that $\mathbb{W}$ is
isometric to $\mathbb{W}_p$, so without further restrictions, this would
be an easy task, but keeping in mind \eqref{comp}, we want to make
sure that the mappings $\Psi_p$ change in a controlled way as we
let $p$ vary in $S$. It would be possible to arrange this even for
isometric $\Psi_p$, but the construction is simpler if we allow
for bilipschitz distortion, and the main ideas are contained in
the following lemma.
\begin{lemma}\label{l:Psi_A}
For $n\geq 2$ and $D:=(a_2,\ldots,a_n,c,b_2,\ldots,b_n)\in
\mathbb{R}^{2n-1}$, define
\begin{displaymath}
\psi_D(x_2,\ldots,x_{2n}):= cx_{n+1}+\sum_{i=2}^n (a_i x_i + b_i
x_{n+i}),
\end{displaymath}
and consider the vertical subgroups
\begin{displaymath} \mathbb{W} :=
\{(x_1,\ldots,x_{2n},t)\in \mathbb{H}^n:\;x_1=0\}\text{ and }\mathbb{W} ':=
\left\{(x_1,\ldots,x_{2n},t)\in
\mathbb{H}^n:\;x_1=\psi_D(x_2,\ldots,x_{2n})\right\}.
\end{displaymath}
Then the map $\Psi_D(0,x_2,\ldots,x_{2n},t):=$
\begin{displaymath}
\left(\psi_D(x_2,\ldots,x_{2n}),b_2 x_{n+1}+x_2,\ldots,b_n
x_{n+1}+x_n,x_{n+1},-a_2 x_{n+1}+x_{n+2},\ldots,-a_n
x_{n+1}+x_{2n},t \right)
\end{displaymath}
has the following properties:
\begin{enumerate}
\item $\Psi_D:(\mathbb{W},\cdot)\to (\mathbb{W}',\cdot)$ is a group isomorphism,
\item $\Psi_D:(\mathbb{W},d_{\mathbb{H}})\to (\mathbb{W}',d_{\mathbb{H}})$ is $L_D$-bilipschitz
with $L_D$ depending continuously on $D$, \item
$d_{\mathbb{H}}(\Psi_D(w),\Psi_{D'}(w))\lesssim
\max\{|D-D'|,|D-D'|^{1/2}\}\,\|w\|$, for all $w\in \mathbb{W}$.
\end{enumerate}
\end{lemma}
\begin{proof}
We start by noting that
\begin{align*}
\psi_D(x_2,&\ldots,x_{2n})= c x_{n+1} +\sum_{i=2}^n (a_i x_i + b_i
x_{n+i} ) \\& = \psi_D(b_2x_{n+1}+x_2,\ldots,b_n
x_{n+1}+x_{n},x_{n+1},-a_2x_{n+1}+x_{n+2},\ldots,-a_n
x_{n+1}+x_{2n}),
\end{align*}
which can be used to show that $\Psi_D(\mathbb{W})=\mathbb{W}'$. It further
follows directly from the definition that $\Psi_D$ is injective,
$\Psi_D(0)=0$, and $\Psi_D(w^{-1})=(\Psi_D(w))^{-1}$. In order to
see that
\begin{equation}\label{eq:GroupHomo}
\Psi_D(w\cdot w') =\Psi_D(w)\cdot \Psi_D(w'),\quad w,w'\in \mathbb{W},
\end{equation}
it suffices to verify the identity for the last components of the
points, as the first components agree obviously by linearity of
$\Psi_D$. For $w=(0,x_2,\ldots,x_{2n},t)$,
$w'=(0,x_2',\ldots,x_{2n}',t')$, we find that
\begin{align*}
[\Psi_D&(w\cdot w')]_{2n+1}= t+t'+\tfrac{1}{2}\sum_{i=2}^n (x_i
x_{n+i}'-x_{n+i}x_i' ) \\
=&t+t'+ \tfrac{1}{2}\left(c x_{n+1} +\sum_{i=2}^n (a_i x_i + b_i
x_{n+i} ) \right)x_{n+1}'- \tfrac{1}{2}\left(c x'_{n+1} +\sum_{i=2}^n
(a_i x_i' +
b_i x_{n+i}' ) \right)x_{n+1}\\
&+\tfrac{1}{2} \sum_{i=2}^n \left((b_i x_{n+1}+x_i)(-a_i
x_{n+1}'+x_{n+i}')- (b_i x'_{n+1}+x_i')(-a_i x_{n+1}+x_{n+i}) \right)\\
=& [\Psi_D(w)\cdot \Psi_D(w')]_{2n+1},
\end{align*}
which shows \eqref{eq:GroupHomo} and thus completes the proof of
the first claim in the lemma. Using the group isomorphism property
and the fact that $\Psi_D$ commutes with Heisenberg dilations, we
next observe that
\begin{align*}
d_{\mathbb{H}}(\Psi_D(w),\Psi_D(w'))&= \|\Psi_D(w')^{-1}\cdot
\Psi_D(w)\|= \|\Psi_D(w'^{-1}\cdot w)\| \in [c_D d_{\mathbb{H}}(w,w'),
C_D d_{\mathbb{H}}(w,w')],
\end{align*}
where \begin{displaymath} c_D:= \min\{\|\Psi_D(v)\|:\;
\|v\|=1\}\quad \text{and}\quad C_D:= \max\{\|\Psi_D(v)\|:\;
\|v\|=1\}.\end{displaymath} This concludes the second part of the
lemma, up to the continuity of $D\mapsto L_D$, which will follow
from the third part. To verify the third part, let us fix $D,D'\in
\mathbb{R}^{2n-1}$, and an arbitrary point $w=(0,x_2,\ldots,x_{2n},t)$ in
$\mathbb{W}$, and compute $\Psi_{D'}(w)^{-1}\cdot \Psi_D(w)=$
\begin{align*}
(\psi_{(D-D')}(w),(b_2-b_2')x_{n+1},\ldots,(b_n-b_{n}')x_{n+1},0,(a_2'-a_2)x_{n+1},\ldots,(a_n'-a_n)x_{n+1},
\tau),
\end{align*}
where
\begin{displaymath}
\tau:= \tfrac{1}{2}x_{n+1}\left[\psi_{(D-D')}(w)+\sum_{i=2}^n
\left((b_i-b_i')x_{n+i}+(a_i-a_i')x_i \right) \right].
\end{displaymath}
This shows that
\begin{displaymath}
d_{\mathbb{H}}(\Psi_{D}(w),\Psi_{D'}(w))\lesssim
\max\{|D-D'|,|D-D'|^{1/2}\} \|w\|,
\end{displaymath}
as claimed.
\end{proof}
The mappings defined in Lemma \ref{l:Psi_A} will be used later in
the case where the components of $D$ are the entries of an
intrinsic gradient $\nabla^{\varphi}\varphi(w)$. For $p=\Phi(w)$,
we then denote
\begin{equation}\label{eq:Psi_p}
\Psi_p:=
\Psi_{(D_2^{\varphi}\varphi(w),\ldots,D_{2n}^{\varphi}\varphi(w))},
\end{equation}
so that $\Psi_p(\mathbb{W})=\mathbb{W}_p$ is the vertical plane appearing in
Proposition \ref{approxProp}.
\begin{remark}
It is important to note that $\Psi_p$ is different from the
obvious parametrization $L_p:\mathbb{W} \to \mathbb{W}_p$ used in the proof of
Proposition \ref{approxProp}. While $L_p$ is intrinsic Lipschitz,
the map $\Psi_p$ is metrically Lipschitz and it is obtained by
precomposing $L_p$ with a map that serves as ``characteristic
straightening map'' in this setting.
\end{remark}
\begin{remark}
The proof of Theorem \ref{mainGraphs} can be modified so that it
yields $2$-bilipschitz maps instead of $2L'$-bilipschitz maps for
a constant $L'=L'(L)>1$. In the case $n=1$ this is due to the
third author in an earlier version of this paper, and it is based
on replacing the bilipschitz map $\Psi_p:\mathbb{W}\to\mathbb{W}_p$ in
\eqref{eq:Psi_p} by the isometry $(0,y,t)\mapsto (y e_p, t)$,
where $e_p$ represents a horizontal unit vector vector (in the
$\{X_1,X_2\}$-frame) perpendicular to the horizontal normal
$\nu_{\mathbb{H}}(p)$.
\end{remark}
Returning to the proof of Theorem \ref{mainGraphs}, we proceed to
construct the mappings $i_{w\to p}:G\to \Phi(\mathbb{W})$ for $w\in G$,
$p\in S=\Phi(\mathbb{W})$. Since $G=\mathbb{H}^{n-1}\times \mathbb{R}$ (as in
\eqref{eq:HeisxR}) is isometrically isomorphic to $\mathbb{W}$ via the map
\begin{displaymath}
F:((z_1,\ldots,z_{2n-2},t),s)\mapsto
(0,z_1,\ldots,z_{n-1},s,z_n,\ldots,z_{2n-2},t)
\end{displaymath}
the idea for the construction of $i_{w\to p}(v)$ is informally
the following: identify $w^{-1}\cdot_G v \in G$ with the point
$F(w^{-1}\cdot_G v)\in \mathbb{W}$, then map this point to the vertical
plane $\mathbb{W}_p$ by means of the bilipschitz map $\Psi_p$ from
\eqref{eq:Psi_p}, left translate by the point $p$, and finally let
$i_{w\to p}(v)$ be a point in $S$ of minimal distance from
$p\cdot \Psi_p(F(w^{-1}\cdot_G v))\in p\cdot \mathbb{W}_p$, keeping in
mind Proposition \ref{approxProp}. Such a point may not be unique,
but this does not matter as long as the choice is made depending
only on the product $w^{-1}\cdot_G v$, and not on the points $v$
and $w$ individually.
We now explain the construction in detail. First, if $u\in G$, let
$$q := q[p,u] \in S$$ be any point satisfying
\begin{equation}\label{form31} d_{\mathbb{H}}\left(p \cdot \Psi_p(F(u)),q\right)
= \operatorname{dist}_{\mathbb{H}}\left(p \cdot \Psi_p(F(u)),S\right). \end{equation}
Then, if $v,w \in {G}$ and $p \in S$, let
\begin{equation}\label{eq:i_1}i_{w \to p}(v) :=
q[p,w^{-1}\cdot_G v].\end{equation} The definition implies that if
$w,w',v,v' \in {G}$ with $w^{-1}\cdot_G v = (w')^{-1}\cdot_G v'$,
then
\begin{equation}\label{form30} i_{w \to p}(v) = i_{w' \to p}(v'). \end{equation}
To simplify notation in the sequel, we define $\mathrm{Tan}^{w}_{p} \colon
({G},d_{{G}}) \to (\mathbb{W}_{p},d_{\mathbb{H}})$ to be the map given by
\begin{equation}\label{eq:unitary_Tan}
\mathrm{Tan}^w_p(v) = \Psi_p (F(w^{-1}\cdot_G v)),\quad v\in G.
\end{equation}
It follows from Lemma \ref{l:Psi_A} that $\mathrm{Tan}^{w}_{p} \colon
({G},d_{{G}}) \to (\mathbb{W}_{p},d_{\mathbb{H}})$ is a bilipschitz map
with bilipschitz constant bounded in terms of
$\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})}$. Evidently $i_{w
\to p}(w) = p \cdot \mathrm{Tan}_{p}^{w}(w) = p$. Also note that the
isomorphism property of $\Psi_p\circ F$ implies the following
"chain rule":
\begin{equation}\label{chain} \mathrm{Tan}^{w_{1}}_{p}(w_{3}) = \mathrm{Tan}^{w_{1}}_{p}(w_{2}) \cdot \mathrm{Tan}^{w_{2}}_{p}(w_{3}), \qquad w_{1},w_{2},w_{3} \in
G, \: p \in S. \end{equation} Since $p \cdot \mathrm{Tan}_{p}^w(v) \in p
\cdot \mathbb{W}_{p}$, one infers from \eqref{form20} and the definition
of $i_{w \to p}(v)$ that
\begin{align} d_{\mathbb{H}}(p \cdot \mathrm{Tan}^{w}_{p}(v),i_{w \to p}(v)) & = \operatorname{dist}_{\mathbb{H}}(p \cdot \mathrm{Tan}^{w}_{p}(v),S) \notag\\
&\label{form18} \leq Ad_{\mathbb{H}}(p \cdot \mathrm{Tan}_{p}^{w}(v),p)^{1 +
\alpha} \lesssim_L Ad_{G}(w,v)^{1 + \alpha}. \end{align} Using
this estimate,
one has
\begin{align} |d_{\mathbb{H}} & (i_{w \to p}(v),i_{w \to p}(v')) - d_{\mathbb{H}}(p\cdot\mathrm{Tan}^{w}_{p}(v),p\cdot\mathrm{Tan}^{w}_{p}(v'))| \notag\\
& \leq d_{\mathbb{H}}(p \cdot \mathrm{Tan}^{w}_{p}(v),i_{w \to p}(v)) + d_{\mathbb{H}}(p \cdot \mathrm{Tan}^{w}_{p}(v'),i_{w \to p}(v')) \notag\\
&\label{form25} \lesssim_L A\max\{d_{G}(w,v)^{1 +
\alpha},d_{G}(w,v')^{1 + \alpha}\}, \qquad v,v' \in G.
\end{align}
Moreover
\begin{displaymath}
d_{\mathbb{H}}(p\cdot\mathrm{Tan}^{w}_{p}(v),p\cdot\mathrm{Tan}^{w}_{p}(v')) =
d_{\mathbb{H}}(\Psi_p(F(w^{-1}\cdot_G v)),\Psi_p(F(w^{-1}\cdot_G v'))),
\end{displaymath}
and since $\Psi_p \circ F$ is bilipschitz with a constant that
depends only on $\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})}$,
condition \eqref{ISO} follows with the help of \eqref{form25}.
It remains to check condition \eqref{comp}. Using the homogeneity
of $G$ (see the discussion around \eqref{form27} for further
details), it suffices to verify the case "$x = 0$" of condition
\eqref{comp}: if $w \in G$ with $\|w\| \leq 2$, $p \in S$, and
$i_{0 \to p}(w) = q \in S$, then
\begin{equation}\label{form24} d_{\mathbb{H}}(i_{0 \to p}(v),i_{w \to q}(v)) \lesssim_{L} AH \max\{\|w\|^{1 + \alpha/2},\|v\|^{1 + \alpha/2},
d_{G}(v,w)^{1 + \alpha/2}\} \end{equation} for all $v \in G$ with
$\|v\|\leq 1$. (In particular, it may be interesting to note that the
$C^{1,\alpha}_{\mathbb{H}}$-hypothesis only gives the condition
\eqref{comp} with exponent $\alpha/2$.) To estimate the left hand
side of \eqref{form24}, the strategy will be to first obtain a
corresponding estimate for
\begin{displaymath} d_{\mathbb{H}}(p \cdot \mathrm{Tan}^{0}_{p}(v), q \cdot \mathrm{Tan}^{w}_{q}(v)), \end{displaymath}
and eventually conclude \eqref{form24} from this bound combined
with \eqref{form18}. Consider $v,w \in G$ with $\|w\| \leq 2$.
Start by applying the "chain rule" \eqref{chain}, and the triangle
inequality, as follows:
\begin{align*} d_{\mathbb{H}} (p \cdot \mathrm{Tan}^{0}_{p}(v), q \cdot \mathrm{Tan}^{w}_{q}(v))
& = d_{\mathbb{H}}([p \cdot \mathrm{Tan}^{0}_{p}(w)] \cdot \mathrm{Tan}^{w}_{p}(v), q \cdot \mathrm{Tan}^{w}_{q}(v))\\
& \leq d_{\mathbb{H}}([p \cdot \mathrm{Tan}^{0}_{p}(w)] \cdot \mathrm{Tan}^{w}_{p}(v), [p \cdot \mathrm{Tan}^{0}_{p}(w)] \cdot \mathrm{Tan}^{w}_{q}(v))\\
& \quad + d_{\mathbb{H}}([p \cdot \mathrm{Tan}^{0}_{p}(w)] \cdot
\mathrm{Tan}^{w}_{q}(v),q \cdot \mathrm{Tan}^{w}_{q}(v)) =: I_{1} + I_{2}.
\end{align*} To estimate $I_{1}$, note that by left-invariance
\begin{displaymath} I_{1} = d_{\mathbb{H}}(\mathrm{Tan}^{w}_{p}(v),\mathrm{Tan}^{w}_{q}(v)) =
d_{\mathbb{H}}(\Psi_p(F(w^{-1}\cdot_G v)),\Psi_q(F(w^{-1}\cdot_G
v))).
\end{displaymath}
Then \eqref{eq:Psi_p} and the third part of Lemma \ref{l:Psi_A}
imply that
\begin{displaymath}
I_1=d_{\mathbb{H}}(\Psi_p(F(w^{-1}\cdot_G
v)),\Psi_q(F(w^{-1}\cdot_G v)))\lesssim_L
|\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))-\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(q))|^{1/2}
d_{G}(v,w).
\end{displaymath}
To proceed, we note that $p= a \cdot \varphi(a)$ and $q=b\cdot
\varphi(b)$ satisfy
\begin{align*}
|\nabla^{\varphi}\varphi(b)-\nabla^{\varphi}\varphi(a)|&=|\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p\cdot
\varphi(a)^{-1}\cdot a^{-1}
\cdot b \cdot\varphi(a)))-\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))|\\
&=|\nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(\varphi(a)^{-1}\cdot a^{-1}\cdot b \cdot\varphi(a))) - \nabla^{\varphi^{(p^{-1})}}\varphi^{(p^{-1})}(0)|\\
&\overset{\eqref{CFOdef}}{\leq} H \|\varphi(a)^{-1}\cdot a^{-1}\cdot b \cdot\varphi(a)\|^{\alpha}\\
&= H \|\varphi(a)^{-1}\cdot a^{-1}\cdot b \cdot\varphi(b) \cdot\varphi(b)^{-1} \cdot\varphi(a)\|^{\alpha}\\
&\leq 2 H d_{\mathbb{H}}(p,q)^{\alpha},
\end{align*}
using Lemma \ref{invintrgrad} when passing to the second line. It
follows for $p\in S$ and $q=i_{0\to p}(w)$ with $\|w\|\leq 2$ that
\begin{align}\label{form39} |\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))-\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(q))| \leq
2 H d_{\mathbb{H}}(p,q)^{\alpha} & = Hd_{\mathbb{H}}(i_{0 \to p}(0),i_{0 \to p}(w))^{\alpha}\\
& \stackrel{\eqref{form25}}{\lesssim}_{L} H[\|w\|^{\alpha} +
A\|w\|^{\alpha}] \lesssim_L AH\|w\|^{\alpha}. \notag\end{align}
One needed here the assumption $\|w\| \leq 2$ since
\eqref{form25} initially gives a term of the form $\|w\|^{1 +
\alpha}$. The estimate above implies that
\begin{displaymath}
I_1 \lesssim_L
|\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(p))-\nabla^{\varphi}\varphi(\pi_{\mathbb{W}}(q))|^{1/2}
d_{G}(v,w) \lesssim_{L} A H\|w\|^{\alpha/2} d_G(v,w).
\end{displaymath}
Thus, the term $I_{1}$ is bounded from above by the right hand
side of \eqref{form24}.
The term $I_{2}$ has the form $I_{2} = \|\mathfrak{b}^{-1}
\cdot\mathfrak{ a} \cdot \mathfrak{b}\|$ with
\begin{displaymath} \mathfrak{a} = q^{-1} \cdot p \cdot \mathrm{Tan}^{0}_{p}(w) \quad \text{and} \quad \mathfrak{b} = \mathrm{Tan}^{w}_{q}(v). \end{displaymath}
Note that $\|\mathfrak{a}\| = d_{\mathbb{H}}(q,p \cdot \mathrm{Tan}^{0}_{p}(w))
\lesssim_L A\|w\|^{1 + \alpha}$ by \eqref{form18} (this is the
place where the relation $q = i_{0 \to p}(w)$ is used), whereas
$\|\mathfrak{b}\| \sim_L d_{G}(w,v)$. Now, writing $\mathfrak{a} =
(x_{a},t_{a})$ and $\mathfrak{b} = (x_{b},t_{b})$, one can easily
compute that
\begin{equation}\label{eq:FundCommRel} \mathfrak{b}^{-1} \cdot \mathfrak{a} \cdot \mathfrak{b}
=\mathfrak{ a} \cdot \left(0,0,\sum_{i=1}^n \left( x_{a,i}x_{b,n+i}-x_{b,i}x_{a,n+i} \right) \right)=:\mathfrak{a}\cdot (0,0,\omega(x_a,x_b)).
\end{equation}
(This is just the fundamental "commutator relation" in $\mathbb{H}^n$.)
Consequently,
\begin{align*} I_{2} & \leq \|\mathfrak{a}\| + \sqrt{|{\omega(x_a,x_b)}|} \lesssim_L A\|w\|^{1 + \alpha} + \sqrt{\|\mathfrak{a}\|\|\mathfrak{b}\|}\\
& \lesssim_L A\|w\|^{1 + \alpha} + A^{1/2}\|w\|^{1/2 + \alpha/2}d_{G}(w,v)^{1/2}\\
& \lesssim_L A\max\{\|w\|^{1 + \alpha/2},d_{G}(w,v)^{1 +
\alpha/2}\}.
\end{align*} This shows that also $I_{2}$ is bounded by the right
hand side of \eqref{form24}. Glancing again at the estimates for
$I_{1}$ and $I_{2}$, one sees that
\begin{equation}\label{form26} d_{\mathbb{H}}(p \cdot \mathrm{Tan}^{0}_{p}(v), q \cdot \mathrm{Tan}^{w}_{q}(v))
\lesssim_{L} AH\max\{\|w\|^{1 + \alpha/2},d_{G}(v,w)^{1 +
\alpha/2}\}, \end{equation} which is even a bit better than
\eqref{form24}. The estimate \eqref{form24} now follows from the
triangle inequality:
\begin{align*} d_{\mathbb{H}}(i_{0 \to p}(v),i_{w \to q}(v)) & \leq d_{\mathbb{H}}(i_{0 \to p}(v),p \cdot \mathrm{Tan}^{0}_{p}(v))\\
&\quad + d_{\mathbb{H}}(p \cdot \mathrm{Tan}^{0}_{p}(v), q \cdot \mathrm{Tan}^{w}_{q}(v))\\
&\quad + d_{\mathbb{H}}(i_{w \to q}(v),q \cdot \mathrm{Tan}^{w}_{q}(v)).
\end{align*} The middle term here is controlled by \eqref{form26},
and the first and third terms are controlled by \eqref{form18},
recalling the bounds for $\|v\|\leq 1$ and $\|w\|\leq 2$, which
ensure that we can replace "$\alpha$" by "$\alpha/2$" in
\eqref{form18}. This concludes the proof of \eqref{form24}.
Finally, we address the point left open above, that \eqref{form24}
looks slightly less general than \eqref{comp}. To check
\eqref{comp} properly, we need to fix $w_{1},w_{2} \in B_{G}(0,1)$
and $p,q \in S$ with $i_{w_{1} \to p}(w_{2}) = q$, and verify that
\begin{equation}\label{form27} d_{\mathbb{H}}(i_{w_{1} \to p}(w_{3}),i_{w_{2} \to q}(w_{3})) \lesssim
\max\{d_{G}(w_{1},w_{2}),d_{G}(w_{1},w_{3})),d_{G}(w_{2},w_{3})\}^{1
+ \alpha/2} \end{equation} for all $w_{3} \in G$ with
$d_G(w_1,w_3)\leq 1$. However, set $w := w_1^{-1}\cdot_G w_{2}$
and $v := w_1^{-1}\cdot_G w_{3}$, and observe that
\begin{displaymath}w_{2}^{-1}\cdot_G w_3 = w^{-1}\cdot_G v\quad \text{and} \quad w_{1}^{-1}\cdot_G w_3 =0^{-1}\cdot_G v. \end{displaymath}
These relations, and \eqref{form30}, show that
\begin{displaymath} i_{w_{1} \to p}(w_{3}) = i_{0 \to p}(v) \quad \text{and} \quad i_{w_{2} \to q}(w_{3}) = i_{w \to q}(v). \end{displaymath}
Thus, \eqref{form27} follows from \eqref{form24} applied with $w$
and $v$, as above.
This completes the verification of the hypotheses of Theorem
\ref{main}, and hence the proof of Theorem \ref{mainGraphs}.
\section{Lipschitz flags and extra vertical H\"older
regularity}\label{s:LipFlag}
This section is devoted to the first Heisenberg group, $\mathbb{H}^1$.
For convenience we use coordinates $(x,y,t)$, instead of
$(x_1,x_2,t)$,
to denote points in $\mathbb{H}^1$. As
usual, we may identify $\mathbb{W}$ with $\mathbb{R}^2$ by mapping
$(0,y,t)$ to $(y,t)$, and we identify $(x,0,0)\in\mathbb{V}$ with
$x\in \mathbb{R}$.
\begin{definition}\label{d:ExtraVerticalHolder}
We say that an intrinsic Lipschitz function $\varphi:\mathbb{W} \to \mathbb{V}$
has \emph{extra vertical H\"older regularity with constants
$0<\alpha\leq 1$ and $H>0$} if
\begin{equation}\label{eq:VertHolDef}
|\varphi(y,t)-\varphi(y,t')|\leq H |t-t'|^{\frac{1+\alpha}{2}},
\end{equation}
for all $y,t,t'\in \mathbb{R}$.
\end{definition}
Intrinsic Lipschitz functions are always $1/2$-H\"older continuous
with respect to the Euclidean metric along vertical lines.
Condition \eqref{eq:VertHolDef} constitutes an amount of extra
regularity at small scales which is not implied by the intrinsic
Lipschitz condition alone, see for instance \cite[Example
1.3]{MR3400438}.
\begin{remark}\label{r:VertHolDilateTranslate}
The definition is left invariant: if $\varphi$ has extra vertical
H\"older regularity with constants $\alpha$ and $H$, then for
every $p\in \mathbb{H}^1$, the function $\varphi^{(p^{-1})}$ whose
intrinsic graph is $p^{-1}\cdot \Phi(\mathbb{W})$ also has extra vertical
H\"older regularity with the same constants. Moreover, for $r>0$,
the function $\delta_\frac{1}{r} \circ \varphi \circ \delta_{r}$,
whose intrinsic graph is $\delta_{\frac{1}{r}}(\Phi(\mathbb{W}))$, has
extra vertical H\"older regularity with constants $\alpha$ and
$Hr^{\alpha}$. So condition \eqref{eq:VertHolDef} improves under
``zooming in''.
\end{remark}
\begin{remark}\label{r:small_scale} An intrinsic Lipschitz
function \textbf{with compact support} has extra vertical H\"older
regularity with constants $0<\alpha\leq 1$ and $H>0$ if and only
if there is $H'>0$ such that
\begin{equation}\label{eq:all_Scale}
|\varphi(y,t)-\varphi(y,t')|\leq \left\{\begin{array}{ll}H'
|t-t'|^{\frac{1+\alpha}{2}},&\text{if }|t-t'|\leq 1,\\H'
|t-t'|^{\frac{1-\alpha}{2}},&\text{if }|t-t'|\geq
1,\end{array}\right.,\quad y,t,t'\in \mathbb{R},
\end{equation}
that is, $\varphi$ has extra vertical H\"older regularity in the
sense of \cite[Theorem 5.1]{fssler2018riesz}.
\end{remark}
Before studying in more detail the intrinsic graphs of functions
that satisfy the conditions in Definition
\ref{d:ExtraVerticalHolder}, we give two examples of such functions.
\begin{ex}\label{ex:EuclC1} Under the identification $\mathbb{W}\triangleq \mathbb{R}^2$ and $\mathbb{V}\triangleq \mathbb{R}$ described
before Definition \ref{d:ExtraVerticalHolder}, every compactly
supported Euclidean Lipschitz function $\varphi:\mathbb{R}^2
\to \mathbb{R}$ is an intrinsic Lipschitz function that satisfies
the extra vertical H\"older regularity condition in Definition
\ref{d:ExtraVerticalHolder} with $\alpha=1$.
\end{ex}
\begin{ex}\label{ex:intrC1alpha}
Let $0<\alpha\leq 1$, $\mathbb{W},\mathbb{V}\subset \mathbb{H}^1$ as above, and let
$\varphi:\mathbb{W} \to \mathbb{V}$ be a compactly supported
$C_{\mathbb{H}}^{1,\alpha}(\mathbb{W})$ function. Since $\operatorname{spt} \varphi$ is
compact, $\nabla^{\varphi}\varphi$ is continuous and compactly
supported, hence $L :=
\|\nabla^{\varphi}\varphi\|_{L^{\infty}(\mathbb{W})} < \infty$. According
to \cite[Lemma 2.22]{CFO2}, this implies that $\varphi$ is
intrinsic Lipschitz. The extra vertical H\"older regularity
condition follows from \cite[Proposition 4.2]{CFO2}, keeping in
mind the characterization of compactly supported
$C_{\mathbb{H}}^{1,\alpha}(\mathbb{W})$ functions stated in Proposition
\ref{equivProp}.
\end{ex}
Intrinsic graphs of intrinsic Lipschitz functions with extra
vertical H\"older regularity turn out to be well approximable at
all points and small scales by intrinsic Lipschitz graphs of a
special form, the \emph{Lipschitz flags}; see Proposition
\ref{p:contenutesSpPsi} for the precise statement. In the proof of
Theorem \ref{mainIntroVertical}, Lipschitz flags will play an
analogous role as vertical tangent planes did in the proof of
Theorem \ref{mainIntro}, so we start with some observations of
general nature that serve as a counterpart for Section
\ref{s:DefPrelim}.
\subsection{Approximation by Lipschitz flags}
\begin{definition}\label{d:FlagSurface}
We say that $F\subset \mathbb{H}^1$ is a \emph{Lipschitz flag} if there
exists a Euclidean Lipschitz map $\psi :\mathbb{R} \to \mathbb{R}$ such that $F$
is the intrinsic graph $\Phi(\mathbb{W})$ of the map $\varphi : \mathbb{W} \to \mathbb{V}$
defined by
\begin{equation}\label{linkvarphipsi}
\varphi (y,t ) = \psi (y).
\end{equation}
\end{definition}
Lipschitz flags are bilipschitz equivalent to the parabolic plane.
This observation appeared already in \cite[Lemma
7.5]{fssler2019singular}, but we include the proof for completeness.
\begin{lemma}\label{flagimpliesbilip} If $F\subset \mathbb{H}^1$ is a
Lipschitz flag given by an $L$-Lipschitz function $\psi :\mathbb{R}\to
\mathbb{R}$, then there is a $\sim (1+L)$-bilipschitz map
\begin{equation*}
\Psi _F: (\mathbb{W}, d_\mathbb{H}) \to (F, d_\mathbb{H}).
\end{equation*}
\end{lemma}
\begin{proof}
Let $F$ be a Lipschitz flag, so $F$ is the intrinsic graph
$\Phi(\mathbb{W})$ of the map $\varphi : \mathbb{W} \to \mathbb{V}$ defined as in
\eqref{linkvarphipsi} for the Euclidean L-Lipschitz function $\psi
:\mathbb{R} \to \mathbb{R}$. We will show that the map $\Psi _F: (\mathbb{W}, d_\mathbb{H}) \to
(F, d_\mathbb{H})$ given by
\begin{equation}\label{eq:psi_F}
\begin{aligned}
\Psi _F (y,t) : = \left(\psi (y), y, t- \frac 1 2 y \psi (y) + \int_0^y \psi (\eta ) \, d \eta \right)\\
\end{aligned}
\end{equation}
is the $\sim (1+L)$-bilipschitz map which we are looking for.
Firstly, we observe that
\begin{equation*}
\Psi _F (y,t) = \Phi \left(y, t + \int_0^y \psi (\eta ) \, d \eta \right),\\
\end{equation*}
where $\Phi$ is the graph map\footnote{The map $\Psi_F$ is the
composition of the graph map $\Phi$ with the ``characteristic
straightening map'' mentioned in the introduction.} of $\varphi$,
hence $\Psi_F(\mathbb{W})=\Phi(\mathbb{W})$. Moreover, since $\psi$ is an
$L$-Lipschitz function, we get
\begin{equation}\label{bilipproof}
\begin{aligned}
d_\mathbb{H} & (\Psi _F (y,t), \Psi _F (y',t') )\\& = \left\| \left(\psi (y')-\psi (y),y'- y, t'- t+ \tfrac{ 1}{ 2} (y-y' )(\psi (y')+\psi (y)) + \int_{y}^{y'} \psi (\eta ) \, d \eta \right) \right\|\\
& = \left\| \left(\psi (y')-\psi (y),y'- y, t'- t + \int_{y}^{y'} \left(\tfrac{\psi (\eta ) -\psi (y)}{2} \right)+ \left(\tfrac{\psi (\eta ) -\psi (y')}{2} \right) \, d \eta \right) \right\|\\
& \lesssim (1+L) \left(|y'-y | + \sqrt{ |t'-t|}\right),
\end{aligned}
\end{equation}
for all $(y,t), (y',t') \in \mathbb{W}$. On the other hand, since
\begin{equation*}
\begin{aligned}
|t'-t| \leq \left| t'- t + \int_{y}^{y'} \left(\tfrac{\psi (\eta )
-\psi (y)}{2} \right)+ \left(\tfrac{\psi (\eta ) -\psi (y')}{2}
\right) \, d \eta \right| + \left| \int_{y}^{y'} \left(\tfrac{\psi
(\eta ) -\psi (y)}{2} \right)+ \left(\tfrac{\psi (\eta ) -\psi
(y')}{2} \right) \, d \eta \right|,
\end{aligned}
\end{equation*}
it follows
\begin{equation*}
\begin{aligned}
d_\mathbb{H} ( (y,t), (y',t')) \lesssim |y-y'| + |t-t'|^{1/2} \lesssim
(1+L) d_{\mathbb{H}}(\Psi _F (y,t), \Psi _F (y',t') ),\end{aligned}
\end{equation*}
for all $(y,t), (y',t') \in \mathbb{W}$. Hence, putting together
\eqref{bilipproof} and the last inequality, we get that $\Psi _F$
is a $\sim(1+L)$-biLipschitz map, as desired.
\end{proof}
Given an intrinsic Lipschitz graph $S\subset \mathbb{H}^1$, we will
define for each $p\in S$ a Lipschitz flag that intersects
$S$ in a Lipschitz curve passing through $p$. We start with some
general definitions that will be used throughout this section. We
state them in terms of the intrinsic Lipschitz function
$\varphi^{(p^{-1})}$, whose intrinsic graph is $p^{-1}\cdot S$,
recall \eqref{form34} .
\begin{definition}\label{d:mappsi_p} Let $S=\Phi(\mathbb{W})=\{ w\cdot \varphi (w)\,:\, w\in \mathbb{W} \}$ be the intrinsic graph of an intrinsic $L$-Lipschitz function $\varphi : \mathbb{W} \to \mathbb{V}$. To each point $p\in S$, we associate the function
\begin{equation}\label{defipsi_P}
\begin{aligned}
\psi _p :\mathbb{R} & \longrightarrow \mathbb{R}\\
s & \longmapsto \varphi ^{(p^{-1})} (s, \tau _p(s)),
\end{aligned}
\end{equation}
where $\tau_p :\mathbb{R} \to \mathbb{R} $ is some solution of the Cauchy problem
\begin{equation}\label{Cauchy problem}
\left\{
\begin{array}{lcl}
\dot{\tau}_p(s)= \varphi ^{(p^{-1})} ( s, \tau _p(s)), \quad \mbox{for } s\in \mathbb{R} \\
\tau _p (0)=0. \\
\end{array}
\right.
\end{equation}
\end{definition}
Note that $\varphi ^{(p^{-1})} $ is intrinsic Lipschitz with the
same constant as $\varphi$, and then the global existence of
$\tau_p $ follows as in \cite[(6.27)]{antonelli2019pauls}.
Proposition 6.10 in \cite{antonelli2019pauls} is only for higher
dimensions, but this part of the argument works also for our
setting $\mathbb{H}^1$. Moreover, using the same proof as for
\cite[(6.30)]{antonelli2019pauls}, it follows that $\tau _p$
satisfies the inequality
\begin{equation}\label{(6.31)new}
|\tau _p (s) | \lesssim _L |s|^2, \qquad s\in \mathbb{R}.
\end{equation}
Notice further that $\gamma_p : s \mapsto (0,s,\tau_p(s))$ is a $C^1$ curve with
\begin{equation*}
\dot \gamma_p (s)= \begin{pmatrix} 0\\ 1\\ \varphi
^{(p^{-1})}(\gamma_p (s))
\end{pmatrix}= D_2^{\varphi^{(p^{-1})}}|_{\gamma_p(s)}, \end{equation*}
where we recall that $D^{\varphi}_{2}$ is the vector field $D^{\varphi}_{2} = \partial_{y} + \varphi \partial_{t}$. As a consequence, from \cite[Proposition 6.6]{antonelli2019pauls} it also follows that
$ s \mapsto \psi_p(s)=\varphi ^{(p^{-1})} (\gamma_p (s))$ is Euclidean Lipschitz with Lipschitz constant depending only
on the intrinsic Lipschitz constant of $\varphi$. The solution $\tau_p$ may not be unique, but
we never need other properties of it than the ones described
above, so any choice will do. For completeness, we also mention
that $\Phi^{(p^{-1})}\circ \gamma_p$ is a Lipschitz curve in
$(\mathbb{H}^1,d_{\mathbb{H}})$ by \cite[Theorem
4.2.16]{kozhevnikov:tel-01178864}.
\begin{definition}\label{d:flagsurfaceF_p}
Let $S= \Phi(\mathbb{W})$ be the intrinsic graph of an intrinsic Lipschitz function $\varphi : \mathbb{W} \to \mathbb{V}$. For each point $p\in S$, we define
\begin{equation}\label{flagsurfaceF_p}
F_p:= \left\{ \left(0,y,t \right) \cdot \left(\psi_p(y) ,0,0
\right)\,:\, (y,t) \in \mathbb{R}^2 \right\}.
\end{equation}
\end{definition}
We note the following properties of $F_p$ defined as in
\eqref{flagsurfaceF_p}:
\begin{enumerate}
\item $F_p$ is a Lipschitz flag, \item $0\in F_p$, \item $p\cdot
F_p$ is also a Lipschitz flag, \item $F_p \cap (p^{-1}\cdot
S)\supseteq \Phi^{(p^{-1})}( \gamma_p(\mathbb{R}))$.
\end{enumerate}
Item (1) follows immediately from the fact that
$\psi_p$ is a Euclidean Lipschitz function. The item (2) follows
since $p\in S$, and so $\psi_p(0)=\varphi ^{(p^{-1})}(0,0)=0$.
Next, (3) follows by computing explicitly that
\begin{displaymath} p \cdot F_{p} = \{(0,y,t) \cdot (\psi(y),0,0) : (y,t) \in \mathbb{R}^{2}\}, \end{displaymath}
where $\psi(y) := \psi_{p}(y - y_{p}) + x_{p}$ and $p = (x_{p},y_{p},t_{p})$. Since $\psi_p:\mathbb{R}\to\mathbb{R}$ is Lipschitz, so is $\psi$. Finally, (4) is clear from the
definitions since $\psi_p(y)=\varphi^{(p^{-1})}(y,\tau_p(y))$ and
$F_p$ is of the form \eqref{flagsurfaceF_p}.
\begin{definition}\label{d:Psi_p}
Let $S= \Phi(\mathbb{W}) $ be the intrinsic graph of an intrinsic
Lipschitz function $\varphi : \mathbb{W} \to \mathbb{V}$. For each point $p\in S$,
we let $\Psi_p:=\Psi_{F_p}:\mathbb{W} \to F_p$ be the map given by the
formula \eqref{eq:psi_F} applied to the Lipschitz flag $F=F_p$
from \eqref{flagsurfaceF_p}, that is,
\begin{equation}\label{defiPsi_p}
\begin{aligned}
\Psi_p(y,t)= \left(\varphi ^{(p^{-1})} ( y, \tau _p(y)) , y, t-
\frac 1 2 y \varphi ^{(p^{-1})} ( y, \tau _p(y)) + \int_0^y
\varphi ^{(p^{-1})} ( \eta , \tau _p(\eta )) \, d \eta \right).
\end{aligned}
\end{equation}
\end{definition}
The reader may think of $\Psi_p$ as a surrogate for the map
that was used in \eqref{eq:unitary_Tan} to
approximate an intrinsic $C^{1,\alpha}$ graph by the vertical
plane $p\cdot \mathbb{W}_p$, and the Lipschitz flag $F_p$ plays the role
of $\mathbb{W}_p$. The analogy is however not perfect: if
$\varphi^{(p^{-1})}$ is not intrinsic linear, the map $\Psi_p$
from Definition \ref{d:Psi_p} is not a group homomorphism and
hence we lack a counterpart for the chain rule \eqref{chain}. The
following properties of $\Psi_p$ defined as in \eqref{defiPsi_p}
will be used:
\begin{enumerate}
\item $\Psi _p(0)=0$ since $\varphi ^{(p^{-1})} (0,0)=0$, \item
$\Psi_p (y,t)= \Phi ^{(p^{-1})} (y,\tau_p(y)) \cdot (0,0,t)$, as
$\int_0^y \varphi ^{(p^{-1})} ( \eta , \tau _p(\eta )) \, d \eta=
\int_0^y \dot{\tau}_p(\eta)\,d\eta = \tau_p(y)$, \item $\Psi_p$
is bilipschitz with a constant that depends only on the intrinsic
Lipschitz constant of $\varphi$ (by Lemma \ref{flagimpliesbilip}
and the paragraph before Definition \ref{d:Psi_p}).
\end{enumerate}
\begin{remark}\label{r:WhatItMeansForFlags}
If $S= \Phi(\mathbb{W}) $ is itself a Lipschitz flag, that is, the
intrinsic graph of an intrinsic Lipschitz
function $\varphi : \mathbb{W} \to \mathbb{V}$ that does not depend on the
$t$-variable, then
\begin{equation}\label{eq:FlagParam}
\Psi_p(y,t)= \Phi^{(p^{-1})}\left(y,t+ \int_0^y
\varphi^{(p^{-1})}(\eta,\tau_p(\eta))\,d\eta\right),
\end{equation}
and in particular, $\Psi_p(\mathbb{W})= \Phi^{(p^{-1})}(\mathbb{W})$ and hence
$S= p \cdot \Psi_p(\mathbb{W})$ in that case. Also note that here
$\varphi^{(p^{-1})}$ does not depend on the $t$-variable, and so
the integral in \eqref{eq:FlagParam} can be written without the
dependence on $\tau_p$.
\end{remark}
As we noted below Definition \ref{d:flagsurfaceF_p}, the surfaces
$F_p=\Psi_p(\mathbb{W})$ and $p^{-1}\cdot S = \Phi^{(p^{-1})}(\mathbb{W})$
intersect at least along a curve. The next lemma shows that they
approximate each other well also in a neighborhood of that curve
if $\varphi$ has extra vertical H\"older regularity.
\begin{lemma}\label{betterProp3.36}
Let $\varphi : \mathbb{W} \to \mathbb{V} $ be an intrinsic Lipschitz function with
extra vertical H\"older regularity with constants $0<\alpha\leq 1$
and $H>0$. Then
\begin{displaymath}
d_\mathbb{H} ( \Psi_p (y,t), \Phi ^{(p^{-1})} (y,\tau_p(y) +t ) ) \leq H
|t|^{(1+\alpha)/2} \leq H\|(y,t)\|^{1 + \alpha}, \qquad \, (y,t) \in \mathbb{R}^2.
\end{displaymath}
\end{lemma}
\begin{remark}\label{r:largeScale}From the proof of Lemma \ref{betterProp3.36} one can
infer that if $\varphi$ has extra vertical H\"older regularity in the stronger sense of \eqref{eq:all_Scale}, then
\begin{equation*}
d_\mathbb{H} ( \Psi_p (y,t), \Phi ^{(p^{-1})} (y,\tau_p(y) +t ) ) \leq
H'\min\{ |t|^{(1+\alpha)/2}, |t|^{(1-\alpha)/2}\}, \quad \mbox{
for }\, (y,t) \in \mathbb{R}^2.
\end{equation*}
\end{remark}
\begin{proof}[Proof of Lemma \ref{betterProp3.36}]
Recall that $\tau_p$ is a solution of the Cauchy problem
\eqref{Cauchy problem}, and we have that
\begin{equation*}
\begin{aligned}
\Psi_p (y,t) =\Phi ^{(p^{-1})} (y,\tau_p(y)) \cdot (0,0,t)
\end{aligned}
\end{equation*}
for all $(y,t) \in \mathbb{R}^2$. As a consequence,
\begin{equation*}
\begin{aligned}
& d_\mathbb{H} ( \Psi_p (y,t), \Phi ^{(p^{-1})} (y,\tau_p(y) +t ) ) = d_\mathbb{H} ( \Phi ^{(p^{-1})} (y,\tau_p(y)) \cdot (0,0,t) , \Phi ^{(p^{-1})} (y,\tau_p(y) +t ) ) \\
& = \|\varphi ^{(p^{-1})} ( y, \tau _p(y) +t )^{-1}\cdot (0, y, \tau _p(y) +t )^{-1} \cdot (0, y, \tau _p(y)) \cdot \varphi ^{(p^{-1})} ( y, \tau _p(y)) \cdot (0,0,t) \|\\
& = | \varphi ^{(p^{-1})} ( y, \tau _p(y) + t ) - \varphi
^{(p^{-1})} ( y, \tau _p(y))| \leq H |t|^{\frac{1+\alpha}{2}},
\end{aligned}
\end{equation*}
as claimed.
\end{proof}
In the following we denote by $[A]_{\delta}^{\mathbb{H}^1}$ the
$\delta$-neighbourhood of a set $A\subset \mathbb{H}^1$ with
respect to $d_{\mathbb{H}}$, and $[B]^{\mathbb{R}^2}_{\delta}$ stands for
the Euclidean $\delta$-neighbourhood of a set $B \subset
\mathbb{R}^2$. A direct corollary of Lemma \ref{betterProp3.36} is
the following result:
\begin{proposition}\label{p:contenutesSpPsi}
Let $\varphi : \mathbb{W} \to \mathbb{V} $ be an intrinsic $L$-Lipschitz function with extra vertical H\"older regularity with constants $0<\alpha\leq 1$
and $H>0$.
Then, there is $c=c(L)\geq 1$ such that for all $r>0$ and all
$p\in S=\Phi(\mathbb{W})$ it follows
\begin{enumerate}
\item $S \cap B(p,r) \subset [p\cdot \Psi
_p(\mathbb{W})]_{c\delta}^{\mathbb{H}^1}$, \item $\left(p\cdot \Psi _p(\mathbb{W})
\right)\cap B(p,r) \subset [S]_{c\delta}^{\mathbb{H}^1}$,
\end{enumerate}
where $\delta := \delta(H,r) := Hr^{1+\alpha}$.
\end{proposition}
\begin{remark}
Continuing Remark \ref{r:largeScale}, we note that if an intrinsic
$L$-Lipschitz function $\varphi$ has extra vertical H\"older
regularity with constants $0<\alpha\leq 1$
and $H'>0$ in the stronger sense \eqref{eq:all_Scale}, then the conclusion of
Proposition \eqref{p:contenutesSpPsi} can be improved by replacing
"$\delta$" with $H'\min\{r^{1+\alpha},r^{1-\alpha}\}$. In other
words, the intrinsic graph of $\varphi$ is well approximated by
Lipschitz flags also at large scales.
\end{remark}
\begin{proof}[Proof of Proposition \ref{p:contenutesSpPsi}]
Fix $r>0$ and $p\in S$ as in the assumptions of the proposition.
Since the metric $d_\mathbb{H}$ is left invariant, it is sufficient to
show that there is $c=c(L)\geq 1$ such that
\begin{enumerate}
\item[(i)] $\left( p^{-1} \cdot S\right) \cap B(0,r) \subset [
\Psi _p(\mathbb{W})]_{c\delta }^{\mathbb{H}^1}$, \item[(ii)] $\Psi _p(\mathbb{W}) \cap
B(0,r) \subset [p^{-1} \cdot S]_{c\delta }^{\mathbb{H}^1}$.
\end{enumerate}
We consider (i). Let $q\in (p^{-1} \cdot S) \cap B(0,r)$. We will
prove that $q \in [ \Psi _p(\mathbb{W})]_{c\delta }^{\mathbb{H}^1}$ for a
constant $c$ depending only on $L$. Firstly, since $q\in \Phi
^{(p^{-1})} (\mathbb{W}) \cap B(0,r)$, we have that $q= \Phi ^{(p^{-1})}
(0,y, \tau _p(y) + t )$ for some $(y,t) \in \mathbb{R}^2$ and $\| q\| = \|
\Phi ^{(p^{-1})} (y, \tau _p(y) + t ) \| \leq r$. More precisely,
by the definition of $\Phi ^{(p^{-1})}$, we find
\begin{equation*}
\begin{aligned}
|\varphi ^{(p^{-1})} ( y, \tau _p(y) +t )| & \leq r,\quad |y| \leq
r,\quad\text{and}\quad
\left|\tau_p(y)+t - \tfrac{ 1}{ 2 }y \varphi ^{(p^{-1})} ( y, \tau _p(y) +t ) \right| \leq \tfrac{r^2}{4},\\
\end{aligned}
\end{equation*}
\begin{equation}\label{stimatsquare}
\begin{aligned}
|t|\leq \left|\tau_p(y)+t - \tfrac {1 }{2} y \varphi ^{(p^{-1})}
( y, \tau _p(y) +t ) \right|+|\tau_p(y)|
+ \left| \tfrac {1}{ 2} y \varphi ^{(p^{-1})} ( y, \tau _p(y) +t )\right| & \stackrel{\eqref{(6.31)new}}{\leq}C_Lr^2+ \tfrac{3r^2}{4}.\\
\end{aligned}
\end{equation}
Now applying Lemma \ref{betterProp3.36} to the point $(y,t)$, we
obtain
\begin{equation*}
\begin{aligned}
d_\mathbb{H} ( \Psi_p (y,t), q ) & = d_\mathbb{H} ( \Psi_p (y,t), \Phi
^{(p^{-1})} (y,\tau_p(y) +t ) )\leq H|t|^{(1+\alpha)/2}
\stackrel{\eqref{stimatsquare}}{\lesssim}_L \delta,
\end{aligned}
\end{equation*}
so $q \in [ \Psi _p(\mathbb{W})]_{c\delta }^{\mathbb{H}^1}$, as desired.
Next we consider (ii). Let $q\in \Psi _p(\mathbb{W}) \cap B(0,r)$. We
want to prove that $q \in [p^{-1} \cdot S]_{c \delta }^{\mathbb{H}^1}$
for a constant $c$ that depends only on $L$. Since $q= \Psi_p (y,
t ) = \Phi ^{(p^{-1})} (y, \tau _p(y) ) \cdot (0,0,t)$ for some
$(y,t) \in \mathbb{R}^2$ and $\|q\| \leq r$, we have that
\begin{equation*}
\begin{aligned}
|\varphi ^{(p^{-1})} ( y, \tau _p(y) )| \leq r,\quad |y| \leq
r,\quad\text{and}\quad
\left|t+ \tau _p(y) - \tfrac{1}{ 2} y \varphi ^{(p^{-1})} ( y, \tau _p(y) ) \right| \leq \tfrac{r^2}{4}.
\end{aligned}
\end{equation*}
and so
\begin{equation}\label{stimatsquare2}
\begin{aligned}
|t| \leq \left|t +\tau _p(y) - \tfrac{y}{ 2} \varphi ^{(p^{-1})}
( y, \tau _p(y) ) \right| + |\tau _p(y)| + \left| \tfrac{y}{ 2}
\varphi ^{(p^{-1})} ( y, \tau _p(y) ) \right|
\stackrel{\eqref{(6.31)new}}{\leq}
\tfrac{3r^2}{4} + C_L r^2.\\
\end{aligned}
\end{equation}
Now we apply Lemma \ref{betterProp3.36} to the point $(y,t)$,
hence
\begin{equation*}
\begin{aligned}
d_\mathbb{H} ( q , \Phi ^{(p^{-1})} (y,\tau_p(y) +t )) = d_\mathbb{H} ( \Psi_p
(y,t), \Phi ^{(p^{-1})} (y,\tau_p(y) +t ) )
\leq H|t|^{(1+\alpha)/2} \stackrel{\eqref{stimatsquare2}}{
\lesssim _{L}} \delta,
\end{aligned}
\end{equation*}
so $q \in [p^{-1} \cdot S]_{c \delta }^{\mathbb{H}^1}$ for $c$ depending
only on $L$, as desired. This completes the proof.
\end{proof}
\medskip
Let $\pi \colon \mathbb{H}^{1} \to \mathbb{R}^{2}$ be the projection $\pi(x,y,t)=(x,y)$. Then $\pi$ is $1$-Lipschitz $(\mathbb{H}^{1},d_{\mathbb{H}}) \to (\mathbb{R}^{2},|\cdot|)$, which easily implies the following statement:
\begin{lemma}\label{l:Notes3}
Assume that $A_1,A_2\subset \mathbb{H}^1$, $p\in \mathbb{H}^1$, $r>0$, and
$\delta>0$ are such that \begin{displaymath} A_1 \cap B(p,r)
\subset [A_2]_{\delta}^{\mathbb{H}^1},\end{displaymath} then
\begin{displaymath}
\pi(A_1 \cap B(p,r))\subset [\pi(A_2 \cap
B(p,r+\delta))]^{\mathbb{R}^2}_{\delta}.
\end{displaymath}
\end{lemma}
The lemma will applied with $A_{2} = F$, a Lipschitz flag. Then $\pi(F) \subset \mathbb{R}^{2}$ is a Lipschitz graph, and the lemma says that $\pi(A_1\cap B(p,r))$ is contained in the Euclidean $\delta$-neighbourhood of the Lipschitz graph $\pi(F)$ whenever if $A_1\cap B(p,r) \subset [F]_{\delta}^{\mathbb{H}^{1}}$.
\subsection{Proof for graphs with extra vertical H\"older regularity}
In this section, we prove Theorem \ref{mainIntroVertical} from the
introduction, which we restate here for the reader's convenience.
\begin{thm}\label{mainVertical} Let $S \subset \mathbb{H}^1$
be the intrinsic graph of a globally defined but compactly
supported intrinsic Lipschitz function with extra vertical
regularity. Then $S$ has big pieces of bilipschitz images of the
parabolic plane $(\Pi,d_{\Pi})$. In particular, $S$ is LI
rectifiable.
\end{thm}
The theorem will be proven as an application of Theorem \ref{main}
and a reduction to unit scale analogous to the one after Lemma
\ref{blowUpLemma}. We will verify the hypotheses of Theorem
\ref{main} for $(G,d_G)=(\Pi,d_{\Pi})$ and $(M,d_M)=(S,d_{\mathbb{H}})$,
with $x_0=0\in G$ and an arbitrary point $p_0\in S$. Since the map
$(y,t)\mapsto (0,y,t)$ is an isometric isomorphism between
$(\Pi,+,d_{\Pi})$ and $(\mathbb{W},\cdot,d_{\mathbb{H}})$, it suffices to
construct maps $i_{w\to p}:\mathbb{W} \to S$ with the desired properties.
As in Section \ref{verification}, the maps $i_{w\to p}$ will be independent of the "scale" parameter $k \in \mathbb{N}$, and can be defined for all $p\in S$ and all $w\in
\mathbb{W}$.
\begin{definition}\label{d:iotanew}
Let $S= \Phi(\mathbb{W})$ be the intrinsic graph of an intrinsic Lipschitz
function $\varphi : \mathbb{W} \to \mathbb{V}$. For each point $p\in S$ and $u \in
\mathbb{W}$, let $q := q[p,u] \in S$ be any point satisfying
\begin{equation}\label{form31new} d_{\mathbb{H}}\left(p \cdot \Psi _p( u) , q \right) =
\operatorname{dist}_{\mathbb{H}}\left(p \cdot \Psi _p( u) ,S\right). \end{equation}
Then, define $i_{w\to p} :\mathbb{W} \to S$ as
\begin{equation}\label{eq:i_2new}i_{w \to p}(v) := q[p,w^{-1}\cdot v].\end{equation}
\end{definition}
\begin{remark}\label{r:iForFlag}
Remark \ref{r:WhatItMeansForFlags} implies that if $S=\Phi(\mathbb{W})$ is
itself a Lipschitz flag, then $i_{w\to p}(v)= p \cdot
\Psi_p(w^{-1}\cdot v)$ for all $v,w\in \mathbb{W}$. In general, the extra vertical H\"older regularity allows
to control the distance between $i_{w\to p}(v)$ and $p \cdot
\Psi_p(w^{-1}\cdot v)$.
\end{remark}
If $\varphi$ has extra vertical H\"older regularity with constants
$0<\alpha\leq 1$ and $H>0$, then Lemma \ref{betterProp3.36}
immediately implies that
\begin{align}\label{eq:Distance estimate} d_{\mathbb{H}}(p\cdot\Psi_p(w^{-1}\cdot w'),i_{w \to p}(w')) &= \operatorname{dist}_{\mathbb{H}}\left(\Psi_{p}(w^{-1} \cdot w'),p^{-1} \cdot S\right) \notag\\
& =
\operatorname{dist}_{\mathbb{H}}
\left(\Psi_p\left(w^{-1}\cdot
w'\right),\Phi^{(p^{-1})}(\mathbb{W})\right) \notag\\& \leq Hd_{\mathbb{H}}(w,w')^{1 + \alpha},
\end{align}
for all $p\in S=\Phi(\mathbb{W})$ and $w,w'\in
\mathbb{W}$.
Once again, $i_{w \to p}(v)$ does not depend on the points $v$ and
$w$ individually, but only on the product $w^{-1}\cdot v$, and by
definition $i_{w\to p}(w)=p$. To apply Theorem \ref{main}, we need to verify the two hypotheses \eqref{ISO} and \eqref{comp}. We start by showing that \eqref{ISO} holds for a
constant which depends only on the bilipschitz constant of
$\Psi_p$, which in turn depends only on the intrinsic Lipschitz
constant of $\varphi$, recall the comment below Definition
\ref{d:Psi_p}. Now \eqref{ISO} follows immediately
from \eqref{eq:Distance estimate} and the triangle inequality:
\begin{equation}\label{eq:FirstCondVertical}
|d_{\mathbb{H}}(i_{w\to p}(w'),i_{w\to p}(w''))-
d_{\mathbb{H}}(\Psi_p(w^{-1}\cdot w'),\Psi_p(w^{-1}\cdot w'))|\lesssim H
\max\{d_{\mathbb{H}}(w,w'),d_{\mathbb{H}}(w,w'')\}^{1+\alpha},
\end{equation}
for all $p\in S$ and $w,w',w''\in \mathbb{W}$. We proceed to verify
condition \eqref{comp} in our situation:
\begin{proposition}\label{p:NotesCor}
Let $\varphi:\mathbb{W} \to \mathbb{V}$ be an intrinsic $L$-Lipschitz function
that has extra vertical H\"older regularity with constants
$0<\alpha\leq 1$ and $H>0$. If $w_1,w_2 \in \mathbb{W}$ satisfy
$\|w_1\|,\|w_2\|\leq 1$, and $p,q\in \Phi(\mathbb{W})$ satisfy $i_{w_1\to
p}(w_2)=q$, then
\begin{equation}\label{form29}
d_{\mathbb{H}}\left(i_{w_1 \to p}(w_3),i_{w_2 \to
q}(w_3)\right)\lesssim_{H,L}
\max\left\{d_{\mathbb{H}}(w_1,w_2),d_{\mathbb{H}}(w_1,w_3),d_{\mathbb{H}}(w_2,w_3)
\right\}^{1+\frac{\alpha}{2}}
\end{equation}
for all $w_3\in \mathbb{W}$
with $d_{\mathbb{H}}(w_1,w_3)\leq 1$.
\end{proposition}
\begin{remark}
If $\varphi$ does not depend on the $t$-variable, that is,
$\Phi(\mathbb{W})$ is itself a Lipschitz flag and the extra H\"older
regularity holds with constant $H=0$, then the left hand side of
\eqref{form29} vanishes for all
$w_1,w_2,w_3\in \mathbb{W}$ with $i_{w_1\to p}(w_2)=q$. Indeed, Remark
\ref{r:iForFlag} implies in this case that
\begin{displaymath}
i_{w_1\to p}(w_3)=p\cdot \Psi_p(w_1^{-1}\cdot w_3),\quad i_{w_2\to
q}(w_3)=q\cdot \Psi_q(w_2^{-1}\cdot w_3),
\end{displaymath}
and
\begin{equation}\label{eq:qForm}
q=i_{w_1\to p}(w_2)=p\cdot \Psi_p(w_1^{-1}\cdot w_2).
\end{equation}
Hence, the left hand side of \eqref{form29} can be written as
\begin{displaymath}
d_{\mathbb{H}}(i_{w_1\to p}(w_3),i_{w_2\to q}(w_3))=
d_{\mathbb{H}}(\Psi_p(w_1^{-1}\cdot w_3),\Psi_p(w_1^{-1}\cdot w_2)\cdot
\Psi_q(w_2^{-1}\cdot w_3)).
\end{displaymath}
We recall from Remark \ref{r:WhatItMeansForFlags} that
\begin{equation}\label{form37} \Psi_q(y,t)= \Phi^{(q^{-1})}\left(y,t+ \int_0^y
\varphi^{(q^{-1})}(\eta,\tau_q(\eta))\,d\eta\right). \end{equation}
Let us spell out the formula for $\Phi^{(q^{-1})}$:
\begin{equation}\label{eq:chain_flag}
\Phi^{(q^{-1})}\overset{\eqref{eq:qForm}}{=}
[\Phi^{(p^{-1})}]^{(\Psi_p(w_1^{-1}\cdot w_2)^{-1})}=
\Psi_p(w_1^{-1}\cdot w_2)^{-1}\cdot
\Phi^{(p^{-1})}(\pi_{\mathbb{W}}(\Psi_p(w_1^{-1}\cdot w_2)\cdot [\cdot])),
\end{equation}
where we have applied the formula $\Phi(\pi_{\mathbb{W}}(p \cdot v)) = p \cdot \Phi^{(p^{-1})}(v)$ from Lemma \ref{l:translated} in the last
equality. Using \eqref{form37}-\eqref{eq:chain_flag}, and writing
$w_i=(0,y_i,t_i)$, it is not difficult to show that
\begin{align*}
&\Psi_p(w_1^{-1}\cdot w_2)\cdot \Psi_q(w_2^{-1}\cdot
w_3)\\&=\Psi_p(w_1^{-1}\cdot w_2)\cdot
\Phi^{(q^{-1})}\left(y_3-y_2,t_3-t_2+\int_0^{y_3-y_2}\varphi^{(q^{-1})}(\eta,\tau_q(\eta))\,d\eta\right)
\overset{\eqref{eq:chain_flag}}{=}\Psi_p(w_1^{-1}\cdot w_3).
\end{align*}
We omit some computations, as the remark only serves to motivate Proposition \ref{p:NotesCor}.
\end{remark}
\begin{proof}[Proof of Proposition \ref{p:NotesCor}]
We first apply the triangle inequality:
\begin{align*}
d_{\mathbb{H}}\left(i_{w_1 \to p}(w_3),i_{w_2 \to q}(w_3)\right) & \leq
d_{\mathbb{H}}\left(p \cdot \Psi_p(w_1^{-1}\cdot
w_3),i_{w_1 \to p}(w_3)\right)\\&\quad + d_{\mathbb{H}}(p \cdot \Psi_p(w_1^{-1}\cdot w_3),q \cdot
\Psi_q(w_2^{-1}\cdot w_3)) \\&\quad + d_{\mathbb{H}}\left(q \cdot
\Psi_q(w_2^{-1}\cdot w_3),i_{w_2 \to q}(w_3)\right).
\end{align*}
The estimate \eqref{eq:Distance estimate} shows that the first and third terms are bounded from above by $H d_{\mathbb{H}}(w_1,w_3)^{1+\alpha}$ and $H d_{\mathbb{H}}(w_2,w_3)^{1+\alpha}$, respectively, which is better than claimed. So, the heart of the matter is to prove an upper bound for the second term. This is the content of the next lemma.
\end{proof}
\begin{lemma}\label{l:NotesLem5}
Under the same assumptions as in Proposition \ref{p:NotesCor}, we
have
\begin{displaymath}
d_{\mathbb{H}}(p \cdot \Psi_p(w_1^{-1}\cdot w_3),q \cdot
\Psi_q(w_2^{-1}\cdot w_3)) \lesssim_{H,L}
\max\left\{d_{\mathbb{H}}(w_1,w_2),d_{\mathbb{H}}(w_1,w_3),d_{\mathbb{H}}(w_2,w_3)
\right\}^{1+\frac{\alpha}{2}}.
\end{displaymath}
\end{lemma}
\begin{proof} We fix points
$ p$ and $q=i_{w_1\to p}(w_2)$ as in the assumptions of
Proposition \ref{p:NotesCor}. We may assume with no loss of generality that $w_{1} \neq w_{2}$, since otherwise $q = p$ and the claimed estimate is clear. By left invariance of $d_{\mathbb{H}}$, we have
\begin{equation}\label{eq:TermToBeBounded}
d_{\mathbb{H}}(p \cdot \Psi_p(w_1^{-1}\cdot w_3),q \cdot
\Psi_q(w_2^{-1}\cdot w_3))= d_{\mathbb{H}}(\Psi_p(w_1^{-1}\cdot
w_3),p^{-1}\cdot q \cdot \Psi_q(w_2^{-1}\cdot w_3)).
\end{equation}
Let us denote $p^{-1}\cdot q =: (x,y,t)$. We then define the
curves
\begin{displaymath}
\gamma_p:\mathbb{R} \to \mathbb{R}^2,\quad \gamma_p(s):=
(\varphi^{(p^{-1})}(s,\tau_p(s)),s)
\end{displaymath}
and
\begin{displaymath}
\gamma_q:\mathbb{R} \to \mathbb{R}^2,\quad \gamma_q(s):=
(\varphi^{(q^{-1})}(s,\tau_q(s))+ x,s+y),
\end{displaymath}
whose traces are the $\pi$-projections of the
corresponding Lipschitz flags:
\begin{displaymath}
\gamma_p(\mathbb{R})= \pi\left(\Psi_p(\mathbb{W})\right)\quad \text{and}\quad
\gamma_q(\mathbb{R})= \pi\left(p^{-1}\cdot q \cdot \Psi_q(\mathbb{W})\right)
\end{displaymath}
We observe that the curves $\gamma_{p}$ and $\gamma_{q}$ come close in at least one point. Writing
\begin{equation}\label{eq:w1w2w3}
w_1 = (0,y_1,t_1),\quad w_2=(0,y_2,t_2),\quad w_3=(0,y_3,t_3),
\end{equation}
and recalling that $q = i_{w_{1} \to p}(w_{2})$ by the assumption in the lemma, we have
\begin{align}
\notag|\gamma_p(y_2-y_1)-\gamma_q(0)|&=
|\gamma_p(y_2-y_1)-(x,y)| \leq d_{\mathbb{H}}(\Psi_p(w_1^{-1}\cdot w_2),p^{-1}\cdot q)\\
\notag &= d_{\mathbb{H}} (p\cdot \Psi_p(w_1^{-1}\cdot w_2), i_{w_1\to p}(w_2))\\
&\overset{\eqref{eq:Distance estimate}}{\leq} H
d_{\mathbb{H}}(w_1,w_2)^{1+\alpha}.\label{eq:close_to_meeting}
\end{align}
We next show, \emph{a fortiori}, that the curves $\gamma_p$ and $\gamma_q$ also stay close
to each other for some time. Precisely, we claim that
\begin{equation}\label{eq:curve_closeness}
\operatorname{dist}_{\mathbb{R}^{2}}(\gamma_p(s+y_2-y_1),\gamma_{q}(\mathbb{R})) \lesssim_{H,L}
\max\left\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2) \right\}^{1+\alpha}, \quad s \in \mathbb{R}.
\end{equation}
Note that for $s=0$, the right hand side of
\eqref{eq:curve_closeness} equals $d_{\mathbb{H}}(w_1,w_2)^{1+\alpha}$, as
expected. To prove the claim for arbitrary $s\in \mathbb{R}$, we first
observe that
\begin{displaymath}
\gamma_p(s+y_2-y_1)= \pi(\Psi_p(s+y_2-y_1,0)). \end{displaymath}
Since $\Psi_{p}$ is Lipschitz according to the remark below Definition \ref{d:Psi_p},
\begin{displaymath}
\|\Psi_p(s+y_2-y_1,0)\|\lesssim_L \|(s+y_2-y_1,0)\|,
\end{displaymath}
and we find
\begin{displaymath}
\gamma_p(s+y_2-y_1) \in \pi\left(\Psi_p(\mathbb{W}) \cap B(0,r)\right)
\end{displaymath}
for some $0 < r\lesssim_L \max\left\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2) \right\}$. It follows from
Proposition \ref{p:contenutesSpPsi} and Lemma \ref{l:Notes3} that
\begin{equation}\label{eq:Incl1}
\gamma_p(s+y_2-y_1) \in \left[\pi\left(\left(p^{-1}\cdot \Phi(\mathbb{W})
\right)\cap B(0,r+c\delta)\right)\right]^{\mathbb{R}^2}_{c
\delta},
\end{equation}
where the constant $c$ depends only on $L$, and
$\delta:=Hr^{1+\alpha}$. Since
\begin{displaymath}
d_{\mathbb{H}}(p,q) \leq d_{\mathbb{H}}(p^{-1}\cdot q,\Psi_p(w_1^{-1}\cdot
w_2))+ \|\Psi_p(w_1^{-1}\cdot w_2)\|\overset{\eqref{eq:Distance
estimate}}{\lesssim_{H,L}} d_{\mathbb{H}}(w_1,w_2)^{1+\alpha}+
d_{\mathbb{H}}(w_1,w_2),
\end{displaymath}
and $d_{\mathbb{H}}(w_1,w_2)\leq 1$, we have
\begin{equation}\label{eq:Incl2}
B\left(0,r+c \delta\right)\subset B(p^{-1}\cdot q,R)
\end{equation}
for some $R\geq r$ with $R\lesssim_{H,L}
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}$. By another instance of
Proposition \ref{p:contenutesSpPsi} (applied to the point $q$),
Lemma \ref{l:Notes3}, and left translation by $p^{-1}$, we find
that
\begin{equation}\label{eq:Incl3}
\left[\pi\left(\left(p^{-1}\cdot \Phi(\mathbb{W}) \right)\cap
B(p^{-1}\cdot q,R)\right)\right]^{\mathbb{R}^2}_{c\delta}\subseteq
\left[\pi\left(p^{-1}\cdot q \cdot
\Psi_q(\mathbb{W})\right)\right]^{\mathbb{R}^2}_{2cHR^{1+\alpha}}=
\left[\gamma_q(\mathbb{R})\right]^{\mathbb{R}^2}_{c_{L,H}R^{1+\alpha}}.
\end{equation}
for a constant $0<c_{L,H}<\infty$ that depends only on $L$ and
$H$. Then the claim \eqref{eq:curve_closeness} follows by
combining the inclusions \eqref{eq:Incl1}, \eqref{eq:Incl2}, and
\eqref{eq:Incl3}.
We now fix $s \in \mathbb{R}$, and let $s'\in
\mathbb{R}$ be any point such that $|\gamma_p(s+y_2-y_1) - \gamma_{q}(s')| \lesssim_{H,L} \max\{|s + y_{2} - y_{1}|,d_{\mathbb{H}}(w_{1},w_{2})\}^{1 + \alpha}$. The existence of $s'$ is guaranteed by \eqref{eq:curve_closeness}. We next show that $s'$ cannot be too far from
$s$. Indeed, considering the first component of
$\gamma_p(s+y_2-y_1)-\gamma_q(s')$, we see immediately from \eqref{eq:curve_closeness} that
\begin{equation}\label{eq:p_q_comparison}
|\varphi^{(p^{-1})}(s+y_2-y_1,\tau_p(s+y_2-y_1))-\varphi^{(q^{-1})}(s',\tau_q(s'))-x|\lesssim_{H,L}
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}.
\end{equation}
Considering the second component, we find the estimate
\begin{equation}\label{eq:first_s_s'_estimate}
|s+y_2-y_1-s'-y|\lesssim_{H,L}
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}.
\end{equation}
By the initial estimate \eqref{eq:close_to_meeting}, we know that
\begin{equation}\label{eq:w1_w2_p_q}
|(y_2-y_1)-y|\leq H d_{\mathbb{H}}(w_1,w_2)^{1+\alpha},
\end{equation}
so \eqref{eq:first_s_s'_estimate} yields
\begin{equation}\label{eq:s_s'_comp}
|s-s'|\lesssim_{H,L}
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}.
\end{equation}
This last estimate allows us to deduce a version of
\eqref{eq:p_q_comparison} with "$s'$" replaced by "$s$". Indeed,
recalling that $s\mapsto\psi_q(s)=
\varphi^{(q^{-1})}(s,\tau_q(s))$ is a Lipschitz function
$\mathbb{R} \to \mathbb{R}$ whose Lipschitz constant depends only
on $L$ (see below Definition \ref{d:mappsi_p}), we find
\begin{align}
\notag|\varphi^{(p^{-1})}&(s+y_2-y_1,\tau_p(s+y_2-y_1))-\varphi^{(q^{-1})}(s,\tau_q(s))-x|\\&\leq
|\varphi^{(q^{-1})}(s,\tau_q(s))-\varphi^{(q^{-1})}(s',\tau_q(s'))|\\\notag&+
|\varphi^{(p^{-1})}(s+y_2-y_1,\tau_p(s+y_2-y_1))-\varphi^{(q^{-1})}(s',\tau_q(s'))-x|\\
\notag&\overset{\eqref{eq:p_q_comparison}}{\lesssim}_{L,H}|s-s'| +
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}\\
&\overset{\eqref{eq:s_s'_comp}}{\lesssim}_{L,H}
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}.\label{eq:bound_two_s}
\end{align}
After these preparations, we are ready to deduce the desired upper
bound for \eqref{eq:TermToBeBounded} by considering $s:=y_3-y_2$.
We will show that
\begin{equation}\label{eq:CompacitbilityGoal}
d_{\mathbb{H}}(\Psi_p(w_1^{-1}\cdot w_3),p^{-1}\cdot q\cdot
\Psi_q(w_2^{-1}\cdot
w_3))\lesssim_{L,H}\max\{d_{\mathbb{H}}(w_1,w_2),d_{\mathbb{H}}(w_1,w_3),d_{\mathbb{H}}(w_2,w_3)\}^{1+\frac{\alpha}{2}}.
\end{equation}
It is convenient to estimate the expression on the left hand side
as follows
\begin{align}
\notag d_{\mathbb{H}}(\Psi_p(w_1^{-1}\cdot w_3)&, p^{-1}\cdot q\cdot
\Psi_q(w_2^{-1}\cdot w_3))\\
\label{i:I1} &\leq
d_{\mathbb{H}}(\Psi_p(w_1^{-1}\cdot w_3),\Psi_p(w_1^{-1}\cdot
w_2)\cdot
\Psi_q(w_2^{-1}\cdot w_3))\\
\label{i:I2} &+ d_{\mathbb{H}}(\Psi_q(w_2^{-1}\cdot w_3),
\Psi_p(w_1^{-1}\cdot w_2)^{-1}\cdot p^{-1}\cdot q\cdot
\Psi_q(w_2^{-1}\cdot w_3))
\end{align}
First, the term \eqref{i:I2} can be bounded using the fundamental
commutator relation as in \eqref{eq:FundCommRel} with
\begin{displaymath}
\mathfrak{a}:=[p\cdot \Psi_p(w_1^{-1}\cdot w_2)]^{-1}\cdot q
\quad\text{and}\quad \mathfrak{b}:=\Psi_q(w_2^{-1}\cdot w_3).
\end{displaymath}
This yields
\begin{align*}
\eqref{i:I2}&\lesssim \|[p\cdot \Psi_p(w_1^{-1}\cdot
w_2)]^{-1}\cdot q\|+ \| [p\cdot \Psi_p(w_1^{-1}\cdot
w_2)]^{-1}\cdot q\|^{\frac{1}{2}}\|\Psi_q(w_2^{-1}\cdot
w_3)\|^{\frac{1}{2}}
\\&\lesssim_{L,H}d_{\mathbb{H}}(w_1,w_2)^{1+\alpha}+
d_{\mathbb{H}}(w_1,w_2)^{\frac{1+\alpha}{2}}d_{\mathbb{H}}(w_2,w_3)^{\frac{1}{2}},
\end{align*}
where the last inequality follows from $q=i_{w_1\to p}(w_2)$, the
estimate \eqref{eq:Distance estimate}, and the Lipschitz
continuity of $\Psi_p$. Hence, recalling that
$d_{\mathbb{H}}(w_1,w_2)\leq 2$, the expression \eqref{i:I2} can be
bounded from above by the right hand side of
\eqref{eq:CompacitbilityGoal}.
Next, we handle the term \eqref{i:I1}. Since points on the
$t$-axis commute with all other elements in $\mathbb{H}^1$, it follows
from the definition of
$\Psi_p$ and $\Psi_q$ that \eqref{i:I1} is independent of the vertical components
of $w_1,w_2,w_3$. Writing these points in coordinates, as in
\eqref{eq:w1w2w3}, and recalling that $s=y_3-y_2$, we thus find
\begin{equation}\label{eq:ProdFormula}
\eqref{i:I1}=\|\Psi_p(s+y_2-y_1,0)^{-1}\cdot
\Psi_p(y_2-y_1,0)\cdot \Psi_q(s,0) \|.
\end{equation}
While $\Psi_p$ and $\Psi_q$ are in general not group
homomorphisms, their second components are linear:
\begin{displaymath}
[\Psi_p]_2(y,t)=[\Psi_q]_2(y,t)=y,\quad (y,t)\in \mathbb{W}.
\end{displaymath}
Thus, the second coordinate of the product in
\eqref{eq:ProdFormula} vanishes by linearity, and it suffices to
consider the first and and third coordinate, which we denote by
$I_1$ and $I_2$, respectively, so that
\begin{equation}\label{eq:finalStepI1}
\eqref{i:I1} \lesssim |I_1|+ |I_2|^{\frac{1}{2}}.
\end{equation}
Using that $\varphi^{(q^{-1})}(0,0)=0$, we may write
\begin{align*}
I_1=\varphi^{(p^{-1})}(y_2-y_1,\tau_p(y_2-y_1&))-\varphi^{(q^{-1})}(0,0)-x\\&+x+\varphi^{(q^{-1})}(s,\tau_q(s))-\varphi^{(p^{-1})}(s+y_2-y_1,\tau_p(s+y_2-y_1)).
\end{align*}
The term $I_1$ is the sum of two expressions of the same form as
in the estimate \eqref{eq:bound_two_s}, and we thus deduce that
\begin{displaymath}
|I_1|\lesssim
\max\{|y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}+
\max\{|s+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}.
\end{displaymath}
Clearly, $|y_2-y_1|\leq d_{\mathbb{H}}(w_1,w_2)$ and by the choice of
$s=y_3-y_2$, we have
\begin{displaymath}
|s+y_2-y_1|=|y_3-y_1|\leq d_{\mathbb{H}}(w_1,w_3).
\end{displaymath}
Thus we see that $|I_1|$ is bounded from above by the right hand
side of \eqref{eq:CompacitbilityGoal}, using again that
$d_{\mathbb{H}}(w_1,w_3),d_{\mathbb{H}}(w_1,w_2)\leq 2$. It remains to bound
$|I_2|$, where $I_2$ denotes the third component of the product in
\eqref{eq:ProdFormula}. A direct computation yields
\begin{displaymath}
I_2=
-\tau_p(s+y_2-y_1)+\tau_p(y_2-y_1)+\tau_q(s)+s\varphi^{(p^{-1})}(y_2-y_1,\tau_p(y_2-y_1)),
\end{displaymath}
and we continue as follows:
\begin{align*}
|I_2|&=|\tau_p(s+y_2-y_1)-\tau_p(y_2-y_1)-\tau_q(s)-s\varphi^{(p^{-1})}(y_2-y_1,\tau_p(y_2-y_1))|\\&=
\left|\int_{y_2-y_1}^{s+y_2-y_1} \dot{\tau}_p(\sigma)\,d\sigma-
\int_0^s \dot{\tau}_q(\sigma)+\varphi^{(p^{-1})}(y_2-y_1,\tau_p(y_2-y_1))\,d\sigma\right|\\
&=\left|\int_0^s
\dot{\tau}_p(\sigma+y_2-y_1)-\dot{\tau}_q(\sigma)-
\varphi^{(p^{-1})}(y_2-y_1,\tau_p(y_2-y_1))\,d\sigma\right|\\
&= \bigg| \int_0^s
\left[ \varphi^{(p^{-1})}(\sigma+y_2-y_1,\tau_p(\sigma+y_2-y_1))-\varphi^{(q^{-1})}(\sigma,\tau_{q}(\sigma))- x \right]\\
& \qquad + \left[x + \varphi^{(q^{-1})}(0,\tau_{q}(0)) - \varphi^{(p^{-1})}(y_2-y_1,\tau_p(y_2-y_1)) \right] \, d\sigma\bigg| \\
&\overset{\eqref{eq:bound_two_s}}{\lesssim}_{L,H} \int_{J_s}
\max\{|\sigma+y_2-y_1|,d_{\mathbb{H}}(w_1,w_2)\}^{1+\alpha}\,d\sigma\\
&\lesssim_{H,L} |s|
\max\{d_{\mathbb{H}}(w_1,w_2),d_{\mathbb{H}}(w_1,w_3)\}^{1+\alpha}\\&\lesssim_{H,L}
\max\{d_{\mathbb{H}}(w_1,w_2),d_{\mathbb{H}}(w_1,w_3),d_{\mathbb{H}}(w_2,w_3)\}^{2+\alpha},
\end{align*}
where $J_s:=[s,0]$ if $s\leq 0$ and $J_s:=[0,s]$ if $s\geq 0$. To
justify the application of \eqref{eq:bound_two_s} above, we have
applied inside the integral an analogous argument as we did to
bound the term $I_1$.
Finally inserting the bounds for $|I_1|$ and $|I_2|$ in
\eqref{eq:finalStepI1}, we conclude that \eqref{i:I1} is bounded
from above by the right hand side of
\eqref{eq:CompacitbilityGoal}. Combined with the bound for
\eqref{i:I2}, this concludes the proof of the lemma.
\end{proof}
\begin{proof}[Proof of Theorem \ref{mainVertical}]
The BP$G$BI condition "at unit scale" follows from Theorem
\ref{main}, whose hypotheses \eqref{ISO} and \eqref{comp} we have
verified in \eqref{eq:FirstCondVertical} and Proposition
\ref{p:NotesCor}, respectively. Here
\begin{equation*}\label{eq:G_Mnew} (G,d_{G},\mu) = (\mathbb{R}^2,d_{\Pi},\mathcal{L}^2), \quad \text{and} \quad (M,d_{M}) = (S,d_{\mathbb{H}}), \end{equation*}
with $x_{0} = 0 \in G$, and $p_{0} \in S$ arbitrary, and we recall
that $(G,d_G)$ is isometric to $(\mathbb{W},d_{\mathbb{H}})$. More precisely,
Theorem \ref{main} yields the existence of $2L$-bilipschitz maps
$f \colon K \to S \cap B(p,1)$, $p \in S$, where $K \subset G$
with $\mathcal{H}^{3}(K) \geq \delta >0$. The constant $L$ only depends
on the intrinsic Lipschitz constant of $\varphi$, and $\delta
> 0$ depends in addition on $\alpha$ and the constant $H$ in
\eqref{eq:VertHolDef}. Since property \eqref{eq:VertHolDef}
improves under ``zooming in'', see Remark
\ref{r:VertHolDilateTranslate}, we can argue analogously as in
Section \ref{ss:red_unit_scale}. Let $p \in S$ and, first, $0 < r
\leq C$, where $C := 2\operatorname{diam}_{\mathbb{H}}(\Phi(\operatorname{spt} \varphi))$. Using
Remark \ref{r:VertHolDilateTranslate} and the support assumption
on $\varphi$, we see that $S_{1/r} := \delta_{1/r}(S)$ is an
intrinsic graph of an intrinsic Lipschitz function (with the same
constant) satisfying \eqref{eq:VertHolDef} with constants $\alpha$
and $H'=H'(H,C)$.
Therefore, by the BP$G$BI property at scale $r = 1$, every ball
$S_{1/r} \cap B(p,1)$ contains the image of a $2L$-bilipschitz map
$g$ from a compact set $K \subset G$ with $\mathcal{H}^{3}(K) \geq
\delta = \delta(C) > 0$. Now, one may simply pre- and post-compose
$g$ with the natural dilations in $G$ and $\mathbb{H}^1$ to produce a
$2L$-bilipschitz map $g_{r} \colon \delta_{r}(K) \to S \cap
B(\delta_{r}(p),r)$ (note also that $\mathcal{H}^{3}(\delta_{r}(K)) =
r^{3}\mathcal{H}^{3}(K) \geq \delta r^{3}$).
Next, consider the case $r > C$. Then, if $p \in S$ is arbitrary,
the set $S \cap B(p,r)$ satisfies
\begin{displaymath} \mathcal{H}^{3}([S \cap B(p,r)] \cap \mathbb{W}) \gtrsim \mathcal{H}^{3}(S \cap B(p,r)). \end{displaymath}
Thus, the restriction of $\mathrm{Id}$ to $[S \cap B(p,r)] \cap
\mathbb{W}$ yields the desired bilipschitz map. The proof of Theorem
\ref{mainVertical} is thus complete.
\end{proof}
\subsection{Application to $C^1$ and intrinsic $C^{1,\alpha}$ surfaces}
As a first application of Theorem \ref{mainVertical}, we deduce
the case $n=1$ of Theorem \ref{mainGraphs}, recalling from Example
\ref{ex:intrC1alpha} that a compactly supported
$C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ function is intrinsic Lipschitz and
satisfies the extra vertical H\"older regularity condition.
\begin{thm}\label{mainGraphsn_1} Let $S = \Phi(\mathbb{W}) \subset \mathbb{H}^1$, where $\varphi \in C^{1,\alpha}_{\mathbb{H}}(\mathbb{W})$ is compactly supported.
Then $S$ has big pieces of bilipschitz images of the parabolic
plane $(\Pi,d_{\Pi})$.
\end{thm}
\begin{remark}
As a corollary of Theorem \ref{mainVertical}, we also obtain that
every Euclidean $C^1$ surface in $\mathbb{H}^1$ is rectifiable by
bilipschitz images of subsets of the parabolic plane. As written
in the introduction, this was known before by the work of
Cole-Pauls and Bigolin-Vittone, cf.\ Theorem \ref{CPBV}, but we
briefly explain how to deduce it from Theorem \ref{mainVertical}.
The reduction uses again the result by Balogh \cite{MR2021034},
which says that the set $\Sigma(S)$ of \emph{characteristic
points} of a Euclidean $C^1$ surface in $\mathbb{H}^1$ has vanishing
$3$-dimensional Hausdorff measure with respect to $d_{\mathbb{H}}$.
We will argue that outside $\Sigma(s)$, the surface $S$ can be
written locally as intrinsic graph of a compactly supported
Euclidean $C^1$ function, and hence as intrinsic Lipschitz graph
with extra vertical H\"older regularity. This will show that $S$
is rectifiable by bilipschitz images of subsets of the parabolic
plane.
We now turn to the details. Let $p \in S\setminus \Sigma(S)$. For
$r>0$ small enough, $S\cap B(p,r)$ is contained in the level set
$\{f=0\}$ of a Euclidean $C^1$ function
$f:\mathbb{R}^3\to\mathbb{R}$ with non-vanishing gradient in
$B(p,r)$. Without loss of generality, we may assume that $f(0)=0$,
$Xf(0)>0$ and
\begin{displaymath}
S\cap B(p,r)=\{q\in B(p,r):\; f(q)=0\}
\end{displaymath}
for $r>0$ with the property that $Xf(q)>0$ for all $q\in B(p,r)$.
Since $Xf(0)=\partial_x f(0)$, we may further assume, by making
$r$ smaller if necessary, that $\partial_x f(q)>0$ for all $q\in
B(p,r)$. In order to write $S\cap B(p,r)$, for small enough $r$,
as intrinsic graph of a Euclidean $C^1$ function, we first
consider the diffeomorphism
\begin{displaymath}
F:\mathbb{R}^3 \to \mathbb{R}^3,\quad
F(x,y,t)=\left(x,y,t+\tfrac{xy}{2}\right).
\end{displaymath}
Then $F(S\cap B(p,r))$ is contained in the level set of $f\circ
F^{-1}$, and hence it is again a Euclidean $C^1$ surface. Since
the derivative of $F$ at the origin is the identity, and
$\partial_x f(q)>0$ for all $q\in B(p,r)$, we can apply the usual
implicit function theorem to deduce that, if $r>0$ is small
enough, there is an open set $U\subset \mathbb{R}^2$, and a
Euclidean $C^1$ function $\psi: U \to \mathbb{R}$ such that
$F(S\cap B(p,r))$ is the Euclidean graph of $\psi$ over the set
$U$ in the $yt$-plane:
\begin{displaymath}
F(S\cap B(p,r))= \{(\psi(y,t),y,t):\; (y,t)\in U\}.
\end{displaymath}
It is easy to see that the preimage of this set under $F$ is then
given by the \textbf{intrinsic} graph of $\psi$,
\begin{displaymath}
S\cap B(p,r)=\left\{
(\psi(y,t),y,t-\tfrac{1}{2}y\psi(y,t)):\,(y,t)\in U\right\}.
\end{displaymath}
We will next modify $\psi$ to obtain a Euclidean $C^1$ function
$\varphi$ that is defined on the entire plane, but compactly
supported. To this end, let $B',B \subset U$ be concentric balls,
relatively open in the $yt$-plane $\mathbb{W}$ (identified with
$\mathbb{R}^2$), such that
\begin{displaymath}
[B(p,r') \cap S] \subseteq \{w\cdot \psi(w):\; w\in B'\} \subseteq
\{w\cdot \psi(w):\, w\in B\} \subseteq [B(p,r)\cap S]
\end{displaymath}
for some $0<r'<r$. We define
\begin{equation*}
\varphi (w) := \left\{\begin{array}{ll} \psi (w),&\text{if } w\in B',\\
\xi (w),&\text{if } w \in B\setminus B',
\\0,&\text{otherwise},\end{array}\right.
\end{equation*}
with a suitable $C^1$ function $\xi$ in order that $\varphi $ is
also $C^1$. More precisely, $\varphi $ is a compactly supported
$C^1$ function defined in $\mathbb{W}$ such that $S \cap B(p,r') =
\Phi(\mathbb{W})\cap B(p,r')$. By Example \ref{ex:EuclC1}, we know that
$\varphi$ is also an intrinsic Lipschitz function with extra
vertical regularity. Finally, it follows from Theorem
\ref{mainVertical} that $\Phi(\mathbb{W})$ is rectifiable by bilipschitz
images, and hence the same holds for $S \cap B(p,r')$. Repeating
the argument for every noncharacteristic point in $S$ proves that
$S$ is rectifiable by bilipschitz images of subsets of the
parabolic plane, and in particular LI rectifiable.\end{remark}
|
\section{Introduction}
The task of emotion recognition (ER) requires understanding the way humans interact to express their emotional state during conversations. Among others, emotions are encoded in both lexical and acoustic information where each modality contributes to the overall emotional state of a given speaker. However, in some situations, one modality can be more insightful to derive emotions than the other. For instance, the phrase \textit{``yeah... of course''} does not have enough lexical information to derive the right emotion, and it may all depend on the acoustic patterns. On the other hand, the phrase \textit{``I really miss my dog!''} does not need acoustic information to detect that the most likely emotion is sadness. Thus, recognizing emotions is not a trivial task because an emotional state can be easily shaped by many factors: context, word content, spectral and prosodic information, among others \citep{DBLP:journals/speech/BarbulescuRB17}.
In this paper, we study the emotion recognition problem from the speech and language perspectives. We formally look into acoustic and lexical modalities with the aim of improving models that only use acoustic information.
In the first part of this work, our goal is to assess the extent to which lexical information benefits acoustic models.
We propose a multimodal method that is inspired by the way humans process emotions in a conversation. That is, lexical and acoustic information is simultaneously perceived at every word step. Hence, we introduce the concept of acoustic words: word-level representations derived from acoustic features in a speech fragment. The acoustic word representations enable a natural combination of the modalities where lexical and acoustic features are aligned at the word level. Additionally, we leverage these representations with two attention mechanisms: modality-based and context-based attentions. The former mechanism prioritizes one of the modalities at each word step, whereas the latter mechanism focuses on the most important word representations across the entire utterance. Our multimodal approach outperforms the current state of the art on the USC-IEMOCAP dataset reported on lexical and acoustic modalities.
In the second part of this work, our goal is to induce semantic information from the proposed multimodal model into an acoustic model.
We study a more challenging scenario where we establish that lexical information is available during training but not during the evaluation phase. Such restriction is commonly found in real-world applications, where transcripts or ASR outputs represent a bottleneck in a deployment pipeline due to computational complexity or privacy-related constraints. To address this challenge, we frame this task as a multi-view learning problem \citep{Blum:1998:CLU:279943.279962}.
We induce lexical information from our multimodal model into the acoustic network during training while still providing a lexical-independent acoustic model for testing or deployment. That is, our acoustic model learns to capture semantic and contextual information without relying on explicit lexical inputs such as ASR or transcripts. This multi-view acoustic network significantly outperforms models that have been exclusively trained on acoustic features.
\section{Related Work}
Recognizing emotions is a complex task because it involves several ambiguous human interactions such as facial expressions, change in pitch or tone of voice, linguistic semantics and meaning, among others \citep{cowie2009perceiving, Provostetal2009}.
Many researchers have approached these challenges by extracting features from visual, acoustic, and lexical information. Early approaches rely on a variation of support vector machine (SVM) classifiers to learn emotional categories such as happiness, sadness, anger, and others \citep{rozgic2012ensemble, perezrosas-mihalcea-morency:2013:ACL2013, jinetal2015}. For instance, \citet{rozgic2012ensemble} use an automatically generated ensemble of trees whose nodes contain binary SVM classifiers for each emotional category. \citet{jinetal2015} also use multimodality, and their study focuses on comparing early and late-fusion methods. Consistently, researchers have found that multimodal approaches outperform unimodal ones.
Recent work has focused on different ways to fuse the acoustic, lexical, and visual modalities. However, we narrow the discussion to the acoustic and lexical modalities to align with the scope of the paper. In most of the cases, researchers have used concatenation to fuse the lexical and acoustic representations at different stages of their models. Other works have proposed multimodal pooling fusion \citep{Aldeneh:2017:PAL:3136755.3136760}, tensor fusion networks \citep{DBLP:journals/corr/ZadehCPCM17}, modality hierarchical fusion \citep{Majumderetal2018}, context-aware fusion with attention \citep{DBLP:journals/corr/abs-1803-07427}, and conversational memory networks (CMN) \cite{hazarika-EtAl:2018:N18-1}. Nevertheless, all the previous fusion techniques have been made at the utterance level, whereas our work focuses on multimodal fusion at the word level by introducing acoustic word representations.
We compare our work to \citet{DBLP:journals/corr/abs-1803-07427} because they document the current best performance on lexical and acoustic information on the IEMOCAP dataset using the standard 10-fold speaker-exclusive cross-validation setting.
Closely related work on acoustic word embeddings has been made by \citet{DBLP:journals/corr/HeWL16}. They induce acoustic information into lexical representations at the character level in a multi-view unsupervised setting. We introduce the concept of acoustic word representations in a different way: we learn vector representations of words out of frame-level acoustic features. This allows us to align lexical and acoustic information at the word level, which simulates the way humans perceive emotions in conversations (i.e., both modalities are simultaneously perceived).
We also explore multi-view settings to overcome the absence of lexical inputs during evaluation \citep{Blum:1998:CLU:279943.279962}. There are multiple options to conduct the experiments in this scenario \citep{DBLP:journals/corr/abs-1304-5634, DBLP:conf/icml/WangALB15}, such as deep cannonical correlation analysis (DCCA) \citep{Andrew:2013:DCC:3042817.3043076} and siamese networks with contrastive loss functions \citep{DBLP:journals/corr/HeWL16}. We use the latter approach in our experiments. To the best of our knowledge, there is no prior work trying to overcome the absence of lexical inputs by inducing lexical information into an acoustic model for the task of emotion recognition.
\section{Methodology}
We describe the data representation and introduce the idea of acoustic words in Sections \ref{sec:input_representation} and \ref{sec:acoustic_words}. Then, we use this concept to define the multimodal architecture in Section \ref{sec:multimodal_model}. Finally, we explain the multi-view learning setting using the proposed multimodal model and our acoustic model in Section \ref{sec:multimodal_model}.
\subsection{Data Representation}
\label{sec:input_representation}
\noindent\textbf{Acoustic features}. We extract frame-level features using OpenSMILE\footnote{\url{audeering.com/technology/opensmile/}} \citep{Eyben:2013:RDO:2502081.2502224}. We use the Computational Paralinguistic Challenge (ComParE) feature-set introduced by \citet{schuller2013interspeech} for the InterSpeech emotion recognition challenge. These features include energy, spectral, MFCC, and other low-level descriptors. The InterSpeech ComParE 2013 features are fairly standard and well-documented. Additionally, we normalize these features using z-standardization before feeding them into our models.
\noindent\textbf{Lexical features}. We use word embeddings to represent the lexical information. Specifically, we employ deep contextualized word representations using the language model ELMo
\citep{peters-EtAl:2018:N18-1}. ELMo represents words as vectors that are entirely built out of characters.
This allows us to overcome the problem of out-of-vocabulary words by always having a vector based on morphological clues for any given word. Additionally, these representations have proven to capture syntax and semantics aspects as well as the diversity of the linguistic context of words (e.g., polysemy).
\subsection{Acoustic Words}
\label{sec:acoustic_words}
Previous studies usually extract features from the modalities in independent modules, and then they concatenate the corresponding utterance representations from the acoustics and lexical features to feed into the next layers of their models. However, we argue that a more natural way to understand emotions is to align lexical and acoustic information, which simulates the way humans process both modalities simultaneously. Thus, we introduce the concept of acoustic word representations (see Figure \ref{fig:multimodal_model}).
These representations are extracted from frame-level features by taking the output of a bidirectional LSTM at every segment. Note that this procedure requires the word alignment information.
Additionally, we exclude frames that do not belong to the words of the speaker.
This reduces any potential bias towards other people's emotional states as well as environmental noise.
\subsection{Hierarchical Multimodal Model}
\label{sec:multimodal_model}
\begin{figure}[t!]
\centering
\includegraphics[width=0.8\linewidth]{images/multimodal_model6.png}
\caption{ The multimodal model. The shadowed box incloses the acoustic word mechanism, whose output is fed into the GMU unit along with the lexical word representation at each timestep. The model can have N layers of BLSTM at the frame and word levels. }
\label{fig:multimodal_model}
\end{figure}
Our goal is to provide a neural network model that efficiently combines acoustic and lexical information for emotion recognition. We propose a hierarchical multimodal model that uses: 1) acoustic word representations derived from frame-level features, 2) a modality-based attention mechanism at the word level that prioritizes one modality over the other, and 3) a context-based attention mechanism that emphasizes the most relevant parts in the entire utterance.
In Figure \ref{fig:multimodal_model}, the shadowed box represents the low level of the hierarchy, where the frame features are used to generate the acoustic word representation. The high level of the model is where the word representations from each modality are combined.
\noindent\textbf{Modality-based attention}. The idea of the modality-based attention is to prioritize one of the modalities at the word level. That is, when the lexical features are more relevant to capture emotions (i.e., informative words are used), the model should prioritize such features and vice versa (i.e., arousal and pitch levels increase). To achieve this behavior, we incorporate the bimodal version of the GMU cell proposed by \citet{arevalo2017gated}. The GMU equations are as follows:
\begin{equation} \label{eq:attention-modality}
\begin{split}
h_a =& ~\mathrm{tanh}(\mathrm{W}_a x_a + b_a) \\
h_l =& ~\mathrm{tanh}(\mathrm{W}_l x_l + b_l) \\
z =&~ \sigma(\mathrm{W}_z [x_a, x_l] + b_z) \\
h =&~ z * h_a + (1 - z) * h_l
\end{split}
\end{equation}
\noindent where $x_a$ and $x_l$ are the acoustic and lexical input vectors, respectively. These inputs are concatenated ($[x_a, x_l]$) and then multiplied by $\mathrm{W}_z$ so that the concatenation can be projected into the same space of the hidden vectors $h_a$ and $h_l$. Finally, $z$ is multiplied by the hidden acoustic vector $h_a$, and $(1- z)$ by the hidden lexical vector $h_l$. By adding the result of these products, the model incorporates a complementary mechanism over the modalities, which allows prioritizing one over the other when necessary.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.9\linewidth]{images/multiview_model5.png}
\caption{ The multi-view models. The view on the left is the acoustic model, and the view on the right is the multimodal model. The shadowed box in the middle is the contrastive loss module.}
\label{fig:multiview_model}
\end{figure*}
\noindent\textbf{Context-based attention}. We use a fairly standard attention mechanism over the entire utterance that was introduced by \citet{DBLP:journals/corr/BahdanauCB14}. The idea is to concentrate mass probability over the words that capture emotional states along the sequence. Our attention mechanism uses the following equations:
\begin{equation} \label{eq:attention-context}
\begin{split}
e_i =&~ v^\intercal ~\mathrm{tanh}(\mathrm{W}_h h_i + b_h) \notag\\
a_i =&~ \frac{\mathrm{exp}(e_i)}{\sum^{N}_{j=1}\mathrm{exp}(e_j)}, ~~~\text{where}~\sum_{i=1}^N a_i = 1 \notag\\
z =&~ \sum_{i=1}^N a_i h_i \notag
\end{split}
\end{equation}
where $\mathrm{W}_h \in \mathbb{R}^{d_a \times d_h}$ and $b_h \in \mathbb{R}^{d_h}$ are trainable parameters of the model. The vector $v \in \mathbb{R}^{d_a}$ is the attention vector to be learned. Also, $d_a$ and $d_h$ are the dimensions of the attention layer and the hidden state, respectively. Then, we multiply the scalars $a_i$ and their corresponding hidden vectors $h_i$ to obtain our weighted sequence. The sum of the weighted vectors, $z$, is fed into a softmax layer.
\subsection{Multi-view Learning}
\label{sec:multiview_model}
A more realistic and challenging scenario happens when lexical information is not available during testing. In this case, our goal is to build an acoustic model that is capable of inferring some notion of semantic and contextual features by taking advantage of lexical information only available during training. To achieve this, we frame the problem as a multi-view learning task, where two disjoint networks share their learned information through the loss function \citep{Lian:2018:SER:3267935.3267946}. The fact that they are disjoint networks allows them to function without each other during evaluation.
Consider the acoustic and multimodal views $V_a$ and $V_m$.
The acoustic view, $V_a$, is comprised of N layers of bidirectional LSTMs followed by an attention and a softmax layers. The multimodal view, $V_m$, follows the architecture described in Section \ref{sec:multimodal_model}.
As shown in Figure \ref{fig:multiview_model}, the view on the left, $V_a$, takes only the raw frame vectors, whereas the view on the right, $V_m$, takes the aligned frame and word vectors as inputs.
Each view learns an utterance representation of the emotions, $h_a$ and $h_m$, which are the outputs of their corresponding attention layers, as defined in Eq. \ref{eq:attention-context}. Since these vectors come from the same source of information (i.e., same speaker utterance), we assume that their emotion representations are similar. In general, we want vectors with similar emotions to be close and dissimilar ones to be far regardless of the modalities they use. To achieve this, we use the following contrastive loss function:
\noindent
\resizebox{\hsize}{!}{
\begin{minipage}{1.2\hsize}
\begin{align}\label{eq:contrastive_loss}
\mathcal{L}_{c} = &\frac{1}{2N}\sum_i^N{\mathrm{max}(0, m + dis(h_{a_i}, h^+_{m_i}) - dis(h_{a_i}, h^-_{m_i}))} \notag \\
+ &\frac{1}{2N}\sum_i^N{\mathrm{max}(0, m + dis(h_{m_i}, h^+_{a_i}) - dis(h_{m_i}, h^-_{a_i}))}
\end{align}
\vspace{0.2\baselineskip}
\end{minipage}
}
where the $+$ and $-$ superscripts refer to positive (i.e., close) and negative (i.e., far) vectors. We force a margin of at least $m$ to keep negative samples separated from positive samples. We define $dis(v,w) = 1 - cos(v, w)$ as the function that calculates the distance between two vectors. Note that we determine cross-view pairs when comparing vectors because we want the models to induce similar information from different modalities.
Additionally, choosing the negative samples can dramatically affect the performance of the models.
For instance, for random samples that may not share acoustic or lexical properties, the models can easily satisfy the margin $m$ without forcing much learning. Instead, we want the models to find the nuances in acoustically similar samples that have different emotion labels.
Thus, besides random sampling, we also consider similar acoustic properties (e.g., valence, arousal, or dominance) that overlap among the emotions.
In addition to the contrastive objective function, we use cross-entropy loss functions for the acoustic and multimodal views:
\begin{align}
\mathcal{L}_{a} =& - \frac{1}{N}\beta_a\sum_i^N y_i log(\hat{y}_i) \label{eq:aco-cross-entropy} \\
\mathcal{L}_{m} =& - \frac{1}{N}\beta_m\sum_i^N y_i log(\hat{y}_i) \label{eq:lex-cross-entropy}
\end{align}
\noindent where $\beta_a$ and $\beta_m$ are used to weight the loss from the acoustic and multimodal views, respectively. These weights can vary along the epochs to facilitate the optimization of the acoustic view. We discuss this in Section \ref{sec:multiview_experiment}, and the training procedure is described in Algorithm \ref{alg:training_algorithm}.
\begin{algorithm}
\caption{Multi-view Training Algorithm}
\label{alg:training_algorithm}
\begin{algorithmic}[1]
\small
\Procedure{GetNegSamples}{$Data, \mathrm{y}$}
\LineComment{Loop through the targets of the batch}
\For{$i \gets 1, \dots, \|\mathrm{y}\|$}
\LineComment{Randomly pick sample with class other than $\mathrm{y}_i$}
\State $y^-_i \gets$ \Call{Rand}{$Data$}
\Statex[4] ~s.t.~ $y^-_i \neq \mathrm{y}_i$ and
\Statex[5] ~~$y^-_i, \mathrm{y}_i$ are acoustically similar
\LineComment{Collect the corresponding negative inputs}
\State $(\mathrm{x}^-_{a_i}, \mathrm{x}^-_{l_i}) \gets getinput(y^-_i)$
\EndFor
\State \textbf{return} $(\mathrm{x}^-_{a}, \mathrm{x}^-_{l})$
\EndProcedure
\Repeat{:}
\LineComment{Loop through the training batches}
\For{$(\mathrm{x}_a, \mathrm{x}_l, \mathrm{y}) \gets nextbatch(Data)$}
\LineComment{Get the negative acoustic and lexical inputs}
\State $(\mathrm{x}^-_a, \mathrm{x}^-_l) \gets$ \Call{GetNegSamples}{$Data, \mathrm{y}$}
\LineComment{Get the neg. hidden vectors from neg. inputs}
\State $\mathrm{h}^-_{a} \gets hidden(V_a, \mathrm{x}^-_a)$
\State $\mathrm{h}^-_{m} \gets hidden(V_m, \mathrm{x}^-_a, \mathrm{x}^-_l)$
\LineComment{Get the pos. hidden vectors and predictions}
\State $(\mathrm{h}_{a}, \mathrm{\hat{y}}_{a}) \gets forward(V_a, \mathrm{x}_a)$
\State $(\mathrm{h}_{m}, \mathrm{\hat{y}}_{m}) \gets forward(V_m, \mathrm{x}_a, \mathrm{x}_l)$
\LineComment{Calculate and add the individual losses}
\State $\mathcal{L}_{c} \gets$ \Call{Contrastive}{$\mathrm{h}_{a}, \mathrm{h}^-_{a}, \mathrm{h}_{m}, \mathrm{h}^-_{m}$}
\State $\mathcal{L}_{a} \gets$ \Call{CrossEntropy}{$\mathrm{y}, \mathrm{\hat{y}}_{a}$}
\State $\mathcal{L}_{m} \gets$ \Call{CrossEntropy}{$\mathrm{y}, \mathrm{\hat{y}}_{m}$}
\State $\mathcal{L} \gets \mathcal{L}_{c} + \beta_a \mathcal{L}_{a} + \beta_m \mathcal{L}_{m}$
\LineComment{Update the parameters using backprop.}
\State $\Theta_{V_m} \gets \Theta_{V_m} - \alpha \partial \mathcal{L} / \partial \Theta_{V_m}$
\State $\Theta_{V_a} \gets \Theta_{V_a} - \alpha \partial \mathcal{L} / \partial \Theta_{V_a}$
\EndFor
\Until{stopping criteria met}
\end{algorithmic}
\end{algorithm}
\noindent \textbf{Teacher-student learning}.
We anticipate two potential problems with the previously described setting:
1) the learning process may predominantly concentrate on the multimodal view because it has more learning capabilities (i.e., large number of parameters) than the acoustic view, leaving the acoustic model to be of secondary importance during training,
and
2) a cross-entropy loss over one-hot vectors ignores informative overlaps among the emotion classes resulting in a very strict objective function. To address these issues, we look into a teacher-student learning approach \citep{softlabelsbetterthanhard}. Given an already-optimized multimodal model $V_m$ (the teacher), we want our acoustic view $V_a$ (the student) to predict probability distributions such as the ones generated by the teacher. We can calculate the difference between the probability distributions of the teacher and the student using Kullback-Leibler (KL) divergence. Then, we minimize the following loss function:
\begin{align} \label{eq:kl_div_loss}
\mathcal{L}_{KL} = - \frac{1}{N}\sum_i^N p(y_i|x_{m_i}, V_m) log\frac{p(y_i|x_{m_i}, V_m)}{p(y_i|x_{a_i}, V_a)}
\end{align}
where $x_{m_i}$ and $x_{a_i}$ are the multimodal and acoustic inputs for sample $i$, respectively, and $V_m$ and $V_a$ represent the parameters of the views.
\section{Experiments}
We describe the dataset used for the experiments in Section \ref{sec:data_and_feats}. Then, we define the experimental models in Section \ref{sec:experimental_models}, which are used in the multimodal and multi-view experiment in Sections \ref{sec:multimodal_experiment} and \ref{sec:multiview_experiment}.
\subsection{Dataset}
\label{sec:data_and_feats}
\begin{table}[t!]
\centering
\renewcommand{\arraystretch}{0.7}
\setlength{\tabcolsep}{5pt}
\small
\begin{tabular}{rllll}
\toprule
\textbf{Utterances} & \textbf{Anger} & \textbf{Happiness} & \textbf{Neutral} & \textbf{Sadness} \\\midrule
F1 - 528~~ & 147 & 132 & 171 & 78 \\
M1 - 556~~ & 82 & 146 & 212 & 116 \\\midrule
F2 - 479~~ & 67 & 166 & 134 & 112 \\
M2 - 542~~ & 70 & 161 & 227 & 84 \\\midrule
F3 - 522~~ & 92 & 128 & 130 & 172 \\
M3 - 624~~ & 148 & 154 & 190 & 132 \\\midrule
F4 - 527~~ & 205 & 185 & 75 & 62 \\
M4 - 501~~ & 122 & 118 & 180 & 81 \\\midrule
F5 - 590~~ & 78 & 159 & 221 & 132 \\
M5 - 651~~ & 92 & 283 & 163 & 113 \\\midrule
5,520~~ & 1,103 & 1,632 & 1,703 & 1,082 \\
\bottomrule
\end{tabular}
\caption{Data distribution of the USC-IEMOCAP dataset. F and M mean female and male speakers followed by their session number. }
\label{tab:eimocap_distr}
\end{table}
\begin{table*}[t!]
\centering
\renewcommand{\arraystretch}{0.9}
\setlength{\tabcolsep}{10pt}
\begin{tabular}{llllll}
\toprule
\textbf{Type} & \textbf{Experiment} & \textbf{Modality} & \textbf{Dev} & \textbf{Test} & \textbf{Comment} \\
\midrule
\multirow{5}{*}{Baseline} & B-ACO-1 & \multirow{2}{*}{Acoustics} & 0.5858 & - & Silence frames \\
& B-ACO-2 & ~ & 0.5729 & - & Silence frames removed \\
\cmidrule{2-6}
& B-LEX & Lexical & 0.6706 & - & - \\
\cmidrule{2-6}
& B-MM-1 & \multirow{2}{*}{Multimodal} & 0.7195 & - & Silence frames \\
& B-MM-2 & ~ & 0.7265 & - & Silence frames removed \\
\midrule
\multirow{5}{*}{Hierarchical} & H-ACO-1 & Acoustics & 0.5697 & - & Acoustic words \\
\cmidrule{2-6}
& H-MM-1 & \multirow{4}{*}{Multimodal} & 0.7316 & - & Aligned words \\
& H-MM-2 & ~ & 0.7341 & - & ~~~+ GMU \\
& H-MM-3 & ~ & 0.7354 & - & ~~~+ Attention \\
& H-MM-4 & ~ & \bf0.7383 & \bf0.7169 & ~~~+ GMU + Attention \\
\midrule
SOTA & - & Multimodal & - & \textbf{0.7079} & \citet{DBLP:journals/corr/abs-1803-07427} \\
\bottomrule
\end{tabular}
\caption{The results of the multimodal experiments. The name of the experiments starts either with B or H referring to baseline or hierarchical models. ACO, LEX, and MM mean acoustic, lexical and multimodal. Our results provide a new state-of-the-art UA when we use the hierarchical model with GMU and attention. Once the models are optimized on the validation set, we evaluate the best ones on the test set.}
\label{tab:multimodal_results}
\end{table*}
We focus our experiments on the USC-EIMOCAP dataset \citep{busso2008iemocap}. This dataset provides conversations between female and male speakers throughout five sessions. Each session involves a different pair of speakers, which accounts for a total of 10 speakers. The conversations are split into small utterances that map to emotion categories. The original emotion categories are merged to mitigate the unbalanced classes into four categories: \textit{anger}, \textit{happiness}, \textit{neutral}, and \textit{sadness}. Table \ref{tab:eimocap_distr} shows the distribution of the dataset. We split the dataset using the one-speaker-out experimental setting. That is, we take four sessions for training, and the remaining session is split by speakers into the validation and test sets. We report our unweighted accuracy scores running 10-fold cross-validation experiments and averaging scores across folds.
\subsection{Defining Experimental Models }
\label{sec:experimental_models}
\noindent \textbf{B-ACO}: The acoustic baseline is composed of two BLSTM layers of 256 dimensions each, followed by average pooling and a softmax layer. B-ACO-1 uses the raw sequence of frames, whereas B-ACO-2 employs the frames that correspond to the speaker.
\noindent \textbf{B-LEX}: The lexical baseline uses word embeddings of 1,024 dimensions from ELMo. We feed these vectors into two BLSTM layers of 256 dimensions followed by average pooling and a softmax layer.
\noindent \textbf{B-MM}: The multimodal baseline uses BLSTMs with average pooling over time on each modality, similar to B-ACO and B-LEX. We concatenate the vectors from each modality and feed them into a softmax layer.
\noindent \textbf{H-ACO}: The hierarchical acoustic model uses acoustic word representations. The acoustic words are generated with two BLSTMs of 256 dimensions using the speaker frames (i.e., no silence). At the word level, we perform average pooling over time and feed the resulting vector into a softmax layer.
\noindent \textbf{H-MM}: The hierarchical multimodal model uses the acoustic word representations in H-ACO, and the lexical word representations in B-LEX, with 256 dimensions each. H-MM-1 uses two layers of BLSTM over the concatenated word representations followed by average pooling and a softmax layer. Based on H-MM-1, H-MM-2 incorporates the GMU unit and H-MM-3 adds the attention layer. H-MM-4 uses both GMU and the attention layer.
\subsection{Multimodal Experiments }
\label{sec:multimodal_experiment}
\begin{figure*}[t!]
\centering
\includegraphics[width=\linewidth]{images/attention_viz6.png}
\caption{ Multimodal Attention. The figure shows the attention mechanisms at the modality and utterance levels. The bars over the words are the average of $z$ in Eq. \ref{eq:attention-modality}, and they show how much acoustic (left bar in blue) or lexical (right bar in red) information was used. The highlights in the background of the words are the attention probabilities, where the higher the probability the darker the word. }
\label{fig:multimodal_global_attention}
\end{figure*}
\noindent\textbf{Impact of silence}.
We experiment with silence and the baselines B-ACO and B-MM. In Table \ref{tab:multimodal_results}, although keeping silence seems better than removing it (B-ACO-1 vs. B-ACO-2), the multimodal model shows a small improvement when silence is ignored (B-MM-1 vs. B-MM-2).
By looking into the predictions, besides the silence and environmental noise in the original frames, we notice that a second speaker can influence the emotions of the speaker being evaluated. This observation, along with the model improvements, suggests that is possible to fuse information more efficiently.
\noindent\textbf{Hierarchical models}. To make better use of the modalities, we align lexical information with acoustic representations at the word level. Based on the silence impact, our acoustic word representations only use frames where the speaker intervenes in the conversation (i.e., no silence or other speakers). Similar to the previous scenario, we see a detrimental behavior in the hierarchical acoustic model compared to the models that use the original sequence of frames (H-ACO-1 vs. B-ACO). However, when we concatenate the lexical and acoustic word representations (H-MM-1), our hierarchical model surpasses the UA of all previous models. In fact, our best model (H-MM-4) outperforms the previous state-of-the-art UA. This serves as strong evidence that fusing information more efficiently can yield a better performance.
\noindent\textbf{Ablation experiment}. Table \ref{tab:multimodal_results} shows the performance of the hierarchical multimodal models with and without the modality- and context-based attention mechanisms (H-MM). Using H-MM-1 as a common ground, the modality-based attention (H-MM-2) provides an improvement of about 1\% on the UA metric. This result suggests that one modality can be more informative than the other, and hence, it is important to prioritize the one that carries more emotional information. Likewise, adding the attention mechanism, H-MM-3, outperforms H-MM-1 by a similar percentage. Our intuition is that weighting the words that provide strong emotional information based on the context allows the model to disambiguate meaning and discriminate more easily the samples. Lastly, H-MM-4 combines both attention mechanisms, which improves over the individual attention models H-MM-2 and H-MM-3 by about 1\% of UA. This means that the attention mechanisms are more complementary than overlapping.
\noindent\textbf{Attention visualization}.
For the modality-based attention, the vector $z$ from Eq. \ref{eq:attention-modality} determines how much acoustic information will go through the next layers, whereas $(1-z)$ is the amount of lexical data allowed.
Figure \ref{fig:multimodal_global_attention} provides a visualization of these vectors. The bars show the amount of information that is captured from one modality versus the other. For instance, the sample \textit{``oh my gosh''} illustrates that the words rely on more acoustic than lexical information. Intuitively, this phrase by itself could describe different emotions, but it is the acoustic modality that mitigates the ambiguity.
Regarding the context-based attention, Figure \ref{fig:multimodal_global_attention} shows the places where the model focuses along the utterance. For large-context utterances, where the acoustic features are more or less similar, the semantics can help to highlight specific spots.
For example, in the second sentence on the right of Figure \ref{fig:multimodal_global_attention}, the model detects the semantics of the words \textit{sense} and \textit{stupid} and associates them with the words \textit{should}, \textit{go}, and \textit{army}. The attention mechanism not only emphasizes semantics but it also takes into account the acoustic features. In the same block of sentences, it is worth noting that the words primarily driven by acoustics (e.g., \textit{sweatheart}, \textit{oh god}, \textit{sorry} and \textit{yeah}) are highlighted by the attention mechanism. These results also align with the intuition that the attention mechanisms are complementary.
\begin{table*}[t!]
\centering
\renewcommand{\arraystretch}{1.1}
\setlength{\tabcolsep}{9pt}
\begin{tabular}{lllll}
\toprule
\bf View1 & \bf View 2 & \bf Dev & \bf Test & \bf Comment\\
\midrule
B-ACO-1 & - & 0.5858 & 0.5443 & Acoustic-exclusive baseline \\
\midrule
\multirow{4}{*}{B-ACO-1}
& B-LEX
& 0.5971
& -
& Loss: $\mathcal{L}_c + \mathcal{L}_a + \mathcal{L}_m$
(Eqs. \ref{eq:contrastive_loss}, \ref{eq:aco-cross-entropy}, \ref{eq:lex-cross-entropy}) \\
& H-MM-4
& 0.5976
& -
& Loss: $\mathcal{L}_c + \mathcal{L}_a + \mathcal{L}_m$
(Eqs. \ref{eq:contrastive_loss}, \ref{eq:aco-cross-entropy}, \ref{eq:lex-cross-entropy}) \\
& H-MM-4 $\dagger$
& 0.5969
& -
& Loss: $\mathcal{L}_c + \mathcal{L}_a$
(Eqs. \ref{eq:contrastive_loss}, \ref{eq:aco-cross-entropy}) \\
& H-MM-4 $\dagger$
& \bf0.6060
& \bf0.5859
& Loss: $\mathcal{L}_c + \mathcal{L}_{KL}$ (Eqs. \ref{eq:contrastive_loss}, \ref{eq:kl_div_loss}) \\
\midrule
B-ACO-1 + Attention
& H-MM-4 $\dagger$
& \bf0.6100
& \bf0.5976
& Loss: $\mathcal{L}_c + \mathcal{L}_{KL}$ (Eqs. \ref{eq:contrastive_loss}, \ref{eq:kl_div_loss}) \\
\bottomrule
\end{tabular}
\caption{The results of the multi-view experiments. We use the acoustic model B-ACO-1 as the first view and evaluate its performance using different second views. $\dagger$ means that the second view is not updated during training and its classification loss is not included.
}
\label{tab:multiview_results}
\end{table*}
\subsection{Multi-view experiments}
\label{sec:multiview_experiment}
Our multi-view experiments use utterance-level representations to calculate the contrastive loss in Eq. \ref{eq:contrastive_loss}. We discard experiments at the word level because 1) contrasting emotions for every word individually poses a complex task\footnote{Negative words are hard to choose because we want properly formed utterances with the same number of words.}, and 2) context helps to disambiguate meaning as well as to convey the overall emotion rather than relying on high emotional words individually. Additionally, our experiments aim at a more practical scenario where there is no need for transcripts or ASR output with forced alignment.
\noindent\textbf{Choosing negative samples}.
To calculate the loss as in Eq. \ref{eq:contrastive_loss}, we randomly choose negative samples in two ways: 1) forcing a different class, and 2) forcing a different class that is acoustically similar to the positive sample (e.g., \textit{sadness} vs. \textit{neutral}, or \textit{anger} vs. \textit{happiness}).
We saw that the model generalizes better using the second option. Our intuition is that the model does not have problems to force the margin $m$ between vectors when the negative input samples come from fairly easy discriminative classes (e.g., \textit{happiness} vs. \textit{neutral}). In contrast, the model struggles to force the margin $m$ between vectors when classes are acoustically similar, which turns into better generalization.
\noindent\textbf{Different views}. We choose B-ACO-1 as the first view because it uses raw frame level features. As shown in Table \ref{tab:multiview_results}, we compare B-LEX and H-MM-4 as simple and elaborated second views by applying the contrastive and the views' cross-entropy loss functions. Indeed, by using B-LEX we show that the acoustic model B-ACO-1 improves its accuracy. Further improvements are made if we use H-MM-4 as a second view. This means that it is better to transfer information to the acoustic model when the modalities are effectively combined rather than when we try to induce only lexical information.
\noindent\textbf{Frozen weights}. We further explore H-MM-4 as a second view by first optimizing it, and then fixing its weights in the multi-view setting.
Experiments with a trainable second view show that the lexical model is prioritized even when the losses are weighted as in Eq. \ref{eq:aco-cross-entropy} and \ref{eq:lex-cross-entropy}. The intuition is that there is nothing new that this second view can learn from the multi-view setting once it has been optimized separately, and thus, it is better to exclude the complexity of learning it from scratch. Table \ref{tab:multiview_results} shows a small improvement over the previous models reaching 59.69\% of UA on the validation set.
\noindent\textbf{Teacher-student learning}. We also experiment with a teacher-student setting where the model H-MM-4 is optimized separately. This model is a non-trainable second view where its class predictions are used as soft labels to evaluate the first view. The idea is to provide informative similitudes among the training samples by evaluating against a probability distribution over the classes rather than hard labels. The model reduces its loss more steadily than previous models, and once optimized, it surpasses previous results. Finally, we consider the case of a more complex student network since previous studies suggest that small student models may not be able to cope with the teacher models \citep{softlabelsbetterthanhard, adversarialteacherstudent}. By adding an attention layer over the acoustic model B-ACO-1, we are able to improve the accuracy of the model by 1\% absolute points, as shown in Table \ref{tab:multiview_results}.
\section{Conclusions}
We presented multimodal and multi-view approaches for emotion recognition. The first approach assumes that lexical information is always available when the speech signal is being processed. For such a scenario, our hierarchical multimodal model outperforms the state-of-the-art score with the aid of modality- and context-based attention mechanisms. The second approach adapts to a more realistic scenario where lexical data may not be available for evaluation. Our multi-view setting has shown that acoustic models can still benefit from lexical information over models that have been exclusively trained on acoustic features.
|
\section{Introduction}
In ordinal computability, "clockability" denotes the property of an ordinal that it is the halting time of some program. The term was introduced in \cite{HL}, which was the paper that started the area of ordinal computability by introducing Infinite Time Turing Machines (ITTMs). By now, a lot is known about clockability for ITTMs. To give a few examples: In \cite{HL}, it was proved that there are gaps in the ITTM-clockable ordinals, i.e., there are ordinals $\alpha<\beta<\gamma$ such that $\alpha$ and $\gamma$ are ITTM-clockable, but $\beta$ is not. Moreover, it is known that no admissible ordinal is ITTM-clockable (Hamkins and Lewis, \cite{HL}), that the first ordinal in a gap is always admissibles (Welch, \cite{W}), that the supremum $\lambda$ of the ITTM-writable ordinals (i.e. ordinals coded by a real number that is the output of some halting ITTM-computation) equals supremum of the ITTM-clockable ordinals (Welch, \cite{W}) and that ITTM-writable ordinals have real codes that are ITTM-writable at the point the next clockable appears. Moreover, it is known that not every ITTM-admissible below $\lambda$ starts a gap, there are admissibles properly inside gaps, and occasinally many of them (Carl, Durand, Lafitte, Ouazzani, \cite{CDLO}). And indeed, clockability turned out to be a central topic in ordinal computability; it was, for example, crucial for Welch's analysis of the computational strength of ITTMs.
Besides ITTMs, clockability was also considered for Infinite Time Register Machines (ITRMs), where the picture turned out to be quite different: In particular, there are no gaps in the ITRM-clockable ordinals (see \cite{CFKMNW}), and in fact, the ITRM-clockable ordinals are exactly those below $\omega_{\omega}^{\text{CK}}$, which thus includes $\omega_{n}^{\text{CK}}$ for every $n\in\omega$, i.e. the first $\omega$ many admissible ordinals.
For other models, clockability received comparably little attention. This work arose out of a question of T. Kihara during the CTFM (International Conference on Computability Theory and Foundations of Mathematics) conference in 2019 in Wuhan who, after hearing that admissible ordinals are never ITTM-clockable, asked whether the same holds for OTMs. After most of the results of this paper had been proved, we found two questions in the report of the $2007$ BIWOC (Bonn International Workshop on Ordinal Computability) \cite{BIWOC} concering this topic: the first (p. 42, question 9), the first, due to J. Reitz, asking whether $\omega_{1}^{\text{CK}}$ was OTM-clockable, the second, due to J. Hamkins, whether gap-starting ordinals for OTMs can be characterized as something stronger than being admissible. Both are considered to be answered by the claim that no admissible ordinal is OTM-clockable, which is attributed to J. Reitz and S. Warner. Upon personal inquiry, Reitz told us that they had a sketch of a proof which, however, did not entirely work; what it does show with a few modifications, though, is that $\Sigma_{2}$-admissible ordinals are not OTM-clockable, and the argument that Reitz sketched in personal correspondence to us in fact resembles the one of Theorem \ref{sigma2 not otm clockable} below. We thus regard Reitz and Warner as the first discoverers of this theorem. Both the argument of Reitz and Warner from 2007 and the one we found during the CTFM in 2019 are adaptations of Welch's argument that admissible ordinals are not ITTM-clockable.
The statement actually made in BIWOC, is, however, false: As we will show below, $\omega_{n}^{\text{CK}}$ is OTM-clockable for any $n\in\omega$. Thus, there are plenty of admissible ordinals that are OTM-clockable, and the answer to the first question is positive. The idea is to use the ITRM-clockability of these ordinals, which follows from Theorem [no gaps] in \cite{CFKMNW}, together with a slightly modified version of the obvious procedure for simulating ITRMs on OTMs. This actually shows that $\omega_{n}^{\text{CK}}$ is clockable on an ITTM with tape length $\alpha$ as soon as $\alpha>\omega$. Thus, the strong connection between admissibility and clockability seems to depend rather strongly on the details of the ITTM-architecture. We remark that this is a good example of how the studies of different models of infinitary computability can fruitfully interact: At least for us, it would not have been possible to find this result while only focusing on OTMs.
Moreover, we will answer the second question in the positive as well by showing that, if $\alpha$ starts a gap in the OTM-clockable ordinals, then $\alpha$ is an admissible limit of admissible ordinals.
Of course, the space between "admissible limit of admissible ordinals" and "$\Sigma_{2}$-admissible" is rather broad. In particular, we do not know whether every gap starting ordinals for OTMs is $\Sigma_{2}$-admissible, though we conjecture this to be false.
\section{Ordinal Turing Machines}
Ordinal Turing Machines (OTMs) were introduced by Koepke in \cite{Koe1} as a kind of "symmetrization" of ITTMs: Instead of having a tape of length $\omega$ and the whole class of ordinals as their working time, OTMs have a tape of proper class length $\text{On}$ while retaining $\text{On}$ as their "working time" structure. We refer to \cite{Koe1} for details.
In contrast to Koepke's definition but in closer analogy with the setup of ITTMs, we allow finitely many tapes instead of a single one. Though models of ordinal computability generally enjoy a good degree of stability under such variations as far as computational strength is concerned, this often makes a difference when it comes to clockability. Intuitively, simulating several tapes with separate read-write-heads on a single tape requires one to check the various head positions to determine whether the simulated machine has halted, which leads to a delay in halting. For ITTMs, this is e.g. demonstrated in \cite{SH}. For OTMs, insisting on a single tape would lead to a theory that is "morally" the same as the one described here, but make the results much less compelling and the proofs more technically involved and harder to follow. Thus, allowing multiple tapes seems to be a good idea.
The following picture of OTM-computations may be useful to some readers: Let us imagine the tape split into $\omega$-block. Then an OTM-computation proceeds like this: The head works for a bit in one $\omega$-block, then leaves it to the right, works for a bit in the new $\omega$-portion, again leaves it to the right and so on, until eventually the computation either halts or the head is moved back from a limit position, i.e., goes back to $0$ and starts over. Thus, if one imagines an $\omega$-portion as single point, then the head moves from left to right, jumps back to $0$, moves right again etc. Moreover, in each $\omega$-portion, we have a classical ITTM-computation (up to the limit rules for the head position and the inner state, which make little difference).
We fix some terminology for the rest of this paper.
\begin{defini}
If $M$ is one of ITRM, ITTM or OTM and $\alpha$ is an ordinal, then $\alpha$ is called $M$-clockable if and only if there is an $M$-program that halts at time $\alpha+1$.\footnote{The $+1$ allows limit ordinals to appear as haltling times and thus simplifies the theory.} $\alpha$ is called $M$-writable if and only if there is a real number coding $\alpha$ that is $M$-computable. An $M$-clockable gap is an interval $[\alpha,\beta)$ of ordinals such that $\alpha<\beta$, no element of $[\alpha,\beta)$ is $M$-clockable and $[\alpha,\beta)$ is maximal in the sense that there are cofinally many $M$-clockable ordinals below $\alpha$ and $\beta$ is $M$-clockable. In this case, we say that $\alpha$ "starts" the gap and call $\alpha$ a "gap starting ordinal" or "gap starter" for $M$.
\end{defini}
\section{Basic observations}
We start with some useful observations that can mostly be obtained by easy adaptations of the corresponding results about ITTM-clockability.
We start by noting that the analogue of the speedup-theorem for ITTMs from \cite{HL} holds for multitape-OTMs. As the proof -- an adaptation of the argument the speedup-theorems for ITTMs -is somewhat messy and the statement is not needed in this paper, we merely sketch the proof. (The main difference is that, in contrast to ITTMs, OTMs do not have their head on position $0$ at every limit time and that the head may make long "jumps" when moved to the left from a limit position.
This generates a few extra complications.) To simplify the proof, we start by building up a few preliminaries.
For the ITTM-speedup, the following compactnes property is used: If $P$ halts in $\delta+n$ many steps and the head is located at position $k$ at time $\delta$, then only the $n$ cells contents before and after the $k$th one at time $\delta$ are relevant for this. Now, this is a fixed string $s$ of $2n$ bits. In \cite{HL}, a construction is described that achieves that the information whether these $2n$ cells currently contain $s$ at a limit time $\gamma$ is coded on some extra tapes at time $\gamma$. Due to the special limit rules for ITTMs that set the head back to position $0$ at every limit time, the Hamkins-Lewis-proof has this information stored at the initial tape cells, but the construction is easily modified to store the respective information on any other tape position.
We will use it in the following way: Suppose that $P$ is an OTM-program that halts at time $\delta+n$, where $\delta$ is a limit ordinal and $n\in\omega$. We want to "speed up" $P$ by $n$ steps, i.e. to come up with a program $Q$ that halts in $\delta$ many steps. Suppose that $P$ halts with the head on position $\gamma+k$, where $\gamma$ is a limit ordinal and $k\in\omega$. $m$ be $k-n$ if $k-n\geq 0$ and $0$, otherwise, and let $s$ be the bit string present on positions $\gamma+m$ until $\gamma+k+n$ at time $\delta$. Then we use the Hamkins-Lewis-construction to take care that the information whether the bit string present on positions $\eta+m$ until $\eta+k+n$ is equal to $s$ on the $\eta+k$th cells of three extra tapes, for each limit ordinal $\eta$.
An extra complication arises from the possibility of a "setback": Within the $n$ steps from time $\delta$ to time $\delta+n$, it may happen that the head is moved left from position $\delta$, thus ending up at the start of the tape. Clearly, it will then take $<n$ many further steps at the start of the tape and only consider the first $n$ bits during this time. However, we need to know what these bits are - or rather, whether they are the "right ones", i.e., the ones present at time $\delta$ - while our head is located at position $\delta+k$. The idea is then to store this information in the inner state of the sped-up program. We thus create extra states: The new state $2i$ will represent the old state $i$ together with the information that the first $n$ bits where the "right ones" (i.e. the same ones as at time $\delta$) and $2i+1$ will represent the old state $i$ together with the information that some of these bits deviated from the one at time $\delta$. To achieve this, we use an extra tape $T_{4}$. At the start of $Q$, a $1$ is written to each of the first $n$ cells of $T_{4}$; after that the, head is set back to position $0$ and then moved along with the head of $P$. In this way, we will always know whether the head of $P$ is currently located at one of the first $n$ cells. Whenever this is the case, we insert some intermediate steps to read out the first $n$ bits, update the inner state and move the head back to its original position. (This requires some additional states, but we skip the details.) Note that, if $\eta$ is a limit time and the first $n$ bits have been changed unboundedly often below $\eta$, then the head will be located at one of these positions at time $\eta$ by the liminf-rule and thus, a further update will take place so that the state will correctly represent the configuration afterwards. On the other hand, if the first $n$ bits were only changed boundedly often before time $\eta$, then let $\bar{\eta}$ be the supremum of these times. We just saw that the state will represent the configuration correctly finitely many steps after time $\bar{\eta}$, after which the first $n$ cell contents remain unchanged, so that the state is still correct at time $\eta$.
In the following construction, we will need to know whether the head is currently located at a cell the index of which is of the form $\delta+k$, where $\delta$ is a limit ordinal and $k$ is a natural number. To achieve this, we add three tapes $T_0$, $T_1$ and $T_2$ to $P$. The tape $T_{0}$ serves as a flag: By having two cells with alternating contents $01$ and $01$, we can detect a limit time as a time at which both cells contain $0$. On $T_2$, we move the head along with the head on $P$ and place a $1$ on a cell whenever we encounter a cell on which a $0$ is written. Thus, the head occupies a certain limit position for the first time if and only if the head on $T_{1}$ reads a $0$ at a limit time. Finally, on $T_{2}$, we more the head along with the heads on $T_{1}$ and the main tape. Whenever the head on $T_{1}$ reads a $0$ at a limit time, we interrupt the computation, move the head on $T_{2}$ for $k$ many steps to the right, write a $1$, move the head $k$ many places to the left, and continue. In this way, the head on $T_{2}$ will read a $1$ if and only if the head on the main tape is at a position of the desired form. As this merely inserts finitely many steps occasionally, running this procedure along with an OTM-program $P$ will still carry out $\delta$ many steps of $P$ at time $\delta$ whenever $\delta$ is a limit ordinal. We will say that the head is "at a $\delta+k$-position" if the index of the cell where it is currently located is of this form with $\delta$ a limit ordinal and, by the construction just described, we can use formulations like "if the head is currently at a $\delta+k$-position" without affecting the running time at limit ordinals.
\begin{lemma}{\label{OTM speedup}}
If $\alpha+n$ is OTM-clockable and $n\in\omega$, then $\alpha$ is OTM-clockable.
\end{lemma}
\begin{proof}
It is clear that finite ordinals are OTM-clockable and that OTM-clockable ordinals are closed under addition (by simply running one program after the other).\footnote{It is folklore (and easy to see) that, for any reasonable model of computation, clockable ordinals are closed under ordinal arithmetic, i.e. under addition, multiplication and exponentiation, see e.g. \cite{HL} or \cite{CFKMNW}. This also holds true for OTMs.} Thus, if $\alpha$ is clockable, then so is $\alpha+m$ for any finite $m$ and hence it suffices to consider the case that $\alpha$ is a limit ordinal. Moreover, we assume for simplicity that $P$ uses only one tape; if $P$ uses several tapes, the construction below is carried out for each of these.
Let $P$ be an OTM-program that runs for $\alpha+n$ many steps, where $\alpha$ is a limit ordinal. We want to construct a program $Q$ that runs for $\alpha$ many steps. Let the head position at time $\alpha$ be equal to $\delta+k$, where $\delta$ is a limit ordinal and $k\in\omega$. As above, let $m$ be $k-n$ if $k-n\geq 0$ and otherwise let $m=0$. Let $s$ be the bit string present on the positions $\delta+m$ until $\delta+k+n$ at time $\alpha$, and let $t$ be the string present on the first $n$ positions.
Using the constructions explained above, $Q$ now works as follows: Run $P$. At each step, determine whether the head is currently at a location of the form $\eta+k$ with $\eta$ a limit ordinal and whether one of the two following conditions holds:
\begin{enumerate}
\item The head is currently at one of the first $n$ positions and the bit string currently present on the positions $\eta+m$ up to $\eta+k+n$ is equal to $s$.
\item The head is currently not on one of the first $n$ positions, the bit string currently present on the positions $\eta+m$ up to $\eta+k+n$ is equal to $s$ and whether the bit string currently present on the first $n$ positions is equal to $t$.
\end{enumerate}
If not, continue with $P$. Otherwise, halt. As described above, the necessary information can be read off from the various extra tapes and the inner state simultaneously. Now it is clear that, if $Q$ halts at time $\beta$, then $P$ will halt at time $\beta+n$. Thus, $Q$ halts at time $\alpha$, as desired.
\end{proof}
\begin{lemma}{\label{OTM multitape simulation}}
Suppose that $\alpha$ is exponentially closed and
clockable by an OTM. The
$\alpha\cdot 2$ is clockable by an OTM using only one tape.
\end{lemma}
\begin{proof}
Usual simulation with various tape portions. Work in stages: simulate the current step on each tape, then set the head back to position $0$, start the next phase. Use markers on extra tape portions to represent head positions (fill with $1$s up to head position, then with $0$s).
Exponential closure guarantees that this is possible in "real time", i.e. time $\alpha$. (This probably only requires multiplicative closure.) However, we need extra steps to see whether the halting configuration was assumed. For this, we potentially need to retrieve the head position of the simulated computation, which may take up to $\alpha$ many further steps.
\end{proof}
\begin{defini}
Let $\sigma$ be the minimal ordinal such that $L_{\sigma}\prec_{\Sigma_{1}}L$.
\end{defini}
\begin{prop}{\label{below sigma}}
Every OTM-clockable ordinals is $<\sigma$, and their supremum is $\sigma$.
\end{prop}
\begin{proof}
`The program $P$ halts' is $\Sigma_{1}$, thus, if it halts in $L$, then it halts in $L_{\sigma}$, and thus, the halting time of $P$, if it exists, is $<\sigma$.
On the other hand, every real number in $L_{\sigma}$ is OTM-computable, including codes for all ordinals $<\sigma$, and thus we can write such a code for any ordinal $\alpha<\sigma$ and then run through this code, which takes at least $\alpha$ many steps. Thus, there is an OTM-clockable ordinal above $\alpha$ for every $\alpha<\sigma$.
\end{proof}
\begin{prop}{\label{OTM gap existence}}
There are gaps in the OTM-clockable ordinals. That is, there are ordinals $\alpha<\beta<\gamma$ such that $\alpha$ and $\gamma$ are OTM-clockable, but $\beta$ is not.
\end{prop}
\begin{proof}
This works like the argument in Hamkins and Lewis \cite{HL} for the existence of gaps in the ITTM-clockable ordinals: Take the OTM-program that simultaneously simulates all OTM-programs and halts as soon as it arrives at a level at which no OTM-program halts. If there were no gap, then this program would halt after all OTM-halting times, which is a contradiction.
\end{proof}
\begin{lemma}{\label{very quick writing}}
If an ordinal $\alpha$ is OTM-clockable, then a real number coding $\alpha$ is OTM-writable in $<\alpha^{\prime}$ many steps, where $\alpha^{\prime}$ denotes the next exponentially closed ordinal after $\alpha$.
\end{lemma}
\begin{proof}
If $\alpha$ is clocked by some OTM-program $P$, then $L_{\alpha+1}$ is minimal with the property that it believes that $P$ halts. Thus, there is a $\Sigma_{1}$-statement that becomes true in $L_{\alpha+1}$ for the first time. Hence $\alpha+1$ is an index. Thus, a real number coding $\alpha+1$ is contained in $L_{\alpha+2}$. But the OTM-program that enumerates $L$ will write $L_{\alpha+2}$ in $<\alpha^{+}$ many steps. So just run this program and "clock along", then it will halt when $L_{\alpha+2}$ has been written; then, we can easily find out the desired real code (a real number coding the ordinal height of the predecessor) in the code for $L_{\alpha+2}$.
\end{proof}
\begin{prop}{\label{sigma1 singularization clockable}}
If $\beta<\alpha$ is exponentially closed and OTM-clockable and there is a total $\Sigma_{1}(L_{\alpha})$-function $f:\beta\rightarrow\alpha$ such that $f$ is cofinal in $\alpha$, then $\alpha$ is OTM-clockable.
\end{prop}
\begin{proof}
This works by the same argument as the "only admissibles start gaps"-theorem for ITTMs, see Welch \cite{W}: Suppose for a contradiction that $\alpha$ starts an OTM-gap, but is not admissible.
Pick $\beta<\alpha$ OTM-clockable and $f:\beta\rightarrow\alpha$ such that $f$ is $\Sigma_{1}(L_{\alpha})$ and cofinal in $\alpha$. Let $B$ be an OTM-program that clocks $\beta$.
By the last lemma, we can compute a real code for $\beta$ in $<\beta^{\prime}\leq\alpha$ many steps.
Run the OTM that enumerates $L$. If $\beta$ is exponentially closed, then we will have a code for $L_{\beta}$ on the tape at time $\beta$. In addition, for each new $L$-level, check which ordinals recieve $f$-images when evaluating the definition of $f$ in that level. Determine the largest ordinal $\gamma$ such that $f$ is defined on $\gamma$. Whenever $\gamma$ increases, say from $\gamma_{0}$ to $\gamma_{1}$, let $\delta$ be such that $\gamma_{0}+\delta=\gamma_{1}$ and run $B$ for $\delta$ many steps. When $B$ halts, all elements of $\beta$ have images, so we have arrived at time $\alpha$.
\end{proof}
This suffices to for an OTM-analogue of Welch's theorem \cite{W}:
\begin{corollary}{\label{OTM gap starters admissible}}
If $\alpha$ starts a gap in the OTM-clockable ordinals, then $\alpha$ is admissible.
\end{corollary}
\begin{proof}
As $\alpha$ starts an OTM-gap, it is exponentially closed.
If $\alpha$ is not admissible, there is a total cofinal $\Sigma_{1}(L_{\alpha})$-function $f:\beta\rightarrow\alpha$ with $\beta<\alpha$. Pick $\gamma>\beta$ OTM-clockable and large enough so that all parameters are contained in $L_{\gamma}$. By Lemma \ref{very quick writing}, we can write a real code for $L_{\gamma}$, and thus for all of its elements
in time $<\gamma^{\prime}\leq\alpha$. We can now use Proposition \ref{sigma1 singularization clockable} to clock $\alpha$, a contradiction.
\end{proof}
\section{$\Sigma_{2}$-admissible ordinals are not OTM-clockable}
We now show that no $\Sigma_{2}$-admissible ordinal $\alpha$ can be the halting time of a parameter-free OTM-computation. The proof is mostly an adapatation of Welch's argument to the extra subtleties of OTMs.
\begin{thm}{\label{sigma2 not otm clockable}}
No $\Sigma_{2}$-admissible ordinal is OTM-clockable.
\end{thm}
\begin{proof}
Let $\alpha$ be $\Sigma_{2}$-admissible and assume for a contradiction that $\alpha$ is the halting time of the parameter-free OTM-program $P$. At time $\alpha$, suppose that the read-write-head is at position $\rho$, the program is in state $s\in\omega$ and the head reads the symbol $z\in\{0,1\}$. As one cannot move the head more than $\alpha$ many places to the right in $\alpha$ many steps, we have $\rho\leq\alpha$.
By the limit rules, $z$ must have been the symbol on cell $\rho$ cofinally often before time $\alpha$ and similarly, $s$ must have been the program state cofinally often before time $\alpha$. By recursively building an increasing `interleaving' sequence of ordinals of both kinds, we see that the set $R$ of times at which the program state was $s$ and the symbol on $\rho$ was $z$, we see that $R$ is closed and unbounded in $\alpha$.
We now distinguish three cases.
\bigskip
\textbf{Case 1}: $\rho<\alpha$ and the head position $\rho$ was assumed cofinally often before time $\alpha$.
Let $\beta$ be the order type of the set of times at which $\rho$ was the head position in the computation of $P$. We show that $\beta=\alpha$. If not, then $\beta<\alpha$; let $f:\beta\rightarrow\alpha$ be the function sending each $\iota<\beta$ to the $\iota$th time at which $\rho$ was the head position. Then $f$ is $\Sigma_{1}$ over $L_{\alpha}$ and thus, by admissibility of $\alpha$, $f[\beta]$ is bounded in $\alpha$, contradicting the case assumption.
Let $T$ be the set of times at which $\rho$ was the head position. Then, by the limit rules and the case assumption, $T$ is closed and unbounded in $\alpha$.
As $S$ and $T$ are both $\Sigma_{1}$ over over $L_{\alpha}$ and $\alpha$ is admissible, it follows that $S\cap T$ is also closed and unbounded in $\alpha$. In particular, there is an element $\gamma<\alpha$ in $S\cap T$, i.e. there is a time $<\alpha$ at which the head was on position $\rho$, the cell $\rho$ contained the symbol $z$ and the inner state was $s$. But then, the situation that prompted $P$ to halt at time $\alpha$ was already given at time $\gamma<\alpha$, so $P$ cannot have run up to time $\alpha$, a contradiction.
\bigskip
\textbf{Case 2}: $\rho<\alpha$ and the head position $\rho$ was assumed boundedly often before time $\alpha$.
By the liminf rule for the determination of the head position at time $\alpha$, this implies that, for every $\iota<\rho$, there is a time $\tau_{\iota}<\alpha$ such that, from time $\tau_{\iota}$ on, the head never occupied a position $<\iota$. The function $f:\iota\mapsto\tau_{\iota}$ is $\Pi_{1}$ over $L_{\alpha}$ (we have $f(\iota)=\tau$ if and only if, for all $\beta>\tau$ and all partial $P$-computations of length $\beta$, the head position in the final state of the partial computation was $\geq\iota$) and thus in particular $\Sigma_{2}$ over $L_{\alpha}$. By $\Sigma_{2}$-admissibility of $\alpha$ and the case assumption $\rho<\alpha$, the set $f[\rho]$ must be bounded in $\alpha$, say by $\gamma<\alpha$. But this implies that, after time $\gamma$, all head positions were $\geq\rho$. As $\rho$ was assumed only boundedly often as the head position, this means that, from some time $<\alpha$ on, all head positions were actually $>\rho$. But then, $\rho$ cannot be the inferior limit of the sequence of earlier head positions at time $\alpha$, contradicting the case assumption that the head is on position $\rho$ at time $\alpha$.
\bigskip
\textbf{Case 3}: $\rho=\alpha$.
This implies that the head is on position $\rho$ for the first time at time $\alpha$, so that we must have $z=0$, as there was no chance to write on the $\rho$th cell before time $\alpha$.
Let $S$ be the set of times $<\alpha$ at which some head position was assumed for the first time during the computation of $P$. By the same reason as above, this newly reached cell will contain $0$ at that time. If we can show that there is such a time $<\alpha$ at which the inner state is also $s$, we are done, because that would mean that the halting situation at time $\alpha$ was already given at an earlier time, contradicting the assumption that $P$ halts at time $\alpha$.
As $\rho>0$, there must be an ordinal $\tau<\alpha$ such that the head was never on position $0$ after time $\tau$ (otherwise, the liminf rule would force the head to be on position $0$ at time $\alpha$). This means that the head was never moved to the left from a limit position after time $\tau$. This further implies that, after time $\tau$, for any position $\beta$ that the head occupied, all later positions were at most finitely many positions to the left of $\beta$ and hence that, if $\beta$ is a limit ordinal, then it never occupied a position $<\beta$ afterwards. In particular, the sequence of limit positions that the head occupied after time $\tau$ is increasing. Note that the set of head positions occupied before time $\tau$ is bounded in $\alpha$, say by $\xi$. Let $S^{\prime}$ be the set of elements $\iota>\tau$ of $S$ such that, at time $\iota$, the head occupied a limit position $>\xi$ for the first time. Then $S^{\prime}$ is a closed and unbounded subset of $S$.
As $s$ is the program state at the limit time $\alpha$, there must be $\gamma<\alpha$ such that, after time $\gamma$, the program state was never $<s$ and moreover, the program state $s$ itself must have occured cofinally often in $\alpha$ after that time.
But now, building an increasing $\omega$-sequence of times starting with $\gamma$ that alternately belong to $S^{\prime}$ and have the program state $s$, we see that its limit $\delta$ is $<\alpha$ and is a time at which the head was reading $z$ and the state was $s$, we have the desired contradiction.
\bigskip
Since each case leads to a contradiction, our assumption on $P$ must be false; as $P$ was arbitrary, $\alpha$ is not a parameter-free OTM-halting time.
\end{proof}
\textbf{Remark}: In the second case, we must have that $\rho$ is a limit ordinal bigger than $0$ (successor ordinals and zero cannot come up as liminfs in any other way). If $\rho$ is not of the form $\beta+\omega$, then the argument for case $2$ applies as well: For as $\rho>0$, there must be some time $\tau<\alpha$ after which $0$ was not the head position any more. However, the only way to move the head to the left is to move it to the left from a limit position, which brings it to position $0$. Thus, the sequence of limit position after time $\tau$ was increasing and we can redefine $f$ to send $\omega\iota<\rho$ to the first time $>\tau$ when the head was on position $\omega\iota$; this will send (the set of limit ordinals below) $\rho$ cofinally into $\alpha$ in a $\Sigma_{1}(L_{\alpha})$-definable way, a contradiction. Thus, in this case, only admissibility is required!
In the case $\rho=\beta+\omega$, a final segment of the computation takes place only in this $\omega$-portion between $\beta$ and $\rho$, and only at the end is the head moved out of this portion to position $\rho$. What happens during this time is thus something like an ITTM-computation with the content $x$ of this $\omega$-portion of the tape at the time the head moves into this position for the first time after $\tau$ serving as an oracle.\footnote{"Something like", because of course the head position rule and the limit state rule are not adhered to; however, this can be simulated, but apparently only at the cost of a time delay.}
\section{Existence of admissible OTM-clockable ordinals}
We will now show that at least the first $\omega$ many admissible ordinals are OTM-clockable, thus answering the first question mentioned in the introduction positively.
We recall Theorem $6$ from \cite{CFKMNW}:
\begin{thm}{\label{ITRM no gaps}}
There are no gaps in the ITRM-clockable ordinals. That is, if $\alpha<\beta$ and $\beta$ is ITRM-clockable, then $\alpha$ is ITRM-clockable.
\end{thm}
Combining this result with the main result of \cite{Koe2} on the computational strength of ITRMs, we obtain:
\begin{lemma}{\label{ITRM clockables}}
The ITRM-clockable ordinals are exactly those below $\omega_{\omega}^{\text{CK}}$. In particular, $\omega_{n}^{\text{CK}}$ is ITRM-clockable for all $n\in\omega$.
\end{lemma}
\begin{proof}
By Theorem $6$ of \cite{Koe2}, every ITRM-halting time is $<\omega_{\omega}^{\text{CK}}$. On the other hand, every real number in $L_{\omega_{\omega}^{\text{CK}}}$ is ITRM-computable. Moreover, there is a procedure for checking whether a real number codes a well-founded relation on an ITRM, and it is easy to check that, for a real number coding a well-ordering of length $\alpha$, this procedure takes at least $\alpha$ many steps. Now, given $\alpha<\omega_{\omega}^{\text{CK}}$, pick an ITRM-computable real code $c$ for $\alpha$. Now run the ITRM-program for computing $c$ and run the well-foundedness check on $c$. This will halt after at least $\alpha$ many steps. Thus, there is an ITRM-clockable ordinals $>\alpha$. Consequently, the ITRM-clockable ordinals are unbounded in $\omega_{\omega}^{\text{CK}}$. By Theorem \ref{ITRM no gaps}, every ordinal $<\omega_{\omega}^{\text{CK}}$ is ITRM-clockable. As we mentioned in the beginning, no other ordinals are ITRM-clockable.
\end{proof}
\begin{lemma}{\label{OTM ITRM clock simulation}}
Let $\alpha$ be ITRM-clockable. Then $\alpha$ is OTM-clockable.
\end{lemma}
\begin{proof}
Let $P$ be an ITRM-program that clocks $\alpha$.
The simulation of ITRMs by OTMs here works like this: Use a tape for each register, have $i$ many $1$s, followed by $0$s, on a tape to represent that the respective register contains $i\in\omega$; in addition, after a simulation step is finished, the head position on this tape represents the register content, i.e. it is at the first $0$ on the tape.
For an ITTM, the simulation takes an extra $\omega$ many steps to halt because it takes time to detect an overflow. For an OTM, one can simply use one extra tape for each register, write $1$ to their $\omega$th positions at the start of the computation, move their heads along with the heads on the register simulating tapes and know that there is an overflow as soon as one of the heads on the extra tapes reads a $1$.
\end{proof}
The fact that more tapes are needed the more registers $P$ uses may be seen as a little defect. (Note that, by the results of \cite{Koe2}, the halting times of ITRM-programs using $n$ registers are bounded by $\omega_{n+1}^{\text{CK}}$ so that indeed arbitrarily large numbers of registers - and thus of tapes - are required to make the above construction work for all $\alpha_{n}^{\text{CK}}$ with $n\in\omega$.) It would certainly be nicer to have a uniform bound on the number of required tapes. This is indeed possible:
\begin{corollary}{\label{OTM ITRM subtle clock simulation}}
Let $\alpha$ be ITRM-clockable. Then $\alpha$ is OTM-clockable by an OTM-program that uses three tapes.
\end{corollary}
\begin{proof}
We can reduce this to three tapes: One simulates all ITRM-registers as described above. One represents the program line, but not via the tape content, but via the head position: Head on $i$ means $i$ is the active program line. Automatically works at limits.
ITRM halts depends on active program line and the content of a certain register, namely whether that register contains $0$ or $1$. If the program halts, it is clear on which register this depends; say wlog it is the first. Extra tape on which the content of the first register is represented via the head position. This tape has $0$ everywhere, expect at $0$ and $\omega$, where we have $1$ (write this at the start of the computation). Now, reading a $1$ on this tape means the register contains $0$ (either as a liminf or due to an overflow - that would be the $1$ on position $\omega$), reading $0$ means that the register contains a positive number. Head on this tape is moved along with the register content, when $1$ is read at a limit time (can be determined with a flag on two extra tapes) move it one place to the left, this simulates the reset in the case of an overflow.
\end{proof}
\begin{corollary}{\label{admissible clockables}}
For every $n\in\omega$, $\omega_{n}^{\text{CK}}$ is OTM-clockable.
\end{corollary}
This answers the first question mentioned above in the positive. By a relativization of the above argument, we can achieve the same for the second (i.e. whether gap starters for OTMs are something "better" than admissible):
\begin{thm}{\label{successor admissible do not start gaps}}
Let $\alpha=\beta^{+}$ be a successor admissible. Then $\alpha$ does not start an OTM-clockable gap.
\end{thm}
\begin{proof}
Suppose for a contradiction that $\alpha=\beta^{+}$ starts an OTM-clockable gap. Then there is an OTM-clockable ordinal $\gamma\in(\beta,\alpha)$; pick one. By Lemma \ref{very quick writing} above, a real code $c$ for $\gamma$ is OTM-writable in $<\alpha$ many steps. Suppose $c$ has been written. Then $\omega_{1}^{\text{CK},c}=\alpha$. Thus, $\alpha$ is ITRM-clockable in the oracle $c$. But now, $\alpha$ is OTM-clockable by first writing $c$ and the ITRM-clocking $\alpha$ relative to $c$, a contradiction to the assumption that $\alpha$ starts a gap.
\end{proof}
\begin{corollary}
Every gap-starting ordinal for OTMs is an admissible limit of admissible ordinals.
\end{corollary}
\section{Conclusion and further work}
We showed that OTM-gaps are always started by limits of admissible ordinals and that, while admissible ordinals can be OTM-clockable, $\Sigma_{2}$-admissible ordinals cannot. This provokes the following questions:
\bigskip
\textbf{Question}: Is every gap-starting ordinal for OTMs $\Sigma_{2}$-admissible?\footnote{Note that Welch's argument that ITTM-gaps are always started by admissible ordinals does not seem to help, as it uses the upwards absoluteness of $\Sigma_{1}$, which $\Sigma_{2}$ does not enjoy. In particular, a $\Sigma_{2}$-formula could define different total and cofinal functions over different $L_{\alpha}$s. We conjecture that the techniques used in the construction of the $\Sigma_{2}$-machine by Friedman and Welch in \cite{BIWOC} and generalized in \cite{COW} could be of help here.}
\bigskip
\textbf{Question}: What is the minimal gap-starting ordinal for OTMs? Does it coincide with first $\Sigma_{2}$-admissible ordinal?
\bigskip
Further worthile topics include clockability for OTMs with a fixed ordinal parameter $\alpha$ and for other models of computability, like $\alpha$-ITTMs or $\alpha$-ITRMs (see \cite{COW}) or the "hypermachines" of Friedman and Welch (see \cite{FW}).
|
\section{Summary and conclusions}
We have presented a model for star formation in galaxies that is motivated by the origin of the stability of nuclear fusion in main sequence (MS) stars. The energy generated by nuclear fusion in a MS star equals the rate at which energy is lost through radiation. This equilibrium is secularly stable because if the star loses energy, it heats up, which increases the rate at which fusion occurs. The analogy with a star forming galaxy is that the rate of energy injection by supernovae (and winds from their massive progenitor stars) equals the rate at which energy is lost due to cosmological accretion. This equilibrium is stable provided the star formation rate increases with the pressure of the star forming gas.
Equation~(\ref{eq:SR}), $(1/2)\dot M_\star v_\star^2=(\kappa/2)\omega_b{\hbox{$\dot M_{\rm h}$}}\,{\hbox{$v_{\rm h}$}}^2$, encapsulates this energy balance. Here, $v_\star^2$ is a measure of the effective energy injected per unit mass of star formed by feedback, so that the left hand side is the rate at which feedback increases the galaxy's energy. The right hand side of the equation is the energy loss term due to cosmological accretion ($\omega_b$ is the cosmological baryon to total mass fraction), with ${\hbox{$v_{\rm h}$}}^2$ a measure of the depth of the dark halo's potential. In our \lq $I\kappa\epsilon\alpha$\rq\ model, the star formation rate is set by the cosmological accretion rate by energy balance. The predicted dependence of $\dot M_\star$ on redshift and virial velocity, {\hbox{$v_{\rm h}$}}, or halo mass, {\hbox{$M_{\rm h}$}}, agrees very well with that measured in the {\sc eagle}\ cosmological hydrodynamical simulation \citep{Schaye15}, as shown in Figs. \ref{fig:exp1} and \ref{fig:exp2}, respectively.
The $I\kappa\epsilon\alpha$\ model has four parameters (I, $\kappa$, $\epsilon $, and $\alpha$; hence the name), which together shape the star forming sequence of galaxies. The parameter \lq I\rq\ stands for the (stellar) {\bf I}nitial mass function (IMF), which sets how much energy is available for feedback from star formation, in particular from the supernovae (SN) associated with star formation, as well as the recycled fraction ${\cal R}$ that relates the time integral of star formation to the stellar mass formed. We have kept the IMF constant in this paper. The dimensionless parameters $\kappa$ and $\alpha$ quantify the rate of cosmological accretion onto a halo ($\kappa$), and the concentration of such halos ($\alpha$, see Eq.~\ref{eq:halo}). We find that $\kappa\approx 5/3$ and $\alpha\approx 1$, and have kept these constant as well.
We think that the main numerical parameter that affects our results is $\epsilon$, which is a measure of the fraction of the energy that is injected by SNe that effectively increases the energy of the star forming gas, rather than being radiated away. It relates $v_\star^2$ to the energy produced by SNe per unit mass (or more generally to the energy injected in the ISM as a result of recent star formation), see Eq.~(\ref{eq:sn}). If feedback is efficient, $\epsilon$ is large, and $\dot M_\star$ is small. The {\sc eagle}\ simulation has a parameter, $f_{\rm th}$, that controls what fraction of the available supernova energy is injected into the star forming gas. This means that $f_{\rm th}\approx \epsilon$, provided radiative loses are small. Because feedback {\em is} efficient\footnote{Gas heated by SNe has its temperature increased by $\Delta T=10^7$~K where its cooling rate is minimal and mostly independent of metallicity.} in {\sc eagle}, radiative loses in SN-heated gas are mostly small, which explains why the $I\kappa\epsilon\alpha$\ model reproduces the {\sc eagle}\ model with $f_{\rm th}$ held constant relatively well. However, in the {\sc eagle}\ {\sc reference}\ model, $f_{\rm th}$ is allowed to vary as a function of density and metallicity in a way that is calibrated so that the simulation reproduces (some) observations. Therefore to improve the agreement of $I\kappa\epsilon\alpha$\ with data, we would need to understand how radiative loses depend on the interstellar medium of a star forming galaxy. It seems unlikely that there is a simple way to do so.
A striking feature of the model is that $\dot M_\star$ does not depend on the gas mass, $M_{\rm gas}$, unlike what is assumed in many self-regulating models \citep[e.g.][]{Bouche10, Lilly13}. We use a star formation law (in our case the Kennicutt-Schmidt law, \citealt{Kennicutt98}) to infer $M_{\rm gas}$ from $\dot M_\star$ - rather than the other way around. Doing so allows us to reproduce the $M_{\rm gas}-\dot M_\star$ relation in {\sc eagle}\, (Fig.~\ref{fig:Mgas}) as well as the mass-metallicity relation (Fig.~\ref{fig:Zvh}).
We tried to incorporate feedback from accreting black holes (AGN) by ({\em i}) identifying when feedback from star formation fails so that a black hole can grow, and ({\em ii}) include AGN in the self-regulation process. Stellar feedback fails in galaxies with deep enough potential wells, so that
energy injected by stars cannot compensate for energy lost through accretion even if {\em all} accreted gas is converted into stars. We showed that this occurs in halos with virial velocity above a nearly redshift independent critical value of $\sim 180~{\rm km}~{\rm s}^{-1}$. Demanding that the AGN regulates galaxy formation results in a relation between the black hole mass and the virial velocity of the halo of the form $M_{\rm BH}\propto {\hbox{$v_{\rm h}$}}^5$, which closely follows the observed relation.
In the Introduction we discussed how gas cooling is thought to play an important role in determining the rate at which a galaxy forms stars, to the extent that it may even be the main property that determines the location of the peak in the redshift evolution of the star formation rate density of the Universe \citep{Hernquist03}. Numerical simulations at first sight support this claim directly: a simulation where the contribution from metals is not included when calculating the cooling rate - and hence where the cooling rate is lower - yields lower values of $M_\star/{\hbox{$M_{\rm h}$}}$ than when metals are included (compare models {\sc NOZCOOL} and model {\sc REF} in Fig.3 of \citealt{Haas13a}). However, a lower metallicity of star forming gas reduces cooling loses of injected feedback energy, increasing $\epsilon$ and hence reducing $\dot M_\star$: that sequence of events is also consistent with the findings from \citealt{Haas13a}). The main impact of metallicity on the cooling rate of the gas may be on the efficiency of feedback, rather than on the accretion rate. Of course this argument breaks down in halos where the virial temperature is so high that most of the gas is and remains hot.
We think that $I\kappa\epsilon\alpha$\ provides a simple way of calculating the properties of a galaxy in terms of those of its host halo - and the results so obtained agree reasonably well with those from much more sophisticated models and importantly also with data. We suggest that a better description of how cooling losses depend on the properties of a galaxy through its history would improve the quality of the theoretical prediction.
\section{Discussion}
\subsection{Comparison to previous work}
The paper by \cite{Bouche10} sparked interest in trying to understand the basic physics underlying self-regulation of galaxies. That paper, and several that followed, contain equations that resemble those of Section~3 - but the underlying assumptions are sometimes strikingly different, as we discuss here. The starting point of \cite{Bouche10} is their realisation that the dependence of $\dot M_\star$ on
stellar mass and redshift, resembles that of the cosmological accretion rate, suggesting that the
gas accretion rate $\dot M_{\rm gas, acc}\propto \dot M_{\rm h}$. The proportionality constant is argued
to be less than $\omega_b$, the cosmological gas to total matter density, because only cold accreted gas is assumed to be eligible for star formation. The resulting star formation rate, is then determined by the efficiency with which gas is converted into stars - that is - by the star formation law.
This reasoning results in $\dot M_\star\propto \omega_b\dot M_{h}$, as in our Eq.~(\ref{eq:dotmstar}), with the important distinction that the efficiency of galaxy formation,
\begin{equation}
\label{eq:epsgal}
\epsilon_g\equiv {\dot M_\star\over \omega_b\dot M_{h}}\,,
\end{equation}
is set by the efficiency of star formation,
\begin{equation}
\epsilon_\star\equiv {\dot M_\star\over M_{\rm gas}/\tau_d}\,,
\label{eq:epsg}
\end{equation}
where $\tau_d$ is a characteristic time that still needs to be determined. The onus of getting the observed $M_\star/{\hbox{$M_{\rm h}$}}$ relation is now wholly on the star formation law, Eq.~(\ref{eq:epsg}). The solution advocated by \cite{Bouche10} , is to assume that halos do not form any stars as long as their halo mass is below some minimum value, $M_{h, min}\approx 10^{10}-10^{11}{\rm M}_\odot$, which conspires to result in $\epsilon_g$ increasing with {\hbox{$M_{\rm h}$}}. They stress repeatedly that their results are completely independent of the efficiency of feedback.
\cite{Lilly13} build on this work, and in their \lq gas regulator\rq\ frame work,
$\dot M_\star$ is regulated by the gas reservoir of the galaxy, $M_{\rm gas}$ in our notation. Rather than assuming a minimum halo mass $M_{h, min}$ below which no stars form, the model introduces two main fitting parameters, which in our notation are the product $\epsilon_\star\,\tau_d$ (their variable $\epsilon$) and $\beta$ (their variable $\lambda$). In the follow-up paper by \cite{Birrer14}, they show how the evolution of galaxies over cosmic time can be modelled well once $\epsilon$ and $\lambda$ are parameterised as functions of $M_\star$. Note that these cannot be independent of $M_\star$, since otherwise the ratio $M_\star/{\hbox{$M_{\rm h}$}}$ is constant as well, since a constant fraction of the accreted gas is converted into stars.
The \lq minimum bathtub\rq\ model described by \cite{Dekel14} has very similar ingredients, in that $\dot M_\star$ is also regulated by $M_{\rm gas}$ through the star formation law. These authors stress that many properties of galaxies follow from this model if it is assumed that the system is in a quasi-steady state, $\dot M_{\rm gas}=0$.
These models \lq self-regulate\rq\ in the sense that the star formation rate is determined by the gas mass by mass conservation, in our notation $\dot M_{\rm gas}=\omega_b\dot {\hbox{$M_{\rm h}$}} - (1-{\cal R}+\beta)\dot M_\star$ (Eq.~\ref{eq:MB}}), so that too much star formation depletes the gas reservoir which ultimately decreases $\dot M_\star$. Conversely too little star formation leads to a build-up of $M_{\rm gas}$, and through the star formation law, this increases $\dot M_\star$. A very nice feature of these models, in addition to prediction correctly the rapid increase in $\dot M_\star/M_\star$ with redshift because the gas accretion rate $\propto {\hbox{$\dot M_{\rm h}$}}$, is that they correctly predict secondary parameter dependencies, for example the fact that galaxies that lie above the main sequence are more gas rich and more metal poor, see also \cite{Dayal13}.
What all these models have in common is that the star formation rate is set by the gas reservoir through the star formation law. The origin of that law is not discussed in detail, but presumably it results from a balance between cooling and heating from star formation, as originally envisioned by \cite{White91}.
In these models, feedback from star formation is only important in setting the
star formation law, basically parameterised by $\epsilon_\star$. Combined with a model for the build-up of dark matter halos, or using dark matter-only simulations that follow the growth of halos, these \lq self-regulation\rq\ models are very successful in building realistic looking mock universes \citep[see e.g.][]{Moster18, Tacchella18}.
In our opinion, there are two major weaknesses to this basic model: ({\em i}) to be predictive the model needs to be able to predict how the efficiency of star formation, $\epsilon_\star$, and the mass loading factor, $\beta$, depend on halo (or stellar mass), a formidable task. More worryingly, ({\em ii}) there is evidence that one of the main assumptions - that the star formation rate depends on the gas mass through the star formation law - is not quite correct.
At first sight it seems impossible that the rate of star formation in a galaxy is not dependent on the star formation law - and in fact it would be if the galaxy were isolated. However a galaxy in a cosmological setting can gain mass through accretion and lose it through winds - and therefore the amount of gas in the reservoir is not some constant, rather $M_{\rm gas}$ too is set by the physics of galaxy formation. Demanding that $\dot M_\star$ depends on $M_{\rm gas}$ through a star formation law, and vice versa, results in a \lq chicken and egg\rq\ problem.
Numerical simulations can be very helpful in distinguishing cause from effect. The {\sc owls} simulations described by \cite{Schaye10} are cosmological hydrodynamical simulations performed with {\sc gadget} \citep{Springel05}, but the parameters of sub-grid models are varied over a very wide range and not calibrated to observations as in {\sc eagle}. In particular, the {\sc owls} simulation suite includes parameter variations in which the efficiency of feedback from stars ({\em i.e.} the value of $v_{\rm \star}$ in our notation) and the star formation law (the values of $A$ and $n$ in Eq.~\ref{eq:KS}), are varied separately. By plotting $M_{\rm gas}$ and $\dot M_\star$ versus a variable that does not depend on either $v_\star$ or the star formation law, such as halo mass, {\hbox{$M_{\rm h}$}}, it becomes possible to test the very core assumption of the gas-regulator or bath-tub models.
\cite{Haas13a} compares models with the same star formation law (same value of $A$ and $n$) but different values of the feedback efficiency. Compare in particular their models REF and WML4: these have identical numerical parameters, except that the value of $v_\star^2$ in simulation WML4 is twice that of REF. Maybe not surprisingly, $\dot M_\star/{\hbox{$M_{\rm h}$}}$ in the simulation with the stronger feedback is about half as large as in REF (their Fig.~4). Because the star formation law in these simulations is the same, this also implies that $M_{\rm gas}/{\hbox{$M_{\rm h}$}}$ is also approximately half in WLM4 compared to REF, as is also born out by the same figure.
However now compare models REF and SFAMPLx3 in \cite{Haas13b}: these have identical feedback parameters, but the value of $A$ (from Eq.~\ref{eq:KS}) in simulation SFAMPLx3 is three times that in simulation REF. Figure~5 in \cite{Haas13b} shows that nevertheless the ratio $\dot M_\star/{\hbox{$M_{\rm h}$}}$ is nearly identical in the two simulations: the star formation rate in a halo of given mass is not, or only very weakly dependent, on $A$: a direct violation of the main assumption in the \lq gas-regulator\rq\ models. Given that the star formation rates are the same in these models, but the star formation law differs, this must imply that the gas reservoir in SFAMPLx3 is less than that in REF at a given value of {\hbox{$M_{\rm h}$}}: the same figure~5 shows that indeed $M_{\rm gas}/{\hbox{$M_{\rm h}$}}$ is about a factor of three lower in model SFAMPLx3 compared to REF.
As stressed by \cite{Haas13b} and confirming what was found by \cite{Schaye10}: stellar feedback regulates the star formation rate by determining the amount of (star forming) gas. In this interpretation, {\em $\dot M_\star$ regulates $M_{\rm gas}$ through stellar feedback, rather than $M_{\rm gas}$ setting $\dot M_\star$ through a star formation law}.
The model presented by \cite{Dave12} incorporate self-regulation through feedback, as envisioned here. Because they limit their analysis to equilibrium states defined by $\dot M_{\rm gas}=0$, their results are actually very similar to the various incarnations of the bath-tub models.
In our interpretation, self-regulation follows from energy conservation, Eq.~(\ref{eq:SR}), and in particular the fact that $\dot E_g=0$ is a secularly stable equilibrium (provided that $\dot\rho_\star$ increases with pressure of the star forming gas). Therefore accretion sets the star formation rate, once the net energy input generated by forming stars is known. This sets the \lq efficiency of galaxy formation\rq\ (the ratio of the star formation rate over the cosmological baryon accretion rate onto a halo) to be
\begin{equation}
{\dot M_\star \over {\omega_b \dot M_{\rm h}}} = \kappa {v_{\rm h}^2\over v_\star^2}\,,
\end{equation}
which does not depend on the star formation law but on the properties of the halo (through {\hbox{$v_{\rm h}$}}) and the efficiency of feedback (through $v_\star$)
This is in contrast to Eq.~(\ref{eq:epsg}). The star formation law then determines the gas reservoir in the $I\kappa\epsilon\alpha$\ model, with any excess accreted gas expelled in a wind.
Combining the main $I\kappa\epsilon\alpha$\ relation between the star formation rate and the accretion rate on a halo of Eq.~(\ref{eq:SR}), with the relation between halo virial velocity and halo mass (Eq.~\ref{eq:vh}) and the equation for the growth of a halo (Eq.~\ref{eq:correaM}), allows us to write the star formation rate in terms of the halo accretion rate in the form
The relation between stellar mass and halo mass (Eq.~\ref{eq:mstar}) can be cast in the form
\begin{eqnarray}
\log_{10}\left({M_\star\over 10^{12}{\rm M}_\odot}\right) &=&{5\over 3}\log_{10} \left({M_h\over 10^{12}{\rm M}_\odot}\right) + \log_{10} {\cal N}(z)\nonumber\\
\log_{10} {\cal N}(z) &=&\log_{10}\left({1.7\times 10^{10}{\rm M}_\odot\over 10^{12}{\rm M}_\odot} {1-{\cal R}\over 0.55} {m_\star(z)\over m_h(z)^{5/3}}\right)\,,\nonumber\\
\end{eqnarray}
which has the form of Eq.~(1) in the paper by \cite{Salcido19}, with their $\epsilon(M_h,z)=5/3$. These author show that a halo mass - stellar mass of this form can
be integrated to give analytical relations for the galaxy stellar mass function and the evolution of the cosmic star formation rate density.
\subsection{Limitations of the model}
A forming galaxy can fail to be able to attain its equilibrium star formation rate given by Eq.~(\ref{eq:SR}) for several reasons. Consider for example what happens if ${\hbox{$\dot M_{\rm h}$}}$ suddenly decreases - for example because the galaxy becomes a satellite. Star formation will nevertheless continue in accordance with the star formation law, depleting the gas reservoir. In such galaxies, the star formation rate is set by the gas consumption time scale, rather than regulated by feedback. A less extreme version of the same phenomenon occurs when ${\hbox{$\dot M_{\rm h}$}}$ for a particular halo is unusually small compared to the ensemble average. The $I\kappa\epsilon\alpha$\ model does not correctly describe this situation and in particular is not applicable to satellite galaxies.
We have neglected the finite lifetimes of massive stars. We think this is unlikely to be a major limitation at lower redshifts when the dynamical time of any galaxy is {\em much} larger than the lifetimes of massive stars. However, the limitation may affect the onset of star formation in small galaxies at high redshift. When {\hbox{$v_{\rm h}$}}\ is very low, gas cannot cool and our self-regulation argument will not correctly predict $\dot M_\star$. When the halo grows in mass it may pass the threshold where gas can cool on a short time scale, and star formation may be unable to self-regulate because of the finite lifetimes of massive stars. This may lead to a star burst which $I\kappa\epsilon\alpha$\ does not model correctly.
Not unrelated is what happens at high values of {\hbox{$v_{\rm h}$}}\ at low redshift. The $I\kappa\epsilon\alpha$\ model predicts that feedback becomes inefficient for
${\hbox{$v_{\rm h}$}}\approx 180$~km~s$^{-1}$ following similar reasoning to \cite{Bower17}. We argued, as did \cite{Bower17}, that the resulting increase in gas mass triggers the AGN, which, once the black hole mass has increased sufficiently, will regulate the galaxy. However, by construction this occurs in the same halos that develop a hot halo of gas, so that it becomes unlikely that the right hand side of Eq.~(\ref{eq:SR}) describes correctly the rate at which gas enters the galaxy: it may simply add to the hot halo instead (see the discussion in \citet{Bouche10} on hot versus cold accretion). We think therefore that it is unlikely that $I\kappa\epsilon\alpha$\ models such galaxies accurately. Moreover, galaxy-galaxy mergers contribute significantly to the mass growth of such galaxies, and we have not attempted to include these in the model either.
We also neglected that stars may form from gas lost by previous generations of stars - such recycling may affect the star formation rate of galaxies at late times
when their stellar masses are high but the cosmological accretion rate low \citep[e.g.][]{Oppenheimer10, Voort17}. Gas lost from galaxies by winds may re-accrete later - again we have neglected this effect. More in general, we have neglected the possibility that the accretion rate differs from $\omega_b{\hbox{$\dot M_{\rm h}$}}$.
If Eq.~(\ref{eq:SR}) is indeed applicable, then it might be possible to estimate the scatter around the main sequence of star forming galaxies from the scatter of {\hbox{$\dot M_{\rm h}$}}\ around the ensemble average. This would provide a good test of the basic assumption in our model.
\section{Simulation details}
We compare the results of the model described in this paper to galaxies from the {\sc eagle}\ simulation, which we briefly describe here. \lq Evolution and Assembly of Galaxies and their Environments\rq\ ({\sc eagle}) is a suite of cosmological, hydrodynamical simulations, performed using an evolution of the {\sc gadget} smoothed particle hydrodynamics code described by \cite{Springel05}. {\sc eagle}\ uses a set of sub-grid modules to encode unresolved physics, described in detail by \cite{Schaye15}, which we briefly summarise here.
The sub-grid modules contain a set of numerical parameters, whose values are calibrated in a {\sc reference} run to reproduce a small number of $z\approx 0$ observables, namely the galaxy stellar mass function, the relation between galaxy stellar mass, $M_\star$, and size, and between $M_\star$ and black hole mass, as detailed by \cite{Crain15}. Given these calibrated values, the simulation also reproduces several observable relations that were not part of the calibration, in particular yielding a \lq main sequence\rq\ of blue star forming galaxies in which $\dot M_\star$ depends on $M_\star$ and redshift as observed \citep{Furlong15}, as well as a \lq red sequence\rq\ of quenched galaxies \citep{Trayford15,Trayford17}. The $z=0$ galaxy colours correlate with galaxy morphology as observed \citep{Correa17,Trayford18}.
Most relevant for the comparisons in this paper are the implementation of star formation, of stellar feedback, and of feedback from accretion blackholes (AGN) in {\sc eagle}:
\begin{itemize}
\item {\em Star formation:} sufficiently dense gas in {\sc eagle}\ is converted into star particles at a rate per unit volume, $\dot\rho_\star$, that depends on the gas pressure, $P$, as
\begin{equation}
\label{eq:dotmstar1}
\dot\rho_\star\propto P^{(n-1)/2}\,.
\end{equation}
The normalisation of this relation and the exponent $n$ are set by the Kennicutt-Schmidt law \citep{Kennicutt98} that relates the surface density of star formation, $\dot\Sigma_\star$ and of gas, $\Sigma_{\rm g}$,
\begin{equation}
\label{eq:KS}
\dot\Sigma_\star = A \left({\Sigma_g\over 1~{\rm M}_\odot\,{\rm pc}^{-2}}\right)^{n}\,.
\end{equation}
The underlying assumption connecting these relations is that volume and surface densities are related by the local Jeans length, as motivated by \cite{Schaye08}. The simulation does not resolve the multi-phase nature of the interstellar medium and
star forming gas is assumed to have a minimum pressure \citep{Schaye15},
\begin{equation}
\label{eq:eos}
P \propto \rho^{4/3} \propto u^4\,,
\end{equation}
where $u$ is the thermal energy per unit mass.
\item {\em Stellar feedback} is implemented as described by \citet{DallaVecchia12}: a newly formed star particle increases the temperature of surrounding gas by an amount $\Delta T$. The quantity of gas heated depends on the effective energy injected by star formation, $f_{\rm th}\Delta E$, where $\Delta E$ is the total energy released by the winds from massive stars and core-collapse supernovae which in turn depends on the assumed stellar initial mass function (IMF). The value of $\Delta T$ is chosen such that gas is heated to a temperature where its cooling rate is small: this makes the feedback efficient. The value of $1-f_{\rm th}$ quantifies the fraction of injected energy that is lost from the star forming region, for example through radiative cooling; $f_{\rm th}$ is one of the main calibration parameters in {\sc eagle}.
\item {\em Black holes and AGN:} The seeding, merging, accretion, and feedback from black holes (BHs) in {\sc eagle}\ is described by \cite{Rosas15}. Seed BHs are inserted in each dark matter halo once it becomes sufficiently well resolved. When a BH accretes mass and becomes an AGN, it injects thermal energy in the surrounding gas.
\end{itemize}
\begin{table}
\begin{center}
\caption{Selected parameters of the {\sc eagle}\ simulations used here. From left-to-right the columns show: simulation name; co-moving box size; initial baryonic particle mass; maximum proper softening length, and comment.}
\label{tbl:sims}
\begin{tabular}{lrrrrrrr}
\hline
Name & $L$ & $m_{\rm g}$ & $\epsilon_{\rm prop}$ & comment\\
& (Mpc) & ($10^6 {\rm M}_\odot$) & (kpc)\\
\hline
REF & 50 & $1.81$ & 0.7 & reference model \\
FBconst & 50 & $1.81$ & 0.7 & {\hbox{$f_{\rm th}=1$}}\\
FBconstnoAGN & 50 & $1.81$ & 0.7 & {\hbox{$f_{\rm th}=1$, no AGN}}\\
DMO & 50 & 0 & 0.7 & dark matter only\\
\hline
\end{tabular}
\end{center}
\end{table}
The origin of red galaxies in {\sc eagle}\ is investigated by \cite{Trayford16}. Ram-pressure stripping and \lq strangulation\rq\ dramatically decreases the star formation rate of satellite galaxies, causing them to leave the blue cloud of star forming galaxies and settle onto the red sequence. The simple self-regulating model described in this paper does not attempt to describe these effects, and we will therefore only compare to central, {\em i.e.} non-satellite, {\sc eagle}\ galaxies. Similarly, AGN feedback suppresses star formation in massive galaxies, causing them to become passive. Since that mechanism is also not included in the model, most of the comparison in this paper i
to {\sc eagle}\ variation FbConstNoAGN, in which $f_{\rm th}$ is a constant, and which does not include AGN feedback.
We also use variation FbConst, in which $f_{\rm th}$ is kept constant and which {\em does} include AGN.
Table~\ref{tbl:sims} contains a list of parameters of the {\sc eagle}\ runs that we used. Simulation \lq REF\rq\ is the default {\sc eagle}\ model from Table~2 of \cite{Schaye15}. The simulation FBconst with $f_{\rm th}=1$ appears in Table~1 of \cite{Crain15},
simulation. Simulation DMO is a dark matter-only version of the same volume. All simulations are initialised from the same Gaussian initial conditions, so that halo masses are nearly identical in all runs.
Galaxies of the {\sc eagle}\ reference model look like observed galaxies in many of their properties. Keeping $f_{\rm th}$ constant, the simulated galaxies have similar stellar masses and star formation rates, but are typically smaller than in the reference model. Therefore this model is not as good a representation of the real galaxy population, but we believe its physics is still reasonable - and it is much easier to compare to our simple model. Many of the properties of the population of {\sc eagle}\ galaxies can be extracted directly from the public database\footnote{\url{http://icc.dur.ac.uk/Eagle/database.php}.} \citep{McAlpine16}, which we used extensively in preparing the figures.
\section{Introduction}
The cold dark matter cosmogony links the small fluctuations detected in the cosmic microwave background ({\sc cmb}) at redshift $z\sim 1000$ to the observed large-scale clustering of galaxies at all observable redshifts. The fluctuations in the {\sc cmb} temperature correspond to density perturbations that grow in amplitude due to gravity, resulting in the formation of dark matter halos that host galaxies (see e.g. \citealt{Springel05a} and reference therein for more background).
Whereas computer simulations can reliably predict virtually all properties of dark halos, the same can not be said for the properties of the galaxies that inhabit these halos. Even though our basic understanding of the underlying physics is probably correct - galaxies form as gas accretes onto a halo, cools, becomes self-gravitating and forms stars \citep{White78, White91} - numerous uncertainties remain. What sets the star formation rate of a galaxy in a given halo at a given redshift? How does the energetic feedback from stars and accreting black holes regulate star formation? What is the role of galaxy interactions such as mergers? Are there any other crucial processes, for example feedback from cosmic rays or reionisation, and what is the role of magnetic fields?
Models that are designed to reproduce a mock universe that looks and evolves like the one we observe may not care about the details of the relevant physical processes. Examples include halo occupation distribution models ({\sc hod}, e.g. \citealt{Peacock00}) or subhalo abundance matching ({\sc sham}, e.g. \cite{Vale04}, see e.g. \cite{Wechsler18} for recent reviews).
Semi-analytical models recognise that the physics of galaxy formation is complex, and use parametrizations to model poorly understood physical processes. Cosmological hydrodynamical simulations try to capture some of these physical processes as accurately as possible (cosmological accretion and cooling of gas onto halos for example), but also rely on more parametrised descriptions of physical processes to capture physics below the resolution scale (see \citealt{Somerville15} and \citealt{Naab17} for recent reviews).
Several of the semi-analytical models and recent hydrodynamical simulations yield mock universe that look impressively similar to the one observed. Even though these models typically all include the same ingredients, the details of how the processes are implemented may be quite different. It is therefore somewhat surprising that the resulting galaxy population is nevertheless very similar. At the very least this suggests some level of degeneracy in the modelling and that such calculations cannot be used to infer how the unresolved processes operate in detail. But it also suggests that many properties of galaxies do not actually depend on the details of many of these processes (see \citealt{Hopkins14} for a similar point of view).
Arguably one of the more striking features of the galaxy population as a whole is the emergence of a \lq star forming main sequence\rq\ (or \lq blue cloud\rq), \cite{Noeske07}, on which galaxies form stars at a specific rate, $\dot M_\star/M_\star$, that depends weakly on stellar mass ($M_\star$), but increases rapidly with redshift. The scatter around the mean trend is small, of order 0.3~dex (see \citealt{Schreiber15} and references therein for more recent observational analysis and discussion).
The appearance of such a main sequence suggests that the rate at which a galaxy forms stars in a halo of given mass, is somehow self-regulating. Several papers argued just that \citep[e.g.][]{Bouche10, Lilly13, Dekel14, Dave12, Dayal13}. The aim of these models is not to be able to predict the properties of galaxies in great detail, but rather understand the origin of self-regulation. The current paper follows this philosophy, adopting simplifications to more clearly expose the feedback loop that operates on the star forming sequence.
This paper is organised as follows: section~2 exposes the basic physics behind self-regulation in our model and tests the central assumptions by comparing to galaxies from the {\sc eagle}\ cosmological hydrodynamical simulation \citep{Schaye15}. Section~3 explores consequences in terms of galaxy scaling relations (such as the galaxy stellar mass function and the mass-metallicity relation), compares these to simulations and data, and discusses successes and failures of the model. Section~4 puts our results into context by comparing to previous work, and discusses what we think are its main limitations. Section~5 summarises our findings and is followed by an appendix that contains a short overview of the {\sc eagle}\ simulations, including a description of the reference model, \lq Ref-L100N1504\rq, in which the subgrid parameters are calibrated to reproduce redshift $z=0$ observations of the galaxy stellar mass function, the relation between galaxy size and mass, the relation between black hole mass and stellar mass, as described by \cite{Crain15}. The appendix also describes the {\sc eagle}\ model \lq FbConstNoAGN\rq, in which the feedback parameters are kept constant and which does not include feedback from AGN, as well as another {\sc eagle}\ variation, \lq FbConst\rq, in which the feedback parameters are kept constant and which does include AGN feedback.
\section{Galactic winds and the failure of self-regulating stellar feedback}
\subsection{Galaxy sizes and gas fractions}
\label{sect:outflows}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_KSlaw_Mo.png}
\caption{The mass of star forming gas in galaxies, $M_{\rm gas}$, versus the star formation rate, $\dot M_\star$, at different redshifts: $z=0$ (blue), $1$ (green) and $3$ (red). The {\em thick solid lines} show the scaling in $I\kappa\epsilon\alpha$\ obtained by using the Kennicut-Schmidt law (Eq.~\ref{eq:KSlaw}), with $n=1.4$ and assuming that the scale-length of the gas disk evolves as in Eq.~\ref{eq:rgas} \citep{Mo98}. {\em Large coloured dots} are the median relation in the {\sc eagle}\ galaxies (simulation FbconstnoAGN), with the {\em shaded regions} encompassing the $25^{\rm th}-75^{\rm th}$ percentile range. {\em Dotted horizontal line} correspond to {\sc eagle}\ galaxies with 100 gas particles. The galaxies for which $R_{\rm gas}$ is less than the gravitational softening length in {\sc eagle}\ lie below
{\rm triangles} on each line, indicating that those galaxies are not well resolved in the {\sc eagle}\ simulation. In the $I\kappa\epsilon\alpha$\ model, the amount of gas in the interstellar medium is set by the star formation rate rather than the other way around.}
\label{fig:Mgas}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_outflow_2.pdf}
\caption{The dependence of the wind mass loading factor $\beta$ ({\em top panels}, from Eq.~\ref{eq:MB}) and the wind-speed, $v_{\rm w}$, at 5~times the wind launching radius ({\em bottom panels}, from Eq.~\ref{eq:vw}) as a function of halo mass, {\hbox{$M_{\rm h}$}} ({\em left panels}), and halo virial velocity, {\hbox{$v_{\rm h}$}} ({\em right panels}), from the $I\kappa\epsilon\alpha$\ model. The launching radius of the wind is taken to be equal to the gas scale radius, $R_{\rm gas}$ (Eq.\ref{eq:rgas}). Coloured lines correspond to $\epsilon=0.1$, $\kappa=5/3$, $\alpha=1$, $1-{\cal R}=0.55$ in $I\kappa\epsilon\alpha$\ model, at different redshifts ({\em blue, green, red and purple} corresponding to $z=0$, 1, 3, and 6, respectively). The redshift dependence is stronger as a function of {\hbox{$M_{\rm h}$}}\ than as a function of {\hbox{$v_{\rm h}$}}. At low values of ${\hbox{$v_{\rm h}$}}\lessapprox 120$~km~s$^{-1}$, the wind speed tracks {\hbox{$v_{\rm h}$}}, and the mass loading decreases from $\beta\sim 30$ at ${\hbox{$v_{\rm h}$}}\sim 50$~km~s$^{-1}$ to $\beta\sim 1$ at ${\hbox{$v_{\rm h}$}}\sim 100$~km~s$^{-1}$. The outflow begins to stall, $\beta\rightarrow 0$, for ${\hbox{$v_{\rm h}$}}\rightarrow 180$~km~s$^{-1}$, at which point the wind speed becomes large, $\sim 10^3$~km~s$^{-1}$. The {\em black dashed line} in the bottom right panel is the one-to-one relation. The {\em thin black curve} labelled \lq MW\rq\ shows the evolution of $\beta$ and $v_{\rm w}$ for a Milky Way-like galaxy, $z=0$ halo mass of $M_{\rm h,0}=10^{12}{\rm M}_\odot$, as it grows in mass.}
\label{fig:beta}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_beta_map.pdf}
\caption{The mass loading factor $\beta$ from Eq.~(\ref{eq:MB2}) as a function of the halo virial velocity, {\hbox{$v_{\rm h}$}}, and redshift, $z$, shown as a color-map, for $\dot M_{\rm gas}=0$ ({\it top panel}) and for $\dot M_{\rm gas} \neq 0$ ({\it bottom panel}). The dashed and solid black lines correspond to $\beta=0.5$ and $1$, respectively. }
\label{fig:betamap}
\end{figure}
The star formation rate in the $I\kappa\epsilon\alpha$\ model does not depend on the gas mass. Instead, we {\em compute} the mass of star forming gas by assuming a star formation law. Taking the Kennicutt-Schmidt \citep{Kennicutt98} star formation law and assuming that star forming gas is in an exponential disk with scale-length $R_{\rm gas}$, the (total) star formation rate of a galaxy is related to its gas mass by
\begin{equation}
\dot M_\star = {2\pi A\,R_{\rm gas}^2\over n^2}\left({M_{\rm gas}/{\rm M}_\odot\over 2\pi (R_{\rm gas}/{\rm pc})^2}\right)^n\,,
\label{eq:KSlaw}
\end{equation}
where $A$ and $n$ are the parameters of the Kennicutt-Schmidt law (Eq.~\ref{eq:KS}). We follow \cite{Mo98}
(see also \citealt{Kravtsov13}) by assuming\footnote{\cite{Mo98} apply this reasoning to the stellar disk; \cite{Navarro17} show that $R_{\rm gas}$ scales better with the scale radius of the halo, but since we neglect variations in halo concentration by taking $\alpha=1$, these are equivalent.} that disk size, $R_{\rm gas}$, is a constant fraction, $\lambda$, of the halo's virial radius. Using Eq.~(\ref{eq:halo}), this yields
\begin{eqnarray}
R_{\rm gas}&=&\lambda\,R_{\rm h}\, = R_{\rm gas, 0}\,r_{\rm gas}(z)\nonumber\\
R_{\rm gas, 0} &=& 2~{\rm kpc}\,{\lambda\over 0.01}\, \left({M_{\rm h,0}\over 10^{12}{\rm M}_\odot}\right)^{1/3}\nonumber\\
r_{\rm gas}(z) &=& {m_{\rm h}(z)^{1/3}\over \mathcal{H}(z)^{2/3}}\,,
\label{eq:rgas}
\end{eqnarray}
where $\lambda=0.01$ yields a reasonable reference scale-length of $R_{\rm gas, 0}=2$~kpc for the galaxy inhabiting a $10^{12}{\rm M}_\odot$ halo at $z=0$. Using the $\dot M_\star-M_{\rm h,0}$ relation from Eq.~(\ref{eq:dotmstar}) and the $M_\star-M_{\rm h,0}$ relation from Eq.~(\ref{eq:mstar}) allows us to relate galaxy size to star formation rate and stellar mass,
\begin{eqnarray}
{R_{\rm gas}(z)\over 2~{\rm kpc}} {0.01\over\lambda}&=&\left({\dot M_\star\over 1.2~{\rm M}_\odot~{\rm yr}^{-1}}\right)^{1/5}{r_{\rm gas}(z)\over \Psi_\star(z)^{1/5}}\nonumber\\
&=&\left({M_\star\over 1.7\times 10^{10}~{\rm M}_\odot}\right)^{1/5}{r_{\rm gas}(z)\over m_\star(z)^{1/5}}\,.\nonumber\\
\label{eq:mgms}
\end{eqnarray}
Sizes of a galaxies with a given $M_\star$ depend on redshift $\propto m_{\rm h}^{1/3}/(\mathcal{H}^{2/3}m_\star^{1/5})$. The ratio $m_{\rm h}^{1/3}/m_\star^{1/5}$ varies by less than a factor 0.75 below $z=6$, meaning that the size scales approximately as $1/\mathcal{H}(z)^{2/3}=1/(1+z)$ for $z\gg 1$, and slower than that at lower $z$. This agrees rather well with the observed scaling: \cite{Allen17} quotes a scaling $\propto (1+z)^{-0.97}$ for redshifts 5-7 and \cite{Vanderwel14} quotes a scaling $(1+z)^{-0.75}$
for redshifts 0-3. At a given value of $\dot M_\star$, $R_{\rm gas}\propto (1+z)^{-1.7}$ for $z\gg 1$, which is steeper than the $(1+z)^{-1.1}$ quoted by \cite{Shibuya15}. The weak dependence of size on mass, $R_{\rm gas}\propto M_\star^{1/5}$, is consistent
the scaling $R_\star\propto M_\star^{0.22}$ for the stellar size-$M_\star$ relation found by \cite{Vanderwel14}.
Substituting Eq.~(\ref{eq:rgas}) into Eq.~(\ref{eq:KSlaw}) yields
\begin{eqnarray}
M_{\rm gas} &=& M_{\rm gas,0}~m_{\rm gas}(z)\nonumber\\
\left[{M_{\rm gas,0}\over 2.45\times 10^9~{\rm M}_\odot}\right]^n &=& {\kappa\over 5/3}{\alpha\over 1}\left[{h \over 0.677}\right]^{5/3}\,\left[{\lambda\over 0.01}\right]^{2n-2}\nonumber\\
&\times &\left[{v_\star\over 400~{\rm km~s}^{-1}}\right]^{-2}\left[{M_{\rm h,0}\over 10^{12}~{\rm M}_\odot}\right]^{1+2n/3}\nonumber\\
m_{\rm gas}(z) &=& [{\hbox{$\Psi_\star$}}(z)]^{1/n}\,[r_{\rm gas}(z)]^{2-2/n}\,.
\label{eq:Mgas}
\end{eqnarray}
Using the $\dot M_\star-M_{\rm h,0}$ relation from Eq.~(\ref{eq:dotmstar}) allows us to compute the $M_{\rm gas}-\dot M_\star$ relation, and the result is compared to the {\sc eagle}\ simulation in Fig.~\ref{fig:Mgas}, where we used the values of $A$ and $n$ from \cite{Kennicutt98}. The $I\kappa\epsilon\alpha$\ prediction reproduces very well the slope of the relation and the normalisation at $z=0$. The simulated evolution is somewhat weaker than predicted. Although pleasing, the excellent agreement between the theoretical prediction and the simulation is not surprising: galaxies in {\sc eagle}\ follow the Kennicutt-Schmidt relation of Eq.~(\ref{eq:KS}), and that relation results in galaxies following Eq.~(\ref{eq:KSlaw}) at least approximately.
The evolution of the gas mass, as governed by the dimensionless function $m_{\rm gas}(z)$ from Eq.~(\ref{eq:Mgas}), is plotted in Fig.~\ref{fig:results}. The ratio of $m_{\rm gas}(z)$ over its value at $z=0$, is 0.5, 0.8, 1.2 and 1.4 at $z=4$, 3, 2 and 1, respectively, meaning that the gas mass of a forming galaxy changes by slightly more than a factor of 2 since $z=4$. Therefore assuming that galaxies form stars at nearly constant gas mass is a relatively good approximation below $z\sim 4$; it forms the basis of the equilibrium model of \cite{Dave12}, see also \cite{Bouche10,Krumholz12}.
\subsection{Galactic winds}
Galactic winds are a natural outcome of a model in which cosmological accretion sets the star formation rate but a star formation law sets the gas mass. Indeed, conservation of baryon mass requires that
\begin{equation}
\omega_{\rm b}{\hbox{$\dot M_{\rm h}$}} = \dot M_{\rm gas} +(1-{\cal R})\dot M_\star +\dot M_{\rm w} \equiv \dot M_{\rm gas} + (1+\beta-{\cal R})\,\dot M_\star\,,
\label{eq:MB}
\end{equation}
where $\dot M_{\rm w}$ is the rate at which the galaxy loses mass through a galactic wind, and the ratio $\beta\equiv \dot M_{\rm w}/\dot M_\star$ is usually called the \lq mass loading factor\rq\ of the wind. Solving for $\beta$ gives
\begin{equation}
\beta = {v_\star^2\over \kappa{\hbox{$v_{\rm h}$}}^2} - (1-{\cal R}) - {\dot M_{\rm gas}\over \dot M_\star}\,,
\label{eq:MB2}
\end{equation}
where we used the main $I\kappa\epsilon\alpha$\ relation of Eq.~(\ref{eq:SR}) to relate $\dot M_\star$ and $\dot{\hbox{$M_{\rm h}$}}$. We can compute $\beta$ as a function of redshift and {\hbox{$v_{\rm h}$}}\ or halo mass by integrating this equation using the relation between the gas mass and $\dot M_\star$ (Eq.~\ref{eq:KSlaw}), the result is shown in Figs.~\ref{fig:beta} and \ref{fig:betamap}. The velocity of the outflow can be estimated by assuming that the wind conserves energy once launched\footnote{This assumption may not be unreasonable because $\epsilon$ already accounts significant radiative losses before the wind is launched.},
\begin{eqnarray}
{1\over 2}\dot M_\star\,v_{\star}^2 = {1\over 2}\dot M_{\rm w}\,\left(1+{2\over {\cal M}^2\gamma(\gamma-1)}\right)v_{\rm w}^2 + {1\over 2}\dot M_{\rm w} v_\phi^2 \nonumber\\
{1\over 2}v_\phi^2 = \frac{c v_{\rm h}^2}{\ln(1+c) - {c \over 1+c}} \left[ {\ln(1+{c R_{\rm L} \over R_{\rm h}}) \over {c R_{\rm L}\over R_{\rm h}}} - {\ln(1+ {c R\over R_{\rm h}}) \over {c R\over R_{\rm h}}} \right]\,.
\label{eq:vphi}
\end{eqnarray}
${\cal M}=v_{\rm w}/c_s$ is the wind's Mach number, $\gamma=5/3$ is the adiabatic index, $v_\phi^2/2$ is the change in potential of a Navarro-Frenk-White (NFW) halo \citep{Navarro97} between the launch cite, $R_L$, and the location $R$ where it is measured \citep[e.g.][]{Lokas01}, $c$ is the halo's concentration parameter which depends on $M_{\rm h}$ and $z$ \citep[e.g.][]{Ludlow14} and we assume the launch radius, $R_{\rm L}=R_{\rm gas}$. $v_\phi\approx 0$ if the wind speed is measured very close to the launch site, and $v_\phi$ equals the escape speed from the halo if the wind speed is measured at infinity. This expression also neglects any ram-pressure the outflow may suffer. If the outflow is cold, ${\cal M}\rightarrow \infty$, and
\begin{equation}
\label{eq:vw}
v_{\rm w}^2 \approx \frac{v_{\star}^2}{\beta} - v_\phi^2\,.
\end{equation}
Clearly this treatment of the wind is quite approximate and in particular it is not obvious how one should compare our value of $v_{\rm w}$ to observations, in which the wind speed is often expressed in terms of the full width half maximum of an emission line. Fortunately, the behaviour of the mass loading $\beta$ is independent of these considerations, although here it is not so clear whether $\beta$ refers to gas leaving the galaxy or gas leaving the halo.
Given these limitations, we plot $\beta$, and the wind speed, $v_{\rm w}$, at a distance of 5~times the gas scale radius, $R_{\rm gas}$, as a function of halo mass, virial velocity, and redshift in Fig.\ref{fig:beta}. The $\beta-{\hbox{$M_{\rm h}$}}$ relation evolves with redshift, as is clear from the left panels of the figure, basically because the relation between $\dot M_\star$ and {\hbox{$\dot M_{\rm h}$}}\ depends on virial velocity according to Eq.~(\ref{eq:SR}). Most of that redshift dependence is removed if we plot $\beta$ as a function of {\hbox{$v_{\rm h}$}}, as is seen from the right panels in the figure. As {\hbox{$v_{\rm h}$}}\ increases, $\beta$ decreases and $v_{\rm w}$ increases. Also notice that as ${\hbox{$v_{\rm h}$}}$ tends to a critical value of around $v_{\rm h,c}\approx 180$~km~s$^{-1}$, $\beta$ drops precipitously whereas the wind speed {\em increases} rapidly.
Winds in low {\hbox{$v_{\rm h}$}}\ galaxies are slow and strongly mass-loaded, $\beta\gg 1$, as can be seen from Fig.~\ref{fig:beta}. When $\beta\gg 1$ and making the further approximation that $|\dot M_{\rm gas}|\ll \dot M_\star$, Eqs~(\ref{eq:MB}) and (\ref{eq:vw}) combine to
\begin{equation}
v_{\rm w} = \left({\kappa (1+\beta-{\cal R})\over \beta}\right)^{1/2}{\hbox{$v_{\rm h}$}}\approx \kappa^{1/2}\,{\hbox{$v_{\rm h}$}}\,.
\end{equation}
Therefore, the wind speed tracks the halo's virial velocity (in low {\hbox{$v_{\rm h}$}}\ galaxies at $z<4$), as is apparent from Fig.~\ref{fig:beta}.
The relation between gas mass and star formation rate that results from the Kennicutt-Schmidt star formation law, Eq.~(\ref{eq:Mgas}), and the
equation for the mass-loading of winds, Eq.~(\ref{eq:MB2}), have interesting consequences, namely ({\em i}) the emergence of a mass-metallicity relation, and ({\em ii}) the existence of a characteristic value of {\hbox{$v_{\rm h}$}}\ above which self-regulation due to feedback from stars fails. We investigate these next.
\subsection{The mass-metallicity relation}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_correa_single_Z_SF.png}
\caption{Metallicity, $Z$, of the star forming gas as function of the halo's circular velocity, {\hbox{$v_{\rm h}$}}, at various redshift. Large solid dots
are the median relation from {\sc eagle}\ (simulation FbconstnoAGN), with the shaded area encompassing the 25-75$^{\rm th}$ percentile for redshifts $z=0$ (blue), 1 (green) and 3 (red). Only halos with at least $10^3$ gas particles are shown. Black lines correspond to the $I\kappa\epsilon\alpha$\ model, from Eq.~(\ref{eq:ZMS}), with $\epsilon=0.05$ and 0.1 shown as a {\em solid} and {\em dashed line}, respectively; the redshift dependence of these lines is negligible, and the results depend very little on the assumed initial metallicity. The dependence of $Z$ on {\hbox{$v_{\rm h}$}}\ is slightly shallower in {\sc eagle}\ compared to $I\kappa\epsilon\alpha$. The $Z-{\hbox{$v_{\rm h}$}}$ relation is almost independent of redshift in both the model and {\sc eagle}.}
\label{fig:Zvh}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_Z_z1.pdf}
\caption{Metallicity, $Z$ in solar units, as a function of redshift, $z$, for halos with different mass; ({\em blue}, {\em green}, {\em red} correspond to $z=0$ halo masses of $\log_{10}M_{\rm h,0}/{\rm M}_\odot$=10, 11 and 12, respectively). The result from integrating Eq.~(\ref{eq:Zdot}) numerically is shown as {\em solid lines}, the approximation $Z=\kappa y{\hbox{$v_{\rm h}$}}^2/v_\star^2$ is shown as {\em dashed lines}. Results are shown taking $v_\star=400$~km~s$^{-1}$, $y = 0.04$ and ${\rm Z_\odot}=0.0127$.}
\label{fig:Zatt}
\end{figure}
The metal mass of the star forming gas, $M_Z\equiv Z\,M_{\rm gas}$, changes due to metals synthesised and released by stars, metals accreted, metals lost in a galactic wind, and metals locked-up in long-lived stars.
Its rate of change is therefore
\begin{eqnarray}
\dot M_Z &=& {d\over dt}(Z M_{\rm gas})\nonumber\\
&=& y \dot M_\star + Z_0\omega_{\rm b}{\hbox{$\dot M_{\rm h}$}} - Z_{\rm w}\dot M_{\rm w} - Z (1-{\cal R})
\dot M_\star\,,\nonumber\\
\label{eq:ZB}
\end{eqnarray}
where $y$ is the stellar yield, $Z_0$ is the metallicity of accreted gas and $Z_{\rm w}$ is the metallicity of
the wind which may differ from that of the gas, for example because enriched gas is more like to be ejected by feedback \citep[see e.g.][]{Creasey15}.
Combining this relation with Eq.~(\ref{eq:MB}), which expresses baryon mass conservation, and the main $I\kappa\epsilon\alpha$\ relation between ${\hbox{$\dot M_{\rm h}$}}$ and $\dot M_\star$ from Eq.~(\ref{eq:SR}), we find that provided $Z_{\rm w}=Z$ and $Z_0 = 0$,
\begin{eqnarray}
\dot Z &=& y\frac{\dot M_\star}{M_{\rm gas}} - Z \frac{\omega_{\rm b} \dot M_{\rm h}}{M_{\rm gas}}\,, \nonumber\\
&=& \frac{\dot M_\star}{M_{\rm gas}} \left( y - Z \frac{v_\star^2}{\kappa v_{\rm h}^2} \right)\,.
\label{eq:Zdot}
\end{eqnarray}
The recycled fraction ${\cal R}$ does not affect $\dot Z$ in the instantaneous recycling approximation, and the wind's mass loading $\beta$ does not affect $\dot Z$ provided $Z_{\rm w}=Z$. Integrating this equation in time,
we compare the relation between $Z$ and {\hbox{$v_{\rm h}$}}\ as a function of redshift to the results from
{\sc eagle}\ (simulation FbConstnoAGN) in Fig.~\ref{fig:Zvh}; the agreement is quite good, with $I\kappa\epsilon\alpha$\ showing a somewhat steeper dependence of $Z$ on {\hbox{$M_{\rm h}$}}\ and a lower normalisation at Milky Way-like values of ${\hbox{$v_{\rm h}$}}\sim 140$~km~s$^{-1}$.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_Z.pdf}
\caption{Metallicity 12+(O/H) as a function of the stellar mass, $M_\star$, from $I\kappa\epsilon\alpha$\ for feedback efficiency $\epsilon=0.01$ ({\em dotted blue}), $0.05$ ({\em solid blue}) and $0.1$ ({\em dashed blue}). The $I\kappa\epsilon\alpha$\ results (solid lines) have been obtained by integrating Eq.~(\ref{eq:Zdot}), for a yield, $y=0.04$ and then converted to the units of 12+(O/H) assuming that a metallicity of $Z=Z_\odot=0.0127$ corresponds to 12+(O/H)$=8.7$. For comparison the observed trend at $z=0$ for SDSS galaxies \citep{Mannucci10} is also shown as a thin magenta line with vertical bars for the scatter. $I\kappa\epsilon\alpha$\ reproduces the slope of the $Z-M_\star$ relation very well.}
\label{fig:Zz}
\end{figure}
Interestingly though, both $I\kappa\epsilon\alpha$\ and {\sc eagle}\ show very little evolution of the $Z-{\hbox{$v_{\rm h}$}}$ relation in Fig.~\ref{fig:Zvh}. Indeed, Eq.~(\ref{eq:Zdot}) shows that the metallicity of a galaxy tends to a value $Z\approx \kappa y{\hbox{$v_{\rm h}$}}^2/v_\star^2$ that in fact only depends on a halo's virial velocity and not explicitly on redshift. In this approximation, the metallicity of a galaxy changes only secularly, tracking the evolution of ${\hbox{$v_{\rm h}$}}^2$. Such a behaviour is an attractor of Eq.~(\ref{eq:Zdot}): given that $\dot M_\star/M_{\rm gas}>0$, $\dot Z$ is positive (negative) when $Z< y\kappa v_{\rm h}^2/v^2_\star$
(when $Z> y\kappa v^2_{\rm h}/v^2_\star$). Therefore $Z$ approaches the secular value,
\begin{equation}
Z = \kappa y{v_{\rm h}^2\over v_\star^2}\,,
\end{equation}
on the gas consumption time-scale, $M_{\rm gas}/\dot M_\star$. This secular value reproduces the evolution from Eq.~(\ref{eq:Zdot}) very well, as shown by the dashed lines in Fig.~\ref{fig:Zatt}.
Using this secular expression for $Z({\hbox{$v_{\rm h}$}})$, taking $v_\star=400$~km~s$^{-1}$, and $y=0.04$, $Z_\odot=0.0127$ for the total metal yield and solar metallicity as done in the
the stellar evolution models collected from the literature by \citet{Wiersma09}, we obtain
\begin{eqnarray}
{Z\over 0.68\times Z_\odot} &=& \left(\frac{M_{\rm h,0}}{10^{12}~{\rm M_\odot}}\right)^{2/3} (m_{\rm h}\mathcal{H})^{2/3} \nonumber\\
&=&
\left({\dot M_{\star}\over 1.2~{\rm M}_\odot{\rm yr}^{-1}}\right)^{2/5}
{(m_{\rm h}\mathcal{H})^{2/3}\over \Psi_\star^{2/5}}\nonumber\\
&=&
\left({M_{\star}\over 1.7\times 10^{10}~{\rm M}_\odot}\right)^{2/5}
{(m_{\rm h}\mathcal{H})^{2/3}\over m_\star^{2/5}}\,.
\label{eq:ZMS}
\end{eqnarray}
The reference values of $\dot M_\star$ and $M_\star$ for the star formation rate and stellar mass, are taken from Eqs.~(\ref{eq:dotmstar}) and (\ref{eq:mstar}). The normalisation of this relation, $0.68Z_\odot$ for $M_{\rm h,0}=10^{12}{\rm M}_\odot$, depends on $I\kappa\epsilon\alpha$\ parameters $\propto \left({\kappa\alpha/(\epsilon\eta)}\right)^{3/5}$. The observed normalisation is uncertain but at face value higher than what we find by a factor of two \cite[e.g.][]{Tremonti04}.
The dependence of $Z$ on ${\hbox{$v_{\rm h}$}}$ implies that $I\kappa\epsilon\alpha$\ galaxies fall on a mass-metallicity relation, as well as on a star-formation rate-metallicity relation. Similarly to the stellar mass-halo mass relation, we can compute how $Z$ depends on $M_\star$ at a given redshift,
\begin{equation}
{d\ln Z\over d\ln M_\star}|_{z={\rm const}}={d\ln Z\over d\ln \dot M_\star}|_{z={\rm const}}={2\over 5}\,,
\end{equation}
independent of redshift, with the value of the exponent resulting from the ${\hbox{$v_{\rm h}$}}^2\propto M_\star^{2/5}$ dependence of Eq.~(\ref{eq:mstar}). As a galaxy grows in mass, its metallicity increases as
\begin{equation}
{d\ln Z\over d\ln M_\star}|_{M_{\rm h,0}={\rm const}}={2\over 3}{d\ln (m_{\rm h}\mathcal{H})/dz\over d\ln m_\star/dz}\,.
\end{equation}
The evolution of $Z$ at a {\em given} stellar mass or star formation rate, is $\propto (m_{\rm h}\mathcal{H})^{2/3}$ according to Eq.~(\ref{eq:ZMS}).
With increasing $z$, $m_{\rm h}(z)$ decreases whereas $\mathcal{H}(z)$ increases, resulting in little evolution in the $Z-M_\star$ relation. At a given value of $M_\star$, $Z$ decreases with increasing $z$ by factors 0.9 and 0.76 compared to its $z=0$ value for $z=2$ and $3$, respectively. The observed evolution is somewhat stronger and better reproduced by the {\sc eagle}\ {\sc reference} model in which $\epsilon$ varies with the local gas properties \citep{DeRossi17}.
Why does $Z$ depend on $M_\star$ in $I\kappa\epsilon\alpha$? The $I\kappa\epsilon\alpha$\ metallicity of a galaxy is $Z\approx y\dot M_\star/(\omega_{\rm b} \dot M_{\rm h})$, the ratio
of the rate at which stars metal enrich the ISM over the rate at which these metals are being diluted by accreting (primordial) gas. The reason this ratio depends on $M_\star$ is that the star formation efficiency
depends on $v_{\rm h}$: $\dot M_\star/ (\omega_{\rm b}\dot M_{\rm h})\propto v_{\rm h}^{2}\propto M_\star^{2/5}$, given that $M_\star\propto v_{\rm h}^5$.
In $I\kappa\epsilon\alpha$\,, {\em the origin of the mass-metallicity relation is the dependence of the star formation efficiency on the halo's virial velocity.
The $M_\star-Z$ relation evolves because the $M_\star-v_{\rm h}$ relation evolves.} The first part of this claim agrees with \cite{Dave12}, but the second part does not: in their model, evolution is caused by the increase in metallicity of accreting gas. Note that, as long as the galaxy self-regulates\footnote{Clearly this would not true in case of a recent merger which might increase $\dot M_\star$ {\em and} dilute $Z$ by gas accretion.} its gas metallicity is set by the instantaneous star formation rate rather than a consequence of the build-up of metals that fail to escape from the potential well of its host halo. In other words, the reason that $Z$ depends on ${\hbox{$v_{\rm h}$}}$ is because $\dot M_\star/\dot M_{\rm h}$ depends on $v_{\rm h}$, rather than that it is \lq easier for metals to escape from halos with low $v_{\rm h}$\rq, as is often claimed. Indeed we have assumed that $Z_{\rm w}=Z$ so that an outflow by itself does not affect $Z$ {\em at all}. Instead, low $v_{\rm h}$ halos have galaxies with low $Z$ because they are inefficient at forming stars.
\subsection{When stellar feedback fails}
\label{sect:AGN}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_nM.pdf}
\caption{The evolution of the galaxy stellar mass function (GSMF). {\em Coloured curves} show the {\sc eagle}\ GSMF (simulation FbConst, in which the feedback efficiency is constant and which includes AGN feedback), with results at $z=1$, 3, 6, and 8 shown in {\em green}, {\em red}, {\em purple} and {\em yellow}, respectively. At high mass, curves are drawn as {\em dashed lines} if there are fewer than 5 galaxies dex in $\log_{10}M_\star$, at low mass when there are fewer than 100 stellar particles per galaxy. {\em Black full lines} are the corresponding $I\kappa\epsilon\alpha$\ results from \S \ref{sect:gsmf}, with a {\em triangle} corresponding to galaxies of mass $M_{\star, {\rm agn}}$ (Eq.~\ref{eq:Magn}) above which feedback from AGN is expected to set in, and a {\em filled circle} at half this mass. The {\em black dotted line} is the halo mass function, Eq.\ref{eq:hmf}. The {\em coloured open circles} indicate the abundance of halos with mass $M_{\rm h,{\rm agn}(z)}/2$, computed from Eq.~(\ref{eq:Magn}). The $I\kappa\epsilon\alpha$\ model predicts the shape and evolution of the normalization of the {\sc eagle}\ GSMF well. The predicted location of the knee in the GSMF is also reasonable. We used $a=\bar a$ and $b=\bar b$ for the accretion history of halos, Eq.~(\ref{eq:correaMp}).}
\label{fig:GSMF}
\end{figure}
The basic $I\kappa\epsilon\alpha$\ relation of Eq.~(\ref{eq:SR}) between the halo accretion rate and the star formation rate results in $\dot M_\star\propto v_{\rm h,0}^{5}$,
where $v_{\rm h,0}$ is the virial velocity of the halo at redshift $z=0$, so that halos with a large virial velocity form stars at a greater rate. For low values of $v_{\rm h,0}$, only a very small fraction of the accreted baryons are converted into stars with the majority of the accreted gas expelled in an outflow, as discussed in \S \ref{sect:outflows}. The rate of gas accretion increases $\propto v^3_{\rm h,0}$ but the star formation rate increases $\propto v_{\rm h,0}^{5}$. Since obviously the star formation rate cannot be higher than when {\em all} accreted gas is converted to stars, $\dot M_\star\le \omega_{\rm b}\dot{\hbox{$M_{\rm h}$}}$, it eventually becomes impossible to satisfy Eq.~(\ref{eq:SR}) when {\hbox{$v_{\rm h}$}}\ is larger than the critical value that results from inserting $\dot M_\star= \omega_{\rm b}\dot{\hbox{$M_{\rm h}$}}$ in Eq.~(\ref{eq:SR}),
\begin{equation}
v_{\rm h, max} = {v_\star\over \kappa^{1/2}}=310 {v_\star/400~{\rm km}~{\rm s}^{-1}\over (\kappa/(5/3))^{1/2}}~{\rm km}~{\rm s}^{-1}\,.
\label{eq:vcrit}
\end{equation}
Equation~\ref{eq:MB2} gives a slightly weaker limit when requiring that the mass-loading factor $\beta\ge 0$ so that any outflow decreases the baryon fraction of the halo rather than spuriously increasing it.
As a halo grows and ${\hbox{$v_{\rm h}$}}$ increases, $\beta$ starts to drop rapidly to values below $1$, as seen in Fig.~\ref{fig:beta}. At $z\le 2$, $\beta\rightarrow 0$ for ${\hbox{$v_{\rm h}$}}\rightarrow v_{\rm h,{\rm max}}$ but $\beta$ already plunges to values $\beta\le 1$
as ${\hbox{$v_{\rm h}$}}$ approaches a somewhat smaller critical velocity. At higher redshifts, this critical halo virial velocity decreases, basically because it is no longer a good approximation to neglect $\dot M_{\rm gas}$.
What is the consequence of this failure of self-regulation for halos with too high ${\hbox{$v_{\rm h}$}}$? The $I\kappa\epsilon\alpha$\ GSMF discussed in \S \ref{sect:gsmf} is a power-law that tracks the power-law shape of the halo mass function. In contrast, the {\em observed} GSMF has an exponential cut-off at stellar masses above a characteristic stellar mass. It is thought that feedback from accreting black holes (AGN) suppresses star formation in such massive galaxies and this is the cause of the observed break in the GSMF \citep[e.g.][]{Bower06,Croton06}.
This motivates us to associate the critical velocity above which stars cannot self-regulate galaxy formation with those halos in which
AGN regulate galaxy formation instead. Using the subscript \lq agn\rq\ as a mnemonic, we see from Fig.~\ref{fig:beta} that the onset of AGN activity takes place at a nearly redshift-independent value of {\hbox{$v_{\rm h}$}}\ of order
\begin{equation}
v_{\rm h, {\rm agn}}\approx 180~{\rm km}~{\rm s}^{-1}\,,
\end{equation}
for which the corresponding virial temperature is\footnote{$m_p$ is the proton mass, $k_{\rm B}$ is Boltzmann's constant and $\mu$ the mean molecular weight.}
\begin{equation}
T_{\rm h,\rm agn} = {\mu m_p v_{\rm h, {\rm agn}}^2\over 5{\rm k_{\rm B}}}
\approx 10^{5.7}\,{\mu\over 0.62}\,
\left({v_{\rm h, {\rm agn}}\over 180~{\rm km}~{\rm s}^{-1}}\right)^2~{\rm K}\,.
\label{eq:Tvir}
\end{equation}
In the model described by \cite{Bower17}, seed black holes start to grow exponentially in mass when the outflow that is powered by feedback from star formation ceases to be buoyant in the hot corona that fills the dark matter halo. This causes a build-up of gas that fuels the growth of the black hole. The episode of exponential growth ends when the black hole is sufficiently massive that its feedback regulates the forming galaxy. In practise this results in a significant decrease in $\dot M_\star/M_\star$. This model describes well the behaviour of galaxies in the {\sc eagle}\ simulation, with the transition between star formation and AGN feedback regulated galaxies occurring in halos with a virial temperature nearly identical to that of Eq.~(\ref{eq:Tvir}) \citep{Mcalpline18}.
At first sight it seems that the reasoning that led to Eq.~(\ref{eq:Tvir}), \lq stellar feedback fails because $v_\star^2$, a measure of the thermal energy of feedback-heated gas, is too low compared to $\kappa{\hbox{$v_{\rm h}$}}^2$\rq\, is very different from that of \cite{Bower17}, \lq stellar feedback fails because outflows are no longer buoyant in the hot corona\rq. However, the build-up of the hot halo is itself depending on the efficiency of stellar feedback \citep{correa18}. Put in terms of $I\kappa\epsilon\alpha$: the higher $\epsilon$, the higher the value of {\hbox{$v_{\rm h}$}}\ above which a hot corona develops (see in particular Fig.~14 in \citealt{correa18}). Within the current interpretation, the failure of stellar feedback is not due to the formation of a hot corona, but rather the formation of a hot halo is facilitated by failing stellar feedback.
The results from previous sections allow us to compute other properties of the halo and the galaxy when ${\hbox{$v_{\rm h}$}}=v_{\rm h, {\rm agn}}$, the onset of AGN activity. The halo mass, stellar mass, and star formation rate in a halo with ${\hbox{$v_{\rm h}$}}=v_{\rm h, {\rm agn}}$ at $z=0$, are
\begin{eqnarray}
M_{\rm h,{\rm agn}}(z=0)&=& 2\times 10^{12}{\rm M}_\odot {(v_{\rm h,{\rm agn}}/180{\rm km}~{\rm s}^{-1})^3\over \alpha/1}\nonumber\\
M_{\star, {\rm agn}}(z=0) &=& 5.3\times10^{10} {\rm M_\odot} \left({v_{\rm h,{\rm agn}}\over 180{\rm km}~{\rm s}^{-1}}\right)^5\nonumber\\
\dot M_{\star, {\rm agn}}(z=0) &=& 3.8\, {\rm M}_\odot{\rm yr}^{-1} \left({v_{\rm h,{\rm agn}}\over 180{\rm km}~{\rm s}^{-1}}\right)^5\,,
\end{eqnarray}
and the corresponding values at redshift $z$ are
\begin{eqnarray}
M_{\rm h,{\rm agn}}(z) &=& {M_{\rm h,{\rm agn}}(z=0)\over \mathcal{H}(z)}\nonumber\\
M_{\star,{\rm agn}}(z) &=& M_{\star,{\rm agn}}(z=0)\,{m_\star(z)\over (m_{\rm h}(z)\mathcal{H}(z))^{5/3}}\nonumber\\
\dot M_{\star, {\rm agn}}(z) &=& \dot M_{\star, {\rm agn}}(z=0) \,{\Psi_\star(z)\over (m_{\rm h}(z)\mathcal{H}(z))^{5/3}}\,.\nonumber\\
\label{eq:Magn}
\end{eqnarray}
We do not expect the $I\kappa\epsilon\alpha$\ GSMF to be correct for halos with ${\hbox{$v_{\rm h}$}}\ge v_{\rm h, {\rm agn}}$. We therefore plot the GSMF discussed in
\S \ref{sect:gsmf} up to halos of mass $M_{\star,{\rm agn}}(z)$, and compare to the {\sc eagle}\ GSMF (simulation FbConst, in which the stellar feedback efficiency is a constant and which {\em does} include feedback from AGN) in Fig.~\ref{fig:GSMF}. The $I\kappa\epsilon\alpha$\ model reproduces the power-law shape of the {\sc eagle}\ mass function up to $M_{\rm h,{\rm agn}}(z)$ well, getting the evolution of the normalisation approximately correct as well. The value of $M_{\star,{\rm agn}}(z)$ is close to where {\sc eagle}\ predicts a rapid decrease in the number density of galaxies, which is due to the action of AGN feedback in the simulation. The number density of galaxies at the knee decreases with increasing $z$. The previous equations elucidate the reason for this in $I\kappa\epsilon\alpha$. Consider two redshifts $z_1$ and $z_2$, with $z_1<z_2$, say. Haloes with ${\hbox{$v_{\rm h}$}}=v_{\rm h,{\rm agn}}$ at a redshift $z_2$ will be more massive at $z=0$ than those that have ${\hbox{$v_{\rm h}$}}=v_{\rm h,{\rm agn}}$ at a redshift $z_1$, by the factor $\mathcal{H}(z_2)/\mathcal{H}(z_1)$, which is $\approx \left((1+z_2)/(1+z_1)\right)^{3/2}$ for $z_1\ge 1$. The corresponding ratio of number densities then follows from the slope of the PS halo-mass function, $\left(\mathcal{H}(z_2)/\mathcal{H}(z_1)\right)^{\alpha_{\rm h}}$. For example the co-moving number density at $z=6$ is lower than at $z=1$ by a factor 4.8.
\subsection{Reality check}
\label{sec:real}
Up to now we have compared $I\kappa\epsilon\alpha$\ to an {\sc eagle}\ simulation in which the feedback parameters are kept constant (simulation FbConst). That simulation does not reproduce the observed properties as well as the {\sc eagle}\ reference simulation. So, how well does $I\kappa\epsilon\alpha$\ describe the observations?
For a fiducial value of $v_\star=400~{\rm km}~{\rm s}^{-1}$, $I\kappa\epsilon\alpha$\ predicts that a $z=0$ galaxy with stellar mass $M_\star=5\times 10^{10}{\rm M}_\odot$ has a star formation rate of $\dot M_\star=3.5~{\rm M}_\odot~{\rm yr}^{-1}$ and is hosted in a dark matter halo of mass $M_{\rm h}=1.9\times 10^{12}{\rm M}_\odot$. For the Milky Way, the inferred values are $M_\star=(5\pm 1)\times 10^{10}{\rm M}_\odot$, $\dot M_\star=(1.65\pm 0.19){\rm M}_\odot~{\rm yr}^{-1}$ and ${\hbox{$M_{\rm h}$}}=(1.1\pm0.3)\times 10^{12}{\rm M}_\odot$ \citep{BlandHawthorn16}, respectively. However, the scatter in $M_\star$ and $\dot M_\star$ for a halo with given $M_{\rm h}$ is substantial, and the $I\kappa\epsilon\alpha$\ value for $M_\star$ is consistent with the abundance matching analysis by \cite{Guo10} and the star formation rate of $\dot M_\star=3.5~{\rm M}_\odot~{\rm yr}^{-1}$ falls well within the blue cloud for a galaxy with that $M_\star$ in the MPA-JHU DR7\footnote{\url{https://wwwmpa.mpa-garching.mpg.de/SDSS/DR7/}} catalogue. This reasonable level of agreement is of course not surprising: we chose $I\kappa\epsilon\alpha$'s feedback efficiency parameter $\epsilon$ which sets $v_\star$ by comparing to these data sets.
The $I\kappa\epsilon\alpha$\ specific star formation rate is $\dot M_\star/M_\star\approx 0.07~{\rm Gyr}^{-1}$ at $z=0$, {\em independent of $\epsilon$}, as compared to an observed value of 0.1~Gyr$^{-1}$ at $M_\star=10^{10}{\rm M}_\odot$ (see the discussion of the data compilation by \citealt{Behroozi18}). The observed sSFR increases to a value of $1~{\rm Gyr}^{-1}$ ($2~{\rm Gyr}^{-1}$) by redshift $z=1$ ($z=2$, \citealt{Behroozi18}), as compared to the $I\kappa\epsilon\alpha$\ values of 0.3 (1). The $I\kappa\epsilon\alpha$\ values are actually very close to those in {\sc eagle}\ (simulation FbConstNoAGN). The faster observed evolution might signal that $\epsilon$ does evolve.
The $M_\star\propto M_{\rm h}^{5/3}$ dependence of stellar mass on halo mass according to Eq.~(\ref{eq:mstar}) results in a redshift-independent low-mass slope of the galaxy stellar mass function of $dn/d\log(M_\star)\propto M_\star^{\approx -0.54}$. The faint-end slope of the Schechter luminosity function \citep{Schechter76},
\begin{equation}
{dn(L)\over d\log L}\propto L^{-\alpha_{\rm g}}\,\exp(-L/L_\star)\,,
\label{eq:Shechter}
\end{equation}
is $\alpha_{\rm g}\approx 0.48$ at redshift $z=0$ in the {\sc gama} \lq z\rq-band \citep{Loveday12}, a long enough wavelength so that stellar mass is approximately proportional to z-band luminosity. The level of agreement between the two slopes, 0.54 versus 0.48, is encouraging, but not surprising given that $I\kappa\epsilon\alpha$\ reproduces the {\sc eagle}\ GSMF at the low mass-end well (Fig.~\ref{fig:GSMF}). Observationally there is no convincing evolution of this slope out to $z\sim 3$ in the $K$-band \citep{Mortlock17}, also consistent with the $I\kappa\epsilon\alpha$\ prediction of no evolution.
The observed evolution in the location of the knee of the Schechter luminosity function is claimed to be consistent with little or no evolution in the value of the {\em stellar mass} at which the transition occurs (e.g. \citealt{Song16}) but an alternative interpretation is that the transition occurs at a nearly constant star formation rate. Indeed, according to \cite{Parsa16}, the absolute 1500\AA\ magnitude of galaxies at the knee of the Schechter luminosity function occurs at $M_{1500,c}=-19.6$, -20.3, -20.6 and -20.68 for redshifts $z=1$, 2, 3, 4, respectively. If we make the reasonable assumption that UV-luminosity is proportional to star formation rate, then the star formation rate $\dot M_{\star}$ of those galaxies increases compared to the value at $z=1$ by factors $\dot M_{\star}(z)/\dot M_{\star}(z=1)=1.9$, 2.5, 2.8 and 2.9 at $z=2$, 3, 4 and 5. The prediction from $I\kappa\epsilon\alpha$\ follows from Eq.~(\ref{eq:Magn}), $\dot M_{\star,{\rm agn}}(z)/\dot M_{\star,{\rm agn}}(z=1)=1.6$, 2.2, 2.8 and 3.4, respectively, impressively close to the observations.
We conclude from this brief comparison to data that $I\kappa\epsilon\alpha$\ reproduces observations of the observed galaxy population and its evolution rather well, although there are some differences too.
\subsection{Incorporating AGN feedback}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_BH.pdf}
\caption{The dependence of black hole mass, $M_{\rm BH}$, on the circular speed of the host halo, $v_c$. The {\it blue line} is the trend from $I\kappa\epsilon\alpha$, Eq.~(\ref{eq:BH}), with $v_{\rm h}$ scaled to $v_c$ following \citealt{Ferrarese02}; the trend originally proposed by \citet{Ferrarese02} is shown as a {\it dashed black line}. Data points are taken from the compilation by \citet{Kormendy13}; {\it blue circles} with error bars are spiral and S0 galaxies and {\it grey circles} are elliptical galaxies. }
\label{fig:BH}
\end{figure}
An important limitation of the model as described so far is the absence of AGN feedback. Following the arguments that led us to stellar-feedback self-regulation,
an obvious way to include AGN in the model is by modifying Eq.~(\ref{eq:SR}) to
\begin{eqnarray}
{1\over 2}\dot M_\star\,v_\star^2 + {1\over 2}\dot M_{\rm BH}\,v_{\rm agn}^2&=& {\kappa\over 2}\omega_{\rm b}\,{\hbox{$\dot M_{\rm h}$}}{\hbox{$v_{\rm h}$}}^2\nonumber\\
v_{\rm agn}^2 &=&{2\epsilon_r\epsilon_f\over 1-\epsilon_r} c^2\,,
\label{eq:vbh}
\end{eqnarray}
with the understanding that AGN feedback sets in when\footnote{$v_{\rm agn}$, which characterises the energy input by the AGN per unit of mass accreted onto the BH, is not to be confused with $v_{\rm h, {\rm agn}}$ - the virial velocity of the halo above which stellar feedback fails.} ${\hbox{$v_{\rm h}$}}\gtrapprox v_{\rm h,{\rm agn}}$. Here, $\epsilon_r\approx 0.1$ is the radiative efficiency of the AGN and $\epsilon_f\approx 0.15$ the fraction of radiated energy that couples to the gas (see the discussion in \citealt{Schaye15}, their section 4.6). As in the case of feedback from star formation discussed in \S \ref{sect:selfreg}, AGN feedback will be self-regulating provided that the black hole accretion rate {\em increases} with the pressure in the ISM.
We can examine what to expect for the black hole mass of a halo with ${\hbox{$v_{\rm h}$}}\gg v_{\rm h,{\rm agn}}$ by
integrating Eq.~(\ref{eq:vbh}) for $v_\star=0$,
\begin{eqnarray}
M_{\rm BH} &=& {3\over 5}{\kappa\omega_{\rm b}\over \alpha^{3/2}} {v_{\rm h}^5\over 10H(z){\rm G}v_{\rm agn}^2}\nonumber\\
&=& 1.2\times 10^7\,{(v_{\rm h}/200~{\rm km}~{\rm s}^{-1})^5\over (\epsilon_r\epsilon_f/(1-\epsilon_r))/(0.1\times 0.07)}{\rm M}_\odot\,,\nonumber\\
\label{eq:BH}
\end{eqnarray}
where the second line is at redshift $z=0$; this scaling is plotted as a blue line in Fig.\ref{fig:BH}. How does this compare to observations? \citet{Ferrarese02} claim that black hole mass scales with the circular speed as $M_{\rm BH} \propto v_{c}^{5.5}$. This scaling is shown as a black dashed line and is close to Eq.~(\ref{eq:BH}). \citet{Kormendy13} argue that, because the scatter in the $M_{\rm BH}\hbox{--}v_c$ relation is large at low $v_c$, the {\it Magorrian relation} \citep{Magorrian98} between black hole mass and bulge mass, is more fundamental. We would argue instead that low-mass black holes are not in a self-regulating regime.
The observed relation between black hole mass and (3D)\footnote{We have assumed that $\sigma_\star^2=3\sigma^2$, where $\sigma$ is the line-of-sight stellar velocity dispersion.} stellar velocity dispersion $\sigma_\star$, is $M_{\rm BH}=1.1\times 10^7\,(\sigma_\star/200~{\rm km}~{\rm s}^{-1})^{5.12}~{\rm M}_\odot$ \citep{McConnell11}. Provided ${\hbox{$v_{\rm h}$}}\sim \sigma_\star$, the observed dependence on velocity is close to our prediction while the normalisation requires reasonable values for $\epsilon_r$ and $\epsilon_f$. The scaling of the $M_{\rm BH}-\sigma_\star$ relation in the model by \cite{Silk98} is identical to ours basically because both are based on energy arguments, however, our normalisation is significantly more realistic, as shown in Fig.~\ref{fig:BH}. The model by \cite{King03} is based on momentum arguments; their scaling, $M_{\rm BH}\propto \sigma_\star^4$, is shallower than observed. \cite{Booth10} obtain a
$M_{\rm BH}\propto M_{\rm h}^{1.55}$ scaling by arguing that the net total energy injected by an AGN is of order of the binding energy of a halo.
This is somewhat similar to our reasoning, except that we argue that it is the {\em rate} of energy injection by the AGN that tracks the {\em rate} of energy accretion by the halo due to self-regulation. The secular growth rate of a black hole - and hence the time-averaged luminosity of the AGN - therefore depends on the cosmological accretion rate onto its host halo and therefore on redshift, and not just on halo properties.
\section*{Acknowledgements}
We thank our colleagues (J. Schaye, M. Schaller, R. Crain and R. Bower) for sharing with us the data from the {\sc eagle}\ simulation; and we are grateful to L. Heck and J. Helly for providing the computing support. This study was funded by the Science and Technology Facilities Council [grant number ST/F001166/1]. The study made use of the DiRAC Data Centric system at Durham University which is run by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk); the equipment was funded by BIS National E-Infrastructure capital grant ST/K00042X/1, STFC capital grant ST/H008519/1, and STFC DiRAC; as a part of the National E-Infrastructure. This research was also supported by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO 3D), through project number CE170100013. The International Centre for Radio Astronomy Research (ICRAR) is a Joint Venture of Curtin University and The University of Western Australia, funded by the Western Australian State government. M.S. is supported by an ASTRO 3D fellowship at the ICRAR, Curtin University.
\bibliographystyle{mnras}
\section{Self-regulation of star formation in galaxies}
The appearance of a star-forming sequence of galaxies is suggestive of the action of a feedback cycle. Such a feedback cycle is also important in understanding the main sequence of {\em stars} in a Hertzsprung-Russel diagram. Indeed: nuclear energy generation in main sequence stars is secularly stable - a prerequisite for their longevity. We begin this section by briefly describing the well-known reason behind this stability (see e.g. any text book on stellar structure, for example \citealt{Prialnik09}). We next investigate whether we can apply similar reasoning to star forming galaxies.
\subsection{The secular evolution of main sequence stars}
The total energy $E$ of a main sequence star of mass $M$ is the sum of its gravitational energy, $\Omega<0$, and its internal energy, $U=Mu$, where $u$ is its mean specific energy per unit mass. Stars are approximately in virial equilibrium, $E=\Omega/2=-Mu$, and as a consequence $dE/du<0$. Therefore, if a star loses energy for example through radiation so that $\dot E<0$, it will {\em heat up}, $\dot u>0$. The effective negative specific heat capacity of a star is a well-known but nevertheless intriguing feature of gravitationally bound systems, see e.g. \cite{Lynden-Bell77}, and is crucial for its longevity.
Indeed, consider a star losing energy through radiation (rate $L$), while gaining internal energy through nuclear fusion (rate $\dot E_{\rm nucl}$),
\begin{equation}
\dot E = \dot E_{\rm nucl} - L\,.
\label{eq:destar}
\end{equation}
In equilibrium, $\dot E=0$, however consider what happens for (small) deviations from equilibrium. Assuming $\dot E_{\rm nucl}<L$, say, $|E|$ increases since $E<0$, meaning $|u|$ increases and hence the temperature $T$ rises. The rate of energy generation through fusion is a rapidly increasing function of $T$, hence increasing $T$ increases $\dot E_{\rm nucl}$, so that $\dot E_{\rm nucl}<L$ results in an increase in $\dot E_{\rm nucl}$ towards equilibrium. Similarly, if $\dot E_{\rm nucl}>L$, the decrease in $T$ results in a decrease in the nuclear burning, until $\dot E_{\rm nucl}=L$. Clearly, the negative specific heat capacity of a star is not just an amusing feature of self-gravitating systems, but is key in understanding stability on the main sequence. As the star's mean molecular weight changes due to fusion, $L$ and hence $\dot E_{\rm nucl}$ evolve secularly on a time scale which vastly exceeds $E/L$.
\subsection{The evolution of a galactic halo}
As a galactic halo\footnote{We will use the term \lq galactic halo\rq\ to refer to a central galaxy (as opposed to a satellite galaxy) with gas and stars, together with its host dark matter halo.} grows in mass due to cosmological accretion, its energy changes in time as well. At first sight there is little in common between the evolution of a galactic halo and that of a main sequence star. Indeed, the total energy of a star changes only secularly, $|\dot E|\ll L$, as self-regulation leads to a near balance between the energy generated by nuclear fusion and lost by radiation, but a galactic halo seems to have no equivalent channel for regulation. Does that mean that it is not secularly stable? The answer is partially yes: we show in the following that the dark matter halo is not secularly stable, in the sense that $\dot E_{\rm h} \neq 0$. However the same may not be true for the galaxy itself, because supernovae inject energy into the interstellar medium. Below we investigate whether that energy injection rate balances the loss of energy due to cosmological accretion, and if such a situation is a stable equilibrium - in analogy with the evolution of main sequence stars describe above. Before we do so we summarise some well known relations for the evolution of dark matter halos.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_virial_mergertree.pdf}
\caption{The redshift evolution of the virial ratio, $-2U_{\rm h}/\Omega_{\rm h}$, of dark matter halos from the {\sc eagle}\ L100N1504 dark matter only simulation tracked along their merger tree. Here, $U_{\rm h}$ is the sum of the kinetic energy of all particles in the centre of mass rest frame, and $\Omega_{\rm h}$ is the gravitational energy. Different colours refer to halos in narrow bins of their $z=0$ halo mass $M_{\rm h,0}$, {\em blue}, {\em red} and {\em purple} correspond to $M_{\rm h, 0}=[0.98\hbox{--}1.02]\times10^{11}$~M$_\odot$, $[0.9\hbox{--}1.1]\times10^{13}$~M$_\odot$ and $[0.8\hbox{--}1.2]\times10^{13}$~M$_\odot$, respectively; solid curves are the median value of the virial ratio, the shaded region encompasses the 25th to 75th percentiles. Halos evolve approximately in virial equilibrium.}
\label{fig:VTz}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_alpha_mergertree_11.pdf}
\caption{ As Fig.\ref{fig:VTz} but for the concentration parameter $\alpha$ from Eq.~(\ref{eq:halo}). As a halo grows, $\alpha$ remains approximately constant. The {\em dashed curve} quantifies the (negligible) effect of the last term on the right hand side in Eq.~(\ref{eq:de}).}
\label{fig:a1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_kappa_123.pdf}
\caption{As Fig.\ref{fig:VTz} but for $\kappa={d(\ln|{E_{\rm h}}|)/ d(\ln{M_{\rm h}})}$, where $E_{\rm h}$ and $M_{\rm h}$ are the total energy and mass of a halo from Eq.~(\ref{eq:halo}).}
\label{fig:k1}
\end{figure}
\subsection{The growth of a dark matter halo}
We begin by investigating the cosmological growth in mass and the associated change in energy of a dark matter halo. The concentration and mass of a dark matter halo may be affected by baryonic processes. Indeed, in the simulations presented by \citealt{Duffy10}, strong cooling and inefficient feedback increases the central dark matter density of galaxy and group halos significantly, whereas strong feedback, for example from an AGN, {\em decreases} that density. Baryonic mass loss, associated with strong feedback, may lead to a decrease in the rate at which a dark matter halo increases its mass. These effects are relatively modest at the scale of galaxies in the {\sc eagle}\ simulations, as shown by \citealt{Schaller15}, and we will neglect them in this paper.
The total energy, $E_{\rm h} <0$, of a dark matter halo with mass {\hbox{$M_{\rm h}$}}\ is the sum of its potential energy, $\Omega_{\rm h}<0$, and its internal energy, $U_{\rm h}$ (the total kinetic energy of all dark matter particles in the centre of mass rest frame, subscript $h$ for halo). Dark matter halos satisfy the virial theorem approximately, $E_{\rm h}\approx {\Omega_{\rm h}/2} \approx -U_{\rm h}$ \citep[e.g.][]{Neto07}, as we show in Fig.~\ref{fig:VTz}. There is clearly some evolution of the ratio $U_{\rm h}/\Omega_{\rm h}$ as the halo grows, but we will neglect this in what follows.
Assuming that the dark matter halo is in virial equilibrium, mass, radius and internal energy are related by,
\begin{eqnarray}
\label{eq:halo}
{\hbox{$E_{\rm h}$}} &=& \Omega_{\rm h} + U_{\rm h} = {\Omega_{\rm h}\over 2} = -U_{\rm h}\nonumber\\
\Omega_{\rm h} &=& -\alpha {{\hbox{\rm G}}\,M^{2}_{h}\over {\hbox{$R_{\rm h}$}}}\,\nonumber\\
U_{\rm h} &=& {1\over 2} {\hbox{$M_{\rm h}$}}\,{\hbox{$v_{\rm h}$}}^2\,\nonumber\\
{\hbox{$R_{\rm h}$}} &=& \left({{\hbox{\rm G}}{\hbox{$M_{\rm h}$}}\over 100\,H^2}\right)^{1/3}\,.
\end{eqnarray}
We used the standard way of assigning a \lq radius\rq, {\hbox{$R_{\rm h}$}}, to a halo, by requiring that the mean density within {\hbox{$R_{\rm h}$}}\ is 200 times the critical density, $\rho_c=3H^2/(8\pi{\hbox{\rm G}})$, where $H(z)$ is the Hubble constant at redshift $z$. The value of the dimensionless parameter $\alpha$ depends on the halo's density profile: $\alpha=3/5$ for constant density, $\alpha = {\hbox{$R_{\rm h}$}}/(6a)$ for the spherical profile with scale radius $a$ described by \cite{Hernquist90}, and $\alpha$ is uniquely related to the concentration parameter, $c$, of a halo with an NFW \citep{Navarro97} profile. Equations~(\ref{eq:halo}) also define a characteristic \lq virial velocity \rq\ of the halo, {\hbox{$v_{\rm h}$}}, also given by
\begin{equation}
{\hbox{$v_{\rm h}$}}^2=\alpha\,({\hbox{\rm G}}{\hbox{$M_{\rm h}$}})^{2/3}(10H)^{2/3}\,.
\label{eq:vh}
\end{equation}
If the accreting halo remains in virial equilibrium, then
\begin{equation}
{d\ln |{\hbox{$E_{\rm h}$}}|\over dz} = {5\over 3} {d\ln{\hbox{$M_{\rm h}$}}\over dz}+{d\ln\alpha\over dz}+{2\over 3}{d\ln H\over dz}\,.
\label{eq:de}
\end{equation}
We will show below that the first term on the right hand side, $|{5\over 3} {d\ln{\hbox{$M_{\rm h}$}}/ dz}|$ is of order unity. How about the other terms? We tracked the evolution of the parameter $\alpha$ of halos in the {\sc eagle}\ L0100N1504 dark matter only simulation along their merger tree. The result is plotted in Fig.~\ref{fig:a1}, where different colours refer to halos in bins of their redshift $z=0$ mass, $M_{\rm h,0}$. As was the case of the virial ratio, there is clearly some evolution in $\alpha$ as a halo grows, but that evolution is relatively weak and we will neglect it.
We also note that the term $(2/3)d\ln H/dz$ is always $<1/2$. Therefore the last two terms in Eq.~(\ref{eq:de}) are small compared to the first term on the right hand side, therefore ${d\ln |E_{\rm h}|/dz} \approx \kappa {d\ln{\hbox{$M_{\rm h}$}}/ dz}$ with $\kappa\approx 5/3$. To test this approximation in more detail, we once more track halos along their merger tree to compute $d\ln |{\hbox{$E_{\rm h}$}}|/d\ln{\hbox{$M_{\rm h}$}}$ directly, the result is plotted in Fig.~\ref{fig:k1}; different colours refer to halos in bins of $M_{\rm h,0}$. As $M_{\rm h}$ increases, $|E_{\rm h}|$ increases, with $d\ln |{\hbox{$E_{\rm h}$}}|/d\ln M_{\rm h}\approx 5/3$. Combining this approximation with Eq.~(\ref{eq:vh}), motivates us to parametrize the rate of change of energy as a halo grows in mass by
\begin{equation}
\label{eq:dedt}
{\hbox{$\dot E_{\rm h}$}} = -{\kappa\over 2}\,{\hbox{$\dot M_{\rm h}$}}{\hbox{$v_{\rm h}$}}^2\,.
\end{equation}
The variables $\alpha$ (Eq.~\ref{eq:vh}) and $\kappa$ are two of the four parameters of the $I\kappa\epsilon\alpha$\ model - and as we just showed, they are approximately independent of halo mass and redshift, and we will simply keep them constant at $\alpha=1$ and $\kappa=5/3$. We proceed by parametrizing the evolution of {\hbox{$M_{\rm h}$}}.
The increase with time of the halo mass in the extended Press-Schecher (EPS) or \lq excursion set\rq\ formalism \citep{Bond91, Lacey93} describes the growth measured in simulations very well. Here we will use the parametrisation described by \cite{Correa15a, Correa15b}, which we write in the form of the value of the halo mass at redshift $z=0$,
${\hbox{$M_{{\rm h},0}$}}$, times a dimensionless function $m_{\rm h}(z)$,
\begin{eqnarray}
{\hbox{$M_{\rm h}$}} &=& {\hbox{$M_{{\rm h},0}$}}\, m_{\rm h}(z)\nonumber\\
m_{\rm h}(z) &\approx &(1+z)^{a}\exp(-bz)\,.
\label{eq:correaM}
\end{eqnarray}
The corresponding logarithmic growth rate is
\begin{eqnarray}
{d\ln{\hbox{$M_{\rm h}$}}\over dz} &=& (a-b)\xi_{\rm h}(z)\nonumber\\
\xi_{\rm h}(z) &=& {1\over a-b}\left({a\over 1+z}-b\right)\,.\nonumber\\
\label{eq:correaMp}
\end{eqnarray}
The dimensionless functions $m_{\rm h}$ and $\xi_{\rm h}(z)$ are both unity at $z=0$. Since halos grow in mass, ${d\ln{\hbox{$M_{\rm h}$}}/dz}\le 0$, and in terms of the previous equation we have that the function $\xi_{\rm h}(z)>0$ but $a-b<0$. The parameters $a$ and $b$ depend on the mass of the halo at some reference redshift which we take to be $z=0$. Averaging over halo masses, \cite{Correa15b} find
\begin{equation}
\bar a\approx 0.24\,,\quad \bar b\approx 0.75\,.
\label{eq:correaf}
\end{equation}
We will use $\bar a$ and $\bar b$ and denote them by $a$ and $b$ in our derivations, but in our figures we will use the more elaborate but more accurate version discussed by \cite{Correa15b} in which $a$ and $b$ are functions of $M_{\rm h,0}$ (except in Figure~\ref{fig:results} and \ref{fig:GSMF} in which we use the constant values from Eq.~\ref{eq:correaf}). Using $\bar a$ and $\bar b$, gives $|(5/3)d\ln{\hbox{$M_{{\rm h},0}$}}/dz|=0.85$ at $z=0$ and $1.25$ for $z\rightarrow\infty$, therefore $|(5/3)d\ln {\hbox{$M_{\rm h}$}}/dz|$ is of order unity, as we used before.
The virial velocity's evolution follows from Eq.~(\ref{eq:vh}),
\begin{equation}
{\hbox{$v_{\rm h}$}}^2(z) = v_{\rm h,0}^2\,\left(m_{\rm h}(z)\mathcal{H}(z)\right)^{2/3}\,,
\label{eq:correavh}
\end{equation}
where the function $\mathcal{H}(z)$ is defined by
\begin{equation}
\mathcal{H}(z)\equiv {H(z)\over H_0}\,.
\label{eq:Ez}
\end{equation}
\subsection{The growth of a galaxy}
\subsubsection{Stability of feedback-regulated galaxy formation}
\label{sect:selfreg}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_EgE.pdf}
\caption{The evolution of the total energy of the dark matter halo, $E_{\rm h}$ ({\em black}) and the total energy of the star forming gas, $E_{\rm g}$ ({\em blue}) along the merger tree of a halo of $z=0$ mass, $M_{\rm h,0}=[0.8\hbox{--}1.2]\times 10^{13}~M_\odot$, selected from the {\sc eagle}\ simulation Ref-L100N1504. The solid curves show the median relation while the shaded area encompasses the 25$^{\rm th}$ to 75$^{\rm th}$ percentiles. While the total energy of the dark halo keeps on decreasing, the energy of the central galaxy decreases (secularly) at a slower rate as it is regulated by feedback from star formation.}
\label{fig:Eg}
\end{figure}
A (central) galaxy too satisfies the equivalent of Eq.~(\ref{eq:dedt}). We neglect any pre-processing of the accreted matter, so that the ratio of gas mass that accretes onto the galaxy to total mass accreted onto the galactic halo, is simply the cosmological ratio $\omega_{\rm b}$ of the baryon to the total matter density,
\begin{equation}
\omega_{\rm b}\equiv {\Omega_{\rm b}\over \Omega_{\rm dm}+\Omega_{\rm b}}={\Omega_{\rm b}\over\Omega_{\rm m}}\,.
\end{equation}
Once more neglecting the effect of the growing galaxy on the dark matter halo leads us to deduce that cosmological accretion decreases the energy of a galaxy at a rate $\dot E_{\rm g} = -(\kappa/ 2)\omega_{\rm b}{\hbox{$\dot M_{\rm h}$}}{\hbox{$v_{\rm h}$}}^2$ (subscript \lq$g$\rq\ for galaxy).
However, unlike the case of the dark matter halo, the growing galaxy {\em generates} energy
through feedback from stars (and AGN, discussed later), therefore
\begin{equation}
\label{eq:edotG}
\dot E_{\rm g} = \dot E_\star -{\kappa\over 2} \omega_{\rm b}\,{\hbox{$\dot M_{\rm h}$}}{\hbox{$v_{\rm h}$}}^2\,.
\end{equation}
In analogy with Eq.~(\ref{eq:destar}), we now speculate that $\dot E_\star\approx {\kappa\over 2} \omega_{\rm b}\,{\hbox{$\dot M_{\rm h}$}}{\hbox{$v_{\rm h}$}}^2$: feedback from star formation compensates the energy loss associated with cosmological accretion so that the galaxy grows at nearly constant energy. Figure~\ref{fig:Eg} supports this {\em Ansatz}: it shows that, whereas the energy $E_{\rm h}$ of the dark matter halo (black curve) increases by almost 2 orders of magnitude from a look-back time of 10~Gyr to the present, the energy of the galaxy, $E_{\rm g}$, (blue curve) changes by less than $\sim 50$ per cent over the same time interval.
Most of the energy injected into the galaxy's interstellar medium (ISM) is associated with star formation (i.e. supernovae and other processes associated with short-lived massive stars), therefore we write $\dot E_\star$ in terms of the star formation rate, $\dot M_\star$, and a characteristic velocity $v_\star$,
\begin{equation}
\dot E_\star = {1\over 2}\dot M_\star\,v_{\star}^2\,.
\label{eq:estar}
\end{equation}
We can obtain an order of magnitude estimate for $v_\star$ by assuming that most of the injected energy is from core collapse supernovae (SNe), which inject $10^{51}$~erg of energy each and occur once per $100/\eta$ solar masses worth of stars formed\footnote{$\eta=1.74$ for a \citet{Chabrier03} stellar initial mass function that consists of stars in the mass range of $[0.1,100]{\rm M}_\odot$, of which those with mass $6-100~{\rm M}_\odot$ explode as a core collapse SN.}, hence
\begin{equation}
\label{eq:sn}
v_{\star} = \left(\epsilon\eta{2\times 10^{51}{\rm erg}\over 100~{\rm M}_\odot}\right)^{1/2} \approx 400 \left({\epsilon\over 0.091}\times {\eta\over 1.74}\right)^{1/2}~{\rm km~s}^{-1}\,.
\end{equation}
The factor $\epsilon$ accounts for radiative loses, with $\epsilon=1$ corresponding to no radiative losses and $\epsilon\ll 1$ when such losses are
substantial. Numerical simulations of SNe going off in a range of gas densities \citep[e.g.][and reference therein]{Thornton98}, and analytical models of the wind in M82 combined with simulations \citep[e.g.][]{Strickland09}, suggest that a large fraction of the injected energy is radiated, $1-\epsilon\approx 90~\%$. The cooling rate of a radiating plasma also depends on its metallicity, therefore $\epsilon$ is is unlikely to be constant in all galaxies and at all times. In this paper we use $\epsilon$ as a fitting parameter when comparing to the simulations; we used a reference value of $\epsilon=0.091$ in Eq.~(\ref{eq:sn}) which is consistent with the expected radiative losses being substantial {\em and} yields a round number for $v_\star$.
Equation~(\ref{eq:edotG}) that describes the rate of change of the energy of a galaxy is reminiscent of Eq.~(\ref{eq:destar}) that describes the rate of change of a main sequence star: whereas the star loses energy (becomes more bound) through radiative losses, the galaxy becomes more bound as the potential well of its host halo deepens due to cosmological accretion. While the star reacts by compensating the energy loss by nuclear fusion, the galaxy reacts by forming stars that inject energy in the galaxy's star forming gas. For stars, this results in
$\dot E=\dot E_{\rm nucl}-L=0$, and we propose here that the same is true for a galaxy, $\dot E_{\rm g}=\dot E_\star-(\kappa/2)\omega_{\rm b}\dot M_{\rm h}{\hbox{$v_{\rm h}$}}^2\approx 0$.
Why would the feedback from star formation be just so that $\dot E_{\rm g}\approx 0$? Is the equilibrium situation stable in the galaxy's case, just as it was for the star? To examine this question, suppose that $\dot E_{\rm g}<0$ - i.e. $|E_{\rm g}|$ is increasing because the galaxy is currently undergoing too little star formation given the current cosmological accretion rate. With gas in the galaxy getting compressed by the deepening potential well, the internal energy $U_{\rm g}$ of the galaxy will increase. How does that affect the star formation rate?
In the {\sc eagle}\ implementation of star formation, an increase in thermal energy per unit mass implies an increase in pressure, $P\propto u^4$ from Eq.~(\ref{eq:eos}), and hence an increase in star formation rate, $\dot\rho_\star\propto u^{4(n-1)/2}\approx u^{0.8}$ from Eq.~(\ref{eq:dotmstar1}) for $n=1.4$ from Eq.~(\ref{eq:KS}). Therefore an increase in the accretion rate results in an increase in the star formation rate (and conversely, a decrease in the accretion rate results in a decrease in the star formation rate), so that the equilibrium\footnote{If the dynamical time-scales are very short, then self-regulation may fail to keep the galaxy in equilibrium. This may happen for example at high redshift,
\citep[e.g.][]{Duffy10}} situation, $\dot E_{\rm g}=0$, is secularly stable - just as in the case of nuclear fusion in a main sequence star, and for a similar reason\footnote{To take the analogy further, the galaxy in its galactic halo plays the role of the stellar core in the main sequence star.}. We note in particular that the increase in star formation rate due to increased accretion, {\em neither} assumes {\em nor} requires that the gas mass - the fuel for star formation - increases. In our model, the gas reservoir is {\em not} regulating the star formation rate in a galaxy. We also note that stability requires that the star formation rate increases with the ISM's pressure, but without requiring any detailed form of the dependence of $\dot\rho_\star$ on $P$: {\em the details of exactly how star formation feedback operates are unimportant for the secular stability of the star formation rate in a star forming galaxy}. Another consequence is that the star formation rate in a cosmological galaxy depends very little, if at all, on the star formation law that relates star formation rate to the gas mass\footnote{We note this is not true in simulations of an {\em isolated} galaxy, for which the simulator specifies the gas fraction.}.
The star formation rate in our model of feedback-regulated galaxy formation, depends on the stellar {\bf I}nitial mass function (through $\eta$ and the recycle fraction ${\cal R}$ discussed below) and the parameters $\mathbold{\kappa}$ (Eq.~\ref{eq:dedt}), $\mathbold{\epsilon}$ (Eq.~\ref{eq:sn}) and $\mathbold{\alpha}$ (Eq.~\ref{eq:halo}), which is why we called it $I\kappa\epsilon\alpha$. By computing the star formation rate and stellar mass as a function of halo mass, we next show that $I\kappa\epsilon\alpha$\ galaxies lie on a star-forming main sequence.
\subsubsection{The main sequence of star forming galaxies}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_scalings.pdf}
\caption{ The evolution of the dimensionless star formation rate $\Psi_\star(z)$ (black curve) from Eq.~(\ref{eq:dotmstar}), stellar mass $m_\star(z)$ (blue curve, we also plot $\log_{10}1/m_\star$ as a dashed blue line) from Eq.~(\ref{eq:mstar}), specific star formation rate ${\hbox{$\Psi_\star$}}(z)/m_\star(z)$ (red curve) from Eq.~(\ref{eq:sSFR}), and the gas mass $m_{\rm gas}$ (magenta curve) from Eq.~(\ref{eq:Mgas}). We used $a=\bar a$ and $b=\bar b$ for the accretion history of halos, Eq.~(\ref{eq:correaMp}).}
\label{fig:results}
\end{figure}
Setting $\dot E_{\rm g}=0$ in Eq.~(\ref{eq:edotG}) for a self-regulating galaxy results in a relation between a galaxy's star formation rate and the cosmological accretion rate onto its host halo at a given redshift,
\begin{equation}
{1\over 2}\dot M_\star v_\star^2 = {\kappa\over 2}\omega_{\rm b}{\hbox{$\dot M_{\rm h}$}}\,{\hbox{$v_{\rm h}$}}^2\,,
\label{eq:SR}
\end{equation}
which is the main result of this paper. The right-hand side is the cosmological {\em energy} accretion rate onto a halo of given mass. The left-hand side sets the corresponding star formation rate in the galaxy, in terms of the effective energy injection rate per stellar mass formed. Substituting the expressions for the accretion rate ${\hbox{$\dot M_{\rm h}$}}$ and virial velocity ${\hbox{$v_{\rm h}$}}$ from Eqs.~(\ref{eq:correaM}) and (\ref{eq:correavh}) allows us to write the star formation rate as a product of its value at $z=0$, $\dot M_{\star, 0}$, times a dimensionless function, ${\hbox{$\Psi_\star$}}(z)$,
\begin{eqnarray}
\dot M_\star(z) &=& \kappa\omega_{\rm b}{\hbox{$\dot M_{\rm h}$}} {{\hbox{$v_{\rm h}$}}^2\over v_\star^2} \equiv \dot M_{\star, 0}{\hbox{$\Psi_\star$}}(z)\nonumber\\
\dot M_{\star, 0} &=& \kappa\omega_{\rm b}(b-a)H_0M_{\rm h,0}{v^2_{\rm h,0}\over v_\star^2}\nonumber\\
&=&1.2~{\rm M}_\odot{\rm yr}^{-1}{\kappa\over 5/3}{\alpha\over 1}\left[{h \over 0.677}\right]^{5/3}\left[{M_{\rm h,0}\over 10^{12}~{\rm M}_\odot}\right]^{5/3}\nonumber\\
&\times &\left[{v_\star\over 400~{\rm km~s}^{-1}}\right]^{-2}\nonumber\\
{\hbox{$\Psi_\star$}}(z) &=& (1+z)\xi_{\rm h}(z)\left(m_{\rm h}(z)\,\mathcal{H}(z)\right)^{5/3}\,.
\label{eq:dotmstar}
\end{eqnarray}
The star formation rate scales $\propto M_{\rm h,0}^{5/3}\propto v_{\rm h,0}^5$; the function ${\hbox{$\Psi_\star$}}(z=0)=1$.
Since stars lose mass during stellar evolution, the time integral of the star formation rate does not equal the total stellar mass at some later time. In the \lq instantaneous recycling approximation\rq,
\begin{equation}
M_\star(t) = (1-{\cal R}) \int_0^t\dot M_\star(t')\,dt' \,,
\label{eq:Mstar}
\end{equation}
where ${\cal R}$ is the fraction of mass originally in stars that is returned back to star forming gas through stellar mass loss; the stellar population models used in {\sc eagle}\ have $1-{\cal R}\approx 0.55$ \citep{Wiersma09}. The stellar mass is in this approximation
\begin{eqnarray}
M_\star(z) &=& (1-{\cal R}){\dot M_{\star,0}\over H_0}\,\int_z^\infty {{\hbox{$\Psi_\star$}}(z') (1+z')^{-1}\mathcal{H}(z')}^{-1}\,dz'\nonumber\\
&\equiv & M_{\star,0}\,m_\star(z)\nonumber\\
M_{\star,0} &=& (1-{\cal R}){\dot M_{\star,0}\over H_0}\,m_{\star,0}\nonumber \\
&=& 1.7\times 10^{10}{\rm M}_\odot {1-{\cal R}\over 0.55}\left[{h\over 0.677}\right]^{2/3}
{\kappa\over 5/3}{\alpha\over 1}\nonumber\\
&\times & \left[{M_{\rm h,0}\over 10^{12}~{\rm M}_\odot}\right]^{5/3}
\left[{v_\star\over 400~{\rm km~s}^{-1}}\right]^{-2}
\nonumber\\
m_\star(z) &=& {1\over m_{\star,0}}\int_z^\infty {{\hbox{$\Psi_\star$}}(z') (1+z')^{-1}\mathcal{H}(z')}^{-1}\,dz'\nonumber\\
m_{\star,0} &=& \int_0^\infty {{\hbox{$\Psi_\star$}}(z') (1+z')^{-1}\mathcal{H}(z')}^{-1}\,dz'=1.78\,,
\label{eq:mstar}
\end{eqnarray}
with $m_\star(z=0)=1$. To evaluate $M_{\star, 0}$ and $\dot M_{\star,0}$ we have used the cosmological parameters $\Omega_{\rm b}=0.0482519$, $\Omega_{\rm m}=0.307$, $\Omega_\Lambda=1-\Omega_{\rm m}$, $\omega_{\rm b}=0.157$ and $h=0.677$ from \cite{Planck14}, and set $a=\bar a$ and $b=\bar b$ for the redshift dependence of the halo accretion rate from Eq.~(\ref{eq:correaf}); numerical values in our figures correspond to the more general accretion histories from \citet{Correa15b}, for which $a$ and $b$ depend on $M_{\rm h,0}$.
The specific star formation rate, ${\rm sSFR}$, follows from combining Eqs.~(\ref{eq:dotmstar}) and (\ref{eq:mstar}),
\begin{eqnarray}
{\rm sSFR}(z) &\equiv& {\dot M_\star(z)\over M_\star(z)} ={H_0\over m_{\star,0}\,(1-{\cal R})}\, {{\hbox{$\Psi_\star$}}(z)\over m_\star(z)}\nonumber\\
&=&{H_0\over m_{\star,0}(1-{\cal R})}\approx 0.07~{\rm Gyr}^{-1}~\hbox{\rm at $z=0$}\,.\nonumber\\
\label{eq:sSFR}
\end{eqnarray}
This ratio would depend on halo mass and hence also on $M_\star$ if ({\em i}) we had taken into account that the halo accretion rate depends on halo mass rather than using the average accretion rate from Eq.~(\ref{eq:correaf}) and ({\em ii}) if one or more of the $I\kappa\epsilon\alpha$\ parameters were to depend on halo mass.
The expression for the sSFR at $z=0$ from Eq.~(\ref{eq:sSFR}) looks suspiciously simple: what sets this numerical value? Tracing back the definitions of the dimensionless functions $\mathcal{H}(z)$ (Eq.~\ref{eq:Ez}), ${\hbox{$\Psi_\star$}}(z)$ (Eq.~\ref{eq:dotmstar}) and $m_\star$ (Eq.~{\ref{eq:mstar}), we see that these {\em only} depend on cosmology and the growth rate of dark matter halos. Changing the growth rate will change the value of the numerical constant $m_{\star,0}$ in Eq.~(\ref{eq:mstar}). The only other $I\kappa\epsilon\alpha$\ parameter that sets the sSFR is ${\cal R}$, the recycled mass fraction, which depends on the IMF. Therefore the value of the sSFR at $z=0$ depends on cosmology (through the accretion history of halos), and on the fraction of mass returned to the ISM during stellar evolution, ${\cal R}$ - and nothing else. This is of course a consequence of assuming that none of the $I\kappa\epsilon\alpha$\ parameters evolve.
The dimensionless functions ${\hbox{$\Psi_\star$}}(z)$, $m_\star(z)$, and ${\hbox{$\Psi_\star$}}(z)/m_\star(z)$ provide the unique connection between the stellar properties of a galaxy and the properties of its host halo - they are plotted in Fig.~\ref{fig:results}. The star formation rate of an $I\kappa\epsilon\alpha$\ galaxy varies over a factor of $\sim 6$ between $z=0$ and $z=6$, peaking at $z\sim 2$, with half the stellar mass forming below $z\sim 1$. The sSFR increases rapidly with redshift, and is higher than its $z=0$ value by factors of 4.6, 13.7 and 30 at redshifts 1, 2 and 3, respectively.
To summarise: $I\kappa\epsilon\alpha$\ predicts a main sequence of star forming galaxies along which the specific star formation rate
does not depend on $M_\star$, provided the $I\kappa\epsilon\alpha$\ parameters themselves do not depend on halo mass. The value of this specific star formation rate increases rapidly with redshift.
\subsubsection{Comparison to {\sc eagle}}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_correa_single_vh_log_1.pdf}
\caption{The dependence of the star formation rate, $\dot M_\star$, on the virial velocity {\hbox{$v_{\rm h}$}}\ of a galaxy's host halo, at different redshifts. The {\em coloured lines} are the predictions from the $I\kappa\epsilon\alpha$\ model (Eq.~\ref{eq:dotmstar}, with $\epsilon=0.2$) based on our self-regulation arguments; {\em large dots} are the median star formation rate in {\sc eagle}\ galaxies (simulation FbconstnoAGN), with the {\em shaded area} encompassing the $25^{\rm th}-75^{\rm th}$ percentile range. Different colours correspond to different redshifts ({\em blue, green, red and purple} correspond to $z=0,1,3$ and 6, respectively). The $I\kappa\epsilon\alpha$\ model captures well the dependence of $\dot M_\star$ on {\hbox{$v_{\rm h}$}}\ and $z$.}
\label{fig:exp1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_correa_single.pdf}
\caption{Same as Fig.~\ref{fig:exp1}, but for the dependence of $\dot M_\star$ on halo mass, {\hbox{$M_{\rm h}$}}. }
\label{fig:exp2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig_correa_single_Mstar_1.pdf}
\caption{The stellar mass-halo mass ratio, $M_\star/M_{\rm h}$, as a function of {\hbox{$M_{\rm h}$}}\ at different redshifts. The {\em coloured lines} are the predictions from the $I\kappa\epsilon\alpha$\ model ($M_\star$ from Eq.~\ref{eq:mstar}); {\em large dots} are the median relation in the {\sc eagle}\ galaxies (simulation FbconstnoAGN), with the {\em shaded area} encompassing the $25^{\rm th}-75^{\rm th}$ percentile range. Different colours correspond to different redshifts ({\em blue, green, red and purple} correspond to $z=0,1,3$ and 6, respectively). The {\em black dashed} and {\em black dotted} lines correspond to {\sc eagle}\ galaxies with approximately 100 and 500 star particles, respectively. }
\label{fig:SHMR}
\end{figure}
We test the ideas put forward in the previous section by comparing the star formation rate of galaxies as a function of halo properties and redshift to that of {\sc eagle}\ galaxies. We emphasize that for a given assumed stellar IMF, the $\epsilon$ parameter of the the $I\kappa\epsilon\alpha$\ model - a measure of the radiative loses in the ISM of the energy injected by SNe - is the central free parameter that sets the star formation rate in a cosmological halo. It does so by setting the characteristic velocity $v_\star$ through Eq.~(\ref{eq:sn}). The parameter $\epsilon$ likely depends on the properties of a galaxy's ISM - presumably $\epsilon$ would be smaller (greater cooling losses) when the ISM is denser and more metal rich. Rather than proposing a more detailed model for this, at this stage we simply keep $\epsilon$ constant. However, the {\sc eagle}\ reference simulation has a parameter $f_{\rm th}$ which explicitly changes the amount of energy injected into the ISM per solar mass of stars formed, depending on density and metallicity of the ISM (see Eq.~(7) in \citealt{Schaye15}). Therefore, to keep the comparison between $I\kappa\epsilon\alpha$\ and {\sc eagle}\ fair, we compare here to the \lq FBconstnoAGN\rq\ {\sc eagle}\ variation, in which $f_{\rm th}$ is kept constant (and which does not include AGN feedback either, see the Appendix for more details). We re-iterate though, that keeping $f_{\rm th}$ constant is not quite equivalent to keeping $\epsilon$ constant, because the cooling losses in {\sc eagle}\ still depend on density and metallicity.
The star formation rate predicted by Eq.~(\ref{eq:dotmstar}) is compared to the {\sc eagle}\ FBconstnoAGN model in Fig.~\ref{fig:exp1}, taking $\alpha=1$, $\kappa=5/3$, $\eta=1.7$ and $\epsilon=0.2$ (so that $v_\star$ is constant, see Eq.~\ref{eq:sn}); coloured lines are the $I\kappa\epsilon\alpha$\ prediction at different redshifts, large dots are the median relation for {\sc eagle}\ galaxies with the shaded region encompassing the $25^{\rm th}-75^{\rm th}$ percentile range. Even when keeping $v_\star$ constant, Eq.~(\ref{eq:dotmstar}) captures accurately the increase in $\dot M_\star$ with the halo's virial velocity {\hbox{$v_{\rm h}$}}\ at fixed $z$, as well as the increase in $\dot M_\star$ with $z$ at fixed {\hbox{$v_{\rm h}$}}. With only one \lq free\rq\ model parameter $\epsilon$ (which sets $v_\star$), we were astonished by the level of agreement between $I\kappa\epsilon\alpha$\ and {\sc eagle}.
In the case of Fig.~\ref{fig:exp1}, the increase in $\dot M_\star$ with $z$ at given {\hbox{$v_{\rm h}$}}\ is due to the increase in the cosmological accretion rate onto a halo with given ${\hbox{$v_{\rm h}$}}$ at given $z$, as is apparent from Eq.~(\ref{eq:SR}). However, plotting $\dot M_\star$ as a function of ${\hbox{$M_{\rm h}$}}$ (Fig.~\ref{fig:exp2}), we see that the redshift dependence is stronger due to the $\mathcal{H}(z)^{2/3}$ dependence of Eq.~(\ref{eq:dotmstar}). This is not surprising within the context of our self-regulation model: the star formation rate depends on virial velocity rather than halo mass.
\subsection{The $M_\star-{\hbox{$M_{\rm h}$}}$ relation}
The stellar mass of a galaxy in Eq.~(\ref{eq:mstar}) is the product of a dimensional number that depends on the galaxy's halo mass at $z=0$, $M_{\rm h,0}$, times a dimensionless function $m_\star(z)$. This functional dependence allows us to answer the question of what is the $M_\star-{\hbox{$M_{\rm h}$}}$ relation in $I\kappa\epsilon\alpha$\ in two different ways, ({\em i}) \lq What is the $M_\star-{\hbox{$M_{\rm h}$}}$ relation for a population of galaxies at a given redshift?\rq, and ({\em ii}) \lq How does the $M_\star-{\hbox{$M_{\rm h}$}}$ ratio of a halo evolve?\rq\ The answer to the first question follows from $M_\star\propto M_{\rm h,0}^{5/3}$ and ${\hbox{$M_{\rm h}$}}\propto M_{\rm h,0}$, therefore
\begin{equation}
{d\ln M_\star\over d\ln{\hbox{$M_{\rm h}$}}}|_{z={\rm const}} = {5\over 3}\,.
\end{equation}
The value of the exponent can be traced back to the ${\hbox{$\dot M_{\rm h}$}}\,{\hbox{$v_{\rm h}$}}^2\propto {\hbox{$M_{\rm h}$}}^{5/3}$ dependence of the star formation rate on halo mass, Eq.~(\ref{eq:dotmstar}). We compare the predicted relation to that measured in {\sc eagle}\ in Fig.~\ref{fig:SHMR}: coloured lines are the theoretical predictions at different redshifts, large dots are the median relation for {\sc eagle}\ galaxies with the shaded region encompassing the $25^{\rm th}-75^{\rm th}$ percentile range. Given that $I\kappa\epsilon\alpha$\ predicts the dependence of $\dot M_\star$ on {\hbox{$M_{\rm h}$}}\ as a function of $z$ in {\sc eagle}\ variation FBconstnoAGN well, it is not very surprising that it also reproduces the relation between $M_\star$ and {\hbox{$M_{\rm h}$}}.
Although $I\kappa\epsilon\alpha$\ galaxies lie along a line with $M_\star/{\hbox{$M_{\rm h}$}}\propto {\hbox{$M_{\rm h}$}}^{2/3}$, they do not evolve along this line. The $M_\star-{\hbox{$M_{\rm h}$}}$ ratio for a given halo evolves as
\begin{equation}
{d\ln M_\star\over d\ln{\hbox{$M_{\rm h}$}}}|_{M_{\rm h,0}={\rm const}}=-{m_{\rm h}(z)\over m_{\star,0}(a-b)m_\star(z)}(m_{\rm h}(z)\mathcal{H}(z))^{2/3}\,.
\end{equation}
This logarithmic slope is $\approx 1.1$ at $z=0$ and increases with $z$ to become nearly constant at a value of 1.4 for $z\ge 4$. If this slope were $5/3$, then (star forming) galaxies would evolve along the $z=0$ $M_\star-M_{\rm h}$ relation so that the stellar mass in a halo of a given mass would be independent of redshift. Because the slope is less than 5/3, the $M_\star/M_{\rm h}$ versus $M_{\rm h}$ relation evolves with redshift, in the sense that the stellar mass increases with redshift at a constant halo mass, however, that evolution is not very strong. This is the redshift evolution seen in Fig.~\ref{fig:SHMR}.
Summarising, we conclude that $I\kappa\epsilon\alpha$\ reproduces the relation between halo mass, star formation rate, and stellar mass measured in the FBconstnoAGN {\sc eagle}\ variation. The fact that $I\kappa\epsilon\alpha$\ reproduces the dependence of $\dot M_\star$ on {\hbox{$M_{\rm h}$}}\ is particularly encouraging, since it directly tests the very basis of the self-regulation argument of Eq.~(\ref{eq:SR}). Interestingly, the star formation rate predicted by Eq.~(\ref{eq:dotmstar}) does not depend {\em at all} on the galaxy's gas mass or indeed the assumed star formation law - as long as $\dot\rho_\star\propto u^\zeta$ for some sufficiently large and positive value of the exponent $\zeta$, so that the star formation rate {\em increases} with the pressure of the galaxy's interstellar medium. Instead the star formation rate depends on the cosmological accretion rate, and on $v_\star$ - that is, on the efficiency of stellar feedback. We will return to this point in the discussion section.
\subsection{The galaxy stellar mass function (GSMF)}
\label{sect:gsmf}
We compute the GSMF by combining the $M_\star-{\hbox{$M_{\rm h}$}}$ relation from $I\kappa\epsilon\alpha$\ with a model for the evolution of the halo mass function. The Press-Schechter (PS, \citealt{Press74}) approximation for the actual number density of halos per dex in halo mass \citep[e.g.][]{Reed07}, at $z=0$, is
\begin{equation}
{dn_{\rm h}\over d\log M_{\rm h,0}} = n_0\,({M_{\rm h,0}\over M_{\rm ps}})^{-\alpha_{\rm h}}\exp(-{M_{\rm h,0}\over M_{\rm ps}})\,,
\label{eq:hmf}
\end{equation}
where $n_0\approx 1\times 10^{-4}{\rm Mpc}^{-3}$ is a normalisation constant, $M_{\rm ps}\approx 2\times 10^{14}{\rm M}_\odot$ a characteristic mass above which the number density of halos falls exponentially, and the exponent $\alpha_{\rm h}\approx 0.9$. In the approximation that all halos grow at the same logarithmic rate, ${\hbox{$M_{\rm h}$}}(z)=M_{\rm h,0}\, m_{\rm h}(z)$, the halo mass function at redshift $z$ is
\begin{equation}
{dn_{\rm h}\over d\log{\hbox{$M_{\rm h}$}}} = n_0\,({{\hbox{$M_{\rm h}$}}\over m_{\rm h}(z)\,M_{\rm ps}})^{-\alpha_{\rm h}}\exp(-{{\hbox{$M_{\rm h}$}}\over m_{\rm h}(z)M_{\rm ps}})\,,
\label{eq:PS}
\end{equation}
where $n_0$ and $M_{PS}$ are redshift independent, and $n_{\rm h}$ is now the co-moving number density of halos per dex in halo mass. Provided the $I\kappa\epsilon\alpha$\ parameters are constants, $M_\star\propto {\hbox{$M_{\rm h}$}}^{5/3}$, and the co-moving number density of galaxies per dex in stellar mass becomes
\begin{eqnarray}
{dn_{\rm g}\over d\log M_\star}&=&{3n_0\over 5}
\left({M_\star\over m_\star(z)M_{\star, {\rm ps}}}\right)^{-\alpha_\star}\,\exp(-
{M_\star\over m_\star(z)M_{\star, {\rm ps}}})\nonumber\\
M_{\star, {\rm ps}} &=&M_{\star,0}\left({M_{\rm ps}\over 10^{12}{\rm M_\odot}}\right)^{5/3}\nonumber\\
&\approx&1.7\times 10^{10}\left({M_{\rm ps}\over 10^{12}{\rm M_\odot}}\right)^{5/3}{\rm M}_\odot\nonumber\\
\alpha_\star &=& {3\over 5}\alpha_{\rm h}\approx 0.54\,.
\label{eq:gsmf}
\end{eqnarray}
In this approximation, the GSMF is just a scaled version of the halo mass function, with a power-law shape at low masses and an exponential cut-off at high masses. However it is well known that the \lq knee\rq\ in the galaxy stellar mass function - above which the exponential sets in - does not correspond to the knee in the halo mass function, but rather is a consequence of AGN feedback \citep{Bower06, Croton06}. We discuss how this can be incorporated in the model in \S \ref{sect:AGN} below.
It is interesting to note that we can make the same argument that lead to Eq.~(\ref{eq:gsmf}) to the star formation rate of a galaxy and compute the \lq galaxy star formation rate function\rq, GSRF, the number density of galaxies per dex of star formation rate.
Since $\dot M_\star \propto {\hbox{$M_{\rm h}$}}^{5/3}$, the GSRF has the same shape as the GSMF,
\begin{eqnarray}
{dn_{\rm g}\over d\log \dot M_\star}&=&{3n_0\over 5}
\left({\dot M_\star\over \Psi_\star(z)\dot M_{\star, {\rm ps}}}\right)^{-\alpha_\star}\,\exp(-
{\dot M_\star\over \Psi_\star(z)\dot M_{\star, {\rm ps}}})\nonumber\\
\dot M_{\star, {\rm ps}} &=&\dot M_{\star,0} \left({M_{\rm ps}\over 10^{12}{\rm M_\odot}}\right)^{5/3}\nonumber\\
&\approx&1.2\,\left({M_{\rm ps}\over 10^{12}{\rm M_\odot}}\right)^{5/3}{\rm M}_\odot~{\rm yr}^{-1}\,.
\label{eq:gsrf}
\end{eqnarray}
The constants $\dot M_{\star,0}(M_{\rm h,0}=10^{12}{\rm M}_\odot, z=0)$ and $M_{\star,0}(M_{\rm h,0}=10^{12}{\rm M}_\odot, z=0)$ are the $z=0$ star formation rate and stellar mass of a galaxy in a halo of mass $10^{12}{\rm M_\odot}$;
the numerical values for these are taken from Eqs.~(\ref{eq:dotmstar}) and (\ref{eq:mstar}), respectively.
At sufficiently low halo mass, these functions are power laws with slope $3\alpha_{\rm h}/5\approx 0.54$. The co-moving number density of galaxies with a given stellar mass increases monotonically with decreasing redshift $\propto m_\star(z)^{\alpha_\star}$. The corresponding evolution of the co-moving number density of galaxies with a given star formation rate is $\propto \Psi_\star(z)^{\alpha_\star}$: this function is not monotonic but peaks around $z\sim 2$. It falls to approximately $0.5$ and $0.54$ times its $z=2$ value at redshifts $z=5.5$ and $z=0$, respectively
The $I\kappa\epsilon\alpha$\ star formation in a halo with low ${\hbox{$v_{\rm h}$}}$ is much less than the rate at which that halo accretes gas. Indeed, according to Eq.(\ref{eq:SR}), only a fraction $\kappa{\hbox{$v_{\rm h}$}}^2/v_\star^2$ of the accreted gas goes into stars. What happens to the remaining gas? Also, self-regulation due to feedback from star formation must eventually fail for sufficiently high values of ${\hbox{$v_{\rm h}$}}\approx v_\star/\kappa^{1/2}\approx 310~{\rm km}~{\rm s}^{-1}$, since then the star formation rate required to self-regulate would {\em exceed} the gas accretion rate. To investigate the consequence of these considerations in more detail, we next examine the gas properties in $I\kappa\epsilon\alpha$.
|
\section*{Abstract}
{\bf For the foreseeable future, the exploration of the high-energy frontier
will be the domain of the Large Hadron Collider (LHC).
Of particular significance will be its high-luminosity
upgrade (HL-LHC), which will operate until the mid-2030s.
In this endeavour, for the full exploitation of the HL-LHC physics potential
an improved understanding of the parton distribution functions (PDFs)
of the proton is critical.
The HL-LHC program would be uniquely complemented by the proposed Large Hadron electron Collider (LHeC),
a high-energy lepton-proton and lepton-nucleus collider based at CERN.
In this work, we build on our recent PDF projections for the HL-LHC
to assess the constraining power of the LHeC measurements of inclusive
and heavy quark structure functions.
We find that the impact of the LHeC would be significant, reducing
PDF uncertainties by up to an order of magnitude in comparison to state-of-the-art
global fits.
In comparison to the HL--LHC projections, the PDF constraints from the inclusive and semi--inclusive LHeC data
are in general more significant for small and intermediate values of the momentum fraction $x$.
At higher values of $x$, the impact of the LHeC and HL--LHC data is expected to be of a
comparable size, with the HL--LHC constraints being more competitive in some cases, and the LHeC ones in others.
}
\clearpage
\vspace{10pt}
\noindent\rule{\textwidth}{1pt}
\tableofcontents\thispagestyle{fancy}
\noindent\rule{\textwidth}{1pt}
\vspace{10pt}
\section{Introduction}
\label{sec:intro}
The particle physics community is busy preparing for the extensive precision and discovery physics programme that will come from Run III of the LHC, and most significantly, for the major upgrade beginning in the mid-2020s, the High-Luminosity LHC (HL-LHC).
Here, protons will be collided with an instantaneous luminosity a factor of five greater than the LHC and will accumulate up to ten times more data, resulting in an integrated luminosity of around $\mathcal{L}=3$ ${\rm ab}^{-1}$ for both the ATLAS and CMS detectors, and 300 ${\rm fb}^{-1}$ for LHCb.
The rich physics prospects of the HL-LHC, which will operate until at least 2035,
have recently been analyzed in detail~\cite{Azzi:2019yne,Cepeda:2019klc,CidVidal:2018eel,Cerri:2018ypt,Citron:2018lsq},
considering physics within and beyond the Standard Model (SM), Higgs, flavour, and heavy ion physics.
In this context, a precise determination of the quark and gluon structure of the proton,
as encoded in the parton distribution functions (PDFs)~\cite{Gao:2017yyd,Rojo:2015acz,Kovarik:2019xvh},
is an essential ingredient for the success of the HL--LHC.
Conversely, the HL--LHC itself offers an unprecedented opportunity to
improve our understanding of proton structure.
We recently analyzed in detail the HL-LHC potential to constrain the PDFs~\cite{Khalek:2018mdn}
by using projected measurements for a range of SM processes,
from Drell-Yan to top quark pair and jet production.
We found that PDF uncertainties on LHC processes can be reduced by a factor between two and five,
depending on the specific flavour combination and on the assumptions about the experimental systematic uncertainties.
Our PDF projections have already been used in a number of related HL--LHC studies,
as reported in~\cite{Azzi:2019yne,Cepeda:2019klc}.
A quite distinct possibility to improve our understanding of proton structure
is the proposal to collide high energy electron and positron beams with the hadron beams from the HL--LHC.
This facility, known as the Large Hadron Electron Collider (LHeC)~\cite{AbelleiraFernandez:2012cc,AbelleiraFernandez:2012ty,Klein:2018rhq}, would run concurrently with the HL--LHC and be based on a new purpose--built detector at the designed interaction point.
A key outcome of the LHeC operations would be a significantly larger and higher--energy dataset of lepton--proton collisions in comparison to the existing HERA structure function measurements~\cite{Abramowicz:2015mha}.
Indeed, the latter to this day form the backbone of all PDF determinations~\cite{Harland-Lang:2014zoa,Dulat:2015mca,Abramowicz:2015mha,Alekhin:2017kpj,Ball:2017nwa}, and thus the LHeC would
provide the opportunity to greatly extend the precision and reach of HERA data in both $x$ and $Q^2$,
highlighting its potential for PDF constraints.
Moreover, these measurements
would be taken in the relatively clean environment of lepton--proton collisions, where the corresponding
theoretical predictions are known to a very high level of precision.
It should also be emphasized here that the LHeC has a broad and exciting physics program which goes
well beyond studies of the proton structure, including topics such as the characterization
of the Higgs sector or the study of cold nuclear matter in the small-$x$ region, where new QCD dynamical
regimes such as saturation are expected to appear.
Quantitative PDF projection studies based on LHeC pseudo--data have been presented
previously~\cite{AbelleiraFernandez:2012cc,AbelleiraFernandez:2012ty,Paukkunen:2017phq,Cooper-Sarkar:2016udp},
where a sizeable reduction in the resultant PDF uncertainties is reported.
These LHeC PDF projections are based upon the HERAPDF-like input PDF parametrization and flavour assumptions~\cite{Abramowicz:2015mha},
with some additional freedom in the input parametrisation
added in the most recent studies~\cite{LHeCtalk1,LHeCtalk2}.
These baseline fits include constraints from the HERA structure function measurements, with in some cases the addition of a limited subset of collider data.
However, different
results may be obtained if a more flexible parametrization or alternative flavour assumptions are used,
as shown for example in the study of~\cite{Ball:2017otu} carried out in the NNPDF framework.
In addition, the interplay of these constraints from the LHeC with the expected sensitivity from the HL--LHC~\cite{Khalek:2018mdn} has not yet been studied.
Thus a natural question to ask is how the projected sensitivity of a state-of-the-art
global PDF determination
will improve with data from the LHeC, and how this will complement the information
contained in the measurements in $pp$ collisions provided by the HL--LHC.
In this paper, we study in detail the projected sensitivity of the LHeC for constraining PDFs within the framework
of a global analysis.
We follow the strategy presented in~\cite{Khalek:2018mdn}, starting from the
PDF4LHC15~\cite{Gao:2013bia,Carrazza:2015aoa,Butterworth:2015oua} baseline set, and quantify the expected impact of the LHeC measurements both individually and combined
with the information provided by the HL--LHC. For the LHeC pseudodata we use the most recent publicly available official
LHeC projections on the expected statistical and systematic errors, and choice of binning~\cite{mklein} (see also~\cite{Klein:1564929} for
further details)
As we will demonstrate, the expected constraints from the LHeC are significant and fully complementary
with those from the HL--LHC.
When included simultaneously, a significant reduction in PDF uncertainties in the entire relevant
kinematical range for the momentum fraction $x$ is achieved, with beneficial implications for LHC phenomenology.
In addition, in order to understand some of the methodological
systematic effects which may be at play when performing such a profiling study, we
also assess the impact of adopting the more restrictive HERAPDF input parameterisation as our baseline PDF set.
The slightly subtle issue here is that when one generates pseudodata assuming an underlying PDF parameterisation one is implicitly assuming that it will be possible to describe the actual data with this parameterisation. If this assumption is too strong, that is the underlying parameterisation is too restrictive, then one is liable to overestimate the impact of the data. Consistent with this, we find a more marked reduction of the
PDF uncertainties for the rather restrictive HERAPDF baseline in comparison to the global PDF4LHC case. To clarify this further, we present pure `LHeC--only' fits, that is modifying the Hessian prior to effectively remove the impact of the data in the baselines, and isolate the impact of the parameterisation alone. These are therefore directly comparable to other LHeC--only projections; we in particular take a tolerance of $T=1$ here. We again find significantly smaller projected errors when assuming the HERAPDF parameterisation in comparison to the more flexible PDF4LHC one.
The outline of this paper is as follows.
In Sect.~\ref{sec:pseudo--data} we present the main features of the projected
LHeC pseudo--data and the theory settings that will be used for the QCD analysis.
In Sect.~\ref{sec:Hessian} we review the Hessian profiling formalism which is used
to quantify the constraints on the PDFs of the LHeC pseudo--data.
In Sect.~\ref{sect:results} we present the PDF projections for the LHeC, both
individually and in combination with the HL--LHC.
In Sect.~\ref{sec:tolstudy} we study in detail the impact of varying the tolerance and the
flexibility of the underlying PDF parameterisation on the projected constraints.
Finally, in Sect~\ref{sec:conc} we conclude.
\section{Pseudo--data generation and theory calculations}
\label{sec:pseudo--data}
In this section we present the details of the LHeC pseudo--data that will be used for our PDF projections,
as well as the settings required to evaluate the corresponding theory predictions
for the LHeC inclusive and heavy quark structure functions.
\subsection{LHeC pseudo--data}
For the LHeC dataset, we use the most recent publicly available official
LHeC projections~\cite{mklein} (see also~\cite{Klein:1564929} for
further details) for electron
and positron neutral-current (NC) and charged-current (CC) scattering.
The main features of the pseudo--data sets we consider are summarised in Table~\ref{tab:pseudo--dataLHeC}, along with the corresponding integrated luminosities and kinematic reach.
While the nominal high energy data ($E_p=7$ TeV) provides the dominant PDF constraints, the
lower energy ($E_p=1$ TeV) data extends the acceptance to higher $x$ and provides a handle on the longitudinal structure function, $F_L$, and hence the gluon PDF (we note that further variations in the electron and/or proton energy will provide additional constraints on this, as well as on other novel low--$x$ QCD phenomena).
The charm and bottom heavy quark NC structure function pseudo--data provide additional constraints on the gluon.
In addition, charm production in $e^- p$ CC scattering provides important information on the anti-strange quark distributions
via the $\overline{s}+ W \to \overline{c}$ process. We do not include charm production in $e^+ p$ CC scattering, as the corresponding pseudo-datasets are not currently publicly available, though this would provide an additional constraint on the strange quark PDF.
We apply a kinematic cut of $Q \ge 2$ GeV to ensure that the fitted
data lie in the range where perturbative QCD calculations can be reliably
applied.
We note that in what follows we will sometimes refer to `data' for brevity, but this is always understood to imply the pseudo-datasets as described above.
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.70}
\small
\begin{tabular}{c|c|c|c|c}
\toprule
Observable & $E_p$ & Kinematics & $N_{\rm dat}$ & $\mathcal{L}_{\rm int}$ [${\rm ab}^{-1}$] \\
\toprule
$\tilde{\sigma}^{\rm NC}$ ($e^- p$) & 7 TeV & $5\times10^{-6}\le x \le 0.8$,
$5 \le Q^2 \le 10^6$ ${\rm GeV}^2$ &
150 & 1.0 \\
\midrule
$\tilde{\sigma}^{\rm CC}$ ($e^- p$) & 7 TeV & $8.5 \times 10^{-5}\le x \le 0.8$,
$10^2 \le Q^2 \le 10^6$ ${\rm GeV}^2$ &
114 & 1.0 \\
\midrule
$\tilde{\sigma}^{\rm NC}$ ($e^+ p$) & 7 TeV & $5\times10^{-6}\le x \le 0.8$,
$5 \le Q^2 \le 5\times 10^5$ ${\rm GeV}^2$ &
148 & 0.1 \\
\midrule
$\tilde{\sigma}^{\rm CC}$ ($e^+ p$) & 7 TeV & $8.5 \times 10^{-5}\le x \le 0.7$,
$10^2 \le Q^2 \le 5\times 10^5$ ${\rm GeV}^2$ &
109 & 0.1 \\
\midrule
$\tilde{\sigma}^{\rm NC}$ ($e^- p$) & 1 TeV & $5\times10^{-5}\le x \le 0.8$,
$2.2 \le Q^2 \le 10^5$ ${\rm GeV}^2$ &
128 & 0.1 \\
\midrule
$\tilde{\sigma}^{\rm CC}$ ($e^- p$) & 1 TeV & $5\times 10^{-4}\le x \le 0.8$,
$10^2 \le Q^2 \le 10^5$ ${\rm GeV}^2$ &
94 & 0.1 \\
\midrule
$F_{2}^{c,\rm NC}$ ($e^- p$) & 7 TeV & $7\times 10^{-6}\le x \le 0.3$,
$4 \le Q^2 \le 2\times 10^5$ ${\rm GeV}^2$ &
111 & 1.0 \\
\midrule
$F_2^{b,\rm NC}$ ($e^- p$) & 7 TeV & $3\times 10^{-5}\le x \le 0.3$,
$32 \le Q^2 \le 2\times 10^5$ ${\rm GeV}^2$ &
77 & 1.0 \\
\midrule
$F_2^{c,\rm CC}$ ($e^- p$) & 7 TeV & $10^{-4}\le x \le 0.25$,
$10^2 \le Q^2 \le 10^5$ ${\rm GeV}^2$ &
14 & 1.0 \\
\bottomrule
Total & & & 945 & \\
\bottomrule
\end{tabular}
\vspace{0.3cm}
\caption{\small \label{tab:pseudo--dataLHeC}
%
Overview of the main features of the LHeC pseudo--data~\cite{mklein} included in our PDF projections.
%
For each process, we indicate the kinematic coverage, the integrated
luminosity, the proton energy, and the number of pseudo--data
points, $N_{\rm dat}$, after the $Q\ge 2$ GeV kinematic cut.
%
Note that in all cases the incoming lepton energy is fixed to be
$E_l=60$ GeV.
%
We ignore the effect of the incoming lepton beam polarization.
}\label{tab:lhecdat}
\end{table}
For the integrated luminosity that is expected to be collected by the LHeC, we take
$\mathcal{L}=1$ ab$^{-1}$ for the high energy $E_{p}=7$ TeV electron and positron cross sections, corresponding to roughly
a three orders of magnitude larger dataset in comparison to HERA.
In fact, as most of these measurements
become quickly dominated by systematic errors, our results do not depend too sensitively on this specific choice.
For the lower energy inclusive structure functions ($E_p=1$ TeV) as well as for the semi-inclusive
measurements we assume $\mathcal{L}=0.1$ ab$^{-1}$.
In total we have $N_{\rm dat}=945$ pseudo--data points that satisfy the $Q\ge 2$ GeV kinematic cut.
For all of the LHeC pseudo--data, we consider the case of unpolarized lepton beams, neglecting the effect
of lepton polarisation.
While beam polarization is, for example, important for precision measurements of electroweak parameters such as
the $W$ mass or the Weinberg angle $\sin\theta_W$, for PDF determination
it is known that the impact of beam polarisation effects is small.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.90\linewidth]{fig/kinplot_fcc.pdf}
\caption{\small The kinematic coverage in the $(x,Q^2)$ plane of the LHeC pseudo--data~\cite{mklein} included in the present analysis:
the inclusive NC and CC structure functions both for high energy (HE)
and low energy (LE) datasets, the NC charm and bottom
semi-inclusive structure functions $F_2^{c\bar{c}}$ and $F_2^{b\bar{b}}$, and the CC charm structure functions
$F_2^c$ providing direct information on the strange content of the proton.
\label{fig:kinplot} }
\end{center}
\end{figure}
The kinematic reach in the $(x,Q^2)$ plane of the LHeC pseudo--data is shown in Fig.~\ref{fig:kinplot}.
The reach in the perturbative region ($Q\ge 2$ GeV)
is well below $x \approx 10^{-5}$ and extends up to $Q^2 \approx 10^6$ ${\rm GeV}^2$
(that is, $Q\simeq 1$ TeV), increasing the HERA coverage
by over an order of magnitude in both cases, via the factor $\sim$ 4
increase in
the collider centre-of-mass energy $\sqrt{s}$.
Due to the heavy quark tagging requirements, the reach for semi-inclusive structure functions
only extends up to $x\simeq 0.3$ in the large-$x$ region.
Note that in addition to providing PDF information, the extended coverage
of the LHeC in the high-$Q$ region would also provide novel opportunities
for indirect searches for new physics beyond the Standard Model
through precision measurements,
see for example~\cite{AbelleiraFernandez:2012cc,AbelleiraFernandez:2012ty,Carrazza:2019sec},
as well as a rich program of Higgs production and decay
studies.
The pseudo--data listed in Table~\ref{tab:lhecdat}
have been generated assuming a detector coverage with lepton
rapidity $|\eta_l|\le 5$ and inelasticity $0.001 \le y \le 0.95$.
Systematic uncertainties due to the scattered electron (positron)
energy scale and polar angle, hadronic energy scale, calorimeter noise, radiative corrections, photoproduction background and a global efficiency error are included in a correlated way across the NC datasets, while a single global source of correlated systematic is taken across all CC datasets.
In addition, an uncorrelated efficiency uncertainty of 0.5\% is taken, while a fully correlated luminosity uncertainty of 1\% is assumed.
In the case of the semi-inclusive heavy-quark structure functions, there are two sources of systematics considered correlated across bins for both NC and CC production respectively.
We note that the statistical errors are generally an order of magnitude or more smaller than the systematic
uncertainties, apart from close to kinematic boundaries, and
hence as discussed above we would not expect
our results to change significantly if somewhat smaller datasets are assumed.
Indeed, we have explicitly verified the validity of this assumption
by using instead an
integrated luminosity of 0.3 ab$^{-1}$ for the case of high energy neutral-current electron scattering.
According to the above considerations, we then produce the pseudo--data values as usual by shifting the corresponding theory predictions by the appropriate experimental errors. In particular, the pseudo--data point $i$ is generated according to
\begin{equation}\label{eq:pseudo--data}
\sigma_i^{\rm exp} = \sigma_i^{\rm th} \left(1+ \delta^{\rm exp}_{{\rm unc},i}\cdot r_i + \sum_k \Gamma_{ik}^{\rm exp} s_{k,i}\right) \;,
\end{equation}
where $s_i$, $r_k$ are univariate Gaussian random numbers, $\Gamma_{ik}^{\rm exp}$ is the $k$-th correlated systematic error and $\delta^{\rm exp}_{{\rm unc},i}$ is the total uncorrelated error for datapoint $i$.
The $\sigma_i^{\rm th}$ are the corresponding theoretical predictions computed
using the baseline PDF set, which we discuss in more detail below.
The $s_{k,i}$ random numbers are the same for all data points
for which the $k$-th systematic error is fully correlated among them.
In addition to the processes listed in Table~\ref{tab:lhecdat}, there are additional PDF-sensitive
measurements from the LHeC that one could consider in such an exercise.
One example is jet production in electron-proton collisions, for which NNLO (and in some cases even N${}^3$LO)
QCD calculations are available~\cite{Gehrmann:2018odt,Currie:2018fgr,Currie:2017tpe}, and that has also been
studied at HERA~\cite{Andreev:2016tgi}, see also~\cite{Britzger:2019kkb,Cooper-Sarkar:2019pkf}.
While such jet production at the LHeC will provide additional information on the large-$x$ gluon,
as the pseudo--data projections are not currently available,
we do not consider these here.
A further example is charm production in $e^+ p$ CC scattering, which would provide a constraint on the strange quark PDF; we currently only include this process in $e^- p$ scattering, as the $e^+ p$ pseudo-data projections are again not currently available.
We note that these same choices have also been adopted
by all other LHeC PDF projections carried out so far.
\subsection{Theoretical calculations}
For all the pseudo--data listed in Table~\ref{tab:lhecdat} we have evaluated
the corresponding theoretical predictions based on the PDF4LHC15 NNLO PDF set.
Specifically, we use
the Hessian version~\cite{Gao:2013bia,Carrazza:2015aoa} {\tt PDF4LHC15\_100}
composed of $N_{\rm eig}=100$ symmetric eigenvectors. This is of course not the only possible choice, and indeed as well as missing some of the more recent LHC and non--LHC data included in global fits, it omits the possibility for the LHeC to for example shed light on the different variable flavour schemes applied by the global fits included in PDF4LHC (see~\cite{Accardi:2016ndt} for a general critical discussion). However, this remains a useful baseline, corresponding to the best available estimate for the current knowledge on proton structure from global PDF fits.
Moreover, this is the same baseline PDF set used for our earlier HL--LHC projections~\cite{Khalek:2018mdn}, allowing us to directly
compare the results of these with those based on the LHeC pseudo--data, as well as to combine
the constraints provided by the two future facilities.
The LHeC structure functions have been evaluated at NLO using the {\tt APFEL} program~\cite{Bertone:2013vaa}
with the FONLL-B general-mass variable-flavor-number scheme~\cite{Forte:2010ta}.
In a real PDF determination it would be important to include NNLO QCD corrections
as well as in principle small-$x$ BFKL resummation corrections~\cite{Ball:2017otu}. However in our case, where by construction the agreement between data and theory is good and we are only interested in evaluating the relative reduction in PDF uncertainties, this is not necessary.
In particular, as the dominant PDF sensitivity is already contained within the NLO calculation, it is not
necessary to include higher-order effects.
On the other hand, theoretical uncertainties such as those arising from missing higher orders in the predictions that enter the PDF fit~\cite{AbdulKhalek:2019bux,Harland-Lang:2018bxd}, which are omitted here, may have some impact, given
that as will be shown below the resultant PDF uncertainties are often
at the per-mile level. In addition, for simplicity we do not consider the contribution from the choice of heavy quark masses or strong coupling on the PDF projections (see e.g.~\cite{Alekhin:2017kpj} for a study of these effects in a global PDF fit). In fact, as discussed most recently in~\cite{LHeCtalk1}, data from the LHeC can reduce the uncertainties on these inputs below the level where they will be expected to give a significant contribution to PDF uncertainties.
\section{Hessian profiling and the role of tolerance}
\label{sec:Hessian}
In this section we review the Hessian profiling formalism~\cite{Paukkunen:2014zia,Schmidt:2018hvu} which is used
to estimate the constraints on the PDFs of the LHeC pseudo--data, following
the general approach presented in~\cite{Khalek:2018mdn}.
After minimising with respect to the experimental nuisance parameters, the total $\chi^2$ due to $N_{\rm dat}$ pseudo--data points can be written as
\begin{eqnarray}
\label{eq:hessianchi2rev}
\chi^2\left( {\rm \beta_{th}}\right)
&=&\sum_{i,j=1}^{N_{\rm dat}}\left( \sigma_i^{\rm exp}
-\sigma_i^{\rm th}
-\sum_k\sigma_i^{\rm th}\Gamma_{ik}^{\rm th}\,\beta_{k,\rm th}\right) C_{ij}^{-1}\left( \sigma_j^{\rm exp}
-\sigma_j^{\rm th}
-\sum_m\sigma_j^{\rm th}\Gamma_{jm}^{\rm th}\,\beta_{m,\rm th}\right) \nonumber \\
&& +T^2\sum_k \beta_{k,\rm th}^2 \; .
\end{eqnarray}
Here $\sigma_i^{\rm exp}$ and $\sigma_i^{\rm th}$ represent the central values
of the pseudo--data and theory predictions, respectively. The $\beta_{k,\rm th}$ are the nuisance parameters
corresponding to movement along the PDF eigenvectors, i.e. such that $\beta_{k,\rm th}=0$ gives the prediction from the baseline PDF set, prior to profiling, while a non--zero value will result from the profiling itself.
The matrix $\Gamma_{ik}^{\rm th}$ corresponds to the rate of change of the theory prediction $i$ with eigenvector $k$, encoding the effect of varying these nuisance parameters on the theory. The tolerance factor $T$ will be discussed further below.
The above expression for the $\chi^2$, Eq.~(\ref{eq:hessianchi2rev}), assumes that the final profiled theory prediction is sufficiently close to the original PDF prediction that we only have to expand to the first order in $\beta_{k,\rm th}$.
That is, a linear approximation is taken, and higher order corrections beyond it are assumed
to be negligible (see~\cite{Eskola:2019dui} for a discussion of such effects).
As we are interested in a closure test, where this will by construction be true, profiling is expected to represent to very good approximation the result of performing an actual fit.
The experimental covariance matrix $C$ that enters the $\chi^2$ definition
Eq.~(\ref{eq:hessianchi2rev}) is given by the following expression:
\begin{equation}
\label{eq:covariancematrix}
C_{ij} = \delta_{ij}\left( \delta^{\rm exp}_{{\rm unc},i}\sigma^{\rm th}_i\right)^2 +
\sum_k\sigma_i^{\rm th}\sigma_j^{\rm th}\Gamma_{ik}^{\rm exp}\Gamma_{jk}^{\rm exp} \, ,
\end{equation}
where $\Gamma_{ik}^{\rm exp}$ and $\delta^{\rm exp}_{{\rm unc},i}$ are defined as in Eq.~\eqref{eq:pseudo--data}.
%
Note that as the impact of the uncorrelated errors are defined in terms of a fixed
theoretical prediction (rather than of the fit output itself), our
results are resilient with respect to
the D'Agostini bias~\cite{Ball:2009qv,Ball:2012wy}.
Profiling then proceeds via the minimisation of Eq.~(\ref{eq:hessianchi2rev})
with respect to the Hessian PDF nuisance parameters $\beta_{k,\rm th}$. This can be performed analytically, resulting in
\begin{equation}\label{eq:betamin}
\beta_{k,\rm th}^{\rm min} = - \sum_l H_{kl}^{-1}a_l\;,
\end{equation}
where we have defined
\begin{align}\label{eq:hkl}
H_{kl} &= \sum_{i,j} \sigma_i^{\rm th}\Gamma_{ik}^{\rm th} C^{-1}_{ij} \sigma_j^{\rm th}\Gamma_{jl}^{\rm th} + T^2 \delta_{kl}\;,\\
a_k & = \sum_{i,j} \sigma_i^{\rm th}\Gamma_{ik}^{\rm th} C^{-1}_{ij}\left(\sigma_j^{\rm th}-\sigma_j^{\rm exp}\right)\;.
\end{align}
The result of Eq.~(\ref{eq:betamin}) represents a new shifted position in eigenvector space, and we can then readily construct a new set of profiled PDF parameters from this.
The matrix $H$ corresponds to the new Hessian matrix for the profiled fit, so that Eq.~\eqref{eq:hessianchi2rev} can be rewritten as
\begin{equation}\label{eq:chiprof}
\chi^2_{\rm profiled} = \chi^2_{\rm profiled}|_{\beta_{k,\rm th}=\beta_{k,\rm th}^{\rm min}} + \sum_{kl} \delta \beta_{k,\rm th} H_{kl}\delta\beta_{l,\rm th}\;,
\end{equation}
where $\delta \beta_{k,\rm th} = \beta_{k,\rm th} - \beta_{k,\rm th}^{\rm min}$.
This can be diagonalised in the usual way in terms of the eigenvectors $\vec{v}^{(k)}$ of $H$, with $k=1,\cdots,N_{\rm eig}$. This then provides a new set of PDF errors via
the following expression
\begin{equation}\label{eq:betaerror}
\delta \beta_{l,\rm th}^{(k)} = T v_l^{(k)} \sqrt{\frac{1}{\hat{\epsilon}_k}}\;,
\end{equation}
where the errors are treated as symmetric, and as above $v_l^{(k)}$ is the $l$th component of $k$th eigenvector of $H$, and $\hat{\epsilon}_k$ is the corresponding eigenvalue.
Now, in Eq.~\eqref{eq:betaerror} we can see that the profiled PDF uncertainty carries
an explicit dependence on the tolerance $T$, which ensures that the corresponding uncertainty as in \eqref{eq:chiprof} is determined via a $\Delta\chi^2=T^2$ criterion.
However, we can see from Eq.~\eqref{eq:hkl} that the new Hessian matrix also depends on the tolerance, and hence there is an additional implicit dependence on this in Eq.~(\ref{eq:betaerror}), via the eigenvalues $\hat{\epsilon}_k$.
In particular, if we for example take the limit of a particularly unconstraining dataset, setting $C_{ij}^{-1} \sim 0$, then we have $\hat{\epsilon}_k \sim T^2$.
This implies that the profiled
PDF errors would be independent of $T$, being unchanged and equal to the original ones, as consistency would dictate.
On the other hand, if we consider the case of a highly constraining dataset, where the first term in Eq.~\eqref{eq:hkl} dominates, the eigenvalues become independent of $T$ and the profiled
PDF uncertainties scale linearly with it.
For the more realistic situation where the impact of a new dataset is of a similar order of magnitude to the PDF errors of the baseline set, i.e. the explicit constraints from the new dataset enters at a similar level to the implicit constraints from the datasets which result in the baseline set, the scaling with $T$ will be somewhere in between.
Note that in this case the impact of the new dataset is also dictated by the size of the original tolerance as in Eq.~\eqref{eq:hkl}; the larger the tolerance, the smaller the corresponding impact.
This is entirely consistent with our physical expectation of the role the tolerance should play
within a Hessian global PDF determination.
As we will see below, the above considerations are relevant for the case of the LHeC dataset, where we will consider different values of the tolerance.
Unless otherwise stated, in the studies presented here, we use $T=3$, which roughly corresponds to the average
tolerance determined in the CT14 and MMHT14 analyses.
This is the same choice as the one we adopted for our PDF projections with the HL--LHC pseudo--data.
The resulting profiled PDF set\footnote{Note that sometimes in this paper we will for brevity use the shorthand `fit', but it always understood that a profiling has been performed rather than a full refit.} can be straightforwardly used
for phenomenology using the uncertainty prescription
of symmetric Hessian sets, and the default output format
is compliant with the {\tt LHAPDF} interface~\cite{Buckley:2014ana}.
\section{Results}\label{sect:results}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhechq.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhechq}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhechq.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhechq}
\caption{\small Impact of LHeC on the 1--$\sigma$ relative PDF uncertainties of the gluon, down quark, anti--up quark and strangeness distributions, with respect to the PDF4LHC15 baseline set.
%
In this comparison, shown at $Q=10$ GeV, we indicate the results of both profiling
with the inclusive LHeC measurements alone and also with the
semi-inclusive heavy quark structure functions.}\label{fig:pdf_lhechq}
\end{center}
\end{figure}
We now present the main results of this work, namely quantifying the impact of the LHeC pseudo--data on the
PDF4LHC15 baseline in different scenarios.
We begin by considering the impact of the LHeC measurements alone with respect to the
baseline, before combining these constraints with those provided by the HL--LHC.
In Fig.~\ref{fig:pdf_lhechq} we indicate the impact of the LHeC structure
function measurements, both for the inclusive dataset alone and in combination with the semi-inclusive
heavy quark datasets, on the PDF uncertainties of the gluon, down quark, anti--up quark and strangeness distributions.
We can see that the effect is in all cases significant, with the heavy quark data placing additional important constraints.
In the case of the gluon PDF, the inclusive pseudo--data already place significant constraints at low to intermediate $x$, through the precise direct measurements of their $Q^2$ slope, $\partial F_2 /\partial \ln Q^2$, as well as indirectly through the constraints on the quarks.
The heavy quark data further reduce the uncertainties on the gluon, with the impact at high $x$ from the charm and beauty structure functions being most significant here. The inclusive data have some impact on the down quark, in particular at high $x$, where the uncertainties on the dominantly valence distribution are significantly reduced.
A qualitatively similar, and somewhat larger, effect is found for the up quark, which we do not show here for brevity.
The impact on the anti--up from the inclusive structure function data is sizeable, in particular at lower $x$. For these quark distributions, the additional constraints from the heavy quark data is less pronounced, though far from negligible. For the strangeness, the inclusive data already place some constraints, which are significantly improved upon by the addition of the charged-current charm data.
It is interesting, in particular from the point of view of comparing with existing LHeC PDF projections,
to investigate the robustness of the results shown in Fig.~\ref{fig:pdf_lhechq} with respect
to the choice of tolerance $T$ used in the analysis.
As discussed in Sect.~\ref{sec:Hessian}, here we take as a baseline a tolerance value of $T=3$, to be roughly consistent with the underlying assumptions of the PDF4LHC15 baseline set.
However, for fits to DIS--only data, as in the LHeC pseudo--data and the existing HERAPDF2.0 sets~\cite{Abramowicz:2015mha}, a tolerance of $T=1$ is often taken, it being argued that in this case the underlying data are cleaner and self--consistent, allowing for this textbook value to be taken. On the other hand, it has been known for some time, see e.g.~\cite{Watt:2012tq}, that fits to the HERA dataset only are found not to be consistent within the quoted uncertainties in comparison to those including collider data when using a textbook tolerance $T=1$.
Indeed, the HERAPDF2.0 result for example for the up quark at high $x$ is found to be in clear tension with global fit results~\cite{Abramowicz:2015mha}, again indicating a tension between the underlying datasets, or more precisely the data/theory comparisons. Moreover, some degree of tension within the HERA dataset itself has also been observed, with the $e^-p$ CC data pulling in a different direction to other HERA measurements~\cite{Thorne:2015caa}, and the final combined charm and beauty structure function data~\cite{H1:2018flt} being in some tension with the inclusive data.
Nonetheless, it is instructive for the purposes of this exercise
to examine the further reduction in uncertainties when a tolerance of $T=1$ is assumed.
In Fig.~\ref{fig:pdf_tolerance} we show a similar comparison to that of
Fig.~\ref{fig:pdf_lhechq} for the complete LHeC dataset,
but with the baseline results obtained with a tolerance $T=1$,
together with those based on $T=3$ instead.
%
We can see that the difference is as expected largest where the LHeC data have more of an impact, generally leading to a further factor of around 2 reduction in the uncertainty.
On the other hand at higher $x$, where the impact of the LHeC data is smaller, the difference is consequently reduced.
Note however that this comparison has been carried out
for illustration purposes alone, as using $T=1$ is actually inconsistent with
the assumptions upon which the Hessian PDF sets in PDF4LHC15 are based.
%
Indeed, the PDF uncertainties in PDF4LHC15 set correspond to an approximate
value of the tolerance of $T\simeq 3$,
and certainly something larger than the textbook $T=1$ case.
%
This implies that taking $T=1$ for the LHeC pseudo--data
in effect corresponds to weighting such data preferentially with respect to the existing data that
leads to the PDF4LHC15 baseline set. This therefore does not accurately reproduce the situation of a
global PDF analysis.
A further point worth clarifying is whether the reduction in uncertainties when including data from the LHeC will be significantly less when compared to a more recent global PDF fit than PDF4LHC. To this end, in Fig.~\ref{fig:pdf_nnpdf} we compare the impact of the LHeC pseudodata on the PDF4LHC baseline to the result found by taking the newer NNPDF3.1~\cite{Ball:2017nwa} set, which in particular includes a range of more recent LHC data, from high precision $W$, $Z$ to differential top quark pair and the $Z$ boson transverse momentum distribution. To be precise, we take the NNLO symmetric Hessian version of this set, to allow us to perform the same profiling exercise as before. We can see that the baseline uncertainties are indeed smaller in comparison to the PDF4LHC case, though it should be emphasised that as the PDF4LHC combination will normally have larger errors than any of the three global inputs, this would also in general be true for the earlier NNPDF set entering into the combination. Even so, the results are qualitatively rather similar to the PDF4LHC case, and the impact of the LHeC data remains clear.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_Tcomp.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhec_Tcomp}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhec_Tcomp.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhec_Tcomp}
\caption{\small Same as Fig.~\ref{fig:pdf_lhechq} for the complete LHeC dataset,
now comparing the baseline results obtained with a tolerance $T=3$
with those based on $T=1$.}\label{fig:pdf_tolerance}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_NNPDF.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhec_NNPDF}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhec_NNPDF.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhec_NNPDF}
\caption{\small Same as Fig.~\ref{fig:pdf_lhechq} for the complete LHeC dataset,
with the result of profiling the NNPDF3.1 Hessian set also shown.}\label{fig:pdf_nnpdf}
\end{center}
\end{figure}
We now consider how the PDF impact of the LHeC pseudo--data compares
to that of the corresponding HL--LHC
projections reported in~\cite{Khalek:2018mdn}.
Moreover, we would also like to quantify the expected PDF uncertainty reduction
if both the HL--LHC and LHeC pseudo--data are simultaneously added to PDF4LHC15 by means
of profiling.
In Figs.~\ref{fig:pdf_lhec-hlllhc} and~\ref{fig:pdf_lhec-hlllhc-abs} we show a similar comparison to
that of Fig.~\ref{fig:pdf_lhechq}, including the LHeC results in addition to the HL--LHC projections
as well as their combination.
We can see
that at low $x$ the LHeC data place in general by far the strongest constraint, in particular for the gluon.
This is to be expected: the LHeC provides an outstanding coverage of the small-$x$ kinematical
region, as illustrated in Fig.~\ref{fig:kinplot}.
In the intermediate $x$ region, the HL--LHC and LHeC pseudo--data are found to place comparable constraints on the PDFs.
At higher $x$ the constraints are again comparable in size, with the HL--LHC resulting in a somewhat larger reduction in the gluon and strangeness uncertainty, while the LHeC has a somewhat larger impact for the down and anti-up quark distributions. To show this more clearly, in Fig.~\ref{fig:pdf_lhec-hlllhc_gluonlin} we show the same plot as before for the gluon PDF, but with a linear $x$ scale.
The combination of both HL--LHC and LHeC pseudo--data
nicely illustrate a clear and significant reduction in PDF uncertainties over a very wide range of $x$, improving upon the constraints from the individual datasets in a non--negligible way.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhec_HLLHCn}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhec_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhec_HLLHCn}
\caption{\small Same as Fig.~\ref{fig:pdf_lhechq}, now
comparing the impact of the LHeC pseudo--data with that of the HL--LHC projections
and to their combination.}\label{fig:pdf_lhec-hlllhc}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_HLLHClin.pdf}
\caption{\small Same as Fig.~\ref{fig:pdf_lhec-hlllhc}, for the gluon PDF alone, with a linear scale.}\label{fig:pdf_lhec-hlllhc_gluonlin}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-lhec_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-lhec_HLLHCn}
\includegraphics[width=0.48\linewidth]{fig/xubar-lhec_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-lhec_HLLHCn}
\caption{\small Same as Fig.~\ref{fig:pdf_lhec-hlllhc}, but with the error relative to each set shown.}\label{fig:pdf_lhec-hlllhc-abs}
\end{center}
\end{figure}
In order to further assess the PDF impact of the LHeC pseudo--data (alone or in combination
with the LHeC measurements), and in particular their relevance for LHC phenomenology, in Fig.~\ref{fig:lumi_lhec-hlllhc} we show the impact on the gluon--gluon, quark--gluon, quark--antiquark and quark--quark partonic luminosities.
In this comparison, we display the relative reduction of the PDF uncertainty in the luminosities compared to the baseline.
%
For example, a value of 0.2 in this plot indicates that the profiled PDF uncertainties are reduced
down to
20\% of the original ones.
Some clear trends are evident from this comparison, consistent with the results from the individual PDFs
shown in Figs.~\ref{fig:pdf_lhec-hlllhc} and~\ref{fig:pdf_lhec-hlllhc-abs}.
We can in particular observe that at low mass the LHeC places the dominant constraint, while at intermediate masses the LHeC and HL--LHC constraints are comparable in size, and at high mass the stronger constraint on the gluon--gluon and quark--gluon luminosities comes from the HL--LHC, with the LHeC dominating for the quark--quark and quark--antiquark luminosities.
As in the case of the PDFs, for the partonic luminosities the combination of the HL--LHC and LHeC constraints
leads to a clear reduction in the PDF uncertainties in comparison
to the individual cases, by up to an order of magnitude over a wide range of invariant masses, $M_X$, of the produced final state.
It is also worth emphasising that the LHeC and HL--LHC will have completely different experimental and theoretical systematics,
thus their complementarity would provide a particularly precious asset
to disentangle possible beyond the Standard Model (BSM) effects.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/ggERR_lhec_LheC_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/qgERR_lhec_LheC_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/qqERR_lhec_LheC_HLLHCn.pdf}
\includegraphics[width=0.48\linewidth]{fig/q2ERR_lhec_LheC_HLLHCn.pdf}
\caption{\small Impact of LHeC, HL--LHC and combined LHeC + HL--LHC pseudo--data on the uncertainties of the gluon--gluon, quark--gluon, quark--antiquark and quark--quark luminosities, with respect to the PDF4LHC15 baseline set.
%
In this comparison we display the relative reduction of the PDF uncertainty in the luminosities compared to the baseline.
%
}\label{fig:lumi_lhec-hlllhc}
\end{center}
\end{figure}
In summary,
the LHeC and HL--LHC datasets both place significant constraints on the PDFs, with some differences
depending on the kinematic region or the specific flavour combination being considered.
Most importantly, we find that these are rather complementary: while the LHeC places the most significant constraint at low to intermediate $x$ in general (though in the latter case the HL--LHC impact is often comparable in size), at high $x$ the HL--LHC places the dominant constraint on the gluon and strangeness, while the LHeC dominates for the up and down quarks.
Moreover, when both the LHeC and HL--LHC pseudo--data are simultaneously included
in the fit, all PDF flavours can be constrained across a wide range of $x$, providing a strong motivation
to exploit the input for PDF fits from both experiments, and therefore for realising the LHeC itself.
Finally, a few important caveats concerning this exercise should be mentioned.
First, the processes included for both the LHeC and HL--LHC, while broad in scope, are by no means exhaustive.
Most importantly, as mentioned in Sect.~\ref{sec:pseudo--data}, for the LHeC no jet production data are included, which
would certainly improve the constraint on the high-$x$ gluon. In addition, the inclusion of charm production in $e^+ p$ CC scattering would further constrain the strange quark.
In the case of the HL--LHC, only those processes which provide an impact at high-$x$ were included, and hence the lack of constraint at low-$x$ that is observed occurs essentially by construction.
In particular, there are a number of processes that will become available
with the legacy HL--LHC dataset, or indeed those in the current LHC dataset that are not currently included in global fits, but which can in principle constrain the low-$x$ PDFs, from low mass Drell--Yan to inclusive $D$ meson production~\cite{Zenaiev:2015rfa,Gauld:2016kpd} and exclusive vector meson photoproduction~\cite{Jones:2016ldq}, though here the theory is not available at the same level of precision to the LHeC case.
A further point of note is the value of the tolerance $T$ used in this analysis.
By performing a closure test using PDF4LHC15 as input, we are implicitly assuming that the final LHeC (and HL--LHC) data will be describable by this set, within the $T=3$ tolerance criteria to allow for some degree of tension among datasets.
That is, one is implicitly assuming that no greater tension between datasets than this will occur.
While this assumption is guided by previous experience with the wide range of measurements included in existing global PDF fits, it may turn out to be too strong and if this is the case we would expect the resulting PDF uncertainties to be larger, though at this point there is no strong motivation for believing that this will indeed be the case.
On the other hand, as shown in Fig.~\ref{fig:pdf_tolerance}, a more aggressive, smaller, choice for the LHeC leads to smaller uncertainties in that case, though the interpretation of such a choice in the context of this global fit to both HL--LHC and LHeC pseudo--data is far from clear,
being inconsistent with the assumptions used to construct the PDF4LHC15 baseline to begin with.
\section{Parameterisation dependence and LHeC--only fits}
\label{sec:tolstudy}
The main aim of this paper is to establish quantitatively the expected impact
that the availability of inclusive and heavy quark structure function measurements from the LHeC
would have on a state-of-the-art global PDF analysis.
Moreover, we also want to assess how the constraints provided by the LHeC
compare with those that will eventually be obtained
from measurements in proton-proton collisions at the HL--LHC.
To achieve these goals,
we have used as a baseline the PDFLHC15 set, which contains
$N_{\rm ev}=100$ symmetric Hessian eigenvectors, as constructed via a
MC combination~\cite{Carrazza:2015hva,Watt:2012tq} of the CT14~\cite{Dulat:2015mca},
MMHT14~\cite{Harland-Lang:2014zoa} and NNPDF3.0~\cite{Ball:2014uwa} sets.
This PDF4LHC15 set can therefore be
interpreted as arising from an underlying Hessian PDF set with $N_{\rm ev}$
free parameters determined from the experimental data.
In the current exercise we are performing a closure test, where the pseudodata are by construction in agreement with the theory generated using this PDF4LHC set. We are therefore by construction assuming that no additional parametric freedom will need to be introduced
in order to describe the (future) data under consideration, or more precisely, to describe the combination of global PDF fits to these data.
Such an assumption may turn out to be too strong, though as with the choice of tolerance above, there is currently no strong motivation for believing this will be true.
Nonetheless, a natural question to ask is the extent to which the type of projection studies we consider here are dependent on the flexibility of the parameterisation adopted in the baseline prior PDF set.
To explore this point further, we will consider the use of a baseline PDF prior set
based on a rather more restrictive parametrisation in comparison
to PDF4LHC15, specifically the HERAPDF2.0 NNLO set~\cite{Abramowicz:2015mha}.
In this HERAPDF case, there are only $\sim$ 14 free parameters, reflecting the lack of constraints coming from the HERA data alone on for example the detailed quark flavour decomposition.
To illustrate how this parametrization is less flexible than the one
used in global fits, we note for example that
the down quark valence and antiquark are parameterised in terms of only 3 free parameters, while the total
strangeness is assumed to be proportional to the antidown quark.
This is in contrast to the CT and MMHT sets, which have each between 2 and 3 times more free parameters in total, while the NNPDF parametric freedom is greater still.
There is therefore significantly less parametric freedom in the HERAPDF2.0 case in comparison to the PDF4LHC15 baseline.
We note in particular that in the the original LHeC studies of PDF
impact~\cite{AbelleiraFernandez:2012cc,Klein:1564929} a close variant of the HERAPDF set is adopted, in terms of PDF parameterisation and quark flavour assumptions, and then a full fit is performed
to the LHeC pseudo--data.
In more recent studies, however, some additional freedom has been included, for example no longer assuming that the anti--up and anti--down are equal to each other at small $x$~\cite{LHeCtalk1,LHeCtalk2}.
Taking into account these considerations, it follows that
when the LHeC pseudo--data are generated with this more restrictive
HERAPDF2.0 parametrization and flavour assumptions, one is by construction making a much stronger assumption that the future, very precise and wide ranging, LHeC data will be describable by such a restrictive parameterisation.
Under this assumption, it may be expected that the corresponding projected PDF uncertainties will be smaller than in our study.
To quantify this possibility, we have performed precisely the same profiling exercise as before to the same LHeC pseudo--data, but in this case using the HERAPDF2.0 NNLO set as the baseline rather than PDF4LHC15.
To be consistent with the HERAPDF methodology, we take a tolerance of $T=1$, and results are compared
to the profiling of PDF4LHC15 when $T=1$ is also used
(see Fig~\ref{fig:pdf_tolerance}), for a direct comparison.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_HERAPDF.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhec_HERAPDF}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhec_HERAPDF.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhec_HERAPDF}
\caption{\small Same as Fig.~\ref{fig:pdf_tolerance}, now
comparing the impact of the LHeC pseudo--data when added
on top of either the PDF4LHC15 or the HERAPDF2.0 sets.
%
In both cases, a tolerance of $T=1$ is used.
%
For HERAPDF2.0, only experimental uncertainties are taken
into account, while the model and parametric ones are not considered.
}\label{fig:pdf_lhec-herapdf}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-lhec_HERAPDF.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-lhec_HERAPDF}
\includegraphics[width=0.48\linewidth]{fig/xubar-lhec_HERAPDF.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-lhec_HERAPDF}
\caption{\small Same as Fig.~\ref{fig:pdf_lhec-herapdf}, but with the error relative to each set shown.
}\label{fig:pdf_lhec-herapdf-abs}
\end{center}
\end{figure}
In Figs.~\ref{fig:pdf_lhec-herapdf} and~\ref{fig:pdf_lhec-herapdf-abs} we display
a similar comparison to that of Fig.~\ref{fig:pdf_tolerance}, now
showing the impact of the LHeC pseudo--data when profiling either the PDF4LHC15 or the HERAPDF2.0 sets.
%
Here we consider only the so-called
`experimental' component of the PDF
uncertainties for the HERAPDF2.0 set,
that is, the one associated with the $N_{\rm ev}=14$
eigenvectors evaluated using the standard
$\Delta \chi^2 =1$ criterion.
From this comparison, a clear systematic trend is observed, with
the resultant PDF uncertainties corresponding to the profiled HERAPDF2.0
set lying significantly below the profiled PDF4LHC15 case.
This effect is most pronounced for the down quark and strangeness
distributions, which as discussed above are precisely those where
the parametric freedom in the HERAPDF2.0 case is the most restrictive.
For the gluon PDF, which in the HERAPDF2.0 case
has the largest freedom, with 5 free parameters,
from Figs.~\ref{fig:pdf_lhec-herapdf} and~\ref{fig:pdf_lhec-herapdf-abs}
we observe that
the differences are smaller though still rather significant.
In particular, at higher $x$, one finds that the PDF uncertainty
in the profiled HERAPDF2.0 case can be up to a factor 10 or more
smaller than when PDF4LHC15 is adopted as the prior set.
Although the uncertainty on the HERAPDF2.0 baseline is generally smaller
to begin with
than for PDF4LHC15, and one might therefore attribute some of the differences
in the profiled
uncertainties in the former case to this fact,
it is worth noting that at higher $x$,
where the baseline uncertainties are comparable in size, the difference
is still present and indeed largest.
This result thus provides a clear indication that the use of
such a restricted parameterisation as in the one adopted in the HERAPDF-based
prior will in general lead to smaller uncertainties in the final analysis
in comparison to the global fit case.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_HERAPDF_BASE.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhec_HERAPDF_BASE}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhec_HERAPDF_BASE.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhec_HERAPDF_BASE}
\caption{\small Same as Fig.~\ref{fig:pdf_lhec-herapdf}, where
now the baseline HERAPDF2.0 includes not only the
experimental PDF uncertainty, but also the corresponding
model and parametrization components.
}\label{fig:pdf_lhec-herapdfbase}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-lhec_HERAPDF_BASE.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-lhec_HERAPDF_BASE}
\includegraphics[width=0.48\linewidth]{fig/xubar-lhec_HERAPDF_BASE.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-lhec_HERAPDF_BASE}
\caption{\small Same as Fig.~\ref{fig:pdf_lhec-herapdfbase}, but with the error relative to each set shown.
}\label{fig:pdf_lhec-herapdfbase-abs}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-lhec_HERAPDF_prior_comb.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-lhec_HERAPDF_prior_comb.pdf}
\includegraphics[width=0.48\linewidth]{fig/xubar-lhec_HERAPDF_prior_comb.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-lhec_HERAPDF_prior_comb.pdf}
\caption{\small Impact of LHeC on the PDF uncertainties of the gluon, down quark, anti--up quark and strangeness distributions, with respect to the PDF4LHC15 and HERAPDF baselines, when the relative impact of the initial PDF prior, that is, of the datasets included in the baseline PDF sets, is reduced by a factor of 100. For the HERAPDF case we show results corresponding to the experimental uncertainties only baseline and that including model and parameterisation uncertainties. A tolerance of $T=1$ is taken in all profiled cases, while the result of the PDF4LHC set (including prior but with no LHeC data) is shown for comparison.
}\label{fig:pdf_lhec-prior}
\end{center}
\end{figure}
To investigate this effect further, in Figs.~\ref{fig:pdf_lhec-herapdfbase} and~\ref{fig:pdf_lhec-herapdfbase-abs} we show the same comparison to that of Figs.~\ref{fig:pdf_lhec-herapdf} and~\ref{fig:pdf_lhec-herapdf-abs}, but in this case also including
the `model' and `parameterisation' components
of the PDF uncertainties for the HERAPDF2.0 set.
Specifically, we have slightly modified the underlying PDF set to include these uncertainties in a symmetric Hessian format, rather than in the native format
that requires an envelope prescription to combine them with the
experimental component.
These model and parameterisation components account for the effects
of some additional parametric freedom, as well as for
the variation of certain model parameters such as the heavy quark masses, the starting scale $Q_0$ and so on.
From this comparison,
we can observe that once the additional components are
accounted for, the total PDF
uncertainties for the baseline HERAPDF2.0 set are now larger, and more in line with the PDF4LHC15 case.
We then
find that the profiled uncertainties are indeed larger than in the previous case, reflecting the additional freedom in the baseline, and lying closer to the PDF4LHC15 result.
Nonetheless, there is still a clear trend for these uncertainties to lie systematically below the PDF4LHC15-based projections, reflecting the fact that even after including these additional uncertainties, the underlying parametric freedom is significantly less in the HERAPDF2.0 case.
The only exception is in the low $x$ gluon (and at very low $x$ in the anti-up), where curiously the uncertainty in the HERAPDF2.0 case is in fact somewhat larger. We can find no convincing explanation of this fact, so simply leave it as an observation.
Now, in the above comparison the varying freedom in the parameterisation is not the only difference between the two baseline sets, which in addition fit to rather different underlying datasets.
To clarify this, we also consider the result of reducing the relative
impact of the PDF prior in the profiling, that is,
the second term in \eqref{eq:hkl}, by a significant factor of $100$, for both HERAPDF and PDF4LHC baselines.
This should effectively remove the impact of the data that are
used as input to construct the baseline sets, and
their underlying PDF uncertainties. In other words, the {\it only} difference between these results should be in the underlying parameterisation flexibility that one assumes will be sufficient to describe the LHeC data.
Moreover, as the effect of the prior datasets have been removed, in such a comparison the constraints will now be purely due to the LHeC dataset and hence in the PDF4LHC case no longer correspond to a global PDF analysis, but rather to a pure LHeC--only one, using the PDF4LHC parameterisation. We recall in particular that these comparisons are performed with a tolerance $T=1$, and hence are directly comparable to existing LHeC--only studies~\cite{AbelleiraFernandez:2012cc,AbelleiraFernandez:2012ty,Paukkunen:2017phq,Cooper-Sarkar:2016udp}.
The results are shown in Figs.~\ref{fig:pdf_lhec-prior} and~\ref{fig:pdf_lhec-prior-err}, including the PDF4LHC case with the original prior to assess the impact relative to the current PDF determination from global fits. For the PDF4LHC parameterisation, the results of the LHeC--only fit are in general significant, resulting in uncertainties that are up to an order of magnitude smaller. The one exception is the down quark, where the uncertainties are comparable; this is consistent with the fact that the determination from DIS data alone is known to be less significant in this case. Comparing with the HERAPDF case, we can again see that in general the projected uncertainties resulting from the HERAPDF baseline is significantly smaller in comparison to the PDF4LHC case. This is most marked for the `experimental' HERAPDF baseline, while for the set including model and parameterisation uncertainties the PDF uncertainties are seen to be somewhat larger, in particular in the regions less constrained by data.
A closer comparison shows that the difference between the PDF4LHC and HERAPDF errors is somewhat larger in this case, then when the baseline prior is included.
This is perhaps not surprising: as the PDF4LHC15 baseline set fits to a larger dataset, the impact of reducing the prior in this case should be to increase the PDF uncertainties by more than in the HERAPDF2.0 case, leading to a larger difference.
To summarise our findings, we have demonstrated in this section how
in general the quantitative
interpretation of the PDF projections based on LHeC pseudo--data
depends sensitively on the choice of input dataset, parametric flexibility,
and flavour assumptions that enter the construction of the prior PDF fit.
A more flexible parametrization, such as the one used in global PDF fits,
as required to give a satisfactory description of all available measurements
from lepton-hadron and hadron-hadron collisions, will in general result in
larger profiled uncertainties than in the case of a HERAPDF2.0 or
similar baseline.
However, the latter results adopt
the implicit strong assumption that the parametric flexibility of
HERAPDF2.0 will be sufficient to describe all future precision
measurements, including those from the LHeC.
We note that such an
assumption is found to be unjustified already when one
accounts for the existing LHC data.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig/xg-ERR-lhec_HERAPDF_prior_comb.pdf}
\includegraphics[width=0.48\linewidth]{fig/xd-ERR-lhec_HERAPDF_prior_comb.pdf}
\includegraphics[width=0.48\linewidth]{fig/xubar-ERR-lhec_HERAPDF_prior_comb.pdf}
\includegraphics[width=0.48\linewidth]{fig/xsp-ERR-lhec_HERAPDF_prior_comb.pdf}
\caption{\small As in Fig.~\ref{fig:pdf_lhec-prior} but with the relative errors shown.}\label{fig:pdf_lhec-prior-err}
\end{center}
\end{figure}
\section{Summary and outlook}\label{sec:conc}
In this study, we have quantified the expected information that the realisation
of the Large Hadron electron Collider (LHeC) at CERN would provide on our knowledge
of the quark and gluon structure of the proton.
This study, based on the pseudodata presented in~\cite{mklein}, is the first time that the impact
of the LHeC measurements has been assessed in the context of a global PDF analysis.
Our results complement and extend our previous study of the PDF projections based
on HL--LHC pseudo-data, and provide a compelling picture of how much better
our understanding of the parton distribution functions can become through the
combination of these
two facilities, one already fully approved (HL--LHC) and the second under consideration (LHeC).
In essence, we have assessed the `ultimate' precision that can be expected
for PDFs from experiments at CERN alone by 2035.
Of course, other related theory developments, such as for example progress
in lattice QCD calculations~\cite{Lin:2017snn,Hobbs:2019gob},
may well have a significant impact in addition.
Our results demonstrate that the LHeC can improve our current precision on PDFs significantly,
with uncertainties dominated by experimental systematics, as statistical errors quickly
become negligible.
For those poorly constrained flavours, e.g., the gluon at both small-$x$
and large-$x$, the sea quark flavour separation,
and the total strangeness, the PDF uncertainties can be expected to be reduced
by up to an order of magnitude.
In particular, we have shown
how the availability of the strange, charm,
and bottom heavy-quark production data play an important role in constraining
the gluon and strange PDFs.
In addition, further LHeC processes not considered here, such as jet production,
could provide additional information in particular on the gluon PDF.
By comparing with the corresponding PDF projections
based on HL-LHC pseudo--data, we find that the LHeC measurements would place stronger constraints
in general from the small-$x$ to the intermediate-$x$ regions for most flavours.
In relation to this, it is also worth emphasising that beyond studies directly relating to PDF constraints, this high precision probe of the low--$x$ region will in particular provide a unique environment to study novel QCD phenomena, such as BFKL and saturation effects.
In the large-$x$ region, which is of course crucial
for BSM searches,
the LHeC and HL-LHC impact is broadly found to be comparable in size, with the HL--LHC resulting in a somewhat larger reduction in the gluon and strangeness uncertainty, while the LHeC has a somewhat larger impact for the down and anti-up quark distributions.
Furthermore, the combination of both the LHeC and HL-LHC pseudo--data leads
to a significantly superior PDF error reduction in comparison to the two facilities individually.
This could become particularly crucial for the interpretation
of possible anomalies in the large $p_T$ region, that might indicate the presence
of new physics beyond the SM.
Moreover, as the complete LHeC dataset
that has been used in this analysis is dominated by systematic errors,
our results strongly suggest that smaller LHeC datasets, which are expected to have a similar
PDF impact as the full
legacy dataset, could be exploited already at relatively early stages in HL--LHC running.
In this study we have also considered the robustness
of our results when using alternative PDF priors in the
profiling, in particular the HERAPDF2.0 PDF set that is based on a more
restrictive parameterisation in comparison to the global sets that enter
the PDF4LHC15 combination.
In this case, the LHeC pseudo--data in general leads to significantly smaller PDF uncertainties in comparison
to the results based on the PDF4LHC15 prior.
This comparison reveals the fact that, when interpreting the projections
for future LHeC and HL-LHC measurements,
the input PDF functional forms should be adjusted to make sure that any possible
parametrization bias and flavour assumptions are minimised.
As a word of caution, in this study we have ignored any possible issues such as data incompatibilities,
limitations of the theoretical calculations, or issues affecting the data
correlation models.
These are already common in PDF fits, but can only be addressed when carrying out
a global fit with real data.
The results presented in this study (as well as in any other projection)
can therefore only be interpreted as providing an estimate
of the possible precision on the PDFs that can be expected from these
future facilities.
In addition, as mentioned before, there are other possible LHeC measurements that have not
been considered here,
such as inclusive jet production, which can place further important constraints
on the PDFs.
Finally, it would be interesting to study in the future the extent to which our
results might change if a full fit, rather than a profiling study, is carried out.
However, it should be emphasised that within the context of the closure tests we are considering here,
where data and theory agree by construction, profiling is known to provide a very good
approximation to the true result.
Finally, one important point that we have not discussed in detail here is the potential for BSM contamination of HL--LHC data at high $x$. Our projections, which are based entirely on a closure test, by construction assume that the future HL--LHC (and LHeC) data will be describable by SM theory. However, it is well known that BSM physics may enter various LHC datasets, in particular at higher $p_\perp$ and invariant masses, and hence may be incorrectly absorbed into PDF fits. A detailed study of these effects is beyond the scope of the current paper, but clearly in future fits it will be essential to disentangle BSM and QCD effects. A first recent study~\cite{Carrazza:2019sec} of a fit in the framework of the SM Effective Field Theory (SMEFT) has demonstrated how this may be achieved, by using the different scaling with the energy of BSM and QCD effects as a means to separate them, complemented with other statistical estimators. Using a similar approach, it should be possible to identify potential BSM effects that might be present in the tails of LHC distributions without the risk of reabsorbing them into the PDFs. Nonetheless, this is clearly a delicate issue, and something that will be largely absent in the case of the LHeC. This provides a further strong motivation for input from such a machine in PDF fits.
The results of this study are made
publicly available in the {\tt LHAPDF6} format~\cite{Buckley:2014ana} by means
of {\tt Zenodo} data repository:
\begin{center}
\url{https://zenodo.org/record/3250580}
\end{center}
Specifically, the following PDF sets can be obtained from this repository:
\begin{center}
{\tt PDF4LHC15\_nnlo\_lhec}\\
{\tt PDF4LHC15\_nnlo\_hllhc\_scen1\_lhec}\\
{\tt PDF4LHC15\_nnlo\_hllhc\_scen2\_lhec}\\
{\tt PDF4LHC15\_nnlo\_hllhc\_scen3\_lhec}
\end{center}
where the first set corresponds to the profiling of PDF4LHC15 using the
entire LHeC dataset listed in Table~\ref{tab:lhecdat}, and the other three
correspond to the simultaneous profiling with both the LHeC and HL--LHC
pseudo--data, for the three different projections of the experimental systematic
errors.
%
In addition,
in the same repository one can also find
the corresponding projections based only on the HL--LHC pseudo--data:
\begin{center}
{\tt PDF4LHC15\_nnlo\_hllhc\_scen1}\\
{\tt PDF4LHC15\_nnlo\_hllhc\_scen2}\\
{\tt PDF4LHC15\_nnlo\_hllhc\_scen3}
\end{center}
In this way, the PDF projections presented in this work can be straightforwardly compared
to other related projections,
and used in various phenomenological
applications, for example in the context
of feasibility studies for future colliders both for lepton-hadron
and hadron-hadron scattering.
\section*{Acknowledgments}
We thank Max Klein for providing us with the pseudo-data
for the LHeC projections.
%
We are grateful to Claire Gwenlan, Francesco Giuli,
and Gavin Pownall for discussions
about the LHeC pseudo--data and the {\tt xFitter}-based PDF projections.
We are also grateful to Nestor Armesto and Fred Olness for
discussions in the context of the LHeC QCD and small-$x$ working
group.
\paragraph{Funding information}
S.~B. acknowledges financial support from the UK Science and Technology Facilities
Council.
L.~H.~L thanks the Science and Technology Facilities Council (STFC)
for support via grant award ST/L000377/1.
R.~A.~K. and J.~R. are
supported by the European Research Council (ERC) Starting Grant ``PDF4BSM'' and by the Dutch
Organization for Scientific Research (NWO).
The work of J.~G. was sponsored by the National Natural
Science Foundation of China under the Grant No. 11875189 and No.11835005.
\input{LHeC_PDFs_sp.bbl}
\nolinenumbers
\end{document}
|
\section{Introduction}
In this article, we parameterize an equivalent medium, resulting from the~\citet{Backus1962} average, with respect to linear inhomogeneity parameters of its constituent isotropic layers.
Sedimentary basins often show thin layers that induce seismic anisotropy.
The parametrization allows us to make a relationship between the anisotropy to layer inhomogeneity, which is, in seismological context, an important relation.
We also assume the medium possesses linear inhomogeneity, which was first introduced by~\citet{Slotnick1959}.
The traveltime expression from a source to a receiver in a linearly inhomogeneous media is shown by~\citet{SlawinskiSlawinski1999}.
To obtain a solution from the analytical relationship between the anisotropy and the linear inhomogeneity parameters,
we require to provide one of the inhomogeneity parameter.
We use two independent seismological methods, 1-D tomography and $ab$ model, to make the analytical relation to be well posed.
The outcomes of the seismological methods are a pair of inhomogeneity parameters,
using one of the parameter from each method allows us to obtain the solution of the analytical relation.
The resultant solution from the analytical relation then can be examined by comparing with the results of the seismological methods.
\section{Equivalent medium parameterization}
We parameterize a transversely isotropic equivalent medium resulting from the~\citeauthor{Backus1962} average of thin, intrinsically homogeneous, isotropic layers, which we refer to as the~{\it\citeauthor{Backus1962} medium}.
We assume that a stack of such constituent layers is inhomogeneous and possesses a constant-velocity gradient that increases linearly with depth~\citep[see e.g.,][]{SlawinskiSlawinski1999}.
Specifically, for both $P$ and $S$ waves,
\begin{equation}
\label{eq:vPvS}
v_{P}\left(z\right) = a_{P} + b_{P}\,z
\quad{\rm and}\quad
v_{S}(z) = a_{S} + b_{S}\,z
\,,
\end{equation}
where~$a_{P,S}$ are the wavespeeds at the top of the medium,~$b_{P,S}$ are positive velocity-gradient constants, and $z$ is the depth.
\subsection{\citeauthor{Backus1962} average}
Following the definition of~\citet[Section~3]{Backus1962}, the average of a function $f(z)$ of width $\ell'$ is the moving average given by
\begin{equation}
\label{eq:BackusAverage}
\overline{f}(z)
=
\int\limits_{-\infty}^{\infty}
w\left(\zeta - z\right)
f\left(\zeta\right)
{\rm d}\zeta
\,,
\end{equation}
where the properties of the weighting function are
\begin{equation*}
w\left(z\right)
\geqslant
0
\,,\quad
w\left(\pm\infty\right)
=
0
\,,\quad
\int\limits_{-\infty}^{\infty}
w\left(z\right){\rm d}z
=
1
\,,\quad
\int\limits_{-\infty}^{\infty}
z\,w\left(z\right){\rm d}z
=
0
\,,\quad
\int\limits_{-\infty}^{\infty}
z^{2}\,w\left(z\right){\rm d}z
=
\left(\ell'\right)^{2}
\,.
\end{equation*}
The result of performing average~\eqref{eq:BackusAverage} on isotropic layers results is a homogeneous TI medium, where the corresponding elasticity parameters, which are referred to as {\it Backus parameters}, are
\begin{subequations}
\begin{align}
\label{eq:cTI1111}
c^{\overline{\rm TI}}_{1111}
&=
\overline{
\left(\dfrac{c_{1111}-2\,c_{2323}}{c_{1111}}\right)
}^{\,2}\,\,
\overline{\left(\dfrac{1}{c_{1111}}\right)}^{\,-1} +
\overline{
\left(
\dfrac{
4\left(c_{1111} - c_{2323}\right)c_{2323}
}{
c_{1111}
}
\right)
}
\,,
\\
\label{eq:cTI1122}
c^{\overline{\rm TI}}_{1122}
&=
\overline{
\left(\dfrac{c_{1111}-2\,c_{2323}}{c_{1111}}\right)
}^{\,2}\,\,
\overline{\left(\dfrac{1}{c_{1111}}\right)}^{\,-1} +
\overline{
\left(
\dfrac{
2\left(c_{1111} - 2\,c_{2323}\right)c_{2323}
}{
c_{1111}
}
\right)
}
\,,
\\
\label{eq:cTI1133}
c^{\overline{\rm TI}}_{1133}
&=
\overline{\left(\dfrac{c_{1111}-2\,c_{2323}}{c_{1111}}\right)}\,\,
\overline{\left(\dfrac{1}{c_{1111}}\right)}^{\,-1}
\,,
\\
\label{eq:cTI1212}
c^{\overline{\rm TI}}_{1212}
&=
\overline{c_{2323}}
\,,
\\
\label{eq:cTI2323}
c^{\overline{\rm TI}}_{2323}
&=
\overline{\left(\dfrac{1}{c_{2323}}\right)}^{\,-1}
\,,
\\
\label{eq:cTI3333}
c^{\overline{\rm TI}}_{3333}
&=
\overline{\left(\dfrac{1}{c_{1111}}\right)}^{\,-1}
\,.
\end{align}
\end{subequations}
Since the weighting function, $w$\,, in integral~\eqref{eq:BackusAverage}, is continuous and symmetric, the~\citeauthor{Backus1962} average may be written as a weighted average~\citep[e.g.,][Section~4.2.2]{Slawinski2018}.
Herein, the~\citeauthor{Backus1962} average is weighted by layer thickness.
For density-scaled VSP measurements, $v_{P} = \sqrt{c_{1111}}$ and $v_{S} = \sqrt{c_{2323}}$\,; this allows for a reparameterization of parameters~\eqref{eq:cTI1111}--\eqref{eq:cTI3333} in terms of linear-inhomogeneity parameters~\eqref{eq:vPvS}.
For example, Backus parameter~\eqref{eq:cTI1212} may be rewritten as
\begin{align}
\nonumber
c^{\overline{\rm TI}}_{1212}
=
\overline{c_{2323}}
=
\frac{1}{h_{2}-h_{1}}
\int\limits_{h_{1}}^{h_{2}}
c_{2323}\,
{\rm d}z
&=
\frac{1}{h_{2}-h_{1}}
\int\limits_{h_{1}}^{h_{2}}
\left(a_{S} + b_{S}\,z\right)^{2}
{\rm d}z
\\
\label{eq:AnIn_cTI1212}
&=
\frac{1}{3}
\left(
3\,a_{S}^{2} +
3\,a_{S}\,b_{S}\left(h_{1} + h_{2}\right) +
b_{S}^{2}\left(h_{1}^{2} + h_{1}\,h_{2} + h_{2}^{2}\right)
\right)
\,.
\end{align}
For Backus parameter~\eqref{eq:cTI1111}, we begin with the first term, where
\begin{equation}
\label{eq:AnIn_cTI1111_t1}
\overline{\left(\dfrac{c_{1111}-2\,c_{2323}}{c_{1111}}\right)}
=
\dfrac{1}{h_{2}-h_{1}}
\int\limits_{h_{1}}^{h_{2}}
\left(
1
-
2\,\dfrac{
{v_{S}}^{2}
}{
{v_{P}}^{2}
}
\right)
{\rm d}z
=
\dfrac{1}{h_{2}-h_{1}}
\int\limits_{h_{1}}^{h_{2}}
\left(
1
-
2\,\dfrac{
\left(a_{S} + b_{S}\,z\right)^{2}
}{
\left(a_{P} + b_{P}\,z\right)^{2}
}
\right)
{\rm d}z
=
1-\dfrac{2\,I_{1}}{h_{2} - h_{1}}
\end{equation}
and
\begin{align}
\nonumber
I_{1}
=
\int\limits_{h_{1}}^{h_{2}}
\dfrac{
\left(a_{S}+b_{S}\,z\right)^{2}
}{
\left(a_{P}+b_{P}\,z\right)^{2}
}\,
{\rm d}z
&=
\dfrac{h_{2}\,{b_{S}}^2}{{b_{P}}^2} -
\dfrac{h_{1}\,{b_{S}}^2}{{b_{P}}^2}
\\
\nonumber
&+
\dfrac{
\ln\left(a_{P} + h_{1}\,b_{P}\right)
\left(
2\,a_{P}\,{b_{S}}^{2} -
2\,a_{S}\,b_{P}\,b_{S}
\right)
}{
{b_{P}}^{3}
}
-
\dfrac{
\ln\left(a_{P}+h_{2}\,b_{P}\right)
\left(
2\,a_{P}\,{b_{S}}^{2} -
2\,a_{S}\,b_{P}\,b_{S}
\right)
}{
{b_{P}}^3
}
\\
\label{eq:I1}
&+
\dfrac{
{a_{P}}^{2}\,{b_{S}}^2 -
2\,a_{P}\,a_{S}\,b_{P}\,b_{S} +
{a_{S}}^2\,{b_{P}}^2
}{
b_{P}
\left(
h_{1}\,{b_{P}}^{3} +
a_{p}\,{b_{P}}^{2}
\right)
}
-
\dfrac{
{a_{P}}^2\,{b_{S}}^{2} -
2\,a_{P}\,a_{S}\,b_{P}\,b_{S} +
{a_{S}}^2\,{b_{P}}^2
}{
b_{P}
\left(h_{2}\,{b_{P}}^{3} + a_{P}\,{b_{P}}^{2}\right)
}
\,.
\end{align}
For the second term in parameter~\eqref{eq:cTI1111},
\begin{equation*}
\overline{\left(\dfrac{1}{c_{1111}}\right)}^{\,-1}
=
\overline{\left(\dfrac{1}{v_{P}^{2}}\right)}^{\,-1}
=
\left(
\dfrac{1}{h_{2}-h_{1}}
\int\limits_{h_{1}}^{h_{2}}
\dfrac{1}{\left(a_{P} + b_{P}\,z\right)^{2}}\,
\mathrm{d}z
\right)^{\!\!\!-1}
=
\left(h_{2} - h_{1}\right)
\left(\,\,
\int\limits_{h_{1}}^{h_{2}}
\left(a_{P} + b_{P}\,z\right)^{-2}\,
\mathrm{d}z
\right)^{\!\!\!-1}
\,.
\end{equation*}
Substituting $u$ for~$a_{P}+b_{P}z$\,, which gives $\mathrm{d}z = \dfrac{\mathrm{d}u}{b_{P}}$\,, and changing the limits, $h_{1}\to a_{P} + b_{P}\,h_{1}$ and $h_{2}\to a_{P} + b_{P}\,h_{2}$\,, we obtain
\begin{align}
\nonumber
\overline{\left(\dfrac{1}{c_{1111}}\right)}^{\,-1}
&=
\left(h_{2} - h_{1}\right)
\left(\,\,
\int\limits_{a_{P} + b_{P}\,h_{1}}^{a_{P} + b_{P}\,h_{2}}
u^{-2}
\dfrac{\mathrm{d}u}{b_{P}}
\right)^{\!\!\!-1}
=
\left(h_{2} - h_{1}\right)
\left(
\left.
\dfrac{-u^{-1}}{b_{P}}
\right|_{a_{P} + b_{P}\,h_{1}}^{a_{P} + b_{P}\,h_{2}}
\right)^{\!\!\!-1}
\\
\nonumber
&=
\left(h_{2} - h_{1}\right)
\left(-b_{P}\right)
\left(
\left(a_{P} + b_{P}\,h_{2}\right)^{-1} -
\left(a_{P} + b_{P}\,h_{1}\right)^{-1}
\right)^{-1}
\\
\label{eq:AnIn_cTI3333}
&=
\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
\,.
\end{align}
For the last term in parameter~\eqref{eq:cTI1111},
\begin{equation}
\label{eq:AnIn_cTI1111_t3}
\overline{
\left(
\dfrac{
4\left(c_{1111} - c_{2323}\right)c_{2323}
}{
c_{1111}
}
\right)
}
=
\dfrac{4}{h_{2}-h_{1}}
\left(\,\,
\int\limits_{h_{1}}^{h_{2}}
\left(a_{S} + b_{S}\,z\right)^{2}
{\rm d}z
-
\int\limits_{h_{1}}^{h_{2}}
\dfrac{
\left(a_{S} + b_{S}\,z\right)^{4}
}{
\left(a_{P} + b_{P}\,z\right)^{2}
}\,
{\rm d}z
\right)
=
\dfrac{4}{h_{2}-h_{1}}\left(I_{2} - I_{3}\right)
\,,
\end{equation}
where
\begin{equation}
\label{eq:I2}
I_{2}
=
\int\limits_{h_{1}}^{h_{2}}
\left(a_{S} + b_{S}\,z\right)^{2}\,
{\rm d}z
=
\frac{1}{3}
\left(h_{2} - h_{1}\right)
\left(
3\,a_{S}{}^2 +
3\,a_{S}\,b_{S}\left(h_{1} + h_{2}\right) +
b_{S}{}^{2}
\left(h_{1}^{2} + h_{1}\,h_{2} + h_{2}^{2}\right)
\right)
\end{equation}
and
\begin{align}
\nonumber
I_{3}
=
\int\limits_{h_{1}}^{h_{2}}
\dfrac{
\left(a_{S} + b_{S}\,z\right)^{4}
}{
\left(a_{P} + b_{P}\,z\right)^{2}
}\,
{\rm d}z
&=
h_{2}
\left(
\dfrac{
2\,a_{p}
\left(
\dfrac{2\,a_{P}\,{b_{S}}^{4}}{{b_{P}}^{3}} -
\dfrac{4\,a_{S}\,{b_{S}}^{3}}{{b_{P}}^{2}}
\right)
}{
b_{P}
}
-
\dfrac{{a_{P}}^{2}\,{b_{S}}^{4}}{{b_{P}}^{4}} +
\dfrac{6\,{a_{S}}^{2}\,{b_{S}}^{2}}{{b_{P}}^{2}}
\right)
\\\nonumber
&-
h_{1}
\left(
\dfrac{
2\,a_{P}
\left(
\dfrac{2\,a_{P}\,{b_{S}}^{4}}{{b_{P}}^{3}} -
\dfrac{4\,a_{S}\,{b_{S}}^{3}}{{b_{P}}^{2}}
\right)
}{
b_{P}
}
-
\dfrac{{a_{P}}^{2}\,{b_{S}}^{4}}{{b_{P}}^{4}} +
\dfrac{6\,{a_{S}}^{2}\,{b_{S}}^{2}}{{b_{P}}^{2}}
\right)
\\\nonumber
&+
{h_{1}}^2
\left(
\dfrac{a_{P}\,{b_{S}}^{4}}{{b_{P}}^{3}} -
\dfrac{2\,a_{S}\,{b_{S}}^{3}}{{b_{P}}^{2}}
\right)
-
{h_{2}}^2
\left(
\dfrac{a_{P}\,{b_{S}}^{4}}{{b_{P}}^{3}} -
\dfrac{2\,a_{S}\,{b_{S}}^{3}}{{b_{P}}^{2}}
\right)
\\\nonumber
&+
\dfrac{
\ln\left(a_{P} + h_{1}\,b_{P}\right)
\left(
4\,{a_{P}}^{3}\,{b_{S}}^{4} -
12\,{a_{P}}^{2}\,a_{S}\,b_{P}\,{b_{S}}^{3} +
12\,a_{P}\,{a_{S}}^{2}\,{b_{P}}^{2}\,{b_{S}}^{2} -
4\,{a_{S}}^{3}\,{b_{P}}^{3}\,b_{S}
\right)
}{
{b_{P}}^5
}
\\\nonumber
&-
\dfrac{
\ln\left(a_{P} + h_{2}\,b_{P}\right)
\left(
4\,{a_{P}}^{3}\,{b_{S}}^{4} -
12\,{a_{P}}^{2}\,a_{S}\,b_{P}\,{b_{S}}^{3} +
12\,a_{P}\,{a_{S}}^{2}\,{b_{P}}^{2}\,{b_{S}}^{2} -
4\,{a_{S}}^{3}\,{b_{P}}^{3}\,b_{S}
\right)
}{
{b_{P}}^5
}
\\\nonumber
&-
\dfrac{{h_{1}}^{3}\,{b_{S}}^{4}}{3\,{b_{P}}^{2}} +
\dfrac{{h_{2}}^{3}\,{b_{S}}^{4}}{3\,{b_{P}}^{2}}
\\\nonumber
&+
\dfrac{
{a_{P}}^{4}\,{b_{S}}^{4} -
4\,{a_{P}}^{3}\,a_{S}\,b_{P}\,{b_{S}}^{3} +
6\,{a_{P}}^{2}\,{a_{S}}^{2}\,{b_{P}}^{2}\,{b_{S}}^{2} -
4\,a_{P}\,{a_{S}}^{3}\,{b_{P}}^{3}\,b_{S} +
{a_{S}}^{4}\,{b_{P}}^{4}
}{
b_{P}
\left(h_{1}\,{b_{P}}^{5} + a_{P}\,{b_{P}}^{4}\right)
}
\\\label{eq:I3}
&-
\dfrac{
{a_{P}}^{4}\,{b_{S}}^{4} -
4\,{a_{P}}^{3}\,a_{S}\,b_{P}\,{b_{S}}^{3} +
6\,{a_{P}}^{2}\,{a_{S}}^{2}\,{b_{P}}^{2}\,{b_{S}}^{2} -
4\,a_{P}\,{a_{S}}^{3}\,{b_{P}}^{3}\,b_{S} +
{a_{S}}^{4}\,{b_{P}}^{4}
}{
b_{P}
\left(h_{2}\,{b_{P}}^{5} + a_{P}\,{b_{P}}^{4}\right)
}
\,.
\end{align}
The first two terms in parameter~\eqref{eq:cTI1122} are given by formul\ae~\eqref{eq:AnIn_cTI1111_t1} and~\eqref{eq:AnIn_cTI3333}, whereas the third term is
\begin{equation}
\label{eq:AnIn_cTI1122_t3}
\overline{
\left(
\dfrac{
2\left(c_{1111} - 2\,c_{2323}\right)c_{2323}
}{
c_{1111}
}
\right)
}
=
\frac{2}{h_{2}-h_{1}}
\left(\,\,
\int\limits_{h_{1}}^{h_{2}}
\left(a_{S}+b_{S}\,z\right)^{2}
{\rm d}z
-
2
\int\limits_{h_{1}}^{h_{2}}
\frac{
\left(a_{S} + b_{S}\,z\right)^{4}
}{
\left(a_{P} + b_{P}\,z\right)^{2}
}\,
{\rm d}z
\right)
=
\frac{2}{h_{2}-h_{1}}
\left(I_{2} - 2\,I_{3}\right)
\,,
\end{equation}
where $I_{2}$ and $I_{3}$ are given by integration constants~\eqref{eq:I2} and~\eqref{eq:I3}.
Finally, in a manner similar to obtaining the second term in parameter~\eqref{eq:cTI1111}, we use $u$ substitution and change limits of integration to obtain the term in parameter~\eqref{eq:cTI2323}, where
\begin{align}
\nonumber
\overline{\left(\dfrac{1}{c_{2323}}\right)}^{\,-1}
&=
\left(
\frac{1}{h_{2}-h_{1}}
\int\limits_{h_{1}}^{h_{2}}
\dfrac{1}{\left(a_{S} + b_{S}\,z\right)^{2}}\,
{\rm d}z
\right)^{\!\!\!-1}
=
\left(h_{2} - h_{1}\right)
\left(\,\,
\int\limits_{h_{1}}^{h_{2}}
\left(a_{S} + b_{S}\,z\right)^{-2}\,
\mathrm{d}z
\right)^{\!\!\!-1}
\\
\label{eq:AnIn_cTI2323}
&=
\left(a_{S} + b_{S}\,h_{1}\right)
\left(a_{S} + b_{S}\,h_{2}\right)
\,.
\end{align}
Thus, using formul\ae~\eqref{eq:AnIn_cTI1212},~\eqref{eq:AnIn_cTI1111_t1},~\eqref{eq:AnIn_cTI3333},~\eqref{eq:AnIn_cTI1111_t3},~\eqref{eq:AnIn_cTI1122_t3},~\eqref{eq:AnIn_cTI2323}, along with integration constants~\eqref{eq:I1},~\eqref{eq:I2},~\eqref{eq:I3}, we may restate Backus parameters~\eqref{eq:cTI1111}--\eqref{eq:cTI3333} as
\begin{subequations}
\label{eq:cTI_abh1h2}
\begin{align}
\label{eq:cTI1111_abh1h2}
c^{\overline{\rm TI}}_{1111}
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
&=
\left(1-\dfrac{2\,I_{1}}{h_{2} - h_{1}}\right)^{2}
\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
+
\frac{4}{h_{2}-h_{1}}
\left(I_{2} - I_{3}\right)
\,,
\\
\label{eq:cTI1122_abh1h2}
c^{\overline{\rm TI}}_{1122}
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
&=
\left(1-\dfrac{2\,I_{1}}{h_{2} - h_{1}}\right)^{2}
\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
+
\frac{2}{h_{2}-h_{1}}
\left(I_{2} - 2\,I_{3}\right)
\,,
\\
\label{eq:cTI1133_abh1h2}
c^{\overline{\rm TI}}_{1133}
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
&=
\left(1-\dfrac{2\,I_{1}}{h_{2} - h_{1}}\right)
\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
\,,
\\
\label{eq:cTI1212_abh1h2}
c^{\overline{\rm TI}}_{1212}
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\right)
&=
\frac{1}{3}
\left(
3\,a_{S}^{2} +
3\,a_{S}\,b_{S}\left(h_{1} + h_{2}\right) +
b_{S}^{2}\left(h_{1}^{2} + h_{1}\,h_{2} + h_{2}^{2}\right)
\right)
\,,
\\
\label{eq:cTI2323_abh1h2}
c^{\overline{\rm TI}}_{2323}
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\right)
&=
\left(a_{S} + b_{S}\,h_{1}\right)
\left(a_{S} + b_{S}\,h_{2}\right)
\,,
\\
\label{eq:cTI3333_abh1h2}
c^{\overline{\rm TI}}_{3333}
\left(h_{1}\,,h_{2}\,,a_{P}\,,b_{P}\right)
&=
\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
\,.
\end{align}
\end{subequations}
\subsection{Anisotropy parameters}
The anisotropy of any transversely isotropic medium may be described by the~\citet{Thomsen1986} parameters.
Thus, using Backus parameters~\eqref{eq:cTI_abh1h2}, we define
\begin{subequations}
\label{eq:ThomsenParam_linInh}
\begin{align}
\label{eq:gamma_linInh}
\gamma
=
\gamma\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\right)
&:=
\frac{
c^{\overline{\rm TI}}_{1212} -
c^{\overline{\rm TI}}_{2323}
}{
2\,c^{\overline{\rm TI}}_{2323}
}
\,,
\\
\label{eq:delta_linInh}
\delta
=
\delta\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
&:=
\frac{
\left(
c^{\overline{\rm TI}}_{1133} +
c^{\overline{\rm TI}}_{2323}
\right)^{2}
-
\left(
c^{\overline{\rm TI}}_{3333} -
c^{\overline{\rm TI}}_{2323}
\right)^{2}
}{
2\,c^{\overline{\rm TI}}_{3333}
\left(
c^{\overline{\rm TI}}_{3333} -
c^{\overline{\rm TI}}_{2323}
\right)
}
\,,
\\
\label{eq:epsilon_linInh}
\varepsilon
=
\varepsilon\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
&:=
\frac{
c^{\overline{\rm TI}}_{1111} -
c^{\overline{\rm TI}}_{3333}
}{
2\,c^{\overline{\rm TI}}_{3333}
}
\,.
\end{align}
\end{subequations}
Explicitly, parameter~\eqref{eq:gamma_linInh} is
\begin{subequations}
\begin{equation}
\label{eq:gamma_abh1h2}
\gamma
=
\gamma\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\right)
=
\frac{
{b_{S}}^{2}\left(h_{1} - h_{2}\right)^{2}
}{
6
\left(a_{S} + b_{S}\,h_{1}\right)
\left(a_{S} + b_{S}\,h_{2}\right)
}
\,.
\end{equation}
For parameter~\eqref{eq:delta_linInh},
\begin{align}
\nonumber
\delta
&=
\delta
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
\\
\label{eq:delta_abh1h2}
&=
\frac{
2\,{\delta}_{k1}
\left(
-b_{P}
\left(h_{1} - h_{2}\right)
\left(
2\,a_{P} + b_{P}\left(h_{1} + h_{2}\right)
\right)
+
{\delta}_{k2}
\right)
\left(
b_{P}
\left(h_{1} - h_{2}\right)
{\delta}_{k3}
+
{\delta}_{k1}\,
{\delta}_{k2}
\right)
}{
{b_{P}}^6
\left(a_{P} + b_{P}\,h_{1}\right)
\left(h_{1} - h_{2}\right)^{2}
\left(a_{P} + b_{P}\,h_{2}\right)
\left(
{a_{P}}^2 +
{\delta}_{k4} +
a_{P}\,b_{P}\left(h_{1} + h_{2}\right) -
a_{S}\,b_{S}\left(h_{1} + h_{2}\right)
\right)
}
\,,
\end{align}
where
\begin{align*}
{\delta}_{k1}
&=
b_{S}
\left(-a_{S}\,b_{P} + a_{P}\,b_{S}\right)
\,,
\\
{\delta}_{k2}
&=
2\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
\ln\left(
\frac{a_{P} + b_{P}\,h_{1}}{a_{P} + b_{P}\,h_{2}}
\right)
\,,
\\
{\delta}_{k3}
&=
\left(
{a_{P}}^{2}\left({b_{P}}^2 - 2\,{b_{S}}^2\right) +
{b_{P}}^{2}\left(
{\delta}_{k4}
\right)
+
a_{P}\,b_{P}\left(
2\,a_{S}\,b_{S} +
\left(b_{P} - b_{S}\right)
\left(b_{P} + b_{S}\right)
\left(h_{1} + h_{2}\right)
\right)
\right)
\,,
\\
{\delta}_{k4}
&=
-a_{S}^2 +
\left(b_{P} - b_{S}\right)
\left(b_{P} + b_{S}\right)\,h_{1}\,h_{2}
\,.
\end{align*}
For parameter~\eqref{eq:epsilon_linInh},
\begin{align}
\label{eq:epsilon_abh1h2}
\varepsilon
&=
\varepsilon
\left(h_{1}\,,h_{2}\,,a_{S}\,,b_{S}\,,a_{P}\,,b_{P}\right)
=
\frac{
2\,b_{S}
\left(
{b_{P}}^{3}
\left(h_{1} - h_{2}\right)^{2}
\left(
-6\,{a_{P}}^{2}\,b_{P}\,b_{S} +
b_{P}
\left(
{\varepsilon}_{k1}
\right)
+
3\,a_{P}
\left(
{\varepsilon}_{k2}
\right)
\right)
+
\left({\varepsilon}_{k3}\right)
\left({\varepsilon}_{k4}\right)
\right)
}{
3\,{b_{P}}^{6}\left(a_{P} + b_{P}\,h_{1}\right)
\left(h_{1} - h_{2}\right)^{2}
\left(a_{P} + b_{P}\,h_{2}\right)
}
\end{align}
\end{subequations}
where
\begin{align*}
{\varepsilon}_{k1}
&=
-12\,{a_{S}}^{2}\,b_{S}
+
\left(b_{P} - b_{S}\right)
b_{S}
\left(b_{P} + b_{S}\right)
\left(h_{1} - h_{2}\right)^{2}
+
3\,a_{S}
\left({b_{P}}^{2} - 2\,{b_{S}}^{2}\right)
\left(h_{1} + h_{2}\right)
\,,
\\
{\varepsilon}_{k2}
&=
2\,a_{S}
\left({b_{P}}^{2} + 2\,{b_{S}}^{2}\right)
+
b_{S}
\left(-{b_{P}}^{2} + 2\,{b_{S}}^{2}\right)
\left(h_{1} + h_{2}\right)
\,,
\\
{\varepsilon}_{k3}
&=
6
\left(a_{S}\,b_{P} - a_{P}\,b_{S}\right)
\left(a_{P} + b_{P}\,h_{1}\right)
\left(a_{P} + b_{P}\,h_{2}\right)
\ln\left(
\frac{a_{P} + b_{P}\,h_{1}}{a_{P} + b_{P}\,h_{2}}
\right)
\,,
\\
{\varepsilon}_{k4}
&=
-b_{P}
\left({b_{P}}^{2} - 2\,{b_{S}}^{2}\right)
\left(h_{1} - h_{2}\right)
+
2\,b_{S}\left(-a_{S}\,b_{P} + a_{P}\,b_{S}\right)
\ln\left(
\frac{a_{P} + b_{P}\,h_{2}}{a_{P} + b_{P}\,h_{1}}
\right)
\,.
\end{align*}
\subsection{Total differentials of anisotropy parameters}
To quantify the uncertainty of expressions~\eqref{eq:gamma_abh1h2}--\eqref{eq:epsilon_abh1h2}, i.e., the sensitivity to changes in model parameters, we require the total differential.
We demonstrate this, we differentiate $\gamma$ with respect to each of its coordinates directions to obtain its linear functional
\begin{subequations}
\begin{equation}
\label{eq:gammaDifferential}
{\rm d}\gamma
=
\left(
\frac{\partial\,\gamma}{\partial h_{1}}
\right)
{\rm d}h_{1}
+
\left(
\frac{\partial\,\gamma}{\partial h_{2}}
\right)
{\rm d}h_{2}
+
\left(
\frac{\partial\,\gamma}{\partial a_{S}}
\right)
{\rm d}a_{S}
+
\left(
\frac{\partial\,\gamma}{\partial b_{S}}
\right)
{\rm d}b_{S}
\,,
\end{equation}
where
\begin{align*}
\frac{\partial\,\gamma}{\partial h_{1}}
&=
\frac{
{b_{S}}^{2}
\left(h_{1} - h_{2}\right)
\left(
2\,a_{S} +
b_{S}\left(h_{1} + h_{2}\right)
\right)
}{
6
\left(a_{S} + b_{S}\,h_{1}\right)^{2}
\left(a_{S} + b_{S}\,h_{2}\right)
}
\,,
\\
\frac{\partial\,\gamma}{\partial h_{2}}
&=
-\frac{
{b_{S}}^{2}
\left(h_{1} - h_{2}\right)
\left(
2\,a_{S} +
b_{S}\left(h_{1} + h_{2}\right)
\right)
}{
6
\left(a_{S} + b_{S}\,h_{1}\right)
\left(a_{S} + b_{S}\,h_{2}\right)^{2}
}
\,,
\\
\frac{\partial\,\gamma}{\partial a_{S}}
&=
-\frac{
{b_{S}}^{2}
\left(h_{1} - h_{2}\right)^{2}
\left(
2\,a_{S} +
b_{S}\left(h_{1} + h_{2}\right)
\right)
}{
6
\left(a_{S} + b_{S}\,h_{1}\right)^{2}
\left(a_{S} + b_{S}\,h_{2}\right)^{2}
}
\,,
\\
\frac{\partial\,\gamma}{\partial b_{S}}
&=
\frac{
a_{S}\,b_{S}
\left(h_{1} - h_{2}\right)^{2}
\left(
2\,a_{S} +
b_{S}\left(h_{1} + h_{2}\right)
\right)
}{
6
\left(a_{S} + b_{S}\,h_{1}\right)^{2}
\left(a_{S} + b_{S}\,h_{2}\right)^{2}
}
\,.
\end{align*}
We may perform similar operations on $\delta$ and $\varepsilon$ to obtain
\begin{equation}
\label{eq:deltaDifferential}
{\rm d}\delta
=
\left(\frac{\partial\,\delta}{\partial h_{1}}\right){\rm d}h_{1} +
\left(\frac{\partial\,\delta}{\partial h_{2}}\right){\rm d}h_{2} +
\left(\frac{\partial\,\delta}{\partial a_{S}}\right){\rm d}a_{S} +
\left(\frac{\partial\,\delta}{\partial b_{S}}\right){\rm d}b_{S} +
\left(\frac{\partial\,\delta}{\partial a_{P}}\right){\rm d}a_{P} +
\left(\frac{\partial\,\delta}{\partial b_{P}}\right){\rm d}b_{P}
\end{equation}
and
\begin{equation}
\label{eq:epsilonDifferential}
{\rm d}\varepsilon
=
\left(\frac{\partial\,\varepsilon}{\partial h_{1}}\right){\rm d}h_{1} +
\left(\frac{\partial\,\varepsilon}{\partial h_{2}}\right){\rm d}h_{2} +
\left(\frac{\partial\,\varepsilon}{\partial a_{S}}\right){\rm d}a_{S} +
\left(\frac{\partial\,\varepsilon}{\partial b_{S}}\right){\rm d}b_{S} +
\left(\frac{\partial\,\varepsilon}{\partial a_{P}}\right){\rm d}a_{P} +
\left(\frac{\partial\,\varepsilon}{\partial b_{P}}\right){\rm d}b_{P}
\,.
\end{equation}
\end{subequations}
We do not list the the resultant expressions since they would require half-a-dozen pages.
However, the expressions for the partial derivatives of differentials~\eqref{eq:deltaDifferential} and~\eqref{eq:epsilonDifferential} may be derived using a symbolic software, such as Mathematica.
\section{Methods for well-posedness}
\label{sec:Methods}
Expressions~\eqref{eq:gamma_abh1h2}--\eqref{eq:epsilon_abh1h2} form an ill-posed system of three equations with nine unknowns, where
\begin{equation}
\label{eq:AnalyticSystem}
\begin{cases}
\gamma
=
\gamma\left(h_{1},h_{2},a_{S},b_{S}\right)
\\
\delta
=
\delta\left(h_{1},h_{2},a_{S},b_{S},a_{P},b_{P}\right)
\\
\varepsilo
=
\varepsilon\left(h_{1},h_{2},a_{S},b_{S},a_{P},b_{P}\right)
\end{cases}
\,.
\end{equation}
We use differentials~\eqref{eq:gammaDifferential}--\eqref{eq:epsilonDifferential} to form a measure of error.
To obtain solutions, we require well-posedness, which necessitates information for six of the nine unknown parameters.
Using the~\citeauthor{Backus1962} average on a region of interest of a well log, we reduce the number of unknowns to four.
In particular, we specify $h_{1}$ and $h_{2}$\,, and calculate values for $\gamma=:\gamma^{\overline{\rm TI}}\,,\delta=:\delta^{\overline{\rm TI}}\,,\,\varepsilon=:\varepsilon^{\overline{\rm TI}}$\,, where~${}^{\overline{\rm TI}}$ denotes a quantity obtained using the~\citet{Backus1962} average.
Information for any of~$a_{S}\,,b_{S}\,,a_{P}\,,b_{P}$ must be obtained using additional methods.
To obtain information for the remaining model parameters, we use 1-D tomography and the~$ab$ model.
\subsection{1-D tomography}
\label{sec:1D_tomo}
To obtain linear inhomogeneity parameters from VSP data, we use a 1-D tomography method.\
The inversion algorithm, therein, is based on the Levenberg-Marquardt least-square solution.
As a result of inversion, we obtain a velocity profile corresponding to discrete depths.
The inverted velocity profile represents the velocity of a vertically inhomogeneous medium.
To calculate the linear inhomogeneity parameters for a particular region, we apply the linear regression to the resultant velocities for the required depths.
The 1-D tomography can provide relatively good results in the case of a better initial model, a sufficient number of data, and that the assumption of linear inhomogeneity holds firmly for the medium.
\subsection{$\boldsymbol{ab}$ model}
The~$ab$ model is a traveltime inversion method.\
In this method, the velocity model is a linear function of depth.
To obtain the traveltime in a depth segment from borehole velocity, we assume the medium is composed of multiple thin isotropic layers.
Based on that assumption, we calculate Fermat's traveltime.
In general, the borehole data contains high-frequency contents, which restricts the ray to travel from a source to a receiver with high offset.
To synthesize sufficient traveltime data, we need traveltime for both near and high offsets.
That is possible only if we remove the higher frequencies from the data.
We apply a weighted average in every twelve consecutive data points to filter out those contents.
As an outcome, we get a smoother velocity, which allows rays to travel from source to receiver with sufficiently more significant offset.
In the inversion, we minimize the difference between the Fermat's traveltime and the model traveltime~\citep{SlawinskiSlawinski1999} and obtain linear inhomogeneity parameters.
\subsection{Mizzen O-16}
The VSP and well log data used, herein, was obtained from the Mizzen O-16 discovery well, which is a site in the Flemish Pass basin and was drilled in 2009 by Statoil~\citep{Enachescu2011}.
The well log data used is supplied by IHS energy; the data description is provided in~\citet{NalcorEnergy2016}.
We use $P$- and $S$-wavespeed measurements for depths of at depth 1865m to 2648.60\,m\,.
The checkshot (VSP) data is provided by the Canada-Newfoundland \& Labrador Offshore Petroleum Board~\citep{CNLPOB2009}.
Therein, the traveltime data corresponds to a single source and multiple receivers.
The source is located at a 26.5\,m offset and the receivers are placed along vertical axis, starting at depth 1865\,m and ending at 2650\,m.
We consider our last receiver at depth 2650m so that the velocity inversion from traveltime merges with the region where the well log data are recorded.
The descriptions of the data are given in Tables~\ref{tab:vel_des} and \ref{tab:cs_des}, which are collected with the permission of the Petroleum Development Section of Natural Resources, Government of Newfoundland and Labrador.
\begin{table}[h]
\centering
\begin{tabular}{c*{6}{c}}
\toprule
Field &
Well &
KB (${\rm m}$) &
TVD (${\rm m}$) &
Water depth (${\rm m}$) &
Spud date &
Log data
\\[2pt]
\toprule
Mizzen & O-16 & 21.15 & 3797 &1095 & 2008 & $\checkmark$ \\
\bottomrule
\end{tabular}
\caption{
Description of the well for log data
}
\label{tab:vel_des}
\end{table}
\begin{table}[h]
\centering
\begin{tabular}{c*{6}{c}}
\toprule
Field &
Well &
Offset (${\rm m}$) &
TVD (${\rm m}$) &
Water depth (${\rm m}$) &
Spud date &
Checkshot
\\[2pt]
\toprule
Mizzen & O-16 & 26.50 & 3797 &1095 & 2009 & $\checkmark$ \\
\bottomrule
\end{tabular}
\caption{
Description of the well for checkshot data
}
\label{tab:cs_des}
\end{table}
\section{Numerical search}
\label{sec:Results}
\subsection{Restrictions}
In our numerical search, we have two restrictions.
First, the stability conditions of isotropy restrict the values for~$a_{S}$ and~$a_{P}$ such that
\begin{equation*}
a_{P} > 2\,a_{S}\,/\sqrt{3}
\,.
\end{equation*}
Second, to remain consistent with the assumption of constantly increasing velocity gradient with depth of~\citet{SlawinskiSlawinski1999}, we restrict our solutions to positive values of~$b_{P}$ and~$b_{S}$\,.
However, it is worth noting that there do exist solutions for negative values of~$b_{P}$ and~$b_{S}$\,.
We may demonstrate this with a numerical example.
For, say, an input of~$a_{P} = 2040.36\,{\rm ms}^{-1}$\,, there exist two solutions, as illustrated by the two instances of triple-point intersection in the left- and right-hand plots of Figure~\ref{fig:TriplePoints}.
Therein, the left-hand plot corresponds to a solution of positive~$b_{P}$ and~$b_{S}$\,, where
\begin{equation*}
\label{eq:VSP_anal_given_bp}
a_{S} = 752.95\,{\rm ms}^{-1}
\,,\quad
b_{S} = 0.3666\,{\rm s}^{-1}
\,,\quad
a_{P} = 2164.68\,{\rm ms}^{-1}
\,,\quad
b_{P} = 0.4081\,{\rm s}^{-1}
\,.
\end{equation*}
However, the right-hand plot corresponds to a solution of negative~$b_{P}$ and~$b_{S}$\,, where
\begin{equation*}
\label{eq:VSP_anal_given_bp}
a_{S} = 906.32\,{\rm ms}^{-1}
\,,\quad
b_{S} = -0.3194\,{\rm s}^{-1}
\,,\quad
a_{P} = 2164.68\,{\rm ms}^{-1}
\,,\quad
b_{P} = -0.3556\,{\rm s}^{-1}
\,.
\end{equation*}
Throughout the entirety of the numerical results, we consider positive solutions only.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{TriplePoint_positive_bsbp.pdf}
\hspace*{\fill}
\includegraphics[width=0.45\textwidth]{TriplePoint_negative_bsbp.pdf}
\caption{%
Contour plot of anisotropy values~\eqref{eq:AnisotropyValues}, where $b_{P}$ is along the vertical axis and $b_{S}$ is along the horizontal axis.
Green lines represent $\gamma = \gamma^{\overline{\rm TI}}$\,; orange lines represents $\delta = \delta^{\overline{\rm TI}}$\,, blue lines represent $\varepsilon = \varepsilon^{\overline{\rm TI}}$\,.}
\label{fig:TriplePoints}
\end{figure}
\subsection{Calculations}
\label{sec:Calcs}
Let us obtain a solution to system~\eqref{eq:AnalyticSystem} using the methodology of Section~\ref{sec:1D_tomo} with the VSP data of Mizzen O-16.
Using results from 1-D tomography on traveltime data, we recover a velocity profile as a function of depth.
For a region of interest, whose depth ranges from 1865\,m to 2648.60\,m\,, we recover two parameters, namely~$a_{P}$ and~$b_{P}$\,.
We may use either parameter as input to system~\eqref{eq:AnalyticSystem} to obtain values for the remaining three unknowns.
For example, we use~$a_{P_{\rm VSP,in}} = 1415\,{\rm ms}^{-1}$ and~$b_{P_{\rm VSP,in}} = 0.70\,{\rm s}^{-1}$ as startup values for the tomography.
For the region of interest, we recover~$a_{P_{\rm VSP,out}} = 1354.9\,{\rm ms}^{-1}$ and~$b_{P_{\rm VSP,out}} = 0.3933\,{\rm s}^{-1}$\,, where, at the top of the region,~$a_{P_{\rm layer}} = a_{P_{\rm VSP,out}} + b_{P_{\rm VSP,out}}\cdot1865\,{\rm m} = 2088.38\,{\rm ms}^{-1}$\,.
We input~$b_{P} = b_{P_{\rm VSP,out}}$ into system~\eqref{eq:AnalyticSystem}, for
\begin{equation*}
h_{1} = 0\,{\rm m}
\,,\quad
h_{2} = \left(2648.6 - 1865\right){\rm m} = 783.6\,{\rm m}\,,
\end{equation*}
\begin{equation}
\label{eq:AnisotropyValues}
\gamma^{\overline{\rm TI}} = 0.017561151400350
\,,\quad
\delta^{\overline{\rm TI}} = -0.005822848520484
\,,\quad
\varepsilon^{\overline{\rm TI}} = 0.002868244418444
\,,
\end{equation}
and obtain
\begin{equation}
\label{eq:VSP_anal_given_bp}
a_{S} = 725.55\,{\rm ms}^{-1}
\,,\quad
b_{S} = 0.3533\,{\rm s}^{-1}
\,,\quad
a_{P} = 2085.91\,{\rm ms}^{-1}
\,,\quad
b_{P} = 0.3933\,{\rm s}^{-1}
\,.
\end{equation}
It is clear that~$a_{P_{\rm layer}}\neq a_{P}$ but we may perform an error analysis to assess the ``closeness'' of our result.
To do so, we recall expressions~\eqref{eq:gammaDifferential}, \eqref{eq:deltaDifferential}, \eqref{eq:epsilonDifferential}, which are total differentials ${\rm d}\gamma\,,{\rm d}\delta\,,{\rm d}\varepsilon$\,.
Using uncertainty measures for parameters~\eqref{eq:VSP_anal_given_bp}, where
\begin{equation}
\label{eq:Aniso_Unc_param}
{\rm d}h_{1} = {\rm d}h_{2} = 0.05\,{\rm m}
\,,\quad
{\rm d}a_{S} = {\rm d}a_{P} = 2\,{\rm ms}^{-1}
\,,\quad
{\rm d}b_{S} = {\rm d}b_{P} = 0.01\,{\rm s}^{-1}
\,,
\end{equation}
we obtain
\begin{equation}
\label{eq:anisotropy_differentials}
{\rm d}\gamma = 0.000790010132156
\,,\quad
{\rm d}\delta = -0.000299948944243
\,,\quad
{\rm d}\varepsilon = 0.000143889902574
\,.
\end{equation}
We obtain a new set of solutions for the lower limit of anisotropy by subtracting values~\eqref{eq:anisotropy_differentials} from the left-hand sides of system~\eqref{eq:AnalyticSystem}, which are
\begin{equation}
\label{eq:VSP_anal_given_bp_sub}
a_{S_{-}} = 742.47\,{\rm ms}^{-1}
\,,\quad
b_{S_{-}} = 0.3522\,{\rm s}^{-1}
\,,\quad
a_{P_{-}} = 2138.75\,{\rm ms}^{-1}
\,,\quad
b_{P} = 0.3933\,{\rm s}^{-1}
\,.
\end{equation}
Similarly, the upper limit of anisotropy is obtained adding values~\eqref{eq:anisotropy_differentials}, which results in
\begin{equation}
\label{eq:VSP_anal_given_bp_add}
a_{S_{+}} = 709.58\,{\rm ms}^{-1}
\,,\quad
b_{S_{+}} = 0.354246\,{\rm s}^{-1}
\,,\quad
a_{P_{+}} = 2036.58\,{\rm ms}^{-1}
\,,\quad
b_{P} = 0.3933\,{\rm s}^{-1}
\,.
\end{equation}
Thus, we find that~$a_{P_{\rm layer}}$ falls within the range determined by~$a_{P_{-}}$ and~$a_{P_{+}}$\,.
In particular, such startup values provide results that are within 0.12\% and 0.13\%\,, respectively, of the theoretical predictions.
We may repeat this entire process using~$a_{P} = a_{P_{\rm VSP,out}}$ as input.
In addition, this process may be repeated for a range of VSP startup values, in order to find the combination of startup values that are closest to the predicted model parameters indicated by the analytical relation.
We tabulate the results of such a process, for values of~$a_{P_{\rm VSP,in}}$ ranging from 1325\,ms$^{-1}$ to 1515\,ms$^{-1}$\,, with increments of 20\,ms$^{-1}$\,, and~$b_{P_{\rm VSP,in}}$ ranging from 0.57\,s$^{-1}$ to 0.7\,s$^{-1}$\,, with increments of 0.01\,s$^{-1}$\,, in Table~\ref{tab:VSP_results} of Appendix~\ref{app:Tables}.
To compare the results from VSP with the~$ab$ model, we repeat this process for the same range of values of~$a_{P_{\rm ab,in}}$ and~$b_{P_{\rm ab,in}}$\,.
The results are tabulated in Table~\ref{tab:ab_results} of Appendix~\ref{app:Tables}.
In comparison, the startup values of
~$a_{P_{ab,{\rm in}}} = 1625\,{\rm ms}^{-1}$ and~$b_{P_{ab,{\rm in}}} = 0.6\,{\rm s}^{-1}$, provide results that are within 0.03\% and 0.03\%\,, respectively, of the theoretical predictions.
We illustrate the entirety of the results of Tables~\ref{tab:VSP_results} and~\ref{tab:ab_results} in Figures~\ref{fig:fig_bp_given_ap_new} and~\ref{fig:fig_ap_given_bp_new}.
Therein, the solid red line represents the solution to system~\eqref{eq:AnalyticSystem} corresponding to the value on the horizontal axis, and the solid black lines represent solutions for the lower and upper limits of anisotropy, corresponding to uncertainty parameters~\eqref{eq:Aniso_Unc_param}.
Orange diamonds represent the solutions from 1-D tomography and blue dots represent solutions from the~$ab$ model.
\section{Discussion}
For each solution method, we obtain some $a_{P_{{\rm VSP},ab}}$ and $b_{P_{{\rm VSP},ab}}$ parameter.
Using the outputted~$a_{P_{{\rm VSP},ab}}$ as input to system~\eqref{eq:AnalyticSystem}, we calculate~$b_{P}$\,,~$b_{P_{-}}$\,,~$b_{P_{+}}$\,, and compare them to the outputted $b_{P_{{\rm VSP},ab}}$\,.
These results are illustrated in Figure~\ref{fig:fig_bp_given_ap_new}; the opposite operation is performed and those results are illustrated in Figure~\ref{fig:fig_ap_given_bp_new}.
\begin{figure}[h]
\centering
\subcaptionbox{\label{fig:fig_bp_given_ap_new}}{\includegraphics[width=0.49\textwidth]{fig_bp_given_ap_new.pdf}}\hfill
\subcaptionbox{\label{fig:fig_ap_given_bp_new}}{\includegraphics[width=0.49\textwidth]{fig_ap_given_bp_new.pdf}}
\caption{%
Solutions of 1-D tomography and~$ab$-model for startup values of $a_{P_{\rm VSP,in}}$ and~$b_{P_{\rm VSP,in}}$ in Table with VSP results and $a_{P_{ab,{\rm in}}}$ and~$b_{P_{ab,{\rm in}}}$ in Table with $ab$ results.
}
\label{fig:fig_bp_given_ap_ap_given_bp_new}
\end{figure}
In both subplots of Figure~\ref{fig:fig_bp_given_ap_ap_given_bp_new}, we observe that the area of solutions for 1-D tomography overlaps the line of~$ab$-model solutions.
To discern which startup values lead to common outputs, we turn our attention to Figure~\ref{fig:fig_bp_given_ap_ap_given_bp_solution}.
Therein, startup values of $a_{P_{{\rm VSP,in}}} = 1415$\,ms$^{-1}$ and $b_{P_{{\rm VSP,in}}} = 0.7$\,ms$^{-1}$ result in$a_{P_{{\rm VSP,out}}} = 2088.38$\,ms$^{-1}$ and $b_{P_{{\rm VSP,out}}} = 0.3932$\,ms$^{-1}$\,.
These results, as stated in Section~\ref{sec:Calcs}, are within 0.12\% and 0.13\%\,, respectively, of the theoretical predictions.
Similarly, startup values of $a_{P_{ab,in}} = 1625$\,ms$^{-1}$ and $b_{P_{ab,in}} = 0.6$\,ms$^{-1}$ result in outputs that are within 0.03\% of the theoretical predictions.
\begin{figure}[h]
\centering
\subcaptionbox{\label{fig:fig_bp_given_ap_solution}}{\includegraphics[width=0.49\textwidth]{fig_bp_given_ap_fixed_ap_solution.pdf}}\hfill
\subcaptionbox{\label{fig:fig_ap_given_bp_solution}}{\includegraphics[width=0.49\textwidth]{fig_ap_given_bp_fixed_ap_solution.pdf}}
\caption{%
Solutions of 1-D tomography and~$ab$-model for startup values of $a_{P_{\rm VSP,in}}=1415$\,ms$^{-1}$ and~$b_{P_{\rm VSP,in}}$ incrementing by 0.01\,s$^{-1}$ from 0.57\,s$^{-1}$ to 0.7\,s$^{-1}$\,, and $a_{P_{ab,{\rm in}}} = 1625$\,ms$^{-1}$ and~$b_{P_{ab,{\rm in}}}$ incrementing by 0.05\,s$^{-1}$ from 0.25\,s$^{-1}$ to 0.65\,s$^{-1}$.
}
\label{fig:fig_bp_given_ap_ap_given_bp_solution}
\end{figure}
\section{Conclusion and future work}
\label{sec:Conclusion}
In this article, we show that we can obtain multiple solution from VSP and $ab$-model that do not lie in the theoretical range.
However, we also show that there exist some common solution.
Based on the initial model, we find that there are some initial values that lead to that common solution, which may allow us to conclude that, by tuning initial model parameters, we can obtain the common solution set.
The number of traveltime data from the field measurement plays an important role in the VSP method.
As a future work, we plan to apply the analytical relation and seismic methods to a site that possesses more data, which may produces a stable solution for different initial models in the VSP method.
By far, we have shown that the relationship between inhomogeneity and anisotropy provides reasonable solution in compare to the seismic methods.
In addition to our future project, we wish to examine the limitation of the analytical relationship by calculating Backus average from a sparse well log data.
\bibliographystyle{apa}
|
\section{Conclusion}
We propose RUBi to reduce unimodal biases learned by Visual Question Answering (VQA) models. RUBi is a simple learning strategy designed to be model agnostic. It is based on a question-only branch that captures unwanted statistical regularities from the question modality. This branch influences the base VQA model to prevent the learning of unimodal biases from the question.
We demonstrate a significant gain of +5.94 percentage point in accuracy over the state-of-the-art result on VQA-CP v2, a dataset specifically designed to account for question biases.
We also show that RUBi is effective with different kinds of common VQA models.
In future works, we would like to extend our approach on other multimodal tasks.
\newpage
\section*{Acknowledgments}
We would like to thank the reviewers for valuable and constructive comments and suggestions. We additionally would like to thank Abhishek Das and Aishwarya Agrawal for their help.
The effort from Sorbonne University was partly supported within the Labex SMART supported by French state funds managed by the ANR within the Investissements d’Avenir programme under reference ANR-11-LABX-65, and partly funded by grant DeepVision (ANR-15-CE23-0029-02, STPGP-479356-15), a joint French/Canadian call by ANR \& NSERC.
\input{main.bbl}
\newpage
\input{sections/suppl.tex}
\end{document}
\section{Introduction}
The recent Deep Learning success in computer vision \cite{NIPS2012_4824} and natural language understanding \cite{mikolov2013efficient} allowed researchers to tackle multimodal tasks that combine visual and textual modalities \cite{kiros2015skip, karpathy2015deep, lu2016visual, Das2016a, Vries2017}.
Among these tasks, Visual Question Answering (VQA) attracts increasing attention. The goal of the VQA task is to answer a question about an image. It requires a high-level understanding of the visual scene and the question, but also to ground the textual concepts in the image and to use both modalities adequately.
Solving the VQA task could have tremendous impacts on real-world applications such as aiding visually impaired users in understanding their physical and online surroundings, searching through large quantities of visual data via natural language interfaces, or even communicating with robots using more efficient and intuitive interfaces.
Several large real image VQA datasets have recently emerged \cite{VQA, VQA2_Goyal_2017_CVPR, vqa-cp, Kafle_2017_ICCV, gurari2018vizwiz, hudson2019gqa, zellers2018recognition}. Each one of them targets specific abilities that a VQA model would need to be used in real-world settings such as fine-grained recognition, object detection, counting, activity recognition, commonsense reasoning, etc. Current end-to-end VQA models \cite{Anderson_2018_CVPR, Cadene_2019_CVPR, BenYounes_2019_AAAI, Hu2018stacknmn, kim2018bilinear, Shi2018,Wu2018,Peng2018} achieve impressive results on most of these benchmarks and are even able to surpass the human accuracy on a specific benchmark accounting for compositional reasoning \cite{johnson2016clevr}.
However, it has been shown that they tend to exploit statistical regularities between answer occurrences and certain patterns in the question \cite{agrawal2016analyzing, vqa-cp, ramakrishnan2018overcoming, johnson2016clevr, hudson2019gqa}. While they are designed to merge information from both modalities, in practice they often answer without considering the image modality.
When most of the bananas are \textit{yellow}, a model does not need to learn the correct behavior to reach a high accuracy for questions asking about the color of bananas. Instead of looking at the image, detecting a banana and assessing its color, it is much easier to learn from the statistical shortcut linking the words \textit{what}, \textit{color} and \textit{bananas} with the most occurring answer \textit{yellow}.
One way to quantify the amount of statistical shortcuts from each modality is to train unimodal models. For instance, a question-only model trained on the widely used VQA v2 dataset \cite{VQA2_Goyal_2017_CVPR} predicts the correct answer approximately 44\% of the time over the test set.
VQA models are not discouraged to exploit these statistical shortcuts from the question modality, because their training set often follows the same distribution as their testing set. However, when evaluated on a test set that displays different statistical regularities, they usually suffer from a significant drop in accuracy \cite{vqa-cp, ramakrishnan2018overcoming}. Unfortunately, these statistical regularities are hard to avoid when collecting real datasets. As illustrated in Figure~\ref{fig:big_picture}, there is a crucial need to develop new strategies to reduce the amount of biases coming from the question modality in order to learn better behaviors.
We propose RUBi, a training strategy to reduce the amount of biases learned by VQA models.
Our strategy reduces the importance of the most biased examples, i.e. examples that can be correctly classified without looking at the image modality.
It implicitly forces the VQA model to use the two input modalities instead of relying on statistical regularities between the question and the answer.
We take advantage of the fact that question-only models are by design biased towards the question modality.
We add a question-only branch on top of a base VQA model during training only.
This branch influences the VQA model, dynamically adjusting the loss to compensate for biases.
As a result, the gradients backpropagated through the VQA model are reduced for the most biased examples and increased for the less biased.
At the end of the training, we simply remove the question-only branch.
We run extensive experiments on VQA-CP v2 \cite{vqa-cp} and demonstrate the ability of RUBi to surpass current state-of-the-art results from a significant margin. This dataset has been specifically designed to assess the capacity of VQA models to be robust to biases by the question modality.
We show that our RUBi learning framework provides gains when applied on several VQA architectures such as Stacked Attention Networks \cite{yang2016stacked} and Top-Down Bottom-Up Attention \cite{Anderson_2018_CVPR}.
We also show that RUBi is competitive on the standard VQA v2 dataset \cite{VQA2_Goyal_2017_CVPR} when compared to approaches that reduce unimodal biases.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{images/big_picture.png}
\caption{\label{fig:big_picture} Our RUBi approach aims at reducing the amount of unimodal biases learned by a VQA model during training. As depicted, current VQA models often rely on unwanted statistical correlations between the question and the answer instead of using both modalities.}
\end{figure}
\section{Related work}
Real-world datasets display some form of inherent biases due to their collection process \cite{gordon2013reporting, chao2017being, Torralba2011}.
As a result, machine learning models tend to reflect these biases because they capture often undesirable correlations between the inputs and the ground truth annotations \cite{Stock2017, Jia2018, Manjunatha2018}.
Procedures exist to identify certain kinds of biases and to reduce them. For instance, some methods are focused on gender biases \cite{Burns2018, biasemnlp}, some others on the human reporting biases \cite{misra2016seeing}, and also on the shift in distribution between lab-curated data and real-world data \cite{gupta2018robot}.
In the language and vision context, some works evaluate unimodal baselines \cite{anand2018blindfold, thomason2019shifting} or leverage language priors \cite{rohrbach2018hallucination}.
In the following, we discuss about related works that assess and reduce unimodal biases learned by VQA models.
\paragraph{Assessing unimodal biases in datasets and models}
Despite being designed to merge the two input modalities, it has been found that VQA models often rely on superficial correlations between inputs from one modality and the answers without considering the other modality \cite{jabri2016revisiting, Manjunatha2018}.
An interesting way to quantify the amount of unimodal biases that can potentially be learned by a VQA model consists in training models using only one of the two modalities \cite{VQA, VQA2_Goyal_2017_CVPR}.
The question-only model is a particularly strong baseline because of the large amount of statistical regularities that can be leveraged from the question modality. With the RUBi learning strategy, we take advantage of this baseline model to prevent VQA models from learning question biases.
Unfortunately, biased models that exploit statistical shortcuts from one modality usually reach impressive accuracy on most of the current benchmarks. VQA-CP v2 and VQA-CP v1 \cite{vqa-cp} were recently introduced as diagnostic datasets containing different answer distributions for each question-type between train and test splits. Consequentially, models biased towards the question modality fail on these benchmarks. We use the more challenging VQA-CP v2 dataset extensively in order to show the ability of our approach to reduce the learning of biases coming from the question modality.
\paragraph{Balancing datasets to avoid unimodal biases}
Once the unimodal biases have been identified, one method to overcome these biases is to create more balanced datasets. For instance, the synthetic datasets for VQA \cite{johnson2016clevr, hudson2019gqa} minimize question-conditional biases via rejection sampling within families of related questions to avoid simple shortcuts to the correct answer.
Doing rejection sampling in real VQA datasets is usually not possible due to the cost of annotations. Another solution is to collect complementary examples to increase the difficulty of the task. For instance, VQA v2 \cite{VQA2_Goyal_2017_CVPR} has been introduced to weaken language priors in the VQA v1 dataset \cite{VQA} by identifying complementary images. For a given VQA v1 question, VQA v2 also contains a similar image with a different answer to the same question.
However, even with this additional balancing, statistical biases from the question remain and can be leveraged \cite{vqa-cp}.
That is why we propose an approach to reduce unimodal biases during training. It is designed to learn unbiased models from biased datasets. Our learning strategy dynamically modifies the loss values to reduce biases from the question. By doing so, we reduce the importance of certain examples, similarly to the rejection sampling approach, while increasing the importance of complementary examples which are already in the training set.
\paragraph{Architectures and learning strategies to reduce unimodal biases}
In parallel of these previous works on balancing datasets, an important effort has been carried out to design VQA models to overcome biases from datasets.
\cite{vqa-cp} proposed a hand-designed architecture called Grounded VQA model (GVQA).
It breaks the task of VQA down into a first step of locating and recognizing the visual regions needed to answer the question, and a second step of identifying the space of plausible answers based on a question-only branch. This approach requires training multiple sub-models separately. In contrast, our learning strategy is end-to-end. Their complex design is not straightforward to apply on different architectures while our approach is model-agnostic. While we rely on a question-only branch, we remove it at the end of the training.
The work most related to ours in terms of approach is \cite{ramakrishnan2018overcoming}. The authors propose a learning strategy to overcome language priors in VQA models. They first introduce an adversary question-only branch. It takes as input the question encoding from the VQA model and produces a question-only loss. They use a gradient negation of this loss to discourage the question encoder to capture unwanted biases that could be exploited by the VQA model. They also propose a loss based on the difference of entropies between the VQA model and the question-only branch output distributions. These two losses are only backpropagated to the question encoder.
In contrast, our learning strategy targets the full VQA model parameters to reduce the impact of unwanted biases more effectively.
Instead of relying on these two additional losses, we use the question-only branch to dynamically adapt the value of the classification loss in order to reduce the learning of biases in the VQA model. A visual comparison between \cite{ramakrishnan2018overcoming} and RUBi can be found in Figure~\ref{fig:comparison-overcoming} in the supplementary materials.
\section{Reducing Unimodal Biases Approach}
We consider the common formulation of the Visual Question Answering (VQA) task as a multi-class classification problem. Given a dataset $\mathcal{D}$ consisting of $n$ triplets $(v_i,q_i,a_i)_{i \in [1,n]}$ with $v_i \in \mathcal{V}$ an image, $q_i \in \mathcal{Q}$ a question in natural language and $a_i \in \mathcal{A}$ an answer, one must optimize the parameters $\theta$ of the function $f: \mathcal{V} \times \mathcal{Q} \rightarrow \mathbb{R}^{|\mathcal{A}|}$ to produce accurate predictions. For a single example, VQA models use an image encoder $e_v: \mathcal{V} \rightarrow \mathbb{R}^{n_v \times d_v}$ to output a set of $n_v$ vectors of dimension $d_v$, a question encoder $e_q: \mathcal{Q} \rightarrow \mathbb{R}^{n_q \times d_q}$ to output a set of $n_q$ vectors of dimension $d_q$, a multimodal fusion $m: \mathbb{R}^{n_v \times d_v} \times \mathbb{R}^{n_q \times d_q} \rightarrow \mathbb{R}^{d_{m}}$, and a classifier $c: \mathbb{R}^{d_{m}} \rightarrow \mathbb{R}^{|\mathcal{A}|}$.
These functions are composed as follows:
\vspace{-0.1cm}
\begin{equation}
f(v_i, q_i) = c(m(e_v(v_i), e_q(q_i)))
\label{eq:mm}
\end{equation}
Each one of them can be defined to instantiate most of the state of the art models, such as \cite{yang2016stacked, lu2016hierarchical, kim2018bilinear,benyounescadene2017mutan, BenYounes_2019_AAAI,yu2018beyond,Cadene_2019_CVPR} to cite a few.
\vspace{-0.3cm}
\paragraph{Classical learning strategy and pitfall}
The classical learning strategy of VQA models, depicted in Figure~\ref{fig:model_question}, consists in minimizing the standard cross-entropy criterion over a dataset of size $n$.
\vspace{-0.1cm}
\begin{equation}
\mathcal{L}(\theta;\mathcal{D}) = - \frac{1}{n} \sum^{n}_{i=1} \log(\text{softmax}(f(v_i, q_i))) [a_i]
\label{eq:mm_loss}
\end{equation}
VQA models are inclined to learn unimodal biases from the datasets \cite{vqa-cp}. This can be shown by evaluating models on datasets that have different distributions of answers for the test set, such as VQA-CP v2. In other words, they rely on statistical regularities from one modality to provide accurate predictions without having to consider the other modality. As an extreme example, strongly biased models towards the question modality always output \textit{yellow} to the question \textit{what color is the banana}. They do not learn to use the image information because there are too few examples in the dataset where the banana is not yellow. Once trained, their inability to use the two modalities adequately makes them inoperable on data coming from different distributions such as real-world data.
Our contribution consists in modifying this cost function to avoid the learning of these biases.
\subsection{RUBi learning strategy}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{images/model.png}
\caption{\label{fig:model_question} Visual comparison between the classical learning strategy of a VQA model and our RUBi learning strategy. The red highlighted modules are removed at the end of the training. The output $\hat{a}_i$ is used as the final prediction.}
\end{figure}
\paragraph{Capturing biases with a question-only branch}
One way to measure the unimodal biases in VQA datasets is to train an unimodal model which takes only one of the two modalities as input.
The key idea of our approach, depicted in Figure~\ref{fig:model_question}, is to adapt a question-only model as a branch of our VQA model, that will alter the main model's predictions. By doing so, the question-only branch captures the question biases, allowing the VQA model to focus on the examples that cannot be answered correctly using the question modality only.
The question-only branch can be formalized as a function $f_Q: \mathcal{Q} \rightarrow \mathbb{R}^{|\mathcal{A}|}$ parameterized by $\theta_Q$, and composed of a question encoder $e_q: \mathcal{Q} \rightarrow \mathbb{R}^{n_q \times d_q}$ to output a set of $n_q$ vectors of dimension $d_q$, a neural network $\mathit{nn}_q$: $\mathbb{R}^{n_q \times d_q} \rightarrow \mathbb{R}^{|\mathcal{A}|}$ and a classifier $c_q$: $\mathbb{R}^{|\mathcal{A}|} \rightarrow \mathbb{R}^{|\mathcal{A}|}$.
\begin{align}
f_Q(q_i) &= c_q(\mathit{nn}_q(e_q(q_i)))
\label{eq:question-only}
\end{align}
During training, the branch acts as a proxy preventing any VQA model of the form presented in Equation~\eqref{eq:mm} from learning biases. At the end of the training, we simply remove the branch and use the predictions from the base VQA model.
\paragraph{Preventing biases by masking predictions}
\begin{figure}
\hspace{0.3cm}
\begin{subfigure}[t]{0.35\linewidth}
\includegraphics[width=0.99\linewidth]{images/model_loss_classic_low.png}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.50\linewidth}
\includegraphics[width=0.99\linewidth]{images/model_loss_rubi_low.png}
\end{subfigure}
\hspace{0.3cm}
\vspace{0.4cm}
\hspace{0.3cm}
\begin{subfigure}[t]{0.35\linewidth}
\includegraphics[width=0.99\linewidth]{images/model_loss_classic_high.png}
\caption{Classical learning strategy}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.5\linewidth}
\includegraphics[width=0.99\linewidth]{images/model_loss_rubi_high.png}
\caption{RUBi learning strategy}
\end{subfigure}
\hspace{0.3cm}
\caption{Detailed illustration of the RUBi impact on the learning. In the first row, we illustrate how RUBi reduces the loss for examples that can be correctly answered without looking at the image. In the second row, we illustrate how RUBi increases the loss for examples that cannot be answered without using both modalities.}
\label{fig:model_loss}
\end{figure}
Before passing the predictions of our base VQA model to the loss function defined in Equation~\eqref{eq:mm_loss}, we merge them with a mask of length $|\mathcal{A}|$ containing a scalar value between 0 and 1 for each answer.
This mask is obtained by passing the output of the neural network $\mathit{nn}_q$ through a sigmoid function $\sigma$.
The goal of this mask is to dynamically alter the loss by modifying the predictions of the VQA model. To obtain the new predictions, we simply compute an element-wise product $\odot$ between the mask and the original predictions as defined in the following equation.
\begin{align}
f_{QM}(v_i, q_i) &= f(v_i,q_i) \odot \sigma(\mathit{nn}_q(e_q(q_i))))
\label{eq:late-fusion-branch})
\end{align}
Our method modifies the predictions in this specific way to prevent the VQA model to learn biases from the question. To better understand the impact of our approach on the learning, we examine two scenarios.
First, we reduce the importance of the most biased examples, i.e. examples that can be correctly classified without using the image modality. To do so, the question-only branch outputs a mask to increase the score of the correct answer while decreasing the scores of the others. As a result, the loss is much lower for these biased examples. In other words, the gradients backpropagated through the VQA model are smaller, thereby reducing the importance of these examples in the learning.
As illustrated in the first row of Figure~\ref{fig:model_loss}, given the question \textit{what color is the banana}, the mask takes a high value of 0.8 for the answer \textit{yellow} which is the most likely answer for this question in the training set. On the other hand, the value for the other answers \textit{green} and \textit{white} are smaller. We see that the mask influences the VQA model to produce new predictions where the score associated with the answer \textit{yellow} increases from 0.8 to 0.94. Compared to the classical learning approach, the loss is smaller with RUBi and decreases from 0.22 to 0.06.
Secondly, we increase the importance of examples that cannot be answered without using both modalities. For these examples, the question-only branch outputs a mask that increases the score of the wrong answer. As a result, the loss is much higher and the VQA model is encouraged to learn from these examples. We illustrate this behavior in the second row of Figure~\ref{fig:model_loss} for the same question about the color of the banana. When the image contains a green banana, RUBi increases the loss from 0.69 to 1.20.
\paragraph{Joint learning procedure}
We jointly optimize the parameters of the base VQA model and its question-only branch using the gradients computed from two losses.
The main loss $\mathcal{L}_{QM}$ refers to the cross-entropy loss associated with the predictions of $f_{QM}(v_i, q_i)$ from Equation~\ref{eq:late-fusion-branch}. We backpropagate this loss to optimize all the parameters $\theta_{QM}$ which contributed to this loss.
$\theta_{QM}$ is the union of the parameters of the base VQA model, the encoders, and the neural network $\mathit{nn}_q$ of the question-only branch.
In our setup, we share the parameters of the question encoder $e_q$ between the VQA model and the question-only branch.
The question-only loss $\mathcal{L}_{QO}$ is a cross-entropy loss associated with the predictions of $f_Q(q_i)$ from Equation~\ref{eq:question-only}.
We use this loss to only optimize $\theta_{QO}$, union of the parameters of $c_q$ and $\mathit{nn}_q$.
By doing so, we further improve the question-only branch ability to capture biases. Note that we do not backpropagate this loss to the question encoder $e_q$ preventing it from directly learning question biases.
We obtain our final loss $\mathcal{L}_{\text{RUBi}}$ by summing the two losses together in the following equation:
\begin{align}
\mathcal{L}_{\text{RUBi}}(\theta_{QM}, \theta_{QO} ; \mathcal{D}) &= \mathcal{L}_{QM}(\theta_{QM} ; \mathcal{D}) + \mathcal{L}_{QO}(\theta_{QO} ; \mathcal{D})
\label{eq:all_loss}
\end{align}
\subsection{Baseline architecture}
Most VQA architectures from the state of the art are compatible with our RUBi learning strategy. To test our strategy, we design a fast and simple architecture inspired from \cite{Cadene_2019_CVPR}. This baseline architecture is detailed in the supplementary material.
As common in the state of the art, our baseline architecture encodes the image as a bag of $n_v$ visual features $\mathbf{v}_i \in \mathbb{R}^{d_v}$ using the pretrained Faster R-CNN by \cite{Anderson_2018_CVPR}, and encodes the question as a vector $\mathbf{q} \in \mathbb{R}^{d_q}$ using a GRU, pretrained on the skipthought task \cite{kiros2015skip}.
The VQA model consists of a Bilinear BLOCK fusion \cite{BenYounes_2019_AAAI} which merges the question representation $\mathbf{q}$ with the features $\mathbf{v}_i$ of each region of the image. The output is aggregated using a max pooling on the $n_v$ regions. The resulting vector is then fed into a MLP classifier which outputs the final predictions.
While most of our experiments are done with this fast and simple baseline architecture, we experimentally demonstrate that the RUBi learning strategy is effective on other VQA architectures.
\section{Experiments}
\paragraph{Experimental setup}
We train and evaluate our models on VQA-CP v2 \cite{vqa-cp}. This dataset was developed to evaluate the models robustness to question biases.
We follow the same training and evaluation protocol as \cite{ramakrishnan2018overcoming}, who also propose a learning strategy to reduce biases. For each model, we report the standard VQA evaluation metric \cite{VQA}. We also evaluate our models on the standard VQA v2 \cite{VQA2_Goyal_2017_CVPR}. Further implementation details are included in the supplementary materials, as well as results on VQA-CP v1 and grounding experiments on VQA-HAT \cite{vqahat}.
\subsection{Results}
\paragraph{State-of-the-art comparison}
In Table~\ref{tab:vqacpv2_sota}, we compare our approach consisting of our baseline architecture trained with RUBi on VQA-CP v2 against the state of the art. To be fair, we only report approaches that use the strong visual features from \cite{Anderson_2018_CVPR}. We compute the average accuracy over 5 experiments with different random seeds.
Our RUBi approach reaches an average overall accuracy of 47.11\% with a low standard deviation of $\pm$0.51. This accuracy corresponds to a gain of +5.94 percentage points over the current state-of-the-art UpDn + Q-Adv + DoE.
It also corresponds to a gain of +15.88 over GVQA \cite{vqa-cp}, which is a specific architecture designed for VQA-CP.
RUBi reaches a +8.65 improvement over our baseline model trained with the classical cross-entropy. In comparison, the second best approach UpDn + Q-Adv + DoE only achieves a +1.43 gain in overall accuracy over their baseline UpDn. In addition, our approach does not significantly reduce the accuracy over our baseline for the answer type \textit{Other}, while the second best approach reduces it by 10.57 point.
\paragraph{Additional baselines}
We compare our results to two sampling-based training methods. In the \textit{Balanced Sampling} method, we sample the questions such that the answer distribution is uniform. In the \textit{Question-Type Balanced Sampling} method, we sample the questions such that for every question type, the answer distribution is uniform, but the question type distribution remains the same overall
Both methods are tested with our baseline architecture.
We can see that the \textit{Question-Type Balanced Sampling} improves the result from 38.46 in accuracy to 42.11. This gain is already +0.94 higher than the previous state of the art method \cite{ramakrishnan2018overcoming}, but remains significantly lower than our proposed method.
\begin{table}[t]
\centering
\caption{State-of-the-art results on VQA-CP v2 \texttt{test}. All reported models use the same features from \cite{Anderson_2018_CVPR}. Models with * have been trained by \cite{ramakrishnan2018overcoming}. Models with ** have been trained by \cite{shrestha2019answer}.}.\vspace{4pt}
\label{tab:vqacpv2_sota}
\centering
\begin{tabular}{l c cccc}
\toprule
\multirow{2}{*}{Model} && \multirow{2}{*}{Overall} & \multicolumn{3}{c}{\footnotesize Answer type} \\
\cmidrule(r){4-6}
& && \footnotesize Yes/No & \footnotesize Number & \footnotesize Other \\
\midrule
Question-Only \cite{vqa-cp} &&
15.95 & 35.09 & 11.63 & 7.11 \\
UpDn \cite{Anderson_2018_CVPR} ** &&
38.01 & . & . & . \\
RAMEN \cite{shrestha2019answer} &&
39.21 & . & . & . \\
BAN \cite{kim2018bilinear} ** &&
39.31 & . & . & . \\
MuRel \cite{Cadene_2019_CVPR} &&
39.54 & 42.85 & 13.17 & 45.04 \\
UpDn \cite{Anderson_2018_CVPR} * &&
39.74 & 42.27 & 11.93 & \textbf{46.05} \\%&
UpDn + Q-Adv + DoE \cite{ramakrishnan2018overcoming} &&
41.17 & 65.49 & 15.48 & 35.48 \\
\midrule
Balanced Sampling && 40.38 & 57.99 & 10.07 & 39.23 \\
Q-type Balanced Sampling && 42.11 & 61.55 & 11.26 & 40.39 \\
\midrule
Baseline architecture (ours) &&
38.46 $\pm$ 0.07 & 42.85 $\pm$ 0.18 & 12.81 $\pm$ 0.20 & 43.20 $\pm$ 0.15 \\
RUBi (ours) &&
\textbf{47.11} $\pm$ 0.51 & \textbf{68.65} $\pm$ 1.16 & \textbf{20.28} $\pm$ 0.90 & 43.18 $\pm$ 0.43 \\
\bottomrule
\end{tabular}
\end{table}
\paragraph{Architecture agnostic}
RUBi can be used on existing VQA models without changing the underlying architecture. In Table~\ref{tab:vqa2_archs}, we experimentally demonstrate the generality and effectiveness of our learning scheme by showing results on two additional architectures, Stacked Attention Networks (SAN) \cite{yang2016stacked} and Bottom-Up and Top-Down Attention (UpDn) \cite{Anderson_2018_CVPR}.
First, we show that applying RUBi on these architectures leads to important gains over the baselines trained with their original learning strategy.
We report a gain of +11.73 accuracy point for SAN and +4.5 for UpDn. This lower gap in accuracy may show that UpDn is less driven by biases than SAN. This is consistent with results from \cite{ramakrishnan2018overcoming}.
Secondly, we show that these architectures trained with RUBi obtain better accuracy than with the state-of-the-art strategy from \cite{ramakrishnan2018overcoming}. We report a gain of +3.4 with SAN + RUBi over SAN + Q-Adv + DoE, and +3.06 with UpDn + RUBi over UpDn + Q-Adv + DoE. Full results splitted by question type are available in the supplementary materials.
\begin{table}[t]
\vspace{-0.3cm}
\centering
\begin{minipage}[t]{0.55\linewidth}
\centering
\caption{Effectiveness of the RUBi learning strategy when used on different architectures on VQA-CP v2 \texttt{test}. Detailed results can be found in the supplementary materials.}
\vspace{4pt}
\label{tab:vqa2_archs}
\centering
\renewcommand{\arraystretch}{1.11}
\setlength{\tabcolsep}{3pt}
\hspace*{-2mm}
\begin{tabular}{c c}
\begin{tabular}{l c}
\toprule
SAN & Overall\\
\midrule
Baseline \cite{yang2016stacked} & 24.96 \\
+ Q-Adv + DoE \cite{ramakrishnan2018overcoming} & 33.29 \\
+ RUBi (ours) & 37.63 \\
\bottomrule
\end{tabular}
\begin{tabular}{l c}
\toprule
UpDn & Overall \\
\midrule
Baseline \cite{Anderson_2018_CVPR} & 39.74 \\
+ Q-Adv + DoE \cite{ramakrishnan2018overcoming} & 41.17 \\
+ RUBi (ours) & \textbf{44.23} \\
\bottomrule
\end{tabular}
\end{tabular}
\end{minipage}
\hfill
\begin{minipage}[t]{0.35\linewidth}
\setlength\tabcolsep{1.5pt}
\centering
\caption{Overall accuracy of the RUBi learning strategy on VQA v2 \texttt{val} and \texttt{test-dev} splits.}.
\vspace{4pt}
\label{tab:vqa2}
\centering
\begin{tabular}{l cc}
\toprule
Model & \texttt{val} & \texttt{test-dev} \\
\midrule
Baseline (ours) & \textbf{63.10} & \textbf{64.75} \\
RUBi (ours) & 61.16 & 63.18 \\
\bottomrule
\end{tabular}
\end{minipage}
\end{table}
\paragraph{Impact on VQA v2}
We report the impact of our method on the standard VQA v2 dataset in Table~\ref{tab:vqa2}. VQA v2 train, val and test sets follow the same distribution, contrarily to VQA-CP v2 train and test sets.
In this context, we usually observe a drop in accuracy using approaches focused on reducing biases. This is due to the fact that exploiting unwanted correlations from the VQA v2 train set is not discouraged and often leads to a higher accuracy on the test set.
Nevertheless, our RUBi approach leads to a comparable drop to what can be seen in the state-of-the-art.
We report a drop of 1.94 percentage points with respect to our baseline, while \cite{vqa-cp} report a drop of 3.78 between GVQA and their SAN baseline. \cite{ramakrishnan2018overcoming} report drops of 0.05, 0.73 and 2.95 for their three learning strategies with the UpDn architecture which uses the same visual features as RUBi.
As shown in this section, RUBi improves the accuracy on VQA-CP v2 from a large margin, while maintaining competitive performance on the standard VQA v2 dataset compared to similar approaches.
\paragraph{Validation of the masking strategy}
We compare different fusion techniques to combine the output of $\mathit{nn}_q$ with the output from the VQA model. We report a drop of 7.09 accuracy point on VQA-CP v2 by replacing the sigmoid with a ReLU on our best scoring model. Using an element-wise sum instead of an element-wise product leads to a further performance drop.
These results confirm the effectiveness of our proposed masking method which relies on a sigmoid and an element-wise sum.
\paragraph{Validation of the question-only loss}
In Table~\ref{tab:ablation-lqo}, we validate the ability of the question-only loss $\mathcal{L}_{QO}$ to reduce the question biases. The absence of $\mathcal{L}_{QO}$ implies that the question-only classifier $c_q$ is never used, and $\mathit{nn}_q$ only receives gradients from the main loss $\mathcal{L}_{QM}$. Using $\mathcal{L}_{QO}$ leads to consistent gains on all three architectures. We report a gain of +0.89 for our Baseline architecture, +0.22 for SAN, +4.76 for UpDn.
\begin{table}[h]
\centering
\begin{tabular}{l ccccc}
\toprule
Model & $\mathcal{L}_{QO}$ & Overall & Yes/No & Number & Other \\
\midrule
\multirow{2}{*}{Baseline + RUBi}
& \cmark & \textbf{47.11} & 68.65 & 20.28 & \textbf{43.18} \\
& \xmark & 46.11 & \textbf{69.18} & \textbf{26.85} & 39.31 \\
\midrule
\multirow{2}{*}{SAN + RUBi}
& \cmark & \textbf{37.63} & 59.49 & \textbf{13.71} & \textbf{32.74} \\
& \xmark & 36.96 & \textbf{59.7}8 & 12.55 & 31.69 \\
\midrule
\multirow{2}{*}{UpDn + RUBi}
& \cmark & \textbf{44.23} & \textbf{67.05} & \textbf{17.48} & \textbf{39.61}\ \\
& \xmark & 39.47 & 60.27 & 16.01 & 35.01 \\
\bottomrule
\end{tabular}
\vspace{0.2cm}
\caption{Ablation study of the question-only loss $\mathcal{L}_{QO}$ on VQA-CP v2.}
\label{tab:ablation-lqo}
\end{table}{}
\subsection{Qualitative analysis}
To better understand the impact of our RUBi approach, we compare in Figure~\ref{fig:qualitative} the answer distribution on VQA-CP v2 for some specific question patterns. We also display interesting behaviors on some examples using attention maps extracted as in \cite{Cadene_2019_CVPR}.
In the first row, we show the ability of RUBi to reduce biases for the \textit{is this person skiing} question pattern. Most examples in the train set have the answer \textit{yes}, while in the test set, they have the answer \textit{no}. Nevertheless, RUBi outputs 80\% of \textit{no}, while the baseline almost always outputs \textit{yes}. Interestingly, the best scoring region from the attention map of both models is localized on the shoes. To get the answer right, RUBi seems to reason about the absence of skis in this region. It seems that our baseline gets it wrong by not seeing that the skis are not locked under the ski boots. This unwanted behavior could be due to the question biases.
In the second row, similar behaviors occur for the \textit{what color are the bananas} question pattern. 80\% of the answers from the train set are \textit{yellow}, while most of them are \textit{green} in the test set. We show that the amount of \textit{green} and \textit{white} answers from RUBi are much closer to the ones from the test set than with our baseline. In the example, it seems that RUBi relies on the color of the banana, while our baseline misses it.
In the third row, it seems that RUBi is able to ground the textual concepts such as \textit{top part of the fire hydrant} and \textit{color} on the right visual region, while the baseline relies on the correlations between the fire hydrant, the yellow color of its core and the answer \textit{yellow}.
Similarly on the fourth row, RUBi grounds \textit{color}, \textit{star}, \textit{fire hydrant} on the right region, while our baseline relies on correlations between \textit{color}, \textit{fire hydrant}, the yellow color of the top part region and the answer \textit{yellow}. Interestingly, there is no similar question that involves the color of a star on a fire hydrant in the training set. It shows the capacity of RUBi to generalize to unseen examples by composing and grounding existing visual and textual concepts from other kinds of question patterns.
\begin{figure}[!h]
\centering
\includegraphics[width=1.0\linewidth]{images/qualitative-blur.png}
\caption{\label{fig:qualitative} Qualitative comparison between the outputs of RUBi and our baseline on VQA-CP v2 \texttt{test}.
On the left, we display distributions of answers for the train set, the baseline evaluated on the test set, RUBi on the test set and the ground truth answers from the test set. For each row, we filter questions in a certain way. In the first row, we keep the questions that exactly match the string \textit{is this person skiing}. In the three other rows, we filter questions that respectively include the following words: \textit{what color bananas}, \textit{what color fire hydrant} and \textit{what color star hydrant}.
On the right, we display examples that contains the pattern from the left. For each example, we display the answer of our baseline and RUBi, as well as the best scoring region from their attention map.}
\end{figure}
\section{Supplementary materials}
\label{seq:visual-comparison}
\begin{figure}[h]
\hspace{0.3cm}
\begin{subfigure}[t]{0.45\linewidth}
\includegraphics[width=0.85\linewidth]{images/model_rubi.png}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.49\linewidth}
\includegraphics[width=0.80\linewidth]{images/model_25.png}
\label{fig:model_overcoming}
\end{subfigure}
\caption{Visual comparison between RUBi and \cite{ramakrishnan2018overcoming}.}
\label{fig:comparison-overcoming}
\end{figure}
\subsection{Additional experiments}
\label{sec:additional-exp}
\paragraph{Results on VQA-CP v1}
\label{sec:vqacp1-resuults}
In Table \ref{vqacp1-full}, we report results on the VQA-CP v1 dataset \cite{vqa-cp}. Our RUBi approach consistently leads to significant gains over the classical learning strategy with a gain of +9.8 overall accuracy point with our baseline architecture, +19.2 with SAN and +7.66 with UpDn. Additionally, RUBi leads to a gain of +2.65 over the adversarial regularization method (AdvReg) from \cite{ramakrishnan2018overcoming} with SAN. A visual comparison between RUBi and \cite{ramakrishnan2018overcoming} can be found in Figure~\ref{fig:comparison-overcoming}. Finally, all three architectures trained with RUBi reach a higher accuracy than GVQA \cite{vqa-cp} which has been hand-designed to overcome biases.
\begin{table}[h!]
\centering
\begin{tabular}{lllll}
\toprule
Model & Overall & Yes/No & Number & Other \\
\midrule
GVQA \cite{vqa-cp} & 39.23 & 64.72 & 11.87 & 24.86 \\
\midrule
Baseline (ours) & 37.13 & 41.96 & 12.54 & 41.35 \\
Baseline + RUBi & \textbf{46.93} & \textbf{66.78} & \textbf{20.98} & \textbf{43.64} \\
\midrule
SAN \cite{ramakrishnan2018overcoming} & 26.88 & 35.34 & 11.34 & 24.70 \\
SAN + AdvReg \cite{ramakrishnan2018overcoming} & 43.43 & 74.16 & 12.44 & 25.32 \\
SAN + RUBi & \textbf{46.08} & \textbf{75.00} & \textbf{13.30} & \textbf{30.49} \\
\midrule
UpDn (ours) & 37.15 & 41.13 & 12.73 & \textbf{43.00 } \\
UpDn + RUBi & \textbf{44.81} & \textbf{69.65} & \textbf{14.91} & 32.13 \\
\bottomrule
\end{tabular}
\vspace{0.1cm}
\caption{Overall accuracy top1 on VQA-CP v1}
\label{vqacp1-full}
\end{table}
\paragraph{Detailed results on VQA-CP v2}
\label{sec:vqacp2-results-full}
In Table~\ref{vqacp2-full}, we report the full results of our experiments for SAN and UpDn architectures on the VQA-CP v2 dataset.
\begin{table}[h!]
\centering
\begin{tabular}{lllll}
\toprule
Model & Overall & Yes/No & Number & Other \\
\midrule
SAN \cite{yang2016stacked} & 24.96 & 38.35 & 11.14 & 21.74 \\
%
SAN + RUBi & \textbf{37.63} & \textbf{59.49} & \textbf{13.71} & \textbf{32.74} \\
\midrule
UpDn \cite{Anderson_2018_CVPR} & 39.74 & 42.27 & 11.93 & \textbf{46.05} \\
UpDn + RUBi & \textbf{44.23} & \textbf{67.05} & \textbf{17.48} & 39.61 \\
\bottomrule
\end{tabular}
\vspace{0.1cm}
\caption{Overall accuracy top1 on VQA-CP v2 for the SAN and UpDn architectures.}
\label{vqacp2-full}
\end{table}
\paragraph{Quantitative study of the grounding ability on VQA-HAT}
We conduct additional studies to evaluate the grounding ability of models trained with RUBi. We follow the experimental protocol of VQA-HAT \cite{vqahat}. We train our models on VQA v1 train set and evaluate them using rank-correlation on the VQA-HAT val set, which is a subset of the VQA v1 val set. This metric compares attention maps computed from a model against human annotations indicating which regions humans found relevant for answering the question.
In Table~\ref{tab:vqahat}, we report a gain of +0.012 with our baseline architecture trained with RUBi, a gain of +0.019 with SAN and a loss of -0.003 with UpDn architecture. In future works, we would like to go beyond these early results in order to further evaluate the impact on grounding induced by RUBi.
\begin{table}[h]
\centering
\begin{tabular}{l cc}
\toprule
Model & RUBi & Rank-Corr. \\
\midrule
Random \cite{vqahat} & & 0.000 \\
Human \cite{vqahat} & & 0.623 \\
\midrule
\multirow{2}{*}{Baseline}
& \xmark & 0.431 \\
& \cmark & \textbf{0.443} \\
\midrule
\multirow{2}{*}{SAN}
& \xmark & 0.191\\
& \cmark & \textbf{0.210}\\
\midrule
\multirow{2}{*}{UpDn}
& \xmark & \textbf{0.449} \\
& \cmark & 0.446 \\
\bottomrule
\end{tabular}
\vspace{0.1cm}
\caption{Correlation with Human Attention Maps on VQA-HAT \texttt{val} set \cite{vqahat}.}
\label{tab:vqahat}
\end{table}{}
\paragraph{Qualitative study of the grounding ability on VQA-HAT}
We display in Figure~\ref{fig:vqahat_good} and Figure~\ref{fig:vqahat_bad} some manually selected VQA triplets associated to the human attention maps provided by VQA-HAT \cite{vqahat} and the attention maps computed from our baseline architecture when trained with and without RUBi. In Figure~\ref{fig:vqahat_good}, we observe that the attention maps with RUBi are closer to the human attention maps than without RUBi. On the contrary, we observe in Figure~\ref{fig:vqahat_bad} some failure to improve grounding ability.
\begin{figure}[!h]
\centering
\includegraphics[width=1.0\linewidth]{images/vqahat/3430380_blur.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/2187511.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/4350371.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/4516741.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/4794400.png}
\caption{Examples of better grounding ability on VQA-HAT implied by RUBi. From the left column to the right: image-question-answer triplet, human attention map from \cite{vqahat}, attention map from our baseline, attention map from our baseline trained with RUBi.}
\label{fig:vqahat_good}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=1.0\linewidth]{images/vqahat-bad/2919811.png}
\includegraphics[width=1.0\linewidth]{images/vqahat-bad/496881_blur.png}
\includegraphics[width=1.0\linewidth]{images/vqahat-bad/4662730.png}
\includegraphics[width=1.0\linewidth]{images/vqahat-bad/5399372.png}
\includegraphics[width=1.0\linewidth]{images/vqahat-bad/5789641.png}
\caption{Examples of failure to improve grounding ability on VQA-HAT. From the left column to the right: image-question-answer triplet, human attention map from \cite{vqahat}, attention map from our baseline, attention map from our baseline trained with RUBi.}
\label{fig:vqahat_bad}
\end{figure}
\iffalse
\begin{figure}[!h]
\centering
\includegraphics[width=1.0\linewidth]{images/vqahat/1394721.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/2187511.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/2296532.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/2232410.png}
\includegraphics[width=1.0\linewidth]{images/vqahat/2770321.png}
\caption{}
\label{fig:vqahat_good2}
\end{figure}
\fi
\clearpage
\subsection{Implementation details}
\paragraph{Image encoder} We use the pretrained Faster R-CNN by \cite{Anderson_2018_CVPR} to extract object features. We use the setup that extracts 36 regions for each image. We do not fine-tune the image extractor.
\paragraph{Question encoder}
We use the same preprocessing as in \cite{Cadene_2019_CVPR}. We apply a lower case transformation and remove the punctuation. We only consider the most frequent 3000 answers for both VQA v2 and VQA CP v2. We then use a pretrained Skip-thought encoder with a two-glimpses self attention mechanism. The final embedding is of size 4800. We fine-tune the question encoder during training.
\paragraph{Baseline architecture}
Our baseline architecture is a simplified version of the MuRel architecture \cite{Cadene_2019_CVPR}.
First, it computes a bilinear fusion between the question vector and the visual features for each region.
The bilinear fusion module is a BLOCK \cite{BenYounes_2019_AAAI} composed of 15 chunks, each of rank 15.
The dimension of the projection space is 1000, and the output dimension is 2048.
The output of the bilinear fusion is aggregated using a max pooling over $n_v$ regions.
The resulting vector is then fed into a MLP classifier composed of three layers of size (2048, 2048, 3000), with ReLU activations.
It outputs the predictions over the space of the 3000 answers.
\paragraph{Question-only branch}
The RUBi question-only branch feeds the question into a first MLP composed of three layers, of size (2048, 2048, 3000), with ReLU activations.
First, this output vector goes through a sigmoid to compute the mask that will alter the predictions of the VQA model.
Secondly, this same output vector goes through a single linear layer of size 3000. We use these question-only predictions to compute the question-only loss.
\paragraph{Optimization process}
We train all our models with the Adam optimizer. We train our baseline architecture with the learning rate scheduler of \cite{Cadene_2019_CVPR}. We use a learning rate of $1.5 \times 10^{-4}$ and a batch size of 256. During the first 7 epochs, we linearly increase the learning rate to $6 \times 10^{-4}$. After epoch 14, we apply a learning rate decay strategy which multiplies the learning rate by 0.25 every two epochs. We train our models until convergence as we do not have a validation set for VQA-CP v2. For the UpDn and SAN architectures, we follow the optimization procedure described in \cite{ramakrishnan2018overcoming}.
\paragraph{Software and hardware}
We use pytorch 1.1.0 to implement our algorithms in order to benefit from the GPU acceleration. We use four NVidia Titan Xp GPU in this study. We use a single GPU for each experiments. We use a dedicated SSD to load the visual features using multiple threads.
A single experiment from Table~1 with the baseline architecture trained with or without RUBi takes less than five hours to run.
\iffalse
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{images/arch.png}
\caption{\label{fig:arch} Illustration of our fast and simple baseline architecture to test our RUBi learning strategy. We obtain the attention map using argmax over the outputs of the BLOCK fusion. }
\end{figure}
\fi
|
\section{Notation and Definitions}
\label{two}
Before the formulation of the main problem we have to say here a few words more about notation used in this manuscript. In this section and in our further considerations by $\mathcal{B}(\mathcal{H})$ we denote the algebra of all bounded linear operators on the Hilbert space $\mathcal{H}$, and we take $\mathcal{H}\cong \mathbb{C}^d$. Using this notation let us define the following set:
\be
\label{Qset}
\mathcal{S}(\mathcal{H})=\{\rho \in \mathcal{B}(\mathcal{H}) \ | \ \rho \geq 0, \ \tr \rho=1\},
\ee
which is the set of all states on a space $\mathcal{H}$. Let us now suppose, that we are given with the state $\rho_{A_1\cdots A_m}$ of $m$ subsystems $A_1,\ldots,A_m$ defined on the Hilbert space $\mathcal{H}_{A_1\cdots A_m}=\mathcal{H}_{A_1}\ot\cdots \ot \mathcal{H}_{A_m}$. Defining the partition $p\equiv \left\lbrace X_1,\ldots,X_k \right\rbrace $, where $X_i$ are disjoint subsets of the indices $X=\{1,\ldots,m\}$, $ \ \cup_{r=1}^k X_r=X$, we can say, that the state $\rho_{A_1\cdots A_m}$ is separable with respect to the partition $p$, if and only if it can be written as
\be
\label{sep1}
\rho_{A_1\cdots A_m}=\sum_{i=1}^Mp_i\rho_{1}^i\ot \cdots \ot \rho_k^i,\qquad \forall i \ p_i \geq 0,\qquad \sum_i p_i=1.
\ee
It is easy to see from the above definition that if $k=n$ we obtain definition of fully separable state, and for $k=m=2$,~\cref{sep1} reduces to the standard bipartite definition of separability. In this paper we consider the multiparitite density matrices on $\mathcal{B}\left(\mathcal{H}_A^{\ot n}\ot \mathcal{H}_B^{\ot n} \right)$ with $\operatorname{dim}\mathcal{H}_A=d_A$ and $\operatorname{dim}\mathcal{H}_B=d_B$.
In the previous papers treating on the similar problem authors deal with separable states acting on $\mathcal{H}_{A_1A_2B_1B_2}$ in the cut $A_1A_2:B_1B_2$. In our multipartite approach ($n \geq 2$) the separable states will be denoted as $\mathcal{SEP}^{(\overline{A},\overline{B})}\equiv \mathcal{\overline{SEP}}$, where $\overline{A}=A_1A_2\ldots A_n$ and similarly for $\overline{B}$.
Moreover, in the further part of this manuscript, whenever we talk about the distance from the set of the separable states we mean the following:
\begin{notation}
Suppose that we are given with a quantum state $\rho \in \mathcal{B}\left(\mathcal{H}_{\bar{A}}\ot \mathcal{H}_{\bar{B}} \right) $ and the set of separable states $\mathcal{\overline{SEP}}$, then by $\operatorname{dist}\left( \rho,\mathcal{\overline{SEP}}\right) $ we understand the following quantity
\be
\label{dd}
\operatorname{dist}\left( \rho,\mathcal{\overline{SEP}}\right) =\frac{1}{2}\mathop{\min}\limits_{\sigma \in \mathcal{\overline{SEP}}}||\rho-\sigma||_1,
\ee
which is the minimal trace distance. In~\cref{dd} by $||\cdot||_1$ we understand the trace norm which is defined as $||X||_1\equiv \tr\left[XX^{\dagger} \right]^{1/2}$ for all matrices $X$.
\end{notation}
As we see later, restriction to bi-separability is not a limiting factor in our problem.
\section{Construction Method}
\label{method}
The main goal of this section is the construction a class of the multipartite entangled operators
\be
\label{1}
\rho \in \mathcal{B}\left(\mathcal{H}_A^{\ot n}\ot \mathcal{H}_B^{\ot n} \right) ,
\ee
having PPT property with respect to any cut of subsystems and are arbitrary far from the set of fully separable states.
Let us define $\forall_k \ \alpha_k \in \{0,1\}$ the binary vector $\mathbold{\alpha}\equiv \left(\alpha_i,\ldots,\alpha_{n-1} \right)$, such that $\mathbold{\alpha} \neq \bf{0}$.
Secondly we define the operators of the partial transposition with respect to an arbitrary cut as
\be
\tau_{\mathbold{\alpha}}\equiv \text{\noindent
\(\mathds{1}\)}\ot \operatorname{T}^{\alpha_1} \ot \cdots \ot \operatorname{T}^{\alpha_{n-1}},
\ee
where $\operatorname{T}^{\alpha_k}$ for $1\leq k \leq n-1$ denotes standard operation of transposition on $k^{\text{th}}$ subsystem, whenever $\alpha_k\neq 0$. If for some $k$ there is $\alpha_k=0$ we define $T^{\alpha_k}=\text{\noindent
\(\mathds{1}\)}_k$.
It is easy to see, that in total we have
\be
\label{N}
N=\sum_{k=1}^{n-1}\binom{n-1}{k}=2^{n-1}-1
\ee
such operators. Next we introduce the operators
\be
E_{ij}^{\ot n}\equiv \underbrace{e_{ij}\ot \cdots \ot e_{ij}}_{n},
\ee
where $\left\lbrace e_{ij}\right\rbrace_{i,j=1}^d$, with $e_{ij}=|i\>\<j|$ denotes the basis of $n \times n$ complex matrices $M(n,\mathbb{C})$. Let us consider the following mixture:
\be
\label{sum1}
\rho=\sum_{l=0}^{ND}\rho^{(l)} \in \mathcal{B}\left(\mathcal{H}_A^{\ot n}\ot \mathcal{H}_B^{\ot n} \right),
\ee
where $D=\frac{1}{2}d_A(d_A-1)$. In the sum given by~\cref{sum1} we can distinguish state $\rho^{(0)}$ in the maximally entangled form
\be
\label{max_form}
\rho^{(0)}=\sum_{i,j=1}^{d_A} E_{ij}^{\ot n}\ot a_{ij}^{(0)},\qquad a_{ij}^{(0)}\in \mathcal{B}\left(\mathcal{H}_B^{\ot n} \right).
\ee
Herein we define the rest of $dN$ operators in such a way, that whole operator $\rho$ admits the PPT property. Observe that $\forall l\neq 0$ there exists a bijection $f:(\mathbold{\alpha},i,j)\rightarrow l$ which allows us to define off-diagonal term of $l^{\text{th}}-$ operators as
\be
\rho_{12}^{(l)}=\tau_{\alpha}\left(E_{ij}^{\ot n} \right) \ot a_{12}^{(l)}, \quad \text{where}\quad a_{12}^{(l)}\in\mathcal{B}\left(\mathcal{H}^{\ot n}_B \right),
\ee
and for given $l$ we have
\be
\rho^{(l)}=\sum_{i,j=1}^2\rho_{ij}^{(l)}.
\ee
Having the form of $\rho_{12}^{(l)}$, together with all sub-blocks $a_{21}^{(l)}, a_{22}^{(l)}, a_{11}^{(l)}\in \mathcal{B}\left(\mathcal{H}_B^{\ot n} \right) $ we have full information about all $\rho^{(l)}$, so we know the structure of $\rho$ from~\cref{sum1}.
Example below illustrates how above construction works in practice.
\begin{example}
Let us consider the case when $n=3$ and the local dimension of the Hilbert space is $d_A=3$, then the total state $\rho \in \mathcal{B}\left( \mathcal{H}_A^{\ot 3} \ot \mathcal{H}_B^{\ot 3}\right) $ from~\cref{1} can be represented as:
\begin{widetext}
\be
\rho=\left(\scalemath{0.65}{\begin{array}{ccc|ccc|ccc !{\vrule width 1.5pt} ccc|ccc|ccc !{\vrule width 1.5pt} ccc|ccc|ccc}
a_{11}^{(0)} & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & a_{12}^{(0)} & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & a_{13}^{(0)}\\
\cdot & a_{11}^{(1)} & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & a_{12}^{(1)} & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & a_{11}^{(2)} & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & a_{12}^{(2)} & \cdot & \cdot\\
\hline
\cdot & \cdot & \cdot & a_{11}^{(3)} & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & a_{12}^{(3)}& \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & a_{11}^{(4)}& \cdot & \cdot& \cdot & \cdot& a_{12}^{(4)} & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot\\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot\\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & a_{11}^{(5)}& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& a_{12}^{(5)} & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & a_{11}^{(6)}& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot& a_{12}^{(6)} & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot\\
\hlinewd{2pt}
\cdot & \cdot & \cdot & \cdot & a_{21}^{(4)}& \cdot & \cdot& \cdot & \cdot& a_{22}^{(4)} & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & a_{21}^{(3)} & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & a_{22}^{(3)}& \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot\\
\hline
\cdot & a_{21}^{(1)} & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & a_{22}^{(1)} & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot \\
a_{21}^{(0)} & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & a_{22}^{(0)} & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & a_{23}^{(0)} \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & a_{11}^{(7)} & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & a_{12}^{(7)} & \cdot\\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot& \cdot & a_{11}^{(8)}& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & a_{12}^{(8)} & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & a_{11}^{(9)}& \cdot & \cdot& \cdot & \cdot& a_{12}^{(9)} & \cdot & \cdot & \cdot & \cdot\\
\hlinewd{2pt}
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & a_{21}^{(6)}& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& a_{22}^{(6)} & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & a_{21}^{(5)}& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot & a_{22}^{(5)} & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot\\
\hline
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot& \cdot & \cdot & a_{21}^{(9)} & \cdot& \cdot & \cdot& \cdot & a_{22}^{(9)} & \cdot & \cdot & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot & a_{21}^{(8)} & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & a_{22}^{(8)} & \cdot & \cdot & \cdot\\
\hline
\cdot & \cdot & a_{21}^{(2)} & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & a_{22}^{(2)} & \cdot & \cdot \\
\cdot & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & a_{21}^{(7)}& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & a_{22}^{(7)} & \cdot \\
a_{31}^{(0)} & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & a_{32}^{(0)} & \cdot & \cdot & \cdot & \cdot& \cdot & \cdot& \cdot & \cdot& \cdot & \cdot & \cdot & \cdot & a_{22}^{(0)}\\
\end{array}}\right),
\ee
\end{widetext}
where dots denote zeros and $a_{ij}^{(l)}\in \mathcal{B}\left(\mathcal{H}_B^{\ot 3} \right)$.
\end{example}
\begin{remark}
From all considerations presented in this section, we see that in particular case, when $n=2$, for particular choice of parameters, we reduce our class of operators to already known examples of pbits ($d_A=2$) and pdits ($d_A=3$), where now $d_A$ plays role of dimension of the key part. For the detailed description we refer reader to~\cite{KH_phd}.
\end{remark}
\begin{remark}
Let us notice that in the recent paper~\cite{Huber} authors consider the similar operator as in~\eqref{sum1} (in unnormalised form it is equivalent to our operator up to swap operation with respect to subsystems) to show that some linear positive maps are non-decomposable.
\end{remark}
\section{Properties of given subclass of states}
To provide our main result we need to consider the following mixture of states:
\be
\label{form}
\rho = p\rho^{(0)}+\frac{q}{ND}\sum_{l=1}^{ND}\rho^{(l)},
\ee
where $p+q=1$, $D=\frac{1}{2}d_A(d_A-1)$ and $N$ is given through eq.~\eqref{N}. Here and later we assume operators $\rho^{(0)},\rho^{(l)}$ to be normalised. Let us notice that states given in~\cref{form} belongs to the class of the states defined in~\eqref{sum1}. Our first results says:
\begin{lemma}
\label{q}
Let us assume that we are given with $\rho\in \mathcal{B}\left(\mathcal{H}_A^{\ot n}\ot \mathcal{H}_B^{\ot n} \right) $ as in~\cref{form}, and the state $\rho^{(0)}$ is in entangled form~\eqref{max_form}, then the following statement holds
\be
||\rho-\rho^{(0)}||_1=2q.
\ee
\end{lemma}
\begin{proof}
We prove above statement by direct calculations, namely we have
\be
\begin{split}
||\rho-\rho^{(0)}||_1&=\left| \left| p\rho^{(0)}+\frac{q}{ND}\sum_{l=1}^{ND}\rho^{(l)}-\rho^{(0)}\right| \right| _1\\
&=\frac{q}{ND}\left| \left| \sum_{l=1}^{ND}\rho^{(l)}-ND\rho^{(0)}\right| \right| _1.
\end{split}
\ee
At this point we use definition of the trace norm and rewrite above equation as
\be
\label{d1}
\begin{split}
&||\rho-\rho^{(0)}||_1=\\
&=\frac{q}{ND}\tr\left[\left(\sum_{l=1}^{ND}\rho^{(l)}-ND\rho^{(0)} \right) \left(\sum_{l=1}^{ND}\rho^{(l)}-ND\rho^{(0)} \right)^{\dagger} \right]^{1/2}.
\end{split}
\ee
Because all operators in~\cref{d1} are hermitian and supported on orthogonal subspaces, we reduce~\cref{d1} to
\be
||\rho-\rho^{(0)}||_1=\frac{q}{ND}\tr\left[\left(\sum_{l=1}^{ND}\rho^{(l)}+ND\rho^{(0)} \right)^2 \right]^{1/2}.
\ee
Making further simplification, finally we obtain:
\be
||\rho-\rho^{(0)}||_1=\frac{q}{ND}\tr\left[\sum_{l=1}^{ND}\rho^{(l)}+ND\rho^{(0)} \right]=2q.
\ee
\end{proof}
Before we go further in our considerations, we make here some specific choice of the interior structure for the states from~\cref{form}. Namely, we assume the following PPT-invariant form of the sub-blocks:
\be
\label{s1}
\begin{split}
&a_{ij}^{(0)}=a=\frac{1}{d_Ad_B^{n}}\begin{pmatrix}
1 & 0 & \cdots & 0 & 0\\
0 & 1 & \cdots & 0 & 0\\
\vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & \cdots & 1 & 0\\
0 & 0 & \cdots & 0 & 1
\end{pmatrix}_{d_B^n \times d_B^n}\\ &a_{rs}^{(l)}=b=\frac{1}{2d_B^{n}}\begin{pmatrix}
1 & 1 & \cdots & 1 & 1\\
1 & 1 & \cdots & 1 & 1\\
\vdots & \vdots & \ddots & \vdots & \vdots \\
1 & 1 & \cdots & 1 & 1\\
1 & 1 & \cdots & 1 & 1
\end{pmatrix}_{d_B^n \times d_B^n},
\end{split}
\ee
where $1\leq i,j \leq d_A$ and for every $ 1\leq l\leq ND$ we have $1 \leq r,s \leq 2$. Matrices $a,b$ have the following spectra:
\be
\label{spectra}
\begin{split}
&\operatorname{spec}\left(a\right)=\left\lbrace \frac{1}{d_Ad_B^n},\ldots, \frac{1}{d_Ad_B^n} \right\rbrace,\\
&\operatorname{spec}\left(b \right) =\left\lbrace \frac{1}{2},0,\ldots,0 \right\rbrace,
\end{split}
\ee
where $0$ in $\operatorname{spec} (b)$ is taken $d_B^n-1$ times.
Now we are in the position to formulate and prove the following:
\begin{lemma}
\label{sq}
Let us consider class of states $\rho$ given by~\cref{form}, together with specific choice of sub-blocks from~\cref{s1}. Then the parameter $q$ describing the trace distance in Lemma~\ref{q} is equal to
\be
q=\frac{1}{1+\frac{d_B^n}{N(d_A-1)}},
\ee
where $N$ is given by expression~\eqref{N}.
\end{lemma}
\begin{proof}
To get the result we need to consider positivity of the state $\rho$ from~\cref{form} before and after partial transposition with respect to every possible cut. Then we use the property that all sub-blocks given in~\cref{s1} are PPT invariant, so their spectra remain unchanged after that partial transposition. Since all the states taken for the construction of $\rho$ are supported on the orthogonal subspaces checking positive semidefiniteness of $\rho$ is equivalent to checking the positive semidefiniteness of the following matrix
\be
\label{X}
X=\begin{pmatrix}
xb & pa\\
pa & xb
\end{pmatrix}\geq 0,
\ee
where $x=\frac{q}{ND}$, $p=1-q$, $D=d_A(d_A-1)/2$ and real matrices $a,b$ are given in by~\cref{s1}. Using analogous observation we deduce condition for positive semidefiniteness of $\rho$ after partial transposition:
\be
\label{Y}
Y=\begin{pmatrix}
pa & xb & \cdots & xb \\
xb & pa & \cdots & xb \\
\vdots & \vdots & \ddots & \vdots\\
xb & xb & \cdots & pa
\end{pmatrix} \geq 0.
\ee
Matrices $X,Y$ can be decomposed as
\be
\label{decomp}
\begin{split}
X&=\text{\noindent
\(\mathds{1}\)}_2\ot xb + \mathbb{I}_2\ot pa - \text{\noindent
\(\mathds{1}\)}_2\ot pa,\\
Y&=\text{\noindent
\(\mathds{1}\)}_n\ot pa + \mathbb{I}_n\ot xb - \text{\noindent
\(\mathds{1}\)}_n\ot xb.\\
\end{split}
\ee
In~\cref{decomp} by $\text{\noindent
\(\mathds{1}\)}_n, \text{\noindent
\(\mathds{1}\)}_2$ we denote $n \times n$, $2 \times 2$ identity matrices respectively By $\mathbb{I}_n,\mathbb{I}_2$ we denote $n \times n, 2\times 2$ matrices filled with ones only. Matrix $\mathbb{I}_n$ has only one non-zero eigenvalue equal to $n$. Because all terms in the decomposition given by~\cref{decomp} commute we can write positivity conditions from~\eqref{X},~\eqref{Y} using components of~\eqref{decomp}:
\be
\label{or}
\begin{split}
\lambda(X)&=x\lambda(b)+2p\lambda(a)-p\lambda(a)\geq 0,\\ \lambda(X)&=x\lambda(b)-p\lambda(a)\geq 0,\\
\lambda(Y)&=p\lambda(a)+nx\lambda(b)-x\lambda(b)\geq 0,\\ \lambda(Y)&=p\lambda(a)-x\lambda(b)\geq 0.
\end{split}
\ee
First and the third inequality in~\eqref{or} are always satisfied. Non-trivial condition we get from the second and the fourth inequality getting
\be
p\lambda(a)-\frac{q}{N(d_A-1)}\lambda(b)= 0.
\ee
Using constraint $p+q=1$ and $D=\frac{1}{2}d_A(d_A-1)$ we derive the parameter $q$:
\be
q=\frac{1}{1+\frac{1}{ND}\frac{\lambda(b)}{\lambda(a)}}.
\ee
Inserting explicit form of eigenvalues $\lambda(a),\lambda(b)$ given in~\eqref{spectra} we get:
\be
q=\frac{1}{1+\frac{d_B^n}{N(d_A-1)}},
\ee
where the number $N$ is given in expression~\eqref{N}.
This finishes the proof.
\end{proof}
The next goal of this section is to compute the trace distance between set of the separable states $\mathcal{\overline{SEP}}$ defined as in~\Cref{method}. Namely, we have the following
\begin{lemma}
\label{fromsep}
The trace distance between class of multipartite states given by
\be
\label{bla}
\rho=p\rho^{(0)}+\frac{q}{ND}\sum_{l=1}^{ND}\rho^{(l)},
\ee
with sub-blocks from~\eqref{s1} is bounded from below as
\be
\operatorname{dist}\left(\rho,\mathcal{\overline{SEP}} \right)\geq 1-\frac{1}{d_A^n}-\frac{1}{1+\frac{d_B^n}{N(d_A-1)}}.
\ee
In the above we take $D=\frac{1}{2}d_A(d_A-1)$, $N=2^{n-1}-1$.
\end{lemma}
\begin{proof}
Adopting the result from~\cite{Badziag_pbits} we arrive at:
\be
\label{sepB}
\operatorname{dist}\left(\rho^{(0)}, \mathcal{\overline{SEP}}\right) \geq 1-\frac{1}{d_A^n}.
\ee
Now let us take the closest separable state $\sigma_{sep}$ to the state $\rho$ given in~\cref{form} with~\cref{s1}, then we have
\be
\label{triangle}
\begin{split}
&||\rho-\sigma_{sep}||_1+||\rho-\rho^{(0)}||_1 \geq \\ &||\sigma_{sep}-\rho^{(0)}||_1\geq 1-\frac{1}{d_A^n},
\end{split}
\ee
so
\be
||\rho-\sigma_{sep}||_1 \geq 1-\frac{1}{d_A^n}-||\rho-\rho^{(0)}||_1.
\ee
In the next step we can use the statement of~\Cref{sq} taking the worst possible choice of $q=1/(1+d_B^n/N(d_A-1))$
\be
\label{final}
\begin{split}
||\rho-\sigma_{sep}||_1\geq 1-\frac{1}{d_A^n}-\frac{1}{1+\frac{d_B^n}{N(d_A-1)}}.
\end{split}
\ee
This finishes the proof.
\end{proof}
\begin{remark}
One can see that since $\overline{\mathcal{SEP}}\subset \mathcal{SEP}$, expression~\eqref{bla} from Lemma~\ref{fromsep} gives us lower bound on the trace distance also from fully separable states as well as from any set of partially separable states.
\end{remark}
Let us notice, that for the generic case for qubits states i.e. $n=2, d_A=2, d_B=2$ our bound achieves minimum.
At the end of this section lest us consider the scenario in which the multipartite PPT sate $\rho$ acts on $\mathcal{B}(\mathbb{C}^d \ot \mathbb{C}^d)$, where $d=d_A^nd_B^n$. Now we show that states from our class can be arbitrary far from the set of separable states $\mathcal{\overline{SEP}}$ in the mentioned cut for the fixed number of subsystems $n$. Namely we have the following:
\begin{theorem}
\label{maxdist}
For every $\epsilon >0 $ there exists PPT state $\rho$ given by~\cref{form} with~\cref{s1} acting on $\mathcal{B}(\mathbb{C}^d \ot \mathbb{C}^d)$, such that
\be
\label{con1}
\operatorname{dist}\left(\rho, \mathcal{\overline{SEP}}\right) \geq 1-\epsilon,
\ee
for $d\leq C(n)/\epsilon^{2+1/n}$, where $C(n)=8N\sqrt[n]{2}$.
\end{theorem}
\begin{proof}
In order to show existence of PPT state which satisfies condition from~\cref{con1} it is enough to consider the worst case scenario by taking $\epsilon/2=1/d_A^n$, and $\epsilon/2=1/(1+d_B^n/(N(d_A-1)))$. Then we get
\be
\begin{split}
d_A^n&=\frac{2}{\epsilon},\\ d_B^n&=N\frac{2-\epsilon}{\epsilon}\left(\sqrt[n]{\frac{2}{\epsilon}}-1 \right).
\end{split}
\ee
Now computing $d=d_k^nd_s^n$ we have
\be
d=d_k^nd_s^n=4N\frac{2-\epsilon}{\epsilon^2}\left(\sqrt[n]{\frac{2}{\epsilon}}-1 \right)\leq \frac{C(n)}{\epsilon^{2+1/n}},
\ee
where $C(n)=8N\sqrt[n]{2}$.
\end{proof}
\begin{remark}
\label{rem10}
To achieve an arbitrary large distance we do not have to consider so called tensoring (or boosting - taking many copies of state) introduced in~\cite{Badziag_pbits}, it is enough only to increase dimension $d_B$ for fixed dimension $d_A$. This is generalizing of theorem given in the paper~\cite{Rut}.
\end{remark}
\begin{remark}
Theorem~\ref{maxdist} can be easily generalized to different cut, i.e. states acting on $\mathcal{B}(\mathbb{C}^d \ot \mathbb{C}^{d'})$, where $d'\neq d$.
\end{remark}
\section{Summary}
In this paper we present a wide class of multipartite entangled density operators which are PPT invariant with respect to every cut, arbitrary far from the set of separable states. Our method of construction in based on mixing multipartite states defined on orthogonal spaces and imposing set of constraints on positive semidefiniteness of the final mixed state before and after the partial transpositions. We discuss distance of any state from our class from the set of separable states as a function of mixing parameter $q$. Afterwards we show, that we are able to construct special, but non-trivial subclass of states, for which we evaluate mentioned distance as the function of dimension $d_B$. As we stated in Remark~\ref{rem10} the novelty of our approach lies in the fact, that to boost the distance between states from our class and the set of separable states we do not have to use many copies of them like in the previous approaches.
Namely we prove, that mentioned distance decreases with growing dimension $d_B$ for fixed $n$. Speaking more precisely, we show that for every $\epsilon >0$, we can find entangled PPT state from our class which is $1-\epsilon$ far from the set of separable states. The scaling of $\epsilon$ with the dimension is $d\propto 1/\epsilon^{2+1/n}$.
\section*{Acknowledgments}
MS is supported by grant "Mobilno{\'s}{\'c} Plus IV", 1271/MOB/IV/2015/0 from Polish Ministry of Science and Higher Education.
\bibliographystyle{unsrt}
|
\section{Introduction}
In \cite{Zar83} Zarhin has shown that the transcendental lattice of K3 surfaces are sometimes endowed with real or complex multiplication that are by definition Hodge endomorphisms. Since then, some progress has been made towards understanding such surfaces: van Geemen in \cite{Gee06} constructed transcendental lattices of rank 6 with real multiplication by a specific kind of real quadratic fields; Schlickewei in \cite{Sch09} computed the endomorphism ring of the associated Kuga-Satake varieties for general cases by a process that is not canonical; and Elsenhans and Jahnel (see \cite{E-J14}, \cite{E-J16} and \cite{E-J18}) used $L$-functions and point counting to write down specific formulae of K3 surfaces with quadratic RM. Using the language of Mumford-Tate group and generalizing Mumford's construction in \cite{Mum69}, we give a slightly different account of the resulting Kuga-Satake variety for K3 surfaces satisfying certain conditions on the rank of the transcendental lattice, which is a canonical construction. We shall also give a general formula for constructing the transcendental lattices of K3 surfaces with RM by arbitrary totally real field up to degree 6 over $\mathbb{Q}$, which will allow us to compute the discriminant form of such lattices. Moreover, we will use these new results to show that all simple Abelian fourfolds with definite quaternionic multiplication are Kuga-Satake varieties of K3 surfaces with quadratic real multiplication, Picard rank 16, and provide a dictionary between the two:
\begin{theorem}\label{main} Any simple Abelian fourfold $A$ with endomorphism by a definite quaternionic order can be realized as a Kuga-Satake variety of a K3 surface $S$ of Picard rank $16$ with real multiplication by a quadratic field: if we let the polarization of $A$ to be given by $$\left (
\begin{array}{cc}
0 & D \\
-D & 0 \\
\end{array} \right )$$ then the Mumford-Tate group of $A$ is given by $SO(D)$ for a positive definite quadratic form $D$. Such $D$'s splits into a product of two conjugating conics $Q \times \overline{Q}$ up to a real quadratic field extension $K/\mathbb{Q}$, with exactly one of which having real points, and the transcendental lattice of the associated K3 surface will have real multiplication by $K$ and its transcendental lattice obtained by the corestriction (to be introduced in section~\ref{sec4}) of $Q$. In particular, the discriminant of the transcendental lattice is $$\disc(T_S)= \disc(K/\mathbb{Q})^2 \cdot \Norm_{K/\mathbb{Q}}(\disc(Q)).$$
Conversely, if $S$ is a Picard rank 16 K3 surface with real multiplication by a quadratic field $K$, then the transcendental lattice $T(S)$ is given by the corestriction of some $K$-ternary quadratic field $(V, Q)$ with signature (defined in~\ref{sec4}) $(2,1), (0,3)$. Up to similarity, $Q$ corresponds to an unique quaternion algebra $B$ defined over $K$ up to isomorphism, and the resulting simple factor of the Kuga-Satake variety of $S$ has rational quaternionic multiplication by $B_{\Cor}$, which represents the same class as $\Cor_{K/\mathbb{Q}}(B)$ in $\Br(\mathbb{Q})$.
\end{theorem}
This paper is organized as follows. In section \ref{sec2}, we recall definitions and some general facts about Mumford-Tate groups, K3 surfaces and Kuga-Satake construction. In section \ref{sec3}, we provide a generalized Mumford's construction that realizes the norm-1 group of some quaternion algebras as Mumford-Tate group of some Abelian varieties with quaternionic multiplication. In section \ref{sec4}, we give a general construction of lattices of K3 type with real multiplication by arbitrary RM field. In section \ref{sec5}, we use the results in section \ref{sec3} and \ref{sec4} to show that all simple Abelian fourfolds with quaternionic multiplication can be realized as the Kuga-Satake variety of some RM K3 surface, and in fact we have fairly nice control over the arithmetic inputs and outputs of this correspondence.
\renewcommand{\abstractname}{Acknowledgments}
\begin{abstract} The author would like to thank Prof. Bert van Geemen for numerous comments, clarifications and suggestions on this paper, Prof. Brendan Hassett for pointing to the subject of K3 surfaces and their Mumford-Tate groups, Prof. Yuri Zarhin for the references on the classification of simple Abelian fourfolds, Prof. John Voight on arithmetic properties of Clifford algebras, and Prof. Frank Calegari for pointing to an error in the draft. The author is supported by the NSF grants DMS-1551514 and DMS-1701659.
\end{abstract}
\section{Background}\label{sec2}
In this section we introduce the definition of Mumford-Tate group and Kuga-Satake construction associated to the transcendental lattice of a K3 surface.
Recall that a $\mathbb{Q}$-Hodge structure of weight $m$ is a finite dimensional $\mathbb{Q}$-vector space $V$ together with a decomposition $V_{\mathbb{C}} = \bigoplus_{p+q=m} V_{\mathbb{C}}^{p,q}$ such that $\overline{V_{\mathbb{C}}^{p,q}} = V_{\mathbb{C}}^{q,p}$. We can define the action of the Deligne torus $\mathbb{S} := \Res_{\mathbb{C}/\mathbb{R}}\mathbb{G}_{m, \mathbb{C}}$ on $V_{\mathbb{C}}^{p,q}$ to be multiplication by $z^{-p}\overline{{z}^{-q}}$. This is equivalent to giving a representation $h: \mathbb{S} \rightarrow \GL(V_{\mathbb{R}})$. (See 1.1, 3.2, 3.3 of \cite{Moo04}) In particular, one can restrict the map $h$ to the unit circle $\mathbb{U} = \{ z | z\overline{z}=1 \}$.
\begin{definition} (4.1 of \cite{Moo04}) Let $V$ be a $\mathbb{Q}$-Hodge structure given by the homomorphism $h: \mathbb{S} \rightarrow \GL(V)_{\mathbb{R}}$. We define the (general) Mumford-Tate group of $V$, notation $\MT(V)$, to be the smallest algebraic $\mathbb{Q}$-subgroup $M$ in $\GL(V)$ such that $h$ factors through $M_{\mathbb{R}} \in \GL(V)_{\mathbb{R}}$.
Similarly, we define the special Mumford-Tate group to be the smallest algebraic $\mathbb{Q}$-subgroup such that the restriction of $h$ onto $\mathbb{U}$ factors through over $\mathbb{R}$.
(\cite{Zar83} 0.3.1.1) An element $g$ in $\End(V)$ is called a Hodge endomorphism if it commutes with the action of of the Mumford-Tate group.
\end{definition}
\begin{remark} In some literature (e.g. \cite{Zar83}), the special Mumford-Tate group is called the Hodge group.
\end{remark}
One key property of the Mumford-Tate group is that it translates the problem of locating sub-Hodge structures in tensorial constructions of a $\mathbb{Q}$-Hodge structure into the problem of locating irreducible $\mathbb{Q}$-subrepresentations:
\begin{proposition} (4.4 of \cite{Moo04}) Let $W \subset V^{\otimes a} \otimes (V^{\vee} )^ {\otimes b}$ be a $\mathbb{Q}$-subspace. Then $W$ is a sub-Hodge structure if and only if $W$ is stable under the action of $\MT(V)$ on $V^{\otimes a} \otimes (V^{\vee} )^ {\otimes b}$.
In particular, $W$ corresponds to the $\mathbb{Q}$-span of a Hodge class if and only if $W$ is 1-dimensional trivial representation of $\MT(V)$.
\end{proposition}
\begin{example} Consider the case when $m=1$ and the $\mathbb{Q}$-Hodge structure $V$ describes an Abelian variety $A$ up to isogeny. The special Mumford-Tate group for $m=1$ always lies inside $\Sp(V, \psi)$, since it has to preserve a symplectic form that is the polarization $\psi$ of the Abelian variety. Then one can prove that the endomorphism algebra of $A$ as a vector space lies inside the space of its Hodge endomorphisms. Indeed, an endomorphism (upon tensoring with $\mathbb{Q}$) on $A$ is given by a correspondence, i.e. a $(1,1)$-class in $H^2(A \times A)$ which then can be realized as a $1$-dimensional invariant subspace in $V^{\otimes 2}$. Since the action of special Mumford-Tate group on $V^{\otimes 2}$ is given by conjugation, $g \in V^{\otimes 2} \cong \End_{\mathbb{Q}}(V)$ commutes with $\gamma \in \MT(V)$ if and only if $\gamma g \gamma^{-1} = g$, i.e. $g$ spans a trivial representation of $\MT(V)$.
\end{example}
In \cite{Zar83} Zarhin has studied extensively the Mumford-Tate group and its actions associated to a K3 surface. Since the Picard group of K3 surfaces are spanned by algebraic cycles (also known to be Hodge cycles), the Mumford-Tate group acts trivially on the Picard group. We recall the definition of K3 surfaces and some facts about its transcendental lattice:
\begin{definition}(\cite{Huy16} Chapter 1, Definition 1.1) A K3 surface over a field $k$ is a complete nonsingular variety $S$ of dimension 2 such that $$\Omega^2_{S/k} \cong \mathcal{O}_S \text{ and } H^1(S, \mathcal{O}_S) = 0.$$
(\cite{Huy16} Chapter 1, Proposition 3.5) The integral cohomology $H^2(S, \mathbb{Z})$ of a complex K3 surface $S$ endowed with the intersection form is abstractly isomorphic to the lattice $E_8(-1)^{\oplus 2} \oplus U^{\oplus 3}$. The Picard lattice embeds primitively into this lattice, its orthogonal complement is called the transcendental lattice of $S$, denoted $T(S)$. $T(S)$ is known to have signature $(2, 20-\rho)$ where $\rho$ is the rank of the Picard group.
\end{definition}
We recall some key properties of $T(S)$ and its Mumford-Tate group proved in \cite{Zar83}:
\begin{proposition} Let $S$ be a complex K3 surface, $T(S)$ its transcendental lattice. Let $G$ be the Mumford-Tate group of $T(S)_{\mathbb{Q}}$ as a weight 2 $\mathbb{Q}$-Hodge structure. Then the following is true:
\begin{enumerate}
\item (1.4.1 of \cite{Zar83}) $T(S)_{\mathbb{Q}}$ is an irreducible representation of $G$.
\item (1.6 of \cite{Zar83}) The ring of Hodge endomorphisms of $T(S)$ tensored with $\mathbb{Q}$ is either a totally real field or a CM field. We say the K3 surface $S$ has RM or CM respectively.
\item (2.1 and 2.2.1 of \cite{Zar83}) In the case when $S$ has RM by a totally real field $K$, the intersection form on $T(S)_{\mathbb{Q}}$ is the trace of a $K$-form $Q$, and the special Mumford-Tate group is the special orthogonal group $\SO_K(Q)$ viewed as a $\mathbb{Q}$ group.
\end{enumerate}
\end{proposition}
We will explain the $K$-form $Q$ in greater detail in section~\ref{sec4} of this paper. The last ingredient we need is the Kuga-Satake construction of a K3 surface $S$, which returns for each K3 surface $S$ an Abelian variety $\KS(T(S))$. A good reference is given in chapter 4 of \cite{Huy16}. We will not go into the detail of this definition, but we will recall two properties of this construction:
\begin{lemma}~\label{ksrep} (\cite{Huy16} Chapter 4, 2.6) There is an inclusion of Hodge structures: $T(S)_{\mathbb{Q}} \subset H^2(\KS(T(S))^2, \mathbb{Q})$. Moreover, the Mumford-Tate groups of $T(S)_{\mathbb{Q}}$ and $H^1(\KS(T(S)))$ share the same Lie algebra.
(\cite{Gee06} Lemma 5.5) In the case when $S$ has real multiplication by a field of degree $d$ and the $K$-rank of the transcendental space $n$, over $\mathbb{C}$ the Mumford-Tate group is given by $SO(n)^d$, and its representation on $H^1(\KS(T(S)), \mathbb{C})$ sums of exterior products of the spin representation of $SO(n)^d$.
\end{lemma}
\section{A generalization of Mumford's construction}\label{sec3}
In this section we generalize the construction of Abelian varieties given by Mumford in \cite{Mum69}.
Throughout the paper let $K$ be a totally real number field, let $d \geq 2$ be its degree. Let $B$ be a quaternion algebra defined over $K$ such that\begin{equation}\label{ram} B \otimes _{\mathbb{Q}} \mathbb{R} \cong \mathbb{H}^{d-1} \oplus M_2(\mathbb{R}) \end{equation} where $\mathbb{H}$ stands for the Hamiltonian. It is a general fact that there exists field extensions $K'$ of $K$ such that $B \hookrightarrow B \otimes_{K} K' \cong M_2(K')$, hence we can define the reduced norm of an element in $B$ to be its determinant up to this embedding, and reduced trace its trace. The collection of reduced norm 1 elements form a group, usuall called the norm-1 group of the quaternion $B$.
Recall that in \cite{Mum69} Mumford wrote down the explicit expression of $B$ under Serre's corestriction functor: Let $B^{(i)} = B \otimes_K K^{(i)}$ be the quaternion algebra over $\overline{\mathbb{Q}}$. Tensoring them together we get a central simple algebra $D = B^{(1)} \otimes B^{(2)} \otimes ... \otimes B^{(d)}$. The Galois group $Gal(\overline{\mathbb{Q}}/\mathbb{Q})$ has a natural semilinear action on $D$ given by the following: if $\tau: \overline{\mathbb{Q}} \rightarrow \overline{\mathbb{Q}}$ is an element in $Gal(\overline{\mathbb{Q}}/\mathbb{Q})$, then $\tau \circ \sigma_i = \sigma_{\pi(i)}$ for some permutation $\pi$. Similarly $\tau$ acts on $\overline{\mathbb{Q}}$, which then gives a semilinear isomorphism $B^{(i)} \xrightarrow{\sim} B^{\pi(i)}$, thus a semilinear automorphism on $D$. We define the elements fixed by such Galois actions in $D$ to be $\Cor_{K/\mathbb{Q}}(B)$. It is now a central simple algebra defined over $\mathbb{Q}$. Notice that at each local place $p$ of $\mathbb{Q}$, the class of $\Cor_{K/\mathbb{Q}}(B)$ in $\Br(\mathbb{Q}_p)$ is equal to the sum of all classes $B_{\pi} \in \Br(K_{\pi})$, where $\pi | p$. This implies that $\Cor_{K/\mathbb{Q}}(B)$ is a 2-torsion class in $\Br(\mathbb{Q})$, and is Brauer equivalent to a quaternion algebra $B_{\Cor}$ defined over $\mathbb{Q}$.
\begin{remark} It is not hard to see that the condition we imposed on $B$ would force $B_{\Cor}$ to be indefinite when $d$ is odd, and definite when $d$ is even.
\end{remark}
We generalize Mumford's construction as follows: consider the central simple algebra $B_{\Cor} \otimes_{\mathbb{Q}} \Cor_{K/\mathbb{Q}}(B)$. It is now of dimension $4^{d+1}$; moreover, $B_{\Cor} \otimes_{\mathbb{Q}} \Cor_{K/\mathbb{Q}}(B) \cong M_{2^{d+1}}(\mathbb{Q})$ exactly because its Brauer classes at all local places add up to 0. Now consider the group of norm-1 elements $G$ in $B$. We define the augmented Nm map to be the following:
$$\begin{array}{cccc}
\widetilde{\Nm}: & B & \rightarrow & B_{\Cor} \otimes \Cor_{K/\mathbb{Q}}(B) \cong M_{2^{d+1}}(\mathbb{Q}) \\
& \gamma & \mapsto & \Id \otimes \gamma^{(1)} \otimes \gamma^{(2)} \otimes ... \otimes \gamma^{(d)} \\
\end{array}$$
Therefore $\widetilde{\Nm}$ is a faithful $\mathbb{Q}$-representation of $G$ into $\GL_{2^{d+1}}(\mathbb{Q})$. Hence it acts on $\mathbb{Q}^{2^{d+1}}$. To see that it can be in fact realized as a special Mumford-Tate group of an Abelian variety of dimension $2^d$, we show that the image of $\widetilde{\Nm} \otimes \mathbb{C}$ lies inside $Sp_{2^{d+1}}(\mathbb{C})$ and give an embedding of Deligne's torus.
\begin{lemma} Over $\mathbb{R}$, $G_{\mathbb{R}} \cong \SU_2(\mathbb{R})^{d-1} \times \SL_2(\mathbb{R})$, and it preserves a symplectic form.
\end{lemma}
\begin{proof} By the definition of $\widetilde{\Nm}$, over $\mathbb{C}$, $\widetilde{\Nm}$ is isomorphic to the direct sum of two copies of $W_{1,...,1}$, i.e. the exterior product of $d$-many standard representations of $SL_2(\mathbb{C})$: $\mathbb{C}^2 \boxtimes \mathbb{C}^2 \boxtimes ... \boxtimes \mathbb{C}^2$. One can compute the decomposition of the exterior product of this representation:$$\wedge^2 (W_{1,...,1} \oplus W_{1, ..., 1}) = 3\wedge^2 W_{1,...,1} \oplus \Sym^2 W_{1,...,1} $$
And $$\Sym^2 W_{1,...,1} = W_{2,2,...,2} \oplus \bigoplus_{\text{even number of 0's in the weight}} W_{2,...,0,...,2} $$
$$\wedge^2 W_{1,...,1} = W_{2,2,...,0} \oplus ... \oplus W_{0,2,...,2} \oplus \bigoplus_{\text{odd number of 0's in the weight}} W_{2,...,0,...,2} $$
When $d$ is even (resp. odd), $\Sym^2 W_{1,...,1}$ (resp. $\wedge^2 W_{1,...,1}$) will have a copy of $W_{0,0,...,0}$ i.e. the trivial representation. Hence $G_{\mathbb{R}}$ preserves a symplectic form.
\end{proof}
Analogous to Mumford's original construction, we can embed the torus $\Res_{\mathbb{C}/\mathbb{R}}(\mathbb{C}^*)$ via the map: $$h: e^{i\theta} \rightarrow \Id_{2^d} \otimes \left ( \begin{array}{cc}
cos\theta & sin\theta \\
-sin\theta & cos\theta \end{array} \right)$$
Notice that $G$ is a $\mathbb{Q}$-simple group, and the above construction gives abelian varieties $A$ up to isogeny with their Mumford-Tate group contained in $G$. This implies that the Mumford-Tate group of $A$ is exactly $G$. Moreover, we can explicitly describe the endomorphism algebra of $A$:
\begin{theorem}Let $(K, B)$ be given. Then the endomorphism algebra $\End(A) \otimes \mathbb{Q}$ of a generic member $A$ of the resulting Abelian varieties is exactly $B_{\Cor}$, i.e. $A$ has quaternionic multiplications.
\end{theorem}
\begin{proof} We first describe how $B_{\Cor}$ acts as Hodge endomorphism. Recall that Hodge endomorphisms are by definition elements in $\End(H^1(A,\mathbb{Q}))$ that are commutative with the action of Mumford-Tate group, and that upon tensoring with $\mathbb{Q}$ the algebra of Hodge endomorphisms and endomorphisms of Abelian varieties coincide.
Note that for an element $g \in B_{\Cor}$, there is a faithful embedding into $M_{2^{d+1}}(\mathbb{Q}) \cong \End(H^1(A,\mathbb{Q}))$: $g \mapsto g \otimes \Id \otimes \Id \otimes ... \otimes \Id$ where $\Id$ is the identity in each $B^{(i)}$. It is obvious that $g$ commutes with the action of $\widetilde{\Nm}(\gamma)$ for any $\gamma \in G$.
To see that $B_{\Cor}$ indeed gives everything, we notice that when $d$ is odd (resp. even), $H^2(A, \mathbb{Q})$ has $3$ (resp. $1$) copies of Hodge cycles, which are the counts of Abelian varieties having either endomorphisms by an indefinite quaternion over $\mathbb{Q}$ (resp. definite quaternion over $\mathbb{Q}$) or RM by a cubic field (resp. no extra endomorphism or CM by an imaginary quadratic field). Since the simple factors of $A$ always have dimension $2^{d-1}$ or $2^{d}$ (resp. the Lie algebra of the special Mumford-Tate group is a Lie algebra over a totally real number field), the latter case is not possible by dimension count (resp. by general facts about special Mumford-Tate groups established in \cite{Z-M95} section 2.6).
\end{proof}
\begin{remark} It is possible for $A$ to split. For example, when $K$ is a cubic field and $\Cor_{K/\mathbb{Q}}(B)$ is already $M_8(\mathbb{Q})$, then the endomorphism algebra of the resulting $A$ is $M_2(\mathbb{Q})$, i.e. it is the product of two isogenous Abelian fourfolds. This is exactly the fourfolds Mumford constructed in his original paper (cf. \cite{Mum69}), and the reason we are calling the construction introduced in this section \say{generalized Mumford construction}.
\end{remark}
\section{Constructing Hodge structure of K3 type with real multiplication}\label{sec4}
We recall the description of the transcendental lattice of a K3 surface with real multiplication structure:
\begin{proposition}
(2.1 in \cite{Zar83}) Let $K$ be a degree $d$ totally real field, $V$ an $n$-dimensional $K$-vector space equipped with a non-degenerate symmetric bilinear form $Q$. A Hodge structure of K3 type with real multiplication by $K$ is a vector space $V_0$ defined over $\mathbb{Q}$ with the following extra structure:
\begin{enumerate}
\item There is a ring embedding $\phi: K \hookrightarrow End(V_0, \mathbb{Q})$ and an isomorphism of $\mathbb{Q}$-vector spaces $f: V \rightarrow V_0$ such that for any $k \in K$ and $v \in V$, we have
$$f(kv) = \phi(k)(f(v))$$
\item There is a quadratic form $Q_0$ on $V_0$ such that $$Q_0(f(v), f(w)) = \Tr_{K/\mathbb{Q}}(Q(v,w))$$
In particular, $Q_0(\phi(k)f(v), f(w)) = \Tr_{K/\mathbb{Q}}(k \cdot Q(v,w))$
\end{enumerate}
Conversely, given a Hodge structure of K3 type with real multiplication by $K$, one can find a unique $(V,Q)$.
\end{proposition}
In \cite{me1} the author introduced a construction called \say{corestriction} of quadratic spaces, which gives rise to a functor:
$$\Cor_{K/\mathbb{Q}}: \{K \text{-quadratic spaces}\} \rightarrow \{\mathbb{Q} \text{-quadratic spaces}\} $$
such that any Hodge structure of K3 type with RM is the corestriction of some $K$-quadratic spaces satisfying certain signature condition and discriminant condition.
We recall the construction of the corestriction of quadratic spaces functor: Let $K,d,V,n,Q$ be as stated above. Let $\sigma_1, ..., \sigma_d$ be the $d$ distinct embeddings of $K$ into $\mathbb{R}$. First we consider the space $$V \otimes_{\mathbb{Q}} \mathbb{R} \cong (K^{\sigma_1})^n \oplus (K^{\sigma_2})^n \oplus ... (K^{\sigma_d})^n$$ where $K^{\sigma_i}$ denotes $\mathbb{R}$ with $K$ embedded into it via $\sigma_i$. Under this isomorphism, a vector $kv = v \otimes k$ in $V$ will be mapped to $ (\sigma_1(k)v^{(1)}, ..., \sigma_d(k)v^{(d)})$.
This space is endowed with a semilinear automorphism by the Galois group $G = Gal(K/\mathbb{Q})$: if $\sigma_j = \sigma_i \circ g$ for a $g \in G$, then $g$ takes $\sigma_i(k)\otimes v^{(i)}$ to $\sigma_j(k) \otimes v^{(j)}$.
\begin{definition}
The corestriction of $V$ from $K$ to $\mathbb{Q}$ is defined to be the $\mathbb{Q}$-subspace in $V \otimes_{\mathbb{Q}} \mathbb{R}$ invariant under the semilinear automorphisms by $G$. i.e.
$$\Cor_{K/\mathbb{Q}} (V) = (V \otimes_{\mathbb{Q}} \mathbb{R})^G $$
Similarly, use $Q^{\sigma_i}$ to denote the $\mathbb{R}$-quadratic form $Q^{\sigma_i}( r_1 \otimes \sigma_i(x), r_2 \otimes \sigma_i(y)) := r_1 \cdot r_2 \sigma_i (Q(x,y))$. The corestriciton of $Q$, denoted $Q_0$ is the quadratic form on $(K^{\sigma_1})^n \oplus (K^{\sigma_2})^n \oplus ... (K^{\sigma_d})^n$ given by $Q^{\sigma_1} \oplus Q^{\sigma_2} \oplus ... Q^{\sigma_d} $. The $\mathbb{Q}$ quadratic space we obtain in this way is denoted $\Cor_{K/\mathbb{Q}}(V, Q)$, called corestriction of the quadratic space $(V,Q)$. Likewise, if $\Lambda$ is a projective $\mathcal{O}_K$ lattice in $(V,Q)$, we can define the corestriction of $\Lambda$ and denote it as $\Cor_{K/\mathbb{Q}}(\Lambda, Q)$.
\end{definition}
We now prove that this space has a compatible $K$-multiplication structure that satisfies the conditions imposed by Zarhin's proposition:
\begin{lemma}
\begin{enumerate}
\item $K$ admits a canonical embedding $\phi$ into $\End(\Cor_{K/\mathbb{Q}}(V), \mathbb{Q})$. Moreover, $\phi(k) \in \End(\Cor_{K/\mathbb{Q}}(V), \mathbb{Z})$ for any $k \in \mathcal{O}_K$.
\item $Q_0$ takes coefficients in $\mathbb{Q}$ on $\Cor_{K/\mathbb{Q}}(V)$. Moreover, let $\phi$ be the canonical imbedding of $K$ into $\End(\Cor_{K/\mathbb{Q}}(V), \mathbb{Q})$, $f$ the canonical isomorphims of $\mathbb{Q}$-vector spaces $f: V \rightarrow \Cor_{K/\mathbb{Q}}(V)$, then we also have $Q_0(\phi(k)f(v), f(w)) = \Tr_{K/\mathbb{Q}}(k \cdot Q(v,w))$.
\end{enumerate}
The discriminant group of $Q_0$ has order $$ | \disc(Q_0) |= \disc(K/\mathbb{Q})^n \cdot \Norm_{K/\mathbb{Q}}(\disc(Q)).$$
\end{lemma}
\begin{proof} Without loss of generality we assume $V$ is $1$-dimensional (i.e. $V = K\cdot e$). The correstriction map in this case recovers the classical setting of Minkowski embeddings:
$$ \begin{array}{cccc}
f: & K & \hookrightarrow & \Pi (\mathbb{R}^{\sigma_i}) \\
& ke & \mapsto & (k^{\sigma_1}e^{(1)}, ..., k^{\sigma_d}e^{(d)})\\
\end{array}
$$
If we assume that $K$ has a primitive element $a$, we can describe $\phi$ simply by describing $a$. In particular, we can let $\phi$ to be the map such that $$\phi(a)f(k \cdot e) := f(ak \cdot e)$$ It is not hard to see that with respect to the $\mathbb{Q}$-basis $\{f(e), f(a \cdot e), ..., f(a^{d-1} \cdot e) \}$ of $\Cor_{K/\mathbb{Q}}(V)$, $\phi$ maps $a$ to its rational canonical form. This completes the proof of the first part.
A high-brow way of proving the rationality of $Q_0$ is by noticing that $Q_0$ is actually invariant under the semilinear automorphisms by the Galois group.
For the purpose of computation, we will write out $Q_0$ for a quadratic form $Q(e,e) = c$ on the 1-dimensional $K$-vector space $V = Ke = \mathbb{Q}(a)e$. If we adopt the basis in the previous lemma, a straighforward computaion will give
$$Q_0 = \left ( \begin{array}{cccc}
\Tr_{K/\mathbb{Q}}(c) & \Tr_{K/\mathbb{Q}}(ac) & ... & \Tr_{K/\mathbb{Q}}(a^{d-1} c) \\
\Tr_{K/\mathbb{Q}}(ac) & \Tr_{K/\mathbb{Q}}(a^2 c) & ... & \Tr_{K/\mathbb{Q}}(a^d c) \\
& ... & ... & \\
\Tr_{K/\mathbb{Q}}(a^{d-1} c) & \Tr_{K/\mathbb{Q}}(a^d c) & ... & \Tr_{K/\mathbb{Q}}(a^{2d-2} c)\\
\end{array} \right )
$$
To prove the relation between $Q_0$ and $Q$, we observe that since we defined $\phi$ such that it satisfies $\phi(k)f(v) = f(k\cdot v)$, and that by the way we set up $Q_0$, we always have $$Q_0(\phi(k)f(v), f(w)) = (k \cdot Q(v,w))^{\sigma_1} + (k \cdot Q(v,w))^{\sigma_2} + ... +(k \cdot Q(v,w))^{\sigma_d}$$
The right hand side gives exactly $\Tr_{K/\mathbb{Q}}(k \cdot Q(v,w))$.
\end{proof}
\begin{remark}
It is not difficult to see that if $\Lambda$ is an $\mathcal{O}_K$-integral lattice in $V$ (namely for any two $v, w \in \Lambda$, $Q(v,w) \in \mathcal{O}_K$), then $f(\Lambda)$ is a $\mathbb{Z}$ lattice in $\Cor_{K/\mathbb{Q}}(V)$. Moreover, with the above description, we can compute the \emph{discriminant form} of the lattice.
\end{remark}
At this point, all that remains is to make sure an $\mathcal{O}_K$-lattice is taken to an even lattice that admits a primitive embedding into the $K3$ lattice with signature $(2, n\cdot d -2)$. This poses some conditions on the quadratic form $Q$.
We say the signature of $Q$ is $(p_1, q_1) ,..., (p_d, q_d)$ if the signature of each embedding $Q^{\sigma_i}$ is given by $(p_i, q_i)$. The weight condition on the vectors in the transcendental lattice imposes the following condition on $Q$:
\begin{lemma}\label{sig} The signature of $Q$ must be of the form $(2, n-2), (0,n), ..., (0,n)$.
\end{lemma}
\begin{proof} We consider where the unique (up to scalar multiplication) $(2,0)$ vector lives. By our construction, over $\mathbb{C}$ it lives inside some specific embedding of $\sigma_{i, \mathbb{C}}: V \hookrightarrow \mathbb{R}^n \otimes \mathbb{C} \subset \Cor_{K/\mathbb{Q}}(V, Q)\otimes \mathbb{C}$. WLOG suppose that $i = 1$. Then we claim that $(0,2)$ vector must live inside the same subspace. This is essentially because the $\mathbb{R}$-forms of $SO(n)$ and $SO_n(\mathbb{C})$ shares the same representation, hence if we consider the complex conjugate of the $(2,0)$ vector, it would be a $(0,2)$ vector living inside $\sigma_{i,\mathbb{C}}: V \hookrightarrow \mathbb{R}^n \otimes \mathbb{C} $. Since up to scalar multiplication there is only one $(0,2)$ vector, we know that the signature of $Q$ is exactly stated as the lemma.
\end{proof}
\begin{theorem}
Any Hodge structure of $K3$ type with real multiplication by a totally real field $K$ of degree $d$ can be obtained through the correstriction of some $K$-vector space $V$ of dimension $n$ equipped with a quadratic form $Q$ such that the following conditions are satisfied:
\begin{enumerate}
\item $Q$ has signature $(2, n-2), (0,n), ..., (0,n) $ and maps to an even quadratic form under $\Cor_{K/\mathbb{Q}}$.
\item The Hodge struture can be made integral if there exists an $\mathcal{O}_K$ lattice $\Lambda \subset V$ such that $\Cor_{K/\mathbb{Q}}(\Lambda)$ admits a primitive embedding into the $K3$ lattice.
\end{enumerate}
\end{theorem}
\section{The case of $n=3$}\label{sec5}
In this section we prove the following correspondence:
$$\left (
\begin{array}{c}
\text{Abelian fourfold with QM} \\
\text{by $B_{\Cor}$, and Mumford-Tate group}\\
\text{preserving a rank 4 quadratic $\mathbb{Q}$-form $Q_0$}
\end{array} \right ) \leftrightarrow \left (
\begin{array}{c}
\text{K3 surfaces $S$ with real multiplication} \\
\text{by the quadratic field $K$ such that}\\
\text{$Q_{0,K} \cong C \times \overline{C}$, $T(S) = \Cor_{K/\mathbb{Q}}(Q)$}
\end{array} \right ) $$
where $\overline{C}$ denotes the Galois conjugation of $C$ over $\mathbb{Q}$ and $Q$ denotes (up to scaling) the ternary quadratic form with the right signture that is determined by $C$.
\begin{theorem}Let S be a K3 surface with real multiplication by arbitrary RM field $K$ with degree no greater than $6$, and let its intersection form on the transcendental space $\Cor_{K/\mathbb{Q}}(Q)$ be given by a ternary $K$-quadratic form $Q$. Then $Q$ is associated to a quaternion algebra $B$ defined over $K$ satisfying the ramifying condition at infinity (see \ref{ram}). Up to isogeny, the Kuga-Satake variety of $S$ is the products of Abelian varieties $A$ given by the generalized Mumford construction we introduced in section~\ref{sec3}. In particular, the endomorphism algebra of a simple factor is either trivial (when $B_{\Cor} = M_2(\mathbb{Q})$) or $B_{\Cor}$.
\end{theorem}
\begin{proof}
The proof of the theorem is mainly a generalization of the results in \cite{Gal00} and section 3 of \cite{me1}. Let $K, Q, B$ be as stated. Recall the classical fact that the norm-1 group $G \subset B$ can be realized as an (at most) double cover of $\SO_K(Q)$ (cf. Proposition 4.5.10 in \cite{Voi17}). Hence $G$ and the special Mumford-Tate group of $T(S)$ shares the same even weight representation. Furthermore, we can locate a copy of $T(S)_{\mathbb{C}}$ in $B_{\Cor}\otimes \Cor_{K/\mathbb{Q}}(B) \otimes \mathbb{C}$ in the similar fashion as \cite{Gal00}:
\begin{equation}\label{split}
T(S)_{\mathbb{C}} = W_{2,0,0,..,0} \oplus W_{0,2,0,...,0} \oplus ... \oplus W_{0,0,...,0,2} \subset H^2(A, \mathbb{C}).
\end{equation}
And a $\mathbb{Q}$ basis of $T(S)$ can be given by $\Id \otimes \Cor_{K/\mathbb{Q}}(B^0) \subset B_{\Cor}\otimes \Cor_{K/\mathbb{Q}}(B)$, where $B^0$ denotes the space of trace 0 quaternion elements. The lemma (lemma \ref{sig}) we have proved on the signature of $Q$ determines the ramifying data of $B$ at infinity, which is exactly the condition stated in ~\ref{ram}.
Moreover, this construction can be made integral. Lattices can be obtained by looking at the intersection of quaternion orders and trace zero space. For specific computations, see section 3 of \cite{me1}.
In \cite{Gee06} van Geemen worked out the representation of the Kuga-Satake variety of RM K3 surfaces over $\mathbb{R}$, which are direct sums of the representations given by $\widetilde{\Nm}$ (see lemma~\ref{ksrep}). Since $T(S)_{\mathbb{C}}$ has Hodge numbers $(1,3d-2,1)$ if and only if one of the subspaces in (\ref{split}) is of Hodge number $(1,1,1)$ which is simultaneously the subspace in $\Cor_{K/\mathbb{Q}}(B^0)$ where $Q^{\sigma_i}$ splits, this necessarily forces the Deligne's torus to be embedded in $G$ exactly the way we specified in section~\ref{sec3}. This implies that the Kuga-Satake variety of $T(S)$ is indeed given by the generalized Mumford construction.
\end{proof}
\begin{remark} This statement is also proven by Schlikewei in \cite{Sch09} by finding copies of $\Cor_{K/\mathbb{Q}}(B)$ in $C^0(T(S))$. Still, one should note that the isomorphism of Hodge substructures in \cite{Sch09} is not canonical. The generalized Mumford's construction provides a canonical way to roundabout the choices made in \cite{Sch09}. In fact, for a general quadratic form $Q$ with signature $(2, n-2), (0,n), ... (0,n)$, we can replace $\Cor_{K/\mathbb{Q}}(B)$ with $\Cor_{K/\mathbb{Q}}(C^0(Q))$, and $B_{\Cor}$ with a minimal dimensional Brauer group representative of $\Cor_{K/\mathbb{Q}}(C^0(Q))$, which by Merkurjev's theorem would become a quaternion over $\mathbb{Q}$. Analogous computations to section \ref{sec3} would prove that the resulting abelian variety has the desired endomorphism, which is indeed Brauer-equivalent to Schlikewei's result.
\end{remark}
We now focus on the case when $d=2$ and $n=3$, and prove the following theorem:
\begin{theorem} Any simple Abelian fourfold $A$ with endomorphism by a definite quaternionic order can be realized as a Kuga-Satake variety of a K3 surface of Picard rank $16$ with real multiplication by a quadratic field, and the relation between the polarization of $A$ and the transcendental lattice of the K3 surface is exactly given by theorem \ref{main}.
\end{theorem}
\begin{proof}To prove the theorem, we show that 1) there is a Hodge structure of K3 type in the second cohomology of $A$, and 2) the special Mumford-Tate group of $A$ is a (double cover of) $\SO_K(Q)$ for some ternary quadratic form $Q$ with signature $(2,1),(0,3)$.
It has been proven (see section 6.1 of \cite{Z-M95}) that for a simple Abelian fourfold with endomorphism by a definite quaternion order, the Lie algebra of the special Mumford-Tate group over $\mathbb{C}$ is isomorphic to $\mathfrak{so}_4$. Moreover, there's a canonical isomorphism between the Lie algebra of $\SL_2 \times \SL_2$ and $\SO(4)$ (cf. pp 274-275 of \cite{F-H91}), and it is established that the standard representation of $\SO(4)$ corresponds to the $W_{1,1}$ of $\SL_2 \times \SL_2$. In this proof we shall adopt the $\SL_2$ notation. We have the following computation: $\Sym^2(H^1(A)) = 3\Sym^2 W_{1,1} \oplus \wedge^2 W_{1,1}$, and $\wedge^2(H^1(A)) = \Sym^2 W_{1,1} \oplus 3 \wedge^2 W_{1,1}$. By comparing the Hodge numbers of $\Sym^2(H^1(A))$ and $H^2(A)$ and counting the copies of sub-representations, it is not hard to work out the Hodge number of $\Sym^2 W_{1,1}$ and $\wedge^2 W_{1,1} = W_{2,0}\oplus W_{0,2}$, which are $(3,3,3)$ and $(1,4,1)$ respectively. Hence there is a Hodge structure of K3 type in $H^2(A \times A)$.
The identification of $\SL_2 \times \SL_2$ with $\SO(4)$ given by Fulton and Harris is geometric: it is obtained by the canonical isomorphism between a quadric surface in $\mathbb{P}^3$ and the two rulings by conics. Hence in our case, the special Mumford-Tate group is a norm-1 subgroup of a quaternion $B$ satisfying the splitting condition if and only if the associated quadric hypersurface does not decompose into the product of conics over $\mathbb{Q}$ but over a quadratic field extension, over which the two conics are conjugate to each other. This quadratic field extension, according to our configuration, is exactly the RM field of the K3 surface, and the corestriction of the conics will give the transcendental lattice of the K3 surface.
It remains to show that the quadric surface cannot split over $\mathbb{Q}$ as the product of two conics $C_1$ and $C_2$. Suppose otherwise. Then we can read off the two factors of the $\mathbb{Q}$-form of $\SL_2$ by associating each $C_i$ with a class in $\Br(\mathbb{Q})$ hence a quaternion algebra's norm-1 subgroup $G_i$. However, in this case the $\wedge^2 W_{1,1}$ decomposes since $W_{2,0}$ and $W_{0,2}$ are now $\mathbb{Q}$-space. The resulting Abelian variety would then be products of Abelian varieties of lower dimension which are no longer simple.
\end{proof}
|
\partial{\partial}
\def{\boldsymbol\psi}{{\boldsymbol\psi}}
\def{\bar{\boldsymbol\psi}}{{\bar{\boldsymbol\psi}}}
\def{\boldsymbol\lambda}{{\boldsymbol\lambda}}
\def{\boldsymbol\sigma}{{\boldsymbol\sigma}}
\def{\bar \nu}{{\bar \nu}}
\def{\bar \psi}{{\bar \psi}}
\def{\bar X}{{\bar X}}
\def{\bar p}{{\bar p}}
\def{\bar q}{{\bar q}}
\def{\bar Q}{{\bar Q}}
\def{\bar q}{{\bar q}}
\def{\bar \ell}{{\bar \ell}}
\def\nu_e{\nu_e}
\def\barnue{{\bar \nu_e}}
\def\nu_\mu{\nu_\mu}
\def{\bar \nu}_\mu{{\bar \nu}_\mu}
\def\nu_\tau{\nu_\tau}
\def{\bar \nu}_\tau{{\bar \nu}_\tau}
\def\nu_L{\nu_L}
\def\nu_R{\nu_R}
\defe_R{e_R}
\defe_L{e_L}
\def\psi_L{\psi_L}
\def\psi_R{\psi_R}
\def{\rm km}{{\rm km}}
\def{\rm kg}{{\rm kg}}
\def{\rm pc}{{\rm pc}}
\def{\rm kpc}{{\rm kpc}}
\def{\rm Mpc}{{\rm Mpc}}
\def{\rm Gpc}{{\rm Gpc}}
\def{\rm sr}{{\rm sr}}
\def{\rm hr}{{\rm hr}}
\def{\rm yr}{{\rm yr}}
\def{\rm Gyr}{{\rm Gyr}}
\defT_{GUT}{T_{GUT}}
\defT_{EW}{T_{EW}}
\defMod. Phys. Lett.\,{m_{Pl}}
\def\m-pl{m_{Pl}}
\defT_{Pl}{T_{Pl}}
\defh_{75}{h_{75}}
\def\Omega h^2_{75}{\Omega h^2_{75}}
\def\Omega h^2_{70}{\Omega h^2_{70}}
\def{\tilde a}{{\tilde a}}
\defa_{eq}{a_{eq}}
\defa_{dec}{a_{dec}}
\def{\bf CP}{{\bf CP}}
\def{\bf C}{{\bf C}}
\def{\bf P}{{\bf P}}
\def{\bf CPT}{{\bf CPT}}
\def{\bf B}{{\bf B}}
\def{\bf L}{{\bf L}}
\def{\rm T}{{\rm T}}
\def\boldmath{\boldmath}
\def\boldsymbol{\boldsymbol}
\def\bs{p}{\boldmath{p}}
\def\bsy{P}{\boldsymbol{P}}
\def{\cal L}{{\cal L}}
\def{\cal S}{{\cal S}}
\def{\cal P}{{\cal P}}
\def{\cal D}{{\cal D}}
\def{\cal V}{{\cal V}}
\def{\cal T}{{\cal T}}
\def{\tilde{\cal D}}{{\tilde{\cal D}}}
\def{\bf A}{{\bf A}}
\def{\bf U}{{\bf U}}
\def{\bf D}{{\bf D}}
\def{\tilde{\bf D}}{{\tilde{\bf D}}}
\def{\bf A}{{\bf A}}
\def{\tilde{\bf A}}{{\tilde{\bf A}}}
\def{\bf F}{{\bf F}}
\def{\tilde{\bf F}}{{\tilde{\bf F}}}
\def{\bf H}{{\bf H}}
\def{\rm G}{{\rm G}}
\def{\mu{\rm G}}{{\mu{\rm G}}}
\def{\rm nG}{{\rm nG}}
\def\hbox{$^\circ$}{\hbox{$^\circ$}}
\def\hbox{$^\prime$}{\hbox{$^\prime$}}
\def\hbox{$^{\prime\prime}$}{\hbox{$^{\prime\prime}$}}
\def\hbox{$.\!\!^{\rm d}$}{\hbox{$.\!\!^{\rm d}$}}
\def\hbox{$.\!\!^{\rm h}$}{\hbox{$.\!\!^{\rm h}$}}
\def\hbox{$.\!\!^{\rm m}$}{\hbox{$.\!\!^{\rm m}$}}
\def\hbox{$.\!\!^{\rm s}$}{\hbox{$.\!\!^{\rm s}$}}
\def\hbox{$.\!\!^\circ$}{\hbox{$.\!\!^\circ$}}
\def\hbox{$.\mkern-4mu^\prime$}{\hbox{$.\mkern-4mu^\prime$}}
\def\hbox{$.\!\!^{\prime\prime}$}{\hbox{$.\!\!^{\prime\prime}$}}
\def\hbox{$.\!\!^{\scriptscriptstyle\rm p}$}{\hbox{$.\!\!^{\scriptscriptstyle\rm p}$}}
\def\fun#1#2{\lower3.6pt\vbox{\baselineskip0pt\lineskip.9pt
\ialign{$\mathsurround=0pt#1\hfil##\hfil$\crcr#2\crcr\sim\crcr}}}
\def\VEV#1{\left\langle #1\right\rangle}
\def\VEV{\sigma_{\rm ann}|v|}{\VEV{\sigma_{\rm ann}|v|}}
\def\mbox{\boldmath$\theta$}{\mbox{\boldmath$\theta$}}
\def\mbox{\boldmath $l$}{\mbox{\boldmath $l$}}
\defJour. Cosmology and Astro-Particle Phys.\,{Jour. Cosmology and Astro-Particle Phys.\,}
\defM.N.R.A.S.\,{M.N.R.A.S.\,}
\defin~prep.\,{in~prep.\,}
\defAstrophys.J.\,{Astrophys.J.\,}
\defAstrophys.J.(Lett.){Astrophys.J.Lett.\,}
\defAstrophys.J.Supp.\,{Astrophys.J.Supp.\,}
\defAstron.J.\,{Astron.J.\,}
\defNature\,{Nature\,}
\defNew Ast.\,{New Ast.\,}
\defNew~Ast.Rev.\,{New~Ast.Rev.\,}
\defNucl. Phys.\,{Nucl. Phys.\,}
\defComm. Math. Phys.\,{Comm. Math. Phys.\,}
\defPhys. Rev. Lett.\,{Phys. Rev. Lett.\,}
\defPhys. Lett.\,{Phys. Lett.\,}
\defRev. Mod. Phys.\,{Rev. Mod. Phys.\,}
\defInt. Jour. Mod. Phys.\,{Int. Jour. Mod. Phys.\,}
\defMod. Phys. Lett.\,{Mod. Phys. Lett.\,}
\defPhys. Rev.\,{Phys. Rev.\,}
\defPhys.Rev.~D{Phys.Rev.D\,}
\defAnnu.Rev.Astron.Astrophys.\,{Annu.Rev.Astron.Astrophys.\,}
\defAstron.Astrophys.\,{Astron.Astrophys.\,}
\defAstron.Astrophys.Supp.\,{Astron.Astrophys.Supp.\,}
\defAstron.Astrophys.\,{Astron.Astrophys.\,}
\defPub.Astr.Soc.Japan\,{Pub.Astr.Soc.Japan\,}
\defPhys.Rep.\,{Phys.Rep.\,}
\defAstrophys. Space Sci.{Astrophys. Space Sci.}
\def\hline{\hline}
\def\tableline{\tableline}
\section*{\normalsize Key Points}
\label{sec:keypoints}
\begin{enumerate}
\item Multi-messenger astrophysics aspires to make use of the information
provided about the astrophysical universe by all four fundamental forces of
Nature, namely { the gravitational, the weak and the strong forces,
besides the electromagnetic force which previously had provided almost all our
information about the Cosmos. These new channels provide previously untapped,
qualitatively different and complementary types of information, capable of probing
down to the densest and energy-richest regions of cosmic objects, which were
hitherto hidden from astronomers' sights}.
\item Diffuse backgrounds of high-energy neutrinos (HENs) with
energies from $\mathrel{\mathpalette\simov >}$10\,TeV to PeV, ultra-high energy cosmic
rays (UHECRs) { at energies of $\mathrel{\mathpalette\simov >} 10^{18}\hbox{~eV}$, and }
$\gamma$-rays with energies between MeV and $\sim$TeV have been
measured, or upper limits have been provided, by Cherenkov
detectors, satellites and ground-based air-shower arrays.
\item Gravitational waves (GWs) from merging stellar mass black hole
and neutron star binaries have been detected at frequencies { in the
$\mathrel{\mathpalette\simov >}$ 10~Hz to $\sim$ 1~kHz range} with laser interferometric
gravitational-wave detectors.
\item The sources of the diffuse UHECR and HEN backgrounds remain
unknown, although a gamma-flaring blazar (a type of active galaxy
with a massive black hole at the center ejecting a relativistic
plasma jet towards the observer) has been tentatively identified
with observed HENs. While up to $\sim$85\% of the $\gamma$-ray
background can be attributed to blazars, it appears that at most
30\% of the HEN background can be due to blazars.
\item Formation channels for the observed stellar mass black hole
binaries, and their possible role as a cosmologically relevant
component of the dark matter, is currently under debate.
\item There is a natural physical connection between high energy
cosmic ray interactions and the resulting very high energy neutrinos
and \mbox{gamma-rays}, which needs to be fully exploited to better understand
the nature of their unknown astrophysical sources. The connection
with gravitational wave emission, while less direct, can be expected
to provide important information about supermassive black hole
populations and dynamics.
\item Even before the arrival of the next generation of gravitational
wave, neutrino, and cosmic ray detectors, the present advanced
LIGO/VIRGO detectors will be able to detect hundreds of binary
mergers up to $\sim$Gpc distances; yet electromagnetic (EM) counterpart
searches rely primarily on the aging space-based facilities \textit{Swift}\ and
\textit{Fermi}, currently operating well beyond their design lifetimes. These EM
counterparts have been found mainly in gamma- or \mbox{X-rays}, and there
is an urgent need for a new generation of EM detectors, also
extending into other frequencies including the UV, optical/IR, and
radio.
\end{enumerate}
\section{Introduction}
\label{sec:intro}
Of the four fundamental forces in nature -- the electromagnetic,
gravitational, weak and strong nuclear forces -- until the middle
of the 20th century it was only messengers of the electromagnetic
force, in the form of optical photons, which allowed astronomers to
study the distant universe. Subsequently, advancing technology added
to these radio, infrared, ultra-violet, \mbox{X-ray}\ and \mbox{gamma-ray}\ photons.
Finally, in the last few decades the messengers of the other three forces,
namely gravitational waves (GWs), neutrinos, and cosmic rays (CRs),
began to be used in earnest. { Thus, we are now finally using the
complete set (as far as known) of forces of Nature, which are revealing
exciting and hitherto unknown details about the Cosmos and its denizens.}
These new non-photonic messengers are generally more challenging to
detect and to trace back to their cosmic sources than most
electromagnetic emissions. When detected, they are usually
associated with extremely high mass or high energy density
configurations, e.g.\ the dense core of normal stars, stellar
explosions occurring at the end of the nuclear burning life of massive
stars, the surface neighborhood of extremely compact stellar remnants
such as white dwarfs, neutron stars or black holes, the strong
and fast varying gravitational field near black holes of either
stellar mass or the much more massive ones in the core of galaxies, or
in energetic shocks in high velocity plasmas associated
with such compact astrophysical sources. This association with the
most violent astrophysical phenomena known means that the interpretation
of multi-messenger observations requires, and can have implications for,
our theories of fundamental physics, including strong-field gravity,
nuclear physics, and particle interactions.
\begin{figure}[ht]
\centerline{\includegraphics[width=6.0in,height=4.0in]{NAT-Figures-MMA-editor-1.jpg}}
\vspace*{-0.5in}
\caption{\footnotesize Examples of current instruments observing
cosmic messengers via the electromagnetic, weak, gravitational, and
strong forces, showing their location. Clockwise from top left:
the LIGO Hanford gravitational wave interferometer;
one of the MAGIC air Cherenkov telescopes;
the Fermi gamma-ray space telescope;
a schemaric of the Pierre Auger cosmic ray observatory in Malarg\"ue, Argentina;
a schematic of the IceCube cubic kilometer neutrino detector in Antarctica;
}
\label{fig:Fermi-IC3-LIGO-Auger}
\end{figure}
The study of such high energy compact objects started in earnest in
the 1950's, after decades of a slower build-up with increasingly large
ground-based optical telescopes. The first major breakthroughs came
from the deployment of large radio-telescopes, followed by the
launching of satellites equipped with \mbox{X-ray}\ and later
\mbox{gamma-ray}\ detectors, which established the existence of active galactic
nuclei, neutron stars, and black holes, and revealed dramatic
high-energy transient phenomena including \mbox{X-ray}\ novae, \mbox{X-ray}\ bursts,
and gamma-ray bursts (GRBs).
Starting in the late 1960's, large underground neutrino detectors
were built, measuring first the neutrinos produced in the Sun
and later those arising from a supernova explosion; and it was only in
the current decade that extragalactic neutrinos in the TeV-PeV range
were discovered. Cosmic rays in the GeV energy range started being
measured in the 1910s, but it was only in the 1960s that large
detectors started measuring higher energies implying an extragalactic
origin, and only in the last decade has it become practical to start
investigating the spectrum and composition in the $10^{18}-10^{20}$ eV
range. Gravitational wave detectors started being built in the 1970s,
but it was not until the 1990s that new technologies and large enough
arrays began to be built approaching the sensitivity required for
detections, the first successes starting in 2015.
For the first time, detectors covering all four fundamental forces of Nature
{\bf (Fig, \ref{fig:Fermi-IC3-LIGO-Auger})} have been thrown into the breach to explore all
the previously hidden aspects of the Cosmos.
\section{\Large Mono- and Multi-Messenger Advances}
\label{sec:currentmm}
The exciting experimental results listed in this section confirm many of the
theoretical/phenomenological predictions and expectations that had been formulated
over the past several decades, while also bringing up new surprises, which we discuss
in the subsequent section.
\subsection{Recent Non-Photonic Mono-Messenger Results}
\label{sec:singlemes}
\noindent$\bullet$\,\,
{\it Ultra-High Energy Cosmic Rays.-}
The Pierre Auger cosmic ray observatory (PAO) \cite{Auger+15revinstr}, located in Argentina, is a
3,000 km$^2$ array of 1660 water Cherenkov stations and 27 air-fluorescence telescopes (one of the
tanks and a set of fluorescence telescopes is in {\bf Fig. \ref{fig:Fermi-IC3-LIGO-Auger}, lower right)},
designed to detect UHECRs at energies between $10^{17}\hbox{~eV}$ to $10^{21}\hbox{~eV}$.
Its measurements of the diffuse UHECR flux energy spectrum, starting in 2009, confirmed conclusively
the existence of a spectral steepening setting in near $6\times 10^{19}\hbox{~eV}$, e.g.
\cite{Auger+17icrc17highlight}, which had been first observed by the HiRes instrument
\cite{Abbasi07hires}.{ This is consistent with the so-called GZK (Greisen, Zatsepin, Kuz'min,
\cite{Greisen66,Zatsepin+66}) feature expected from CR proton energy losses or from heavy ion
photo-dissociation \cite{Gerasimova+62nucphotodis} due to interactions with cosmic microwave background
photons, although it could also be due to reaching a maximum acceleration at the sources.}
From 2010 onwards, Auger also started showing evidence for an UHECR chemical composition becoming
heavier above $\mathrel{\mathpalette\simov >} 10^{18.5}\hbox{~eV}$. The statistical significance of these results has become stronger
over the years \cite{Auger+16mixedcomp,Gora+18augerrev,Petrera+19augerrev}. The spectral results are
consistent, within statistical uncertainties with those obtained with the smaller Telescope Array (TA)
UHECR observatory \cite{Kawai+08-TA}, which is important because Auger is in the Southern hemisphere while
TA is in the Northern. { The chemical composition issue \cite{TA+18masscomp} is still under debate,
although a joint Auger-TA paper \cite{AbuZayyad+18augerTAchem} shows results which agree within the errors.}
The angular resolution in the arrival direction of UHECRs is below $1^o$ above $\sim 10^{19}\hbox{~eV}$ for
both protons and heavy elements, although the magnetic deflection increases with mass;
around $10^{19}\hbox{~eV}$ it is $\mathrel{\mathpalette\simov <} 5^o$ for protons, while for heavy nuclei it could be tens of degrees.
At these energies, due to
the energy losses caused by the GZK effect mentioned, these UHECRs must have originated within
distances of $\mathrel{\mathpalette\simov <} 100{\rm Mpc}$. So far, all attempts at finding angular spatial correlations between
UHECRs and any type of known cosmic sources have been unsuccessful \cite{Auger+17icrc17highlight}.
\\
\noindent$\bullet$\,\,
{\it High Energy Neutrinos.-}
The IceCube neutrino observatory \cite{Achterberg+06-ICECUBE} consists of a cubic kilometer (roughly a
Gigaton) of ice at a depth between 1.4 and 2.4 km below the South Pole, instrumented with 86 strings
connecting 5,160 optical phototubes (see schematic in {\bf Fig. \ref{fig:Fermi-IC3-LIGO-Auger},
lower right)}, which measure the light radiated from { charged particles} produced by passing high energy
neutrinos interacting with the ice. Its construction was finished in 2010, and in 2012-2013 it discovered
a diffuse flux of neutrinos in the range $100 \hbox{~TeV} \mathrel{\mathpalette\simov <} E_\nu \mathrel{\mathpalette\simov <}` 1 \hbox{PeV}$ \cite{IC3+13pevnu1,
IC3+13pevnu2}, later extended down to $\mathrel{\mathpalette\simov <} 100\hbox{~TeV}$. The energy spectrum $dN/dE_\nu$ can be fitted
with a $\sim -2.5$ index power law, but there may be an indication for two components, steeper below $\sim
200\hbox{~TeV}$ and flatter (index $\sim -2$ above that, the highest energy so far being $\sim 10\hbox{PeV}$.
IceCube detects all neutrino flavors, with muon neutrino charged current interactions resulting in
elongated Cherenkov tracks and all other neutrino flavors and interactions largely producing
{near-spherical optical Cherenkov signals from secondary particle cascades,}
the direction of arrival being uncertain by $\sim 10^o -15^o$ for cascades
and $\sim 0.5^o-1.0^o$ for tracks.
Tau neutrinos can also be identified at sufficiently higher energies, where the statistics are lower,
and these have not yet been identified,
{ although a suggested tau-like candidate has been discussed \cite{Kistler+18multipev}.}
The observed flavor distribution is compatible with complete flavor mixing having occurred due to
the neutrino oscillation phenomenon over cosmological distances \cite{Halzen17nurevnat,Ahlers+18ic3nurev}.
So far there is no evident correlation of the observed neutrinos with any type of known cosmic
objects \cite{IC3+17icrc17nusources}, except for one interesting case discussed below.
{The smaller underwater Cherenkov telescopes ANTARES \cite{ANTARES+19results} and Baikal-GVD
\cite{Baikal+18status} have also been in operation and providing upper limits.}
At much higher energies, the high altitude balloon experiment ANITA \cite{Allison+18anita3res},
flying in a circumpolar orbits in Antarctica, has used a radio
technique to measure neutrinos at $\mathrel{\mathpalette\simov >} 10^{17}$ eV, which is starting to provide constraints on
cosmological neutrino sources and the GZK-related cosmogenic neutrino fluxes, complementary with
those provided by Auger \cite{Auger+15cosmonu}.
\\
\noindent$\bullet$\,\,
{\it Gravitational Waves.- }
The Laser Interferometric Gravitational Wave Observatory (LIGO) consist of two detectors, in Louisiana
and Washington state, each with 4 km long L-shaped arms,
{\bf (Fig. \ref{fig:Fermi-IC3-LIGO-Auger}, upper left)} which in 2015 began operation in the
$\sim 10 - 10^3$ Hz frequency range \cite{Abbott+09-LIGO}.
Another array, VIRGO \cite{Acernese+15-VIRGO}, located near Pisa, Italy, and similarly L-shaped with
3 Km long arms, has been operating at epochs coincident with LIGO. Both are actively being commissioned
and will achieve design sensitivity in the coming years.
The long-awaited first discovery of gravitational waves from a
stellar mass binary black hole merger (labeled GW150914) was announced by LIGO in 2016
\cite{LIGO+16-gw150914disc} {\bf (Fig. \ref{fig:BBHGW-crnugam-BNS-blazarnu}(a)}.
This was soon followed by a number of other binary black hole (BBH)
mergers detected both by LIGO and, with lower statistical significance, by VIRGO as well
\cite{LIGO+18bbhGWTC-1}. These BHs weigh up to several tens of solar masses, and have low spins.
\noindent
However, despite intensive searches, no other messengers associated with BBH mergers have been detected
so far, except for a possible $\gamma$-ray burst \cite{Connaughton+16-gbmgw150914} in GW150914.
\\
\noindent$\bullet$\,\,
{\it Electromagnetic Detections.- }
Except for the binary black holes, all the other sources detected with other messengers had been
previously extensively studied through their EM emissions at various wavelengths.
Of major recent relevance are the observations in the optical, X-ray and up to 150 keV
$\gamma$-rays with the Swift satellite, and between 10 keV X-rays to $\mathrel{\mathpalette\simov <} \hbox{~TeV}$ $\gamma$-rays
with the Fermi satellite {\bf (Fig. \ref{fig:Fermi-IC3-LIGO-Auger}, upper right)}
\cite{Thompson15spacegam}, which detected a large number of
Gamma-Ray Burst (GRB) sources, active galactic nuclei (AGNs) including blazars, supernovae, etc.,
as well as a diffuse cosmic $\gamma$-background.
Of increasing importance for such sources are the ground-based air Cherenkov imaging telescopes,
e.g. MAGIC {\bf (Fig. \ref{fig:Fermi-IC3-LIGO-Auger}, upper middle)}, HESS, VERITAS
\cite{Montaruli19ctalhaasogam} and the High Altitude Water Cherenkov observatory HAWC
\cite{HAWC-instr18,Casanova+18hawcrev}, which measure gamma-rays in the 100 GeV to multi-TeV range.
These have been amply supported by ground and space observations with multiple radio, infra-red, optical
and UV telescopes.
\subsection{Developments in Joint Multi-Messenger Astrophysics}
\label{sec:recentmultimes}
\noindent$\bullet$\,\,
{\it Solar and Supernova Neutrinos and Photons.- }
The two earliest multi-messenger detections involved neutrinos in the MeV range. Davis and collaborators,
starting in the 1960's, detected the electron neutrinos produced by the nuclear reactions that are
the energy source for the Sun's light, using a 600 ton perchlorethylene (cleaning fluid) tank
located deep underground in the Homestake mine in South Dakota, US. This neutrino flux was
confirmed by various other experiments including the one in the Kamioka mine in Japan, by Koshiba and
collaborators.
The other early multi-messenger detection was that of neutrinos from a core-collapse supernova,
SN 1987a, resulting from inverse beta decay as protons are converted into neutrons. This was
detected by three different underground detectors, Kamiokande in Japan, Baksan in the Soviet Union,
and Irvine-Michigan-Brookhaven in the US
\cite{Hirata+87nusn1987a-Kamioka,Alexeyev+88nusn1987a-Baksan,Haines+88nusn1987a-IMB}.
The neutrino detection preceded significantly the spectacular optical brightening characterizing
supernovae. These discoveries earned Davis and Koshiba the Physics Nobel Prize in 2002
\cite{Davis03Nobel,Koshiba03Nobel}.
\\
\begin{figure}[ht]
\begin{minipage}[t]{0.43\textwidth}
\centerline{\includegraphics[width=3.0in,height=1.5in]{GW150914wavevelsep.jpg}}
\end{minipage}f
\hspace{-40mm}
\begin{minipage}[t]{0.43\textwidth}
\centerline{\includegraphics[width=3.0in,height=1.5in]{PhiCR_neu_mix.jpg}}
\end{minipage}
\hspace{3mm}
\begin{minipage}[t]{0.43\textwidth}
\centerline{\includegraphics[width=3.0in,height=1.5in]{gw170817ligoferint.jpg}}
\end{minipage}
\hspace{5mm}
\vspace*{-1.1in}
\begin{minipage}[t]{0.43\textwidth}
\centerline{\includegraphics[width=3.0in,height=1.5in]{IC318-TXSnuwhitebkg.jpg}}
\end{minipage}
\vspace*{1.0in}
\caption{\footnotesize
Some recent cosmic multi-messengers advances involving the electromagnetic,
weak, gravitational and strong forces.
(a) The first GW detection from LIGO/VIRGO in the O1-O2 observing run, of the GW150914 binary black hole
merger \cite{LIGO+16-gw150914disc}, showing for the inspiral, merger and ringdown phases the theoretical
and measured waveform, the separation and the relative velocity.
(b) Interrelation expected between (from left to right) the energy spectrum of the diffuse backgrounds
in gamma-rays, high energy neutrinos and ultra-high energy cosmic rays, based on a black hole jet
source model \cite{Fang+18crnuagnjet};
(c) LIGO, VIRGO and Fermi simultaneous multi-messenger discovery of the binary neutron
star merger GW/GRB 170817;
(d) Light track of the muon produced by a 290 TeV muon neutrino coming from the direction of the
blazar TXS 0506+056, detected on 22 September 2017 by IceCube.
}
\label{fig:BBHGW-crnugam-BNS-blazarnu}
\end{figure}
\noindent$\bullet$\,\,
{\it Cosmic ray, Gamma-ray and Neutrino Background Interdependences.- }
The measurements of the diffuse UHECR energy spectrum by the Pierre Auger Observatory starting in 2008
put on a firm ground the detection of a spectral cutoff above $10^{19.5}$ eV, compatible with the
GZK energy losses due to the cosmic microwave background photons \cite{Auger+17icrc17highlight},
after earlier work by HiRes.
Then, starting in 2008, the Fermi satellite (following on previous work by COS-B and other missions)
measured a diffuse gamma-ray background extending into the sub-TeV range \cite{Ackermann+15igbfermispec}.
And starting in 2012-2013 IceCube discovered, with increasing detail, a diffuse high energy neutrino (HEN)
background of astrophysical origin at multi-TeV to PeV energies \cite{IC3+13pevnu1,IC3+13pevnu2}.
There is so far no firm identification of the sources of either the UHECR, HEN or gamma-ray diffuse
backgrounds, although the extragalactic gamma-ray background is known to be dominated by blazars
\cite{Ackermann+16igbfermi}. However, theoretical relationships and mutual constraints are expected
from the basic physics of these three radiations, e.g. {\bf Fig. \ref{fig:BBHGW-crnugam-BNS-blazarnu} (b)}.
\noindent
The HENs are produced when UHECRs collide with low energy
target photons and nuclei resulting in charged and neutral pions, which decay in a predictable
fraction of high energy neutrinos and gamma rays. The resulting energy spectra of neutrinos
and photons imply corresponding diffuse backgrounds which must fit the observed results,
including also the constraint provided by the observed UHECR background.
The fact that the energetics of these three messengers is comparable has led to the idea of a
unification of high-energy cosmic particles, e.g., \cite{Murase+16uhenurev,Fang+18crnuagnjet}.
On the other hand, significant constraints are also placed on generic $pp$ hadronuclear production
models of HENs and gamma-rays when when one compares them to the Fermi diffuse $\gamma$-ray flux,
especially accounting for the $\gamma\gamma$ cascades { initiated by $\gamma$-rays scattering
off cosmic radiation backgrounds} \cite{Murase+13pev,Murase+16hidden}.
The constraints are more stringent for Galactic sources \cite{Ahlers+14pevnugal}.
HAWC \cite{HAWC-instr18} is expected to uniquely contribute to measurements of the $\gamma$-ray background
in the 10 to 100 TeV energy range, which could strongly constrain the fraction of IceCube neutrinos from
Galactic origin.
Among $p\gamma$ photomeson production models of HENs, valuable constraints have been put on the
contribution of the simpler classical GRB neutrino emission models \cite{Abbasi+12-IC3grbnu-nat,
IC3+15promptnugrb,IC3+18grbnuconstraint}, while leaving open the possibility of contribution by
choked GRBs or supernovae driven by choked jets \cite{Meszaros+01choked,Murase+13choked,
IC3+18grbnuconstraint}.
\\
\noindent$\bullet$\,\,
{\it Gravitational Waves and Photons from Binary Neutron Star (BNS) Mergers.- }
{ As the culmination of a long series of previous BNS GW/multi-messenger searches, e.g.
\cite{LIGO+05grb030329,LIGO-VIRGO+08gwcounter,Kanner+08gwopt,LIGO+08grb070201}, the}
joint GW/EM detection of the transient GW/GRB 170817 was the first high significance proof of
the strength of the joint multi-messenger technique in the GW realm \cite{LIGO+17gw170817disc},
e.g. see {\bf Fig. \ref{fig:BBHGW-crnugam-BNS-blazarnu} (c)}.
\noindent
The GWs in GW/GRB 170817 showed that this was a neutron star binary merger, providing a measurement
of their masses, the distance \cite{LIGO+18-170817properties} and gave constraints on the neutron star
equation of state \cite{LIGO+18-170817nsradeos}, while $\gamma$-ray and X-ray measurements by Fermi
and Swift showed it was an off-axis short GRB, e.g. \cite{Troja+17gw170817xr}.
The near simultaneous observation of EM and GW signals from GW170817 showed that they both travel at the
speed of light to better than 1 part in $10^{15}$, thereby ruling out many alternative theories of gravity.
Optical observations with various telescopes showed that it also manifested itself as a Kilonova,
which is an outflow rich in the so-called r-process high atomic number nuclear elements,
providing a { rich interlocking picture, e.g. \cite{Coulter+17gw170817,Kasliwal+17gw170817concord,
Abbott+17-170817multi,Margutti+17-170817xr}. }
{ For their role in the discovery of binary gravitational wave sources Barish, Thorne and Weiss
received the 2017 Nobel Prize in Physics \cite{Weiss19Nobel,Barish19Nobel,Thorne19Nobel}.}
\\
\noindent$\bullet$\,\,
{\it High Energy Neutrinos and Gamma-rays from Blazars.-}
The joint neutrino \cite{IC3multi+18txs0506} and electromagnetic detection
\cite{Magic+17blazarnu,Keivani+18txs0506,NuSTAR+17blazarnu} of the flaring blazar TXS 0506+056 was an
extremely exciting result, being the first time that a known source was shown to be associated
{ (albeit at the $\sim 3\sigma$ level)} with a high significance astrophysical high energy neutrino
{\bf (Fig. \ref{fig:BBHGW-crnugam-BNS-blazarnu}(d)}.
\noindent
Blazars are active galactic nuclei (AGNs), which are galaxies with a massive central black hole
powering a relativistic jet outflow pointing close to the observer line of sight; they are classified
into BL Lac objects and flat spectrum radio-quasars (FSRQs), TXS 0506+056 appearing to be of the BL Lac
type \footnote{Recently, however, arguments have been presented \cite{Padovani+19txs0506fsrq}
indicating that TXS 0506+056 may be an FSRQ instead of a BL Lac object as thought previously.}.
Blazars are notorious for exhibiting sporadic and intense gamma-ray flaring episodes, one of which
was in progress at the time the track-type neutrino was observed. Further analysis indicated that
in previous years other neutrinos may have been associated with this source \cite{IC3+18txs0506a}.
This provided valuable constraints on the radiation mechanisms and the sources of the diffuse
HEN background. Based on simple one-zone emission models where both HENs and gamma-rays originate
from the same region, the neutrino is a low probability event \cite{Gao+19txs0506,Keivani+18txs0506,
Cerruti+19txs0506} and based on a stacking analysis of HENs and blazars it appears that the blazar
population as a whole may account for $\mathrel{\mathpalette\simov <} 10-30\%$ of the entire IceCube neutrino background
{ \cite{IC3+17fermiblaznustack}},
so other sources may in any case need to be appealed to.
\section{\large Emerging Questions and Challenges}
\label{sec:newquations}
\noindent$\bullet$\,\, {\it The Lack of EM or HEN counterparts of binary black hole mergers} is
frustrating, with ten binary black hole mergers detected in GWs so far
(as of March 2019). Such emissions are expected to be faint at best
in BBHs, e.g. \cite{Perna+16gw150914grb,Murase+16bbh}, but they would
be very useful for a better understanding of the binary origin and environment,
{ as well as to get a far better localization than provided by the GWs,
e.g. \cite{Murase+16bbhmulti,Bartos+17bbh,Ford+19bbhagn,LIGO+19multisearch}.}
A much larger sample of BBHs will be needed extending to both
smaller and larger masses to test the hypothesis that BBHs
provide a cosmologically important dark matter component,
e.g. \cite{Bird+16gw150914,Magee+17bhdm,Carr19bhdm}. \\
\noindent$\bullet$\,\, {\it Detection of HEN from GW/EM-detected binary neutron stars}
would provide an example of a ``triple-messenger" source, and would
clarify major open questions in our understanding of these objects.
{ Possible signals and observing strategies have been discussed in,
e.g., \cite{Bartos+11multigwnu,Ando+13multinu,LIGO+19multisearch}.}
Expected HEN fluxes are low, especially for off-axis jet
viewing \cite{Kimura+17sgrbnu,Kimura+18transejnu}, but in the best
case they may be marginally detectable by IceCube (or more plausibly, by a
future IceCube-Gen 2), and would greatly aid in clarifying the physics
of the relativistic jet and the larger-angle slower outflows which give
rise to the GRB, the afterglow, and the kilonova emission of these events. \\
\noindent$\bullet$\,\, {\it Confirmation or refutation of the occurrence of HEN flares in blazars}
{ through additional observations of TXS 0506+056 and other AGNs}
is urgently needed to address the origin of
the IceCube background and illuminate any possible connection between
the HEN and UHECR backgrounds. Progress in these studies will also
require more targeted calculation of AGN neutrino production models,
yielding detailed predictions for \mbox{X-ray}\ and other EM constraints. The
stacking analyses of blazar EM flares against observed HENs
\cite{IC3+17fermiblaznustack,Hooper+18nuagn} as well as theoretical
arguments \cite{Murase+18txs0506} indicate that sources other than
blazars must provide the dominant contribution to the HEN background,
and observational correlation studies involving alternative source
candidates may need to be undertaken, e.g. \cite{Murase+16uhenurev}. \\
\noindent$\bullet$\,\, {\it The masses and spins of the GW-detected compact mergers} offers
new puzzles. One is the origin of the "heavy" binary stellar black holes
($\mathrel{\mathpalette\simov >}$ 30 solar masses), it is not clear how they form and evolve. Another
question is why do the LIGO BHs have very low spins or spins mis-aligned with
the orbital angular momentum. This is in contrast to X-ray BH candidates,
some of which have very large spins. Also the fate of the remnant in the BNS
merger GW 170817 is unknown, e.g. how long did the remnant last before turning
into a black hole, if it finally did. { These and related questions are
discussed in, e.g., \cite{LIGO+18bbhGWTC-1,LIGO+18bbho1o2}.}
Future GW observations could resolve this issue. \\
\noindent$\bullet$\,\, {\it UHECR arrival direction uncertainties are large}, and UHECR
arrival times are delayed by $\sim$10$^4$ to 10$^5$ years relative to
any simultaneous EM or neutrino emission, so direct correlation
attempts have been made only against quasi-steady, non-bursting
sources, so far unsuccessfully \cite{Auger+17icrc17highlight,
IC3+16-nucrcorrel,Moharana+15=IC3correlcr}. At the highest
energies, UHECR positional correlations with muon neutrino tracks, UHE
neutrinos, and/or \mbox{gamma-rays}\ could lead to a better pinpointing of the sources.
This will require { much better instruments and} more sensitive neutrino/EM
correlation analyses as well as much more detailed production models for likely
source candidates. \\
\noindent$\bullet$\,\, {\it Statistically significant measures { of UHECR/EM/HEN/GW }
correlations (or lack thereof)} are urgently needed, and it is also
necessary to explain the UHECR spectrum and chemical composition with
an appropriate distribution of specific sources,
e.g.\ \cite{Aloisio+13comp,Alves+19cosmonu}. Observations must be
fitted in statistical detail to model predictions of possible
candidates, such as AGNs, GRBs, tidal disruption events, clusters of
galaxies, etc., and more sophisticated models must be calculated, and
tested against the observed diffuse neutrino and \mbox{gamma-ray}\ backgrounds. \\
\noindent$\bullet$\,\, {\it Theory and simulations are still in their infancy}, as far
as UHECR, HEN and GW sources.
{ While the HEN/EM inter-relation is in principle straightforward,
aside from non-linearities introduced by EM and hadronic cascades, an
understanding of the HEN/EM inter-relations with UHECRs is significantly
complicated by the fact that the latter are charged, and hence travel in
complicated paths, dependent on the magnetic fields, e.g. \cite{Murase+19crnurevfuku};
in addition, at energies above $\mathrel{\mathpalette\simov >} 10^{17}-10^{18}\hbox{~eV}$, cross-section
uncertainties start to set in.}
For BBH and BNS mergers, a lot of progress
is urgently required to understand post-merger dynamics, the final state
of the remnant, the physics of the ejecta and how BH-NS mergers differ
{ from BNS mergers \cite{Radice+18bnsmerg}.
Supernova simulations have also been a challenge \cite{Glas+19sn3dsimul,Radice+19sncollgw}. }
Lack of reliable GW waveforms means that we have to rely on sub-optimal techniques
for their detection, and it also makes it far more difficult to distinguish
between different collapse models/scenarios.\\
\noindent$\bullet$\,\, {\it The ultra-high energy neutrino range $10^{17}-10^{20}\hbox{~eV}$
explorations by ANITA} \cite{Allison+18anita3res} and other future
experiments need to achieve at least an order of magnitude greater
sensitivity to probe the { cosmogenic neutrino background.
This is due to UHECRs interacting with the cosmic photon backgrounds,
and degeneracies are induced by the effect of the UHECR source luminosity
function and redshift distribution as well as the CR chemical composition,
e.g. \cite{Stanev+05diffnu,Kotera+11uhecr,Globus+17inbigb}.}
The ANITA anomalous upward-going events \cite{Gorham+18anitataunu}, if
confirmed, are very exciting for what they may tell us about cosmic
tau-neutrinos, or possibly about beyond the standard model
physics. Concordance studies between ANITA, IceCube and Auger will
need to be carried out, together with significantly more detailed
theoretical investigations, e.g. \cite{Alvarez+18taunu,
Romero-Wolf+18anitaup,Connolly+18anitatau,Fox+18anitabsm}.
\section{\large Looking Ahead: New Instruments \& Results Expected}
\label{sec:forward}
The spectacular results achieved, mainly in the last decade, by
multi-messenger facilities such as those in the first row of Table 1,
has opened wide new vistas in high energy astrophysics. This has spurred the
building and planning of more sophisticated and more powerful experiments and
missions, geared towards the elucidation of the key new questions raised.
The second row of Table 1 shows some of the new experiments currently
under constructions, while the third row shows some of the next generation of
experiments planned for the period between approximately 5 to 15 years from now.
\begin{figure}[ht]
\centerline{\includegraphics[width=6.0in,height=4.0in]{NAT-Figures-MMA-editor-3.jpg}}
\vspace*{-0.5in}
\caption{\footnotesize
Some major new detectors in the planning stage:
(a) SVOM China-France GRB multi-wavelength follow-up satellite, exp. 2022 \cite{Dagoneau+18svom};
(b) Schematic of the CTA Cherenkov Telescope Array gamma-ray ground array, exp. 2024 \cite{Wild+18-CTA};
(c) IceCube-Gen2, including current IceCube and DeepCore, and the planned high energy array,
super-dense PINGU sub-array and extended surface array (larger ARA radio array not shown) \cite{IC3+17icrcgen2};
(d) KM3NeT planned 3-4 km$^3$ neutrino detector planned in the Mediterranean sea,
which will include also the high-energy ARCA and low energy ORCA sub-arrays \cite{Km3net+18sens};
(e) Schematic of the planned EU next generation Einstein gravitational wave interferometer
\cite{Sathyaprakash+12-EINSTEIN}.
}
\label{fig:SVOM-CTA-IC3Gen2-KM3NeT-ET}
\end{figure}
\begin{figure}[H]
\vspace*{-0.25in}
\begin{minipage}[t]{\textwidth}
\centerline{\includegraphics[width=8in,height=6in,angle=90]{Tab-1-rev-new-crop.jpg}}
\end{minipage}
\caption{TABLE 1 }
\label{tab:1}
\end{figure}
\eject
\noindent$\bullet$\,\, {\it New Electromagnetic Detectors .-} Among the major
space-based electromagnetic facilities coming online within the next 5
to 10 years are the Chinese-French Space Variable Object Monitor
(SVOM) \cite{Dagoneau+18svom} {\bf (Fig. \ref{fig:SVOM-CTA-IC3Gen2-KM3NeT-ET}(a))},
designed for detecting gamma-ray, X-ray
and optical transients, scheduled for launch in late 2022.
{ Two other Chinese mission in preparation are { GECAM \cite{ZHANG+19GECAM},
with an all-sky coverage of the sky aimed at detecting GW counterparts in the
6 keV to 5 MeV energy range, scheduled for the mid-2020s}; and the time-domain
Explorer-class Einstein Probe (EP) \cite{Yuan+15-EinsteinProbe}, with a 3600
sq.deg. field of view sensitivity at 0.5-5 KeV, scheduled for end of 2022.}
{ An Israeli-US mission called ULTRASAT \cite{Sagiv+14-ULTRASAT}
has been proposed, with an ultraviolet (250--280 nm), fast slewing
($\sim$ minutes) imaging detector of 250\,deg$^2$ field of view,
which could detect hundreds of supernovae, $\sim$10 BNS counterparts per year,
and $\sim$100 tidal disruption events per year.}
{ There is a very strong case to be made for a US X-ray-\mbox{gamma-ray} ~satellite for
providing real-time triggers and data, which would be critical for multi-messenger studies.}
A possible NASA mission that recently completed Phase~A
study is the ISS-TAO ``Transient Astrophysics Observer'' \cite{Yacobi+18-ISS-TAO},
on the International Space Station, with a GTM
\mbox{gamma-ray}\ transient monitor and WFI wide field (350 sq.deg.) lobster-type
\mbox{X-ray}\ imager, whose prime target would be EM counterparts of GW
sources, and which might fly by 2032.
A significant role in detecting or confirming transients of multi-messenger
importance will be played by the ZTF (Zwicky Transient Facility) \cite{Patterson+19-ZTF})
and the ASAS-SN facility \cite{Kochanek+17-ASAS-SN}.
Also in the 5 to 10 year timeframe, the multi-national Cherenkov Telescope Array (CTA)
\cite{CTA+17science} {\bf (Fig. \ref{fig:SVOM-CTA-IC3Gen2-KM3NeT-ET}(b))}
and the Chinese Large High Altitude Air Shower Observatory (LHAASO)
\cite{Disciascio+16-LHAASOrev} ground-based facilities will survey the sky at
TeV-PeV gamma-ray energies, e.g.\ \cite{Montaruli19ctalhaasogam}.
{ Both of these use the air Cherenkov technique, but while CTA includes steerable dishes
that provide good angular localization, useful for point sources, LHAASO largely observes
as the sky goes by, as does HAWC, which works better for extended or diffuse emission.}
A major optical/IR survey facility instrument is the Spectroscopic Survey
Telescope (LSST) \cite{LSST+17obsstrat}, while the Square Kilometer Array (SKA)
\cite{Mcpherson+18-SKA} will provide milliarcsecond spatial resolution images at
radio frequencies. { Also in preparation are the 30-meter class TMT, ELT and GMT
optical/IR ground-based telescopes \cite{Sanders+13-TMT,Varela+14-ELT,Johns+12}.}
For the early 2030s, the European Space Agency ESA is preparing a
major flagship \mbox{X-ray}\ mission called ATHENA \cite{Nandra+11-ATHENA}, which will
trace the galaxy formation and metallicity evolution of the Universe
with its large area detectors, and can study Population III GRBs. Due for
final ESA selection in 2022 for a launch in 2032 is the smaller but
nimbler, fast-slewing ($\sim$ minute) satellite THESEUS
\cite{Amati+18theseus}, designed to discover long GRBs at redshifts
$z\mathrel{\mathpalette\simov >} 9$ and seek BNS counterparts with a soft \mbox{X-ray}\ imager, an X-
and \mbox{gamma-ray}\ spectroscopic imager and an \mbox{0.7-m} class infra-red
telescope, which will also provide triggers for ATHENA. The NASA
AMEGO satellite \cite{AMEGO+17icrc}, sensitive to \mbox{gamma-rays}\ from 0.2\,MeV to
$\mathrel{\mathpalette\simov >}$10\,GeV and the German eROSITA \cite{Predehl+16-eROSITA} 0.3--10 keV space
detector will play important roles in in X- and \mbox{gamma-ray}\ astronomy, as
well as in the EM detection of hidden neutrino sources.
{ To complement the above large facilities, it is extremely important to have
also various fast, large field-of-view robotic ground-based telescope systems
which can follow up transient candidates within seconds after an alert is triggered.}
\\
\noindent$\bullet$\,\, {\it High and Low Energy Neutrino Detector Improvements and Plans.-}
High energy neutrino detector planned upgrades include the
IceCube High Energy Array and the denser PINGU sub-array, as part of an extended (10 Gtons)
IceCube Gen2 \cite{IC3+17icrcgen2} {\bf (Fig. \ref{fig:SVOM-CTA-IC3Gen2-KM3NeT-ET}(c))}.
In the Northern hemisphere, the completion in the Mediterranean sea of the
KM3NeT \cite{Km3net+18sens} 3 to 4 Gton EU detector is expected by
$\sim$ 2026 {\bf (Fig. \ref{fig:SVOM-CTA-IC3Gen2-KM3NeT-ET}(d))}; the error box
improvements for muon tracks are expected to be $\mathrel{\mathpalette\simov <} 0.3-0.5$ deg$^2$.
{ The relative advantages/disadvantages of ice vs. water as a Cherenkov
detector medium are that the light absorption length in ice is $\sim 100$ m
vs. $\sim 15$ m for clear ocean water; while the scattering length for
ice is $\sim 20$ m vs. $\sim 100$ m for water. Also ocean water contains
radioactive $^{40}$K, affecting the energy resolution signal to noise ratio.
Thus one gets relative better/worse energy resolution, and worse/better
angular resolution in ice/water, e.g. \cite{IC3+17icrcgen2,Km3net+18sens}.}
Another Gigaton water-based neutrino detector, Baikal-GVD \cite{Baikal+18-Baikal-GVD},
in the lake Baikal, Russia, is expected by $\sim$ 2021-22.
Goals include determining large scale anisotropies and
individual source identifications by neutrinos alone or in tandem with
other multi-messengers. They will facilitate the use of doublets and
multiplets for source population studies, and increase the prospects
for identifying galactic sources, reliably identify tau neutrinos, and
determine the flavor composition of the high energy neutrino background.
The Hyper-Kamiokande (\mbox{Hyper-K}) \cite{Hyper-K+18instr} next generation
megaton water Cherenkov detector, operating at MeV to GeV energies, is
located in the Kamioka mine (Japan) and is scheduled to begin
construction in 2020. It will be an order of magnitude larger than its
predecessor instrument \mbox{Super-K}, where the addition of Gadolinium to
the water is providing significantly greater sensitivity. It will be
able to detect individual supernova explosions out to $\sim$4\,Mpc,
roughly one every 3 to 4 years, and in 10 to 20 years it could
measure the relict supernova diffuse neutrino flux in the 16--30 MeV
energy range \cite{Migenda+17Hyper-K}.
At the highest energies, $10^{17}\hbox{~eV}$ to $10^{21}\hbox{~eV}$, the ANITA
balloon experiment \cite{Allison+18anita3res} will over the next several years
undergo further sensitivity improvements. On a longer timescale of
2022-2032, there is ongoing work and plans for much larger
ground-based detectors using the Askaryan radio technique for
detecting neutrinos and UHECRs in the same $10^{17}-10^{21}\hbox{~eV}$ range,
such as ARIANNA \cite{Barwick+17arianna} and ARA
\cite{Allison+16-ARASproto}, or a possible combination of parts of
these efforts (RNO/ARA). These are aimed at detecting the cosmogenic
neutrino background component produced by GZK UHECRs, as well as for
probing more deeply the nature of the decline of the UHECR spectrum
beyond $10^{20}\hbox{~eV}$. Another large-scale detector proposal aimed at this
goal is the Chinese-led GRAND 10,000 km$^2$ array being planned for the
2025 to 2030s \cite{GRAND+18scides}, as well as the POEMMA
\cite{POEMMA+17instr} and Trinity \cite{Otte18-trinity} projects.
\\
\noindent$\bullet$\,\, {\it UHECR Detector Improvements and Extensions.-} UHECR
facilities undergoing major upgrades include the Auger-Prime addition
of 1600 surface scintillation detectors on top of the existing water
Cherenkov tanks as well as updated electronics \cite{Auger+17-ICRC}; and the
Telescope Array upgrade to the TAx4 configuration four-fold surface
enlargement \cite{Sagawa+15-TAx4}. In the next 5-10 years the planned
K-EUSO experiment \cite{Casolino+17-K-EUSO} on the International Space Station
(ISS) could achieve uniform exposure across the Southern and Northern
hemispheres of $4\times 10^4\,{\rm km}^2\,{\rm sr}\,{\rm yr}$ per year, an order
of magnitude larger than Auger or TAx4. Radio observations with LOFAR
(Low Frequency Radio Array) \cite{Winchen+19-LOFARcr} may also contribute
substantially to an understanding of UHECRs.
In the lower energy range of $10^{12}$ to $>$10$^{15}$\,eV the ISS-CREAM
\cite{Seo+14-ISSCREAM} on the International Space Station, building on
earlier work of the Cosmic Ray Energetics And Mass (CREAM) balloon flights,
{ as well as the future Chinese mission HERD \cite{Zhang+14HERD}},
could greatly increase knowledge about the spectrum and
compositions of CR nuclei with charges in the $Z=1-26$ range.
In the next 10 to 15 years the planned POEMMA \cite{POEMMA+17instr} is
expected to achieve an increase in the exposure by 100$\times$, while
the planned FAST ground array \cite{FAST+17uhecr} would provide 10x
the exposure with high quality events. These advances will address the
chemical composition and anisotropy issues of UHECR, the
{ interpretation of the surmised three CR components making up
the entire spectrum (e.g. see Fig. 4 of \cite{Gaisser+13crsp}) },
the maximum energy of galactic CRs, and will probe
in much more detail the nature of the spectral cutoff, the transition from
galactic to extragalactic components, the strength of the galactic and
intergalactic magnetic fields, etc. \\
\noindent$\bullet$\,\, {\it Gravitational Wave Detectors Improvements \& Planned New
Facilities.-} The upgrades of LIGO and VIRGO are continuing
\cite{Abbott+18gwlivingrev}, and this will reduce the $90\%$ median
credibility error box size for source identifications down to $120-180$
deg$^2$ (with 12-21\% with $\leq 20$ deg$^2$) by the 2019+ O3 run. The
Japanese KAGRA detector being built in the Kamioka mine in Japan is
expected to reach a sensitivity comparable to aLIGO/aVIRGO by 2024. A
new LIGO observatory is under construction in India to house the third
detector (LIGO-India). With these additional facilities, the expected
median 90\% localization error box sizes will be 9-12 deg$^2$. The
number of expected binary black hole and neutron star detections and
the limiting distances are shown in Table 2 {\bf (Fig. \ref{tab:LVKrange})}.
\begin{figure}[H]
\centerline{\includegraphics[width=6.0in,height=1.7in]{Abbott+18gwlivrevtab1-crop.jpg}}
\caption{\footnotesize
TABLE 2. Plausible target detector sensitivities, giving the average detection distance (range)
at which a $2\times 1.4\mbox{$M_\odot$}$ BNS and $2\times 30\mbox{$M_\odot$}$ BBH may be detected with
LIGO, VIRGO and KAGRA \cite{Abbott+18gwlivingrev}.
}
\label{tab:LVKrange}
\end{figure}
Among the next generation of ground-based GW detectors planned for the
2020-30s \cite{Abbott+17gwnextgen}, { in the US, the ``A+'' upgrade
to the LIGO facilities has been approved, which should provide a further factor
of two increase in detection range beyond the detectors' advanced sensitivity.
{ A further upgrade, termed LIGO Voyager, would have a new detector operating in
the existing LIGO, with cryogenic mirrors in the existing LIGO vacuum envelope.
This could bring a further factor of 3 increase in the BNS range (to 1100 Mpc),
with a low frequency cut-off down to 10 Hz.}
In the EU, the planned underground Einstein Telescope \cite{Sathyaprakash+12-EINSTEIN}
{\bf (Fig. \ref{fig:SVOM-CTA-IC3Gen2-KM3NeT-ET}(e))},
with three arms of 10 km length, will be able to measure the GW polarization }
of BBHs and BNS sources from distances 3 to 10 times farther than with the
current design sensitivity of the LIGO/VIRGO designs.
Further in the future, the Cosmic Explorer ground-based observatory
\cite{Abbott+17-COSMICEXPLORER-GWNextGen} would use 40~km arms to achieve
a further order of magnitude improvement in sensitivity over the 10 to 10$^4$ Hz
frequency range and detect compact binary in-spirals throughout our Hubble volume.
In order to study the merger of the much larger ($10^6$ to
$10^8\mbox{$M_\odot$}$) ``supermassive'' black holes located at the center of
galaxies, the EU is planning a large space-based GW detector called
eLISA \cite{Klein+16-eLISA}, consisting of an interferometer using three
small satellites in Solar orbit. A different technique for the
exploration of supermassive BH mergers is provided by the Pulsar Timing
Arrays (PTAs) \cite{Schutz+16ptambh}, such as NANOGrav, PPTA, and
EPTA, e.g. \cite{Hobbs+17ptagw,Arzoumanian+18nanograv11}. This
technique relies on measuring the time delays in the EM radio signals
from distant pulsars caused by the space-time variations induced by
the GW field of merging BH binaries, and is expected to yield the
first stochastic (population-integrated) detections in the near future. \\
\noindent$\bullet$\,\, {\it Exploiting the Multi-Messenger Synergy.-} The Astrophysical
Multi-messenger Observatory Network (AMON) \cite{Keivani+17AMON} is a
multi-institution consortium which has signed MoUs with a number of
observatories using different messengers. { One major purpose is to
combine disparate rare signals appearing in coincidence, so that even sub-threshold
detections in one messenger, when combined with other sub-threshold signals
in other messengers, can yield a reliable above-threshold detection.
The other purpose is to rapidly distribute interesting transient alerts
arising from any observatory to all the other observatories and the community,
to facilitate rapid follow-up.}
The architecture of AMON consists of a
central hub with radial spokes connecting to individual observatories,
from which it receives individual sub-threshold (and also above threshold)
alerts, which are then subjected to analysis and/or redistributed to
other observatories. This greatly increases the speed and
effectiveness of reaction to a trigger, compared to the large number
of traditional individual observatory-to-observatory connections.
Observatories that have signed up to the AMON network
include, so far, ANTARES, Auger, \textit{Fermi}\ LAT, \textit{Fermi}\ GBM, HAWC,
IceCube, Swift-BAT, VERITAS, and others. Live data streams from
IceCube, HAWC, \textit{Fermi}\ and \textit{Swift}\ are being received by AMON,
triggering alerts as in the \mbox{TXS~0506+056} blazar neutrino and
gamma-ray flare discovery \cite{Keivani+18txs0506,Ayala+19-AMONmulti},
and other coincident sub-threshold analyses are being carried out,
e.g.\ for LIGO + \textit{Swift}\ BAT, HAWC, and others. The unique and
most promising aspect of AMON is its massive emphasis on leveraging
multiple live sub-threshold alerts. In addition, sub-threshold
analyses are also being carried out using archival data from different
individual observatories, e.g.\ \cite{Turley+18nugamsearch}. Other
groups are also developing algorithms and strategies for
multi-messenger searches, e.g. \cite{Countryman+19llama}.
\\
\noindent$\bullet$\,\, {\it Theory and Simulations.-} The high-quality data provided by
the facilities outlined above will only yield fruit insofar as it
is thoroughly analyzed and interpreted through realistic source models
satisfying state-of-the-art physics (while keeping in mind the ultimate
possibility of beyond the standard model physics). Such models must
be considered at three levels. The basic level is based on an overall
conceptual picture, using analytical or semi-analytical source models,
with their constitution, dynamics and multi-messenger radiation or
micro-physical properties, e.g.\ \cite{Murase+13pev,Ahlers+14pevnugal,Murase+18txs0506}, see
\cite{Ahlers+18ic3nurev,Schutz18gwrev,Murase+16uhenurev,Meszaros17arnpnurev,Abbott+18gwlivingrev}
for reviews.
The second level involves detailed numerical simulations of the dynamics of the formation of the sources
and their evolution leading to the state at which the various types of multi-messenger emission occurs,
e.g. \cite{Shibata+17grhydrobns,Easter+18gwbns,Parsotan+18grbphotsim,Vaneerten+18grbagrev}.
The third level involves detailed calculations and simulations of the radiation physics, using
large-scale numerical codes to describe the emission of multi-messengers, followed by their possible
changes during propagation from the source to the observer, and their detailed effects on particular
types of detectors, e.g. \cite{Fang+18crnuagnjet,Senno+16hidden,Keivani+17AMON,Hotokezaka+18nsmerg,
Biehl+18crnugrb,Alves+19cosmonu}.
For low source numbers or low signal rates, diffuse backgrounds must be calculated using simulated
source signals convolved over cosmological luminosity and redshift distributions,
e.g. \cite{Alves+19cosmonu}.
All three of these types of calculations will have to be considerably expanded and refined to address
and exploit the potential of the much more detailed data expected from the above new facilities.
Even for the semi-analytical studies, the increasing sensitivity and range of the detectors will
make it necessary to make use of farther and fainter reaching source luminosity functions, more
extensive redshift and mass distributions, intervening plasma and radiation field spectral densities, etc.
The source formation and dynamics studies, which are increasingly incorporating general relativistic
and magnetohydrodynamic (MHD) effects, will need to be extended to the 3-dimensional regime much more
commonly than before, and the use of adaptive mesh and shock capture methods will have to be exploited
and developed further. There will be an increasing need for a better cross-calibration of the various
Monte Carlo methods used in simulating high energy particle interactions and cascades, incorporating
the errors due to theoretical uncertainties, especially in the extrapolation to energies beyond those of
accelerator data. As another example, the atomic and nuclear physics of very heavy elements is poorly
understood and sparsely studied in the lab, yet to reliably elucidate the sources of r-process elements
in the Universe (lately ascribed largely to BNS mergers), the error estimates arising from these
theoretical and experimental lab uncertainties will need to be quantified and taken into account.
{ We must emphasize that nuclear and atomic physics lab experiments that can shed light on the
r-process details are crucial for our understanding of this important phenomenon.}
\section{\large Conclusions and Perspectives}
\label{sec:conclusion}
Some of the most important questions that will be addressed in the
next 5 to 10 years with upgraded GW detectors such as LIGO, VIRGO and
KAGRA, as they improve sensitivity at frequencies above 0.1\,kHz, are
to explore in detail the lower mass range of binary black hole
mergers, to test whether the final outcome of neutron star mergers is
a massive neutron star or a black hole, to probe the final ringdown of
space-time around a newly formed black hole, to determine the maximum
mass of neutron stars, to test whether General Relativity remains valid
under extreme density and pressure conditions, and to explore the nature
of the central engine of GRBs. As they push towards lower frequencies
below 10 Hz it will be possible to probe intermediate mass BHs of 100
to 500 solar masses, important for understanding how the most massive
BHs at the center of galaxies form. To get more reliable results, larger
interferometer arm lengths such as the $\sim$ 10 km of the Einstein or
Cosmic Explorer experiments will be needed. On the 10-20 year timescale,
even longer arms, such as the 2.5 million km in the space-based eLISA
interferometer, will measure frequencies $\sim 0.1$ Hz which can probe
the merger of $\mathrel{\mathpalette\simov >} 10^6\mbox{$M_\odot$}$ black hole mergers, important for
understanding the growth of galaxies and clusters of galaxies in the
Universe, and the existence of a primordial GW background left over
from the inflationary era of the early Bing Bang.
The neutrino detector upgrades such as PINGU in IceCube and ORCA in
KM3NeT will get better statistics in the 1-10 GeV energy range to
probe fundamental neutrino physics issues such as the hierarchy of the
mass ordering, as well as the mixing angles between flavors,
constraining the neutrino masses and testing the existence of sterile
neutrinos. These issues are likely to be resolved in the next 5-10
years { with the help of these and other detectors},
while the next generation IceCube Gen-2 is likely to identify the sources
of the origin of high energy neutrino, e.g.\ \cite{Murase+16uhenurev}.
In the Northern hemisphere, the 3 to 4\,Gton KM3NeT detector will be
able to observe HENs from our galactic center, where most of the
(so far undetected) galactic neutrino sources reside.
The upgraded ANITA balloons and, if funded, the ARIANNA and RNO/ARA
experiments in Antarctica will
{ make progress towards detecting the $10^{17}-10^{19}$ eV cosmogenic
neutrinos produced by UHECR protons interacting with the cosmic microwave background,
or by UHECR nuclei interacting with the diffuse starlight. These experiments will
also complement IceCube-Gen2 in the pursuit of the identification of tau-neutrinos.}
More reliable determinations may need to wait for larger experiments such as
GRAND and POEMMA, which will also address the chemical composition of the
highest energy UHECRs.
The next Galactic supernova should be an ideal and important event for
multi-messenger astrophysics, which can be exquisitely studied by the
Hyper-K \cite{Hyper-K+18instr} and JUNO \cite{Lu+16junosnnu}
experiments to address fundamental issues of neutrino oscillation and
supernova physics, e.g.\ \cite{Beacom10sndifnu,Tamborra+18ccsnnu}.
The next upgrades of the Auger and TA UHECR arrays are expected to
answer the chemical composition question independently of the answers
obtained through the above neutrino detectors, providing a much needed
consistency check. When the TAX4 array is completed its area will be
comparable to Auger's, and being in the Northern hemisphere while
Auger is in the Southern, will prove or disprove possible anisotropies
of arrival, also testing whether the UHECR production is dominated by
a few nearby sources or more numerous distant ones. Together, both of
them will probe the details of the properties of hadronic interactions
at energies three orders of magnitude higher than what is achievable in
laboratory accelerators. These questions can be more thoroughly
investigated from space with K-EUSO and POEMMA in the 10-15 years.
Future all-sky monitors such as the Large Spectroscopic Survey
Telescope (LSST) in the optical and the Square Kilometer Array (SKA)
in the radio, as well as the \textit{Fermi}, \textit{Swift}\ and expected (2022) SVOM
satellites at keV to GeV energies will provide rapid EM triggers and
accurate sky localization as well as follow-up capability for studying
cataclysmic events such as BH or NS mergers, supernovae, gamma-ray
bursts, AGN flares, etc., where CRs, neutrinos and GWs are also
expected in varying amounts. The strength and mix of these different
messengers is model-dependent, and multiple triggers in different
messengers, as well as model development and extensive simulations,
will be the key for understanding the physics of these energetic
sources.
The key to fully exploit the power of these new facilities is the
multi-messenger approach, due to the complementary advantages and
limitations of the different messenger particles. Cosmic rays provide
unique information about particle accelerators well beyond terrestrial
laboratory capabilities, and about source magnetic fields and total
energetics, but they do not reach us from beyond $\sim$100\,Mpc and have at
best poor angular resolution. Neutrinos reach us from the most
distant reaches of the Universe, and probe the inner, denser regions
of the most energetic and cataclysmic events, and ultra-high energy
neutrinos, being intimately linked to UHECRs, can provide unique clues
as to how the latter reach their enormous energies. Gravitational wave
observations probe the most compact regions of high energy sources; the
GW wave strain amplitude at earth is directly proportional to the source
compactness, measured in terms of $GM/c^2R$, and the GW luminosity goes
as the fifth power of the compactness. GWs provide excellent information
about central object masses, angular momenta, and distances, and they will
eventually be detected from the farthest distances and earliest epochs in the
Universe, but they have modest angular angular resolution at best, and do
not probe the bulk of the stellar or baryonic mass of their sources.
Electromagnetic waves provide excellent angular resolution, velocity
and redshift determination capabilities, but the opacity of matter
prevents them from probing the inner, denser regions of astronomical
sources, while at higher gamma-ray frequencies they cannot reach us
directly from the much grater distances probed by neutrinos or
gravitational waves. By using several messengers in conjunction,
astronomers and physicists can hedge the foibles of each against the
advantages of the others, forging them into a formidable toolkit
for probing the highest energies and densest, most violent corners of
the Universe, and in this fashion, putting our physical theories of
the universe to their most extreme and exacting possible tests.
\section{Explanatory Boxes}
\label{sec:boxes}
{\bf BOX 1: Neutrino Production and Oscillations}\\
\noindent
Cosmic ray protons $p_{cr}$ interacting with target protons $p_t$ and target photons $\gamma_t$
lead to a reduced energy proton $p$ or neutron $n$, and to intermediate charged or neutral
short-lived unstable particles, such as pions $\pi^{\pm,0}$, muons $\mu^\pm$, neutrons $n$
and (at higher energies) Kaons $K^\pm$, whose decay results in neutrinos $\nu_i$ of different
flavors $i$, $\gamma$-rays and electrons or positrons $e^\pm$.
{ CR nucleons, on the other hand, are primarily subject to spallation against target protons and
photo-disintegration against target photons, both resulting in smaller nucleons including protons,
the latter subsequently undergoing the same above mentioned interactions.}
\\
\begin{equation}
p_{cr} + p_t/\gamma_t \to p/n + \pi^\pm + \pi^0 + K^\pm + \cdots
\end{equation}
\begin{eqnarray}
\pi^+ \to & \mu^+ + \nu_\mu , ~~~~~~~~~~~~~\cr
& \mu^+ \to e^+ + \nu_e + {\bar \nu}_\mu \cr
\pi^- \to & \mu^- + {\bar \nu}_\mu , ~~~~~~~~~~~~~\cr
& \mu^- \to e^- + \barnue + \nu_\mu
\end{eqnarray}
\begin{eqnarray}
\pi^0 ~~\to & \gamma + \gamma \cr
K^+ /K^- \to & \mu^+ / \mu^- + \nu_\mu /{\bar \nu}_\mu \cr
n \to & p + e^- +\barnue
\end{eqnarray}
\noindent
The primary CR proton's mean relative energy loss per interaction, called the inelasticity, is
$\kappa_{pp}\sim 0.5$ for $pp$ (or $pn$) and $\kappa_{p\gamma}\sim 0.2$ for $p\gamma$ interactions.
For $p_{cr}$ interactions with target
protons (or target neutrons) the mean ratio of charged to neutral pion secondaries is $r_{\pm/0}\sim 2$,
and for interactions with target photons it is $r_{\pm/0}\sim 1$. The electrons and positrons $e^\pm$
quickly lose their energy via synchrotron or inverse Compton scattering, resulting in further $\gamma$-rays,
so the final result of the $p_{cr}+ p_t/\gamma_t$ interactions are high energy neutrinos $\nu_i$ and
$\gamma$-rays. The mean energy of the resulting neutrinos and $\gamma$-rays is $\sim 0.05$ and $\sim 0.1$
of the initial CR proton's energy.
Once neutrinos of any flavor are produced, during their travel from the source to the observer
the neutrinos of any flavor can change into neutrinos of any of the three flavors, in the so-called
neutrino oscillation phenomenon. As a result, independently of the ratio of neutrino flavors
initially produced a the source where the CR interactions took place, after traveling over astronomical
distances typically all three neutrino flavors are present at the observer.
{ For the neutrino sources considered here, the observable neutrino flavors are expected to be
oscillation-averages due to the large source distances and the finite energy resolution of
the neutrino observatories.}
\\
\eject
{\bf BOX 2: Multi-messengers and their inter-relation~ }
\begin{figure}[H]
\begin{minipage}{0.56\textwidth}
\centerline{\includegraphics[width=4in]{multi-transient-fig-v4-small.jpg} }
\end{minipage}
\hspace{9mm}
\begin{minipage}[t]{0.32\textwidth}
\vspace*{-1.3in}
\caption{\footnotesize
Multiple messenger particles { possibly emanating from
(A) a blazar flare;
(B) a tidal disruption event (gravitational waves detectable by eLISA for some events);
(C) a long gamma-ray burst;
(D) an engine-driven supernova, or
(D) a supernova (gravitational waves detectable for Galactic events);
(F) a double black hole merger, or
(G) a double neutron star merger leading to a short gamma-ray burst.} }
\label{fig:multi-transient-fig}
\end{minipage}
\end{figure}
A multi-messenger source might emit two, three, or even all four
different types of messengers.
{
For instance, panel (G) of Fig.\ref{fig:multi-transient-fig} shows a binary
neutron star merger such as the GW/GRB 170817 event, from which two
types of multi-messengers, gravitational waves (GW) and photons ($\gamma$),
were observed \cite{LIGO+17gw170817disc,Troja+17gw170817xr,
Kasliwal+17gw170817concord}. Such sources may also emit high energy
neutrinos (HENs) and cosmic rays (CRs) e.g.
\cite{Kimura+17sgrbnu,Kimura+18transejnu,Kimura+18bnscr}, although for
this particular source theories predict fluxes too low for current
detectors; if so, even closer events or next-generation HEN facilities
will be required to observe HEN from these sources.
Another panel, (B), shows a tidal disruption event (TDE) of a star}
by a massive black hole; in this case shocks in the disrupted gas can
accelerate particles and lead to CRs and HENs,
e.g.\ \cite{Guepin+17tdecrnu,Biehl+17tdecrnu,Senno+17tde,Wang+16tdecrnu}.
TDEs involving white dwarf stars and $\sim$1000\,\mbox{$M_\odot$}\ black holes lead to strong
low-frequency ($\sim$1\,mHz) gravitational wave emission that could be
observed by the forthcoming eLISA mission.
{ A solitary supermassive black hole with a jet may emit \mbox{gamma-rays}, HEN,
and cosmic rays (panel B)},
as we suspect occurred during the 2017 flaring episode of the BL~Lac-type
blazar \mbox{TXS~0506+056} \cite{Magic+17blazarnu,Keivani+18txs0506,NuSTAR+17blazarnu,
IC3multi+18txs0506,Gao+19txs0506,Cerruti+19txs0506}.
Here and in related sources, the coproduction of CRs, HEN, and
high-energy \mbox{gamma-rays}\ is anticipated, as the physics of these three
messengers are closely connected -- high-energy particle acceleration
and shocks lead to the interaction of highly-relativistic protons (or
nuclei) with ambient gas or intense radiation fields, resulting in
neutrinos, \mbox{gamma-rays}, and $e^\pm$.
For single objects, even those of extreme mass and undergoing
substantial accretion, relatively weak gravitational wave emission is
expected as the time-varying quadrupole moment (which requires the
breaking of azimuthal symmetry) in these cases are thought to be
minimal. { The sole exception would be a Galactic engine-driven
supernova or a Galactic supernova (panels D and E), which would be}
sufficiently nearby that detection of coherent or incoherent
gravitational waves by current and future ground-based detectors is
anticipated. { As far as thermal ($\sim 10$ MeV) neutrino detection
from such supernovae, IceCube is well matched, having for such emission
also a roughly galactic reach.
A challenge for theory is to predict the amplitude and
spectrum of GW and neutrinos from different types of supernovae. }
Strong GW emissions have been observed from the mergers of compact
binary systems, either from { two merging stellar-mass black holes (panel F)
\cite{LIGO+18bbhGWTC-1}, two merging neutron stars (panel G)}
\cite{LIGO+17gw170817disc}, or (in the future) BH-NS mergers, because
the final in-spiral to coalescence yields a strong gravitational wave
signal in the ``sweet spot'' frequency range for ground-based
gravitational wave detectors. In the case of 30\,\mbox{$M_\odot$}\ +
30\,\mbox{$M_\odot$}\ black hole binary systems, such coalescence events can
already be observed out to $\sim$500\,Mpc distances
\cite{Abbott+18gwlivingrev}.
However, in the case of BH-BH mergers little EM flux is expected,
because the ambient matter density (protons, electrons) in the
vicinity of the binary, at the time of the merger, is typically very
low.
A key exception is the accreting supermassive black holes at the
centers of massive galaxies, which are expected to merge in the wake
of the coalescence of their component galaxies. These SMBH mergers are
key targets for the eLISA mission, and may well exhibit accompanying
EM, CR, and HEN emission \cite{Klein+16eLISA-MBH}.
\\
\noindent
{\it Acknowledgments:} We are grateful to Stephane Coutu, Douglas Cowen, Miguel Mostaf\'a and
Bangalore Sathyaprakash and the referees for useful discussions and comments.
\bibliographystyle{revtex}
\footnotesize
\input{nat-arx2.bbl}
\end{document}
|
\section*{Acknowledgments}
This work is supported by National Science Foundation (NSF) under NSF grants CNS- 1920182, CNS-1532061, CNS-1551221.
\bibliographystyle{IEEEtran}
\section{Conclusion}
\label{sec-conclusion}
Of all the many existing medical imaging modalities, X-ray imaging remains the most widely used modality as it is the most cost effective and one of the easiest to administer. Chest X-ray remains an essential imaging modality for the diagnosis and follow up treatment of many diseases affecting the lungs, heart and bone structure within the chest area. In this study, a new deep learning based \textit{image-to-images} approach was proposed that simultaneously suppresses the bones that hinder the visibility and scrutiny of organs and nodules and segments the organs within the chest area. Essentially, and for the first time, the architectural design of this deep learning-based model exploits in the most effective way the interplay of parameters in between the two tasks to optimize the outcome for both tasks at once. In order to perform these two essential tasks of bone suppression and organs segmentation, the well-established pix2pix network is extended to generate two output images simultaneously (an image with bones suppressed, and a second image showing the segmented organs), yielding the new \textit{image-to-imag\underline{es}} with an automated end-to-end framework instead of the traditional \textit{image-to-image} approach that deals with each task separately. The proposed method was trained via an end-to-end process and is evaluated by cross validation and significance testing with several standard metrics, resulting in highly accurate results for both tasks. Through two additional empirical evaluations involving low-dose CT images and neuroimaging, the proposed architectural design of the model is shown to be amenable for generalization to other domains of application, although developed around CXR imaging.
In summary, the contributions of this work can be summarized as follow:
\begin{itemize}
\item This work is the first to try to extend the application of \textit{image-to-image} network to \textit{image-to-images} network, while optimizing the use of parameters and securing computational efficiency.
\item The network is improved through the inclusion of dilated convolutions in some specific layers, which are shown to improve the accuracy of the results significantly.
\item The \textit{image-to-images} network is used to accomplish simultaneously the two common and most needed tasks of bone suppression and organ segmentation in CXR images while optimizing the outcomes for both.
\item All of the developed code is shared publicly online for validation purposes and for use by the research community interested in the automated diagnosis and in treatment follow up using chest X rays.
\end{itemize}
\section{Results and discussion}
\label{sec-results}
In this section, the experimental results for both tasks are reported and compared with state-of-the-art methods in the first two subsections. This is followed by a subsection which provides the performance and qualitative analysis, investigates the ablation study and demonstrates how this model is amenable to other application domains (i.e. generalization of the model) with preliminary results obtained on low-dose CT image processing and on a neuroimage translation problem. The generalization subsection shows the benefits of multitasking in comparison with single tasking and the potential for future work of \textit{pix2pix MTdG} towards resolving other \textit{image-to-imag\underline{es}} translation problems. Notice that, for simplification in plotting the figures \textit{p2p} notation is used instead of \textit{pix2pix}.
\subsection{Task 1: Organ Segmentation}
\begin{figure}[!t]
\centering
\captionsetup{justification=centering}
\minipage{0.4\textwidth}
\includegraphics[width=\linewidth]{Figures/Dice_res.pdf}
\endminipage\vfill
\minipage{0.4\textwidth}
\includegraphics[width=\linewidth]{Figures/rfn_res.pdf}
\endminipage
\caption{The box plot of the segmentation accuracy achieved by \textit{u-net} and \textit{pix2pix MTdG}. \\ up) Dice Scores, Down) False Negative Rates.}
\label{fig:compare-T1}
\end{figure}
The segmented regions of the heart, the left lung and right lung generated by the model as output masks, are associated with the same regions in the ground-truth by using standard evaluation metrics in image processing, namely the Dice and Jaccard scores, false-negative rate (FNR) and false-positive rate (FPR).
The Jaccard index is a metric that measures the percent overlap between the target ground-truth mask (GT) and our prediction mask (PM).
The overlap area between GT and PM would be the true positive area (TP). The area which is predicted in PM but is not in PM is a false-positive (FP) and inversely for false-negative (FN) defining the area which is in GT but not in PM.
Jaccard metric is closely related to the Dice coefficient, which is not as easily described geometrically.
The false-positive rate indicates the area ratio of predicted mask which had no associated ground truth mask, and similarly false-negative indicates the area ratio of the ground truth mask which had no associated predicted mask.
These standard evaluations metrics can be expressed mathematically as follow where symbol ! defines the binary negation operator:
\begin{gather}
Dice=\frac{2 \times (PM \cap GT)}{PM + GT}= \frac{2 \times TP}{TP + FP + TP + FN} \\
Jaccard=\frac{PM \cap GT} {PM \cup GT} = \frac{TP}{TP + FP+ FN}
\\
FNR=\frac{!PM \cap GT}{GT}= \frac{FN}{TP + FN}
\\
FPR= \frac{PM \cap !GT}{GT} = \frac{FP}{TP + FN
\label{eq:SegScores}
\end{gather}
The well-known U-net method \cite{u-net} is used as the competitor due to its state-of-the-art performance in various segmentation applications \cite{u-net-nature} as well as for the reported CXR segmentation studies with promising results \cite{seg-u-net-brazil, seg-u-net-scia, seg-u-net-islam2018towards}.
The average and standard deviation of the segmentation results, by all metrics, are summarized in Table \ref{tab:segmentation}.
The best achieved results, highlighted in blue, demonstrate that the \textit{multitask pix2pix} with an embedded dilation in the generator (\textit{MTdG}) surpasses the \textit{u-net} method.
For further statistical investigation, figure \ref{fig:compare-T1} shows the box plots of the segmentation scores evaluated using the Dice and false-negative rate of the heart, left and right lung for the \textit{u-net} and \textit{pix2pix MTdG}.
The inner line, the bottom and top edges of the box indicate median, the 25th, and 75th percentiles, respectively. While the whiskers are extended to the most extreme not outlier data points, while the outliers are plotted individually using the '+' symbol.
As an example, the segmentation result of the \textit{pix2pix MTdG} for the best and worst achieved Dice scores are shown in Figure \ref{fig:image-dice}.
\begin{table}[]
\centering
\captionsetup{justification=centering}
\minipage{0.5\textwidth}
\centering
\caption{Segmentation results of different methods while the best scores are colored blue. }
\label{tab:segmentation}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& & \textbf{u-net} & \textbf{pix2pix MTdG} & \textbf{p-value} \\ \hline
\multirow{4}{*}{\textbf{Dice}} & Left lung & 0.983 $\pm$ 0.005 & 0.990 $\pm$ 0.006 & 1.2 e-93 \\
& Heart & 0.965 $\pm$ 0.014 & 0.977 $\pm$ 0.015 & 3.6 e-57 \\
& Right lung & 0.980 $\pm$ 0.009 & 0.988 $\pm$ 0.006 & 5.5 e-99 \\
& \textbf{Average} & 0.976 $\pm$ 0.007 & \textcolor{blue}{ 0.985 $\pm$ 0.007 } & - \\ \hline
\multirow{4}{*}{\textbf{Jaccard}} & Left lung & 0.967 $\pm$ 0.010 & 0.980 $\pm$ 0.011 & 7.7 e-97 \\
& Heart & 0.933 $\pm$ 0.026 & 0.956 $\pm$ 0.027 & 3.8 e-59 \\
& Right lung & 0.961 $\pm$ 0.017 & 0.977 $\pm$ 0.012 & 6.9 e-105 \\
& \textbf{Average} & 0.953 $\pm$ 0.013 & \textcolor{blue}{ 0.971 $\pm$ 0.013 } & - \\ \hline
\multirow{4}{*}{\textbf{FNR}} & Left lung & 0.010 $\pm$ 0.005 & 0.008 $\pm$ 0.005 & 8.8 e-18 \\
& Heart & 0.028 $\pm$ 0.017 & 0.018 $\pm$ 0.017 & 7.1 e-24 \\
& Right lung & 0.015 $\pm$ 0.008 & 0.010 $\pm$ 0.006 & 3.0 e-45 \\
& \textbf{Average} & 0.017 $\pm$ 0.006 & \textcolor{blue}{ 0.012 $\pm$ 0.007 } & - \\ \hline
\multirow{4}{*}{\textbf{FPR}} & Left lung & 0.024 $\pm$ 0.013 & 0.013 $\pm$ 0.013 & 3.0 e-73\\
& Heart & 0.043 $\pm$ 0.033 & 0.028 $\pm$ 0.029 & 2.3 e-22 \\
& Right lung & 0.026 $\pm$ 0.017 & 0.013 $\pm$ 0.012 & 8.3 e-67 \\
& \textbf{Average} & 0.031 $\pm$ 0.015 & \textcolor{blue}{0.018 $\pm$ 0.013} & - \\ \hline
\end{tabular}
\endminipage
\end{table}
\begin{table*}[]
\minipage{0.9\textwidth}
\begin{center}
\begin{threeparttable}
\caption{Comparison the segmentation results of different methods on JSRT dataset.}
\label{tab:seg-comp}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\textbf{Method} & \textbf{Image Size} & \textbf{Augmentation} & \textbf{Evaluation scheme} & \textbf{Dice} & \textbf{Jaccard} & \textbf{Dice} & \textbf{Jaccard} \\ \hline
& & & & \multicolumn{2}{c|}{Lungs} & \multicolumn{2}{c|}{Heart} \\ \hline
\textbf{Human observer} \cite{seg-human+ASM+voting} & 2048$\times$2048 & No & - & - & 0.946 & - & 0.887 \\ \hline
\textbf{InvertedNet} \cite{seg-InvertedNet} & 256$\times$256 & No & 3-fold CV & 0.974 & 0.950 & 0.937 & 0.882 \\ \hline
\textbf{u-net} by \cite{seg-u-net-brazil} & 256$\times$256 & No & 5-fold CV & 0.976 & 0.962 & - & - \\ \hline
\textbf{u-net} by \cite{seg-u-net-scia} & 256$\times$256 & No & 5-fold CV & - & 0.959 & - & 0.899 \\ \hline
\textbf{u-net} by us & 512$\times$512 & Yes & 5-fold CV * & 0.981 & 0.964 & 0.965 & 0.933 \\ \hline
\textbf{MTdG } (proposed) & 512$\times$512 & Yes & 5-fold CV * & \textcolor{blue}{0.989} & \textcolor{blue}{0.978} & \textcolor{blue}{0.977} & \textcolor{blue}{0.956} \\ \hline
\textbf{u-net} by \cite{seg-u-net-islam2018towards} & 512$\times$512 & Yes & train/test split (80\%/20\%)** & 0.986 & - & - & - \\ \hline
\textbf{SegNet} by \cite{seg-u-net-brazil} & 256$\times$256 & No & 5-fold CV & 0.979 & 0.955 & 0.944 & 0.896 \\ \hline
\textbf{SCAN} \cite{seg-SCAN} & 400$\times$400 & No & train/test split (209/38) & 0.973 & 0.947 & 0.927 & 0.866 \\ \hline
\textbf{FCN} by \cite{seg-u-net-brazil} & 256$\times$256 & No & 5-fold CV & 0.974 & 0.950 & 0.942 & 0.892 \\ \hline
\textbf{MTdG } (proposed) & 256$\times$256 & No & 5-fold CV & 0.974 & 0.962 & 0.934 & 0.928 \\ \hline
\textbf{MTdG } (proposed) & 256$\times$256 & No & 3-fold CV & 0.962 & 0.953 & 0.921 & 0.916 \\ \hline
\textbf{ASM tuned} \cite{seg-human+ASM+voting} & 256$\times$256 & No & 2-fold CV & - & 0.927 & - & 0.814 \\ \hline
\textbf{Hybrid voting} \cite{seg-human+ASM+voting} & 256$\times$256 & No & 2-fold CV & - & 0.949 & - & 0.86 \\ \hline
\textbf{Seghers et. al} \cite{seg-Seghers} & 256$\times$256 & No & train/test split (50/44) & - & 0.951 & - & - \\ \hline
\textbf{Ibragimov et. al} \cite{seg-Ibragimov} & - & No & - & - & 0.953 & - & - \\ \hline
\end{tabular}
\begin{tablenotes}
\footnotesize
\item *: p-values from significance test are reported in Table \ref{tab:segmentation}.
\item **: JSRT dataset is not used for this work.
\end{tablenotes}
\end{threeparttable}
\end{center}
\endminipage\hfill
\end{table*}
\begin{figure}[]
\centering
\captionsetup{justification=centering}
\minipage{0.5\textwidth}
\includegraphics[width=1\linewidth]{Figures/best-dice.png}
\endminipage\vfill
\minipage{0.5\textwidth}
\includegraphics[width=1\linewidth]{Figures/worst-dice.png}
\endminipage\vfill
\caption[Caption for LOF]{Results by proposed method \textit{pix2pix MTdG} regarding to the best (top) and worst (bottom) Dice scores. \\
\footnotesize{Columns left to right are, input image, segmentation result, segmentation target, bone suppression result, and bone suppression target. \\
Top) Best Dice: average of Dice, Jaccard, FPR and FNR are 0.99, 0.99, 0.01 and 0.01. \\
Bottom) Worst Dice: average of Dice, Jaccard, FPR and FNR are 0.94, 0.90, 0.12 and 0.01. } \\
}
\label{fig:image-dice}
\end{figure}
To the best of our knowledge, no multitask framework has been found in the literature to benchmark the proposed multitask network for our tasks. A comparison between \textit{u-net} (implemented by us) and \textit{pix2pix MTdG} along with the p-values of student's t-test is provided in Table \ref{tab:segmentation} which clearly shows that the \textit{pix2pix MTdG} method yields better results with a significant difference ($p < 0.001$). However, to contrast these results with other methods in the literature, due to the variations in utilizing different folding schemes and image sizes, a fair comparison of the results is not a straightforward process.
The results of different state-of-the-art algorithms on the JSRT dataset are summarized in Table \ref{tab:seg-comp}. The settings reported from each method are also provided in Table \ref{tab:seg-comp} with `-' to mean that the value is not reported. For a fair comparison, the proposed method has also been tested using $256\times256$ image size, 3-fold cross-validation and without any augmentation. As presented in the Table, while the scores are really close to each other, the best-achieved results are the ones provided by \textit{pix2pix MTdG} in $512\times512$ image resolution. In all these other settings, the \textit{pix2pix MTdG} performance is still reasonable and is comparable to the performance of other techniques.
\subsection{Task 2: Bone Suppression}
The second task of the \textit{pix2pix MTdG} network is bone suppression. The results of this task are evaluated via the structural similarity index (SSIM) metric for similarity estimation, and the root mean squared error (RMSE) metric to measure the difference between predicted and actual values \cite{ssim}. The RMSE measure between two X and Y images is expressed by Equation \eqref{eq:RMSE}, where $N$ in the total number of pixels in image and $i$ is the pixel index.
SSIM is a reference-based quality assessment metric, which compares the local patterns of pixel intensities between the reference and output images. The maximum value of 1 implies that the two images are structurally similar, while a zero value indicates that there is no structural similarity between them. Usually, the SSIM index is calculated via windowing on the images with 8x8 window size and 1 pixel striding. At the end, the mean of the computed values (M-SSIM) would be reported. The M-SSIM measure between two images $X$ and $Y$ and the default SSIM measure between two windows $W^X_i$ and $W^Y_i$ are as defined by Equations \eqref{eq:MSSIM} and \eqref{eq:SSIM} where $\mu_{W^X_i}$, $\mu_{W^Y_i}$, $\sigma_{W^Y_i}^2$ , $\sigma_{W^X_i}^2$, and $\sigma_{W^X_iW^Y_i}$ show the average of $W^X_i$, the average of $W^Y_i$, the variance of $W^X_i$, the variance of $W^Y_i$ and the Covariance of $W^X_i$ and $W^Y_i$, respectively. The default settings of $c_1=(0.01 \ L)^2$ and $c_2=(0.03 \ L)^2$ are considered with $L$ being the dynamic range of the pixel-values (i.e. 255 in our experiment).
\begin{gather}
RMSE(X,Y)=\sqrt{\frac{1}{N} \sum_{i=1}^N (x_i-y_i)^2}
\label{eq:RMSE}
\end{gather}
\begin{gather}
M{\text -}SSIM= \frac{1}{N} \sum_{i=1}^N SSIM(W^X_i,W^Y_i)
\label{eq:MSSIM}
\\
SSIM(W^X_i,W^Y_i)= \hspace{120pt} \nonumber \\ \hspace{50pt} \frac{(2\mu_{W^X_i}\mu_{W^Y_i}+c_1)(2\sigma_{W^X_iW^Y_i}+c_2)}{(\mu_{W^X_i}^2+\mu_{W^Y_i}^2+c_1)(\sigma_{W^X_i}^2+\sigma_{W^Y_i}^2+c_2)}
\label{eq:SSIM}
\end{gather}
\begin{figure}[]
\centering
\captionsetup{justification=centering}
\includegraphics[width=0.9 \linewidth,trim={0 0 0 0cm},clip]{Figures/BSE_res.pdf}
\caption{Bone suppression task showing boxplots of the results. \\ Left) M-SSIM similarity score. Right) RMSE difference.}
\label{fig:SSIM}
\end{figure}
\begin{figure}[]
\centering
\captionsetup{justification=centering}
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{Figures/best-rmse.png}
\endminipage\vfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{Figures/worst-rmse.png}
\endminipage\vfill
\caption{Results by proposed method \textit{pix2pix MTdG} regarding to the best (top) and worst RMSEs (bottom). \\
\footnotesize{Columns left to right: input image, segmentation result, segmentation target, bone suppression result and bone suppression target. \\
Top) Best RMSE: SSIM: 0.99, RMSE : 2.66. \\
Bottom) Worst RMSE: SSIM: 0.94, RMSE : 10.55.} }
\label{fig:image-rmse}
\end{figure}
\begin{table}[!t]
\centering
\caption{Results of bone suppression task via different methods.}
\label{tab:Bone}
\begin{tabular}{|c|c|c|c|}
\hline
& \textbf{AutoEncoder} \cite{bonesup-new-1} & \textbf{pix2pix MTdG} & \textbf{p-value} \\ \hline
\textbf{MSSIM} & 0.970 $\pm$ 0.011 & \textcolor{blue}{0.976 $\pm$ 0.006} & 8.6 e-24 \\ \hline
\textbf{RMSE} & 5.096 $\pm$ 1.812 & \textcolor{blue}{4.297 $\pm$ 1.046} & 5.3 e-45 \\ \hline
\end{tabular}
\end{table}
\begin{table*}[!h]
\centering
\caption{Summary of the properties of different methods.}
\label{tab:summary}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\textbf{Task} & & \textbf{pix2pix ST} & \textbf{pix2pix STdG} & \textbf{U-net} & \textbf{AutoEncoder} & \textbf{pix2pix MT} & \textbf{pix2pix MTdG} \\ \hline
\multirow{2}{*}{\textbf{Segmentation}} & \textbf{Average Dice} & 0.977 $\pm$ 0.008 & 0.984 + 0.008 & 0.976 $\pm$ 0.007 & - & 0.978 $\pm$ 0.008 & \textcolor{blue}{0.985 $\pm$ 0.007} \\ \cline{2-8}
& \textbf{Average FNR} & 0.018 $\pm$ 0.008 & 0.014 + 0.009 & 0.017 $\pm$ 0.006 & - & 0.018 $\pm$ 0.008 & \textcolor{blue}{0.012 $\pm$ 0.007} \\ \hline
\multirow{2}{*}{\textbf{Rib Suppression}} & \textbf{Average MSSIM} & 0.969 $\pm$ 0.009 & 0.974 $\pm$ 0.007 & - & 0.970 $\pm$ 0.011 & 0.968 $\pm$ 0.007 & \textcolor{blue}{0.976 $\pm$ 0.006} \\ \cline{2-8}
& \textbf{Average RMSE} & 5.296 $\pm$ 1.688 & 4.697 $\pm$ 1.587 & - & 5.096 $\pm$ 1.812 & 5.382 $\pm$ 1.470 & \textcolor{blue}{4.297 $\pm$ 1.046} \\ \hline
\textbf{} & \textbf{No. Parameters} & \multicolumn{2}{c|}{2 $\times$ 57,190,084} & 31,084,008 & 64,400 & \multicolumn{2}{c|}{57,199,303} \\ \hline
\textbf{} & \textbf{Minimum Epochs} & \multicolumn{2}{c|}{250 } & 48 & 150 & \multicolumn{2}{c|}{300 } \\ \hline
\textbf{} & \textbf{Training time} & \multicolumn{2}{c|}{786 = 2 $\times$ 393 min } & 248 min & 852 min & \multicolumn{2}{c|}{662 min } \\ \hline
\end{tabular}
\end{table*}
While there is much literature about bone suppression in CXR images and even with the availability of proprietary software, there is no shared dataset available nor is there any open source codes or models that are shared with the research community. There is just one competitor study for bone suppression with shared source code in which the authors used an interesting \textit{AutoEncoder} network architecture \cite{bonesup-new-1}.
The box plots of the achieved M-SSIMs and RMSEs for the \textit{pix2pix MTdG} and \textit{AutoEncoder} are shown in Figure \ref{fig:SSIM} and Table \ref{tab:Bone}, proving the good performance of the proposed \textit{pix2pix MTdG}. The best and worst results with regards to the RMSE measurement are shown in Figure \ref{fig:image-rmse}.
\subsection{Discussion}
As discussed earlier, the proposed \textit{pix2pix MTdG} method provides promising results in accomplishing both segmentation and bone suppression tasks simultaneously. In this subsection, other characteristics of the proposed method are discussed and a summary of these results is provided in Table \ref{tab:summary}.
\begin{figure*}[]
\captionsetup{justification=centering}
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/loss_ST_t1.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/loss_ST_t2.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/loss_STdG_t1.pdf}
\endminipage\hfill
\vfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/loss_STdG_t2.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/loss_MT.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/loss_MTdG.pdf}
\endminipage\hfill
\caption{The loss curves of the different schemes for a training session as a function of epochs. \\
\footnotesize{p2p: \textit{pix2pix}, ST: single task, STdG: single task with dilation in generator, MT: multitask, MTdG: Multitask with dilation in generator.} }
\label{fig:loss}
\end{figure*}
\subsubsection{Performance Analysis}
The performance of the proposed method is assessed here with respect to the results obtained and the network parameters that were considered. The following remarks can be made: \\
\textbf{a}) As shown in Tables \ref{tab:segmentation} and \ref{tab:summary}, the proposed multitask pix2pix with dilation (\textit{pix2pix MTdG}) achieves the best results for both organs segmentation and bone suppression tasks. \\
\textbf{b}) Using multitask pix2pix without dilation was not as effective for improving the results. \\
\textbf{c}) Another advantage of \textit{pix2pix MTdG} is in the number of required parameters of the network, which is an intrinsic requirement for the multitask pix2pix scheme. The trainable parameters of pix2pix, u-net, and multitask pix2pix are $57,190,084$; $31,084,008$ and $57,199,303$, respectively. Note that the u-net architecture can only perform the segmentation task. To employ two separate pix2pix networks for the two tasks, a large number of training parameters $2 \times 57,190,084$ would be needed. In other words, the performance of \textit{multitask pix2pix} is reasonable in the number of parameters used while it maintains comparable good results to the state-of-the-art techniques. \\
\textbf{d}) Nonetheless, the\textit{ multitask pix2pix }framework has one drawback, which is the number of required iteration/epochs for the training phase. In the experiments conducted in this study, \textit{pix2pix MTdG} required almost $300$ epochs while u-net and single task \textit{pix2pix} converged in only 50 and 250 epochs, respectively, meaning that the multitask framework requires more training time in comparison with other methods. It is worth noting that the above concern is just for training and since a pre-trained network is used to generate the output in the testing phase, both methods perform similarly fast requiring only 1.2 seconds.
\begin{figure*}[]
\captionsetup{justification=centering}
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/dice_left.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/dice_heart.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/dice_right.pdf}
\endminipage\hfill
\vfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/rfn_left.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/rfn_heart.pdf}
\endminipage\hfill
\minipage{0.33\textwidth}
\includegraphics[width=\linewidth]{Figures/rfn_right.pdf}
\endminipage\hfill
\caption{Segmentation task evaluation showing box-plots for different schemes. \\ (Left) left lung, (Middle) heart, (Right) right lung. \\
\footnotesize{(\textit{pix2pix ST}, \textit{pix2pix STdG}, \textit{pix2pix MT} and \textit{pix2pix MTdG})}
}
\label{fig:ablation-T1}
\end{figure*}
\begin{figure}[]
\centering
\captionsetup{justification=centering}
\minipage{0.35\textwidth}
\includegraphics[width=\linewidth]{Figures/Ablation_ssim.pdf}
\endminipage\vfill
\minipage{0.35\textwidth}
\includegraphics[width=\linewidth]{Figures/Ablation_rmse.pdf}
\endminipage
\caption{Bone suppression task evaluation showing the box-plots for the different schemes. \\ \footnotesize{(\textit{pix2pix ST}, \textit{pix2pix STdG}, \textit{pix2pix MT} and \textit{pix2pix MTdG})} }
\label{fig:ablation-T2}
\end{figure}
\begin{figure}[]
\centering
\captionsetup{justification=centering}
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{Figures/comp-3.png}
\endminipage\vfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{Figures/comp-5.png}
\endminipage\vfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{Figures/comp-2.png}
\endminipage\vfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{Figures/comp-6.png}
\endminipage\vfill
\caption{Segmentation results of different schemes for different subjects. Top to bottom are subjects and left to right are input, target, \textit{pix2pix ST}, \textit{u-net}, \textit{pix2pix MT} and \textit{pix2pix MTdG}.}
\label{fig:comparison}
\end{figure}
\subsubsection{Ablation Study}
In order to investigate the effectiveness of dilation and multitasking, the ablation study is considered. For this reason, six processing schemes are addressed: two \textit{single task pix2pix} (\textit{p2p ST}) for two tasks, two \textit{single task pix2pix with dilation} (\textit{p2p STdG}), \textit{multi task pix2pix} (\textit{p2p MT}) and \textit{multitask pix2pix with dilation} (\textit{p2p MTdG}). Figure \ref{fig:loss} shows the loss curves for the training session of the schemes with respect to epochs. The loss GAN, loss L1 and loss discriminator expressed as $\mathbb{E} [-log(F +\epsilon )]$, $\mathbb{E} [\mid Y - \hat{Y} \mid_1 ] $ and $\mathbb{E} [-( \ log(R + \epsilon) + log(1-F + \epsilon ) \ )]$, respectively, were included in the loss functions \eqref{eq:OurLoss-G} and \eqref{eq:OurLoss-D}. The training is stopped when the L1 loss reaches and stabilizes at almost $0.005$. As shown in Fig. \ref{fig:loss}, all of the schemes are able to converge and make the generator the declared winner. Furthermore, while the multitask schemes need more epochs for reaching the desired L1 loss value, there is no significant difference between the achieved final loss values.
\begin{figure*}[!hb]
\centering
\captionsetup{justification=centering}
\includegraphics[width=\linewidth, trim={0 23cm 0cm 0},clip]{Figures/Figure_generalization.pdf}
\caption{Preliminary results of two more applications of \textit{multitask pix2pix} for joint tasks. \\
Top row) Low dose CT: image enhancement and segmentation. \\
Mid row) MRI Neuroimage translation: T1 to T1-inverse and T2-flair. \\
Bottom row) Magnification of the selected region (area inside yellow rectangle). \\
\footnotesize{p2p: \textit{pix2pix}, ST: single task, MTdG: Multitask with dilation in generator.} }
\label{fig:generalization}
\end{figure*}
In contrast to the almost similar final training loss values of the different schemes, as shown in Table \ref{tab:summary}, there are differences in the test sessions and in the outputs of the schemes' models. The best results are achieved by \textit{pix2pix MTdG} while the multitasking without dilation was not helpful. This could be inferred as a hyper-parameter selection criterion, as the dilation makes the receptive fields bigger to become more suitable to the nature of the CXR input images and the organs' shapes and sizes. On the other hand, when dilation is exploited in the network, multitasking (\textit{p2p MTdG}) assumes the benefits and generates better results than those obtained from a single task with dilation (\textit{p2p STdG}); this outcome is achieved despite the fact that the multitask network has half of the weights in comparison to two single-task networks. Figures \ref{fig:ablation-T1} and \ref{fig:ablation-T2} demonstrate the results of segmentation and bone suppression tasks for different schemes as can be observed from the box-plots of Dice score, false negative rates, RMSE and M-SSIM score in concurrence to the results shown in table \ref{tab:summary}, especially when comparing for outliers, percentile edges and median line.
\subsubsection{Qualitative Analysis}
Figure \ref{fig:comparison} presents the segmentation results of the different methods for various subjects. From left to right, the images shown are the input image, the target image, single task \textit{pix2pix} output, \textit{u-net} output, \textit{pix2pix MT} output, and \textit{pix2pix MTdG} output. In retrospect, the following assessments can be made:\\
\textbf{a}) As presented in the first row, the \textit{pix2pix MTdG} framework delivered the best outcomes in comparison with the other methods. \\
\textbf{b}) However, the result of all the other methods in the second row is almost the same. \\
\textbf{c}) The third row shows that the \textit{multitask pix2pix} without dilation achieved better results in contrast to the other methods. This could be related to the fact that this dilation process may not be as effective in some specific cases. We intend to combine the layers with and without dilation as future work to see if the accuracy could be improved further. \\
\textbf{d}) Moreover, as it is illustrated in the fourth row of Figure \ref{fig:comparison}, the \textit{pix2pix} based methods suffer from false-positive segments such as isolated islands. Although this is the case for the u-net method as well, u-net demonstrates better results in this case. This could be associated with the fact that the loss function of u-net is addressing the segmentation constraints while the loss function of \textit{pix2pix} is constructed to perform the pixel-wise comparison. This drawback could simply be removed by employing post-processing techniques such as connected components and considering the island area, which is not the aim of this paper at this juncture, but could be exploited in future work. Another approach to deal with this drawback is to implement segmentation related loss functions such as dice score to \textit{pix2pix} loss function. While this technique was discussed by some authors in the literature, our investigations did not prove its efficiency in delivering any improvement to our case.
\subsubsection{Generalization of Method and Future Work}
\label{subsec:generalization}
The size of the input CXR images considered in this study are 512x512. This resolution chosen for the proposed method is also exploited in the Stanford's \textit{CheXNext} network \cite{general-Chest-Andrew} with promising results. Nonetheless, manufacturers of X-ray radiography continue to improve its resolution and hence 512x512 could be limiting in obtaining similar optimal results sought with a higher resolution. Since more pixels mean more revealing information that could enhance the prospects for more accurate diagnosis, it is thus reasonable to extend the proposed network for dealing with higher resolution images such as 2048x2048. It should be noted that since the generator of the proposed method is a holistic network (not locally and patch or block-based), it is expected to yield more promising results with higher resolution images, but at the expense of more convolutional layers and with a higher demand for more variables to contend with and hence more taxing computational requirements.
With these contending challenges in mind, and in order to show the potential benefits of the \textit{image-to-images} translation and \textit{multitask pix2pix} with respect to the generalization aspect to other domains of application, even under the 512x512 resolution, this section provides preliminary results on two more experiments that were conducted. The first experiment involves low dose CT image enhancement and segmentation of kidneys. The second is a neuroimaging translation for cross-modality generation of T2-flair and T1-inverse from the T1 input image.
Close attention is given in the literature to LDCT imaging because of the use of a lower dose of radiation and its wider availability for being affordable with faster scanning time making it suitable for screening, diagnosis and follow up visits \cite{ldct-nature-lung}. While segmentation is an intrinsic problem in imaging, likewise in LDCT, the image enhancement aspect that could lead to better image quality is also a state-of-the-art issue which is being addressed in the literature \cite{ldct-nature-rec}.
In the second experiment, while \textit{Cross-modality generation} can serve as an auxiliary method in clinical diagnosis \cite{n2n_general}, it also has great potential for multimodal registration \cite{n2n_registration, n2n_arxiv,n2n_arxiv_seg_treg}, segmentation \cite{n2n_registration, n2n_arxiv_seg_treg, n2n_segmentation}, super-resolution \cite{n2n_superresolution} and structural information improvement \cite{n2n_enhancement} (e.g. MRI 3T to 7T). For this second experiment, an MRI T1 volume of the brain is translated into an MRI T1-inverse and a T2-flair, in slice by slice fashion.
Figure \ref{fig:generalization} shows the qualitative and quantitative results (as seen in one slice) of single task and \textit{multitask pix2pix} using leave one subject out (LOSO) as a test evaluation scheme. The network is trained using all of the 2D axial slices from the training subjects and is tested on all of the 2D axial slices of the test subjects not seen in the training phase. For these studies, the hyper-parameters and networks are the same as the aforementioned for CXR analysis and are not optimized for these applications. For both applications, the multitask method outperforms the single task methods while having only half of the network weights. It is worth mentioning that there is a significant difference in segmentation task for LDCT application and T2-flair task for the neuroimaging application. The magnification area is shown to emphasize the differences. These results clearly show the potential benefits of using \textit{image-to-images} translation to other domains of application involving different imaging modalities even when the proposed model as used for these additional applications are implemented on the model solely based on CXR images.
\section{Introduction}\label{Introduction}
Chest radiography, also called chest X-ray or CXR, is one of the most affordable and widely used medical imaging modality, which has significant practical implications in the diagnosis and screening of the thorax region, the organs and bone structure within it. Over 2 billion procedures per year are performed using this technology for the purpose of medical diagnosis of a variety of diseases, such as pneumonia, tuberculosis, lung cancer, and heart failure. Moreover, chest radiography remains the most prevalent screening test for pulmonary disorders \cite{general-1,general-nature, general-2, pulmonary_abnormalities_1,pulmonary_abnormalities_3}.
However, due to overlapping organs, low resolution and subtle anatomical shape and size variations, interpreting CXR images accurately remains challenging and requires a well-trained staff. On the other hand, managing a large number of CXR images each day results in high workloads for the radiography staff, yielding a tedious process fraught with setbacks and errors in diagnosis and in assessing adequately treatment follow up. It is reported that almost 90 percent of mistakes in pulmonary tumor diagnosis could be associated with the CXR screening of images \cite{chest-512}.
Therefore, many efforts have been devoted to the development of automated computer-based methods to improve accuracy in diagnosis and in finding any abnormalities that may otherwise be left undetected \cite{general-diamant2017chest, general-qin2018computer, general-tmi,pulmonary_abnormalities_2}.
There is considerable literature focusing on CXR image analysis. Among the more recent work on chest radiography, a team from Stanford \cite{general-Chest-Andrew} proposed a convolutional neural network called \textit{CheXNeXt} as a deep learning algorithm to concurrently detect the presence of 14 different pathologies such as pneumonia, fibrosis, emphysema, and nodules in frontal-view chest radiographs, among others. The \textit{CheXNeXt} algorithm achieved promising results in identifying abnormalities at a performance level that was comparable with the diagnostic accuracy of radiologist practitioners.
Four different deep learning based methods are investigated in \cite{net_radiology} and compared with radiology experts.
In another study \cite{chest-512}, Gozes and Greespan proposed a method to improve the contrast of lung structures in CXR images leading to better accuracy in nodule(s) detection; while Wang and Chia proposed a deep neural network they named \textit{ChestNet} \cite{general-wang2018chestnet} for enhanced diagnosis of diseases on chest radiography.
Moreover, an interesting multi-resolution convolutional network for chest X-ray radiograph lung nodule detection is proposed by Li et al. \cite{nodule_deep}.
The aim of this work is to construct a multitask learning framework using deep learning techniques that address in an effective way the two challenging tasks of organ segmentation and bone suppression simultaneously.
Organ segmentation is used for computer-aided detection and diagnosis while bone suppression enhances the visibility of the disease effects, e.g. nodules particularly on the lung region. Baltruscha et al. show that both of these tasks can improve the diagnosis rate by machine \cite{BSE-4}.
In order to incorporate the multitask objective, a new \textit{image-to-imag\underline{es}} translation machine is proposed based on the well-known \textit{pix2pix} network which is known for its promising results in the domain of \textit{image-to-image} translation and segmentation \cite{pix2pix}.
For this reason, the \textit{pix2pix} network and its implementation are modified to fit the need for multitasking (\textit{pix2pix MT}). As far as the authors know, the proposed network is a first attempt at expanding the application of \textit{image-to-image} network to \textit{image-to-imag\underline{es}} with the ability to generate more than one desired output at once.
Furthermore, the dilated convolution technique \cite{dilation} is employed in specific layers of the generator, which is shown to improve further the results.
Hence, this design is referred to as the \textit{pix2pix MTdG} model given the implication of \textit{multitask pix2pix} and the inclusion of the dilation property in the generator.
More specifically, by feeding a CXR image to the \textit{pix2pix MTdG} network, the proposed model will generate automatically two output images simultaneously, which are the image of the bone suppressed lungs and the image containing the segmentation masks of the heart and lungs.
Experimental results show that the \textit{MTdG} network yields promising and comparable results to the state-of-the-art methods that deal with these tasks individually. Results which are evaluated with several metrics and verified using 5-fold cross-validation along with the significance test exhibit promising outputs for both tasks.
Moreover, the conceptual design of the model can be generalized to extend to other applications. To confirm this assertion, two different applications involving 1) neuroimage modality conversion for cross-modality generation of T2-flair and T1-inverse from the T1 input image, and 2) low-dose Computed Tomography (LDCT) for image enhancement and segmentation of kidneys are addressed to showcase the merits of the proposed \textit{image-to-images} network in such critical applications where two desired outputs can be obtained simultaneously with improved accuracy and with better system efficiency.
The contributions of this work can be summarized as follows: 1) Design and implementation of a multitask network that for the first time augments the traditional \textit{image-to-image} translation model to an \textit{image-to-images} translation model while improving both accuracy and computational efficiency of the \textit{multitask pix2pix} model.
2) An architectural design of the model that allows for two critical tasks of CXR image analysis, namely bone suppression and organ segmentation, to be performed simultaneously through the use of efficient network parameters verified and augmented by an ablation study.
3) A generalized construct making the model more amenable to other application domains, where the results of two more medical applications, brain MRI cross-modality generation and low-dose CT image enhancement and segmentation, are provided in support of this assertion.
4) All the software code for the different variations of this work are publicly shared online including \textit{multitask pix2pix} for the research community to replicate such work or extend its research potential to other applications.
The rest of this paper is organized as follows: Section \ref{sec-related} reports the literature review of related work. Section \ref{sec-method} explains the proposed method, specifies the material used in conducting this study and the evaluation strategy that was used to assess its merits. The experimental results are presented and discussed in section \ref{sec-results} which include a section on method generalization and future work. Finally, the conclusion section \ref{sec-conclusion} provides a retrospective of what was accomplished through this proposed novel approach.
\section{Materials \& Methods}
\label{sec-method}
\begin{figure*}[]
\captionsetup{justification=centering}
\minipage{0.49\textwidth}
\includegraphics[width=\linewidth, trim={4.5cm 4.cm 5.5cm 3.5cm},clip]{Figures/drawing_gen.pdf}
\caption*{a) Generator.}
\endminipage\hfill
\minipage{0.49\textwidth}
\includegraphics[width=\linewidth, trim={4cm 3.5cm 5cm 3cm},clip]{Figures/drawing_disc.pdf}
\caption*{b) Discriminator.}
\endminipage\hfill
\caption{Architecture of the presented image to images translation, \textit{multitask pix2pix}. In this figure, $X, Y_1, Y_2,$ $\hat{Y}_1, \hat{Y}_1$ are the images of input CXR, targets of task 1 and task 2, output for task 1 and task 2, respectively. Notice that, all the images, input, output and target have three channels.}
\label{fig:Model}
\end{figure*}
\subsection{Methodology}
\subsubsection{Background}
In recent years, generative adversarial networks (GANs) and conditional generative adversarial networks (cGANs) have gained a lot of attention because of their superior performance in generation, segmentation, and translation empowered by an adversarial scheme \cite{cGAN}. The GAN architecture consists of two ‘adversarial’ models trained to work against each other: the generator aiming at generating an output and deceiving the discriminator and a discriminator component aiming at segregating the real output from the fake ones. In conditional mode (cGAN), both generator and discriminator are conditioned on ground truth labels or images. For example, in this study, the segmented organs and the bone suppressed images are the conditions and the generator is set up to generate this type of images.
Generators of GANs are intended to learn the mapping from a random noise vector $z$ to an output image $y$, \textit{i.e.,} $G: z \Rightarrow y$ while cGANs are conditioned by an observed image $x$ \textit{i.e.,} $G: \ \{x,z\} \Rightarrow y$. The generator $G$ would learn to produce outputs, which could not be distinguished as “fake” images by an adversarially trained discriminator, $D$.
The objective of a GAN and of a conditional GAN can be expressed through equations \eqref{eq:Gan} and \eqref{eq:cGan} respectively, where $\mathbb{E}$ is the \textit{Expectation} over the population.
Generator $G$ tries to minimize an objective function against an adversarial
$D$ which tries to maximize it, \textit{i.e.,} a \textit{minimax} game as $\hat{G}=arg \ min_G \ max_D \ \mathcal{L}_{GAN}$ and similarly $\hat{G}=arg \ min_G \ max_D \ \mathcal{L}_{cGAN}$.
\begin{gather}
\mathcal{L}_{GAN}(G,D) = \hspace{150pt} \nonumber \\
\mathbb{E} \ [log D(x,y)]
+
\mathbb{E} \ [log(1-D(G(x,z)))] \ \ \
\label{eq:Gan} \\
\mathcal{L}_{cGAN}(G,D) = \hspace{150pt} \nonumber\\
\mathbb{E} \ [log D(x,y)]
+
\mathbb{E} \ [log(1-D(x,G(x,z)))]
\label{eq:cGan}
\end{gather}
\subsubsection{Proposed Model}
A conditional generative adversarial network, called \textit{pix2pix}, is selected as an \textit{image-to-image} translation network to use and modify to meet the intended objectives of bone suppression and image segmentation \cite{pix2pix}. The \textit{pix2pix} is used in the proposed method, because 1) It is the main general \textit{image-to-image} translation method; 2) It shows promising prospects for accurate organ segmentation \cite{pix2pix,GanSeg2,GanSeg6,GanSeg9}; 3) It is intrinsically a collection of filters and would be reasonable to perform bone and rib suppression as an \textit{image-to-image} translation task \cite{BSE-1}.
The generator of \textit{pix2pix} contains an auto-encoding network of convolutional layers with skip connections. The discriminator is also a convolutional neural network (CNN) called \textit{PatchGAN} discriminator as introduced in \cite{pix2pix}, which attempts to determine whether each patch with size $n \times n$ in an image is real or fake, where $n$ can be much smaller than the full size of the image. Specifically, \textit{PatchGAN} discriminator is a CNN which produces a matrix of size $k \times k \times 1$ from an input tensor (or image) of size $N \times N \times *$ where $k=N/n$ and each element in the output matrix indicates the status of the corresponding receptive field on the input tensor (\textit{i.e.} a $k \times k$ \textit{PatchGAN} classifies $k \times k$ patches of the input image/tensor as real or fake). The input tensor for the \textit{PatchGAN} discriminator is a tensor built by concatenation of the input-target pair and the input-output pair for the discriminator to produce an estimation on how realistic they look \cite{pix2pix}.
Figure \ref{fig:Model} shows our model aiming to convert the input CXR image ($X$) into the desired output ($Y$), which is the concatenation of desired targets, $Y_1$ as the organs' segmentation masks and $Y_2$ as the bone suppressed CXR image, \textit{i.e.,} $Y: Y_1 \parallel Y_2$ where $\parallel$ shows concatenation in the channel axis.
The input and output tensors of the generator network are $X$ and $\hat{Y}: \hat{Y}_1 \parallel \hat{Y}_2$ where $\hat{Y}_2$ is the output image corresponding to the bone suppression task and $\hat{Y}_1$ is the output image corresponding to the organ segmentation task which include the masks for the heart (colored red), left lung (colored blue), right lung (colored green) and background (colored black). In fact, the generator creates a tensor with 6 channels, which are the concatenation of $\hat{Y}_1$ and $\hat{Y}_2$.
\begin{comment}
with tf.name_scope("discriminator_loss"):
# minimizing -tf.log will try to get inputs to 1
# predict_real => 1
# predict_fake => 0
discrim_loss = tf.reduce_mean(-(tf.log(predict_real + EPS) + tf.log(1 - predict_fake + EPS)))
discrim_loss_real=tf.reduce_mean(-(tf.log(predict_real + EPS)))
discrim_loss_fake=tf.reduce_mean(-(tf.log(1 - predict_fake + EPS)))
with tf.name_scope("generator_loss"):
# predict_fake => 1
# abs(targets - outputs) => 0
gen_loss_GAN = tf.reduce_mean(-tf.log(predict_fake + EPS))
gen_loss_L1 = tf.reduce_mean(tf.abs(targets - outputs))
\end{comment}
The discriminator network acts in a similar fashion to \textit{PatchGAN} in order to produce two output matrices, $D_R$ and $D_F$ corresponding to the real and fake input tensors ($R$ and $F$). The fake input tensor ($F$) is a concatenation of CXR input image and outputs ($F: X \parallel \hat{Y}_1 \parallel \hat{Y}_2$), and the real input tensor ($R$) is the concatenation of the CXR input image and targets ($R: X \parallel Y_1 \parallel Y_2$). If the discriminator is trained perfectly, it will create $D_R$ matrix of 1 values and $D_F$ matrix of 0 values. On the other hand, if the generator is successful in fooling the discriminator, $D_F$ would be a matrix of 1 values. The loss functions for training the generator and discriminator are as expressed in equations \eqref{eq:OurLoss-G} and \eqref{eq:OurLoss-D} where $\mid \ \mid_1$ defines the $L1$ distance or norm. In the training phase of the networks, for each batch feeding step: 1) The generator generates output images, 2) The discriminator looks at the real pair tensor ($R$) and the fake pair tensor ($F$) and produces an estimate on how realistic they look ($D_R$ and $D_F$), 3) The weights of the discriminator are then adjusted based on the $\mathcal{L}_D$, and 4) The generator's weights are then adjusted based on $\mathcal{L}_G$.
\begin{gather}
\mathcal{L}_G = \mathbb{E} [-log(D_F +\epsilon )] + \lambda \ \mathbb{E} [\mid Y - \hat{Y} \mid_1 ]
\label{eq:OurLoss-G}
\\
\mathcal{L}_D = \mathbb{E} [-( \ log(D_R + \epsilon) + log(1-D_F + \epsilon ) \ )]
\label{eq:OurLoss-D}
\end{gather}
Furthermore, in order to to produce more efficient receptive fields, dilated convolutions \cite{dilation} are utilized in some specific layers of the generator. The encoder of the generator consists of 8 layers and dilated convolutions are used with dilation rate 2 in layers 2 through 7 in the proposed structure of the \textit{MTdG} network. The effects of using and not using the dilated convolutional layers are contrasted in the results section.
\subsection{Data}
The \textit{Japanese Society of Radiological Technology (JSRT)} is the only publicly available database where both desired tasks are available and hence most suitable for training and evaluating the proposed model. This dataset consists of CXR images collected by JSRT \cite{jsrt} and is publicly available in \cite{jsrt-data}. Segmentation masks for lungs and the heart were created later by \cite{jsrt-segments} and are now available in \cite{jsrt-segments-data}. The JSRT dataset comprises 247 CXRs, including images with and without lung nodules. All images have a resolution of $2048 \times 2048$ in gray scale with a color depth of 12 bits.
While there is no publicly available dataset for bone and rib suppression based on DES, Juhász et al. developed a method for bone suppression \cite{bse-jsrt-alg}. Their results on the JSRT dataset have become publicly available in \cite{bse-jsrt-data}. This dataset is used to accomplish the second task of bone suppression as well.
As noted in CheXNeXt and through other investigations, the $512\times512$ resolution is sufficient for classifying lung-related diseases and for localizing nodule(s) \cite{general-Chest-Andrew, chest-512, chest-512-2, chest-512-3}. This resizing of images helped to significantly minimize the computational requirements and, as the results will prove, high accuracy is maintained. Moreover, we anticipate that the resizing of images to $512\times 512$ pixels could help in their effective use towards the development of new pre-processing methods for improving computer-aided diagnosis.
The image intensities are set up with an 8-bit grayscale resolution in the range from 0 to 255.
In order to train the machine learning and especially the deep learning networks, it is essential to have enough number of samples that cover the different variations \cite{Augmentation1, Augmentation2}. Therefore, the original images along with their corresponding masks and suppressed bone images were augmented by rotating them via 10 and -5 degrees, along with translations of (30, 10) and (-20,-10) pixels in reference to the (x,y) coordinates.
Through this process, the size of the dataset has been increased by 5 times, to a total number of 1,235 images along with their corresponding ground-truths for the two tasks.
Furthermore, in order to assess the effectiveness of multitask \textit{image-to-images} translation in terms of its generalization to other applications, two additional experiments are included in the \textit{Generalization and Future Work} section. The dataset used for the low-dose CT (LDCT) experiment is from \textit{"Multi-Atlas labeling beyond the cranial vault"} challenge containing CT scans and corresponding segmentation labels of 13 abdominal organs of 50 subjects \cite{LDCT_data}. In order to simulate the LDCT scans from CT scans, the method based on additive Poisson noise on sonograms of CT scans is used \cite{LDCT_simulation}.
The dataset for the neuroimaging experiment is \textit{MRBrainS18} challenge dataset containing multi-modal MRI brain images (T1, T1-inverse and T2-flair) with segmentation labels of gray matter, white matter, cerebrospinal fluid, and other structures on 3T MRI scans of the brain of seven subjects \cite{MRBrainS18}. Augmentation with the same configuration is also exploited for the LDCT and neuroimaging experiments.
\subsection{Implementation \& Evaluation}
The proposed model has been implemented and modified to comply with the multitasking scenario based on the publicly available \textit{pix2pix} code \cite{pix2pixCode}.
For validation purposes, both the extended code that supports all these different variations and the video showing this process at work are made available through the Internet to the research community.
The intensity channel of the input CXR image is replicated to support the CNN 3-channel RGB input data expectations.
The size of the input/output images and kernel (or filter) are $512\times512\times3$ and $4\times4$, respectively and 5-fold cross-validation on subjects are considered as default. To make a fair comparison with the results of the current state-of-art-techniques, the resolution of $256\times256$ and 3-fold cross-validation has been considered in this study as well.
The network has been implemented using Python and the Tensorflow library. All computations for training the network have been performed on a system equipped by NVIDIA GPU Quadro M6000 with 24 GB memory.
The parameter $\lambda$ in equation \ref{eq:OurLoss-G} is set to 10 and the learning rate of the Adam optimizer is set to 0.0002. The training is stopped when the L1 loss reaches almost $0.005$.
Because of a limited number of subjects, leaving-one-subject-out cross-validation scheme is used for the LDCT and neuroimaging experiments.
For comparative assessment of the results obtained with the different methods, the average, standard deviation, box plots including median, percentiles, and outliers along with t-test for statistical significance are considered regarding the different metrics.
\section{Related Work}
\label{sec-related}
\subsection{Task 1: Organ Segmentation}
Organ segmentation is one of the most difficult tasks in medical imaging due in large part to the elusive thresholding process and the ubiquitous presence of noise \cite{malek-1,malek-2}, but remains an essential task for delineating the anatomical structures of organs and hence for detecting abnormalities such as enlarged heart or collapsed lungs. It should be noted that when performing segmentation in chest radiography, one would also need to contend with the different shape variations in organs due to age, gender, disease and other health-related issues.
Mansoor et al. \cite{seg-mansoor2015segmentation} presented a comprehensive survey discussing the challenges and accomplishments of the different segmentation methods for lungs which are reported in the literature. By considering the CXR as the imaging modality, several deep learning models based on fully convolutional networks have also been investigated. For instance, a network called \textit{InvertedNet} is proposed to segment the heart, left and right clavicles, and lungs \cite{seg-InvertedNet}.
The well-established \textit{U-Net} architecture has been utilized for segmenting the chest region yielding promising results \cite{seg-u-net-scia, seg-u-net-brazil, seg-u-net-islam2018towards}.
A model called structure correcting adversarial network (\textit{SCAN}) was proposed as a generative adversarial network that uses convolutional layers for heart and lungs segmentation \cite{seg-SCAN}. Another method which incorporates two networks is proposed by \cite{seg_2019} with one network used for the initial segmentation process and the second for fine-tuning and correcting the initial results.
Moreover, traditional feature extraction methods are widely used for CXR imaging applications \cite{seg-mansoor2015segmentation}. In \cite{seg-Ibragimov}, Ibragimov et al. proposed an approach for lung segmentation and landmark detection based on Haar-like features, a random forest classifier, and spatial relationships among landmarks. A hierarchical lung field segmentation based on the joint shape and appearance sparse learning is proposed in \cite{seg_sparse}, and an atlas-based method is presented in \cite{seg_atlas}.
\subsection{Task 2: Bone and Rib Suppression}
In chest X-ray images, the bone structure in the chest area is usually visible, which makes it hard for a radiologist to examine thoroughly the organs and assess any effects of a given disease accurately. Organs' visibility is effective for pulmonary abnormalities screening and detection \cite{pulmonary_abnormalities_3}. Consequently, bone and rib suppression is an essential pre-processing step in order to suppress the appearance of bones in the chest X-ray images.
One way to tackle the aforementioned problem is to utilize dual-energy subtraction (DES) imaging \cite{DES}. The DES imaging technique captures two or three radiography scans with two or three different energy level of X-ray exposures.
The captured images either highlight the soft tissues or bones based on the energy levels. Thus, the suppressed bone image will be estimated by combining the acquired images which include both the soft tissue-selective images and the bone-selective images \cite{DES-2}. Although effective in delineating the bone structure in the chest area, the DES imaging process has a number of shortcomings, among them is its more invasive nature due to the higher radiation dose and the presence of artifacts introduced in the acquisition process due to the effect of heartbeats.
Because of these aforementioned reasons, suppressing the bones in CXR images via traditional image processing techniques is considered safer and is shown to be more effective at overcoming the main challenges faced in CXR images.
Along this line of research, a cascaded convolutional neural networks architecture (called \textit{CamsNet}) \cite{BSE-2} is proposed to predict the bone gradients in CXR images progressively with the ability of suppressing the bones as a consequence of these determined gradients. Convolutional neural filters are exploited in \cite{bse_CT} and are shown to be effective for bone suppression as well.
Another recent method is developed by Chen \textit{et al.} \cite{bse-tmi} which anatomically compensates for the ribs and clavicles by specific multiple massive-training artificial neural networks (\textit{MTANNs}) combined with total variation (TV) minimization smoothing along with a post-processing by histogram-matching.
In another study, Gusarev et al. proposed two deep learning architectures that perform bone suppression and create a soft tissue image. Considering bones as a noise level that is affecting these chest images \cite{BSE-3}, they tried to minimize the presence of this noise (i.e., bone) while still preserving the sharpness of the image for the eventual organ segmentation. In \cite{malek-1}, many of the noise suppressing methods reviewed shared the objective of removing as much of the noise as possible while preserving most of the relevant details in the image.
Another bone suppression method, based on deep adversarial networks and 2D Haar wavelet decomposition, has been proposed in \cite{BSE-1}. Their method was mostly based on the theory of \textit{pix2pix} network \cite{pix2pix}, a well-known conditional generative adversarial network. The \textit{pix2pix} network is also used as the cornerstone of our proposed multitask model, which will be described in section \ref{sec-method}.
Bone suppression is also used as a pre-processing step, where bone suppressed CXR images are then feed as input images to algorithms such as \textit{CheXNet} in order to enhance the segmentation process and improve as a consequence the results of the machine (automated) diagnosis \cite{bse-jsrt-usedin-BSE}.
The impact of bone suppression on machine diagnosis using deep learning networks have been thoroughly investigated and detailed in \cite{BSE-4} and \cite{bse-jsrt-usedin-deep}. There are also some commercially available computer-aided detection (CAD) systems such as \textit{Phillips} \cite{commercial_3}, \textit{ClearRead} \cite{commercial_1} and \textit{Caresteam} \cite{commercial_2}.
\subsection{Joint Tasks via Multitask Learning}
In multitask learning, multiple tasks are solved at the same time by exploiting commonalities and differences across tasks. In comparison to training separate single task models, the multitask scheme can result in the following improvements \cite{MTL-Survey}:
1) Improvement in results: Most often, coupling tasks makes the overall system achieve better results with respect to the desired accuracy. For example, in \cite{HyperFace}, a multitask learning approach based on deep convolutional networks is proposed for facial landmark detection with the auxiliary tasks of head pose estimation, gender classification and facial visibility, yielding more accurate results for each of the tasks.
2) Improvement in learning efficiency: Efficiency of learning could include other important implementation aspects such as the number of required parameters, memory or storage requirements, computational time and training convergence rate. Obviously, fewer but optimal parameters and lower memory requirements are desirable in deploying such algorithms on conventional devices such as mobile phones and PCs \cite{hardware}.
Aside from the potential improvement in learning efficiency in the interplay between tasks, and the need for only half the weights required of the multitask model in contrast to the two tasks run separately, other benefits acquired through the intrinsic functions of the multitasking model do not have straightforward reasons. For instance, an additional function can be used to act as a regularization mechanism in other machine learning problems and push the algorithm to find a solution on a smaller area of representations at the intersection of all tasks. Also, the feature selection and filter values can be reassessed and made more sufficient for addressing the nature of the inputs to the model. The overall motivation here is to be able to perform the two tasks of bone suppression and organ segmentation jointly via one deep network with the ability for improving both system efficiency and accuracy in the results.
\subsection{Image-to-Image translation: Applications and methods}
In general, the \textit{image-to-image} translation (I2I) is the process of translating an input image $X$ to a corresponding output image $Y$, and this correspondence could mean different things for the different context of the application at hand. Such I2I techniques could involve translations such as low-resolution $\Leftrightarrow$ high resolution, blurry $\Leftrightarrow$ sharp, thermal or grayscale $\Leftrightarrow$ color, synthetic $\Leftrightarrow$ real, low-dose rate (LDR) $\Leftrightarrow$ high-dose rate (HDR), noisy $\Leftrightarrow$ clean, image $\Leftrightarrow$ painting, day $\Leftrightarrow$ night, summer $\Leftrightarrow$ winter, bad weather $\Leftrightarrow$ good, foggy $\Leftrightarrow$ clear, semantic labeling (segmentation) $\Leftrightarrow$ realistic photo, aerial $\Leftrightarrow$ map, edges and sketch $\Leftrightarrow$ photo and so on, in which symbol $\Leftrightarrow$ shows the bidirectionality between desired task and context \cite{pix2pix, I2I-review-med}. I2I techniques are also used in medical imaging for segmentation, denoising, super-resolution, modality conversion, CT and MRI reconstruction, among others \cite{I2I-review-med}.
I2I has been studied for decades, and different approaches are reported on the basis of filtering, optimization, dictionary learning, deep learning, and more recently generative adversarial network (GAN). While deep learning methods omit the hand-crafted features and GAN methods omit the hand-crafted objective functions, they both remain the most promising methods in data science. GAN-based I2I research in computer vision has yielded different learning models, with a myriad of applications and promising outcomes. In general, there are two categories of methods and applications based on the relation between input and output images: \textit{1) Unsupervised/Unpaired/Unaligned/Unregistered} such as style changing, photo to painting, hair/face and color-changing, weather changing, and \textit{2) Supervised/Paired/Aligned/Registered} such as supervised segmentation and labeling, denoising and super-resolution.
The unpaired category is not relevant in the proposed application of CXR image analysis because the problem at hand is supervised with paired and aligned input/output images. Until now, \textit{pix2pix}\cite{pix2pix}, \textit{CRN}\cite{I2I-CRN}, \textit{BicycleGAN}\cite{I2I-BicycleGAN}, \textit{SIMS}\cite{I2I-SIMS}, \textit{SPADE}\cite{I2I-SPADE} and \textit{pix2pixHD}\cite{I2I-pix2pixHD} remain the most important methods for the paired category. While \textit{pix2pix} and \textit{BicycleGAN} are dealing with a family of applications, others are just considering semantic labels to realistic photo translation. \textit{BicycleGAN} is an \textit{image-to-image} translation with potentially multiple outputs. For example, \textit{BicycleGAN} is able to analyze and translate a given night image to synthesized day images with different types of lighting, sky, and clouds. Each different possibility is generated by feeding a random noise sampled from a known distribution (e.g., a standard normal distribution) along with the input image. Therefore, \textit{pix2pix} is the only general method relevant to our problem, with bone suppression and organ segmentation being a paired/supervised/aligned/registered problem with no randomized output possibilities.
|
\section{Introduction}
Prediction is often considered as one of the key and fundamental components of intelligence \cite{bubic}. Visual prediction often deciphers much useful information about the environment in an information-rich but high dimensional format which presents both opportunities and challenges \cite{ebert}. A successful mechanism capable of predicting future video frames would have many applications in the automation industry. For robotics, the path planning problem in an unknown dynamic environment with moving obstacles still largely remains an unsolved challenge. A key component of the problem that remains unsolved is to make predictions about the motion of the obstacles. To illustrate the complexity of the problem, consider an Unmanned Aerial Vehicle (UAV) flying at a speed of 5m/s or higher (which translates to a speed of $\geq$ 18km/hr) in a cluttered environment, such as a forest trail in a high wind condition. In this scenario, the perception of the camera not only depends upon the dynamics of the objects present in the scene but also on the control actions taken by the UAV. The interaction between the robot's state and action with the dynamics of the scene renders the motion prediction problem almost impossible to solve using conventional vision-based methods such as visual servoing \cite{visp}, \cite{vkumar}. While making pixel level prediction on the motion on robotic agents is a challenging task, designing a motion planner based on raw predicted image frames is an even harder task. However, it is imperative to have a mechanism to predict the motion of the other objects present in the environment in order to solve the motion planning problem for autonomous agents operating in a rapidly changing environment.
Recent works (\cite{villegas}, \cite{xu}) on motion prediction delved into forecasting human motion but these models use very deep architectures that ultimately renders them computationally expensive. Given that human motion is much slower than automated vehicles, predicting higher speed motion will have a much higher level of difficulty and computation cost. Even with the recent advancements in mobile graphics units such as NVIDIA Jetson boards, implementation of deep architectures to solve path planning problems on small mobile agents still remains a challenge.
Designing a light-weight motion prediction framework is only the first step in addressing the challenge. Once the prediction network is designed we need to devise a mechanism to transfer raw predicted image frames into control commands for the robot. Recent advancements in deep reinforcement learning (DRL) \cite{mnih}, \cite{levine} have shown us ways to convert raw sensory inputs into meaningful control commands. While a few model-free learning algorithms have out-performed human operators \cite{mnih}, these frameworks were designed for very limited simulated environments of video games. Learning tasks in the real world present a wide range of challenges as the environment becomes dynamic with sparsely available reward feedback while the agent can only access a partial state of the world. Finally, we need to design the entire framework in such a manner that learning can be enabled without the supervision of human operators.
In this work, we present a novel light-weight framework that can forecast the trajectory of an object moving in the robot's work-space. The proposed Predicting Robot Motion Network (PROM-Net) can easily be trained on raw video data without supervision. Once trained, this network can be implemented on an autonomous mobile agent. The network generates the visual prediction of the surrounding environment from the first-person perspective of the robotic agent. In order to train and test the network we also created our own data set, using two LEGO Mindstorms under 4 different scenarios. To the best of our knowledge, this is the first data set of its kind where the motion of a robotic agent is captured from the first-person view of another robot. We also discuss how these predicted frames can be used to design a model-based reinforcement learning algorithm that would be able to translate the raw predicted image frames into a meaningful reward function to optimize the trajectories of the control policies.
The paper is organized as follows: We first discuss the existing literature on video prediction networks and model predictive controllers (MPC) using raw image frames as input and introduce the Predicting Robot Motion Network (PROM-Net) model. Then we discuss the virtual experimental setup we created in the OpenAI-Gym framework and give a detailed description of the real robotic dataset that was created for testing the performance of PROM-Net. A detailed analysis of the performance of PROM-Net is presented next, followed by a discussion on the future scope of the work.
\section{Related Work}
The problem of video frame prediction \cite{mathieu}, \cite{vondrick2}, \cite{villegas}, \cite{xu} has gained considerable popularity in the computer vision community in recent years. However, video data comes with the issue of larger dimensionality with complex spatio-temporal dynamics in raw pixel values which makes the pixel label frame prediction task very challenging \cite{finn}. While Convolutional Neural Networks (CNN) have proven to be very successful at learning features from static images, \cite{imagenet}, \cite{resnet}, the idea of Convolutional Long Short Term Memory (LSTM) networks that were designed specifically to capture the spatial and temporal dependencies in video data was proposed by \cite{convlstm}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig11.png}
\caption{ Visual motion planning framework}
\label{ICAPS_fig1}
\end{figure}
The paper \cite{oh}, designed an action conditional encoder-decoder network predicting future frames for Atari games. The work in \cite{mathieu} employed a new adversarial loss function for additional regularization and sharper frame prediction. The paper \cite{vondrick2} designed a multi-scale feedforward architecture combined with an adversarial objective to generate a foreground-background mask to create realistic looking video sequences. The work in \cite{casas} presented a framework that predicts the intention of autonomous cars from 3D point clouds and HD maps. The paper \cite{walker} proposed a framework that generated a coarse hallucination of the possible future events from a wide angle view. In \cite{xu}, a framework that balances between reconstruction and adversarial loss for the predicted frames is designed. However, most of the current state of the art video prediction models often require a high-end GPU enabled system to train and test the networks which is not often a feasible option for robotic applications.
\begin{figure*}[h]
\centering
\includegraphics[width=.98\textwidth, height=4cm]{convlstm1.png}
\caption{Schematic architecture of the PROM- Network}
\label{ICAPS_fig2}
\end{figure*}
While considerable progress has been made in DRL \cite{mnih},\cite{mnih2} that learns meaningful skills directly from high dimensional raw sensory data (especially images), most of these are restricted to simulated applications of computer games. Only a few works like \cite{finn2}, \cite{ebert} talks about the application of a model based RL algorithm for robotic manipulation tasks using visual foresight. To the best of our knowledge, there is no existing work that addresses the problem of an end to end motion planning for autonomous mobile agents using visual prediction from a first person (robot) perspective.
\section{Our Approach}
Figure \ref{ICAPS_fig1} shows the schematic representation of the visual prediction based motion planning framework. It shows that the MPC algorithm takes the frames generated by the prediction network as input. The prediction network also generates a reduced dimensional state representation of the world from the raw image inputs for the model-based controller. This is possible as the architecture of the prediction network is based on the encoder-decoder network philosophy. We started with designing the prediction network with a goal of predicting the next $N$ image frames from the past $N$ number of frames. Furthermore, we aim to design a very light-weight network that can be easily implemented on a GPU denied environment. We have successfully designed a motion prediction network that can approximate frames up to 10 time stamps ahead of time. In the following sub-section we present a detailed description of our proposed network.
\subsection{Predicting Robot Motion Network (PROM-Net)}
The architecture of the model is shown in figure \ref{ICAPS_fig2}. This model roughly follows the encoder-decoder philosophy of autoencoder networks. The encoder network is built using 8 2D convolutional filters of size $3\times 3$. The outputs are down-sampled using a maxpooling layer of stride 2. A second 16 channel convolution layer with filter size $5\times5$ and stride 2, further maps the input to a 3 dimensional tensor of size $(16\times16\times16)$. These spatial feature tensors are then passed through two consecutive Convolutional LSTM layer having kernel size of $(5\times5)$ and mapped into a 32 channel feature space of size $(8\times 8)$. The mathematical model of Convolutional LSTM is described in \cite{convlstm}. The two ConvLSTM layers capture the spatio-temporal correlations present in the sequence of image frames and pass them to the next decoder layer for inference.
The decoding network consists of 3 Convolutional LSTM layers. After each ConvLSTM layer we have a deconvolution or transpose convolution layer that upsamples the size of each feature channel and downsamples the total number of feature channels. For example, after the first deconvolution operation, the $(32\times8\times8)$ feature tensor is mapped into a $(16\times16\times16)$ feature tensor. We have used skip connections at intermediate layers to recover from the lossy convolution operations (shown with dotted lines in fig. \ref{ICAPS_fig2}).
We apply batch normalization operation after each Convolutional LSTM layer. We also upsample the number of feature channels each time we apply a downsampling operation on the 2 dimensional spatial feature space. This kind of convention has been followed in designing various previous networks such as \cite{unet},\cite{finn}. All the convolutional filters use the ReLU activation function. The entire network is trained using the RMSProp algorithm that minimizes the mean square loss.
PROM-Net has about 6 million trainable parameters and once trained the network weighs only about 5 Megabytes.
\begin{figure*}
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.8\textwidth]{fig2.png}
\caption{A ROS-Gazebo based virtual experimental environment has been set up in the OpenAI-gym framework.}
\label{ICAPS_fig3a}
\end{subfigure}%
\hspace{1in}
\begin{subfigure}{.3\textwidth}
\centering
\includegraphics[width=0.75\linewidth,keepaspectratio]{fig_wicv19.png}
\caption{LEGO Mindstorms with a \\ GoPro Hero 5 Black camera}
\label{ICAPS_fig3b}
\end{subfigure}
\caption{}
\end{figure*}%
\section{Virtual Experimental Setup}
\label{sec4}
For initial analysis of the networks, we set up a ROS-Gazebo based virtual experimental environment in the OpenAI-Gym framework for robotics \cite{openai} to obtain the training and test data for the network. Figure \ref{ICAPS_fig3a} shows a snapshot of the same. The virtual setup has two turtlebots, Tb1 and Tb2. During the data collection phase, Tb1 remains stationary while tracking and recording the movement of Tb2 using a monocular camera. Tb2 moves in front of Tb1, from point A to point B using a Proportional Integral Derivative (PID) controller that corrects the positional and angular error of the robot. We introduce variation in the PID parameters so that no two trajectories are the same even when Tb2 is moving towards the same goal point. This introduces variance in the local neighbourhood of the trajectories even when the goal point is same. We also recorded video of Tb2 moving towards 4 different target points. These 4 different target points are $(1,0.8)$, $(1.5, -0.8)$, $(2-0.8)$ and $(0.5, -0.5)$ where the position of Tb1 is taken as the origin of the inertial frame. The recorded image frames are converted to gray scale images before being used to train the networks. Altogether we collected about 80 different trajectories (20 trajectories in the local neighbourhood of each of the 4 goal points).
\begin{figure*}
\centering
\includegraphics[width=0.92\linewidth, keepaspectratio]{fig_arm.png}
\caption{The 4 environments from left- Atrium (daylight), Atrium (artificial light), Pavement and Airstrip, respectively}
\label{ICAPS_fig4}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=.75\linewidth, keepaspectratio]{PSNR_wicv19}
\caption{PSNR comparison plot between 2 videos of equal length from two different environment (Atrium daytime and Pavement).}
\label{ICAPS_fig5}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=.91\textwidth,keepaspectratio]{result1.png}
\caption{Qualitative comparison on the performance of Fully Connected LSTM network and PROM network on simulated data set. The first row represents the ground truth, second and third row show the estimates by PROM network and the fully connected LSTM network, respectively for time stamps $10$, $15$, $20$, $25$, $30$, $35$ and $40$.}
\label{ICAPS_fig6}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=.7\textwidth,keepaspectratio]{fig5.png}
\caption{Qualitative analysis on the performance of PROM-Net trained on ARM data set. The first and thrid row from top represents the ground truth, second and fourth row show the estimates generated by PROM-Network for time stamps $20$, $30$, $40$, $50$ and $60$. The first 2 rows represent data under artificial lighting conditions and the last 2 rows are from the outdoor environment.}
\label{ICAPS_fig7}
\end{figure*}
\section{Real Robot Motion Data-set}
To evaluate the performance of PROM-Net with the real-life data, we created our own Actual Robot Motion (ARM) data-set, using two LEGO Mindstorms under different lighting conditions in 4 different environmental settings- indoor (Atrium) daylight, indoor artificial light, and outdoor (pavement) daylight and outdoor (Airstrip) sunlight. To the best of our knowledge, this is the first of its kind data set where the motion of a mobile robot is captured from the first-person view of another robot. In this section, we present details on the real robot motion data set that was collected from the first person perspective of a LEGO Mindstorms robot (see figure \ref{ICAPS_fig3b}) observing another Mindstorms moving in its field of view.
We recorded the videos using a GoPro Hero 5 Black camera at 30 fps, with a resolution of 720$\times$1280. We later down-sampled it to a resolution of 320$\times$240. For the initial phase of data gathering, we mounted the camera on a LEGO Mindstorms robot to observe the environment and kept it in a stationary state.
In future, we will add motion to the recording platform to add more versatility to the data which would closely resemble the practical cases seen in a robot path planning problems. The average speed of the moving agent was kept at about 0.665 km/hr (approximately 11 m/s). The recorded videos do not contain any labelled data as they are meant for unsupervised learning algorithms.
We recorded about 1.5 hours of robot motion of the other LEGO-bot along various trajectories consisting of approximately 120K frames without excluding any particular segments. The GoPro camera offers digital stabilization. We used the narrow-angle shot setting during the recordings. The wide-angle lens of this particular camera produces a significant amount of fish-eye effect for any object moving relatively close to the camera. A wide angle lens will be used in future when we incorporate recordings of unmanned aerial vehicles (UAV) into the data set. Unlike the autonomous ground vehicles, the high speed operation of UAVs (Average speed of 5m/s) demands long range visual data for effective path planning. Below we describe the various scenarios of the recorded data. The videos are segregated in a 3:1 ratio between training and test data. The data set can be accessed at \url{https://sites.google.com/view/meenakshis/dataset}
\begin{table}
\centering
\begin{tabular}{ |p{1.7cm}||p{1.25cm}|p{1cm}|p{1.2cm}|p{1cm}| }
\hline
Types of trajectory & Atrium Daylight & Atrium Night & Pavement & Airstrip\\
\hline
St. Line & 4 &4& 4&4\\
Arc & 4 &4& 4&4\\
Incline L-R & 4 &4& 4&4\\
Incline R-L & 4 &4& 4&4\\
\hline
\end{tabular}\caption{Arrangement of no. of videos in the Data set}
\label{table1}
\end{table}
\section{Scenarios}
Among the two Lego Mindstorms bots, one was remote-controlled via a Bluetooth module to execute four different types of trajectories- Straight path, Inclined path (left to right and right to left) and Arc; each with three different depths (distance from the mounted-camera) in all of the four different environmental settings (Figure \ref{ICAPS_fig4}). The logistics of the recorded videos in each of the environment for each the 4 different types of trajectories are given in table \ref{table1}
This was done to incorporate diversity (Figure \ref{ICAPS_fig5} shows the distribution of PSNR between 2 videos of equal length from 2 different environment) in the data set and to facilitate efficient training of deep networks. Each trajectory in a particular setting was repeated twice for redundancy in a single video.
\subsection{Environment 1, 2: Atrium (Daylight and Artificial Light at Night)}
This setting was used for collecting two different sets of recordings. One was during daytime using natural light (Figure \ref{ICAPS_fig4},$1^{st}$ frame from left ) and the other at night using multiple light sources of white halogens (Figure \ref{ICAPS_fig4}, $2^{nd}$ frame from left). The smooth floor of the atrium results in consistent motion without any jerks. However, the artificially lit night-scene introduces complexity due to multiple shadow formations (different intensities) of the same object.
\subsection{Environment 3: Pavement}
This was recorded in a sun-lit scene with nearby tree canopy (shadows in the backdrop, Figure \ref{ICAPS_fig4}, $3^{rd}$ frame from left) . The ground (lock-tiles) adds intrinsic inconsistency in motion and is bright-colored.
\subsection{Environment 4: Airstrip}
This was recorded in twilight (resulting in, elongated shadows) and the motion was the most jittery here due to coarseness of the asphalt (Figure \ref{ICAPS_fig4}, $4^{th}$ frame from left). Also, there are tiny insects moving in the background which adds a naturally dynamic clutter.
\section{Results and Analysis}
\label{results}
Initially, we trained the network in the simulated environment. In order to maintain uniformity during training, we used the RMSProp optimizer with a batch size of 64 and learning rate 0.001 for all the networks. Our initial investigation with the simulated data set revealed that even though fully connected LSTM networks (\cite{srivastava}) generates moderately accurate predictions for trajectories in the close neighborhood of the ones it has been trained on, it fails to generalize the robot motion when the test trajectories are unlike any training data it has seen before. The same can be inferred from figure \ref{ICAPS_fig6}. Figure \ref{ICAPS_fig6} also shows that PROM-Net can efficiently approximate the future robot motion for unforeseen test scenarios.
For each of the test cases, we have given the network 10 image frames as input and the network predicted the next 10 frames in future. The reconstructed frames by PROM-Net on the real robot data set for two different environments (indoor with artificial lights and outdoor with sunlight) are shown in figure \ref{ICAPS_fig7}. Even though for this paper we have only presented results with grey-scale images, our network can be very easily modified for RGB inputs.
We have given the variation in structural similarity index (SSIM) for all the 10 predicted frames on the real world data set in figure \ref{ICAPS_fig8}. To compare the performance of the proposed prediction network with a FC LSTM network we have given the Peak Signal to Noise Ratio (PSNR) plots for both simulated and real data set in figure \ref{ICAPS_fig9}. It can be easily inferred from the plots that PROM-Net performs well with both the simulated and real data sets.
From figure \ref{ICAPS_fig7}, we can infer that the blurriness in the predicted frames arises due to the regression losses in convolution layers. As our application is focused on solving path planning problems for robotic agents, we can easily accommodate minor reconstruction loss in the predicted frames. Our intentions are to infer the future direction of motion for the moving objects and PROM-Net has proven to be very effective for that purpose.
\section{Conclusion}
\label{sec7}
We presented a novel light-weight unsupervised learning framework for robot motion prediction problems. A new robot motion data set has been introduced to train and test deep architectures for motion and path planning problems with small scale mobile agents. While the present model is capable of predicting robot motions under stationary condition, a more robust framework is needed in order to estimate future frames where the motion of the robot influences the data collected by the camera sensor. We are already working towards building such models. In our future work, we plan on designing and testing a vision based MPC on a mobile agent for motion planning in a cluttered dynamic environment. We envisage that reward function would penalize the controller for actions that would move the agent closer to any obstacle and reward it when the area of the obstacle reduces in the predicted frames. We also plan on extending our robot motion data-set with multiple mobile agents (human and robots) moving in the robot workspace.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth]{fig2a.png}
\caption{SSIM distribution between predicted frames and the ground truth for the 10 time stamps on the ARM data-set.}
\label{ICAPS_fig8}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.4\textwidth, keepaspectratio]{psnr1}
\caption{PSNR plots for PROM-Net with Real data (red line), Simulated Data (blue line) and Fully Connected LSTM Network with simulated data (green line).}
\label{ICAPS_fig9}
\end{figure}
\input{icaps2019.bbl}
\end{document}
|
\section{Introduction}
After the discovery of a Higgs boson, the Standard Model (SM) has been completed in a sense that
the existence of all the predicted particles was confirmed experimentally.
In the SM, the minimal form with an isospin doublet scalar field is introduced as the Higgs sector.
Although the discovered Higgs boson shows similar properties to that of the SM under the current experimental and theoretical
uncertainties, the possibility that the Higgs sector takes a non-minimal form is not excluded at all, and
its exploration is one of the central interests of current and future high-energy physics.
If the Higgs sector is extended from the minimal form, it has a different structure which can be classified
by the number of scalar fields, their representations, symmetries and strength of coupling constants.
There should be strong connection between these properties of extended Higgs sectors and the physics behind
the electroweak (EW) symmetry breaking.
Furthermore, a non-minimal structure of Higgs sectors could solve the problems which cannot be explained in the SM
such as neutrino masses, dark matter and baryon asymmetry of the Universe.
Therefore, the Higgs sector is one of the most important probes of new physics beyond the SM.
The most important property of a non-minimal Higgs sector is the prediction of multiple scalar bosons.
Thus, discovery of additional scalar bosons will be a clear evidence of extended Higgs sectors.
At the LHC, direct searches for a new particle are being performed continuously.
On the other hand, existence of such additional scalar fields generally affects the couplings of the SM-like Higgs boson
to the particles in the SM
by the effect of mixing and the quantum correction, yielding deviations from the predictions of the SM.
Therefore, detecting such deviations by precision measurements is also a strong signature for models with extended Higgs sectors.
Moreover, from the pattern of deviations in various Higgs boson couplings we can indirectly
distinguish the shape of the Higgs sector and further determine new physics models~\cite{Kanemura:2014bqa}.
At the LHC, direct searches for additional Higgs bosons have been performed
via bosonic channels~\cite{Khachatryan:2016yec,Aaboud:2017yyg,Sirunyan:2017nrt,Aaboud:2017itg,Aaboud:2017fgj,Khachatryan:2016cfx,Aaboud:2017cxo} and
fermionic channels~\cite{Khachatryan:2015tra,Khachatryan:2016qkc,Aaboud:2017sjh}.
From the non-observation of the signature, parameters of each extended Higgs sector such as masses and couplings are constrained.
In addition, some of the Higgs boson couplings have been measured at the LHC~Run-I~\cite{Khachatryan:2016vau}
and Run-II~\cite{Sirunyan:2018koj,ATLAS-CONF-2018-031} experiments.
Although the current data from these measurements are consistent with the SM predictions, the experimental and theoretical
uncertainties are not small yet; e.g., about 20\% for the Higgs boson couplings to weak bosons and
typically 20--50\% for the Yukawa couplings of the third generation at the 95\% confidence level.
Above experimental uncertainties can be much reduced at future colliders;
e.g., at the High-Luminosity LHC (HL-LHC)~\cite{CMS:2013xfa,ATLAS:2013hta},
the International Linear Collider (ILC)~\cite{Baer:2013cma,Fujii:2017vwa,Asai:2017pwp},
the Future Circular Collider (FCC)~\cite{Gomez-Ceballos:2013zzn},
the Circular Electron Positron Collider (CEPC)~\cite{CEPC-SPPCStudyGroup:2015csa},
the Compact LInear Collider (CLIC)~\cite{CLIC:2016zwp} and so on.
For example, at the ILC with the collision energy of 250 GeV and the luminosity of 2 ab$^{-1}$,
some of the Higgs boson couplings are expected to be measured with ${\cal O}(1\%)$ level or better~\cite{Fujii:2017vwa}.
In order to extract information on new physics from
these precision measurements at future experiments, accurate calculations with higher-order corrections
are required in models with various extended Higgs sectors.
Radiative corrections to the SM-like Higgs boson vertices have been studied in various Higgs sectors such as,
for example, a model with a real isospin singlet Higgs field (HSM)~\cite{Bojarski:2015kra,Kanemura:2015fra,Kanemura:2016lkz,He:2016sqr},
two Higgs doublet models (THDMs)~\cite{Arhrib:2003ph,Kanemura:2004mg,LopezVal:2010vk,Kanemura:2014dja,Kanemura:2015mxa,Krause:2016oke},
the inert doublet model (IDM)~\cite{Arhrib:2015hoa,Kanemura:2016sos}, the Higgs triplet model~\cite{Aoki:2012yt,Aoki:2012jj}
and the Georgi-Machacek model~\cite{Chiang:2017vvo,Chiang:2018xpl}.
In order to see differences of the prediction among these models,
it is quite important to calculate the renormalized Higgs boson vertices with a consistent and systematic way.
Recently, we have published a numerical program {\tt H-COUP} (version 1.0)~\cite{Kanemura:2017gbi}
to compute a set of SM-like Higgs boson vertices at one-loop level
in various extended Higgs models; i.e., the HSM, four types of THDMs and the IDM.
Other numerical tools are also available
to calculate Higgs boson decays with radiative corrections in models with extended Higgs sectors; e.g.,
{\tt Prophecy4f}~\cite{Altenkamp:2017kxk,Altenkamp:2018bcs} and {\tt 2HDECAY}~\cite{Krause:2018wmo,Krause:2019oar}.
In this paper, we present a complete set of the decay rates of the SM-like Higgs boson ($h$)
including the $h \to WW^*$ mode at the full next-to-leading order (NLO) in QCD
and EW as well as scalar interactions in the HSM, four types of THDMs and the IDM\footnote{The decay rates of the
one-loop induced processes, i.e, $h \to \gamma\gamma$, $Z\gamma$ and $gg$, are calculated at NLO in QCD and at leading order in EW. }.
Some important results have already been highlighted in our letter paper~\cite{Kanemura:2018yai}, in which
the calculation of the partial decay rate of the $h \to WW^*$ mode was not yet included.
We then calculate the branching ratios of the SM-like Higgs boson at NLO in these models.
One-loop calculations are consistently performed based on the on-shell renormalization scheme
for EW parameters~\cite{Hollik:1988ii,Kniehl:1993ay,Kanemura:2017wtm} and the modified minimal subtraction ($\overline{\text{MS}}$) scheme
for QCD corrections~\cite{tHooft:1973mfk} in these models with the extended Higgs sectors.
We discuss the amount of the NLO corrections of the Higgs boson decay rates in each model with detailed descriptions of the computation.
We show various correlations of the deviation in the branching ratios from the SM predictions
under constraints of the perturbative unitarity~\cite{Lee:1977eg}, vacuum stability~\cite{Deshpande:1977rw},
conditions to avoid wrong vacua~\cite{Frere:1983ag} and experimental constraints.
Finally, we investigate the possibility to discriminate the extended Higgs sectors from the difference of the prediction among the models.
This paper is organized as follows.
In Sec.~\ref{sec:model}, we introduce the HSM, the THDMs and the IDM.
In Sec.~\ref{sec:decay}, we present analytic formulae for the decay rates of the Higgs boson at NLO.
EW corrections in each decay mode are discussed in detail.
In Sec.~\ref{sec:numerical}, we show numerical results of the total width, the branching ratios and
correlations of the branching ratios.
Conclusions are given in Sec.~\ref{sec:conc}.
In Appendix, explicit formulae for the NLO calculations are presented.
\section{Models with non-minimal Higgs sectors\label{sec:model}}
In this section, we define the HSM, the THDM and the IDM in order.
Before moving on to the discussion on each extended Higgs sector,
let us briefly explain constraints on a parameter space, as their basic notion are common to models with the extended Higgs sectors.
First of all, the size of dimensionless parameters in the potential can be constrained
by imposing the perturbative unitarity bound which has originally
been introduced in Ref.~\cite{Lee:1977eg} to obtain the upper limit on the Higgs boson mass in the SM.
Using the equivalence theorem~\cite{Cornwall:1974km},
this bound requires that the magnitude of partial wave amplitudes for the elastic scatterings
of 2 body to 2 body scalar boson processes, including the Nambu-Goldstone (NG) bosons, does not exceed a certain value.
Each eigenvalue of the $s$-wave amplitude denoted as $a_0^i$ is required to satisfy:
\begin{align}
|a_0^i| \leq \xi,
\end{align}
where $\xi=1$~\cite{Lee:1977eg} or 1/2~\cite{Gunion:1989we}.
We here take $\xi = 1/2$.
We note that each $a_0^i$ only depends on the scalar quartic couplings as
only scalar contact interactions contribute to the scattering process at the high-energy limit.
Next, the vacuum stability bound provides an independent constraint on scalar quartic couplings.
It requires that the Higgs potential is bounded from below in any direction with large field values.
This condition is schematically written by
\begin{align}
V^{(4)} >0, \label{eq:vs}
\end{align}
where $V^{(4)}$ represents quartic terms of the Higgs potential.
Although in the SM this condition is trivially satisfied by taking the scalar quartic coupling to be positive,
in models with non-minimal Higgs sectors it is given by a set of inequalities in terms of scalar quartic couplings~\cite{Deshpande:1977rw}.
Furthermore, in extended Higgs sectors, wrong local
vacua can generally appear in addition to the true vacuum giving the correct value of the Fermi constant $G_F$.
Thus, we have to avoid parameter regions which realize the depth of such wrong vacua to be deeper than that of the true one \footnote{There is still a possibility of a meta-stable electroweak vacuum even if such situations are realized. {By taking into account such a possibility a constraint from wrong vacua is more moderate. For example, see Ref.~\cite{Branchina:2018qlf}.} }.
The condition to avoid the wrong vacua can be written by combinations of dimensionful and dimensionless parameters in the potential~\cite{Frere:1983ag},
so that it can provide an independent constraint from the above two constraints.
Apart from these theoretical constraints, we need to take into account bounds from experimental data.
At the LEP/SLC experiments, various EW observables have been precisely measured such as the masses and widths of the weak gauge bosons.
These precise measurements can be used to constrain new physics effects which can indirectly be entered into the self-energy diagrams for weak gauge bosons.
Such indirect effect, so called oblique corrections, is conveniently parameterized by the $S$, $T$ and $U$ parameters introduced by Peskin and Takeuchi~\cite{Peskin:1990zt,Peskin:1991sw}, which
are expressed in terms of two point functions of the weak bosons.
From the global fit of EW parameters~\cite{Baak:2012kk}, new physics effects on the $S$ and $T$ parameters under $U = 0$ are constrained by
\begin{align}
S = 0.05\pm 0.09, \quad T = 0.08\pm 0.07,
\end{align}
with the correlation factor of $+0.91$ and the reference values of the masses of SM Higgs boson and top quark being $m_h^{\text{ref}} = 126$ GeV and $m_t^{\text{ref}} = 173$ GeV, respectively.
Flavor experiments also provide important constraints on a parameter space of extended Higgs models, particularly models with a multi-doublet structure.
We will discuss these constraints in more detail in Sec.~\ref{sec:2hdm} about THDMs.
As mentioned in Introduction, additional scalars have been directly searched at the LHC~\cite{Khachatryan:2016yec,Aaboud:2017yyg,Sirunyan:2017nrt,Aaboud:2017itg,Aaboud:2017fgj,Khachatryan:2016cfx,Aaboud:2017cxo,Khachatryan:2015tra,Khachatryan:2016qkc,Aaboud:2017sjh},
and some of the constraints are interpreted in THDMs.
Moreover, the Higgs coupling measurements also constrain the mixing parameters in THDMs~\cite{Khachatryan:2016vau,Sirunyan:2018koj,ATLAS-CONF-2018-031}
by using the so-called $\kappa$ framework~\cite{Heinemeyer:2013tqa}.
We note that the $\kappa$ framework is constructed by the leading order (LO) relation,
and hence the interpretation of such constraints at higher-order level might not be straightforward.
The application of these constraints to each extended Higgs sector will be discussed in the following subsections.
\subsection{Higgs singlet model \label{sec:hsm}}
The Higgs sector of the HSM is composed of an isospin doublet scalar field $\Phi$ with the hypercharge $Y = 1/2$
and a real singlet field $S$ with $Y = 0$.
These scalar fields are parameterized as
\begin{align}
\Phi=\left(\begin{array}{c}
G^+\\
\frac{1}{\sqrt{2}}(v+\phi+iG^0)
\end{array}\right),\quad
S=v_S^{} + s, \label{hsm_f}
\end{align}
where $v$ is the vacuum expectation value (VEV) of the doublet filed which is related to the Fermi constant $G_F$ by
$v = (\sqrt{2}G_F)^{-1/2}\simeq 246$ GeV, while $v_S^{}$ is the VEV
of the singlet field.
Because the singlet field does not contribute to the EW symmetry breaking,
the component fields $G^\pm$ and $G^0$ in the doublet field correspond to the NG bosons which are absorbed into the weak bosons.
The most general Higgs potential is written as
\begin{align}
V_{\text{HSM}} =&\, m_\Phi^2|\Phi|^2+\lambda_\Phi |\Phi|^4
+\mu_{\Phi S}^{}|\Phi|^2 S+ \lambda_{\Phi S} |\Phi|^2 S^2
+t_S^{}S +m^2_SS^2+ \mu_SS^3+ \lambda_SS^4,\label{Eq:HSM_pot}
\end{align}
where all the parameters are real. By the reparameterization of the Higgs potential,
we can take any value of $v_S^{}$ without changing physical results~\cite{Chen:2014ask}.
Hence, we take $v_S^{} = 0$ throughout the paper.
In the HSM, there are two physical neutral Higgs bosons. Their mass eigenstates are defined as
\begin{align}
\begin{pmatrix}
s \\
\phi
\end{pmatrix} = R(\alpha)
\begin{pmatrix}
H \\
h
\end{pmatrix}~~\text{with}~~R(\theta) =
\begin{pmatrix}
c_\theta & -s_ \theta \\
s_\theta & c_\theta
\end{pmatrix}, \label{mat_r}
\end{align}
where $\alpha$ is the mixing angle, and we define the domain of $\alpha$ by
$-\pi/2 \leq \alpha \leq \pi/2$.
Throughout the paper, we use the shorthand notation for the trigonometric function as
$s_\theta \equiv \sin\theta$ and $c_\theta \equiv \cos\theta$.
We identify $h$ as the discovered Higgs boson with a mass of 125 GeV.
After solving the tadpole conditions, the squared masses of these Higgs bosons are expressed as
\begin{align}
&m_H^2=M_{11}^2c^2_\alpha +M_{22}^2s^2_\alpha +M_{12}^2s_{2\alpha} , \label{mbh}\\
&m_h^2=M_{11}^2s^2_\alpha +M_{22}^2c^2_\alpha -M_{12}^2s_{2\alpha} , \label{mh}\\
&\tan 2\alpha=\frac{2M_{12}^2}{M_{11}^2-M_{22}^2}, \label{tan2a}
\end{align}
where $M^2_{ij}$ ($i,j=1,2$) are the squared mass matrix elements in the basis of ($s,\phi$).
Each element is given by
\begin{align}
M^2_{11}= M^2+ v^2\lambda_{\Phi S} ,\quad M^2_{22}=2v^2\lambda_\Phi ,\quad M^2_{12}=v\mu_{\Phi S}^{}, \label{mij}
\end{align}
with $M^2 = 2m_S^2$.
The seven parameters in the potential are then expressed by the following five input parameters
\begin{align}
&m_H^{},~~M^2,~~\mu_S,~~\lambda_S,~~c_\alpha, \label{input:hsm}
\end{align}
{and the two parameters $m_h$ and $v$ that are fixed by experiments. }
Let us briefly discuss the other relevant parts of the Lagrangian in the HSM.
The kinetic term is given by
\begin{align}
{\cal L}_{\text{kin}}^{\text{HSM}} = |D_\mu \Phi|^2 +\frac{1}{2}(\partial_\mu S)^2,
\end{align}
where $D_\mu$ is the covariant derivative for the Higgs doublet. Because the singlet field $S$ does not have the gauge interaction,
the additional Higgs boson $H$ couples to weak bosons only through the $\phi$ component of $H$.
Thus, the gauge-gauge-scalar type interactions are given as
\begin{align}
{\cal L}_{\text{kin}}^{\text{HSM}} \supset gm_W^{}(c_\alpha W_\mu^+ W^{-\mu}h + s_\alpha W_\mu^+ W^{-\mu}H) + \frac{g_Z^{}m_Z^{}}{2}(c_\alpha Z_\mu Z^\mu h + s_\alpha Z_\mu Z^\mu H), \label{eq:hsm-v}
\end{align}
where $g$ is the weak gauge coupling and $g_Z^{} = g/\cos\theta_W$ with $\theta_W$ being the weak mixing angle.
The Yukawa interactions are written by the same form as those in the SM:
\begin{align}
{\cal L}_Y^{\text{HSM}} = -Y_u\bar{Q}_L \tilde{\Phi}\,u_R^{} - Y_d\bar{Q}_L \Phi \, d_R^{} - Y_e\bar{L}_L \Phi \, e_R^{} + \text{h.c.},
\end{align}
where $\tilde{\Phi} = i\sigma_2 \Phi^*$, and we do not show the flavor indices here.
In the above equation, $Q_L$, $L_L$, $u_R$, $d_R$ and $e_R$ are respectively the left-handed quark doublet, lepton doublet, right-handed up-type quark singlet, down-type quark singlet and charged lepton singlet.
The singlet field does not couple to fermions, so that the interaction terms for $h$ and $H$ are extracted as
\begin{align}
{\cal L}_Y^{\text{HSM}} \supset -\frac{m_f}{v}(c_\alpha \bar{f}f h + s_\alpha \bar{f}f H). \label{eq:hsm-f}
\end{align}
As it is seen in Eqs.~(\ref{eq:hsm-v}) and (\ref{eq:hsm-f}), the SM-like Higgs boson $h$ couplings are universally suppressed by the factor of $c_\alpha$ as compared to the corresponding SM values.
As we already mentioned at the beginning of this section,
the parameters in the potential can be constrained by the unitarity, the vacuum stability and the condition to avoid wrong vacua.
For the unitarity bound, there are four independent eigenvalues given in Refs.~\cite{Cynolter:2004cq,Kanemura:2016lkz}.
In this paper, we use the expression for the eigenvalues given in Ref.~\cite{Kanemura:2016lkz},
where the same notation of the potential parameters as that in this paper is applied.
The necessary and sufficient condition to satisfy the vacuum stability is given by~\cite{Pruna:2013bma}
\begin{align}
\lambda_\Phi > 0, \quad \lambda_S > 0, \quad 2\sqrt{\lambda_\Phi \lambda_S} + \lambda_{\Phi S} > 0.
\end{align}
For the condition to avoid these wrong vacua is found in {Ref.~\cite{Espinosa:2011ax,Chen:2014ask,Lewis:2017dme}}.
We use the expression given in Ref.~\cite{Kanemura:2016lkz}.
In the HSM, one-loop corrected two point functions for weak bosons are found in Ref.~\cite{Lopez-Val:2014jva}.
Imposing the bound from the $S$ and $T$ parameters, we can obtain the upper limit on $m_H^{}$ depending on the value of $c_\alpha$.
Constraints on the mass of the additional Higgs boson and the mixing angle from the LHC data have been studied in Refs.~\cite{Robens:2015gla,Robens:2016xkb,Blasi:2017zel,Gu:2017ckc}.
\subsection{Two Higgs doublet model \label{sec:2hdm}}
The Higgs sector of the THDM is composed of two isospin doublet scalar fields $\Phi_1$ and $\Phi_2$ with $Y = 1/2$.
These doublets are parameterized as
\begin{align}
\Phi_i = \left(\begin{array}{c}
w_i^+\\
\frac{1}{\sqrt{2}}(v_i + h_i + iz_i)
\end{array}\right),\hspace{3mm}(i=1,2), \label{eq:doublets}
\end{align}
where $v_1$ and $v_2$ are the VEVs of two doublets with $v = \sqrt{v_1^2 + v_2^2}$, and their ratio is expressed by $\tan\beta = v_2/v_1$.
Having two doublet fields with the same quantum charges causes
dangerous flavor changing neutral currents (FCNCs) at tree level, because both doublets couple to each type of fermions.
In order to avoid such FCNCs, we impose a discrete $Z_2$ symmetry, where two doublets transform as
$\Phi_1 \to +\Phi_1$ and $\Phi_2 \to -\Phi_2$.
One can introduce the soft breaking term of the $Z_2$ symmetry in the potential, retaining the
good property of the flavor sector. In the following, we discuss the THDM with the softly-broken $Z_2$
symmetry and the CP-conservation.
The most general Higgs potential is given by
\begin{align}
V_{\text{THDM}}=& m_1^2|\Phi_1|^2+m_2^2|\Phi_2|^2-m_3^2(\Phi_1^\dagger \Phi_2 +\text{h.c.})\notag\\
& +\frac{\lambda_1}{2}|\Phi_1|^4+\frac{\lambda_2}{2}|\Phi_2|^4+\lambda_3|\Phi_1|^2|\Phi_2|^2+\lambda_4|\Phi_1^\dagger\Phi_2|^2
+\frac{\lambda_5}{2}\left[(\Phi_1^\dagger\Phi_2)^2+\text{h.c.}\right], \label{eq:pot-thdm}
\end{align}
where the $m_3^2$ term softly breaks the $Z_2$ symmetry.
The $m_3^2$ and $\lambda_5$ parameters are taken to be real as we consider the CP-conserving case.
The scalar mass eigenstates can then be defined as follows:
\begin{align}
\left(\begin{array}{c}
w_1^\pm\\
w_2^\pm
\end{array}\right)&=R(\beta)
\left(\begin{array}{c}
G^\pm\\
H^\pm
\end{array}\right),\quad
\left(\begin{array}{c}
z_1\\
z_2
\end{array}\right)
=R(\beta)\left(\begin{array}{c}
G^0\\
A
\end{array}\right),\quad
\left(\begin{array}{c}
h_1\\
h_2
\end{array}\right)=R(\alpha)
\left(\begin{array}{c}
H\\
h
\end{array}\right), \label{mixing}
\end{align}
where $H^\pm$ and $A$ are the charged and CP-odd Higgs bosons, respectively,
while $H$ and $h$ are the
CP-even Higgs bosons. Similar to the HSM case, we identify $h$ {as} the discovered Higgs boson {with a} mass of 125 GeV.
{We define the domain of $\alpha$ and $\beta$ to be $-\pi/2 \leq \alpha \leq 0$ and $0 \leq \beta \leq \pi/2$, respectively,
so that the viable range for $\beta-\alpha$ is expressed as $0\leq \beta-\alpha \leq \pi$.}
After solving two tadpole conditions for $h_1$ and $h_2$, squared masses of the charged and CP-odd
Higgs bosons are given by
\begin{align}
m_{H^\pm}^2 = M^2-\frac{v^2}{2}(\lambda_4+\lambda_5),\quad m_A^2&=M^2-v^2\lambda_5, \label{mass1}
\end{align}
where $M^2=m_3^2/(s_\beta c_\beta)$ which describes the soft-breaking scale of the $Z_2$ symmetry.
For {the} two CP-even Higgs bosons, {the} squared mass matrix elements $M_{ij}^2$
in the Higgs basis~\cite{Davidson:2005cw} are given by
\begin{align}
M_{11}^2&=v^2(\lambda_1c^4_\beta+\lambda_2s^4_\beta +2\lambda_{345}s^2_{\beta}c^2_{\beta}) , \notag \\
M_{22}^2&=M^2 + \frac{v^2}{4}s^2_{2\beta}(\lambda_1+\lambda_2-2\lambda_{345}), \label{m22} \\
M_{12}^2&=\frac{v^2}{2} s_{2\beta}( -\lambda_1c^2_\beta + \lambda_2s^2_\beta + \lambda_{345}c_{2\beta}), \notag
\end{align}
with $\lambda_{345}\equiv \lambda_3+\lambda_4+\lambda_5$.
The squared masses of {the} two CP-even Higgs bosons and the mixing angle $\beta-\alpha$ are
expressed in terms of the matrix elements $M_{ij}^2$ as
\begin{align}
&m_H^2= M_{11}^2c^2_{\beta-\alpha} + M_{22}^2s^2_{\beta-\alpha} - M_{12}^2s_{2(\beta-\alpha)} , \label{111} \\
&m_h^2 = M_{11}^2s^2_{\beta-\alpha} + M_{22}^2c^2_{\beta-\alpha} + M_{12}^2s_{2(\beta-\alpha)}, \label{222} \\
&\tan 2(\beta-\alpha)= -\frac{2M_{12}^2}{M_{11}^2-M_{22}^2}. \label{333}
\end{align}
The eight parameters in the potential are then expressed by the following six input parameters
\begin{align}
m_H^{},~~m_A^{},~~m_{H^\pm}^{},~~M^2,~~\tan\beta,~~s_{\beta-\alpha}, \label{input:thdm}
\end{align}
and {the} two parameters $m_h$ and $v$ {that are fixed} by experiments.
In addition to these parameters, there is a degree of freedom of the sign of $c_{\beta-\alpha}$.
Let us discuss the other relevant parts of the Lagrangian.
The kinetic term for the Higgs doublets are written as
\begin{align}
{\cal L}_{\text{kin}}^{\text{THDM}} = |D_\mu \Phi_1|^2 + |D_\mu \Phi_2|^2.
\end{align}
In the mass eigenbasis of the Higgs bosons, the gauge-gauge-scalar type interaction terms are extracted as
\begin{align}
{\cal L}_{\text{kin}}^{\text{THDM}} \supset gm_W^{}(s_{\beta-\alpha} W_\mu^+ W^{-\mu}h + c_{\beta-\alpha} W_\mu^+ W^{-\mu}H) + \frac{g_Z^{}m_Z^{}}{2}(s_{\beta-\alpha} Z_\mu Z^\mu h + c_{\beta-\alpha} Z_\mu Z^\mu H).
\end{align}
\begin{table}[t]
\begin{center}
{\renewcommand\arraystretch{1.2}
\begin{tabular}{l|ccccccc|ccc}\hline\hline
&$\Phi_1$&$\Phi_2$&$Q_L$&$L_L$&$u_R$&$d_R$&$e_R$&$\zeta_u$ &$\zeta_d$&$\zeta_e$ \\\hline
Type-I &$+$&
$-$&$+$&$+$&
$-$&$-$&$-$&$\cot\beta$&$\cot\beta$&$\cot\beta$ \\\hline
Type-II&$+$&
$-$&$+$&$+$&
$-$
&$+$&$+$& $\cot\beta$&$-\tan\beta$&$-\tan\beta$ \\\hline
Type-X (lepton specific)&$+$&
$-$&$+$&$+$&
$-$
&$-$&$+$&$\cot\beta$&$\cot\beta$&$-\tan\beta$ \\\hline
Type-Y (flipped) &$+$&
$-$&$+$&$+$&
$-$
&$+$&$-$& $\cot\beta$&$-\tan\beta$&$\cot\beta$ \\\hline\hline
\end{tabular}}
\caption{$Z_2$ charge assignments in four types of Yukawa interactions,
and the $\zeta_f$ ($f=u,d,e$) factors appearing in Eq.~(\ref{eq:yuk-thdm}).
}
\label{tab:z2}
\end{center}
\end{table}
The most general Yukawa interactions under the $Z_2$ symmetry are written as
\begin{align}
{\cal L}_Y^{\text{THDM}}
&= -Y_u\bar{Q}_L \tilde{\Phi}_i \,u_R^{} - Y_d\bar{Q}_L \Phi_j \, d_R^{} - Y_e\bar{L}_L \Phi_k \, e_R^{} + \text{h.c.},
\end{align}
where the subscripts $i$, $j$ and $k$ are 1 or 2.
These indices are fixed when we determine the $Z_2$ charge for right-handed fermions.
As in Table~\ref{tab:z2}, there are four independent types of Yukawa interactions depending on the assignment of the
$Z_2$ charge~\cite{Barger:1989fj,Grossman:1994jb,Aoki:2009ha}.
The interaction terms for the physical Higgs bosons are then extracted as
\begin{align}
{\cal L}_Y^{\text{THDM}} \supset
&-\sum_{f=u,d,e}\frac{m_f}{v}\left[ (s_{\beta-\alpha} + \zeta_f c_{\beta-\alpha})\bar{f}fh
+ (c_{\beta-\alpha} - \zeta_f s_{\beta-\alpha}){\overline f}fH-2iI_f \zeta_f\bar{f}\gamma_5fA\right)\notag\\
&-\frac{\sqrt{2}}{v}\left[V_{ud}\bar{u}
\left(m_d\zeta_d\,P_R-m_u\zeta_uP_L\right)d\,H^+
+m_e\zeta_e\bar{\nu}P_Re^{}H^+ +\text{h.c.}\right], \label{eq:yuk-thdm}
\end{align}
with $I_f = 1/2\,(-1/2)$ for $f = u\,(d,e)$ and $V_{ud}$ is the Cabibbo-Kobayashi-Maskawa matrix element.
Similar to the HSM, parameters in the THDMs can be constrained by both the theoretical and experimental constraints.
For the unitarity bound, there are 12 independent eigenvalues of the
$s$-wave amplitude matrix~\cite{Kanemura:1993hm,Akeroyd:2000wc,Ginzburg:2005dt,Kanemura:2015ska}.
We use the expression for the eigenvalues given in Ref.~\cite{Kanemura:2015mxa}.
The vacuum stability bound is sufficiently and necessarily satisfied by imposing the following {conditions}~\cite{Deshpande:1977rw, Sher:1988mj, Nie:1998yn, Kanemura:1999xf}
\begin{align}
\lambda_1 > 0,\quad
\lambda_2 > 0,\quad
\sqrt{\lambda_1\lambda_2} + \lambda_3 + \text{MIN}(0,\lambda_4+\lambda_5,\lambda_4 - \lambda_5) > 0.
\end{align}
In addition, the wrong vacua can be avoided by taking $M^2 \geq 0$~\cite{Branchina:2018qlf}. We thus only take the positive value of $M^2$ in the following discussion.
The expressions of the two point functions for the weak bosons in the THDM are found in Refs.~\cite{Toussaint:1978zm,Bertolini:1985ia,Peskin:2001rw,Grimus:2008nb,Kanemura:2011sj}.
Constraints on the parameters in THDMs from the LHC data have been discussed in Refs.~\cite{Bernon:2015qea,Dorsch:2016tab,Han:2017pfo,Arbey:2017gmh,Chang:2015goa,Blasi:2017zel,Gu:2017ckc}.
Differently from the HSM, constraints from flavor experiments are important to be taken into account in the THDM.
These bounds particularly provide the lower limit on the mass of {the} charged Higgs boson $m_{H^\pm}$ depending on the type of Yukawa interaction and $\tan\beta$.
For example, from the $B_s \to X_s\gamma$ data, $m_{H^\pm}$ hs to be greater than about 600 GeV at 95\% confidence level in the Type-II and Type-Y THDMs with $\tan\beta \gtrsim 2$, while
${\cal O}(100)$ GeV of $m_{H^\pm}$ is allowed in the Type-I and Type-X THDMs with $\tan\beta \gtrsim 2$~\cite{Misiak:2017bgg}.
Constraints on $m_{H^\pm}$ and $\tan\beta$ from various flavor observables are also shown in Ref.~\cite{Haller:2018nnx}
in four types of the THDMs.
\subsection{Inert doublet model}
The contents of the scalar sector in the IDM are the same as those in the THDM.
In the THDM, the $Z_2$ symmetry can be softly-broken by introducing the $m_3^2$ term in the potential,
while in the IDM it is assumed to be unbroken even after the EW symmetry breaking.
Thus, the potential is obtained from Eq.~(\ref{eq:pot-thdm}) with $m_3^2 = 0$.
In addition, the second Higgs doublet $\Phi_2$ is supposed not
to develop the nonzero VEV, otherwise the $Z_2$ symmetry is spontaneously broken.
In the IDM, {the} component scalar fields of $\Phi_1$ and $\Phi_2$ do not mix with each other due to the unbroken $Z_2$ symmetry.
Therefore, we can identify these scalar bosons $(w_2^\pm,z_2,h_2,h_1)$ defined in Eq.~(\ref{eq:doublets})
with the mass eigenstates $(H^\pm,A,H,h)$.
The lightest neutral inert scalar boson can be a candidate of dark matter as it cannot decay into a pair of SM particles.
The mass formulae for the scalar bosons are changed from those of the THDMs, not just because of the absence of the $m_3^2$ term, but
also the absence of the tadpole condition for $h_2$.
These are given as follows:
\begin{align}
m_h^2 &= \lambda_1 v^2,\\
m_{H}^2 &= M^2 + \frac{v^2}{2}(\lambda_3+\lambda_4+\lambda_5),\\
m_{A}^2 &= M^2 + \frac{v^2}{2}(\lambda_3+\lambda_4-\lambda_5), \\
m_{H^\pm}^2 &= M^2 + \frac{v^2}{2}\lambda_3,
\end{align}
where $M^2 = m_2^2$.
We then choose the following five {parameters to be free input parameters} of the IDM
\begin{align}
m_H^{},~~m_A^{},~~m_{H^\pm}^{},~~M^2,~~\lambda_2, \label{input:idm}
\end{align}
{and the two fixed} parameters $m_h$ and $v$.
The same conditions for the perturbative unitarity and vacuum stability in the THDM can be applied to
the IDM, because these bounds are given in terms of the scalar quartic couplings.
The condition to guarantee the inert vacuum with $(\langle \Phi_1^0 \rangle, \langle \Phi_2^0 \rangle = (v/\sqrt{2},0))$ is given by~\cite{Ginzburg:2010wa},
\begin{align}
\frac{m_1^2}{\sqrt{\lambda_1}} < \frac{M^2}{\sqrt{\lambda_2}} . \label{eq:wv_idm}
\end{align}
Since the tadpole condition makes $m_1^2$ negative, and the vacuum stability condition constraints $\lambda_1$ and $\lambda_2$ to be positive,
the condition given in Eq.~(\ref{eq:wv_idm}) is satisfied by taking $M^2 > 0$.
We refer {to} this condition as the one to avoid wrong vacua, according to the other two models discussed above.
For the constraints of the $S$ and $T$ parameters, we can use the same {expressions} as those in the THDM with $s_{\beta-\alpha} = 1$.
In the IDM, constraints on the masses of the additional Higgs bosons from collider experiments are relatively weak
since the additional scalars do not couple to SM fermions.
The constraints from the LEP and the LHC have been studied in Refs.~\cite{Lundstrom:2008ai,Gustafsson:2010zz} and Refs.~\cite{Belanger:2015kga,Belyaev:2016lok}, respectively.
Dark matter constraints from relic density and direct detection also limit the parameter space;
see, e.g., Refs.~\cite{Belyaev:2016lok,Ilnicka:2018def} for details.
\section{Decay rates of the SM-like Higgs boson at one loop\label{sec:decay}}
In this section, we discuss the decay rates of the SM-like Higgs boson; i.e.,
$h \to f\bar{f}$,
$h \to ZZ^* \to Zf\bar{f}$ and $h \to WW^* \to Wf\bar{f}'$ at NLO in EW and QCD.
The loop induced decay rates $h \to \gamma\gamma$, $h \to Z\gamma$ and $h \to gg$ are also discussed at NLO in QCD.
We outline our one-loop calculations.
For the computation of the EW corrections, we adopt the modified on-shell renormalization scheme which has been defined in Ref.~\cite{Kanemura:2017wtm}, while
for the QCD corrections we apply the $\overline{\text{MS}}$ scheme.
In the on-shell scheme, all the counterterms appearing in the decay rates of the $h \to f\bar{f}$ and $h \to VV^* \to Vf\bar{f}$ modes are determined in terms of the
one particle irreducible (1PI) diagrams for one- and two-point functions of Higgs bosons, gauge bosons and fermions by imposing a set of the renormalization conditions.
Adding these counterterms, one can obtain the ultra-violet (UV) finite one-loop corrected vertices.
The on-shell scheme is a physically quite natural renormalization scheme, and it is suitable to apply to EW observables as they include well defined scales such as the weak boson masses.
However, it has been known that the on-shell scheme introduces gauge dependent counterterms, particularly in some mixing parameters~\cite{Yamada:2001px}.
In extended Higgs sectors, a mixing between Higgs bosons can generally appear.
We thus apply the so-called pinch technique to remove the gauge dependence in the renormalized vertex functions to our computations~\cite{Bojarski:2015kra,Krause:2016oke,Kanemura:2017wtm}.
Apart from the UV divergences, there appear infrared (IR) divergences when we calculate virtual photon loop contributions.
Such IR divergences can exactly be cancelled by adding contributions of real photon emissions, where
the finite QED corrections are common to those in the SM.
The analytic expressions of QED corrections are known for $h\to f\bar f$~\cite{Kniehl:1991ze,Dabelstein:1991ky,Bardin:1990zj} and $h\to Zf\bar{f}$~\cite{Kniehl:1993ay}.
Thus, we simply switch off the photon-loop contributions, and use these analytic formulae in our computation as the QED correction part.
For $h\to Wf\bar{f}'$, on the other hand,
we cannot separate EW corrections into QED and weak corrections.
Therefore, by using the phase-space slicing method~\cite{Harris:2001sx},
we numerically evaluate both the virtual and real corrections to $h\to Wf\bar{f}'$, see Appendix~\ref{sec:real_photon} for details.
For QCD corrections, we use the similar technique to remove IR divergences coming from virtual gluon loop contributions.
In our renormalization calculation, we choose
the fine structure constant $\alpha_{\rm em}$, the Fermi constant $G_F$ and the $Z$ boson mass $m_Z$ as the input parameters for the EW parameters.
In this scheme, all the other EW parameters such as $v$, $m_W^{}$ and $s_W^{}$ are outputs.
Using the on-shell definition of the weak mixing angle; i.e., $s_W^2 = 1 - m_W^2/m_Z^2$~\cite{Sirlin:1980nh} and
the modified relation among the EW parameters:
\begin{align}
G_F =\frac{\pi \alpha_{\text{em}}}{\sqrt{2}s_W^2m_W^2}\frac{1}{1-\Delta r}=\frac{1}{\sqrt{2}v^2}\frac{1}{1-\Delta r}, \label{mod-gf}
\end{align}
we can calculate the renormalized squared W boson mass as
\begin{align}
(m_W^2)_{\rm reno} = \frac{m_Z^2}{2}\left[1+\sqrt{1-\frac{4\pi\alpha_{\rm em}}{\sqrt{2}G_Fm_Z^2(1-\Delta r)}} \right]. \label{rmw}
\end{align}
In Eq.~(\ref{mod-gf}), $\Delta r$ is calculated by~{\cite{Sirlin:1980nh}}
\begin{align}
\Delta r = \frac{\text{Re}\hat{\Pi}_{WW}(0)}{m_W^2} + \frac{\alpha_{\text{em}}}{4\pi s_W^2}\left(6 + \frac{7-4s_W^2}{2s_W^2}\log c_W^2 \right), \label{delr}
\end{align}
{where $\hat{\Pi}_{WW}$ is the renormalized $W$ boson two-point function and the second term corresponds to vertex corrections and box diagram contributions to the muon decay rate.}
Including these three EW parameters, we choose the following parameters as the SM inputs:
\begin{align}
\alpha_{\rm em},~~m_Z,~~G_F,~~\Delta\alpha_{\rm em},~~\alpha_s,~~m_t,~~m_b,~~m_c,~~m_\tau,~~m_h, \label{inputs}
\end{align}
where $\Delta \alpha_{\rm em}$ is the shift of the fine structure constant $\alpha_{\text{em}}$ from zero energy to $m_Z$.
We also input the parameters in the potential given in Eqs.~(\ref{input:hsm}), (\ref{input:thdm}) and (\ref{input:idm}) in the HSM, THDMs and IDM, respectively.
We note that the parameters $c_{\alpha}$ and $s_{\beta-\alpha}$ in the HSM and THDMs, respectively, do not physically mean the mixing angles for the CP-even Higgs bosons after the renormalization.
\subsection{Renormalized vertices\label{sec:ren-vert}}
Important ingredients for calculations of decay rates of the Higgs boson are renormalized Higgs boson vertices.
In our computations, the $hf\bar{f}$, $hV^\mu V^{\nu}$ $(V=W,Z)$ and $h {\cal V}^\mu {\cal V}^{\prime\nu}$ $({\cal V V'} = \gamma\gamma,~Z\gamma,~gg)$ vertices are relevant,
where the $h {\cal V}^\mu {\cal V}^{\prime\nu}$ vertices are one-loop induced.
Each of these vertices can be decomposed into several form factors depending on their Lorentz structure as written below.
The {\tt H-COUP} program (ver. 1.0)~\cite{Kanemura:2017gbi} provides numerical values of these renormalized form factors in the extended Higgs sectors.\footnote{
We are now preparing the next version of {\tt H-COUP} program (ver. 2.0)~\cite{future} providing numerical values of decay rates of $h$ at NLO which are calculated in this paper. }
We fully use {\tt H-COUP} in our numerical evaluation of the form factors.
The renormalized $hf\bar{f}$ vertices can be decomposed into the following form factors:
\begin{align}
\hat{\Gamma}_{h ff}(p_1^2,p_2^2,q^2)&=
\hat{\Gamma}_{h ff}^S+\gamma_5 \hat{\Gamma}_{h ff}^P+p_1\hspace{-3.5mm}/\hspace{2mm}\hat{\Gamma}_{h ff}^{V_1}
+p_2\hspace{-3.5mm}/\hspace{2mm}\hat{\Gamma}_{h ff}^{V_2}\notag\\
&\quad +p_1\hspace{-3.5mm}/\hspace{2mm}\gamma_5 \hat{\Gamma}_{h ff}^{A_1}
+p_2\hspace{-3.5mm}/\hspace{2mm}\gamma_5\hat{\Gamma}_{h ff}^{A_2}
+p_1\hspace{-3.5mm}/\hspace{2mm}p_2\hspace{-3.5mm}/\hspace{2mm}\hat{\Gamma}_{h ff}^{T}
+p_1\hspace{-3.5mm}/\hspace{2mm}p_2\hspace{-3.5mm}/\hspace{2mm}\gamma_5\hat{\Gamma}_{h ff}^{PT}, \label{eq:hff-form}
\end{align}
where $p_1^\mu$ $(p_2^\mu)$ is the incoming four-momentum of the fermion (anti-fermion),
and $q^\mu (= p_1^\mu + p_2^\mu)$ is the outgoing four-momentum of the Higgs boson.
For the case with on-shell fermions; i.e., $p_1^2 = p_2^2 = m_f^2$, the following relations hold:
\begin{align}
\hat{\Gamma}_{h ff}^{P} &= \hat{\Gamma}_{h ff}^{PT} = 0, \quad
\hat{\Gamma}_{h ff}^{V_1} = -\hat{\Gamma}_{h ff}^{V_2},\quad
\hat{\Gamma}_{h ff}^{A_1} = -\hat{\Gamma}_{h ff}^{A_2}.
\end{align}
These relations are used for the calculation of the Higgs boson decay into fermions discussed in Sec.~\ref{sec:hff}.
Next, the renormalized $hV^\mu V^\nu$ vertices are defined in terms of three renormalized form factors:
\begin{align}
\hat{\Gamma}_{hVV}^{\mu\nu}(p_1^2,p_2^2,q^2)&=g^{\mu\nu}\hat{\Gamma}_{h VV}^1 + \frac{p_1^\nu p_2^\mu}{m_V^2}\hat{\Gamma}_{h VV}^2 + i\epsilon^{\mu\nu\rho\sigma}\frac{p_{1\rho} p_{2\sigma}}{m_V^2}\hat{\Gamma}_{h VV}^3,
\end{align}
where $p_1^\mu$ and $p_2^\mu$ are incoming four-momenta of the weak bosons, and $q^\mu$ is the outgoing four-momentum of the Higgs boson.
Similarly, we can define the loop induced vertices as
\begin{align}
\hat{\Gamma}_{h{\cal V V'}}^{\mu\nu}(p_1^2,p_2^2,q^2)&=g^{\mu\nu}\hat{\Gamma}_{h {\cal V V'}}^1
+\frac{p_1^\nu p_2^\mu}{q^2}\hat{\Gamma}_{h {\cal V V'}}^2
+i\epsilon^{\mu\nu\rho\sigma}\frac{p_{1\rho} p_{2\sigma}}{q^2}\hat{\Gamma}_{h {\cal V V'}}^3.
\end{align}
For the on-shell photon and gluon with a four-momentum $p_i^\mu$, the Ward identity holds, i.e, $p_{i\mu}\hat{\Gamma}_{h{\cal V V}'}^{\mu\nu} = 0$.
This gives the following relation
\begin{align}
\hat{\Gamma}_{h {\cal V V'}}^2 = -\frac{q^2}{p_1\cdot p_2}\hat{\Gamma}_{h {\cal V V'}}^1. \label{ward}
\end{align}
This relation can be applied to the computation of the loop induced decays of the Higgs boson.
For an off-shell photon appearing in the $h \to Z\gamma^* \to Zf\bar{f}$ decay mode at NLO,
Eq.~(\ref{ward}) cannot be used, so that $\hat{\Gamma}_{h {\cal V V'}}^1$ and $\hat{\Gamma}_{h {\cal V V'}}^2$ separately appear,
as it will be discussed in Sec.~\ref{sec:hzz}.
We note that the form factor $\hat{\Gamma}_{h {\cal V V}'}^3$ is non-zero only when the Higgs boson is a CP-mixed state.
In this paper, we consider the case with CP-conservation in the Higgs sector, so that this form factor becomes zero.
All the renormalized form factors for the $hXX$ vertices defined above
are further decomposed into the tree level and one-loop level parts as follows:
\begin{align}
\hat{\Gamma}^i_{hXX}(p_1^2,p_2^2,q^2)&=\Gamma^{i,{\rm tree}}_{hXX} + \Gamma^{i,{\rm loop}}_{hXX}(p_1^2,p_2^2,q^2) . \label{eq:form-loop}
\end{align}
The tree level contribution to each form factor{,} denoted as $\Gamma^{i,{\rm tree}}_{hXX}${,} is given as
\begin{align}
\Gamma^{S,{\rm tree}}_{hff} = \kappa_f m_f(\sqrt{2}G_F)^{1/2},\quad
\Gamma^{1,{\rm tree}}_{hVV} = 2\kappa_V m_V^2(\sqrt{2}G_F)^{1/2}, \label{form_tree}
\end{align}
where the scaling factors $\kappa_f^{}$ and $\kappa_V^{}$ are given in Table~\ref{tab:kappa} for each extended Higgs model.
All the other form factors are zero at tree level.
The one-loop contributions ($\Gamma^{i,{\rm loop}}_{hXX}$) are decomposed by
\begin{align}
\Gamma^{i,{\rm loop}}_{hXX}(p_1^2,p_2^2,q^2) &= \Gamma^{i,{\rm 1PI}}_{hXX}(p_1^2,p_2^2,q^2) + \delta \Gamma^i_{hXX}. \label{gam_hff_loop}
\end{align}
The first and second terms of the right-hand side are the contribution from 1PI diagrams and counterterms, respectively.
As we mentioned at the beginning of this section, counterterms are determined by a set of on-shell conditions by which
these are written in terms of 1PI diagrams for one- and two-point functions with a fixed value of squared momenta.
Analytic expressions for the contributions from 1PI diagrams and counterterms to these renormalized Higgs boson vertices
are presented in {Ref.~\cite{Kanemura:2015fra} for the HSM, Refs.~\cite{Kanemura:2015mxa, Kanemura:2018esc} for the THDMs, and Refs.~\cite{Kanemura:2016sos,Kanemura:2018esc} for the IDM.}
For the computation of the partial decay rates of $h \to VV^* \to Vf\bar{f}$,
we also need to calculate the one-loop corrected $Vf\bar{f}$ vertices and box diagrams in addition to the above Higgs boson vertices.
In the massless limit for the external fermions, the renormalized $Vf\bar{f}$ vertices are the same as those in the SM, while
the contribution from the box diagrams is simply given by the SM expression multiplied by the scaling factor $\kappa_V^{}$.
For the completeness, we present the analytic expressions for one-loop corrections to the $Vf\bar{f}$ vertices in Appendix~\ref{sec:vff} and those for the box corrections in Appendix~\ref{sec:box}.
\begin{table}[t]
\begin{center}
\begin{tabular}{c|ccccccc|}\hline\hline
& HSM & THDMs & IDM \\\hline
$\kappa_f$ & $c_\alpha$ &$s_{\beta-\alpha} + \zeta_f c_{\beta-\alpha}$ & 1\\\hline
$\kappa_V$ & $c_\alpha$ &$s_{\beta-\alpha}$ & 1 \\\hline\hline
\end{tabular}
\caption{Scaling factors for the Higgs boson couplings to fermions ($\kappa_f^{}$) and weak bosons ($\kappa_V^{}$) in the extended Higgs models at tree level.
The $\zeta_f$ factor in the THDMs is given in Table~\ref{tab:z2}. }
\label{tab:kappa}
\end{center}
\end{table}
\subsection{$h \to f\bar{f}$\label{sec:hff}}
At NLO, the partial decay rate of the $h\to f\bar{f}$ $(f \neq t)$ process can be written in terms of the EW correction part $\Delta_{\text{EW}}^f$
and the QCD correction part $\Delta_{\text{QCD}}^f$ as
\begin{align}
\Gamma(h\to f\bar{f})=\Gamma_0(h\to f\bar{f}) \left[1 + \Delta_{\text{EW}}^f + \Delta_{\text{QCD}}^f \right],
\end{align}
where $\Gamma_0$ is the decay rate at LO expressed as
\begin{align}
\Gamma_0(h\to f\bar{f}) = \frac{N_c^f}{8\pi}m_h(\Gamma_{hff}^{S,\text{tree}})^2\left(1-\frac{4m_f^2}{m_h^2}\right)^{3/2}, \label{eq:hff-lo}
\end{align}
with $N_c^f$ being the color factor{,} i.e., $N_c^f = 3\,(1)$ for $f$ to be quarks (leptons).
The expression for the tree level form factor $\Gamma_{hff}^{S,\text{tree}}$ is given in Eq.~(\ref{form_tree}).
The EW corrections $\Delta_{\text{EW}}^f$ can be further decomposed into weak corrections $\Delta_{\text{weak}}^f$ and QED corrections $\Delta_{\text{QED}}^f$ as:
\begin{align}
\Delta_{\text{EW}}^f = \Delta_{\text{weak}}^f + \Delta_{\text{QED}}^f.
\end{align}
Here, the weak correction means contributions from $W$, $Z$ and scalar boson loops, namely all the loop contributions except for the photon and gluon loops.
We also use this terminology in the later discussion.
The expression for $\Delta_{\text{weak}}^f$ is given in terms of the form factors defined in Eqs.~(\ref{eq:hff-form}) and (\ref{eq:form-loop}) as
\begin{align}
\Delta_{\text{weak}}^f = \frac{2}{\Gamma_{hff}^{S,{\rm tree}}}{\rm Re}\left\{\left[\Gamma_{hff}^{S,{\rm loop}}
+2m_f\Gamma^{V_1,{\rm loop}}_{hff}
+m_h^2\left(1-\frac{m_f^2}{m_h^2}\right)\Gamma_{hff}^{T,{\rm loop}}\right](m_f^2,m_f^2,m_h^2)\right\} - \Delta r , \label{eq:delew}
\end{align}
where $\Delta r$ is given in Eq.~(\ref{delr}).
The QED (QCD) corrections are obtained by taking into account contributions from the virtual photon (gluon) loops and
those from the real photon (gluon) emissions.
We can obtain simple expressions for these corrections by neglecting the term proportional to $m_f^2/m_h^2$.\footnote{In the numerical computation,
we use the exact formula for the QED correction with the $m_f^2/m_h^2$ term, which is given in Ref.~\cite{Dabelstein:1991ky}. }
For the leptonic decays, $f = \ell$, {the QED correction in the on-shell scheme is given by}~\cite{Kniehl:1991ze,Dabelstein:1991ky,Bardin:1990zj}
\begin{align}
\Delta_{\text{QED}}^\ell = \frac{\alpha_{\text{em}}}{\pi}Q_\ell^2\left(\frac{9}{4} + \frac{3}{2}\log\frac{m_\ell^2}{m_h^2} \right). \label{eq:del_qed_f}
\end{align}
For the hadronic decays, $f = q$, {the QED and QCD corrections in the $\overline{\text{MS}}$ scheme~\cite{Mihaila:2015lwa} with the renormalization scale $\mu$ are given by}
\begin{align}
\Delta_{\text{QED}}^q = \frac{\alpha_{\text{em}}}{\pi}Q_q^2\left(\frac{17}{4} +\frac{3}{2}\log\frac{\mu^2}{m_h^2} \right),\quad
\Delta_{\text{QCD}}^q = \frac{\alpha_s(\mu)}{\pi}C_F\left(\frac{17}{4} +\frac{3}{2}\log\frac{\mu^2}{m_h^2}\right), \label{eq:del_qcd_f}
\end{align}
where $C_F = 4/3$. In the numerical evaluation, we set $\mu = m_h$, and replace the quark mass in the tree level form factor $\Gamma_{hff}^{S,{\rm tree}}$ in Eq.~(\ref{eq:hff-lo}) by
the running mass $\bar{m}_q (\mu = m_h)$.
From Eqs.~(\ref{eq:del_qed_f}) and (\ref{eq:del_qcd_f}), we can see that there is no additional Higgs boson mass dependence in their expression, so that
these corrections do not provide deviations in the Higgs boson decay rate from the SM prediction at NLO.
\subsection{$h \to ZZ^* \to Zf\bar{f}$ \label{sec:hzz}}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.2]{NLO-hVV-diag.png}
\vspace{1.0cm}
\caption{Diagrams contributing to the $h\to ZZ^\ast\to Zf\bar{f}$ ($h\to WW^\ast\to Wf\bar{f}'$) mode at NLO. Each diagram denotes
the contributions from $hVV$ vertex corrections (a), oblique corrections (b), $Vf\bar{f}$ vertex corrections (c), $hf\bar{f}$ vertex corrections
(d) and box corrections (e). }
\label{FIG:hzff}
\end{figure}
We calculate the partial decay rate of the Higgs boson into a pair of weak bosons at NLO in this and next subsections.
Because the mass of the discovered Higgs boson is about 125 GeV, at least one of the weak bosons must be off-shell. We thus calculate the process with 3-body final states{,} i.e.,
$h \to VV^* \to Vf\bar{f}$.
Throughout this paper, we neglect the masses of external fermions in the $h \to VV^* \to Vf\bar{f}$ processes.
In Fig.~\ref{FIG:hzff}, all the diagrams contributing to the process are shown.
Similar to the $h \to f \bar{f}$ mode, the decay rate for the $h\to ZZ^\ast \to Z f\bar{f}$ mode at NLO is expressed as
\begin{align}
\Gamma(h\to Z f\bar{f}) =\Gamma_0(h\to Z f\bar{f})\left[1+\Delta_{\rm EW}^Z +\Delta_{\rm QCD}^Z \right],
\end{align}
and the EW correction can separately be expressed by the weak corrections and the QED corrections:
\begin{align}
\Delta_{\rm EW}^Z = \Delta_{\rm weak}^Z + \Delta_{\rm QED}^Z.
\end{align}
The LO contribution to the decay rate of $h\to Z f\bar{f}$ is expressed by
\begin{align}
\Gamma_0(h\to Z f\bar{f}) &= \int_0^{(m_h-m_Z)^2} |\overline{{\cal M}_0^Z}|^2\, ds, \label{eq:gam0}
\end{align}
where $s$ is the Mandelstam variable defined by $(p_f^\mu + p_{\bar{f}}^\mu)^2$, and another variable $u$ defined by $(p_Z^\mu + p_{\bar{f}}^\mu)^2$ is already
integrated out in the squared tree level amplitude $|\overline{{\cal M}_0^Z}|^2$ expressed as
\begin{align}
|\overline{{\cal M}_0^Z}|^2 = \frac{g_Z^2(\Gamma_{hZZ}^{1,\text{tree}})^2}{256\pi^3m_h^3}\frac{v_f^2 + a_f^2}{(x_s - x_Z)^2}\frac{\lambda(x_Z,x_s)+12 x_Z x_s}{3x_Z}\lambda^{1/2}(x_Z,x_s),
\label{eq:m0sq}
\end{align}
with $x_Z^{} = m_Z^2/m_h^2$, $x_s = s/m_h^2$ and $\lambda(x,y) = (1-x-y)^2-4xy$.
The tree level form factor $\Gamma_{hZZ}^{1,\text{tree}}$ is given in Eq.~(\ref{form_tree}), and that for the $Zf\bar{f}$ vertex
$v_f$ and $a_f$ is given in Eq.~(\ref{eq:vff_tree}).
The QED ($\Delta_{\rm QED}^Z$) and QCD ($\Delta_{\rm QCD}^Z$) corrections only enter the $Zf\bar{f}$ vertex correction depicted {in} the diagram (c) in Fig.~\ref{FIG:hzff}.
Their expressions are common to the SM given as~\cite{Kniehl:1993ay}
\begin{align}
\Delta_{\rm QED}^Z = Q_f^2\frac{3\alpha_{\text{em}}}{4\pi},\quad
\Delta_{\rm QCD}^Z = C_F\frac{3\alpha_s(\mu)}{4\pi}. \label{eq:qcd}
\end{align}
Although the diagrams (d) and (e) can also receive QED and QCD corrections, they vanish in the massless limit for the external fermions.
The weak corrections $\Delta_{\rm weak}^Z$ are given by
\begin{align}
\Delta_{\rm weak}^Z & = \frac{2}{\Gamma_0}\int_0^{(m_h-m_Z)^2}ds\, |\overline{{\cal M}_0^Z}|^2 \Bigg\{
\text{Re}\left[ \frac{\Gamma_{hZZ}^{1,{\rm loop}}}{\Gamma_{hZZ}^{1,{\rm tree}}} + \frac{\bar{\lambda}(x_Z,x_s)}{x_Z^{}} \frac{\Gamma_{hZZ}^{2,{\rm loop}} }{\Gamma_{hZZ}^{1,{\rm tree}}}\right](m_Z^2,s,m_h^2) \notag \\
& +\frac{v_fQ_fc_Ws_W}{v_f^2+a_f^2}\frac{s - m_Z^2}{s} \text{Re}\left[\frac{\hat{\Gamma}_{hZ\gamma}^1}{\Gamma_{hZZ}^{1,{\rm tree}}}
+ \bar{\lambda}(x_Z,x_s)\frac{\hat{\Gamma}_{hZ\gamma}^{2}}{\Gamma_{hZZ}^{1,{\rm tree}}} \right](m_Z^2,s,m_h^2) \notag\\
&+\frac{\text{Re}[v_f\Gamma^{V,{\rm loop}}_{Zff} + a_f\Gamma_{Zff}^{A,{\rm loop}}](0,0,s)}{v_f^2 + a_f^2}
-\frac{\text{Re}\,\hat{\Pi}_{ZZ}(s)}{s-m_Z^2} - \frac{v_fQ_fs_W^{}c_W^{}}{v_f^2 + a_f^2}\frac{\text{Re}\,\hat{\Pi}_{Z\gamma}(s)}{s} \Bigg\}\notag \\
& + \frac{1}{\Gamma_0}\int_0^{(m_h-m_Z)^2}ds \int_{u_{\text{min}}}^{u_{\text{max}}} du\, \text{Re}\left(T_{hff}^Z + B_Z\right) \notag\\
& - 2\Delta r - {\rm Re}\hat{\Pi}_{ZZ}^\prime(m_Z^2), \label{eq:del_ew_z}
\end{align}
where
\begin{align}
\bar{\lambda}(x,y) = \frac{1- x - y}{2}\frac{\lambda(x,y)}{\lambda(x,y) +12 x y}.
\end{align}
The similar expression in the SM is found in Ref.~\cite{Kniehl:1993ay}.
The first and second lines correspond to the contribution from the diagram (a) in Fig.~\ref{FIG:hzff}, where
$\hat{\Gamma}_{hZ\gamma}^{1,2}$ are the renormalized form factors for the loop induced $hZ\gamma$ vertex.
The analytic expressions for $\hat{\Gamma}_{hZ\gamma}^{1,2}$ are presented in Appendix~\ref{sec:hgg}.
The third line corresponds to the contribution from the diagrams (b) and (c).
The diagram (c) contains the $Vf\bar{f}$ vertex corrections, so that we need to prepare the calculation of the renormalized $Vf\bar{f}$ vertex denoted as $\hat{\Gamma}_{Vff}$ which
will be implemented in the {\tt H-COUP} ver. 2.0~\cite{future}.
In the massless limit of the external fermions, this correction becomes the same as the SM prediction.
Details of the calculation of $\hat{\Gamma}_{Vff}$ are given in Appendix~\ref{sec:vff}.
In the fourth line, the $T_{hff}^Z$ and $B_Z$ terms represent the contribution from the $hf\bar{f}$ vertex corrections and the box diagrams shown as the diagrams (d) and (e) in Fig.~\ref{FIG:hzff}, respectively.
Both $T_{hff}^Z$ and $B_Z$ depend on the Mandelstam variable $u$ in loop functions, which has to be integrated out.
The integration range of $u$ is given by
\begin{align}
u_{\text{max},\text{min}} = \frac{m_h^2}{2}[1+x_Z^{} - x_s \pm \lambda^{1/2}(x_Z,x_s)].
\end{align}
The explicit formulae for $T_{hff}^Z$ and $B_Z$ are given in Appendix~\ref{sec:box}.
Although the $hf\bar{f}$ vertex corrections can be calculated by using {\tt H-COUP} ver. 1.0~\cite{Kanemura:2017gbi}, we present the {explicit} analytic formulae
for the contribution from the diagram (d) in Eq.~(\ref{eq:thff1}) in the massless limit of the external fermions.
In this limit, both contributions from (d) and (e) become the SM predictions times the scaling factor $\kappa_V^{}$.
We note that we need to add the contribution from the wave function renormalization of the external $Z$ boson{,} i.e., $\hat{\Pi}'_{ZZ}$, because
in our on-shell scheme, the counterterm for the wave function renormalization (normally denoted as $\delta Z_Z$) is not fixed by the condition
{which requires a vanishing derivative} of the renormalized $Z$ boson two-point function, but it is determined by the other conditions, see Ref.~\cite{Kanemura:2017wtm}.
\subsection{$h \to WW^* \to Wf\bar{f}'$}
We compute the partial decay rate of $h \to WW^* \to Wf\bar{f}'$ mode at NLO.
Feynman diagrams are shown in Fig.~\ref{FIG:hzff}.
The decay rate is expressed as
\begin{align}
\Gamma(h\to Wf\bar{f^\prime})&=\Gamma_0(h\to Wf\bar{f^\prime})\left[ 1+\Delta_{\rm EW}^W +\Delta_{\rm QCD}^W \right],
\end{align}
where $\Gamma_0$, $\Delta_{\text{EW}}^W$ and $\Delta_{\text{QCD}}^W$ are
the contributions from LO, EW corrections and QCD corrections, respectively.
The expression for the LO contribution is obtained from Eqs.~(\ref{eq:gam0}) and (\ref{eq:m0sq}) by replacing $Z \to W$ with $g_W^{} \equiv g/\sqrt{2}$, and that
for the QCD corrections is the same as that given in Eq.~(\ref{eq:qcd}), because
the gluon loop corrections only appear in the $W\bar{f}f'$ vertex similar to the $Z\bar{f}f$ vertex in the $h \to Zf\bar{f}$ decay~\cite{Denner:1991kt}.
The EW corrections $\Delta_{\text{EW}}^W$ are given in a similar way to Eq.~(\ref{eq:del_ew_z}) as follows
\begin{align}
\Delta_{\rm EW}^W & = \frac{1}{\Gamma_0}\int_0^{(m_h-m_W)^2}ds\, |\overline{{\cal M}_0^W}|^2\Bigg\{
\text{Re}\left[ \frac{2\Gamma_{hWW}^{1,{\rm loop}}}{\Gamma_{hWW}^{1,{\rm tree}}} + \frac{\bar{\lambda}(x_W,x_s)}{x_W^{}}
\frac{\Gamma_{hWW}^{2,{\rm loop}} }{\Gamma_{hWW}^{1,{\rm tree}}}\right](m_W^2,s,m_h^2) \notag \\
&+2\text{Re}[\Gamma^{V,{\rm loop}}_{Wff} + \Gamma_{Wff}^{A,{\rm loop}}](0,0,s)
-\frac{2\text{Re}\,\hat{\Pi}_{WW}(s)}{s-m_W^2}\Bigg\}\notag \\
&+\frac{1}{\Gamma_0} \left[\int_0^{(m_h-m_W)^2}ds \int_{u_{\text{min}}}^{u_{\text{max}}} du\, \left(T_{hff}^W + B_W\right) +\Gamma({h\to Wf\bar{f}^\prime \gamma}) \right]\notag\\
& - 2\Delta r - {\rm Re}\hat{\Pi}_{WW}^\prime(m_W^2), \label{eq:del_ew_w}
\end{align}
where the expressions for the $Wf\bar{f}'$ vertex corrections $\Gamma^{V,{\rm loop}}_{Wff}$ and $\Gamma_{Wff}^{A,{\rm loop}}$
are given in Appendix~\ref{sec:vff}.
The $hf\bar{f}$ vertex corrections $T_{hff}^W$ and the box diagrams $B_W$ are given in Appendix~\ref{sec:box}.
In the {third} line, $\Gamma({h\to Wf\bar{f}^\prime \gamma})$ denotes the contribution from real photon emissions.
Differently from the $h \to Zf\bar{f}$ mode, we cannot separate the QED corrections from the EW one,
because the virtual photon loop can appear together with the $W$ boson loop in vertex corrections.
Thus, we cannot obtain a simple expression for the QED correction such as Eq.~(\ref{eq:qcd}) in this decay mode.
However, such IR divergence can be cancelled by adding the real photon emissions $\Gamma({h\to Wf\bar{f}^\prime \gamma})$.
Detailed discussions for the treatment of the IR divergence are given in Appendix~\ref{sec:real_photon}.
Similar to the case of $h \to Zf\bar{f}$, in the massless limit of the external fermions
the contributions from {the} $Wf\bar{f}$ vertex become the same as those in the SM, while $T_{hff}^W$ and
$B_W$ are given by the SM prediction multiplied by $\kappa_V^{}$.
\subsection{$h \to \gamma\gamma$, $Z\gamma$, $gg$}
In addition to the $h \to f\bar{f}$ and $h \to VV^* \to Vf\bar{f}$ decays, the Higgs boson can also decay into $\gamma\gamma$, $Z\gamma$ and $gg$.
The LO contributions to the decay rates arise from one-loop diagrams, and they can be expressed in terms of the renormalized
$h{\cal VV}'$ (${\cal VV}' = \gamma\gamma,~Z\gamma,~gg$) vertices defined in Sec.~\ref{sec:ren-vert} as
\begin{align}
\Gamma_0(h\to {\cal VV}') &= \frac{|\hat{\Gamma}_{h{\cal VV}'}^1(m_{\cal V}^2,m_{\cal V'}^2,m_h^2)|^2}{8\pi m_h} \lambda^{1/2}\left(\frac{m_{\cal V}^2}{m_h^2},\frac{m_{{\cal V}'}^2}{m_h^2} \right), \label{eq:loop-induced}
\end{align}
where Eq.~(\ref{ward}) is used.
The analytic expressions for $\hat{\Gamma}_{h{\cal VV}'}^{1,2}$ are given in Appendix~\ref{sec:hgg}.
Let us discuss QCD corrections to these loop induced decay rates at NLO.
For $h\to gg$, there are two sources of {the QCD corrections:} virtual gluon exchanges in the quark loop diagrams and real gluon emissions.
In the $\overline{\rm MS}$ scheme, the QCD corrected decay rate is given as~\cite{Djouadi:2005gi}
\begin{align}
\Gamma(h\to gg)=\Gamma_0(h\to gg)\left[1 + \frac{\alpha_s(\mu)}{\pi}\left(\frac{95}{4}-\frac{7}{6}N_f+\frac{33-2N_f}{6}\log\frac{\mu^2}{m_h^2}\right)\right],
\end{align}
in the limit of $m_t\to \infty$ with $N_f$ being the number of light flavors.
Numerically, the magnitude of NLO correction is about 70\% for $N_f = 5$ and $\mu = m_h$.
For $h\to\gamma\gamma$ and $h\to Z\gamma$, quark loop diagrams are modified by QCD corrections.
We only take into account the NLO QCD correction for the top loop contributions, because those to the other quark loops are negligible.
In the limit of $m_t\to \infty$, the QCD correction is easily implemented in the $\overline{\rm MS}$ scheme as~\cite{Djouadi:2005gi}
\begin{align}
\hat{\Gamma}_{h{\cal V}\gamma}^1(m_{\cal V}^2,0,m_h^2)_{t} \to \hat{\Gamma}_{h{\cal V}\gamma}^1(m_{\cal V}^2,0,m_h^2)_t\left[1-\frac{\alpha_s(\mu)}{\pi}\right],~~~({\cal V}=\gamma,~Z),
\end{align}
where $\hat{\Gamma}_{h{\cal V}\gamma}^1(m_{\cal V}^2,0,m_h^2)_t$ is the top quark loop contribution to the renormalized $h\gamma\gamma$ and $hZ\gamma$ vertices.
The typical magnitude of the QCD corrections is a few percent level with respect to the LO result.
\subsection{New physics effects in loops\label{sec:newphys}}
\begin{table}[!t]
\begin{center}
{\renewcommand\arraystretch{1.2}
\begin{tabular}{cccccc}\hline\hline
$\Delta_{\text{EW}}^{b}$ & $\Delta_{\text{EW}}^{c}$ & $\Delta_{\text{EW}}^{\tau}$ & $\Delta_{\text{EW}}^{Z}$ & $\Delta_{\text{EW}}^{W}$ \\\hline
1.67\% & 1.78\% & 4.91\% & 6.87\% & 3.14\% \\\hline\hline
\end{tabular}}
\caption{EW corrections $\Delta_{\text{EW}}^X$ ($X=b,c,\tau,Z,W$) in the SM. }
\label{tab:sm}
\end{center}
\end{table}
We show numerical values of the EW corrections $\Delta_{\text{EW}}^X$ including weak boson and scalar boson loop effects to the decay rates discussed in the previous subsections.
The numerical values of our inputs shown in Eq.~(\ref{inputs}) are fixed to be the default values implemented in the {\tt H-COUP} code~\cite{Kanemura:2017gbi}.
In order to extract the new physics effects of the EW corrections to the partial decay rates,
we introduce
\begin{align}
\overline{\Delta}_{\text{EW}}^X = \Delta_{\text{EW}}^X\big|_{\rm NP} - \Delta_{\text{EW}}^X\big|_{\rm SM}, \label{eq:delbar}
\end{align}
where $\Delta_{\text{EW}}^X|_{\rm NP}$ $(\Delta_{\text{EW}}^X|_{\rm SM})$ denotes the prediction of $\Delta_{\text{EW}}^X$ in the models with the extended Higgs sectors (SM).
Our results for $\Delta_{\text{EW}}^X|_{\rm SM}$ are summarized in Table~\ref{tab:sm}.
It is important to mention here that the dominant contribution to $\overline{\Delta}_{\text{EW}}^X$ comes from the first term of Eqs.~(\ref{eq:delew}), (\ref{eq:del_ew_z}) and (\ref{eq:del_ew_w})
for $\overline{\Delta}_{\text{EW}}^f$, $\overline{\Delta}_{\text{EW}}^Z$ and $\overline{\Delta}_{\text{EW}}^W$, respectively.
In the case with $\kappa_V^{} \simeq 1$ and $m_\varphi \gg m_h$, additional Higgs boson loop effects of $\Delta_{\text{EW}}^X$ are approximately expressed as~\cite{Kanemura:2015mxa}
\begin{align}
\overline{\Delta}_{\text{EW}}^X \simeq -\frac{1}{16\pi^2}\frac{1}{6}\sum_\varphi c_\varphi \frac{m_\varphi^2}{v^2}\left(1 - \frac{M^2}{m_\varphi^2} \right)^2 , \label{eq:nondec}
\end{align}
where $c_\varphi= 2(1)$ for additional charged (neutral) scalar loop contributions.\footnote{In the THDMs, the charged Higgs boson and top quark loop contribution can also be important for $\overline{\Delta}_{\rm EW}^b$.
The analytic expression for the top quark mass dependence due to charged Higgs boson loops is found in Ref.~\cite{Kanemura:2015mxa}. }
The above equation indicates that scalar loop effects become non-decoupling when $m_\varphi$ is mostly determined by $v$, or equivalently the case with
$M^2/v^2 \ll 1$. In such a non-decoupling case,
the right-hand side of Eq.~(\ref{eq:nondec})
is nearly proportional to $m_\varphi^2$. Of course, there must be an upper limit on $m_\varphi$ from the unitarity bound, under which
the magnitude of the non-decoupling effect can be typically a few percent level, {as we will see in the plots below.}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.7]{delbar_b_HSM.png} \hspace{3mm}
\includegraphics[scale=0.7]{delbar_b_IDM.png}
\caption{New physics effects in the EW corrections $\overline{\Delta}_{\text{EW}}^{b}$ as a function of a mass of the extra scalar boson
in the HSM (left) and the IDM (right).
We take $c_\alpha = 1$, $\mu_S = 0$ and $\lambda_S = 0.1$ in the HSM, while set $m_H^{} = m_A^{} = m_{H^\pm}$ and $\lambda_2 = 0.1$ in the IDM.
The solid (dashed) curves denote the case with the maximal (minimal) value of $M^2$ allowed by the perturbative unitarity, vacuum stability bounds and $S,~T$ parameters. }
\label{fig:delb-hsm}
\end{center}
\end{figure}
Now, we show the plots of $\overline{\Delta}_{\text{EW}}^X$ in each model with the extended Higgs sectors discussed in Sec.~\ref{sec:model}.
Although the quantity $\overline{\Delta}_{\text{EW}}^X$ cannot directly be measured at collider experiments,
studying the prediction of $\overline{\Delta}_{\text{EW}}^X$ turns out to be important to understand the behavior of the deviation in branching ratios from the SM prediction, which will be discussed in the next section.
In the following discussion, we impose the bounds from the perturbative unitarity, the vacuum stability, the conditions to avoid wrong vacua (taking $M^2\geq 0$) and the $S,T$ parameters discussed in Sec.~\ref{sec:model}.
The flavor constraints discussed in Sec.~\ref{sec:2hdm} are also important to be taken into account particularly in the THDMs, but
{we do not impose them here in order to study and} compare the behavior of $\overline{\Delta}_{\text{EW}}^X$ among the extended Higgs sectors.
In the later discussion given in Sec.~\ref{sec:correlations}, we discuss the branching ratios imposing the flavor constraints as well.
In Fig.~\ref{fig:delb-hsm}, predictions of $\overline{\Delta}_{\text{EW}}^b$ are shown in the HSM (left) and the IDM (right),
{results for the other fermions $f$ look almost the same as what are shown in these figures.}
In the HSM, we take $c_\alpha = 1$, $\mu_S = 0$ and $\lambda_S =0.1$, where $\overline{\Delta}_{\text{EW}}^b$ does not directly depend on $\lambda_S$, but it
indirectly determines the allowed size of $\overline{\Delta}_{\text{EW}}$ via the unitarity and vacuum stability bounds.
In the IDM, we take $m_H^{} = m_A^{} = m_{H^\pm}$ and $\lambda_2 = 0.1$, where $\overline{\Delta}_{\text{EW}}^b$ does not directly depend on $\lambda_2$.
As we see from the plots, the magnitude of $\overline{\Delta}_{\text{EW}}^b$ becomes larger when the mass of the extra scalar boson is taken to be larger up to around 900 GeV and 600 GeV in the
HSM and IDM, respectively.
The maximal deviation is found to be $|\overline{\Delta}_{\text{EW}}^b| \simeq 2.5$ (5\%) in the HSM (IDM), which is given at $M^2 = 0$.
The larger maximal amount of the deviation in the IDM as compared with the HSM is due to more than one additional scalar {boson} running in the loop in the IDM.
{In the case of larger values} of $m_H^{}$, the magnitude of $|\overline{\Delta}_{\text{EW}}^b|$ monotonically decreases, because $M^2 = 0$ cannot be taken due to the unitarity constraint.
We then can see the decoupling behavior, $\overline{\Delta}_{\text{EW}}^b \to 0$, at the large mass limit {in both models}, where
loop effects of additional Higgs bosons vanish.
The behavior of $\overline{\Delta}_{\text{EW}}^Z$ and $\overline{\Delta}_{\text{EW}}^W$ in the HSM and the IDM is almost the same as that of $\overline{\Delta}_{\text{EW}}^b$ shown in Fig.~\ref{fig:delb-hsm}.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.7]{delbar_b_TypeI_sba1.png} \\
\includegraphics[scale=0.7]{delbar_b_TypeI_sba099_cbam.png} \hspace{5mm}
\includegraphics[scale=0.7]{delbar_b_TypeI_sba099_cbap.png}
\caption{New physics effects in the EW corrections $\overline{\Delta}_{\text{EW}}^{b}$
as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$ in the Type-I and Type-X THDMs with fixed values of $\tan\beta = 1.5$ (red), 3 (blue) and 5 (green).
The upper panel shows the case with $s_{\beta-\alpha} = 1$ and the lower left (right) panel shows
the case with $s_{\beta-\alpha} = 0.99$ and $c_{\beta-\alpha} <0$ ($c_{\beta-\alpha} > 0$).
The solid (dashed) curves denote the case with the maximal (minimal) value of $M^2$ allowed by
the perturbative unitarity, vacuum stability bounds and $S,~T$ parameters. }
\label{fig:delb-1}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.7]{delbar_b_TypeII_sba1.png} \\
\includegraphics[scale=0.7]{delbar_b_TypeII_sba099_cbam.png} \hspace{5mm}
\includegraphics[scale=0.7]{delbar_b_TypeII_sba099_cbap.png}
\caption{Same as Fig.~\ref{fig:delb-1}, but for the Type-II and Type-Y THDMs. }
\label{fig2}
\end{center}
\end{figure}
In Fig.~\ref{fig:delb-1}, the value of $\overline{\Delta}_{\text{EW}}^{b}$ is plotted as a function of the common additional Higgs boson mass $m_\Phi$ defined by
$m_\Phi =m_H = m_A = m_{H^\pm}$ in the Type-I THDM with fixed values of $\tan\beta$.
Here, we take $s_{\beta-\alpha} = 1$ (upper panel) and $s_{\beta-\alpha} = 0.99$ (lower panels), where the lower left and right panels show the cases of $c_{\beta-\alpha} < 0$ and $c_{\beta-\alpha} > 0$, respectively.
We note that the results in the Type-X THDM are almost the same as those in the Type-I THDM.
From the upper panel, the decoupling behavior can clearly be seen in the large mass region as in the HSM and the IDM.
On the other hand, in the case with $s_{\beta-\alpha}=0.99$ shown in the lower panels,
the decoupling limit cannot be taken, so that there appears the upper limit on the mass of the extra Higgs boson
from the theoretical constraints depending on the value of $\tan\beta$ and the sign of $c_{\beta-\alpha}$.
At $m_\Phi \sim 2m_t$,
the threshold effects of $t\bar{t}$ appear, which push $\overline{\Delta}_{\text{EW}}^b$ into the positive direction.
This peak comes from the top quark loop contribution to the $Z$--$A$ mixing diagram which appears in the counterterm of the $\beta$ parameter.
More detailed discussions have been given in Ref.~\cite{Kanemura:2014dja}.
We can also see the dip above the $t\bar{t}$ threshold for $\tan\beta = 1.5$.
The origin of this dip can be explained {in} the same way as in Fig.~\ref{fig:delb-hsm}.
Namely, {the point, where the dip appears,} corresponds to the maximal value of the mass of the extra Higgs boson with $M^2 = 0$ allowed by the unitarity bound.
Similar to the results in the HSM and the IDM seen in Fig.~\ref{fig:delb-hsm}, the non-decoupling effect of the extra Higgs boson, which can be more significant
for smaller $M^2$, pushes down the value of $\overline{\Delta}_{\text{EW}}^b$.
For larger values of $\tan\beta$, allowed regions of $\overline{\Delta}_{\text{EW}}^{b}$ for a fixed value of $m_\Phi$ are significantly shrunk as compared to the case with $\tan\beta =1.5$,
while {the behavior explained above} does not change so much.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.65]{delbar_bctau_TypeI_sba1.png} \hspace{5mm}
\includegraphics[scale=0.65]{delbar_bctau_TypeII_sba1.png} \\
\includegraphics[scale=0.65]{delbar_bctau_TypeX_sba1.png} \hspace{5mm}
\includegraphics[scale=0.65]{delbar_bctau_TypeY_sba1.png}
\caption{New physics effects in the EW corrections $\overline{\Delta}_{\text{EW}}^{f}$ ($f=b,c,\tau$)
as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$ in the Type-I (upper-left), Type-II (upper-right), Type-X (lower-left)
and Type-Y (lower-right) THDM with $s_{\beta-\alpha} = 1$ and $\tan\beta = 1.5$.
The solid (dashed) curves denote the case with the maximal (minimal) value of $M^2$ allowed by
the perturbative unitarity, vacuum stability bounds and $S,~T$ parameters. }
\label{fig:delbct}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.65]{delbar_Z_TypeI_sba1_tb1.png} \\
\includegraphics[scale=0.65]{delbar_Z_TypeI_sba099_cbam.png} \hspace{5mm}
\includegraphics[scale=0.65]{delbar_Z_TypeI_sba099_cbap.png}
\caption{New physics effects in the EW corrections $\overline{\Delta}_{\text{EW}}^{Z}$ as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$ in the Type-I THDM with fixed values of $\tan\beta = 1.5$ (red), 3 (blue) and 5 (green).
The upper panel shows the case with $s_{\beta-\alpha} = 1$ and the lower left (right) panel shows
the case with $s_{\beta-\alpha} = 0.99$ and $c_{\beta-\alpha} <0$ ($c_{\beta-\alpha} > 0$).
The solid (dashed) curves denote the case with the maximal (minimal) value of $M^2$ allowed by
the perturbative unitarity, vacuum stability bounds and $S,~T$ parameters. }
\label{fig:delz}
\end{center}
\end{figure}
In Fig.~\ref{fig2}, {we show similar plots as in Fig.~\ref{fig:delb-1}, however, for the case {of} the Type-II THDM.}
The results in the Type-Y THDM are almost the same as those in the Type-II THDM.
Again, we can see the decoupling behavior for $s_{\beta-\alpha} = 1$, and observe the upper limit on $m_\Phi$ for $s_{\beta-\alpha} = 0.99$, where
the value of the upper limit does not depend on the types of Yukawa interaction.
Although the behavior of the additional Higgs boson loop contribution{,} i.e., pushing down the value of $\overline{\Delta}_{\text{EW}}^b$, can also be seen as in the Type-I case,
the effect of the $t\bar{t}$ threshold appears in the opposite direction as compared to the case {of} the Type-I THDM.
This can be understood by the difference of the $\tan\beta$ dependence {on} the $\zeta_f$ factor, see Table~\ref{tab:z2}.
In addition, for larger values of $\tan\beta$, the magnitude of $\overline{\Delta}_{\text{EW}}^b$ tends to get larger.
For example, for $s_{\beta-\alpha} = 1$, the maximally allowed value of $|\overline{\Delta}_{\text{EW}}^b|$
is about 4.5, 5.5 and 7.5\% for $\tan\beta = 1.5$, 3 and 5, respectively.
Differently from the HSM and IDM, the value of $\overline{\Delta}_{\text{EW}}^{f}$ can be drastically changed depending not only on the type of Yukawa interaction but also the type of fermion.
Thus, in Fig.~\ref{fig:delbct} we show the results for $\overline{\Delta}_{\text{EW}}^{b}$, $\overline{\Delta}_{\text{EW}}^{c}$ and $\overline{\Delta}_{\text{EW}}^{\tau}$ in four types of the THDMs.
Here, we show the case of $s_{\beta-\alpha} = 1$ and $\tan\beta = 1.5$ for all the types of the THDMs.
It is seen that the direction of the peak at around $m_\Phi^{} = 2m_t$ is determined to be positive (negative) if $\zeta_f = \cot\beta$ ($-\tan\beta$), see Table~\ref{tab:z2}.
The behavior of $\overline{\Delta}_{\text{EW}}^{c}$ and $\overline{\Delta}_{\text{EW}}^{\tau}$ is also classified by the factor of $\zeta_f${,} e.g. that of $\overline{\Delta}_{\text{EW}}^{\tau}$ in the Type-II THDM
is common to the Type-X THDM.
{Concerning $\overline{\Delta}_{\text{EW}}^{b}$, the} behavior is different from e.g. $\overline{\Delta}_{\text{EW}}^{\tau}$ in the Type-II THDM even though these two depend on the same factor of $\zeta_f$.
This can be understood by the fact that the top mass dependence enters in $\overline{\Delta}_{\text{EW}}^{b}$ when charged Higgs bosons run in the loop, while for the other $\overline{\Delta}_{\text{EW}}^{f}$
a dependence {on} fermion masses in loops is negligibly small.
Detailed discussions have been given in Ref.~\cite{Kanemura:2014dja} for the behavior of the radiative correction to the Yukawa couplings in the THDMs.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.65]{delbar_W_TypeI_sba099_cbam.png} \hspace{5mm}
\includegraphics[scale=0.65]{delbar_W_TypeI_sba099_cbap.png}
\caption{New physics effects in the EW corrections $\overline{\Delta}_{\text{EW}}^{W}$ as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$ in the Type-I THDM with fixed values of $\tan\beta = 1.5$ (red), 3 (blue) and 5 (green).
The left (right) panel shows
the case with $s_{\beta-\alpha} = 0.99$ and $c_{\beta-\alpha} <0$ $(c_{\beta-\alpha} > 0)$.
The solid (dashed) curves denote the case with the maximal (minimal) value of $M^2$ allowed by
the perturbative unitarity, vacuum stability bounds and $S,~T$ parameters. }
\label{fig:delw}
\end{center}
\end{figure}
In Fig.~\ref{fig:delz}, we show the value of $\overline{\Delta}_{\text{EW}}^{Z}$ in the Type-I THDM with a fixed value of $\tan\beta$.
The results in all the other types of the THDMs are almost the same as those in the Type-I THDM.
Similar to Fig.~\ref{fig:delb-1}, we can see the decoupling behavior for $s_{\beta-\alpha} = 1$ (upper panel) at the large mass region, while
for $s_{\beta-\alpha} = 0.99$ (lower panels), there appears the upper limit on the additional Higgs boson mass $m_\Phi$ depending on $\tan\beta$.
In addition, the position of {the dip at $m_\Phi \simeq 500$, 250 and 150 GeV} for $s_{\beta-\alpha} = 1$ with $\tan\beta = 1.5$, 3 and 5, respectively, is the same
as that shown in the upper panel of Fig.~\ref{fig:delb-1}, because it is determined by the unitarity bound.
It is seen that for $s_{\beta-\alpha} = 1$, the possible values of $\overline{\Delta}_{\text{EW}}^{Z}$ with larger $\tan\beta$ are included in those with smaller $\tan\beta$.
This is simply because the unitarity bound more strongly constrains the possible non-decoupling effect for a larger value of $\tan\beta$.
The lowest value of $\overline{\Delta}_{\text{EW}}^{Z}$ is found to be around $-3\%$, $-1\%$ and $-0.5\%$ for $\tan\beta = 1.5$, 3 and 5, respectively,
and the largest value corresponds to the SM prediction{, i.e.} $\overline{\Delta}_{\text{EW}}^{Z}\simeq 0$.
For $s_{\beta-\alpha} = 0.99$, we find that $\overline{\Delta}_{\text{EW}}^{Z}$ can be positive.
This is because of the contribution from the virtual photon propagation shown as the diagram (a) in Fig.~\ref{FIG:hzff}, which
is proportional to $\hat{\Gamma}_{hZ\gamma}^{1,\text{loop}}/\hat{\Gamma}_{hZZ}^{1,\text{tree}}$, see Eq.~(\ref{eq:del_ew_z}).
Because the tree level $hZZ$ vertex $\hat{\Gamma}_{hZZ}^{1,\text{tree}}$ is now suppressed by the $\kappa_V^{}(=s_{\beta-\alpha})$ factor,
this contribution can be larger than the case with $s_{\beta-\alpha} = 1$.
This behavior should be compared with the results for $\overline{\Delta}_{\text{EW}}^{W}$ shown in Fig.~\ref{fig:delw}.
Because there is no {virtual-photon-propagation diagram} in the $h \to WW^*$ process{,} as seen in Fig.~\ref{FIG:hzff},
the value of $\overline{\Delta}_{\text{EW}}^{W}$ is negative.
For $s_{\beta-\alpha} = 1$, the value of $\overline{\Delta}_{\text{EW}}^{W}$ is almost the same as that of $\overline{\Delta}_{\text{EW}}^{Z}$,
so that we do not show the result of $\overline{\Delta}_{\text{EW}}^{W}$ with $s_{\beta-\alpha} = 1$.
\section{Numerical results for the Higgs boson decay rates\label{sec:numerical}}
In this section, we numerically show predictions of the total width and the branching ratios of the Higgs boson at NLO in the HSM, the THDMs and the IDM.
After we show these quantities, we demonstrate if these extended Higgs models can be distinguished by the difference of the pattern of
deviations in the branching ratios from those of the SM predictions.
Similar to the analysis in Sec.~\ref{sec:newphys}, we take into account the constraints from the unitarity, the vacuum stability, the conditions to avoid wrong vacua and the $S$, $T$ parameters.
Except for Sec.~\ref{sec:correlations}, we dare not to impose the flavor constraints in order to just see the predictions of deviations in total width and branching ratios.
For the THDMs, we introduce the common mass of the additional Higgs bosons $m_\Phi^{}$; i.e., $m_\Phi^{} = m_H^{} = m_A^{} = m_{H^\pm}^{}$.
\subsection{Total widths}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.8]{sadep_total_width_HSM.png}
\includegraphics[scale=0.8]{mbHdep_total_width_HSM.png}
\caption{Deviation in the total width from the SM prediction as a function of $c_\alpha$ (left) and $m_H^{}$ (right) in the HSM with $\mu_S = 0$ and $\lambda_S = 0.1$.
The values of $c_\alpha$, $m_H$ and $M^2$ are scanned within $0.95 < c_\alpha < 1$, $300\leq m_H \leq 5000$ GeV and $0 \leq M^2 \leq m_H^2$, respectively. }
\label{fig:width_hsm}
\end{center}
\end{figure}
We first {discuss} the total width of the Higgs boson $h$.
In Fig.~\ref{fig:width_hsm}, we show the deviation in the total width from the SM prediction in the HSM.
We scan the parameters $c_\alpha$, $m_H$ and $M^2$ within $0.95 < c_\alpha < 1$, $300\leq m_H \leq 5000$ GeV and $0 \leq M^2 \leq m_H^2$, respectively.
The dependences on $c_\alpha$ and $m_H^{}$ are then displayed in the left and right panels, respectively.
At tree level, the deviation in the width is determined by $s_\alpha^2$, and it almost corresponds to the upper edge of
the distribution in the left panel.
The loop effects typically reduce the width by at most about 2\% level.
In the right panel, it is seen that the magnitude of allowed deviations becomes smaller for larger mass regions, because the large mixing is excluded by the theoretical bounds.
We note that the information of the width is important to identify the HSM, because
the branching ratios of the Higgs boson are almost the same as those of the SM due to nearly universal suppression of the partial decay rates in the HSM.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.8]{tbdep_total_width_sba099_cbam.png}
\includegraphics[scale=0.8]{tbdep_total_width_sba099_cbap.png}
\caption{Deviation in the total width from the SM prediction in four types of the THDMs with
$s_{\beta-\alpha} = 0.99$ as a function of $\tan\beta$.
The left and right panels show the case of $c_{\beta-\alpha} < 0$ and $c_{\beta-\alpha} > 0$, respectively.
The values of $m_\Phi^{}$ and $M^2$ are scanned within $300 \leq m_\Phi \leq 1000$ GeV and $0 \leq M^2 \leq m_\Phi^2$, respectively.
}
\label{fig:width_thdm}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.7]{mbHdep_TyIII_total_width_tb1p5_cbam.png} \hspace{3mm}
\includegraphics[scale=0.7]{mbHdep_TyIII_total_width_tb1p5_cbap.png} \\
\includegraphics[scale=0.7]{mbHdep_TyXY_total_width_tb1p5_cbam.png} \hspace{3mm}
\includegraphics[scale=0.7]{mbHdep_TyXY_total_width_tb1p5_cbap.png}
\caption{Deviation in the total width from the SM prediction in the Type-I and Type-II THDMs (upper panels) and in the Type-X and Type-Y THDMs (lower panels) with
$\tan\beta = 1.5$ as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$.
The left and right panels show the case of $c_{\beta-\alpha} < 0$ and $c_{\beta-\alpha} > 0$, respectively.
The values of $M^2$ and $s_{\beta-\alpha}$ are scanned with the ranges of $0 \leq M^2 \leq m_\Phi^{2}$ and $0.95 \leq s_{\beta-\alpha} \leq 1$.
}
\label{fig:width_thdm2}
\end{center}
\end{figure}
In Fig.~\ref{fig:width_thdm}, the deviation in the total width is shown as a function of $\tan\beta$ in four types of the THDMs with
$s_{\beta-\alpha} = 0.99$ and $c_{\beta-\alpha} < 0~(>0)$ in the left (right) panel.
We scan the values of $M^2$ and $m_\Phi^{}$.
In the left plot, it is seen that except for the Type-I THDM, the width becomes larger as $\tan\beta$ increases, because
some of the partial widths have a $\tan\beta$ enhancement{,} e.g., the $h \to b\bar{b}$ ($h \to \tau \bar{\tau}$) mode in the Type-II and Type-Y (Type-II and Type-X) THDMs{.
In} the Type-I THDM on the contrary, the total width approaches to the SM prediction, more precisely $s_{\beta-\alpha}^2\Gamma_{\rm SM}$, at the large $\tan\beta$ region.
We note that the curves are truncated at around $\tan\beta = 11$ (the same thing also happens in the Type-II and Type-Y THDMs), because of the theoretical constraints.
In the case with $c_{\beta-\alpha}>0$ (the right panel),
the situation is drastically different from the case with $c_{\beta-\alpha}<0$.
The total width has the minimal value at $\tan\beta \sim 7$ in the Type-II, Type-X and Type-Y THDMs, due to
the cancellation between the $s_{\beta-\alpha}$ term and the $c_{\beta-\alpha}$ term in $\kappa_f$, see Table~\ref{tab:kappa}.
This behavior is remarkably observed in the Type-II and Type-Y THDMs, because the $h \to b\bar{b}$ mode, which is the biggest partial width of $h$ in the SM,
follows the behavior explained above.
We can also see that at $\tan\beta \simeq 14$, the deviation in the total width becomes zero, as we have $\kappa_f^2 \simeq 1$ for all the types of Yukawa interaction.
In the Type-I THDM, the width approaches to the SM value at a large value of $\tan\beta$ as also seen in the case with $c_{\beta-\alpha} < 0$.
The typical amount of the loop corrections to the total width is a few percent level, which is shown by a width of each curve.
In Fig.~\ref{fig:width_thdm2}, we show the $m_\Phi$ dependence on the deviation in the total width in four types of the THDMs.
{Here, we scan} the values of $M^2$ and $s_{\beta-\alpha}$, {while we fix} $\tan\beta$ to be 1.5.
For $c_{\beta-\alpha} < 0$ (left panels),
the deviation is distributed in the positive (negative) direction in the Type-II and Type-Y (Type-I and Type-X) THDMs, while
for $c_{\beta-\alpha} > 0$ (right panels), the situation is opposite.
This can be understood by focusing on the deviation in the decay rate of $h \to b\bar{b}$ which is expressed
by $\kappa_b^2 -1 = 2\zeta_b s_{\beta-\alpha}c_{\beta-\alpha} + c_{\beta-\alpha}^2(\zeta_b^2 -1)$ at tree level.
As we are considering $s_{\beta-\alpha} \sim 1$, the $2\zeta_b s_{\beta-\alpha}c_{\beta-\alpha}$ term dominantly determines the behavior.
We see that the allowed magnitude of the deviation is shrunk at around $m_\Phi = 700 ~(450)$ GeV for $c_{\beta-\alpha}< 0$ ($c_{\beta-\alpha}> 0$), {because
in the region above $m_\Phi = 700 ~(450)$ GeV the unitarity and/or vacuum stability bounds constrain $s_{\beta-\alpha}$ to be closer to 1.}
As it is expected from the decoupling theorem, for larger $m_\Phi$ the magnitude of the deviation is getting smaller, but it can still be ${\cal O}(10)\%$ level at around $m_\Phi =1$ TeV.
\begin {figure}[!t]
\begin{center}
\includegraphics[scale=0.9]{mbHdep_total_width_IDM.png}
\caption{Deviation in the total width from the SM prediction in the IDM with $\lambda_2 = 0.1$.
The value $M^2$ is scanned within $0\leq M^2 \leq m_{H^\pm}^2$ under the constraints from the perturbative unitarity and the vacuum stability for the case of $m_H^{} =m_{H^\pm}^{}$ shown by dots.
The black curve shows the case for $m_H^{} = 63$ GeV.
}
\label{fig:idm}
\end{center}
\end{figure}
Finally in Fig.~\ref{fig:idm}, we show the total width in the IDM as a function of the charged scalar boson mass $m_{H^\pm}^{}$ with $m_A^{} = m_{H^\pm}^{}$.
We here take two cases; i.e., (i) $m_H^{}$ is fixed to 63 GeV and (ii) $m_H = m_{H^\pm}$.
The case (i) is motivated by the dark matter physics~\cite{Ilnicka:2015jba,Belyaev:2016lok,Ilnicka:2018def,Belyaev:2018ext}, where $H$ can be a dark matter candidate.
In this case, the value of $M^2$ is taken such that the $HHh$ coupling normalized to $v$ becomes around $10^{-3}$ to avoid constraints from dark matter direct detection experiments.
In the IDM, the total width does not change from the SM value at tree level, so that any deviation is purely due to loop effects.
We can see that in the case (i), the total width monotonically decreases and the deviation is larger as $m_{H^\pm}$ is getting larger.
The black curve is truncated at around $m_{H^\pm} = 700 $ GeV, because of the unitarity constraint.
In the case (ii), the magnitude of the deviation becomes larger up to $m_{H^\pm} \simeq 600$ GeV, while
it becomes smaller above $m_{H^\pm} \simeq 600$ GeV.
The maximal deviation is given at $M^2 = 0$ for $m_{H^\pm} < 600$ GeV, while the unitarity constrains the minimal value of $M^2$ above $m_{H^\pm}\simeq$ 600 GeV
and possible deviations become smaller.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.4]{BR_TyI_sba099_cbam.png} \hspace{-3mm}
\includegraphics[scale=0.4]{BR_TyII_sba099_cbam.png} \hspace{-3mm}
\includegraphics[scale=0.4]{BR_TyX_sba099_cbam.png} \hspace{-3mm}
\includegraphics[scale=0.4]{BR_TyY_sba099_cbam.png} \\
\includegraphics[scale=0.4]{BR_TyI_sba099_cbap.png} \hspace{-3mm}
\includegraphics[scale=0.4]{BR_TyII_sba099_cbap.png}\hspace{-3mm}
\includegraphics[scale=0.4]{BR_TyX_sba099_cbap.png} \hspace{-3mm}
\includegraphics[scale=0.4]{BR_TyY_sba099_cbap.png}
\caption{Branching ratios as a function of $\tan\beta$
in the Type-I, Type-II, Type-X and Type-Y THDMs (from left to right) with $s_{\beta-\alpha} = 0.99$.
We take $c_{\beta-\alpha} < 0$ $(c_{\beta-\alpha} > 0)$ for the upper (lower) panels.
The values of $M^2$ and $m_\Phi (=m_H = m_A = m_{H^\pm})$ are scanned with the ranges of $0 \leq M^2 \leq m_\Phi^{2}$ and $300 \leq m_\Phi \leq 1000$ GeV.
The dashed lines show the predictions in the SM.}
\label{fig:br1}
\end{center}
\end{figure}
\subsection{Branching ratios \label{sec:br}}
\begin{table}[!t]
\begin{center}
{\renewcommand\arraystretch{1.2}
\begin{tabular}{cccccccc}\hline\hline
BR($h \to b\bar{b}$) & BR($h \to c\bar{c}$) & BR($h \to \tau\bar{\tau}$) & BR($h \to WW^*$) & BR($h \to ZZ^*$) \\\hline
59.5\% & 2.60\% & 7.16\% &20.3\% &2.47\% \\\hline\hline
\end{tabular}}
\caption{The SM predictions of the branching ratios of the Higgs boson $h$ at NLO. }
\label{tab:smbr}
\end{center}
\end{table}
We move on to the discussion of the branching ratios of the Higgs boson $h$ at NLO.
For reference, in Table~\ref{tab:smbr} we give our results for the branching ratios in the SM.
In the HSM and the IDM, the branching ratios for $h$ are almost the same as those in the SM predictions, because the partial decay rates are
universally suppressed by the radiative corrections and the mixing, where the latter does not happen in the IDM.
Thus, in the following discussion, we concentrate on the THDMs.
In the THDMs, branching ratios can be modified from those in the SM {by both tree-level mixing effects} parameterized by the scaling factor $\kappa_X^{}$ and loop effects.
When $s_{\beta-\alpha} \neq 1$ is taken, branching ratios can significantly be modified from the SM predictions due to the tree level mixing effects, and
the pattern of deviations strongly depends on the type of Yukawa interactions.
In this case, we may be able to determine the type from the pattern of deviations.
On the other hand, loop contributions to deviations in the branching ratios are relatively smaller than the tree level mixing effects, so that
it would be relatively difficult to extract the loop effects.
When $s_{\beta-\alpha} = 1$, however, the pure loop effect can be extracted, because the tree level mixing vanishes.
Therefore, in the following discussion, we first show the predictions of branching ratios in the THDMs with $s_{\beta-\alpha} \neq 1$ to see
how the mixing effects modify them.
We then show those with $s_{\beta-\alpha} = 1$ in order to extract the size of loop effects.
In Fig.~\ref{fig:br1}, we show the branching ratios as a function of $\tan\beta$ in four types of the THDMs
with $s_{\beta-\alpha} = 0.99$ and $c_{\beta-\alpha} < 0$ ($c_{\beta-\alpha} > 0$)
in the upper (lower) panels. The values of $M^2$ and $m_\Phi$ are scanned.
The typical behavior can be explained by the tree level results, see {e.g.} Ref.~\cite{Kanemura:2014bqa}.
For example, except for the Type-I THDM, some of the branching ratios fall down at $\tan\beta\simeq 7$ with $c_{\beta-\alpha} > 0$,
because of the fact that some $\kappa_f$ factors {become zero, e.g. $\kappa_b$} in the Type-II and Type-Y THDMs, and
it makes the value of the total width to be minimal as we saw it in the right panel of Fig.~\ref{fig:width_thdm}.
{Loop effects due to additional Higgs bosons} appear as a width of each curve.
In order to see the deviation in the ratio of the branching ratio from the SM prediction,
we introduce the following quantity for the $h \to XX$ mode
\begin{align}
\Delta\mu_{XX}^{} \equiv \frac{\text{BR}(h \to XX)_{\text{NP}}}{\text{BR}(h \to XX)_{\text{SM}}} -1. \label{delta_nlo}
\end{align}
Using the formulae of the partial decay rates at NLO discussed in Sec.~\ref{sec:decay},
$\Delta \mu_{XX}$ can be written in terms of the EW corrections $\overline{\Delta}_{\text{EW}}^X$ defined in Eq.~(\ref{eq:delbar}) in the alignment limit as
\begin{align}
\Delta \mu_{XX}^{} &\simeq \overline{\Delta}_{\text{EW}}^X- \sum_f\text{BR}^0(h \to f \bar{f})\overline{\Delta}_{\text{EW}}^f - \sum_V\text{BR}^0(h \to VV^*)\overline{\Delta}_{\text{EW}}^V , \label{muxx_app}
\end{align}
where $\text{BR}^0$ denotes the tree level branching ratio in the SM.
We note that the second term of the right hand side is dominantly determined by $\overline{\Delta}_{\text{EW}}^b$ because the branching ratio of $h \to b\bar{b}$ typically has the
largest value among all the decay modes.
This expression is helpful to understand the behavior of some plots which will be shown below.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.65]{mbHdep_TyI_BR_ff_tb1p5.png} \hspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyII_BR_ff_tb1p5.png} \\
\includegraphics[scale=0.65]{mbHdep_TyX_BR_ff_tb1p5.png} \hspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyY_BR_ff_tb1p5.png}
\caption{Predictions of $\Delta\mu_{ff}$ ($f=b,c,\tau$) defined in Eq.~(\ref{delta_nlo})
in the Type-I (upper-left), Type-II (upper-right), Type-X (lower-left) and Type-Y (lower-right) THDMs with $s_{\beta-\alpha} = 1$ and
$\tan\beta = 1.5$ as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$.
The solid and dashed curves show the case with $\lambda v^2 = 0$ and $(200~\text{GeV})^2$, respectively, where $\lambda v^2 = m_\Phi^2 - M^2$. }
\label{fig:dmu3}
\end{center}
\end{figure}
In Fig.~\ref{fig:dmu3}, we show $\Delta\mu_{ff}$ ($f = b,c,\tau$) as a function of $m_\Phi$ in four types of the THDMs with $s_{\beta-\alpha} = 1$ and $\tan\beta = 1.5$.
{Here, we fix the value of $\lambda v^2 = m_\Phi^2 - M^2$ to be 0} (solid curves) and (200 GeV)$^2$ (dashed curves).
We see the decoupling behavior in the large mass region, and observe the peak at around $m_\Phi = 2m_t$ depending on the type of Yukawa interaction and the type of fermions, where
the direction of the peak is the same as that for the plots of $\overline{\Delta}_{\rm EW}^f$ shown in Fig.~\ref{fig:delbct}.
Notice that the peak appearing at $m_\Phi > 2m_t$ in Fig.~\ref{fig:delbct} does not appear in this plot, as we here fix the value of $\lambda v^2$.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.65]{mbHdep_TyI_BR_ff_tb3.png} \hspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyII_BR_ff_tb3.png} \\
\includegraphics[scale=0.65]{mbHdep_TyX_BR_ff_tb3.png} \hspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyY_BR_ff_tb3.png}
\caption{Same as Fig.~\ref{fig:dmu3}, but for $\tan\beta = 3$. }
\label{fig:dmu4}
\end{center}
\end{figure}
Fig.~\ref{fig:dmu4} shows the case for $\tan\beta = 3$, and all the other choices are the same as in Fig.~\ref{fig:dmu3}.
We see that some of $\Delta\mu_{ff}$ values with $\lambda v^2 = (200~\text{GeV})^2$ are largely different from {those} with $\lambda v^2 = 0$.
For example, the dashed curve for $\Delta\mu_{\tau\tau}$ (green) in the Type-X THDM is located lower than the corresponding solid curve
because $\overline{\Delta}_{\rm EW}^\tau$ appearing in the first term of Eq.~(\ref{muxx_app}) has a smaller value in the case with $\lambda v^2 = (200~\text{GeV})^2$.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.65]{mbHdep_TyI_BRVV_tb1p5.png} \hspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyII_BRVV_tb1p5.png}
\caption{Predictions of $\Delta\mu_{VV}$ ($V=W,Z$) defined in Eq.~(\ref{delta_nlo})
in the Type-I (left) and Type-II (right) THDMs with $s_{\beta-\alpha} = 1$ and $\tan\beta = 1.5$ as a function of $m_\Phi (=m_H = m_A = m_{H^\pm})$.
The solid and dashed curves show the case with $\lambda v^2 = 0$ and $(200~\text{GeV})^2$, where $\lambda v^2 = m_\Phi^2 - M^2$. }
\label{fig:dmu1}
\vspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyI_BRVV_tb3.png} \hspace{3mm}
\includegraphics[scale=0.65]{mbHdep_TyII_BRVV_tb3.png}
\caption{Same as Fig.~\ref{fig:dmu1}, but for $\tan\beta=3$. }
\label{fig:dmu11}
\end{center}
\end{figure}
The $m_\Phi$ dependence of $\Delta\mu_{VV}$ $(V=W,Z)$ is shown in the THDMs with $s_{\beta-\alpha} = 1$ and $\tan\beta = 1.5$ (Fig.~\ref{fig:dmu1}) and 3 (Fig.~\ref{fig:dmu11}),
where the values of $\Delta\mu_{WW}$ and $\Delta\mu_{ZZ}$ are almost the same {of} each other in this case.
As in Fig.~\ref{fig:dmu3}, the value of $\lambda v^2$ is fixed to 0 (solid curves) and (200 GeV)$^2$ (dashed curves).
The left and right panels show the results in the Type-I and Type-II THDMs, respectively, while the results in the Type-X (Type-Y) THDM
are almost the same as those in the Type-I (Type-II) THDM.
In all the panels, the value of $\Delta \mu_{VV}$ approaches to zero in the large $m_\Phi$ region, because of the decoupling property of the additional Higgs bosons.
In the left (right) panel, a peak appears at around $m_\Phi = 2m_t$, because the EW correction to the partial width of $h \to b\bar{b}$ mode has a peak
in the Type-I and Type-X (Type-II and Type-Y) THDMs, see Figs.~\ref{fig:delb-1} and \ref{fig2}.
We can also see that in the Type-I THDM with $\tan\beta = 1.5$
the value of $\Delta\mu_{VV}$ with $\lambda v^2=(200~\text{GeV})^2$
is almost the same as that with $\lambda v^2 = 0$, because the change of $\overline{\Delta}_{\text{EW}}^b$ due to taking different values of $\lambda v^2$ is
accidentally cancelled by that of $\overline{\Delta}_{\text{EW}}^V$.
For $\tan\beta = 3$,
the value of $\Delta\mu_{VV}$ with $\lambda v^2=(200~\text{GeV})^2$ is slightly smaller than that with $\lambda v^2=0$, because the change
of $\overline{\Delta}_{\text{EW}}^b$ becomes smaller than that for $\tan\beta = 1.5$, while $\overline{\Delta}_{\text{EW}}^V$ does not depend on $\tan\beta$ so much.
On the other hand, in the Type-II THDM with $\tan\beta = 3$
the value of $\Delta\mu_{VV}$ with $\lambda v^2=(200~\text{GeV})^2$ is larger than that for $\lambda v^2=0$,
because of the larger negative shift of $\overline{\Delta}_{\text{EW}}^b$.
\subsection{Correlations \label{sec:correlations}}
\begin{table}[!t]
\begin{center}
{\renewcommand\arraystretch{1.2}
\begin{tabular}{cccccccc}\hline\hline
$h \to b\bar{b}$ & $h \to c\bar{c}$ & $h \to \tau\bar{\tau}$ & $h \to WW^*$ & $h \to ZZ^*$ & $h \to gg$ & $h \to \gamma \gamma$& $h \to \mu\mu$\\\hline
0.89\% & 3.2\% & 1.4\% & 1.9\% & 6.7\% & 2.7\% & 13\% & 27\% \\\hline\hline
\end{tabular}}
\caption{Expected 1$\sigma$ uncertainty for the measurements of the branching ratios of the Higgs boson $h$ at ILC250~~\cite{Fujii:2017vwa}. }
\label{tab:br-ilc}
\end{center}
\end{table}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_bb_muc0_mud2p0.png} \hspace{3mm}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_bb_muc0_mud4p0.png}
\caption{Correlation between $\Delta\mu_{\tau\tau}^{}$ and $\Delta\mu_{bb}^{}$ in the Type-I (red), Type-II (blue),
Type-X (green), Type-Y (magenta) THDMs, the HSM (orange) and the IDM (black).
The left (right) panel shows the case with $\Delta\mu_{WW}=0 \pm 2\%$ $(0 \pm 4\%)$.
In the THDMs, we scan $1.5 \leq \tan\beta \leq 10$, $0\leq M^2 \leq m_\Phi^2$ and $300~(600)\leq m_\Phi \leq 1000$ GeV for (darker) colored points.
In the HSM, we scan $300 \leq m_H \leq 5000$ GeV and $0\leq M^2 \leq m_H^2$,
while in the IDM we fix $m_H$ to be 63 GeV and scan $100 \leq m_A^{}(=m_{H^\pm}^{}) \leq 1000$ GeV. }
\label{fig:corr1}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_bb_muc5p0_mud4p0.png} \hspace{3mm}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_bb_mucm5p0_mud4p0.png}
\caption{Same as Fig.~\ref{fig:corr1}, but for the case with
$\Delta\mu_{WW}$ to be $5 \pm 4\%$ (left) and $-5 \pm 4\%$ (right). }
\label{fig:corr2}
\end{center}
\end{figure}
In the discussions so far, we have seen the deviation in the total width and those in branching ratios in each extended Higgs sector.
We now see correlations among the deviations in branching ratios in all the extended Higgs sectors discussed in this paper in order to clarify how we can distinguish
extended Higgs sectors from the precise measurements of the branching ratios.
The branching ratios of the Higgs boson will be measured as accurately as possible at future collider experiments.
In particular at the ILC, we can measure the cross section of $e^+e^- \to Zh$ without depending on the decay of $h$ by using the recoil method~\cite{Yan:2016xyx,Asner:2013psa}.
This makes the measurements of the branching ratios possible without depending on the cross section.
At the ILC with the collision energy of 250 GeV
and the integrated luminosity of 2 ab$^{-1}$ (ILC250), the branching ratios are expected to be measured as shown in Table~\ref{tab:br-ilc}.
We thus consider the situation where the branching ratios are measured to some extent at ILC250,
namely we impose the further constraint on the value of $\Delta \mu_{XX}$ with a given central value and an error in addition to the theoretical constraints which are imposed in the discussion above.
In order to take into account the constraints from $B_s \to X_s \gamma$~\cite{Misiak:2017bgg,Haller:2018nnx},
we consider the case with larger masses of extra Higgs bosons, i.e, $m_\Phi \geq 600$ GeV
in the THDMs as well as that for $m_\Phi \geq 300$ GeV.
As discussed in Sec.~\ref{sec:model}, the lower bound on $m_{H^\pm}^{}$ is about 600 GeV in the Type-II and Type-Y THDMs.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_cc_muc0_mud4p0.png} \\
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_cc_muc5p0_mud4p0.png} \hspace{3mm}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_cc_mucm5p0_mud4p0.png}
\caption{Correlation between $\Delta\mu_{\tau\tau}^{}$ and {$\Delta\mu_{cc}^{}$} in the Type-I (red),
Type-II (blue), Type-X (green) and Type-Y (magenta) THDMs. The upper panel shows the case with
$\Delta\mu_{WW}^{}=0\pm 4\%$, while the lower left (right) panel shows the case with
$\Delta\mu_{WW}^{} = 5 \pm 4\%$ ( $\Delta\mu_{WW}^{} = -5 \pm 4\%$).
The ranges of scanning parameters are the same as those of Fig.~\ref{fig:corr1}.}
\label{fig:corr3}
\end{center}
\end{figure}
In Fig.~\ref{fig:corr1}, we show the correlation between $\Delta\mu_{\tau\tau}^{}$ and $\Delta\mu_{bb}^{}$ in the HSM, four types of the THDMs and the IDM under the additional constraint from $\Delta \mu_{WW} = 0\pm 2\%$ (left) and $\Delta \mu_{WW} = 0\pm 4\%$ (right).
The errors of 2\% and 4\% are taken to consider about 1$\sigma$ and 2$\sigma$ level region at ILC250, respectively~\cite{Fujii:2017vwa}.
Parameters of each model are scanned as it is described in the caption.
We see that the predictions in the HSM and the IDM are given almost at the origin of this plane, because in these models the partial decay rates are almost universally suppressed as we already mentioned in Sec.~\ref{sec:br}.
On the other hand, the predictions of the THDMs are spread out into the different directions depending on the type of Yukawa interactions.
In the Type-Y THDM, two allowed regions appear in the fourth quadrant {if we choose values for $m_\Phi$ starting from $300$ GeV} (lighter points).
This can be explained from Fig.~\ref{fig:br1} (lower-rightmost panel), where we can find the two values of $\tan\beta$ providing the same value of BR$(h \to \tau\bar{\tau})$ while different values of BR$(h \to b\bar{b})$.
This, however, vanishes when we take $m_\Phi \geq 600$ GeV (darker points), as such configuration is favored by the flavor experiments, particularly in the Type-II and Type-Y THDMs.
Remarkably, only in the Type-X THDM, the allowed points are distributed in the wide (small) range of $\Delta\mu_{\tau\tau}$ $(\Delta\mu_{bb})$.
This can also be understood from the third panels of Fig.~\ref{fig:br1}, where only the BR($h \to \tau\bar{\tau}$) mode can significantly be changed depending on $\tan\beta$, while all the other branching ratios
do not change so much.
{In contrast in the Type-II THDM large variations of BR$(h\to\tau\tau)$ appear together with large variations of BR$(h\to bb)$.}
Thus, the other decay branching ratios, particularly the $h \to WW^*$, also strongly vary at the same time, {and then
such configurations are constrained by the bound on $\Delta \mu_{WW}$. }
From this figure, we find that if $\Delta \mu_{\tau\tau}$ is found to be a several percent, we can distinguish the models considered in this paper.
In the following discussion, we focus on the case with $\Delta \mu_{WW}$ to be constrained at the 2$\sigma$ level{,} i.e., allowing 4\% uncertainty.
Let us also consider the case where the central value of $\Delta\mu_{WW}$ is found to be nonzero, and $\Delta\mu_{WW} = 0$ is excluded {at} the 2$\sigma$ level.
In Fig.~\ref{fig:corr2}, we show {such situations with $\Delta\mu_{WW} = 5.0\pm 4.0\%$ (left) and $\Delta\mu_{WW} = -5.0\pm 4.0\%$ (right).}
In this setup, predictions of $\Delta\mu_{WW}$ in the HSM and the IDM are almost zero, so that these models are
excluded, while four types of THDMs can explain such a deviation.
If the value of $\Delta\mu_{WW}$ is given to be $5.0 \pm 4.0\%$ (left), then four types of Yukawa interactions are {well separated of one another}, so
that we can determine the type by measuring $\Delta\mu_{\tau\tau}$ and $\Delta\mu_{bb}$ in addition to $\Delta\mu_{WW}$.
We note that the positive value of $\Delta\mu_{WW}$ essentially comes from the reduction of the other decay rates, especially the $h \to b\bar{b}$ mode,
because the partial width of $h \to WW^*$ reduces by the tree level scaling factor $\kappa_V^{} \leq 1$ and the one-loop effect as seen from Fig.~\ref{fig:delw}.
In fact, in the left panel of Fig.~\ref{fig:corr2}, the allowed points mainly appear in the regions with
$\Delta\mu_{bb} < 0$.
For the case with $\Delta\mu_{WW} = -5.0\pm 4.0\%$, some of the THDMs are overlapped in this plane.
This is because the negative value of $\Delta\mu_{WW}$ can be explained by either decreasing the partial width of the $h \to WW^*$ mode
or increasing the other partial widths.
Therefore, the branching ratio of $h \to b\bar{b}$ can be either larger or smaller than the SM prediction, as seen in the right panel.
This makes discrimination among the four types of Yukawa interactions difficult as compared to the case with positive $\Delta \mu_{WW}$.
In Fig.~\ref{fig:corr3}, we show the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$ under the constraint on $\Delta\mu_{WW}$ with the 4\% uncertainty.
The central value of $\Delta\mu_{WW}$ is supposed to be 0 in the upper panel, and to be $+5\%$ and $-5\%$ in the lower left and right panels, respectively.
As compared to Fig.~\ref{fig:corr1} (right), the allowed points on the upper panel are widely distributed in the $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$ plane, because
the branching ratio of $h \to c\bar{c}$ typically has a smaller portion of the total width than that of $h \to b\bar{b}$.
If we only look at the correlation shown in the upper panel, it seems difficult to distinguish the models unless $\Delta\mu_{\tau\tau}$ is given to several percent level.
However, we would like to emphasize that by looking also at the corresponding plot shown in Fig.~\ref{fig:corr1} (right),
we can distinguish the models.
For example, if $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$ are measured {to be small negative and positive respectively, i.e., the second quadrant in this plane,
both the Type-II and Type-X THDMs} can explain such situation, but these two models may not be distinguished {from} each other.
However, by looking at the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{bb}$ with a negative value of $\Delta \mu_{\tau\tau}$, the Type-II and Type-X THDMs have allowed points {in the different directions compared to each other}.
Therefore, we can distinguish all the four THDMs by using the combination of these correlations even if the central value of $\Delta \mu_{WW}$ is measured to be close to zero.
Similar to Fig.~\ref{fig:corr2}, we show the case for nonzero $\Delta \mu_{WW}$ at the 2$\sigma$ level in the lower two plots in Fig.~\ref{fig:corr3}.
We see that for the case with the central value of $\Delta \mu_{WW}$ to be $+5\%$,
four types of THDMs are clearly {separated from one another}, while the case with the central value of $\Delta \mu_{WW}$ to be $-5\%$
two models are overlapping at some regions of this plane.
However, again the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{bb}$ helps for further discrimination of the models.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_WW_muc0p0_mud2p0_bb.png} \\
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_WW_muc2p5_mud2p0_bb.png} \hspace{3mm}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_WW_mucm2p5_mud2p0_bb.png}
\caption{Correlations between $\Delta\mu_{\tau\tau}^{}$ and $\Delta\mu_{WW}^{}$ in the Type-I (red), Type-II (blue),
Type-X (green) and Type-Y (magenta) THDMs.
The upper panel shows the case with $\Delta\mu_{bb}=0 \pm 2\%$, while the lower-left (right) panel shows the
case with $\Delta\mu_{bb}=2.5 \pm 2\%$ ($\Delta\mu_{bb}=-2.5 \pm 2\%$).
The values of $\tan\beta$, $M^2$ and $m_\Phi (=m_H = m_A = m_{H^\pm})$ are scanned with the ranges of
$1.5 \leq \tan\beta \leq 10$, $0\leq M^2 \leq m_\Phi^2$ and $300 \leq m_\Phi \leq 1000$ GeV, respectively,
under the constraints from the perturbative unitarity, the vacuum stability and the $S$, $T$ parameters. }
\label{fig:corr4}
\end{center}
\end{figure}
In the above analyses, we constrained the value of $\Delta \mu_{WW}$.
We now constrain $\Delta \mu_{bb}$ instead of $\Delta \mu_{WW}$.
In Fig.~\ref{fig:corr4}, we show the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{WW}$ under the constraint on
$\Delta \mu_{bb} = 0\pm 2\%$ (upper panel),
$\Delta \mu_{bb} = 2.5\pm 2\%$ (lower left panel) and
$\Delta \mu_{bb} = -2.5\pm 2\%$ (lower right panel).
The error of 2\% is taken to consider about the 2$\sigma$ level region at ILC250.
For the case with $\Delta \mu_{bb} = 0\pm 2\%$, if $\Delta \mu_{\tau\tau}$ is measured to be a few percent level,
it might be difficult to distinguish the models shown in this figure, particularly for the case with $m_\Phi > 600$ GeV shown as darker points.
If we consider the case with $300 < m_\Phi < 600$ GeV, it is seen that the points with $-10\% \lesssim \Delta \mu_{WW} \lesssim -5\%$ are also allowed in
all the four types of THDMs.
When we consider the case with $\Delta \mu_{bb} = 2.5\pm 2\%$, the situation is drastically changed, where most of the allowed points
are distributed in the region with $\Delta \mu_{WW} <0$, because of the compensation of the positive deviation in the branching ratio of the $h \to b\bar{b}$ mode.
The opposite situation can be seen in the lower right panel showing the case with $\Delta \mu_{bb} = -2.5\pm 2\%$.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_cc_mucm0p0_mud2p0_bb.png} \\
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_cc_muc2p5_mud2p0_bb.png} \hspace{3mm}
\includegraphics[scale=0.75]{NLO_BR_dev_tautau_cc_mucm2p5_mud2p0_bb.png}
\caption{Same as Fig.~\ref{fig:corr4}, but for the correlation between $\Delta\mu_{\tau\tau}$ and $\Delta\mu_{cc}$.
}
\label{fig:corr5}
\end{center}
\end{figure}
Finally, we show the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$ in Fig.~\ref{fig:corr5} using {the same setup as in} Fig.~\ref{fig:corr4}.
The shape of the upper panel looks similar to that seen in the upper panel of Fig.~\ref{fig:corr3}, while the allowed points in Fig.~\ref{fig:corr5} are distributed in
smaller regions than those shown in Fig.~\ref{fig:corr3}.
This is simply because the foreseen uncertainty of the measurements of $\Delta \mu_{bb}$ is smaller as compared to that of $\Delta \mu_{WW}$.
Interestingly, for both the cases of $\Delta\mu_{bb}= 2.5 \pm 2\%$ (left) and $\Delta\mu_{bb}= -2.5 \pm 2\%$ (right)
four types of the THDMs are well separated, and it becomes clearer when $m_\Phi$ is taken to be greater than 600 GeV.
In this subsection, we have discussed various correlations between the deviations in branching ratios from the SM predictions at NLO.
First, if we observe a percent level deviation in one of the decay modes of $h$,
then the HSM and IDM could be ruled out as the branching ratios are almost the same as the SM predictions in these models.
Second, the discrimination of four types of the THDMs strongly depends on the situation.
If we observe a positive deviation in the branching ratio for the $h \to WW^*$ mode,
the discrimination is possible by looking at the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{bb}$ or
$\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$.
On the contrary, if we measure a negative deviation in the branching ratio of the $h \to WW^*$ mode, then
the discrimination of four types becomes more complicated as two of four models can overlap with each other
in the correlation of $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{bb}$ or
$\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$.
However, using three observables $\Delta \mu_{\tau\tau}$, $\Delta \mu_{bb}$ and $\Delta \mu_{cc}$ with the results from the direct searches of
additional scalar bosons and from flavor experiments,
we may be able to separate four types of the THDMs.
\section{Conclusions\label{sec:conc}}
We have discussed the total width and the branching ratios of the 125 GeV Higgs boson $h$ at NLO in EW and QCD
in the HSM, four types of the THDMs and the IDM.
These quantities can be measured at collider experiments as precisely as possible under a given machine performance.
Thus, accurate calculations for the total width and branching ratios are quite important to compare them with future precision data{,} e.g. at the HL-LHC and the ILC.
For the one-loop computation, we systematically applied the on-shell renormalization scheme for each model, in which
we adopted the {\tt H-COUP} program to evaluate numerical values of the renormalized Higgs boson vertices.
The analytic expressions for the decay rates of $h \to f\bar{f}$, $h \to ZZ^* \to Zf\bar{f}$ and $h \to WW^* \to Wf\bar{f}'$ are
presented at NLO, among which the $h \to WW^* \to Wf\bar{f}'$ mode is newly computed in this paper.
We also provided the decay rates of the loop induced processes; i.e., $h \to \gamma\gamma$, $h \to Z\gamma$ and $h \to gg$ with QCD corrections at NLO.
We have shown that in the HSM and the IDM, all the partial decay rates are almost universally suppressed by both the tree level mixing (for the HSM) and the one-loop effect,
so that the branching ratios remain almost the same values as those in the SM.
Thus, if deviations in the branching ratio from the {SM prediction (denoted as $\Delta \mu_{XX}^{}$ for the decay $h\to XX$)} are found, then we may be able to exclude the HSM and IDM.
On the contrary, when we observe the deviation in the total width but not in the branching ratios, then it could be a smoking gun signature to identify these two models.
We also have found that if we observe a positive deviation in the branching ratio of the $h \to WW^*$ mode, four types of the THDMs
can be well {separated from one another} from the correlation between $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{bb}$ or
$\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$.
If we observe a negative deviation in the branching ratio of the $h \to WW^*$ mode,
some of the THDMs can overlap in the $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{bb}$ plane, but this can be disentangled by further looking at another correlation, such as $\Delta \mu_{\tau\tau}$ and $\Delta \mu_{cc}$.
While the branching ratios are measured with a percent level at future precision experiments,
direct searches for additional Higgs bosons are expected to make progress at the LHC Run-III and the HL-LHC.
If additional Higgs bosons are discovered, we can give stronger predictions of the correlation among the branching ratios
by using their masses as inputs.
Even if additional Higgs bosons are not directly discovered, stronger mass bounds obtained from the direct searches provide narrower allowed regions
in the correlations.
On the other hand, if deviations in the Higgs boson couplings, widths and/or branching ratios from the SM predictions are found at future precision experiments,
we can obtain upper limits on masses of additional Higgs bosons; see e.g., Figs.~\ref{fig:width_hsm}, \ref{fig:width_thdm2} and \ref{fig:idm}.
Therefore, indirect searches for extended Higgs models using deviations in the Higgs boson properties play the complementary role to the direct searches as well as flavor constraints.
{Using the synergy between the direct and indirect searches, the parameter space of {extended Higgs models} can be effectively narrow down. }
Finally, we would like to mention
that the {\tt H-COUP} version 2.0, where all the NLO computations for the decay rates presented
in this paper are implemented, will be publicly available in near future~\cite{future}.
\vspace*{-4mm}
\begin{acknowledgments}
This work is supported in part by the Grant-in-Aid on Innovative Areas, the Ministry of Education, Culture,
Sports, Science and Technology, No.~16H06492 and No.~18H04587, and Grant H2020-MSCA-RISE-2014 No.~645722 (Non-Minimal Higgs) [S.K.],
JSPS KAKENHI Grant No.~18K03648 [K.M.],
JSPS KAKENHI Grant No.~18J12866 [K.S.] and Early-Career Scientists, No.~19K14714 [K.Y.].
\end{acknowledgments}
\begin{appendix}
\section{Loop induced $h\gamma\gamma$, $hZ\gamma$, and $hgg$ vertices \label{sec:hgg}}
In this section, we give analytic expressions of loop induced vertices{,} i.e., $h\gamma\gamma$, $hZ\gamma $ and $hgg$ at one-loop level,
which are required to calculate not only the decay rates of the loop induced processes given in Eq.~(\ref{eq:loop-induced}) but also those of $h\to ZZ^* \to Vf\bar{f}$ at NLO given in Eq.~(\ref{eq:del_ew_z}).
{Here, we} present the formulae for the THDMs and the IDM.
Those for the HSM are simply obtained by removing the charged scalar loop contribution denoted as
$\hat{\Gamma}_{h\gamma\gamma/hZ\gamma}^i(p_1^2,p_2^2,q^2)_S$.
The loop induced $h\gamma\gamma$ and $hZ\gamma$ vertices can be decomposed into the contribution from charged scalar loops, fermion loops and weak boson loops as follows:
\begin{align}
\hat{\Gamma}_{h\gamma\gamma}^i(p_1^2,p_2^2,q^2) &=
\hat{\Gamma}_{h\gamma\gamma}^i(p_1^2,p_2^2,q^2)_S
+ \sum_F\kappa_F \hat{\Gamma}_{h\gamma\gamma}^i(p_1^2,p_2^2,q^2)_F
+ \kappa_V^{} \hat{\Gamma}_{h\gamma\gamma}^i(p_1^2,p_2^2,q^2)_V, \\
\hat{\Gamma}_{hZ\gamma}^i(p_1^2,p_2^2,q^2) &=
\hat{\Gamma}_{hZ\gamma}^i(p_1^2,p_2^2,q^2)_S
+ \sum_F\kappa_F\hat{\Gamma}_{hZ\gamma}^i(p_1^2,p_2^2,q^2)_F
+ \kappa_V^{}\hat{\Gamma}_{hZ\gamma}^i(p_1^2,p_2^2,q^2)_V.
\end{align}
The analytic expressions for each contribution to the $h Z \gamma$ vertex are given by
\begin{align}
\hat{\Gamma}_{hZ\gamma}^1(p_1^2,p_2^2,q^2)_S &= -\frac{eg_Z^{}}{16\pi^2}(c_W^2-s_W^2)\lambda_{hH^+H^-}[4C_{24}(m_{H^\pm}^{}) - B_0(q^2; m_{H^\pm},m_{H^\pm})], \\
\hat{\Gamma}_{hZ\gamma}^2(p_1^2,p_2^2,q^2)_S &= -\frac{4eg_Z^{}}{16\pi^2}(c_W^2-s_W^2)\lambda_{hH^+H^-}q^2C_{1223}(m_{H^\pm}^{}), \\
\hat{\Gamma}_{hZ\gamma}^1(p_1^2,p_2^2,q^2)_F &= -\frac{4eg_Z^{}}{16\pi^2}\frac{m_F^2}{v}N_c^Fv_FQ_F [8C_{24}(m_F^{})-2B_0(q^2;m_F^{},m_F^{}) \notag\\
&\hspace{3cm} + (p_1^2+p_2^2 -q^2)C_0(m_F^{}) ], \\
\hat{\Gamma}_{hZ\gamma}^2(p_1^2,p_2^2,q^2)_F &= -\frac{8eg_Z^{}}{16\pi^2}\frac{m_F^{2}}{v}N_c^Fv_FQ_Fq^2\left[C_0(m_F^{}) + 4C_{1223}(m_F^{}) \right], \\
\hat{\Gamma}_{hZ\gamma}^1(p_1^2,p_2^2,q^2)_V&=\frac{2eg_Z^{}}{16\pi^2}\frac{m_W^2}{v}\Big\{
c_W^2\Big[3(m_W^2 +p_1^2 + 2p_2^2-2q^2)C_0(m_W) \notag\\
&+5[4C_{24}(m_W) - B_0(q^2;m_W,m_W)] +2B_0(p_2^2;m_W,m_W)-2B_0(0;m_W,m_W)\Big]\notag\\
& + s_W^2[(3m_W^2 -2p_1^2 + p_2^2 +q^2)C_0(m_W^{}) +2B_0(p_2^2;m_W^{},m_W^{}) \notag\\
& -2B_0(0;m_W^{},m_W^{})+ B_0(q^2;m_W^{},m_W^{}) - 4C_{24}(m_W^{})]\notag\\
&-\frac{m_h^2(c_W^2-s_W^2)}{2m_W^2}[B_0(q^2,m_W^{},m_W^{})-4C_{24}(m_W^{})] +m_h^2s_W^2C_0(m_W^{}) \Big\}, \label{jjj} \\
\hat{\Gamma}_{hZ\gamma}^2(p_1^2,p_2^2,q^2)_V&=\frac{8eg_Z^{}}{16\pi^2}\frac{m_W^2}{v} q^2\Big\{
c_W^2[3C_0(m_W^{}) + 5C_{1223}(m_W^{})] - s_W^2[C_0(m_W^{}) + C_{1223}(m_W^{})] \notag\\
& + \frac{m_h^2(c_W^2-s_W^2)}{2m_W^2} C_{1223}(m_W^{}) \Big\}, \label{kkk}
\end{align}
where $B_i$, $C_{i}$ and $C_{ij}$ are the Passarino-Veltman's functions~\cite{Passarino:1978jh}. In this paper,
we follow the convention of these functions given in Ref.~\cite{Kanemura:2015mxa}.
{Here, we} use the shorthand notation for the $C$ functions defined by $C_{i,ij}(m) \equiv C_{i,ij}(p_1^2,p_2^2,q^2;m,m,m)$ and
$C_{1223} \equiv C_{12} + C_{23}$.
The form factors for the renormalized $h\gamma\gamma$ vertex is obtained from the expressions of the $h Z\gamma$ vertex by the replacement of $(g_Z^{},c_W^2,s_W^2,v_F) \to (e,1,-1,Q_F)$.
Finally, the $hgg$ vertex is induced only from the quark loop. Thus, it is expressed by
\begin{align}
\hat{\Gamma}_{hgg}^1(p_1^2,p_2^2,q^2) &= -\frac{\alpha_s}{\pi}\frac{m_q^2}{v} \left[8C_{24}(m_q)-2B_0(q^2;m_q,m_q) + (p_1^2+p_2^2 -q^2)C_0(m_q) \right]\delta^{ab}, \\
\hat{\Gamma}_{hgg}^2(p_1^2,p_2^2,q^2) &= -\frac{\alpha_s}{\pi}\frac{m_q^2}{v} \left[C_0(m_q) + 4C_{1223}(m_q) \right]\delta^{ab},
\end{align}
where $a$ and $b$ represent the color index.
\section{Renormalized $Vf\bar{f}$ vertices \label{sec:vff}}
We give analytic formulae of the renormalized $Vf\bar{f}$ ($V=Z,W$) vertices, which appear in the decay rates of $h\to VV^* \to Vf\bar{f}$ at NLO
given in Eq.~(\ref{eq:del_ew_z}) and (\ref{eq:del_ew_w}).
{In the limit of massless external fermions,} expressions of these vertices are common to those in the SM.
The renormalized $Vf\bar{f}$ ($V=Z$, $W$) vertices can be decomposed in the massless limit for external fermions as
\begin{align}
\hat{\Gamma}_{Vff}^\mu(p_1^2,p_2^2,q^2)&= g_V^{}\gamma^\mu (\hat{\Gamma}_{Vff}^V - \gamma_5 \hat{\Gamma}_{Vff}^A),
\end{align}
where $p_1^\mu$ $(p_2^\mu)$ is the incoming four-momentum of the fermion (anti-fermion),
and $q^\mu $ is the outgoing four-momentum of the gauge boson.
The gauge coupling $g_V^{}$ is $g_Z$ and $g/\sqrt{2}$ for $Z$ and $W$, respectively.
Similar to Eqs.~(\ref{eq:form-loop}) and (\ref{gam_hff_loop}), we can further decompose these vertices into the tree level part and 1-loop part:
\begin{align}
\hat{\Gamma}_{Vff}^i = \Gamma_{Vff}^{i,\text{tree}} + \Gamma_{Vff}^{i,\text{loop}},~~\text{with}~~
\Gamma_{Vff}^{i,\text{loop}} = \Gamma_{Vff}^{i,\text{1PI}} + \delta \Gamma_{Vff}^i,~~(i=V,A).
\end{align}
The tree level contribution is given by
\begin{align}
&\Gamma^{V,{\rm tree}}_{Zff} = v_f = \frac{I_f}{2} -s_W^2 Q_f,\quad
\Gamma^{A,{\rm tree}}_{Zff} = a_f = \frac{I_f}{2}, \quad
\Gamma^{V,{\rm tree}}_{Wff} = \Gamma^{A,{\rm tree}}_{Wff} = \frac{1}{2}. \label{eq:vff_tree}
\end{align}
The counterterm contribution is determined by imposing the on-shell renormalization condition as
\begin{align}
\delta \Gamma_{Zff}^V &= \frac{1}{16\pi^2}
\Big[ e^2Q_f^2v_f(2B_1 + 1)(m_f^2;m_f,m_\gamma)
+g_Z^2v_f(v_f^2 + 3a_f^2)(2B_1 + 1)(m_f^2;m_f,m_Z) \notag\\
& +\frac{g^2}{4}(v_f+a_f)(2B_1 + 1)(m_f^2;m_{f'},m_W)
-g^2I_fc_W^2B_0(0;m_W,m_W)\Big], \label{delgamv} \\
\delta \Gamma_{Zff}^A &= \frac{1}{16\pi^2}
\Big[ e^2Q_f^2a_f(2B_1 + 1)(m_f^2;m_f,m_\gamma)
+g_Z^2v_f(3v_f^2 + a_f^2)(2B_1 + 1)(m_f^2;m_f,m_Z) \notag\\
& +\frac{g^2}{4}(v_f+a_f)(2B_1 + 1)(m_f^2;m_{f'},m_W)
-g^2I_fc_W^2B_0(0;m_W,m_W)\Big], \label{delgama} \\
\delta \Gamma_{Wff}^V &= \delta \Gamma_{Wff}^A
= \frac{1}{16\pi^2}\Big[\frac{e^2}{4}Q_f^2(2B_1 + 1)(m_f^2;m_f,m_\gamma) + \frac{g_Z^2}{4}(v_f + a_f)^2(2B_1 + 1)(m_f^2;m_f,m_\gamma)\notag\\
&+ \frac{g^2}{8}(2B_1 + 1)(m_f^2;m_{f'},m_W) \Big] + (f \leftrightarrow f') -\frac{g^2}{16\pi^2}B_0(0;m_W,m_W),
\end{align}
where $m_f = m_{f'} = 0$.
The 1PI diagram contributions to these vertices are calculated as
\begin{align}
\Gamma_{Zff}^{V,\text{1PI}} = \frac{1}{16\pi^2} &\Big[v_fe^2 Q_f^2 F_{FVF}(f,\gamma,f) + g_Z^2v_f(v_f^2 + 3a_f^2)F_{FVF}(f,Z,f) \notag\\
& + \frac{g^2}{4}(v_{f'} + a_{f'})F_{FVF}(f',W,f')
+ I_f g^2c_W^2F_{VFV}(W,f',W) \Big], \\
\Gamma_{Zff}^{A,\text{1PI}} = \frac{1}{16\pi^2} & \Big[a_fe^2 Q_f^2 F_{FVF}(f,\gamma,f) + g_Z^2a_f(3v_f^2 + a_f^2)F_{FVF}(f,Z,f) \notag\\
&+ \frac{g^2}{4}(v_{f'} + a_{f'})F_{FVF}(f',W,f')
+ I_f g^2c_W^2F_{VFV}(W,f',W) \Big], \\
\Gamma_{Wff}^{V,\text{1PI}} = \frac{1}{16\pi^2} &\Big\{\frac{e^2}{2}Q_fQ_{f'}F_{FVF}(f,\gamma,f') + \frac{g_Z^2}{2}(v_f + a_f)(v_{f'} + a_{f'})F_{FVF}(f,Z,f') \notag\\
& + \frac{g^2}{2}[F_{VFV}(W,f',Z) + F_{VFV}(Z,f,W)]\notag\\
& + 2e^2 Q_fI_f[F_{VFV}(\gamma,f,W) - F_{VFV}(Z,f,W)] \notag\\
& + 2e^2 Q_{f'}I_{f'}[F_{VFV}(W,f',\gamma) - F_{VFV}(W,f',Z)] \Big\}, \\
\Gamma_{Wff}^{A,\text{1PI}} = \Gamma_{Wff}^{V,\text{1PI}},
\end{align}
where
\begin{align}
F_{FVF}(X,Y,Z) &= 2q^2[C_{11} + C_{23}](p_1^2,p_2^2,q^2;m_X,m_Y,m_Z) \notag\\
&+ 4C_{24}(p_1^2,p_2^2,q^2;m_X,m_Y,m_Z) -2 , \\
F_{VFV}(X,Y,Z) &= q^2[C_0 + C_{11} + C_{23}](p_1^2,p_2^2,q^2;m_X,m_Y,m_Z) \notag\\
&+ 6C_{24}(p_1^2,p_2^2,q^2;m_X,m_Y,m_Z) -1.
\end{align}
\section{Contributions from $hf\bar{f}$ vertex corrections and box diagrams \label{sec:box}}
In the calculation of the decay rate of $h \to VV^* \to Vf\bar{f}$,
there are contributions from $hf\bar{f}$ vertex corrections denoted as $T_{hff}^V$ and box diagrams denoted as $B_V$ in Eqs.~(\ref{eq:del_ew_z}) and (\ref{eq:del_ew_w}).
The analytic expressions for $T_{hff}^V$ are given as follows:
\begin{align}
T_{hff}^Z&=\frac{\kappa_V}{256\pi^3m_h^3}\frac{g_Z^6m_Z^2}{16\pi^2(s-m_Z^2)} \Big\{c_W^4(v_f+a_f)^2[C_{12}(0,u,m_h^2,m_W,0,m_W) \notag\\
& - (C_0+C_{11})(t_Z^{},0,m_h^2;m_W,0,m_W) ]\notag\\
& + 4(v_f^4+6v_f^2a_f^2+a_f^4)\left[C_{12}(0,u,m_h^2;m_Z^{},0,m_Z^{}) - (C_0 + C_{11})(t_Z^{},0,m_h^2,m_Z^{},0,m_Z^{})\right] \Big\}\notag\\
&\times \left[s + \frac{(m_h^2-s-u)(u-m_Z^2)}{m_Z^2} \right], \label{eq:thff1}
\end{align}
\begin{align}
T_{hff}^W& = \frac{\kappa_V}{256\pi^3m_h^3}\frac{g^6m_W^2}{32\pi^2(s-m_W^2)m_h^2}\Bigg\{C_{12}(0,u,m_h^2;m_W^{},0,m_W^{})-(C_0+C_{11})(t_W,0,m_h^2;m_W^{},0,m_W^{}) \notag\\
& + \frac{2}{c_W^4}\left[(v_f+a_f)^2C_{12}(0,u,m_h^2;m_Z^{},0,m_Z^{}) - (v_{f^\prime}+a_{f^\prime})^2(C_0+C_{11})(t_W,0,m_h^2;m_Z^{},0,m_Z^{})\right] \Bigg\}\notag\\
&\times \left[s + \frac{(m_h^2-s-u)(u-m_W^2)}{m_W^2} \right],
\end{align}
where $t_V = m_h^2 + m_V^2 -s -u $.
For the calculation of box diagrams, we define the Passarino-Veltman's $D$ functions:
\begin{align}
&\frac{i}{16\pi^2}[D_0,D^\mu,D^{\mu\nu}](p_1^2,p_2^2,p_3^2,(p_1+p_2+p_3)^2,(p_1+p_2)^2,(p_2+p_3)^2;m_1,m_2,m_3,m_4) \notag\\
& = \int \frac{d^4k}{(2\pi)^4}\frac{[1,k^\mu, k^\mu k^\nu]}{N_1N_2N_3N_4},
\end{align}
where $N_1 = k^2 - m_1^2$,
$N_2 = (k+p_1)^2 - m_2^2$,
$N_3 = (k+p_1+p_2)^2 - m_3^2$,
$N_4 = (k+p_1+p_2+p_3)^2 - m_4^2$.
We note that in our calculation up to the second rank tensors $D^{\mu\nu}$ appear, and these functions are UV finite.
The first and second rank tensor functions are decomposed into the following scalar coefficients:
\begin{align}
D^\mu &= p_1^\mu D_{11} + p_2^\mu D_{12} + p_3^\mu D_{13}, \\
D^{\mu\nu} &= p_1^\mu p_1^\nu D_{21} + p_2^\mu p_2^\nu D_{22} + p_3^\mu p_3^\nu D_{23}
+(p_1^\mu p_2^\nu + p_2^\mu p_1^\nu) D_{24}\notag\\
& +(p_1^\mu p_3^\nu + p_3^\mu p_1^\nu) D_{25}
+(p_2^\mu p_3^\nu + p_3^\mu p_2^\nu) D_{26}
+g^{\mu\nu}D_{27}.
\end{align}
In the following, we shortly express the $D$ functions by
$D_{0,i,ij}(p_1^2,p_2^2,p_3^2,(p_1+p_2+p_3)^2,(p_1+p_2)^2,(p_2+p_3)^2;a,b,c,d) \equiv D_{0,i,ij}(p_1^2,p_2^2,p_3^2,(p_1+p_2+p_3)^2,(p_1+p_2)^2,(p_2+p_3)^2;m_a,m_b,m_c,m_d)$.
The contribution from box diagrams $B_V$ can be expressed as
\begin{align}
B_V &= \frac{\kappa_V^{}}{256\pi^3 m_h^3}\frac{1}{16\pi^2}\frac{c_V^{} g_V^2}{(s-m_V^2)m_V}
\Big[(m_V^2-t)(tu - m_h^2 m_V^2)(\Gamma_{Vff}^{V,{\rm tree}} {\cal B}_V^1 + \Gamma_{Vff}^{A,{\rm tree}} {\cal \bar{B}}_V^1) \notag\\
& + (m_V^2-u)(tu-m_h^2 m_V^2)(\Gamma_{Vff}^{V,{\rm tree}} {\cal B}_V^2 + \Gamma_{Vff}^{A,{\rm tree}} {\cal \bar{B}}_V^2)
+2 (tu - m_h^2 m_V^2 + 2sm_V^2)(\Gamma_{Vff}^{V,{\rm tree}} {\cal B}_V^\gamma + \Gamma_{Vff}^{A,{\rm tree}} {\cal \bar{B}}_V^\gamma)\Big],
\end{align}
where $c_V^{} = 1 (\sqrt{2})$ for $V = Z(W)$.
Each factor ${\cal B}_V^i$ and ${\cal \bar{B}}_V^i$ ($i = 1,2,\gamma$) for $V =Z$ is given as follows:
\begin{align}
{\cal B}_{Z}^1&= g^4m_W\Big[-2c_W^{}I_f(D_0 + D_{11} + D_{13} + D_{25})(0,0,m_Z^2,m_h^2,s,u;W,f',W,W)\notag\\
& \hspace{18mm}-c_W^{}I_f(D_{13}-D_{12}+2D_{26})(0,0,m_Z^2,m_h^2,s,t;W,f',W,W) \notag\\
&\hspace{18mm}+\frac{s_W^{2}}{c_W^{}}I_f(D_{13} - D_{12})(0,0,m_Z^2,m_h^2,s,t;W,f',W,W) \notag\\
&\hspace{18mm}+\frac{v_{f'} + a_{f'}}{c_W} D_{26}(0,m_Z^2,0,m_h^2,u,t;W,f',f',W) \notag\\
&\hspace{18mm} +\frac{4}{c_W^5}(v_f^3 + 3v_fa_f^2)D_{26}(0,m_Z^2,0,m_h^2,u,t;Z,f,f,Z)\Big], \\
{\cal B}_{Z}^2&= g^4m_W\Big[-c_W^{}I_f (D_{13}-D_{12}+2D_{26})(0,0,m_Z^2,m_h^2,s,u;W,f',W,W)\notag\\
&\hspace{18mm}-2c_W^{}I_f(D_0 + D_{11} + D_{13} + D_{25})(0,0,m_Z^2,m_h^2,s,t;W,f',W,W)\notag\\
&\hspace{18mm} +\frac{s_W^{2}}{c_W^{}}I_f (D_{13} - D_{12})(0,0,m_Z^2,m_h^2,s,u;W,f',W,W)\notag\\
&\hspace{18mm}+\frac{v_{f'} + a_{f'}}{c_W}(D_0 + D_{11} + D_{12} + D_{24})(0,m_Z^2,0,m_h^2,u,t;W,f',f',W)\notag\\
&\hspace{18mm}+\frac{4}{c_W^5}(v_f^3 + 3v_fa_f^2) (D_0 + D_{11} + D_{12} + D_{24})(0,m_Z^2,0,m_h^2,u,t;Z,f,f,Z)\Big], \\
{\cal B}_{Z}^\gamma & = g^4m_W\Big\{-\frac{I_f}{2}c_W^{}[-2C_0(s,m_Z^2,m_h^2;m_W,m_W,m_W)
+(t-s-m_Z^2)(D_0+D_{11}) \notag\\
&\hspace{18mm}+2(s+t-m_h^2)D_{12} + (s+t-m_h^2-m_Z^2)D_{13} -4D_{27} ](0,0,m_Z^2,m_h^2,s,u;W,f',W,W) \notag\\
&\hspace{18mm}+ \frac{I_f}{4}\frac{s_W^{2}}{c_W^{}}[C_0(s,m_Z^2,m_h^2;m_W,m_W,m_W)+2(s+t-m_Z^2)(D_0+D_{11}) \notag\\
&\hspace{18mm}- 2(s+t-m_h^2-m_Z^2)D_{13}](0,0,m_Z^2,m_h^2,s,u;W,f',W,W) + (t \leftrightarrow u)\Big\} \notag\\
& +g^4m_W\Big\{
\frac{v_{f'} + a_{f'}}{2c_W}[C_0(u,0,m_h^2;m_W,0,m_W)+(m_h^2-s-t)(D_0+D_{11}) + m_Z^2D_{12} \notag\\
&\hspace{18mm} -2D_{27}](0,m_Z^2,0,m_h^2,u,t;W,f',f',W) \notag\\
&+\frac{2}{c_W^5}(v_f^3 + 3v_fa_f^2)[C_0(u,0,m_h^2;m_Z,0,m_Z)+ (m_h^2-s-t)(D_0+D_{11}) + m_Z^2D_{12} \notag\\
&\hspace{18mm} -2D_{27}](0,m_Z^2,0,m_h^2,u,t;Z,f,f,Z) \Big\}, \\
{\cal \bar{B}}_Z^i &= {\cal B}_{Z}^i\big|_{v_f \leftrightarrow a_f}~~(i = 1,2,\gamma),
\end{align}
Those for $V =W$ are given by
\begin{align}
{\cal B}_{W}^1&= -\sqrt{2}g^4m_W\Big[2I_{f'}(v_{f'}+a_{f'}) (D_0 + D_{11} + D_{13} + D_{25})(0,0,m_W^2,m_h^2,s,u;W,f',Z,W)\notag\\
&\hspace{18mm}+2s_W^2 I_{f'}Q_{f'} (D_0 + D_{11} + D_{13} + D_{25})(0,0,m_W^2,m_h^2,s,u;W,f',\gamma,W)\notag\\
&\hspace{18mm}+I_f(v_f+a_f)(D_{13}-D_{12}+2D_{26})(0,0,m_W^2,m_h^2,s,t;W,f,Z,W)\notag\\
&\hspace{18mm}+s_W^2 I_{f}Q_{f}(D_{13}-D_{12}+2D_{26})(0,0,m_W^2,m_h^2,s,t;W,f,\gamma,W)\notag\\
&\hspace{18mm}+2I_f\frac{v_f+a_f}{c_W^2}(D_0 + D_{11} + D_{13} + D_{25})(0,0,m_W^2,m_h^2,s,u;Z,f,W,Z)\notag\\
&\hspace{18mm}+I_{f'}\frac{v_{f'}+a_{f'}}{c_W^2}(D_{13}-D_{12}+2D_{26})(0,0,m_W^2,m_h^2,s,t;Z,f',W,Z)\notag\\
&\hspace{18mm}+\frac{s_W^2}{c_W^2}I_{f}(v_{f} + a_{f})(D_{13} - D_{12})(0,0,m_W^2,m_h^2,s,t;W,f,Z,W)\notag\\
&\hspace{18mm}- s_W^2I_fQ_f(D_{13} - D_{12})(0,0,m_W^2,m_h^2,s,t;W,f,\gamma,W)\notag\\
&\hspace{18mm}-\frac{ (v_f + a_f)(v_{f'} + a_{f'})}{c_W^4} D_{26}(0,m_W^2,0,m_h^2,u,t;Z,f',f,Z) \Big],
\end{align}
\begin{align}
{\cal B}_{W}^2&= -\sqrt{2}g^4m_W\Big[I_{f'}(v_{f'}+a_{f'})(D_{13}-D_{12}+2D_{26})(0,0,m_W^2,m_h^2,s,u;W,f',Z,W)\notag\\
&\hspace{18mm}+s_W^2 I_{f'}Q_{f'}(D_{13}-D_{12}+2D_{26})(0,0,m_W^2,m_h^2,s,u;W,f',\gamma,W)\notag\\
&\hspace{18mm}+2I_{f}(v_{f}+a_{f}) (D_0 + D_{11} + D_{13} + D_{25})(0,0,m_W^2,m_h^2,s,t;W,f,Z,W)\notag\\
&\hspace{18mm}+2s_W^2 I_fQ_f (D_0 + D_{11} + D_{13} + D_{25})(0,0,m_W^2,m_h^2,s,t;W,f,\gamma,W)\notag\\
&\hspace{18mm}+I_f\frac{v_f+a_f}{c_W^2}(D_{13}-D_{12}+2D_{26})(0,0,m_W^2,m_h^2,s,u;Z,f,W,Z)\notag\\
&\hspace{18mm}+2I_{f'}\frac{v_{f'} + a_{f'}}{c_W^2} (D_0 + D_{11} + D_{13} + D_{25})(0,0,m_W^2,m_h^2,s,t;Z,f',W,Z)\notag\\
&\hspace{18mm}+\frac{s_W^2}{c_W^2}I_{f'}(v_{f'} + a_{f'}) (D_{13} - D_{12})(0,0,m_W^2,m_h^2,s,u;W,f',Z,W)\notag\\
&\hspace{18mm}-s_W^2I_{f'}Q_{f'}(D_{13} - D_{12})(0,0,m_W^2,m_h^2,s,u;W,f',\gamma,W)\notag\\
&\hspace{18mm}-\frac{(v_f + a_f)(v_{f'} + a_{f'})}{c_W^4} (D_0 + D_{11} + D_{12} + D_{24})(0,m_W^2,0,m_h^2,u,t;Z,f',f,Z)\Big],
\end{align}
\begin{align}
{\cal B}_{W}^\gamma &=
-\frac{g^4m_W}{\sqrt{2}}\Big\{
I_{f'}(v_{f'}+a_{f'})[ -2C_0(s,m_W^2,m_h^2;m_W,m_Z,m_W)+(t-s-m_W^2)(D_0+D_{11}) \notag\\
&\hspace{18mm} +2(s+t-m_h^2)D_{12} + (s+t-m_h^2-m_W^2)D_{13}-4D_{27}](0,0,m_W^2,m_h^2,s,u;W,f',Z,W)\notag\\
&\hspace{18mm}+s_W^2I_{f'}Q_{f'} [-2C_0(s,m_W^2,m_h^2;m_W,0,m_W) + (t-s-m_W^2)(D_0+D_{11}) \notag\\
&\hspace{18mm} +2(s+t-m_h^2)D_{12}+ (s+t-m_h^2-m_W^2)D_{13}-4D_{27} ](0,0,m_W^2,m_h^2,s,u;W,f',\gamma,W)\notag\\
&\hspace{18mm}+ I_f\frac{v_f+a_f}{c_W^2} [-2C_0(s,m_W^2,m_h^2;m_W,m_Z,m_W) + (t-s-m_W^2)(D_0+D_{11}) \notag\\
&\hspace{18mm} +2(s+t-m_h^2)D_{12} + (s+t-m_h^2-m_W^2)D_{13}-4D_{27} ](0,0,m_W^2,m_h^2,s,u;Z,f,W,Z) \notag\\
&\hspace{18mm} +\frac{s_W^2}{2c_W^2}I_{f'}(v_{f'} + a_{f'})[C_0(s,m_W^2,m_h^2;m_W,m_Z,m_W)+2(s+t-m_W^2)(D_0+D_{11}) \notag\\
&\hspace{18mm} - 2(s+t-m_h^2-m_W^2)D_{13} ](0,0,m_W^2,m_h^2,s,u;W,f',Z,W)\notag\\
&\hspace{18mm} -\frac{s_W^2}{2}I_{f'}Q_{f'}[C_0(s,m_W^2,m_h^2;m_W,0,m_W) + 2(s+t-m_W^2)(D_0+D_{11}) \notag\\
&\hspace{18mm} - 2(s+t-m_h^2-m_W^2)D_{13} ](0,0,m_W^2,m_h^2,s,u;W,f',\gamma,W)\notag\\
&\hspace{18mm} + (t \leftrightarrow u,~f \leftrightarrow f')\Big\} \notag\\
& +g^4m_W\frac{(v_f + a_f)(v_{f'} + a_{f'})}{\sqrt{2}c_W^4} [C_0(u,0,m_h^2;m_Z,0,m_Z) + (m_h^2-s-t)(D_0+D_{11}) \notag\\
&\hspace{18mm} + m_W^2D_{12} -2D_{27}](0,m_W^2,0,m_h^2,u,t;Z,f',f,Z), \\
{\cal \bar{B}}_W^i & = {\cal B}_W^i~~(i = 1,2,\gamma).
\end{align}
\section{Real photon emissions in $h \to WW^* \to Wf\bar{f}'$ \label{sec:real_photon}}
\begin{figure}[t]
\centering
\includegraphics[scale=1.0]{lam_dep_Wffp.png}\\
\vspace{+0.8cm}
\caption{Numerical check of cancellation of the soft divergence in the decay rate of $h \to Wf\bar{f}$ at NLO.
The horizontal axis is the photon mass $\lambda$ { as a regulator }{to avoid the divergence}.
The blue and orange curves show the contribution from the virtual corrections and real photon emissions, respectively. }
\label{FIG:lam_dep_Wffp}
\end{figure}
The contribution from the real photon emission in the $h \to WW^* \to Wf\bar{f}'$ process
can be separately written by the soft-photon and hard-photon emission parts as
\begin{align}
\Gamma({h\to Wff^\prime \gamma}) = \frac{1}{2m_h}\int_S|\mathcal{M}|^2d\Phi_4+\frac{1}{2m_h}\int_H|\mathcal{M}|^2d\Phi_4,
\end{align}
where $\Phi_4$ is the four body phase space function.
The first integral denoted as $\int_S$ is performed up to the cutoff of the photon energy $\Delta E$, while the second integral is done from $\Delta E$ to the maximal value of the photon energy.
The $\Delta E$ dependence in the NLO decay rate, of course, disappears after summing up the soft and hard photon parts.
The soft-photon part is calculated using the eikonal approximation by which the amplitude can be expressed by the product of
the Born amplitude and the soft-photon factor.
{Then, we can} separately perform the integration with respect to the 3-body phase space and the photon phase space.
Therefore, the soft-photon part is expressed as
\begin{align}
\frac{1}{2m_h}\int_S|\mathcal{M}|^2d\Phi_4 = \int d \Gamma_0(h \to Wf\bar{f}')\delta_{W}^{\rm soft},
\end{align}
where
\begin{align}
\notag
\delta_W^{\rm soft}&=-\frac{\alpha_{\rm em}}{2\pi}\Bigg[
Q_f^2\Big\{\Big(\log\frac{m_f^2}{s}+1\Big)\log\frac{4\Delta E}{\lambda^2}+\frac{1}{2}\Big(\log\frac{m_f^2}{4E_f^2}\Big)^2+\log\frac{m_f^2}{4E_f^2}+\frac{\pi^2}{3}\Big\} \\ \notag
&+Q_{f^\prime}^2\Big\{\Big(\log\frac{m_{f^\prime}^2}{s}+1\Big)\log\frac{4\Delta E}{\lambda^2}+\frac{1}{2}\Big(\log\frac{m_{f^\prime}^2}{4E_{f^\prime}^2}\Big)^2+\log\frac{m_{f^\prime}^2}{4E_{f^\prime}^2}+\frac{\pi^2}{3}\Big\} \\ \notag
&+\Big\{\Big(\log\frac{m_{W}^2}{s}+1\Big)\log\frac{4\Delta E}{\lambda^2}+\frac{1}{2}\Big(\log\frac{E_W-|\vec{p}_W|}{E_W+|\vec{p}_W|}\Big)^2+\frac{E_W}{|\vec{p}_W|}\log\frac{m_{W}^2}{(E_W+|\vec{p}_W|)^2}\Big\} \\ \notag
&+4Q_fI_f \Big\{\frac{1}{2}\log\frac{s^2}{(m_W^2-t)^2}\log\frac{4\Delta E}{\lambda^2}+{\rm Li_2}\left(1-\frac{2E_f(E_W-|\vec{p}_W|)}{t-m_W^2}\right) \\ \notag
&+{\rm Li_2}\left(1-\frac{2E_f(E_W+|\vec{p}_W|)}{t-m_W^2}\right)
\Big\} \\ \notag
&-4Q_{f^\prime}I_f\Big\{\frac{1}{2}\log\frac{s^2}{(m_W^2-u)^2}\log\frac{4\Delta E}{\lambda^2}+{\rm Li_2}\left(1-\frac{2E_{f^\prime}(E_W-|\vec{p}_W|)}{u-m_W^2}\right) \\
&+{\rm Li_2}\left(1-\frac{2E_{f^\prime}(E_W+|\vec{p}_W|)}{u-m_W^2}\right)
\Big\}\Bigg],
\end{align}
with $t = m_h^2 + m_V^2 -s -u $, $E_W = m_h(1-x_s+x_W)/2$ and $|\vec{p}_W| = m_h\lambda^{1/2}(x_s,x_W)/2$.
This expression agrees with~\cite{Ciccolini:2003jy}.
{Here, we introduced} the photon mass $\lambda$ to regularize the IR divergence and the fermion masses $m_f$
to regularize the collinear singularities.
{We numerically evaluate the hard-photon part by} using {\tt Madgraph5\_aMC@NLO}~\cite{Alwall:2014hca}.
In order to check the cancellation of the IR divergence, we show the partial decay rate at NLO in Fig.~\ref{FIG:lam_dep_Wffp}
as a function of the photon mass $\lambda$ {being a regulator} {to avoid the divergence}, where we take $\Delta E=$ 1 GeV.
We clearly see that the sum of the virtual corrections and the real emissions (denoted as ``Total'') does not depend on $\lambda$.
\end{appendix}
|
\section{Introduction and motivation}
\label{sec:Introduction}
Non-Hermitian quantum mechanics has been extremely useful in the description of
a great deal of physical systems, from scattering, resonance phenomena and
ionization, to effective models of open systems \cite{Gamow1928, Siegert1939,
Peierls1959, Hatano1996, Nelson1998, Moibook2011, SchomerusPTRS2013}.
In particular, non-Hermitian systems described by $\mathcal{PT}$-symmetric
Hamiltonians~\cite{BenderPRL1998, Benderbook2018} have found many
applications \cite{HeissJPA2012,BenderJP2015, FengNature2017, El-GanainyNature2018, Miri2019}.
$\mathcal{PT}$-symmetric Hamiltonians may have
real or complex eigenvalues, corresponding to unbroken or broken
$\mathcal{PT}$-symmetry phases. The transition between these phases occurs at
the so called exceptional points, at which two (or more)
eigenvalues and eigenfunctions coalesce and the Hamiltonian becomes defective
\cite{Katobook1995}. Several remarkable phenomena have been reported recently
in the vicinity of these points, for example, topological
states \cite{Lee2016, MartinezEPJST2018, FoaPRB2018, Ni2018, Yuce2018, Lin2019, Caspel2019},
chirality \cite{DembowskiPRL2001, Mailybaev2005, Peng2016}, unidirectional
invisibility \cite{Regensburger2012, Lin2011, Feng2012}, unidirectional zero sonic
reflection \cite{Merkel2018}, enhanced sensing \cite{Chen2017} and the possibility
to stop light \cite{GoldzakPRL2018}.
In this broad context, simple models are valuable as they permit a thorough
understanding of the phenomena taking place in the system. Indeed, analysis of even the
simplest $2\times 2$ $\mathcal{PT}$-symmetric matrices has led to important insights
\cite{BenderJPA2004, WangPTRS2013}, though certain
aspects cannot be captured with such simple model, such as the simultaneous coalescence of
more than two eigenvalues~\cite{FoaPRB2018,GraefeJPA2012}.
In this paper we study a simple
one-dimensional tight-binding chain, with gain and loss at arbitrary
($\mathcal{PT}$-symmetric) positions along the chain.
This system, and extensions of it, have already been extensively
studied \cite{KottosPRL2009, JinPRA2009, YogeshPRA2010, YogeshPRA2011, LonghiOL2014, ElenewskiPRB2014,
GarmonPRA2015, ZhangPRA2017, HarterNature2018}. We focus here in the
$\mathcal{PT}$-symmetric tight-binding chain from the perspective of quantum
transport. The outline and main results of our work are as follows: First we
provide the necessary definitions
in Sec.~\ref{sec:ModelSolution}. In Sec.~\ref{sec:LocEigstates} we outline the
derivation of a continuity equation for the density on the
$\mathcal{PT}$-symmetric tight-binding chain. We also propose a parameter, the
transport coefficient $\xi$, that measures the efficiency
of transport through the system.
The explicit analytical expressions of the eigenvalues and eigenvectors of the system, which
will be used to study transport through the chain, are presented in
Sec.~\ref{sec:egvalsegvecs}, while a detailed derivation can be found in the
Appendix \ref{sec:egvalsol}. It is worth noting that this derivation does not
make use of the Bethe ansatz as in previous works \cite{JinPRA2009,
YogeshPRA2010}, but uses straightforward algebra in the ring of
semi-infinite sequences. This is a simpler, or at least alternative,
solution of the problem. From the explicit results for the eigenvalues and eigenvectors of the system, we show that under certain circumstances, the system may have a set of eigenstates characterized by having eigenvalues that
are independent of the strength of the gain and loss in the system. Some of these states do not couple to the gain and loss, and thus, are non-conducting or ``opaque'', whereas another subset ---that we call ``transparent''---
always conduct efficiently. We show that the
condition for the appearance of such states depends on the divisors of the length of the
system plus one, and of the positions of the loss and gain. This implies that the number of both opaque and transparent states
varies very irregularly dependent on the size of the system, and the precise position of the leads, even for
large sizes. Next, some specific cases of how the eigenvalues behave
in the $\mathcal{PT}$-unbroken
(-broken) phases are analyzed thoroughly. We also develop a simple
perturbation scheme to obtain the eigenvalues and eigenvectors
around an exceptional point. With the full solution to the problem, in
Sec. \ref{sec:transportPT} we analyze analytically and numerically the
transport in a $\mathcal{PT}$-symmetric tight-binding chain as a function of
the chain length and position of the gain and loss.
We analyze some eigenfunctions along the parameter space and show how
the $\mathcal{PT}$-unbroken (-broken) phase affect their behaviour.
We give our conclusions in Sec. \ref{sec:Conclusions} and provide an outlook
based on our results.
\section{The PT-symmetric tight-binding chain}
\label{sec:ModelSolution}
A system is $\mathcal{PT}$-symmetric if the Hamiltonian commutes
with the operator $\mathcal{PT}$, where $\mathcal{P}$ and
$\mathcal{T}$ are the parity and the time reversal operators,
respectively. This is commonly referred
to as space-time reflection symmetry (see e.g.
\cite{BenderCP2005, GraefeJPA2008}). Following
Bender~\cite{Bender2007}, for a $\mathcal{PT}$-symmetric operator
we say that $\mathcal{PT}$ symmetry is unbroken
if all the eigenfunctions of the Hamiltonian are also eigenfunctions
of $\mathcal{PT}$; otherwise, we say that $\mathcal{PT}$ symmetry is broken.
In this work we consider spinless particles for which the effect
of the time reversal operator $\mathcal{T}$ can be defined simply
as complex conjugation \cite{WangPTRS2013}:
\begin{equation}
\begin{matrix}
\mathcal{T} \equiv {}^*, & \mathcal{T}^2=\mathbb{1},
\end{matrix}
\end{equation}
where $\mathbb{1}$ is the identity. For a matrix $M$, the action of
$\mathcal{T}$ is $\mathcal{T}M\mathcal{T} = M^*$. The
parity operator $\mathcal{P}$ is defined by the properties
\begin{equation}
\begin{matrix}
\mathcal{P} = \mathcal{P}^* & \text{and} & \mathcal{P}^2 = \mathbb{1}.
\end{matrix}
\end{equation}
Fixing a basis in a Hilbert space, we choose $\mathcal{P}$ as the
matrix $J$ with components
\begin{equation}
J_{ij} = \delta_{i,N-j+1},
\label{eq:J}
\end{equation}
which is commonly known as the exchange matrix in the mathematical
literature \cite{CantoniLAA1976}, sometimes called sip matrix
\cite{GraefeJPA2008}.
\begin{figure}
\centering
\hspace*{-6mm}
\includegraphics[width=0.5\linewidth]{fig1.pdf}
\caption{One-dimensional tight-binding chain, Eq. (\ref{eq:1}), with
gain and loss in a $\mathcal{PT}$-symmetric configuration. We
illustrate the case $N=8$ with the contacts at sites $k=2$ and
$k'=N-k+1=7$; the strength of the coupling of the contacts is
characterized by $\eta$.}
\label{fig:modeldiagram}
\end{figure}
In the site basis, a tight-binding chain in one dimension with gain and
loss in a symmetric configuration (see Fig. \ref{fig:modeldiagram})
is described by the $\mathcal{PT}$-symmetric Hamiltonian
\begin{equation}
H = t \sum_{i=1}^{N-1} \Big(\ket{i}\bra{i+1} + \ket{i+1}\bra{i}\Big)
+ \mathrm{i} \eta \Big(\ket{k}\bra{k} - \ket{N-k+1}\bra{N-k+1}\Big),
\label{eq:1}
\end{equation}
where $t$ is the nearest-neighbor coupling, $N$ is the length of
the chain and corresponds to the dimension of the Hilbert space of
the system, $k$ is the position of the gain, $k'=N-k+1$ the position
of the loss, and $\eta$ is a real number that describes the strength
of the gain and loss. Without loss of generality we fix $t=1$.
In what follows, we will denote $\ket{E_\theta}=\sum_{j=1}^N u_j(\theta)
\ket{j}$ the eigenstates of the Hamiltonian Eq. (\ref{eq:1}), corresponding
to the energy $E_\theta$, where $\theta$ is the pseudo momentum that
characterizes each state.
\section{Transport properties}
\label{sec:LocEigstates}
\subsection{The continuity equation}
\label{ssec:Continuity}
In this section we derive the continuity equation for the density in
the chain described by the effective Hamiltonian Eq.~(\ref{eq:1}). We illustrate the derivation for the case
where the contacts are in the end-to-end configuration, i.e., at
$k=1$ and $k'=N$, and write its generalization to other configurations.
Let $\ket{\Psi(t)}$ be a solution of the Schr{\"o}dinger equation.
Then we have
\begin{equation}
\frac{\partial\ket{\Psi(t)}}{\partial t} = \frac{H}{\mathrm{i}}\ket{\Psi(t)},
\end{equation}
and its adjoint
\begin{equation}
\frac{\partial\bra{\Psi(t)}}{\partial t} = -\frac{1}{\mathrm{i}}\bra{\Psi(t)}H^\dagger,
\label{eq:cs}
\end{equation}
where we have set $\hbar=1$. Using the site basis we write
\begin{equation}
\ket{\Psi(t)} = \sum_{j=1}^N c_j (t) \ket{j},
\label{eq:Psi}
\end{equation}
where the expansion coefficients are given by
\begin{equation}
\begin{split}
c_n(t) &= \langle n|\Psi(t)\rangle,\\
c_n^*(t) &= \langle n|\Psi(t)\rangle^* = \langle\Psi(t)| n\rangle.
\end{split}
\end{equation}
In order to derive the continuity equation, we consider the time derivative of the diagonal elements of the
density matrix $\rho_{nn}(t) = \langle n|\Psi(t)\rangle \langle \Psi(t)|n\rangle$, obtaining
\begin{equation}
\begin{split}
\frac{\partial\rho_{nn}}{\partial t} &= \frac{1}{\mathrm{i}}\big( \langle n|H|\Psi(t)\rangle
\langle\Psi(t)|n\rangle - \langle n|\Psi(t)\rangle \langle \Psi(t)|H^\dagger|n\rangle\big) \\
& = \frac{1}{\mathrm{i}}\sum_m \big( H_{nm} c_m(t) c_n^*(t) - H_{mn}^* c_n(t) c_m^*(t) \big).
\end{split}
\label{eq:Hs}
\end{equation}
By using the explicit form
of the Hamiltonian, Eq.~(\ref{eq:1}), assuming that the gain and loss are at the end points of the chain, we find
\begin{equation}
\frac{\partial\rho_{nn}}{\partial t}=\frac{1}{\mathrm{i}}\begin{cases}
c_{2}(t)c_1^*(t)- c_1(t) c_{2}^*(t)+ 2\mathrm{i}|c_1(t)|^2\eta, & n=1, \\
c_{N-1}(t)c_N^*(t) - c_N(t) c_{N-1}^*(t) - 2\mathrm{i} |c_N(t)|^2\eta, & n=N,\\
c_{n+1}(t)c_n^*(t) + c_{n-1}(t)c_n^*(t) - c_n(t) c_{n+1}^*(t) - c_n(t)
c_{n-1}^*(t), & \textrm{otherwise}
\end{cases}
\label{eq:continuity1}
\end{equation}
Equation~(\ref{eq:continuity1}) can succinctly be written as
\begin{equation}
\frac{\partial\rho_{nn}}{\partial t}+J_{n+1}-J_n= 2\eta
|c_1(t)|^2\delta_{n,1} - 2\eta|c_N(t)|^2\delta_{n,N},
\label{eq:continuity2}
\end{equation}
where we have introduced the local fluxes
\begin{equation}
J_{n}(t) \equiv \mathrm{i} \big(c_n(t) c_{n-1}^*(t) - c_{n-1}(t) c_n^*(t))
= -2 \operatorname{Im}(c_{n}(t)c_{n-1}^*(t)),
\label{flux}
\end{equation}
which represent the density flux from site $n-1$
to site $n$ ($n\neq 1, N+1$), with the boundary conditions
$J_1= J_{N+1}= 0$.
Equation~(\ref{eq:continuity2}) is a continuity equation with
source and sink terms representing the
inflow and outflow due to the presence of gain and loss in the
chain. Its generalization to other gain and loss configurations
is straight forward and reads
\begin{equation}
\frac{\partial\rho_{nn}}{\partial t}+J_{n+1}-J_n= 2\eta
|c_k(t)|^2\delta_{n,k} - 2\eta|c_{N+1-k}(t)|^2\delta_{n,N+1-k}.
\label{eq:continuity_gen}
\end{equation}
It should be stressed that the inflow and outflow terms are,
as expected, proportional to $\eta$, which are the components
that make the Hamiltonian non-Hermitian.
\subsection{The transport coefficient}
\label{ssec:TransportCoeff}
Now consider $|\Psi(t)\rangle$ to be a time-dependent eigenstate
which we write in the site basis
\begin{equation}
\ket{E_\theta(t)}\equiv e^{-\mathrm{i} E_\theta t}\ket{E_\theta}=e^{-\mathrm{i} E_\theta t}\sum_{j=1}^N u_j(\theta) \ket{j}.
\end{equation}
For these states, $c_n(t)=e^{-\mathrm{i} E_\theta t}u_n(\theta)$.
When all eigenvalues are real, products of the form
$c_n(t)c^*_m(t)=u_n(\theta)u^*_m(\theta)$, as those appearing in
the definition for the flux in Eq.~(\ref{flux}), are independent
of time.
In the broken $\mathcal{PT}$-symmetric phase the
corresponding eigenvalues are complex, and terms of the form
$c_n(t)c^*_m(t)=e^{-2\operatorname{Im}(E_\theta) t} u_n(\theta)u^*_m(\theta)$
increase or decrease exponentially in time.
In view of this, we define
\begin{equation}
\xi_{E_\theta}=\left\vert\frac{c_{k'}(t)}{c_k(t)}\right\vert^2,
\label{transp_coeff}
\end{equation}
which corresponds to the
ratio of the outflow to the inflow in Eq.~(\ref{eq:continuity_gen}) with $k'=N-k+1$.
Evaluated in the states corresponding to the eigenfunctions
$\ket{E_\theta(t)}$, the transport coefficient is independent of
time and of the normalization. If $\xi > 1$,
the inflow at $k'$ is larger than the outflow at $k$ and there is a
buildup of density within the chain. Conversely, if $\xi<1$ the outflow is
larger than the inflow and the system becomes depleted. When $\xi=1$, the
gain and loss are equally coupled, the inflow and outflow are the same,
and, in this sense, transport is efficient.
Now, given that $\ket{E_\theta}$
are eigenvectors of a $\mathcal{PT}$-invariant Hamiltonian, then
\begin{equation}
H\mathcal{PT}\ket{E_\theta} =E^*_\theta\,\mathcal{PT}\ket{E_\theta}.\label{PTket}
\end{equation}
Thus, since in the unbroken
$\mathcal{PT}$-symmetry phase the eigenvalues are real,
the eigenfunctions fulfill
\begin{equation}
\mathcal{PT}\ket{E_\theta} \propto \ket{E_\theta},
\label{sym1}
\end{equation}
while in the broken $\mathcal{PT}$ symmetry phase,
some eigenvalues $E_\theta$ are
complex and come in conjugate pairs. The corresponding
eigenstates satisfy
\begin{equation}
\mathcal{PT}\ket{E_\theta} \propto \ket{E^*_\theta}.
\label{eq:sym2}
\end{equation}
Consequently, in the unbroken $\mathcal{PT}$-symmetry
phase we have $\xi_{E_\theta}=1$. This indicates that
transport in the eigenstates with real eigenvalues
is efficient if the gain and loss couple with such states.
On the other hand, for states in the
$\mathcal{PT}$-broken symmetry
phase with complex eigenvalues, their eigenstates
localize around the gain and decouple from the loss
or viceversa; in this case $\xi_{E_\theta}$ is no
longer equal to one, indicating that transport between
loss and gain is deficient. Yet, in view of
Eq.~(\ref{eq:sym2}), it is straight forward to see that
in this phase $\xi_{E_\theta}\xi_{E_\theta^*}= 1$.
As we shall see below, even if there are some states
with complex eigenvalues, others may still have real
eigenvalues, and therefore efficient transport is still
possible.
Further, under certain circumstances
there may be states with real eigenvalues that are independent
of $\eta$. These can be divided depending on whether (or not)
their amplitudes vanish at the gain and loss positions. If the
amplitude vanishes at the position of the leads, the state does
not couple to the gain and loss and there is no transport through
this state in the system ($\xi_{E_\theta}$ is undefined). We refer
to these as opaque states. If, on the other hand, the amplitude
does not vanish at the leads, then $\xi_{E_\theta}=1$ for all
values of $\eta$. We call these transparent states.
\section{Eigenvectors and eigenvalues}
\label{sec:egvalsegvecs}
To investigate the transport properties of the $\mathcal{PT}$
symmetric tight binding chain we require the spectra and eigenvectors
of the system. For this particular system, the eigenvalues have been
obtained using the Bethe ansatz \cite{YogeshPRA2010, YogeshPRA2011}.
Here, however, we obtain the eigenvalues and eigenvectors
of the Hamiltonian using symbolic calculus
\cite{Losonczi1992, Yueh2005, Chengbook2003};
see Appendix \ref{sec:egvalsol} for the full derivation.
The eigenvalues are given by
\begin{equation}
E_\theta = 2\cos{\theta},
\label{eq:2}
\end{equation}
where the values of the pseudo momentum $\theta$ are those non-trivial
solutions ($\theta \neq m \pi$, with $m\in \mathbb{Z}$) that fulfill
the equation
\begin{equation}
\sin{(N+1)\theta} +
\frac{\eta^2}{\sin^2\theta}\sin[(N-2k+1)\theta]\sin^2(k\theta) = 0,
\label{eq:3}
\end{equation}
where the gain and loss are located at sites $k$ and $k'=N-k+1$,
respectively.
Writing the eigenvectors in the site basis
$\ket{E_\theta} = \sum_{j=1}^N u_j(\theta) \ket{j}$, as required
above to calculate the fluxes and transport coefficients in the system,
the $u_j(\theta)$ component of the eigenvector is given by
\begin{align}
\begin{split}
u_j(\theta) &= \langle j \ket{E_\theta} =
\frac{u_1(\theta)}{\sin\theta}\Big[ \sin j\theta
- \mathrm{i}\eta \Theta(j-k-1) \frac{\sin k\theta\,\sin(j-k)\theta}{\sin\theta} \\
& + \Theta(j-N+k-2) \frac{\sin(j-N+k-1)\theta}{\sin\theta} \\
& \qquad \times \Big( \mathrm{i}\eta \sin(N-k+1)\theta
- \eta^2 \frac{\sin(N-2k+1)\theta\,\sin k\theta}{\sin\theta} \Big
\Big],
\end{split}
\label{eq:4}
\end{align}
where $\Theta(x)$ is the unit step function defined by $\Theta(x)=1$
if $x\geq0$
and $\Theta(x)=0$ otherwise. In Eq.~(\ref{eq:4}), $u_1(\theta)$ is the
first component of each eigenvector, which can be used to fix
the normalization.
Before presenting results for specific configurations of
the system, we discuss some general properties that follow
directly from expressions (\ref{eq:3}) and (\ref{eq:4}).
First of all, clearly, in the limit $\eta\to0$, the
Hamiltonian becomes a symmetric (real Hermitian)
matrix, actually a {\it centrosymmetric} matrix, and the
results in \cite{CantoniLAA1976} hold. From Eq. (\ref{eq:3})
we obtain $\theta=\frac{r\pi}{N+1}$ ($r=1,2, \dots,N$),
which using Eq.~(\ref{eq:2}) yields the well-known
solution for the eigenvalues
$E_{\theta, \eta=0}$~\cite{Losonczi1992, Yueh2005}. From the
centrosymmetry of $H$ it follows that the eigenvectors are
symmetric or skew-symmetric with
respect to the exchange matrix $J$, i.e., they fulfill
$J \ket{E_{\theta,\eta=0}} = \pm \ket{E_{\theta,\eta=0}}$.
It has been shown~\cite{OrtegaAdP2015, OrtegaPRE2016, OrtegaPRE2018}
that centrosymmetry is relevant in achieving good transport properties
in disordered systems.
In the opposite limit, when $\eta\to\infty$, we expect the system
to be divided into several subsystems depending on the positions
$k$ and $k'$ of the gain and loss: two of them correspond to
the uncoupled gain and loss, and the remaining ones to disjoint
tight-binding chains. From Eq.~(\ref{eq:3}), the real parts of
$\theta$ for the disjoint tight-binding chains are given by
$\theta=r\pi/(N-2k+1)$ for $r=1, 2, \dots N-2k$, and the
double-roots $\theta = r\pi/k$ for $r = 1, \dots k-1$. All
these $N-2$ eigenvalues have, in the limit $\eta\rightarrow\infty$,
an imaginary part which is or tends asymptotically to zero.
The asymptotic behavior of the two remaining eigenvalues, which
are purely imaginary, can be obtained by writing $\theta=\mathrm{i}\phi$.
Thus, Eq. (\ref{eq:3}) is transformed to
\begin{equation}
\sinh(N+1)\phi\,\sinh^2\phi + \eta^2\sinh(N-2k+1)\phi\,\sinh^2 k\phi = 0,
\end{equation}
which in the limit $\operatorname{Re}(\phi)\gg1$, reduces to
$e^{2\phi} \sim -\eta^2$. Taking the logarithm we obtain
$\phi \sim \log \eta + \frac{\mathrm{i}\pi}{2} + \mathrm{i}\pi m$, $m\in\mathbb{Z}$
which, using Eq. (\ref{eq:2}), yields
\begin{equation}
E_{\theta,\eta\rightarrow\infty} =
E_{\mathrm{i}\phi,\eta\rightarrow\infty} \sim \pm \mathrm{i}\Big(\eta - \frac{1}{\eta}\Big).
\label{eq:5}
\end{equation}
We note that this equation holds independently of $k$, i.e., for
any symmetric configuration of the gain and loss.
We now discuss the conditions for the presence of opaque and transparent states, and the criteria to distinguish between them. Consider the pseudo momentum $\theta^\textrm{Op}_r = r \pi / M$, where both $r$ and $M$ are
integers, $r=1,\dots,M-1$, and $M>1$ is a divisor
of $N+1$ and $k$ simultaneously. It follows that $M$
divides $k$ and $k'=N-k+1$ as well.
In this case, it is clear that $\theta^\textrm{Op}_r$ are
solutions of Eq.~(\ref{eq:3}), {\it independently} of the
value of $\eta$, and the corresponding eigenvalues are
real. Also, using Eq.~(\ref{eq:4}), it is straight
forward to verify that the corresponding eigenvectors satisfy
$u_k(\theta^\textrm{Op}_r) = u_{N-k+1}(\theta^\textrm{Op}_r)=0$.
Thus, the gain and loss are not coupled to these states and,
as mentioned previously, these states are opaque. Notice, for example, that for
$k=1$ and $k'=N$, the end-to-end configuration, there
is no such $M$ and there will be no opaque states. Whereas in configurations in which $k$ and $N+1$ are not relative primes, there will exist one or more integers $M$ that divide both $k$
and $N+1$, giving rise to opaque states in the system.
Similarly, for the transparent states we define $\theta^\textrm{T}_r=r\pi/A$
for $r=1,\dots, A-1$, as those solutions of Eq.~(\ref{eq:3})
such that $A$ simultaneously divides $N+1$ and $N-2k+1=k'-k$
but does not divide $k$ (which then would fulfill the
definition of an opaque state). If $A$ divides simultaneously
$N+1$ and $k'-k$, it also divides $2k$ and $2k'$. Arguing as above,
the solutions $\theta^\textrm{T}_r$ are also independent from $\eta$ and real. Now the corresponding states are coupled to the gain and loss, transport is efficient through these states, and their eigenvalues are insensitive to the strength of the coupling.
\subsection{Spectra and exceptional points for $N=10$ and $N=23$}
\label{sec:spectra}
In the following, we discuss the spectra for two specific chains of length
$N=10$ and $N=23$, varying (symmetrically) the position of the contacts
$k$ and $k'$. The choice of these values of $N$ is to
illustrate the case in which $N+1$ is a prime number ($N=10$),
and no solutions $\theta^\textrm{Op}_r$ or $\theta^\textrm{T}_r$
exist for any position
of the contacts. In turn, when $N=23$, $N+1=24$ is a highly composite
number (i.e., it has more divisors than any smaller integer),
and we encounter the opposite situation.
\subsubsection{Results for $N=10$}
\label{sec:N10}
We begin with the spectra for $N=10$.
Figure~\ref{fig:2} shows the real and imaginary part
of the spectra for all values of the contact positions $k$ and $k'$. As stated above,
for this value of $N$ all eigenvalues depend on $\eta$, and
there are no opaque states.
\begin{figure}
\centering
\includegraphics[scale=0.3]{fig2a.pdf}
\includegraphics[scale=0.3]{fig2b.pdf}\\
\includegraphics[scale=0.3]{fig2c.pdf}
\includegraphics[scale=0.3]{fig2d.pdf}\\
\includegraphics[scale=0.3]{fig2e.pdf}
\caption{Real and imaginary parts of the
eigenvalues $E_\theta$ as a function of the
coupling parameter $\eta$ for $N=10$. The values of $k$ which
define the specific configuration of the contacts,
are indicated in the figures.
The dashed curves included in the end-to-end
configuration (a) show the approximation given by
Eq.~(\ref{eq:Eplusmin})
around $\eta\approx 1$, which is enlarged in the inset of
the right panel. The triangles illustrate
the asymptotic result for $\eta\to\infty$, Eq.~(\ref{eq:5}).}
\label{fig:2}
\end{figure}
In Fig.~\ref{fig:2}(a) we show
the case $k=1$ and $k'=10$, where the contacts are in
the configuration. In this case there
is only one exceptional point, located at $\eta_0=1$;
the behavior
of the eigenvalues close to the exceptional point is detailed in
Sect.~\ref{sec:perturb}. The blue triangles in this figure
correspond to the asymptotic results for $\eta\to\infty$; c.f.
Eq.~(\ref{eq:5}). Figure~\ref{fig:2}(b) displays the spectrum
for $k=2$ and $k'=9$; we note that there are two
exceptional points for a value
of $\eta < 1$, signaled by the coalescence of two pairs of
eigenvalues and the appearance of two (doubly degenerate)
imaginary parts that branch out. In addition, there are
two (real) eigenvalue crossings
that do not correspond to exceptional points. The real
part of the four
complex eigenvalues that emanated from the exceptional points
coalesce again at a value $\eta > 2$ and $\operatorname{Re}
(E_\theta)=0$ thereafter. After this coalescence, two
of the imaginary parts tend to zero as $\eta$ increases, while the other
two tend to infinity according to Eq.~(\ref{eq:5}).
As we move inwards the position of $k$ and $k'$, richer
behavior of the eigenvalues is observed, with more
occurrences of exceptional
points, some of them again involving coalescences of complex
eigenvalues as well as some crossing of eigenvalues
which do not represent exceptional points. Interestingly,
for $k=5$ and $k'=6$, when the contacts are at the center of
the chain, there are 5 distinct coalescences leading to
exceptional points, all appearing at the same value $\eta=1$.
\subsubsection{Results for $N=23$}
\label{sec:N23}
We now consider the case $N=23$. As mentioned above, $N+1=24$ can be
divided by more integers than any smaller integer; in this case, it can
be divided by $M,A=2, 3, 4, 6, 8, 12$. The spectra for this value
of $N$ are shown in Fig.~\ref{fig:3}. In contrast to
Fig.~\ref{fig:2}, we observe that $E_{\theta_0}=0$
is an eigenvalue of the Hamiltonian, independently of the
location of the contacts and the value of $\eta$. This is a
consequence of the fact that $\theta_0=\pi/2$ always
satisfies Eq.~(\ref{eq:3}) for odd $N$. As we shall
see later, this state is either an opaque or a transparent state,
depending if $k$ is even or odd, respectively.
\begin{figure}
\centering
\includegraphics[scale=0.25]{fig3a.pdf}
\includegraphics[scale=0.25]{fig3b.pdf}\\
\includegraphics[scale=0.25]{fig3c.pdf}
\includegraphics[scale=0.25]{fig3d.pdf}\\
\includegraphics[scale=0.25]{fig3e.pdf}
\includegraphics[scale=0.25]{fig3f.pdf}\\
\includegraphics[scale=0.25]{fig3g.pdf}
\includegraphics[scale=0.25]{fig3h.pdf}\\
\includegraphics[scale=0.25]{fig3i.pdf}
\includegraphics[scale=0.25]{fig3j.pdf}\\
\includegraphics[scale=0.25]{fig3k.pdf}
\caption{Real and imaginary parts of the
eigenvalues $E_\theta$ as a function of the
coupling parameter $\eta$ for all
configurations of $N=23$.}
\label{fig:3}
\end{figure}
Figure~\ref{fig:3}(a) shows that the end-to-end configuration
($k=1$ and $k'=23$) exhibits one exceptional point
at $\eta=1$. The remaining eigenvalues have
a smooth dependence on $\eta$. In this configuration,
$E_{\theta_0}=0$, being independent of $\eta$, is the
only transparent state, because $k$ is odd and $A=2$
divides $N+1=24$ and $2k$ simultaneously. It is easy
to see from Eq.~(\ref{eq:4}) that $u_1(\pi/2)$ and
$u_N(\pi/2)$ are both non zero.
Figure~\ref{fig:3}(b) shows the case $k=2$ and $k'=22$. In this
case there are various eigenvalue crossings, and two values
of $\eta$ where the coalescence corresponds to (pairs of)
exceptional
points; note that the second one involves a coalescence
of the real part of four complex eigenvalues. In this case,
the only integer
$M>1$ that divides simultaneously $k=2$ and $N+1=24$ is $M=2$.
Therefore, $\theta^\textrm{Op}_{r=1} = \pi/2$ satisfies
Eq.~(\ref{eq:3}) and
$u_2(\theta^\textrm{Op}_{r=1}) = u_{22}(\theta^\textrm{Op}_{r=1})=0$
corresponding to the only opaque state of this case. Similarly,
$A=2,4$ divides simultaneously $N+1$ and $2k$; the states with $A=2$
are opaque states and therefore we have two transparent states
$\theta^\textrm{T}_{r=1}=\pi/4$ and $\theta^\textrm{T}_{r=3}=3\pi/4$.
The different panels in Fig.~\ref{fig:3} illustrate how
the spectra become more complex in terms of
eigenvalue crossings, exceptional points,
and opaque or transparent states, as the contacts are moved.
In some cases, one can also observe
avoided level crossings; see Fig.~\ref{fig:3}(j) or~(k).
We compute the number of opaque and transparent states for $N=23$
for some specific configurations of the leads;
Fig.~\ref{fig:4} shows the complete picture.
For this $N$, the divisors of $N+1$ are $2,3,4,6,8,12$; these are
all the possible values $M$ and $A$ we may have. As a first example
we consider $k=8$, and $M=2,4,8$ divide both $N+1$ and $k$; the
spectrum of this configuration is illustrated in Fig.~\ref{fig:3}(h).
Since 8 is a multiple of 2 and 4 it suffices to consider $M^*=8$.
The opaque states are then $\theta^\textrm{Op}_r=r\pi/M^*$ for
$r=1, \dots M^*-1$, and we have $7$ opaque states for this
configuration. Likewise, $A=2,4,8$ divide both $N+1$ and $2k=16$,
but since $A$ also divides $k$, there are no transparent
states in this configuration.
Consider now the configuration $k=6$ as a second example. In
this case we have that $M=2,3,6$ divide both $N+1$ and $k$,
and, using the same arguments as for $k=8$, we conclude that
the opaque states are $\theta^\textrm{Op}_r=r\pi/6$,
for $r=1,\dots 5$. With
regards to the number of transparent states, $A=2,3,4,6,12$ divide
simultaneously $N+1$ and $2k$. In this case, $A^*=12$ is a multiple
of the remaining values, and it suffices to consider it. From
the 11 states arising from $A^*$ we subtract the 5 opaque states,
finally obtaining that there are 6 transparent states in this
configuration, $\theta^\textrm{T}_r=r\pi/12$ for $r=1,3,5,7,9,11$.
\begin{figure}[t]
\hspace*{-12mm}
\includegraphics[scale=0.7]{fig4}
\caption{Number of opaque and transparent eigenstates as a function of the contact
positions $k$ for two examples of small (left) and large (right) $N$.}
\label{fig:4}
\end{figure}
The left panel of Fig.~\ref{fig:4} shows the number of opaque and transparent states in
terms of $k$ for $N=23$. It is worth stressing that the number of these states
is a very irregular function of both the system size $N$ and of the position
of the leads. Indeed, even for large system sizes, this number depends on the
divisibility properties of $N+1$ and of $k$. For example, a system of length $N=839$ shows
up to 279 opaque states and no transparent states when $k=280$; there
are 210 transparent states and 209 opaque states when
$k=210$. The right panel of Fig.~\ref{fig:4} shows the number of opaque
and transparent states for all values of $k$ for a system of size $N=839$.
However, a system of size $N=838$ will have neither opaque nor transparent
states independently of where the leads are placed, because $N+1=839$ is a
prime number.
\subsection{Perturbation theory around an exceptional point}
\label{sec:perturb}
Before we turn to the discussion of the transport coefficient in this system, we
address the behavior of the eigenstates close to the exceptional
points. In order to simplify the discussion, we shall consider the case of $N=10$
with contacts in the end-to-end configuration. In this case, there is only
one exceptional point which occurs at $\eta_0=1$ (see
Fig.~\ref{fig:2}(a)). The eigenvalues are given by
(\ref{eq:2}), where $\theta$ is determined by
(cf. Eq.~(\ref{eq:3}))
\begin{equation}
\sin(N+1)\theta + \eta^2 \sin(N-1)\theta = 0.
\label{eq:exconf}
\end{equation}
To calculate analytically the behavior of the eigenvalues
around an exceptional point, we use a simple perturbation
scheme~\cite{Benderbook1978}. We write Eq.~(\ref{eq:exconf})
generically as $F(\theta, \eta^2)=0$, from which we determine
the values of $\theta$ given the strength of the coupling
constant $\eta$. In the present case, the equation defining
$\theta$ for $\eta=\eta_0=1$ can be rewritten as
\begin{equation}
F(\theta,\eta_0^2)= 2\sin N\theta\cos\theta = 0,
\end{equation}
with the solutions $\theta_r=\frac{r\pi}{N}$ ($r=1,2,\dots,N-1$)
and $\theta_0 = \pi/2$. Since in this case $N$ is even, the
root $\theta_{N/2}$ is identical to $\theta_0$, hence at
$\eta=\eta_0=1$ these roots coalesce.
If we follow the usual perturbation scheme, we write
$\eta^2=\eta_0^2+\epsilon$, propose a solution of the form
$\theta=\theta_0+\epsilon\theta_{(1)}+\dots$ in powers of
$\epsilon$, and solve $F(\theta, \eta^2)=0$, which is also
written as a series expansion in $\epsilon$. Each term
of that series must be equal to zero, which is used to
obtain $\theta_{(1)}$. However, this procedure breaks
down when $\theta_0$ and $\eta_0$ define an exceptional
point. Indeed, to first order in $\epsilon$ the
expansion reads
\begin{equation}
F(\theta, \eta^2) \approx F(\theta_0, \eta_0^2) +
\epsilon\left(\theta_{(1)} \frac{\partial F}{\partial\theta}(\theta_0,\eta_0^2)
+ \frac{\partial F}{\partial (\eta^2) }(\theta_0,\eta_0^2) \right),
\end{equation}
but at the exceptional point we have
\begin{eqnarray}
\label{eq:excep_points}
\frac{\partial F}{\partial \theta}(\theta_0, \eta_0^2) & = &
0, \\
\frac{\partial F}{\partial (\eta^2)}(\theta_0, \eta_0^2) & \neq & 0,
\end{eqnarray}
implying that we cannot choose $\theta_{(1)}$ such that
the first
order term vanishes. We emphasize that the condition given by
Eq.~(\ref{eq:excep_points}) defines the exceptional point.
To overcome the failure of the usual perturbation scheme,
we write the solution for $\theta$ as
$\theta=\theta_0+\epsilon^{1/2}\theta_{(1)}+\epsilon\theta_{(2)}\dots$,
which is analogous to the expansion proposed in
\cite{MoiseyevPRA1980, GarmonRotter2012}.
Then, to first order in $\epsilon$ we have
\begin{equation}
\label{eq:pert}
F(\theta, \eta^2) \approx F(\theta_0,\eta_0^2)
+ \epsilon^{1/2} \theta_{(1)}\frac{\partial F}{\partial \theta}(\theta_0, \eta_0^2)
+ \epsilon\left(\frac{1}{2}\theta_{(1)}^2
\frac{\partial^2 F}{\partial\theta^2}(\theta_0,\eta_0^2) +
\frac{\partial F}{\partial(\eta^2)}(\theta_0,\eta_0^2) +\theta_{(2)}
\frac{\partial F}{\partial \theta}(\theta_0,\eta_0^2)\right).
\end{equation}
The term of order $\epsilon^{1/2}$ and the $\theta_{(2)}$ term vanish identically
at the exceptional point, due to Eq.~(\ref{eq:excep_points}) above,
and from the rest of the term of order $\epsilon$ we can obtain
$\theta_{(1)}$. Explicitly, the first order term in $\epsilon$
leads to
\begin{equation}
-2N\theta_{(1)}^2\cos{\frac{N\pi}{2}} + \sin{\frac{(N-1)\pi}{2}} = 0,
\end{equation}
and we have $\theta_{(1)} = \pm\frac{\mathrm{i}}{\sqrt{2N}}$.
Consequently, near the exceptional point we obtain
\begin{equation}
\theta^\pm \approx \frac{\pi}{2} \pm\mathrm{i}\frac{\epsilon^{1/2}}{\sqrt{2N}}
= \frac{\pi}{2} \pm\mathrm{i}\sqrt{\frac{\eta^2-1}{2N}},
\label{eq:phipm1}
\end{equation}
in terms of which, the eigenvalues read
\begin{equation}
\label{eq:Eplusmin}
E_{\theta^\pm} = 2\cos\theta^\pm \approx \mp 2\mathrm{i}\sinh{\sqrt{\frac{\eta^2-1}{2N}}}.
\end{equation}
Notice that the two solutions of $\theta^\pm$ indicate a coalescence
at the exceptional point, and a ``complexification'' of the
eigenvalues in the $\mathcal{PT}$-symmetric broken phase
($\eta>1$). The correction term proportional to the square
root of $\eta^2 -1$,
describes well the results obtained by direct numerical
diagonalization of the Hamiltonian matrix in the proximity of
the exceptional point; see Fig.~\ref{fig:2}(a).
The treatment described above and Eq.~(\ref{eq:excep_points}) can
be generalized to the case when more eigenvalues coalesce at
the exceptional point. Indeed, if the second (and higher)
derivatives of $F(\theta,\eta)$ also
vanish at the exceptional point, the appropriate perturbation
expansion should be proportional to $\epsilon^{1/p}$, where the
first non-vanishing derivative is
$\partial^{p} F/ \partial\theta^{p}$, and $p$ points coalesce at
the exceptional point. In this situation $p$ distinct eigenvalues
coalesce to the same value \cite{MoiseyevPRA1980, GarmonRotter2012}.
\section{Transport}
\label{sec:transportPT}
Having now a thorough description of the spectra and
eigenvectors of the $\mathcal{PT}$ symmetric chain, we turn
to the discussion of the transport properties of the system.
First of all, for all non-opaque states in the unbroken
$\mathcal{PT}$-symmetry
phase we have $\xi_{E_\theta}=1$, as can be checked
directly using the explicit expression of the coefficients,
Eq.~(\ref{eq:4}). This indicates that transport in the
eigenstates with real eigenvalues
is efficient. On the other hand, for states in the $\mathcal{PT}$-broken
symmetry phase, some eigenvalues become complex and for the corresponding
eigenstates $\xi_{E_\theta}$ is no longer equal to one.
In view of Eq.~(\ref{eq:sym2}), it is straight forward
to see that $\xi_{E_\theta}\xi_{E_\theta^*}=1$ as mentioned previously.
The actual values of the transport coefficient for the
eigenstates near the exceptional point can be evaluated
using the perturbation expansion for $\theta$. From
Eq.~(\ref{eq:phipm1}) we can write
\begin{equation}
\xi_{E_\theta,E_\theta^*} \sim 1\pm \sqrt{N(\eta^2-1)/2}\sim e^{\pm \sqrt{N(\eta^2-1)/2}},
\label{eq:xi1}
\end{equation}
for $\eta > 1$, where the exponential form was chosen
merely to enforce the fact
that $\xi_{E_\theta,{E_\theta}^*}$ are positive, and
that their product
$\xi_{E_\theta} \xi_{E_\theta^*}$ must
be equal to one.
\begin{figure}[t]
\centering
\includegraphics[scale=0.32]{fig5a.pdf}
\includegraphics[scale=0.32]{fig5b.pdf}
\includegraphics[scale=0.32]{fig5c.pdf}\\
\includegraphics[scale=0.32]{fig5d.pdf}
\includegraphics[scale=0.32]{fig5e.pdf}
\caption{Transport coefficient $\xi_{E_\theta}$ as a
function of $\eta$ for all eigenstates and all
configurations of the contacts for $N=10$.
Sub-figure (a) includes the
analytical expansion of the transport coefficient (dashed red curve
in the inset) for $\eta\approx 1$ Eq. \ref{eq:xi1}, as well as the asymptotic expansion (blue triangles) for $\eta\rightarrow \infty$ Eq. \ref{eq:zetaetainf}.}
\label{fig:5}
\end{figure}
In the limit $\eta\rightarrow\infty$ the
$\xi_{{E_\theta,E_\theta^*}}$ are given by
\begin{equation}
\xi_{E_\theta,E_\theta^*} \sim [4\eta^{2(N-1)}]^{\pm1}.
\label{eq:zetaetainf}
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[scale=0.32]{fig6a.pdf}
\includegraphics[scale=0.32]{fig6b.pdf}
\includegraphics[scale=0.32]{fig6c.pdf}\\
\includegraphics[scale=0.32]{fig6d.pdf}
\includegraphics[scale=0.32]{fig6e.pdf}
\includegraphics[scale=0.32]{fig6f.pdf}\\
\includegraphics[scale=0.32]{fig6g.pdf}
\includegraphics[scale=0.32]{fig6h.pdf}
\includegraphics[scale=0.32]{fig6i.pdf}\\
\includegraphics[scale=0.32]{fig6j.pdf}
\includegraphics[scale=0.32]{fig6k.pdf}
\caption{Same as Fig.~\ref{fig:5}, for $N=23$ and all
configurations of the contacts. Only the non-opaque states
are included.}
\label{fig:6}
\end{figure}
Figure~\ref{fig:5} shows the transport coefficient
$\xi_{E_\theta}$ as a function of $\eta$ for all the
eigenfunctions and all contact configurations for $N=10$.
As we have seen, since $N+1$
is prime there are no opaque nor transparent states, and all eigenstates
have a well-defined transport coefficient. The perturbative
approximation, Eq.~(\ref{eq:xi1}), is shown
in Fig.~\ref{fig:5}(a) by the red dashed curves in the inset;
the asymptotic limit of $\xi_{E_\theta}$,
Eq.~(\ref{eq:zetaetainf}), is illustrated by the
blue triangles. We note that there are some configurations
that display states with efficient transport; see
Fig.~\ref{fig:5}(b), (c) and~(d). These states with efficient
transport are not transparent states, but still have real eigenvalues, despite the
fact that some eigenvalues for other states are complex for the same value of $\eta$.
Similarly, Fig.~\ref{fig:6} shows the transport coefficient
$\xi_{E_\theta}$ as a function of $\eta$ for all non-opaque
states and all contact configurations for $N=23$.
Again, we note that in most configurations
there are some transport coefficients
which are identical to $1$ for all values of $\eta$, i.e.,
transport is effective for some eigenstates in those
configurations. Some of those states correspond to
transparent states, but not necessarily all of them.
Notice that for
$k=8$ in Fig.~\ref{fig:6}(h) the opposite is observed:
beyond certain value of $\eta$, all transport
coefficients are different from $1$ and transport
is deficient; this configuration ($k=8$, $k'=16$)
corresponds to the maximum number of opaque states
for $N=23$, having no transparent states. In this
case all states are either
opaque, or have complex eigenvalues, and transport
is always deficient beyond $\eta \geq 1$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.3]{fig7a.pdf}
\includegraphics[scale=0.3]{fig7b.pdf}
\includegraphics[scale=0.3]{fig7c.pdf}\\
\includegraphics[scale=0.3]{fig7d.pdf}
\includegraphics[scale=0.3]{fig7e.pdf}
\includegraphics[scale=0.3]{fig7f.pdf}
\caption{Absolute squared value of eigenstates as function of the site position
$n$ that undergo a single (upper panels) or a pair (lower panels) of exceptional points, for different values of $\eta$ and $N=23$ in the configuration $k=6$ and $k'=18$. For each eigenstate, the corresponding value of the energy at a certain $\eta$ is given with $E_i$. In a similar way, we have
included the values of the transport coefficient $\xi_i$. (a) and~(d) correspond to $\eta=0.5$, where all corresponding eigenvalues are real; (b) and~(e) $\eta=1.5$, and all illustrated states have experienced one exceptional point; (c) and~(f) correspond to $\eta=2.5$.}
\label{fig:7}
\end{figure}
The behavior of the transport coefficients
$\xi_{E_\theta,E_\theta^*}$ is consistent with the strong
localization towards the contacts for the
$\mathcal{PT}$-broken symmetry states. This is
illustrated in
Fig.~\ref{fig:7} where we show the modulus squared of
two pairs of eigenfunctions whose eigenvalues coalesce
at different values of $\eta$, in the $k=6$ and $k'=18$
configuration for $N=23$. In Figs.~\ref{fig:7}(a)
and~(d) for $\eta=0.5$, the eigenvalues
are real, and their eigenfunctions are extended.
Figures~\ref{fig:7}(b) and~(e) display the states at
$\eta=1.5$ after crossing an exceptional point;
localization around the contacts is apparent. In
Figs.~\ref{fig:7}(c) and~(f)
we illustrate the case for $\eta=2.5$. The pair of states
in (c) localize in one or the other side of the chain,
between the edge of the chain and the contact.
A similar situation occurs with the states
illustrated in (e). The states
in panel (f) are strongly localized in the gain or
in the loss. These states correspond to energies approximated
asymptotically by Eq.~(\ref{eq:5}).
\begin{figure}[h!]
\centering
\includegraphics[scale=0.31]{fig8a.pdf}
\includegraphics[scale=0.31]{fig8b.pdf}
\includegraphics[scale=0.31]{fig8c.pdf}
\caption{Absolute squared value of three transparent states
for $N=23$ in the configuration $k=6$ and $k'=18$.}
\label{fig:8}
\end{figure}
As we established above, in several configurations of the
contacts some states have
real eigenvalues that are independent of $\eta$.
The modulus squared of some examples of transparent
states is illustrated
in Fig.~\ref{fig:8} for the configuration $k=6$
and $k'=18$ ($N=23$) and $\eta=0.5$, $\eta=1.5$ and
$\eta=2.5$, respectively. These states
become increasingly
concentrated in the center of the chain, between the
gain and the loss, as the
magnitude of the corresponding eigenvalues increases. In
Fig.~\ref{fig:9} we show the modulus square
of the 5 distinct opaque states of this configuration,
illustrating that all of them have nodes at the
gain and loss. In addition, we observe that in this
case they also have nodes at the central site $j=12$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{fig9.pdf}
\caption{Modulus square of the 5 distinct opaque states
for $N=23$ and $k=6$. As shown, all the opaque states
have nodes at the gain and loss ($k=6$ and $k'=18$).}
\label{fig:9}
\end{figure}
\section{Summary and Conclusions}
\label{sec:Conclusions}
We have presented a detailed analysis of the transport
properties in a one dimensional
$\mathcal{PT}$ tight binding chain. This was achieved by
first deriving a generalized continuity equation for the
density. The terms responsible for the non-hermiticity of
the Hamiltonian appear as gain and loss terms in the
continuity equation. Transport can be quantified via a single
number which we called the transport
coefficient. In the $\mathcal{PT}$-unbroken symmetry
phase this coefficient is
equal to one, which implies that transport is efficient in
the sense that the inflow of the density equals the
outflow. For states with broken $\mathcal{PT}$ symmetry
(complex eigenvalues),
the transport coefficient is different from one, implying
that density either accumulates or is depleted within the
system. To study the detailed behavior of the chain, we
obtained general expressions for the eigenvalues and
eigenvectors of the system. This analysis led us to note
that if $N+1$ (where $N$ is the length of the chain) and
$k$ (the position of the gain) have common divisors, the
system may have eigenstates that do not couple to the gain
and loss, and thus do not transport density through
the chain. We call these states opaque. Similarly,
if $N+1$ and $2k$ have common divisors which do not divide $k$,
the eigenstates have real eigenvalues independently of the
coupling $\eta$ and transport is efficient; we call
these states transparent. To illustrate these
phenomena we analyze the eigenvalues for chain
lengths $N=10$ and $23$. In the first case $N+1$ is prime
and there are no opaque nor transparent states; interestingly,
for $k=5$ and $\eta>1$ transport is deficient for all states. In the
second case $N+1$ is a highly composite integer,
opaque and transparent eigenstates may be present and the behaviour
is richer in terms of the transport.
For instance, we
find that in addition to transparent states, there are some
states with real eigenvalues
for all values of $\eta$, which have a weak dependence on
$\eta$, and
their eigenfunctions do not vanish at the contacts.
The existence of such states as well as transparent states allows
having efficient transport beyond the value of $\eta$ at which $\mathcal{PT}$-symmetry is broken.
Interestingly, for $N=23$ and $k=8$, which corresponds to
the case with the maximum number of opaque states, beyond
certain $\eta$ transport is deficient in all states. This suggests
that opaque states play a role inhibiting transport.
For completeness, we have
presented a simple perturbation scheme to study the
eigenvalues and eigenvectors around the exceptional
points. The development of the perturbation scheme
provides a simple rule to obtain the value of
$\eta$ for which exceptional points appear.
Our results show that the simple $\mathcal{PT}$-symmetric
tight binding chain displays rather complex spectral
properties in
terms of the position of the gain and the loss. These imply rich
transport behavior depending on the presence of opaque and
transparent states, as well as
possibly other states whose eigenvalues remain real
independently of the strength of the contacts. This amounts
to a classification of eigenstates in terms of transport
which should be observed in experiments using, for example, optical
wave-guides or microwave resonators .
\begin{acknowledgments}
The authors gratefully acknowledge financial support from
CONACyT Proyecto Fronteras 952, CONACyT Posdoctoral Fellowship
Programme (AO), CONACyT Proyecto A1-S-13469 and the UNAM-PAPIIT
research grants IG-100819 and IA-103020.
\end{acknowledgments}
|
\section{Introduction}
Training reinforcement learning (RL) agents often requires millions of steps on the target task~\cite{duan2016bdr,kalashnikov2018qsd}.
While work is progressing on more sample-efficient algorithms~\cite{gu20162016cdq,chebotar2017cmm}, more progress is needed before agents can learn efficiently.
If we restrict the target tasks to be \emph{near} a distribution of training tasks, then we can apply meta-learning to train an agent to rapidly adapt to a task~\cite{schmidhuber1987eps,bengio1991lsl}.
As an example, for a straightening robot, the target task might consist of the desired destinations of toys, dishes, and clothing for a specific home, and the training distribution would consists of a set of such tasks.
In this paper, we make use of meta-learning and focus on task distributions where the reward function varies across tasks, or equivalently where the desired behavior varies across tasks and is optimal under a corresponding reward function; e.g. tasks for the straightening robot.
This is distinct from task distributions in which the reward function is constant across tasks but the environment varies, such as maze solving~\cite{jaderberg2017rlu}.
One class of meta-learning techniques for adapting to reward-varying tasks adapts the policy to rewards received during the task~\cite{finn2017mam,duan2016rl2,levine2018rlc}.
Unfortunately, these techniques require a reward function for the target task and specifying a reward function is often tedious, particularly as the task complexity grows.
Another class of meta-learning methods for adapting to reward-varying tasks adapts the policy to mimic expert demonstrations of the target task~\cite{yu2018osi,james2018tec}.
Unfortunately, providing demonstrations can be inefficient. %
For example, without interaction it is not always obvious what the agent is unclear about and what type of a demonstration, how complete of a demonstration, or how many demonstrations to give.
In this paper, we investigate if it is possible to train an agent to interactively maximize another agent's reward, without access to the reward function and without explicitly receiving demonstrations. We call the agent whose reward function is being maximized the "prime" agent, suggesting that it is the primary agent, and we call the other agent the "helper" agent.
In our proposed method, a helper agent is simultaneously meta-learned with a prime agent, who serves as a surrogate for a human prime.
By making the prime agent aware of their joint task-specific reward function, communication emerges allowing the helper to rapidly infer the reward function and adapt to the task.
Our work is most similar to that of~\citet{mordatch2018egc}.
While they address a different task distribution, the main difference is the asymmetry of knowledge and roles in our prime-helper tasks.
Our primary contribution is a method for training a helper agent to, through interaction, rapidly adapt its policy to maximize another agent's reward. We evaluate our method on simulated foraging tasks. Our preliminary results demonstrate that 1) the helper infers the task from a handful of prime agent actions and in some cases zero prime actions (which is itself a communication), 2) the prime agent receives more reward with the helper agent than without, and 3) the emerged communication and the prime agent's delegation to the helper agent is explicit and intentional.
\section{Methodology} \label{sec:methodology}
\begin{figure}
\begin{center}
\includegraphics[height=2in]{meta_training.eps}
\end{center}
\caption{
Training episode $i$ begins by randomly selecting a task and resetting the recurrent state for the prime ($p$) and helper ($h$) agents. On every step, $t$, of the episode, the agents receive their respective observations, $o^{p}_{i,t}$ and $o^h_{i,t}$, and select their respective actions, $a^{p}_{i,t}$ and $o^h_{i,t}$, and a joint reward is stored, $r_{i,t}$. The prime agent's observation, $o^{p}_{i,t}$, informs it of the the task for the episode, depicted with color in the gridworld example. At the end of the episode, the stored rewards are used to improve the prime and helper policies.
}
\label{fig:method}
\end{figure}
Similar to recent work in meta-learning, we train agents whose policies are functions of a recurrent neural network~\cite{duan2016rl2}.\footnote{This technique is commonly described as making use of a recurrent network but the main characteristic is that the policy is modified because of experience and not a change in network weights. For example attention over observations could also be used~\cite{sukhbaatar2015eem,vaswani2017aay}.}
Figure~\ref{fig:method} illustrates our proposed training method.
Each episode consists of the helper agent and the prime agent acting and receiving agent specific observations in a task drawn from a task distribution.
During an episode, the policy weights are fixed and actions are a function of the hidden states.
At the end of each training episode, the policy weights of both agents are moved in a direction that would have increased the prime's reward for that episode's task.
Key to our method is that, during the episode, only the prime has knowledge of the task's reward function.
Since the prime is the only agent with knowledge of the reward function yet the helper is available to help, communication emerges, and the helper is trained to infer the task rewards from the prime and adapt its policy to contribute to the prime's reward.
We simultaneously train the prime and the helper specifically to allow communication to emerge. By emerging the communication, we allow the agents to communicate in ways that a programmer may not have thought of. This is particularly true when optimal performance requires the prime to maintain beliefs about the helper.
As an example, in our experiments, we noticed the prime occasionally communicating after sub-optimal helper behavior indicating that the prime was monitoring the helper's performance.
\section{Experiment: 1-D Foraging} \label{sec:experiment}
\subsection{Task} \label{sec:task}
\begin{figure}
\begin{center}
\includegraphics[width=5.5in]{communication.eps}
\end{center}
\caption{
Two agents collect "good" objects and avoid "bad" objects in a 5 cell gridworld. The prime agent observes which objects are good and which are bad, depicted in color above the gray objects, and must communicate this to the helper agent. Learned policy: (a) if the first object to appear in the episode is a good object, then the prime agent collects that object, (b) if the first object to appear in the episode is a bad object, then the prime agent does not move. In both cases the helper infers the task and collects all future good objects.
}
\label{fig:communication}
\end{figure}
Figure~\ref{fig:communication} depicts two tasks drawn from our distribution of "foraging" tasks. Each task consists of the helper and prime located on a one dimensional gridworld, consisting of five cells.
There are three actions for each agents: move left, move right, and don't move.
Objects randomly drawn from two object classes appear alternately in the right-most and left-most cells, persist for 9 time steps, and then disappear.
An observation, $o_t$, consists of a set of binary strings, one for each cell.
A cell's binary string is of length 5 for the helper and 6 for the prime, with one bit for the presence of the prime, one bit for the presence of the helper, one bit for the presence of an object, two bits for the class of the object, and, for the prime, one bit indicating whether it is good to pick up objects of this class.
If an action, including no action, moves an agent to a cell that contained an object then the object disappears and a reward is stored for that timestep; $+1$ is stored for collecting a good object and $-1$ for collecting a bad object.
Finally, a reward of $-0.1$ is stored for all timesteps in which the prime moves.
This last penalty is added to encourage minimal prime movement, so that the ability to communicate and delegate to the helper is clear in the results.
Each episode runs for $100$ time steps, with exactly $20$ objects appearing in an episode.
As a baseline, we also train a prime agent acting alone without a helper agent, all other aspects of the model and training are identical.
\subsection{Model} \label{sec:model}
We use identical, but independent, deep recurrent Q-networks as the policy for each agent~\cite{mnih2013pad,hausknecht2015drq}. A single 200 unit LSTM is used as the recurrent network~\cite{hochreiter1997lst}. For each agent, on each timestep, their observation, $o_t$, described in section~\ref{sec:task}, is flattened and fed to the LSTM, and a linear layer is applied to the LSTM output, $h_t$, resulting in a 3 dimensional vector, $q_t$, which estimates the expected future reward for each action.
\begin{equation}
h_t = LSTM(o_t,h_{t-1})
\end{equation}
and
\begin{equation}
q_t = Wh_t+b,
\end{equation}
where $LSTM(o_t,h_{t-1})$ represents the LSTM update equations~\cite{hochreiter1997lst}. Either $\argmax(q_t)$ is chosen for $a_t$, or, during training, a random action is chosen with probability $0.05$ ($\epsilon$-greedy exploration).
At the end of each episode a gradient step is taken for each agent, using Adam with the default parameters, on the Q-Learning objective function\cite{kingma2015ams}:
\begin{equation}
\mathcal{L}(\Theta) := \sum_t[q_t\cdot{\vec{a}}_t - (r_t + \gamma \max{}q_{t+1})]^2,
\end{equation}
where $\Theta$ are the parameters of the LSTM and linear layer, one set for each agent, and $\vec{a}_t$ is $a_t$ in one-hot form. We set $\gamma=0.95$ and we do not use a target network. We use a batch size of 100 episodes for each gradient update, and we perform 10,000 gradient updates.
\subsection{Results} \label{sec:results}
\begin{figure}
\begin{minipage}[t]{2.64in}
\centering
\includegraphics[width=1\textwidth]{episode_actions.eps}
\caption{
The prime agent moves less and mostly at the start of an episode when trained with a helper. The aliasing is due to the regular appearance of objects in an episode.
}
\label{fig:episode_actions}
\end{minipage}
\hfill
\begin{minipage}[t]{2.64in}
\centering
\includegraphics[width=1\textwidth]{episode_rewards.pdf}
\caption{
The episode reward due to the prime agent's actions rises, peaks, and then drops during training as the agents learn to communicate and delegate tasks to the helper.
}
\label{fig:episode_rewards}
\end{minipage}
\end{figure}
The baseline prime, which acts alone, takes an average of $29.48$ actions per episode and receives an average reward of $5.99$ per episode out of a maximum expected reward of $10$, compared with $3.98$ actions per episode and an average reward of $9.29$ when assisted by a helper\footnote{20 objects appear per episode, with 10 being good objects on average, resulting in $+10$ reward if all good objects are collected and the prime does not move}.
Thus, the helper is helping the prime.
See figure~\ref{fig:episode_actions} for a plot of the prime's movement during an episode vs. the baseline prime.
Figure~\ref{fig:communication} shows the learned policies, which we constructed by observing trained agents.
If the first object to appear is a "good" object, then the prime collects the object, otherwise it does not move.
In either case, the helper collects all future good objects, even inferring from the prime's lack of motion on an initial bad object that the other object class is good.
Figure~\ref{fig:episode_actions} shows that, when assisted by the helper, the prime acts early in the episode and is largely dormant thereafter.
This, coupled with the observed policies of figure~\ref{fig:communication} and the high reward due to the helper's actions of figure~\ref{fig:episode_rewards}, demonstrates that the helper quickly infers and adapts to the task.
The emerged communication and policies demonstrate that the prime agent intentionally delegates to the helper. The delegation can be seen by the prime's lack of movement late in episodes, figure~\ref{fig:episode_actions}, and the prime's low contribution to the total reward, figure~\ref{fig:episode_rewards}. Further, even when a good object is next to the prime, it will wait for the helper to cross the grid rather than collecting it itself.
Lastly, we note that the agents move through three phases during training, as seen in figure~\ref{fig:episode_rewards}: i) first the prime learns to collect good objects, ii) then the prime and the helper learn to communicate and jointly collect good objects, iii) finally the agents learn to have the helper collect the good objects.
\section{Conclusion and Future Work} \label{sec:conclusion}
We presented a method for training a helper agent to maximize the reward of another agent, without requiring a reward function or explicit demonstrations for the target task.
The results demonstrated that the helper agent adapts quickly and provides a net benefit to the prime agent.
We are currently working to achieve the optimal joint policy of the prime agent taking a single step when the first object to appear is a good object, instead of collecting the object.
We are also working on more complex tasks, including a 3D environment with first person pixel observations.
\iffalse
\begin{figure}
\begin{subfigure}[b]{1.0\textwidth}
\includegraphics[width=5in]{occupancy_0.eps}
\end{subfigure}
\begin{subfigure}[b]{1.0\textwidth}
\includegraphics[width=5in]{occupancy_1.eps}
\end{subfigure}
\begin{subfigure}[b]{1.0\textwidth}
\includegraphics[width=5in]{occupancy_2.eps}
\end{subfigure}
\begin{subfigure}[b]{1.0\textwidth}
\includegraphics[width=5in]{occupancy_3.eps}
\end{subfigure}
\caption{
This doesn't really show what I wanted it to.
TODO:
- pick better colors,
- properly scale,
- add axis labels
}
\end{figure}
\fi
|
\section{Introduction}
The quasar phenomenon has been studied for more than 50 years \citep{Schmidt63}.
Quasar central engines are believed to be
accretion disks feeding material onto supermassive black
holes \citep{Rees84}.
In the standard picture of SMBH growth,
the rest mass energy of accreted mass is divided between radiation
and black hole mass growth via the radiative efficiency $\epsilon$,
implying that emission of quasar light is concomitant with black hole growth.
In the local Universe, dormant
supermassive black holes reside at the centers of all massive
galaxies, and galaxies with more stellar mass in a spheroidal bulge
host more massive black holes \citep{FM00,Gebhardt00}.
The connection between distant quasar ``progenitors'' and ``relic'' supermassive black holes is
encapsulated in the Soltan argument, which states that the
integrated emission from quasars over cosmic time is proportional
to the total mass in supermassive black holes today via the radiative efficiency \citep{Soltan82},
\begin{equation} \label{eqn:soltan}
\rho_{\rm BH}(z=0) = \int_0^\infty \frac{dt}{dz} dz \int_{L_{\rm min}}^\infty \frac{1-\epsilon}{\epsilon} \frac{L_{\rm q}}{c^2} \Phi(L,z) dL,
\end{equation}
where $\Phi(L,z)$ is the quasar luminosity function in some observed band and $L_{\rm q}$ represents the bolometric luminosity of a quasar with observed luminosity $L$.
It then follows that the radiative efficiency can be inferred
by
commensurating the energy density of quasar light with
the inferred mass density of supermassive black holes in the
local Universe.
Applications of this argument by various groups
have measured radiative efficiencies of $\approx10\%$ (e.g. \citealt{YT02,Shankar09,Ueda14})
after statistically correcting for the obscured quasar population,
consistent with predictions of analytic thin disk
accretion models
in general relativity (e.g. \citealt{Thorne74}).
However, current understanding is that these thin disk models represent an idealization
as they fail to reproduce quasar spectral energy distributions (e.g. \citealt{KB99}).
Numerical simulations of quasar accretion disks reveal a more complex
picture of geometrically thick disks supported by radiation pressure,
within which a substantial fraction of the radiation can be advected into the central black
hole, potentially dramatically lowering the radiative
efficiency (e.g. \citealt{Sadowski14}).
The bolometric luminosity of a quasar accretion disk is typically written as
\begin{equation} \label{eqn:acc}
L_{\rm bol} = \epsilon \dot{M} c^2 = \frac{\epsilon}{1-\epsilon} \dot{M}_{\rm BH} c^2,
\end{equation}
where $\dot{M}$ is the total mass inflow rate and $\dot{M_{\rm BH}}$ is the growth rate of the black hole.
The maximum luminosity of a quasar can be estimated by equating the gravitational acceleration from the black hole with radiation pressure on electrons in the infalling gas, known as the Eddington luminosity,
\begin{equation}
L_{\rm edd} = \frac{4\pi GM_{\rm BH}cm_{\rm p}}{\sigma_{\rm T}}.
\end{equation}
From equation~(\ref{eqn:acc}), the characteristic timescale for growing a black hole at the Eddington limit, the Salpeter time $t_{\rm S}$, is then
\begin{equation} \label{eqn:salp}
t_{\rm S} = \frac{1-\epsilon}{\epsilon} \frac{4\pi G m_{\rm p}}{c\sigma_{\rm T}} \approx 45\,{\rm Myr} \times \frac{(1-\epsilon)/\epsilon}{9}.
\end{equation}
Assuming a fixed Eddington ratio $L/L_{\rm Edd}$, a black hole with seed mass $M_{\rm seed}$ at time $t_{\rm seed}$ will then grow as
\begin{equation}
M_{\rm BH}(t) = M_{\rm seed} e^{(L/L_{\rm Edd})([t-t_{\rm seed}]/t_{\rm S})}.
\end{equation}
A lower radiative efficiency would decrease $t_{\rm S}$, and thus could alleviate the tension with growing
supermassive black holes at the highest redshifts.
Luminous quasars with
$\gtrsim10^9$ $M_\odot$ black holes have been discovered at $z > 7$
when the Universe was less than 800 Myr
old \citep{Mortlock11,Banados18,Yang18,Wang18}.
Growing the observed $10^9$ $M_\odot$ black holes at $z\gtrsim7$ from a
$100$ $M_\odot$ initial seed requires $\approx16$
e-foldings, which for $\epsilon=0.1$ corresponds to continuous Eddington-limited accretion for
roughly the entire age of the Universe at that time.
It seems implausible that these black holes have been growing since
the Big Bang. However, demanding a later formation epoch, consistent with expectations
for the death of the first stars in primordial galaxies at $z\sim 20$--$50$ (e.g. \citealt{Tegmark97}),
implies seeds in excess of $1000$ $M_\odot$ \citep{Mazzucchelli17,Banados18} which are then inconsistent with being
stellar remnants \citep{Heger03}.
Two classes of models have been proposed to resolve this tension. In the
first, the black holes grow faster, either by explicitly violating the Eddington luminosity limit (e.g. \citealt{VR05})
or by accreting at a much lower radiative efficiency (e.g. \citealt{Madau14}). In the
second, the initial seeds were much more massive than stellar remnants,
either by forming monolithically via direct collapse of primordial
gas (e.g. \citealt{BL03}) or by coalescence of a dense Population III stellar cluster (e.g. \citealt{Omukai08}).
A method to directly measure the radiative efficiency of the highest redshift
quasars would shed light on this tension and distinguish between these models.
Indeed, the radiative efficiency inferred from the Soltan argument is both
indirect and has negligible contribution from the rare $z > 7$ quasar population.
The highest redshift quasars known reside within the ``epoch of
reionization,'' when the first stars, galaxies, and accreting black
holes ionized the hydrogen and helium in the Universe for the first
time after cosmological recombination \citep{LF13,DF18}. During reionization, abundant neutral hydrogen
in the intergalactic medium (IGM) is expected to imprint two distinct Ly$\alpha$ absorption
features on the rest-frame ultraviolet (UV) spectra of quasars. First,
the ``proximity zone'' of enhanced Ly$\alpha$ transmission resulting from the quasar's own ionizing radiation
will be truncated by neutral hydrogen along our line-of-sight \citep{CH00}. Second, a damping wing
signature redward of rest-frame Ly$\alpha$ will be present, arising
from the Lorentzian wings of the Ly$\alpha$ resonant absorption
cross-section \citep{ME98}.
The two highest redshift quasars known, ULAS
J1120+0641 \citep{Mortlock11} (henceforth J1120+0641) at $z=7.09$, and
ULAS J1342+0928 \citep{Banados18} (henceforth J1342+0928) at $z=7.54$,
both exhibit truncated proximity zones \citep{Bolton11,Davies18b}
(compared to similarly-luminous quasars at
$z\sim6$--$6.5$, \citealt{Eilers17}) and show strong evidence for damping
wing absorption \citep{Mortlock11,Bolton11,Greig17b,Banados18,Davies18b}.
As we show below, an extension of the Soltan argument to \emph{individual quasars} is uniquely
possible at $z\gtrsim7$ due to the presence of neutral hydrogen in the
IGM along our line of sight to the quasar.
\begin{figure*}[htb]
\begin{center}
\resizebox{18cm}{!}{\includegraphics[trim={1em 1em 1em 1em},clip]{both_spectra.pdf}}
\end{center}
\caption{Quasar spectra and model fits close to rest-frame Ly$\alpha$. The grey and pink curves show the observed spectra and corresponding $1\sigma$ noise for J1120+0641 (left, VLT/FORS2, \citealt{Mortlock11}) and J1342+0928 (right, Magellan/FIRE, \citealt{Banados18}). The black curves show the 500 km/s binned spectra used in our statistical analysis. The vertical shaded bands correspond to foreground metal absorption systems which were masked prior to binning. The thick blue curves show the PCA models for the intrinsic quasar spectra, while the thin blue curves show 100 draws from the distribution of covariant prediction uncertainty. The orange curves show the best-fit mean absorption models, corresponding to ($\langle x_{\rm HI} \rangle$, $N_{\rm ion}$) = (0.65, $1.2\times 10^{71}$) for J1120+0641 and ($\langle x_{\rm HI} \rangle$, $N_{\rm ion}$) = (0.90, $7.0\times 10^{70}$) for J1342+0928. The red curves show models assuming a fully neutral Universe ($\langle x_{\rm HI} \rangle=1$) and $N_{\rm ion} = 5.7\times10^{72}$ and $1.8\times10^{72}$ for J1120+0641 and J1342+0928, respectively, corresponding to the maximum Ly$\alpha$ absorption with $N_{\rm ion}$ from equation~(\ref{eqn:lrsa}) assuming $\epsilon=0.1$. The shaded regions around the orange and red curves correspond to the central 68\% scatter of forward modeled mock spectra.}
\label{fig:spectra}
\end{figure*}
\begin{figure*}[htb]
\begin{center}
\resizebox{18cm}{!}{\includegraphics[trim={4em 1em 4em 1em},clip]{both_posterior.pdf}}
\end{center}
\caption{Joint posterior PDF of IGM neutral fraction and number of emitted ionizing photons. The lower panels show the two-dimensional posterior PDFs of $\langle x_{\rm HI} \rangle$ and $N_{\rm ion}$ inferred from the spectra of J1120+0641 (left) and J1342+0928 (right). The inner and outer grey contours enclose the central 68\% and 95\% probability, respectively. The top panels show the corresponding marginalized posterior PDFs for $N_{\rm ion}$. The vertical lines in the top panels indicate the expected $N_{\rm ion}$ from equation~(\ref{eqn:lrsa}), assuming $\epsilon=0.1$ and the measured black hole masses of J1120+0641 ($M_{\rm BH}=2.47\times10^9$ $M_\odot$) and J1342+0928 ($M_{\rm BH}=7.8\times10^8$ $M_\odot$), with a shaded region indicating a $1\sigma$ systematic uncertainty of 0.4 dex in the mass measurements.}
\label{fig:2dpost}
\end{figure*}
\section{The Local Reionization Soltan Argument} \label{sec:soltan}
The simple form of our analogy to the Soltan argument is as follows.
The imprint of the neutral IGM on a reionization-epoch quasar spectrum
constrains the total number of ionizing photons that the quasar ever emitted, $N_{\rm ion}$,
which is proportional to the total accreted black hole mass, $\Delta M_{\rm BH}$,
via the radiative efficiency $\epsilon$.
From measurements of $N_{\rm ion}$ and $M_{\rm BH}$, we can thus constrain the
average radiative efficiency during the entire growth history of the central SMBH.
Below we explain this argument in more detail.
The total number of ionizing photons emitted by a quasar can be written as
$N_{\rm ion} = \int \dot{N}_{\rm ion}(t) dt$,
where $\dot{N}_{\rm ion}(t)$ is the quasar's ionizing photon emission rate. Assuming unobscured emission along our
line of sight, a constant bolometric correction $L_{\rm bol} = C_{\rm ion} \dot{N}_{\rm ion}$, and a constant radiative efficiency $\epsilon$, we can write
\begin{equation} \label{eqn:nion2}
N_{\rm ion} = \int \frac{L_{\rm bol}(t)}{C_{\rm ion}} dt = \frac{\epsilon}{1-\epsilon} \frac{c^2}{C_{\rm ion}} \int \dot{M}_{\rm BH}(t) dt \propto \Delta M_{\rm BH}.
\end{equation}
That is, given a bolometric correction and radiative efficiency, we can translate the number of ionizing photons into the mass growth of the black hole, $\Delta M_{\rm BH} \propto N_{\rm ion}$.
We assume the luminosity-dependent bolometric correction from $M_{1450}$ given in Table 3 of \citet{Runnoe12}\footnote{We additionally include the factor of 0.75 advocated by \citet{Runnoe12} to correct for viewing angle bias.}, and convert from $M_{1450}$ to ionizing luminosity following the \citet{Lusso15} spectral energy distribution (SED), resulting in
$C_{\rm ion}=8.63\times10^{-11}$ erg per ionizing photon. In the following, we neglect the uncertainty in this conversion because the uncertainty in the quasar SED is substantially smaller than the systematic uncertainty in the black hole mass.
For J1120+0641 and J1342+0928, we compute ionizing photon emission rates of $\dot{N}_{\rm ion}^{\rm J1120}=1.2\times10^{57}$\,s$^{-1}$ and $\dot{N}_{\rm ion}^{\rm J1342}=1.4\times10^{57}$\,s$^{-1}$, and bolometric luminosities of $L_{\rm bol}^{\rm J1120}=2.7\times10^{13}$\,$L_\odot$ and $L_{\rm bol}^{\rm J1342}=3.1\times10^{13}$\,$L_\odot$, respectively.
Similar to analyses of the original Soltan argument, the census of ionizing photons recorded by the surrounding IGM
must be modified to account for obscured phases when ionizing photons are absorbed before escaping the quasar host.
That is, any ionizing photons emitted by the quasar which did not reach the IGM along our particular line of sight will be absent from
our accounting of $N_{\rm ion}$ from the spectrum.
Accordingly, we predict that black hole mass, radiative efficiency, and the number of ionizing photons should obey
\begin{equation} \label{eqn:lrsa}
\Delta M_{\rm BH} = 10^9\,M_\odot \times \left(1-f_{\rm obsc}\right)^{-1} \left(\frac{(1-\epsilon)/\epsilon}{9}\right) \left(\frac{N_{\rm ion}}{2.3\times10^{72}}\right),
\end{equation}
where $f_{\rm obsc}$ is the fraction of emitted ionizing photons that never reached the IGM along our line of sight.
This ``local reionization Soltan argument''
thus enables one to constrain
the radiative efficiency of an \emph{individual} quasar via
its spectrum close to rest-frame
Ly$\alpha$.
\section{Constraints on the Radiative Efficiency of $z>7$ Quasars} \label{sec:constrain}
\begin{figure}[htb]
\begin{center}
\resizebox{8.5cm}{!}{\includegraphics[trim={1em 1em 1em 1em},clip]{both_epsilon.pdf}}
\end{center}
\caption{Marginalized posterior PDF of the radiative efficiency $\epsilon$. The orange and purple curves show the posterior PDFs of $\epsilon$ for J1120+0641 and J1342+0928, respectively, marginalized over obscuration and the systematic uncertainty in their black hole masses. The black curve shows the combined constraint under the assumption that both quasars have the same true radiative efficiency. Arrows indicate 95\% credibility upper limits on $\epsilon$. The blue shaded region shows the range of radiative efficiencies predicted for thin accretion disks in general relativity \citep{Thorne74}. }
\label{fig:epspost}
\end{figure}
\begin{figure*}[htb]
\begin{center}
\resizebox{15cm}{!}{\includegraphics[trim={1em 1em 1em 1em},clip]{both_fobsc.pdf}}
\end{center}
\caption{Constraints on the obscuration of $z>7$ quasars assuming $\epsilon=0.1$. Left: Posterior PDF for $f_{\rm obsc}$, combining the constraints from J1120+0641 and J1342+0928. Right: Combined posterior PDF for the relative number of obscured versus unobscured quasars, $f_{\rm obsc}/(1-f_{\rm obsc})$.}
\label{fig:fobsc}
\end{figure*}
We measured $N_{\rm ion}$
for the two highest
redshift quasars known, J1120+0641 and J1342+0928, by analyzing the
Ly$\alpha$ absorption in their rest-frame UV spectra in a very similar fashion to \citet{Davies18b}.
The intrinsic, unabsorbed quasar spectrum close to rest-frame Ly$\alpha$
was estimated via a predictive principal
component analysis (PCA) approach from \citet{Davies18a}.
In Figure~\ref{fig:spectra} we show the two quasar spectra close to
rest-frame Ly$\alpha$ (grey and black curves) compared to their respective
PCA continuum models (blue curves). Both quasars show compelling evidence for an IGM damping wing
and truncated proximity zones, as previously shown by \citet{Davies18b}.
We model reionization-epoch quasar spectra via a multi-scale approach following \citet{Davies18b} (see also Appendix~\ref{sec:method}). The large-scale topology of reionization around massive dark matter halos was computed in a (400 Mpc)$^3$ volume using a modified version of the \texttt{21cmFAST} code (\citealt{Mesinger11}; Davies \& Furlanetto, in prep.), and we stitched lines of sight through this ionization field onto skewers of baryon density fluctuations from a separate (100~Mpc$/h)^3$ \texttt{Nyx} hydrodynamical simulation \citep{Lukic15}. Finally, we performed 1D ionizing radiative transfer to model the ionization and heating of the IGM by the quasar \citep{Davies16,Davies19a}.
Through a Bayesian analysis on a grid of forward-modeled mock Ly$\alpha$ spectra from our simulations (Appendix~\ref{sec:method}), we jointly constrained the total number of
ionizing photons emitted by the quasars ($N_{\rm ion}$) and the
volume-averaged neutral fraction
of the IGM ($\langle x_{\rm HI} \rangle$). The mean Ly$\alpha$ absorption profiles of our best-fit
models and their 68\% scatter in the mock spectra are shown as the orange curves and shaded regions in Figure~\ref{fig:spectra}.
The red curves in Figure~\ref{fig:spectra}
show alternative models where the IGM is fully neutral and $N_{\rm ion}$ for each quasar is instead determined via equation~(\ref{eqn:lrsa}),
assuming $\epsilon=0.1$, $f_{\rm obsc}=0$, and $\Delta M_{\rm BH}=M_{\rm BH}$.
These curves thus correspond to the maximum Ly$\alpha$ absorption in the standard
view of UV-luminous radiatively efficient SMBH growth.
The canonical radiative efficiency appears to be highly inconsistent with the data.
More quantitatively, in the bottom panels of Figure~\ref{fig:2dpost} we show the joint
posterior probability distribution functions (PDFs) for $N_{\rm ion}$
and $\langle x_{\rm HI} \rangle$ from our analysis of J1120+0641 (left) and
J1342+0928 (right). In the top panels of Figure~\ref{fig:2dpost} we show the marginalized posterior PDFs for
$N_{\rm ion}$.
Through the lens of equation~(\ref{eqn:lrsa}), we can view these marginalized posterior PDFs
as constraints on the total
accreted black hole mass, indicated by the upper axes, where we assume $\epsilon=0.1$.
The vertical lines show the
measured black holes masses for J1120+0641 and
J1342+0928, with shaded regions indicating their
systematic uncertainty.
For both quasars the inferred accreted mass is in
strong disagreement with the measured black hole mass,
or equivalently, a radiative efficiency much lower than 10\% is required to match the observations.
At face value, the results above indicate a serious inconsistency between
standard thinking about the radiative efficiency -- informed by general relativity,
accretion disk models, and the Soltan argument -- and our measurements
for these two reionization-epoch quasars. How can we reconcile the smaller than expected
number of ionizing photons emitted towards Earth with the observed
black hole masses? One possibility is that the bulk of the black hole
growth resulted in fewer ionizing photons
escaping into the IGM toward our line-of-sight due to obscuration by
gas and dust in the quasar host galaxy (e.g. \citealt{Hopkins05}).
If the black holes grew appreciably during obscured phases, then
this is clearly degenerate with $N_{\rm ion}$
as indicated in equation~(\ref{eqn:lrsa}).
Observations of similarly luminous quasars at
lower redshifts $z\gtrsim 2$ find that $\sim 50\%$ of them are
obscured \citep{Polletta08,Merloni14}, with some indication for increased obscuration at higher redshifts (\citealt{Vito18}, see also \citealt{Trebitsch19}).
To quantitatively constrain the radiative efficiency,
accounting for both the degeneracy with obscuration and
uncertainties in the
black hole masses, we re-map our 1D constraint on $N_{\rm
ion}$ (i.e. the upper panels of Figure~\ref{fig:2dpost}) to a 3D space
of radiative efficiency, black hole mass, and the obscured fraction (see Appendix~\ref{sec:pdfmap}).
Marginalizing the 3D distributions over obscuration (assuming a uniform linear prior from 0 to 100\%)
and black hole mass uncertainty (lognormal prior with $\sigma=0.4$~dex, \citealt{Shen13}) yields posterior PDFs for the radiative
efficiency as shown in Figure~\ref{fig:epspost}. The posterior median radiative efficiencies of J1120+0641 and J1342+0928
are 0.08\% and 0.1\%, respectively, and the canonical 10\%
is ruled out at greater than 98 per cent probability by each quasar.
The combined posterior PDF for both quasars, assuming both quasars have the same true radiative efficiency, is shown by the black curve in
Figure~\ref{fig:epspost}, which is inconsistent with $\epsilon=0.1$ at 99.8 per cent
probability.
We can also assess what an assumed radiative efficiency of $10\%$ would imply for the obscuration of $z>7$ quasars. The left panel of Figure~\ref{fig:fobsc} shows the combined posterior PDF of $f_{\rm obsc}$ from both quasars assuming $\epsilon=0.1$, implying $f_{\rm obsc}>82\%$ at 95\% credibility. Such a high obscured fraction implies that there are many more similarly-luminous obscured quasars at $z>7$ which have not yet been identified. The right panel of Figure~\ref{fig:fobsc} shows the posterior PDF for the ratio of obscured to unobscured quasars, $f_{\rm obsc}/(1-f_{\rm obsc})$, which we computed from the $f_{\rm obsc}$ posterior PDF by a probability transformation. We constrain this ratio to be $25.7^{+49.6}_{-16.5}$ (posterior median and 68\% credible interval), with a 95\% credible lower limit of $4.4$.
\section{Discussion \& Conclusion} \label{sec:discus}
\begin{figure*}[htb]
\begin{center}
\resizebox{18cm}{!}{\includegraphics[trim={7em 3em 4em 0},clip]{solutions.pdf}}
\end{center}
\caption{Possible solutions to the measured ionizing photon deficiency. The first two columns show schematic representations of the immediate environment of the $z>7$ quasars suggested by the measured deficit of ionizing photons. The blue regions are illuminated by the quasar while the black regions are not. The line of sight towards Earth is in the negative vertical direction as indicated by the orange dashed line. The leftmost column shows the quasar environment as it was 10 Myr ago, while the next column shows the quasar as it is observed today. The third column shows the UV (blue) and bolometric (red) luminosity history of the quasar, where we assume that the bolometric luminosity is proportional to the Eddington luminosity and thus its evolution indicates the growth of the black hole. The rightmost column shows the integrated number of ionizing photons that escaped into the IGM along our line of sight. The top and middle rows show two possible obscuration scenarios -- time-variable obscuration with a large covering fraction (top), or full obscuration followed by blowout (middle). The bottom row shows the case where the radiative efficiency is low, leading to rapid black hole growth and a much less luminous quasar 10 Myr ago.}
\label{fig:solutions}
\end{figure*}
If this radiatively inefficient mode of growth that we have uncovered
applies to quasars at later cosmic epochs, the Soltan argument implies previous
analyses have underestimated the present day SMBH mass density
by at least an order of magnitude.
Without invoking extra SMBH mass locally, the only solution is that
$z\gtrsim 7$ quasars grow or emit their radiation qualitatively differently
from their lower redshift counterparts. A few possibilities are
illustrated in Figure~\ref{fig:solutions}. It could be that $z>7$ quasar accretion disks are
actually radiatively efficient with $\epsilon \simeq 0.1$ but simply have
much more obscuration than inferred from studies of quasar demographics
at lower redshift (see also \citealt{Comastri15}). As our analysis only constrains the integrated
number of ionizing photons emitted in our direction, it is agnostic to
the exact nature of the obscuration. It could have been highly time-variable
with obscured phases lasting $\gtrsim 10$ times longer than UV
luminous ones (top row of Figure~\ref{fig:solutions}), or the black hole could have grown while
fully enshrouded until a ``blowout'' event $\sim 1$ Myr ago when it
transitioned to a UV luminous phase (middle row of
Figure~\ref{fig:solutions}) \citep{Hopkins05}. Either of these obscuration scenarios
predicts many of comparably-luminous obscured quasars for every
UV luminous one at $z > 7$, as discussed above (Figure~\ref{fig:fobsc}). The obscured fraction would then have to
evolve very rapidly to avoid overproducing luminous obscured quasars at later
times. Nevertheless, if such a population exists at $z> 7$, future
mid-IR observations with JWST have the potential to uncover them.
Finally, let us not exclude the possibility that $z>7$ quasar accretion
disks are truly radiatively inefficient (bottom row of Figure~\ref{fig:solutions}).
This would
allow for rapid super-critical mass accretion rates with e-folding
timescales much shorter than 45 Myr without violating the
Eddington limit (equation~\ref{eqn:salp}), and has the appeal that it would easily explain
the existence of $\gtrsim 10^9$ SMBHs at early cosmic times $z > 7$ without
requiring overly massive seeds.
This last scenario poses an intriguing
question: if the radiative efficiencies of the highest redshift quasars
are radically different from those at lower redshift, why do their
spectra appear nearly identical over eight decades in frequency \citep{Banados15b,Shen18,Nanni17}?
Similar to the original Soltan argument, the radiative efficiency that we have derived is a luminosity-weighted
average over the growth of the SMBH, which may differ from the efficiency of the currently
observed accretion flow.
Past phases of extremely super-critical accretion
cannot be ruled out, provided that they only occur at $z > 7$ -- in the same vein, however,
neither can exotic formation scenarios, e.g. direct collapse to $10^9 M_\odot$, as long as they do not liberate UV photons.
Future analyses of additional reionization-epoch quasars, combined with analogous measurements of the impact of luminous quasars on the IGM at lower redshifts \citep{Eilers18J1335,Schmidt18echo,Khrykin19,Davies19b}, will thus greatly improve our understanding of how SMBHs grew.
\section*{Acknowledgements}
We thank Matthew McQuinn and Steven Furlanetto for comments on an early draft of this manuscript. FBD acknowledges support from the Space Telescope Science Institute, which is operated by AURA for NASA, through the grant HST-AR-15014.
|
\section{Introduction}
Francis's implicitly-shifted QR algorithm \cite{Fra61b,Wat11} is still the
standard tool for computing the eigenvalues of a small to medium-sized non-Hermitian matrix
$A\in\mbox{$\cplxs^{n \times n}$}$. Eigenvalue problems often arise naturally as generalized eigenvalue problems for
a pencil $A - \lambda B$, and for these the Moler-Stewart variant \cite{MolSte73}
of Francis's algorithm, commonly called the QZ algorithm, is used. The Francis and Moler-Stewart
algorithms are prime examples of what we call \emph{bulge-chasing} algorithms.
In recent years some generalizations of the QZ algorithm have been proposed, e.g.\
\cite{VanWat12g} and, more generally, the rational QZ (RQZ) algorithm of Camps, Meerbergen, and Vandebril \cite{CaMeVa19a}, which
is the first example of what we call a \emph{pole-swapping} algorithm.
This work has been extended in various directions in \cite{Cam19,CaMaVaWa19,CaMeVa19b}.
In this paper we extend in another direction,
introducing structure-preserving pole-swapping algorithms for two classes of
structured generalized eigenvalue problems that arise in optimal control theory \cite{Meh91},
namely palindromic and alternating eigenvalue problems.
Kressner, Schr\"{o}der, and Watkins \cite{KrScWa09} proposed
structure-preserving bulge-chasing (QZ-like) algorithms for palindromic and alternating eigenvalue problems.
We show that our structured pole-swapping algorithms are generalizations of these bulge-chasing algorithms.
Our algorithms include a refinement step, which can also be incorporated into the algorithm of \cite{KrScWa09},
to ensure backward stability.
We have opted to cover only the complex case. If the matrices $A$ and $B$ happen to be real, the algorithms
introduced here will not preserve the real structure. It is possible, and not difficult, to build pole-swapping algorithms
that preserve real structure, as was shown in \cite{Cam19,CaMeVa19b}. We could have incorporporated
that extension in this paper at the expense of making it longer and less readable.
\section{Basic definitions and facts}
In this paper we will refer sometimes to a pencil $A - \lambda B$ and other times to a pair $(A,B)$.
Either way, we are talking about the same object. The pencil $A - \lambda B$ is called \emph{singular} if
the polynomial $\det(A - \lambda B)$ is identically zero. Otherwise the pencil is \emph{regular}. Throughout the paper
we make the blanket assumption that we are dealing with a regular pencil.
The pair $(A,B)$ is called \emph{palindromic} (or \emph{$*$-palindromic}) if $A$ and $B$ are related by
\begin{equation}\label{eq:palindromic}
A^{*} = B.
\end{equation}
The pair $(A,B)$ is \emph{alternating} (or \emph{even}) if
\begin{equation}\label{eq:alternating}
A^{*} = A \quad\mbox{and}\quad B^{*} = - B.
\end{equation}
These two structures are equivalent in principle, since the generalized Cayley transform
$(A,B) \to (A+B,A-B)$ maps a palindromic pair to an alternating pair and vice versa.
Each of these structures exhibits a spectral symmetry. In the palindromic case,
$\lambda$ is an eigenvalue if and only if $1/\overline{\lambda}$ is. This is shown by writing down the equation
$Ax = \lambda Bx$, taking the conjugate transpose, and applying (\ref{eq:palindromic}). Clearly $\lambda =
1/\overline{\lambda}$ if and only if $\lambda$ is on the unit circle. Eigenvalues on the unit circle need not occur in
pairs, but those off the unit circle must appear in $(\lambda,1/\overline{\lambda})$ pairs, one inside and one outside the
unit circle.
In the alternating case, $\lambda$ is an eigenvalue if and only if $-\overline{\lambda}$ is, as can be shown in the same
way as for the palindromic case. Since $\lambda = -\overline{\lambda}$ if and only if $\lambda$ is on the imaginary
axis, eigenvalues on the imaginary axis need not occur in pairs, but those off of the imaginary axis must occur in
$(\lambda,-\overline{\lambda})$ pairs, one on each side of the imaginary axis.
The spectral symmetry in alternating pencils is exactly the same as that possessed by Hamiltonian matrices.
Recall that a matrix $H\in\mbox{$\mathbb{C}$}^{2m\times 2m}$ is called \emph{Hamiltonian} if $JH$ is Hermitian, where
\begin{displaymath}
J = \left[\begin{array}{cc} & I_{m} \\ -I_{m} & \end{array}\right].
\end{displaymath}
It is well known \cite{Meh91} that the continuous-time
linear-quadratic control problem can be solved by computing the eigensystem
of a Hamiltonian matrix. One can equally well formulate the problem as an eigenvalue problem for an alternating
pencil. In fact, the Hamiltonian eigenvalue problem $H - \lambda I$ is clearly equivalent to the alternating
eigenvalue problem $JH - \lambda J$. Hamiltonian eigenvalue problems are also sometimes studied in the
guise of Hamiltonian/skew-Hamiltonian pencils, but note that the Hamiltonian/skew-Hamiltonian pencil
$H - \lambda S$ is equivalent to the alternating pencil $JH - \lambda JS$. Alternating pencils are discussed
in various guises and contexts in
\cite{MehWat01,MehWat02,ApMeWa02,MaMaMeMe06a,MaMaMeMe06b}, for example.
The spectral symmetry in palindromic pencils is the same as that possessed by symplectic matrices.
The discrete-time linear-quadratic optimal control problem can be solved by computing the eigensystem
of a symplectic matrix or pencil \cite{Meh91},
which can also be formulated as a palindromic eigenvalue problem
\cite{MaMaMeMe06b,MaMaMeMe09,KrScWa09}.
In \cite{KrScWa09} we considered pairs $(A,B)$ for which both $A$ and $B$ have anti-Hessenberg form:
\begin{displaymath}
\parbox{3.2cm}{
\begin{tikzpicture}[scale=1.66,y=-1cm]
\draw (-.15,-.1) -- (-.2,-.1) -- (-.2,1.5) -- (-.15,1.5);
\draw (1.55,-.1) -- (1.6,-.1) -- (1.6,1.5) -- (1.55,1.5);
\foreach \j in {0,...,7}{
\foreach \i in {\j,...,7}{\node at (\i/5,1.4-\j/5)
[align=center,scale=1.0]{$\times$};}}
\foreach \j in {0,...,6}{\node at (\j/5,1.4-\j/5-.2)
[align=center,scale=1.0]{$\times$};}
\end{tikzpicture}
},
\end{displaymath}
shown here in the case $n=8$. We will consider the same form here. We could study instead
the equivalent pair $(FA,FB)$, where $F$ is the \emph{flip} or
\emph{anti-identity} matrix. $FA$ and $FB$ are both
upper Hessenberg, but the anti-Hessenberg form is more convenient for the special structures that
we are considering here.
This form is admittedly quite special. Given a general pair $(A,B)$ that is either palindromic or alternating,
there is no known efficient method for transforming the pair to anti-Hessenberg form while preserving the
relevant structure, except in special cases.
For example, if the pencil corresponds to a control system that has either a single
input or a single output, such a reduction is possible \cite{KrScWa09}.
Following Camps, Meerbergen, and Vandebril \cite{CaMeVa19a} we associate $n-1$ \emph{poles} with the
anti-Hessenberg pair $(A,B)$. For $k=1$, \ldots, $n-1$ the \emph{pole in position $(n-k,k)$} is the
ratio $\sigma_{k} = a_{n-k,k}/b_{n-k,k}$. We assume that for each $k$, $\absval{a_{n-k,k}} + \absval{b_{n-k,k}} > 0$,
since otherwise it would be possible to reduce the problem immediately to two or more smaller eigenvalue problems.
Thus every $\sigma_{k}$ is well defined (but might equal~$\infty$).
In the two structured cases that we are considering, the poles exhibit the same symmetry as the eigenvalues do.
In the palindromic case we have
$a_{n-k,k} = \overline{b}_{k,n-k}$, and it follows that
\begin{displaymath}
\sigma_{k} = 1/\overline{\sigma}_{n-k},\quad k=1,\, \ldots,\, n-1.
\end{displaymath}
In the case of even $n$, there is one unpaired pole $\sigma_{n/2}$, which must satisfy
$\absval{\sigma_{n/2}} = 1$.
In the alternating case we have
\begin{displaymath}
\sigma_{k} = -\overline{\sigma}_{n-k}, \quad k=1,\, \ldots,\, n-1.
\end{displaymath}
If $n$ is even, there is
one unpaired pole $\sigma_{n/2}$, which must lie on the imaginary axis.
\section{Operations on anti-Hessenberg pairs}\label{sec:operations}
Introducing terminology that we have used in some of our recent work
\cite{AuMaRoVaWa18,AuMaRoVaWa18g,AuMaRoVaWa19,AuMaVaWa15}, we define a \emph{core transformation}
(or \emph{core} for short) to be a unitary matrix that acts only on two adjacent rows/columns, for example,
\begin{displaymath}
Q_{3} = \left[\begin{array}{ccccc}
1 & & & & \\ & 1 & & & \\ & & {*} & {*} & \\ & & {*} & {*} & \\ & & & & 1
\end{array}\right],
\end{displaymath}
where the four asterisks form a $2 \times 2$ unitary matrix. Givens rotations are examples of core transformations.
Our core transformations always have subscripts that tell where the action is: $Q_{j}$ acts on rows/columns
$j$ and $j+1$.
In \cite{CaMeVa19a} two types of operations on upper Hessenberg pencils were introduced. These are
unitary equivalence transformations, which we called \emph{moves} of types I and II in \cite{CaMaVaWa19}.
Obviously we can do the same sorts of moves on anti-Hessenberg pencils, but if we want to preserve the
palindromic or alternating structure, we need to use special unitary equivalence transformations, namely
congruences $A - \lambda B \to Q^{*}(A - \lambda B)Q$.
\subsection*{Move of Type I}
This move replaces the pole $\sigma_{1}$ (located at position $(n-1,1)$) by any other value $\rho$.
At the same time, since the symmetry must be preserved, the pole $\sigma_{n-1}$ is changed appropriately.
In the palindromic case, $\sigma_{n-1} = 1/\overline{\sigma}_{1}$, and it gets changed to $1/\overline{\rho}$.
In the alternating case, $\sigma_{n-1} = -\overline{\sigma}_{1}$, and it gets changed to $-\overline{\rho}$.
To see how this is done, we at first ignore the need to preserve structure and consider how we would
insert a pole $\rho$ at position $(n-1,1)$. Because of the anti-Hessenberg form, the vector $(A - \rho B)e_{1}$
consists of zeros, except for the last two entries.\footnote{Here and in what follows, the notation $A - \rho B$ is shorthand
for $\beta A - \alpha B$, where $\alpha$ and $\beta$ are any scalars satisfying $\rho = \alpha/\beta$. This allows us to
include the case $\rho = \infty$ by taking $\beta = 0$.} Therefore a core transformation $Q_{n-1}$, acting on rows
$n-1$ and $n$, can be constructed so that $Q_{n-1}^{*}$ zeros out the entry in position $n-1$, that is,
\begin{displaymath}
Q_{n-1}^{*}(A - \rho B)e_{1} = \gamma\,e_{n}
\end{displaymath}
for some nonzero $\gamma$. Now let $\tilde{A} - \lambda \tilde{B}
= Q_{n-1}^{*}(A - \lambda B)$. This new pencil has the pole $\rho$ in position $(n-1,1)$, as desired, since
$\tilde{a}_{n-1,1} - \rho\tilde{b}_{n-1,1} = 0$. This is exactly the move of type I described in \cite{CaMaVaWa19} and
earlier in \cite{CaMeVa19a}, turned over for the anti-Hessenberg case.
But this transformation does not preserve the symmetry of the pencil. What is needed is a congruence:
$\hat{A} - \lambda \hat{B} = Q_{n-1}^{*}(A - \lambda B)Q_{n-1}$. The right multiplication by $Q_{n-1}$ does not
affect what happens in the lower left-hand corner of the pencil, so the pole in position $(n-1,1)$ is $\rho$, as shown
above. But the right multiplication by $Q_{n-1}$ does affect the pole $\sigma_{n-1}$ at position $(1,n-1)$, changing it
(by symmetry) to $1/\overline{\rho}$ in the palindromic case and $-\overline{\rho}$ in the alternating case.
\subsection*{Move of Type II}
This move swaps two adjacent poles. If we delete the $n$th row and column from the anti-Hessenberg pencil
$A - \lambda B$, we get an anti-triangular pencil $A_{\pi} -\lambda B_{\pi}$, which we call the \emph{pole pencil}
because its eigenvalues are exactly the poles of $A - \lambda B$. Swapping two adjacent poles in $A - \lambda B$
is the same as swapping two eigenvalues in $A_{\pi} - \lambda B_{\pi}$, and there are well-known procedures for
doing this \cite{BaiDem93}, \cite{CaMaVaWa19}, \cite{KagPor96a,KagPor96b},
\cite{VanD81}, \cite[\S\S~4.8,\,6.6]{Wat07}.
Suppose
we want to swap two adjacent poles $\sigma_{k-1}$ and $\sigma_{k}$
located at positions $(n-k+1,k-1)$ and $(n-k,k)$. Temporarily ignoring symmetry, this can
be accomplished by a transformation $Q_{n-k}^{*}(A - \lambda B)Z_{k-1}$. All of the action is
in the subpencil
\begin{equation}\label{eq:2pencil}
\left[\begin{array}{cc} 0 & a_{n-k,k} \\ a_{n-k+1,k-1} & a_{n-k+1,k}\end{array}\right]
- \lambda \left[\begin{array}{cc} 0 & b_{n-k,k} \\ b_{n-k+1,k-1} & b_{n-k+1,k}\end{array}\right].
\end{equation}
This is the principal subpencil of the bulge pencil that contains the two poles $\sigma_{k-1} =
a_{n-k+1,k-1}/b_{n-k+1,k-1}$ and $\sigma_{k} = a_{n-k,k}/b_{n-k,k}$.
In order to preserve the symmetry we will also have
to swap the poles $\sigma_{n-k}$ and $\sigma_{n-k+1}$
in positions $(k-1,n-k+1)$ and $(k,n-k)$. The appropriate transformation is, by symmetry,
$Z_{k-1}^{*}(A - \lambda B)Q_{n-k}$. This can be done provided the two transformations do not
interfere with each other, i.e. $Q_{n-k}$ and $Z_{k-1}$ act on non-overlapping columns. This is the case
if $k < n-k$ or $n-k+1 < k-1$, that is, $k < n/2$ or $k > n/2 + 1$.
The total transformation is $Q^{*}(A - \lambda B)Q$, where $Q = Z_{k-1}Q_{n-k} = Q_{n-k}Z_{k-1}$.
\subsection*{A hint at where we are heading}
In \cite{CaMeVa19a,CaMaVaWa19} it was shown how moves of types I and II can be used to build algorithms
for computing eigenvalues of an upper Hessenberg pencil. In the simplest case a
shift $\rho$ is chosen, and a move of
type I is used to insert it as a pole at the top of the pencil.
Then a sequence of moves of type II is used to exchange the pole $\rho$ downward until it gets to the bottom
of the pencil. Then it is removed from the bottom
(replaced by some new pole $\sigma_{n}$) by a move of type I. This procedure
can be shown \cite{CaMeVa19a} to be a generalization of (one iteration of) the QZ algorithm on a Hessenberg-triangular
pencil.
In our current scenario the matrices are anti-Hessenberg, not Hessenberg, but that is a trivial difference. More
importantly, in the moves described here, everything is doubled up for preservation of
structure. If we introduce a pole $\rho$ at one end of the pencil, we must simultaneously introduce a pole
$\tilde{\rho}$ ($=1/\overline{\rho}$ or $-\overline{\rho}$, depending on the structure) at the other end. Now if we want to move $\rho$ from one end of the pencil to the other by moves of type II, we must simultaneously move $\tilde{\rho}$ in the opposite
direction. There comes a point in the middle where we have to swap $\rho$ and $\tilde{\rho}$, so that they can continue
their journey to the opposite end of the pencil. This requires
a special symmetric version of the move of type II.
\subsection*{Move of Type IIo}
Suppose the dimension $n$ is odd. Then there is an even number of poles $\sigma_{1}$, \ldots, $\sigma_{n-1}$.
The two poles in the middle are $\sigma_{k-1}$ and $\sigma_{k}$, where $k = (n+1)/2$.
These are the two eigenvalues of (\ref{eq:2pencil}) with $k=(n+1)/2$.
In the interest of simplicity and non-proliferation of notation, we rename this subpencil
\begin{displaymath}
A - \lambda B =
\left[\begin{array}{cc} & \alpha_{1} \\ \alpha_{2} & \alpha_{21} \end{array}\right]
- \lambda
\left[\begin{array}{cc} & \beta_{1} \\ \beta_{2} & \beta_{21} \end{array}\right].
\end{displaymath}
We are temporarily re-assigning the symbols $A$ and $B$ to stand for the submatrices that are our current focus.
This little pole pencil has the same structure as the original pencil, either palindromic or alternating, but we will
ignore the structure at first. The eigenvalues are $\alpha_{2}/\beta_{2}$ and $\alpha_{1}/\beta_{1}$, and we would like to
swap them. If we flip the rows and columns we get
\begin{displaymath}
FAF - \lambda FBF =
\left[\begin{array}{cc} \alpha_{21} & \alpha_{2} \\ \alpha_{1} & \end{array}\right]
- \lambda
\left[\begin{array}{cc} \beta_{21} & \beta_{2} \\ \beta_{1} & \end{array}\right],
\end{displaymath}
which has the poles in the desired positions but the ``wrong'' triangularity. We have to fix this. We
find it convenient to work with the partially flipped form
\begin{displaymath}
AF - \lambda BF =
\left[\begin{array}{cc} \alpha_{1} & \\ \alpha_{21} & \alpha_{2} \end{array}\right]
- \lambda
\left[\begin{array}{cc} \beta_{1} & \\ \beta_{21} & \beta_{2} \end{array}\right],
\end{displaymath}
which is easier to study because the matrices are triangular, not anti-triangular.
Next we set up and solve a Sylvester equation to diagonalize the pencil. Specifically, we will find
unit lower triangular matrices
\begin{displaymath}
X = \left[\begin{array}{cc} 1 & \\ x & 1 \end{array}\right]
\quad\mbox{and}\quad
Y = \left[\begin{array}{cc} 1 & \\ y & 1 \end{array}\right]
\end{displaymath}
such that
\begin{equation}\label{eq:sylv}
(AF)X = Y(\check{A}F) \quad\mbox{and}\quad (BF)X = Y(\check{B}F),
\end{equation}
where $\check{A}$ and $\check{B}$ are anti-diagonal matrices with the same anti-diagonals as $A$ and $B$,
respectively. Writing the first of these equations out in detail, we have
\begin{displaymath}
\left[\begin{array}{ccc} \alpha_{1} & \\
\alpha_{21} & \alpha_{2} \end{array}\right]
\left[\begin{array}{ccc} 1 & \\ x & 1 \end{array}\right] =
\left[\begin{array}{ccc} 1 & \\ y & 1 \end{array}\right]
\left[\begin{array}{ccc} \alpha_{1} & \\
& \alpha_{2} \end{array}\right],
\end{displaymath}
and similarly for the $B$ equation. This is a system of two linear equations in two unknowns
\begin{displaymath}
\alpha_{21} + \alpha_{2}x = y \alpha_{1}, \quad \beta_{21} + \beta_{2}x = y \beta_{1}
\end{displaymath}
or
\begin{equation}\label{eq:sylv0}
\left[\begin{array}{cc}
\alpha_{1} & \alpha_{2} \\ \beta_{1} & \beta_{2}
\end{array}\right]
\left[\begin{array}{c}
\phantom{-}y \\ -x
\end{array}\right] =
\left[\begin{array}{c} \alpha_{21} \\ \beta_{21} \end{array}\right].
\end{equation}
This has a unique solution if and only if $\alpha_{1}\beta_{2} - \beta_{1}\alpha_{2} \neq 0$, i.e.\ the
poles are distinct.
Thus, assuming distinct poles, (\ref{eq:sylv}) has a unique solution, which we can easily and stably compute.
Rewriting (\ref{eq:sylv}) we have the equivalence
\begin{equation}\label{eq:nonuneq2}
FY^{-1}(A - \lambda B)(FX) = F(\check{A} - \lambda\check{B})F =
\left[\begin{array}{cc} & \alpha_{2} \\ \alpha_{1} & \end{array}\right]
- \lambda
\left[\begin{array}{ccc} & \beta_{2} \\ \beta_{1} & \end{array}\right],
\end{equation}
which (still) has the poles in the desired locations, but this equivalence is not unitary. To make
a unitary equivalence we will introduce some QR decompositions. Since we are going to work with
triangular matrices, we again remove the $F$ from the left-hand side to obtain
\begin{displaymath}
Y^{-1}(A - \lambda B)(FX) = (\check{A} - \lambda\check{B})F =
\left[\begin{array}{cc} \alpha_{1} & \\ & \alpha_{2} \end{array}\right]
- \lambda
\left[\begin{array}{ccc} \beta_{1} & \\ & \beta_{2} \end{array}\right].
\end{displaymath}
Let
\begin{equation}\label{eq:2qrd}
FX = QR \quad\mbox{and}\quad Y = PS,
\end{equation}
where $Q$ and $P$ are unitary, and $R$ and $S$
are upper triangular with positive entries on the main diagonal. Then
\begin{displaymath}
P^{*}(A - \lambda B)Q = S(\check{A}F -\lambda \check{B}F)R^{-1} = T_{A} -\lambda T_{B},
\end{displaymath}
where $T_{A}$ and $T_{B}$ are upper triangular:
\begin{displaymath}
T_{A} = S(\check{A}F) R^{-1} = \left[\begin{array}{cc}
s_{11}\alpha_{1} r_{11}^{-1} & {*} \\ & s_{22}\alpha_{2}r_{22}^{-1}
\end{array}\right],
\end{displaymath}
\begin{displaymath}
T_{B} = S(\check{B}F) R^{-1} = \left[\begin{array}{cc}
s_{11}\beta_{1} r_{11}^{-1} & {*} \\ & s_{22}\beta_{2}r_{22}^{-1}
\end{array}\right].
\end{displaymath}
Since we want anti-triangular matrices, we now restore the $F$ on the left to obtain
\begin{equation}\label{eq:finaleq}
FP^{*}(A - \lambda B) Q =
\left[\begin{array}{cc}
& s_{22}\alpha_{2} r_{22}^{-1} \\
s_{11}\alpha_{1}r_{11}^{-1} & {*}
\end{array}\right] - \lambda \left[\begin{array}{cc}
& s_{22}\beta_{2} r_{22}^{-1} \\
s_{11}\beta_{1}r_{11}^{-1} & {*}
\end{array}\right].
\end{equation}
This unitary equivalence gives the right anti-triangular form with the poles
$\alpha_{1}/\beta_{1}$ and $\alpha_{2}/\beta_{2}$ in the desired positions.
So far we have assumed no special relationship between $A$ and $B$. To finish the story we must
show that in our two structured cases (\ref{eq:finaleq}) is a congruence, that is $PF = Q$.
In preparation
for this we go back to (\ref{eq:sylv}) and take complex conjugates of both equations to obtain
\begin{displaymath}
X^{*}FA^{*} = F\check{A}^{*}Y^{*}, \quad X^{*}FB^{*} = F\check{B}^{*}Y^{*}.
\end{displaymath}
Further simple manipulations yield
\begin{equation}\label{eq:sylvstar}
(A^{*}F)(FY^{-*}F) = (FX^{-*}F)(\check{A}^{*}F), \quad (B^{*}F)(FY^{-*}F) = (FX^{-*}F)(\check{B}^{*}F).
\end{equation}
Now consider what these equations look like in the alternating case $A^{*} = A$, $B^{*} = - B$
(which also implies $\check{A}^{*} = \check{A}$ and $\check{B}^{*} = - \check{B}$). Clearly
we have
\begin{equation}\label{eq:sylvstarstruc}
(AF)(FY^{-*}F) = (FX^{-*}F)(\check{A}F), \quad (BF)(FY^{-*}F) = (FX^{-*}F)(\check{B}F).
\end{equation}
Now consider the palindromic case $B = A^{*}$. Inserting this equation into (\ref{eq:sylvstar}), we again
get (\ref{eq:sylvstarstruc}), just as in the alternating case. Thus, in either case, (\ref{eq:sylvstarstruc}) holds.
Noting that $FY^{-*}F$ and $FX^{-*}F$ are both
unit lower triangular, and comparing (\ref{eq:sylvstarstruc}) with (\ref{eq:sylv}), we see that the pair
$(FY^{-*}F,FX^{-*}F)$ is a solution of (\ref{eq:sylv}). By uniqueness of the solution we deduce that
$(X,Y) = (FY^{-*}F,FX^{-*}F)$, and in particular $X = FY^{-*}F$. Writing this
as $FX = Y^{-*}F$ and inserting the decompositions $FX = QR$ and $Y= PS$, we obtain
\begin{displaymath}
QR = PS^{-*}F = (PF)(FS^{-*}F).
\end{displaymath}
By uniqueness of the QR decomposition,
\begin{displaymath}
Q = PF.
\end{displaymath}
Making this substitution into (\ref{eq:finaleq}), we obtain
\begin{displaymath}
Q^{*}(A - \lambda B)Q = FT_{A} - \lambda FT_{B} = \hat{A} - \lambda\hat{B},
\end{displaymath}
where
\begin{displaymath}
\hat{A} = \left[\begin{array}{cccc} & \alpha_{2}s_{22}r_{22}^{-1} \\
\alpha_{1}s_{11}r_{11}^{-1} & {*}
\end{array}\right]
\quad\mbox{and}\quad
\hat{B} = \left[\begin{array}{cccc} & \beta_{2}s_{22}r_{22}^{-1} \\
\beta_{1}s_{11}r_{11}^{-1} & {*}
\end{array}\right].
\end{displaymath}
The actual computation is quite simple. From (\ref{eq:2qrd}) we see that we need to build $Q$
satisfying $FX = QR$. This means we need
\begin{equation}\label{eq:qfirstcol}
\left[\begin{array}{c} x \\ 1\end{array}\right] = Q\left[\begin{array}{c} r_{11} \\ 0 \end{array}\right],
\end{equation}
where $x$ is obtained by solving (\ref{eq:sylv0}). In other words, we need a unitary $Q$ whose
first column is proportional to $\left[\begin{array}{cc} x & 1 \end{array}\right]^{T}$.
Partially solving (\ref{eq:sylv0}) by Cramer's rule, we obtain
\begin{displaymath}
x = \frac{\beta_{1}\alpha_{12} - \alpha_{1}\beta_{12}}{\alpha_{1}\beta_{2} - \beta_{1}\alpha_{2}}.
\end{displaymath}
We don't actually need $\left[\begin{array}{cc} x & 1 \end{array}\right]^{T}$ in (\ref{eq:qfirstcol}), as
any vector proportional to $\left[\begin{array}{cc} x & 1 \end{array}\right]^{T}$ will suffice, so
we use instead
\begin{displaymath}
v = \left[\begin{array}{c} v_{1} \\ v_{2}\end{array}\right] =
\left[\begin{array}{c} \beta_{1}\alpha_{12} - \alpha_{1}\beta_{12} \\
\alpha_{1}\beta_{2} - \beta_{1}\alpha_{2}
\end{array}\right].
\end{displaymath}
Thus $Q$ is obtained by
\begin{displaymath}
r = \norm{v}_{2}, \qquad c = \frac{v_{1}}{r}, \qquad
s = \frac{v_{2}}{r},
\end{displaymath}
and
\begin{displaymath}
Q = \left[\begin{array}{cc} c & -\overline{s} \\
s & \phantom{-}\overline{c}\end{array}\right].
\end{displaymath}
We remark that in both palindromic and alternating cases, $s$ is a real number.
In terms of the original
(big) matrices $A$ and $B$, the congruence is
$$Q_{k-1}^{*}(A - \lambda B)Q_{k-1},$$
where $Q_{k-1}$ is
a core transformation built from the $2\times 2$ unitary matrix $Q$.
We emphasize that this procedure succeeds if the two poles are distinct. Of course there is nothing
to be gained by interchanging two poles that are equal. In our application below we will use a move of this
type to interchange two shifts $\rho$ and $\tilde{\rho}$. In the palindromic (resp.\ alternating) case,
$\tilde{\rho} = 1/\overline{\rho}$ (resp.\ $\tilde{\rho} = - \overline{\rho}$),
and the equation $\rho \neq \tilde{\rho}$ just means that $\rho$ does not lie on the unit circle
(resp.\ imaginary axis).
\subsection*{Refinement of a move of type IIo}
After the move of type IIo, the resulting $2 \times 2$ pole pencil has the form
\begin{displaymath}
\hat{A} - \lambda\hat{B} =
\left[\begin{array}{cc} 0 & \hat{\alpha}_{2} \\ \hat{\alpha}_{1} & \hat{\alpha}_{12} \end{array}\right]
- \lambda
\left[\begin{array}{cc} 0 & \hat{\beta}_{2} \\ \hat{\beta}_{1} & \hat{\beta}_{12} \end{array}\right]
\end{displaymath}
in principle. In practice the numbers that are supposed to be zero will not be exactly zero because of
roundoff errors in the computation.
If those numbers are small enough (a modest multiple of the unit roundoff), they can
simply be set to zero without compromising stability. If, on the other hand, they are not small enough,
a correction step can make them smaller. Now let's rewrite the pencil to take the nonzero entries into account.
At the same time we will recycle the notation, leaving off the hats for simplicity. We have
\begin{displaymath}
A - \lambda B = \left[\begin{array}{cc} \epsilon & \alpha_{2} \\ \alpha_{1} & \alpha_{12} \end{array}\right]
- \lambda \left[\begin{array}{cc} \eta & \beta_{2} \\ \beta_{1} & \beta_{12} \end{array}\right],
\end{displaymath}
where $\absval{\epsilon}$ and $\absval{\eta}$ are tiny but not small enough to be set to zero.
The correction step is explained in detail, and in greater generality, in Section~\ref{sec:refine}. Here we provide
a brief description. We look for tiny corrections $x$ and $y$ such that
\begin{displaymath}
\left[\begin{array}{cc} 1 & y \\ & 1 \end{array}\right]
\left[\begin{array}{cc} \epsilon & \alpha_{2} \\ \alpha_{1} & \alpha_{12} \end{array}\right]
\left[\begin{array}{cc} 1 & \\ x & 1 \end{array}\right] =
\left[\begin{array}{cc} 0 & \check{\alpha}_{2} \\ \check{\alpha}_{1} & \check{\alpha}_{12} \end{array}\right],
\end{displaymath}
and similarly for $B$. This yields a pair of equations
\begin{displaymath}
\epsilon + y\alpha_{1} + \alpha_{2}x + y\alpha_{12}x = 0, \quad
\eta + y\beta_{1} + \beta_{2}x + y\beta_{12}x = 0,
\end{displaymath}
which can be simplified by deleting the insignificant quadratic terms to yield the linear equations
\begin{equation}\label{eq:sylv1}
\left[\begin{array}{cc} \alpha_{1} & \alpha_{2} \\ \beta_{1} & \beta_{2} \end{array}\right]
\left[\begin{array}{c} y \\ x\end{array}\right] = - \left[\begin{array}{c} \epsilon \\ \eta \end{array}\right].
\end{equation}
Since the poles $\alpha_{1}/\beta_{1}$ and $\alpha_{2}/\beta_{2}$ are distinct, this system has a unique solution.
It is easy to show that in both the alternating and the palindromic cases, $y = \overline{x}$, so the contemplated
transformation is in fact a congruence. To make a unitary congruence we do a QR decomposition
\begin{displaymath}
\left[\begin{array}{cc} 1 & \\ x & 1 \end{array}\right] = QR.
\end{displaymath}
Because $x$ is tiny, $Q$ is close to the identity matrix. The orthogonal congruence by $Q$ is the desired correction:
\begin{displaymath}
Q^{*} \left( \left[\begin{array}{cc} \epsilon & \alpha_{2} \\ \alpha_{1} & \alpha_{12} \end{array}\right]
- \lambda \left[\begin{array}{cc} \eta & \beta_{2} \\ \beta_{1} & \beta_{12} \end{array}\right] \right) Q =
\left[\begin{array}{cc} \hat{\epsilon} & \hat{\alpha}_{2} \\ \hat{\alpha}_{1} & \hat{\alpha}_{12} \end{array}\right]
- \lambda \left[\begin{array}{cc} \hat{\eta} & \hat{\beta}_{2} \\ \hat{\beta}_{1} & \hat{\beta}_{12}
\end{array}\right],
\end{displaymath}
where $\hat{\epsilon}$ and $\hat{\eta}$ are now normally small enough to be set to zero.
In the unlikely event that they are not, the correction step can be repeated.
Again the computation is quite simple. Solving for $x$ in (\ref{eq:sylv1}), we obtain
\begin{displaymath}
x = \frac{\beta_{1}\epsilon - \alpha_{1}\eta}{\alpha_{1}\beta_{2} - \beta_{1}\alpha_{2}}.
\end{displaymath}
We need a unitary $Q$ whose first column is proportional to $\left[\begin{array}{cc} 1 & x\end{array}\right]^{T}$
or equivalently
\begin{displaymath}
v = \left[\begin{array}{c} v_{1} \\ v_{2}\end{array}\right] =
\left[\begin{array}{c} \alpha_{1}\beta_{2} - \beta_{1} \alpha_{2} \\ \beta_{1}\epsilon - \alpha_{1}\eta \end{array}\right].
\end{displaymath}
Thus we can take
\begin{displaymath}
Q = \left[\begin{array}{cc} c & -\overline{s} \\ s & \phantom{-}\overline{c}\end{array}\right],
\end{displaymath}
where $r = \norm{v}_{2}$, $c = v_{1}/r$, and $s = v_{2}/r$.
\subsection*{Move of Type IIe}
Now consider the case when $n$ is even. There is an odd number of poles $\sigma_{1}$, \ldots, $\sigma_{n-1}$.
The three poles in the middle are $\sigma_{k-1}$, $\sigma_{k}$, and $\sigma_{k+1}$, where $k = n/2$. Recall that
the pole $\sigma_{k}$ is a special unpaired pole, while $\sigma_{k-1}$ and $\sigma_{k+1}$ are linked by the
structure (palindromic or alternating). We need a move that swaps $\sigma_{k-1}$ with $\sigma_{k+1}$, while
leaving $\sigma_{k}$ where it is.
All of the action takes place in a $3 \times 3$ subpencil
\begin{displaymath}
A - \lambda B
= \left[\begin{array}{ccc} & & \alpha_{1} \\
& \alpha_{2} & \alpha_{21} \\
\alpha_{3} & \alpha_{32} & \alpha_{31} \end{array}\right]
- \lambda
\left[\begin{array}{ccc} & & \beta_{1} \\
& \beta_{2} & \beta_{21} \\
\beta_{3} & \beta_{32} & \beta_{31} \end{array}\right].
\end{displaymath}
Again we borrow the symbols $A$ and $B$ to denote the relevant submatrices. The subpencil inherits the
palindromic or alternating structure of the big pencil, but we will proceed at first
as if $A$ and $B$ were unrelated.
The poles are $\sigma_{k-1} = \alpha_{3}/\beta_{3}$, $\sigma_{k} = \alpha_{2}/\beta_{2}$,
and $\sigma_{k+1} = \alpha_{1}/\beta_{1}$. We want to make a unitary
congruence transformation that swaps the positions of $\sigma_{k-1}$ and $\sigma_{k+1}$ while
leaving $\sigma_{k}$ in the middle. We proceed just as in the case of a $2 \times 2$ swap.
We flip the rows and columns to get
\begin{displaymath}
FAF - \lambda FBF =
\left[\begin{array}{ccc}
\alpha_{31} & \alpha_{32} & \alpha_{3} \\ \alpha_{21} & \alpha_{2} & \\ \alpha_{1} & &
\end{array}\right]
- \lambda
\left[\begin{array}{ccc}
\beta_{31} & \beta_{32} & \beta_{3} \\ \beta_{21} & \beta_{2} & \\ \beta_{1} & &
\end{array}\right],
\end{displaymath}
which has the poles in the desired locations but the ``wrong'' triangularity.
As in the $2 \times 2$ case, we find it convenient to work with partially flipped forms
\begin{displaymath}
AF - \lambda BF =
\left[\begin{array}{ccc} \alpha_{1} & & \\
\alpha_{21}& \alpha_{2} & \\ \alpha_{31} & \alpha_{32} & \alpha_{3} \end{array}\right]
- \lambda
\left[\begin{array}{ccc} \beta_{1} & & \\
\beta_{21} & \beta_{2} & \\ \beta_{31} & \beta_{32} & \beta_{3} \end{array}\right].
\end{displaymath}
Next we set up and solve some Sylvester equations to diagonalize the pencil. Specifically, we will find unit
lower triangular
\begin{displaymath}
X = \left[\begin{array}{ccc} 1 & & \\ x_{21} & 1 & \\ x_{31} & x_{32} & 1 \end{array}\right]
\quad\mbox{and}\quad
Y = \left[\begin{array}{ccc} 1 & & \\ y_{21} & 1 & \\ y_{31} & y_{32} & 1 \end{array}\right]
\end{displaymath}
such that
\begin{equation}\label{eq:sylvic}
(AF)X = Y(\check{A}F) \quad\mbox{and}\quad (BF)X = Y(\check{B}F),
\end{equation}
where $\check{A}$ and $\check{B}$ are anti-diagonal matrices with the same anti-diagonals as $A$ and $B$,
respectively. This is the $3 \times 3$ analog of (\ref{eq:sylv}). Writing the first of these equations out in detail, we have
\begin{displaymath}
\left[\begin{array}{ccc} \alpha_{1} & & \\
\alpha_{21} & \alpha_{2} & \\ \alpha_{31} & \alpha_{32} & \alpha_{3} \end{array}\right]
\left[\begin{array}{ccc} 1 & & \\ x_{21} & 1 & \\ x_{31} & x_{32} & 1 \end{array}\right] =
\left[\begin{array}{ccc} 1 & & \\ y_{21} & 1 & \\ y_{31} & y_{32} & 1 \end{array}\right]
\left[\begin{array}{ccc} \alpha_{1} & & \\
& \alpha_{2} & \\ & & \alpha_{3} \end{array}\right],
\end{displaymath}
and similarly for the $B$ equation. Altogether this is a system of six linear equations in
the six unknowns $x_{21}$, $x_{31}$, $x_{32}$, $y_{21}$, $y_{31}$, and $y_{32}$, but fortunately
it turns out to be three systems of two equations that are nearly independent of one another:
\begin{displaymath}
\left[\begin{array}{cc} \alpha_{1} & \alpha_{2} \\ \beta_{1} & \beta_{2} \end{array}\right]
\left[\begin{array}{c} \phantom{-}y_{21} \\ -x_{21} \end{array}\right] =
\left[\begin{array}{c} \alpha_{21} \\ \beta_{21} \end{array}\right],
\end{displaymath}
which has a unique solution if and only if $\alpha_{1}/\beta_{1} \neq \alpha_{2}/\beta_{2}$,
\begin{displaymath}
\left[\begin{array}{cc} \alpha_{2} & \alpha_{3} \\ \beta_{2} & \beta_{3} \end{array}\right]
\left[\begin{array}{c} \phantom{-}y_{32} \\ - x_{32} \end{array}\right] =
\left[\begin{array}{c} \alpha_{32} \\ \beta_{32} \end{array}\right],
\end{displaymath}
which has a unique solution if and only if $\alpha_{2}/\beta_{2} \neq \alpha_{3}/\beta_{3}$, and
\begin{displaymath}
\left[\begin{array}{cc} \alpha_{1} & \alpha_{3} \\ \beta_{1} & \beta_{3} \end{array}\right]
\left[\begin{array}{c} \phantom{-}y_{31} \\ -x_{31} \end{array}\right] =
\left[\begin{array}{c} \alpha_{31} + \alpha_{32}x_{21} \\ \beta_{31} + \beta_{32}x_{21} \end{array}\right],
\end{displaymath}
which has a unique solution if and only if $\alpha_{1}/\beta_{1} \neq \alpha_{3}/\beta_{3}$.
We conclude that there are unique unit lower triangular $X$ and $Y$ such that (\ref{eq:sylvic}) holds
if and only if the poles are distinct. We can easily and stably compute $X$ and $Y$.
Rewriting (\ref{eq:sylvic}) we have the equivalence
\begin{displaymath
(YF)^{-1}(A - \lambda B)(FX) = F(\check{A} - \lambda\check{B})F =
\left[\begin{array}{ccc} & & \alpha_{3} \\ & \alpha_{2} & \\ \alpha_{1} & & \end{array}\right]
- \lambda
\left[\begin{array}{ccc} & & \beta_{3} \\ & \beta_{2} & \\ \beta_{1} & & \end{array}\right],
\end{displaymath}
which is the $3 \times 3$ analog of (\ref{eq:nonuneq2}). From here on the argument is exactly the
same as in the $2 \times 2$ case, starting from (\ref{eq:nonuneq2}), with obvious trivial modifications.
In the end we get a congruence $Q^{*}(A - \lambda B)Q = \hat{A} - \lambda \hat{B}$
that makes the desired swap.
The computation is again straightforward. We need a unitary $Q$ such that $FX = QR$. The entries of $X$
can be computed by three applications of Cramer's rule. Then the $QR$ decomposition can be computed, producing
$Q$ as a product of three core transformations. $Q$ depends on the first two columns of $FX$, and these can
be rescaled to avoid the divisions implied by Cramer's rule without altering $Q$.
We emphasize that this move succeeds if the three poles in question are all distinct. In our application below,
the poles will be $\rho$, $\sigma_{n/2}$, $\tilde{\rho}$. In the palindromic case, $\sigma_{n/2}$ is an unpaired pole
that must lie on the unit circle, and $\tilde{\rho} = 1/\overline{\rho}$. Thus $\rho \neq \tilde{\rho}$ if and only if $\rho$
and $\tilde{\rho}$ are not on the unit circle. Thus, if $\rho \neq \tilde{\rho}$, then all three poles are
automatically distinct. The same is true in the alternating case by a similar argument involving the imaginary axis
instead of the unit circle.
\subsection*{Refinement of a move of type IIe}
We showed above that a move of type IIo can be refined if necessary, and the same is true of a move of type IIe.
After the move of type IIe, the resulting $3 \times 3$ pole pencil has the form
\begin{displaymath}
A - \lambda B = \left[\begin{array}{ccc}
\epsilon_{31} & \epsilon_{32} & \alpha_{3} \\
\epsilon_{21} & \alpha_{2} & \alpha_{23} \\ \alpha_{1} & \alpha_{12} & \alpha_{13} \end{array}\right]
- \lambda \left[\begin{array}{ccc}
\eta_{31} & \eta_{32} & \beta_{3} \\
\eta_{21} & \beta_{2} & \beta_{23} \\ \beta_{1} & \beta_{12} & \beta_{13} \end{array}\right].
\end{displaymath}
We have simplified the notation by leaving off the hats. The numbers
$\absval{\epsilon_{ij}}$ and $\absval{\eta_{ij}}$ are tiny and would have been zero except for roundoff
errors. If they are not small enough to be ignored, we must do a refinement step. The ``upside down'' notation
used here reflects the fact that in the analysis above, we flipped the rows of the matrices. We could do the
same thing here (flip the rows), but for this brief summary we will not bother. For details see Section~\ref{sec:refine}.
For the correction step we look for matrices
\begin{displaymath}
X = \left[\begin{array}{ccc} 1 & & \\ x_{21} & 1 & \\ x_{31} & x_{32} & 1 \end{array}\right] \quad\mbox{and}\quad
Y = \left[\begin{array}{ccc} 1 & y_{12} & y_{13} \\ & 1 & y_{23} \\ & & 1 \end{array}\right]
\end{displaymath}
that set the tiny numbers to zero:
\begin{equation}\label{eq:riccati3}
Y(A - \lambda B)X = \left[\begin{array}{ccc}
0 & 0 & \check{\alpha}_{3} \\
0 & \check{\alpha}_{2} & \check{\alpha}_{23} \\
\check{\alpha}_{1} & \check{\alpha}_{12} & \check{\alpha}_{13} \end{array}\right]
- \lambda \left[\begin{array}{ccc}
0 & 0 & \check{\beta}_{3} \\
0 & \check{\beta}_{2} & \check{\beta}_{23}
\\ \check{\beta}_{1} & \check{\beta}_{12} & \check{\beta}_{13} \end{array}\right].
\end{equation}
Since the corrections to be made are tiny, we expect all of the the numbers $x_{ij}$ and $y_{ij}$ to be tiny.
Writing out (\ref{eq:riccati3}) in detail we get a system of six quadratic equations in six unknowns. The
quadratic terms and a few others are negligible. Eliminating negligible terms we get six linear equations
that have a unique solution as long as the poles are distinct. See Section~\ref{sec:refine} for details.
In both alternating and palindromic cases, the transforming matrices $X$ and $Y$ are related by $Y = X^{*}$,
which simplifies the situation further. In either case, the equations are easily and stably solved.
Once $X$ has been computed, the decomposition $X = QR$ supplies the needed unitary transforming
matrix. The congruence $Q^{*}(A - \lambda B)Q = \hat{A} - \lambda \hat{B}$ yields the desired refinement.
Remark: The moves of types IIo and IIe are special cases of constructions presented in \cite[\S\S 4.1,\,8.1]{KrScWa09}, but we
have taken a different approach here. In \cite{KrScWa09}
the palindromic and alternating cases were considered separately. Here we have looked at the unstructured case first
and shown how to do the swap using a unitary equivalence. Then we have shown that in both of our structured cases
the equivalence is actually a congruence.
We have also added a refinement step, which was not contemplated in \cite{KrScWa09}, to make the algorithm
more robust.
\section{Stability}
We can construct a variety of algorithms from the moves. If each move is backward stable, then any algorithm
built from moves must also be backward stable. We therefore take a moment to consider this question.
Standard backward error analysis \cite{Wil65} shows that moves of type I are backward stable.
If the moves of type II are implemented as shown in \cite{CaMaVaWa19}, they never fail and are
always backward stable.
Moves of types IIo and IIe require the solution of small linear systems that are nonsingular if and
only if the poles involved in the swap are distinct. As we will see below, there are other good reasons
(involving convergence rates) for keeping these poles distinct and preferably far apart. Assuming this
is done, we can expect these moves to be stable. There is a natural stability test associated with these moves:
check that the numbers that are supposed to be zero really are (almost) zero. For the
event that they are not, we have described a refinement step that can be used to make them smaller.
The refinement can be repeated if necessary, though this should be vary rare. Because of the refinement
step, we can say for sure that moves of types IIo and IIe are backward stable, provided that we do not attempt
to swap two equal poles.
\section{Building an algorithm using the moves}\label{sec:algorithm}
First suppose our pair $(A,B)$ has odd dimension $n$, and its poles are
$\sigma_{1}$, $\sigma_{2}$, \ldots, $\sigma_{m}$, $\tilde{\sigma}_{m}$, \ldots, $\tilde{\sigma}_{2}$, $\tilde{\sigma}_{1}$,
where $m = (n-1)/2$, and
$\tilde{\sigma}_{i} = 1/\overline{\sigma}_{i}$ (resp.\ $-\overline{\sigma}_{i}$) in the palindromic (resp.\ alternating) case.
One iteration of the most basic algorithm would proceed as
follows. First a shift $\rho$ is chosen and inserted in place of $\sigma_{1}$ by a move of type I. This move also inserts
a shift $\tilde{\rho}$ in place of $\tilde{\sigma}_{1}$ at the other end. A simple choice of $\rho$ would be the
\emph{Rayleigh quotient shift} $\rho = a_{1,n}/b_{1,n}$. Then a move of type II is used to interchange
$\rho$ with $\sigma_{2}$ and $\tilde{\rho}$ with $\tilde{\sigma}_{2}$. Then another move of type II is used to interchange
$\rho$ with $\sigma_{3}$, and so on. After $m-1$ such moves, the poles will be
$\sigma_{2}$, \ldots, $\sigma_{m}$, $\rho$, $\tilde{\rho}$, $\tilde{\sigma}_{m}$, \ldots, $\tilde{\sigma}_{2}$, with $\rho$
and $\tilde{\rho}$ side by side in the middle.
Then a move of type IIo can be used to swap them, provided that $\rho \neq \tilde{\rho}$. Then additional moves of type II are used to push $\rho$ and $\tilde{\rho}$ further along, that is, $\rho$ is swapped with $\tilde{\sigma}_{m}$,
then $\tilde{\sigma}_{m-1}$, and so on, while $\tilde{\rho}$ is swapped with $\sigma_{m}$, $\sigma_{m-1}$, etc.
Once the shifts arrive at the edge of the pencil, they can be removed by a move of type I, which would replace them by new poles, which could be the original poles $\sigma_{1}$, $\tilde{\sigma_{1}}$, or they could be different. This completes
one iteration of the basic algorithm.
The case of even dimension is the same, except that there is an extra unpaired pole $\sigma_{n/2}$ in the middle.
Type II operations push the shifts $\rho$ and $\tilde{\rho}$ toward the middle, as in the odd case, until the configuration
of poles is $\sigma_{2}$, \ldots, $\rho$, $\sigma_{n/2}$, $\tilde{\rho}$, \ldots, $\tilde{\sigma_{2}}$. Then a move of
type IIe is used to swap $\rho$ and $\tilde{\rho}$ while leaving $\sigma_{n/2}$ fixed. This can be done
provided that $\rho \neq \tilde{\rho}$. Once this exchange has been made, the iteration is completed in the same way
as in the odd case.
Repeated iterations of the basic algorithm with good choices of shifts $\rho$ and $\tilde{\rho}$ will cause the pencil
to tend toward triangular form, exposing the eigenvalues on the anti-diagonal. Not all eigenvalues appear at once.
Good shifts can (usually) cause $a_{n-1,1} \to 0$ and $b_{n-1,1} \to 0$ in just a few iterations; the convergence rate is
typically quadratic. By symmetry we must also have $a_{1,n-1} \to 0$ and $b_{1,n-1} \to 0$ at the same rate.
This exposes a pair of eigenvalues $\lambda = a_{n,1}/b_{n,1}$ and $\tilde{\lambda} = a_{1,n}/b_{1,n}$ at the ends.
Then the problem can be deflated to size $n-2$, and we can go after the next pair of eigenvalues, and so on. All of
this is a consequence of the convergence theorem stated below. We will not present a detailed explanation
because the arguments are the same as in the unstructured case.
\subsection*{A convergence theorem}
The mechanism that drives all variants of Francis's algorithm, including the QZ algorithm, is nested subspace
iteration with changes of coordinate system. See
\cite[p.~431]{Wat10}, \cite[p.~399]{Wat11}, or \cite[Theorem~2.2.3]{AuMaRoVaWa18}.
This is also true of our basic algorithm sketched above. We just need to take a few lines to
set the scene. We make the (generically valid) assumption that none of the poles or shifts is exactly an eigenvalue
of the pencil $A - \lambda B$. We continue to cover the alternating and palindromic cases simultaneously.
Each iteration begins with the choice of a shift $\rho$ and a companion shift $\tilde{\rho}$
($= 1/\overline{\rho}$ in the palindromic case and $-\overline{\rho}$ in the alternating case).
The result of the iteration is a new structured
anti-Hessenberg pencil $\hat{A} - \lambda \hat{B}$ satisfying
\begin{equation}\label{eq:palcoord}
\hat{A} - \lambda \hat{B} = Q^{*}(A - \lambda B)Q.
\end{equation}
We need to define two nested sequences of subspaces.
For $k=1$, \ldots, $n$, define
\begin{displaymath}
\mbox{$\mathcal{E}$}_{k} = \spn{e_{1}, \ldots, e_{k}},
\end{displaymath}
where $e_{1}$, \ldots, $e_{n}$ are the standard basis vectors. Then define
\begin{displaymath}
\mbox{$\mathcal{Q}$}_{k} = Q\mbox{$\mathcal{E}$}_{k},
\end{displaymath}
the space spanned by the first $k$ columns of $Q$.
\begin{theorem}\label{thm:palspace}
A single step of the basic algorithm described above with shift $\rho$ effects nested subspace iterations
\begin{displaymath}
\mbox{$\mathcal{Q}$}_{k} = (A - \tilde{\rho}B)^{-1}(A - \rho B)\mbox{$\mathcal{E}$}_{k}, \qquad \qquad k = 1,\ldots,n-1.
\end{displaymath}
The change of coordinate system (\ref{eq:palcoord}) transforms $\mbox{$\mathcal{Q}$}_{k}$ back to $\mbox{$\mathcal{E}$}_{k}$.
\end{theorem}
This theorem makes no mention of convergence, but we call it a \emph{convergence theorem} anyway.
This result and ones like it can be used together with the convergence theory of
subspace iteration to draw conclusions about
the convergence of the algorithm, as explained in \cite{Wat07,Wat10,Wat11} and elsewhere.
\begin{proof}
We sketch the proof, relying on Theorem~5.2 of \cite{CaMaVaWa19}. That theorem applies to upper-Hessenberg
pencils, so we can apply it to the flipped pencil $FA - \lambda FB$. Notice that
$(FA - \tilde{\rho}FB)^{-1}(FA - \rho FB) = (A - \tilde{\rho}B)^{-1}(A - \rho B)$.
First suppose $n$ is odd and $k < (n-1)/2$. According to (the ``$Z$'' part of)
Theorem~5.2 of \cite{CaMaVaWa19}, the action on $\mbox{$\mathcal{E}$}_{k}$
depends on the two moves of type II that take place at the ``$k$th position'', by which we mean the spot originally
occupied by poles $\sigma_{k}$ and $\sigma_{k+1}$. The basic algorithm inserts the shift $\rho$, swapping it with
$\sigma_{1}$, \ldots, $\sigma_{k}$. The first swap at the $k$th position is an exchange of $\rho$ with $\sigma_{k+1}$,
which (see \cite[Theorem~5.2]{CaMaVaWa19}) generates a factor $(z-\rho)/(z - \sigma_{k+1})$. The only other swap
at the $k$th position happens when $\sigma_{k+1}$ is swapped with $\tilde{\rho}$, moving $\sigma_{k+1}$ back to
its original position. This introduces a factor $(z - \sigma_{k+1})/(z - \tilde{\rho})$. The action on $\mbox{$\mathcal{E}$}_{k}$ is determined
by the product of the factors, which is $r(z) = (z - \rho)/(z - \tilde{\rho})$. Specifically, $\mbox{$\mathcal{E}$}_{k}$ is transformed to
$\mbox{$\mathcal{Q}$}_{k} = r(B^{-1}A)\mbox{$\mathcal{E}$}_{k}$. Since $r(B^{-1}A) = (B^{-1}A - \tilde{\rho}I)^{-1}(B^{-1}A - \rho I) =
(A - \tilde{\rho}B)^{-1}(A - \rho B)$, we have $\mbox{$\mathcal{Q}$}_{k} = (A - \tilde{\rho}B)^{-1}(A - \rho B)\mbox{$\mathcal{E}$}_{k}$, as claimed.
The case $k > (n-1)/2$ is the same, except that the order of the swaps is reversed. This does not change the outcome.
The case $k = (n-1)/2$ is different because this is the spot in the middle where only a single swap
takes place, interchanging the shifts $\rho$ and $\tilde{\rho}$.
Using \cite[Theorem~5.2]{CaMaVaWa19} again, we see that we
just have a single factor $(z - \rho)/(z - \tilde{\rho})$, so the result is the same as in the other cases.
This case can also be deduced directly from \cite[Theorem~4.3]{CaMaVaWa19}, which is a precursor of
\cite[Theorem~5.2]{CaMaVaWa19}.
In the case of even $n$, the argument is the same as in the odd case if $k < n/2 -1$ or $k > n/2$. The only question
mark is in the cases $k = n/2-1$ and $k= n/2$, which are affect by (and only by) the move of type IIe in the middle.
Since this move, which affects three adjacent poles instead of two, is different from a standard move of type II, we
must check what happens here. We leave it to the reader to verify that the proof of
\cite[Theorem~4.3]{CaMaVaWa19},
which applies to moves of type II, is also valid for moves of type IIe. Once this has been checked, the proof is
complete. \hfill\end{proof}
In our discussion of moves of type IIo and IIe we had to make the assumption $\rho \neq \tilde{\rho}$ in order to
ensure that the moves are possible. Theorem~\ref{thm:palspace} gives us another reason for this assumption:
If $\rho = \tilde{\rho}$, we have $(A - \tilde{\rho}B)^{-1}(A - \rho B)=I$, and the iteration goes nowhere. The action of the shift $\rho$
traveling in one direction is exactly cancelled by the shift $\tilde{\rho}$ traveling in the opposite direction.
In the alternating case the requirement $\rho \neq \tilde{\rho}$ means that the shifts should not lie on the imaginary axis.
It follows that this method will only be useful for finding eigenvalues that are not purely imaginary. This is not necessarily
a weakness. For the continuous-time optimal control problems mentioned in the introduction \cite{Meh91}, mild assumptions
guarantee that no eigenvalues are on the imaginary axis; exactly half are in the open left half plane and half are in the
open right half plane. Our basic algorithm will have no problem computing all of these eigenvalues; they will be
extracted in $(\lambda,-\overline{\lambda})$ pairs.
In the palindromic case the requirement $\rho \neq \tilde{\rho}$ means that the shifts should not lie on the unit
circle, so this method will only be useful for finding eigenvalues that have modulus different from 1. For the
discrete-time control problems mentioned above, mild assumptions guarantee that no eigenvalues lie on the
unit circle; exactly half are inside and half are outside. Our algorithm will easily compute all of these eigenvalues,
and they will be extracted in $(\lambda,1/\overline{\lambda})$ pairs.
The bulge-chasing algorithm in \cite{KrScWa09} has the same ($\rho \neq \tilde{\rho}$) restriction. In fact we can
show that our basic algorithm is a generalization of the algorithm in
\cite{KrScWa09}.\footnote{In \cite{KrScWa09} we considered multi-shift bulge-chasing algorithms of arbitrary degree.
Here we are considering only a single-shift algorithm, and this generalizes the single-shift version of the algorithm in
\cite{KrScWa09}.}
Let's take a look at this algorithm, beginning with the palindromic case.
The pencil is $A - \lambda A^{*}$ with $A$ assumed to be \emph{anti-Hessenberg},
i.e.\
\begin{displaymath}
A =
\parbox{3.2cm}{
\begin{tikzpicture}[scale=1.66,y=-1cm]
\draw (-.15,-.1) -- (-.2,-.1) -- (-.2,1.5) -- (-.15,1.5);
\draw (1.55,-.1) -- (1.6,-.1) -- (1.6,1.5) -- (1.55,1.5);
\foreach \j in {0,...,7}{
\foreach \i in {\j,...,7}{\node at (\i/5,1.4-\j/5)
[align=center,scale=1.0]{$\times$};}}
\foreach \j in {0,...,6}{\node at (\j/5,1.4-\j/5-.2)
[align=center,scale=1.0]{$\times$};}
\end{tikzpicture}
}.
\end{displaymath}
The algorithm presented in \cite{KrScWa09}
begins with a reduction step that introduces some zeros above the anti-diagonal,
transforming $A$ to a partially anti-triangular form
\begin{equation}\label{eq:partanttri}
\parbox{3.2cm}{
\begin{tikzpicture}[scale=1.66,y=-1cm]
\draw (-.15,-.1) -- (-.2,-.1) -- (-.2,1.5) -- (-.15,1.5);
\draw (1.55,-.1) -- (1.6,-.1) -- (1.6,1.5) -- (1.55,1.5);
\foreach \j in {0,...,7}{
\foreach \i in {\j,...,7}{\node at (\i/5,1.4-\j/5)
[align=center,scale=1.0]{$\times$};}}
\foreach \j in {0,...,3}{\node at (\j/5,1.4-\j/5-.2)
[align=center,scale=1.0]{$\times$};}
\end{tikzpicture}
}.
\end{equation}
A bulge-chasing algorithm is then applied to this partially reduced form to expose the
eigenvalues $\{\lambda,\, \overline{\lambda}^{-1}\}$ satisfying $\absval{\lambda} \neq 1$
in pairs.
From our current vantage point we can see that the preliminary reduction step is just
a process of introducing zero and infinite poles into the pencil $A - \lambda A^{*}$ by moves
of types I and II. This step is necessary for the bulge-chasing algorithm in \cite{KrScWa09},
but it is not needed for the pole swapping algorithm discussed in this paper;
we can go to work right away on the anti-Hessenberg pencil $A - \lambda A^{*}$.
The algorithm in \cite{KrScWa09} for the alternating case is a little bit different. Starting with
an alternating pencil $A - \lambda B$ in anti-Hessenberg form, this algorithm requires a preliminary
step to transform $B$ to anti-triangular form, leaving $A$ anti-Hessenberg. Then a bulge-chasing
algorithm is applied. The modified pencil has all poles equal to infinity.
Again, from our new viewpoint, we can see that the preliminary reduction
is nothing but a process of introducing infinite poles by moves of types I and II. This is necessary
for the bulge-chasing algorithm in \cite{KrScWa09}, but it is not needed for our pole-swapping algorithm.
In both cases the flop count for this additional reduction is $O(n^{3})$ ($O(n^{2})$ moves at $O(n)$ flops per move),
so the cost is not insignificant.
For a single-shift iteration of the algorithm in \cite{KrScWa09} we have a theorem just like Theorem~\ref{thm:palspace}.
If we introduce a shift $\rho$ at one end, we automatically introduce a complementary shift $\tilde{\rho}$
( $= 1/\overline{\rho}$ or $-\overline{\rho}$) as always, at the other end.
The setup is the same as for Theorem~\ref{thm:palspace}, and the new result looks like this:
\begin{theorem}\label{thm:palspace2}
A single step of the algorithm in \cite{KrScWa09} with shift $\rho$ effects nested subspace iterations
\begin{displaymath}
\mbox{$\mathcal{Q}$}_{k} = (A - \tilde{\rho}B)^{-1}(A - \rho B)\mbox{$\mathcal{E}$}_{k}, \qquad \qquad k = 1,\ldots,n-1.
\end{displaymath}
The change of coordinate system (\ref{eq:palcoord}) transforms $\mbox{$\mathcal{Q}$}_{k}$ back to $\mbox{$\mathcal{E}$}_{k}$.
\end{theorem}
This is an immediate consequence of \cite[Theorem~7.3.1]{Wat07}. Note that Theorem~\ref{thm:palspace2} is identical to
Theorem~\ref{thm:palspace}. The only difference is that the bulge-chasing
algorithm requires a special Hessenberg-triangular form, as described immediately above. Since the action of the
algorithms is the same, we deduce that the pole-swapping algorithm is a generalization of the bulge-chasing
algorithm.
Remark: For clarity we have focused on the most basic possible pole-swapping algorithm for the palindromic and alternating
problems. One can consider variants that introduce multiple shifts and draw the same conclusions.
For more ideas see \cite{CaMaVaWa19}, which also contains a more explicit demonstration of the
connection between bulge-chasing and pole-swapping algorithms.
\section{Justification of the refinement step}\label{sec:refine}
In Section~\ref{sec:operations} we briefly described refinement procedures that can be applied (occasionally)
after moves of types IIo and IIe. Here we provide complete justifications for those procedures.
The algorithm in \cite{KrScWa09} also has a move in the middle to which a refinement step could be applied.
In that paper it was acknowledged that a failure might
occasionally occur, but it was reported that in the course of the various tests, no failures were observed.
However, since a failure might occur at any time, it would make sense to add the refinement step to the algorithm
of \cite{KrScWa09} to make it more robust. In order to accommodate its use in the context of \cite{KrScWa09},
we have made our discussion of the refinement procedure more general than is strictly required for this paper.
First we consider the case of no middle pole(s), as in a move of type IIo. Suppose we have just swapped $m$ poles
with $m$ other poles. After the swap we have a bulge pencil
\begin{displaymath}
\left[\begin{array}{cc}
E_{11} & A_{12} \\ A_{21} & A_{22}
\end{array}\right] - \lambda
\left[\begin{array}{cc}
G_{11} & B_{12} \\ B_{21} & B_{22}
\end{array}\right],
\end{displaymath}
where the submatrices are $m \times m$.
The matrices $E_{11}$ and $G_{11}$ would be zero if there were no roundoff errors,
so $\norm{E_{11}} \ll \norm{A}$ and $\norm{G_{11}} \ll \norm{B}$. We assume that the eigenvalues (the poles)
of the subpencil $A_{21} - \lambda B_{21}$ are disjoint from those of $A_{12} - \lambda B_{12}$. In the case
of a move of type IIo we have $m=1$, but we are now allowing larger $m$ to take into account the scenario of
\cite{KrScWa09}.
If $\norm{E_{11}}$ and $\norm{G_{11}}$ are not small enough, we must do a refinement step.
To this end we seek $X$ and $Y$ such that
\begin{equation}\label{eq:genric1}
\begin{array}{c}
\left[\begin{array}{cc} I & Y \\ & I \end{array}\right]
\left[\begin{array}{cc} E_{11} & A_{12} \\ A_{21} & A_{22} \end{array}\right]
\left[\begin{array}{cc} I & \\ X & I \end{array}\right] =
\left[\begin{array}{cc} 0 & \check{A}_{12} \\
\check{A}_{21} & \check{A}_{22} \end{array}\right], \\ \\
\left[\begin{array}{cc} I & Y \\ & I \end{array}\right]
\left[\begin{array}{cc} G_{11} & B_{12} \\ B_{21} & B_{22} \end{array}\right]
\left[\begin{array}{cc} I & \\ X & I \end{array}\right] =
\left[\begin{array}{cc} 0 & \check{B}_{12} \\ \check{B}_{21} & \check{B}_{22}
\end{array}\right].
\end{array}
\end{equation}
By straightforward computation we find that (\ref{eq:genric1}) holds if and only if the algebraic Riccati equations
\begin{equation} \label{eq:genric2}
\begin{array}{ccc}
A_{12}X + YA_{21} + E_{11} + YA_{22}X & = & 0 \\
B_{12}X + YB_{21} + G_{11} + YB_{22}X & = & 0
\end{array}
\end{equation}
hold.
Since $\norm{E_{11}}$ and $\norm{G_{11}}$ are tiny, we expect that the corrections
$X$ and $Y$ will be tiny as well. Therefore the quadratic terms $YA_{22}X$ and $YB_{22}X$ in (\ref{eq:genric2})
should be negligible, and (\ref{eq:genric2}) should be well approximated by the Sylvester equations
\begin{equation} \label{eq:gensyl1}
\begin{array}{ccc}
A_{12}X + YA_{21} + E_{11} & = & 0 \\
B_{12}X + YB_{21} + G_{11} & = & 0.
\end{array}
\end{equation}
These linear equations have a unique solution $(X,Y)$ if and only if the eigenvalues of the pencil
$A_{12} - \lambda B_{12}$ are disjoint
from those of $A_{21} - \lambda B_{21}$ \cite[Theorem~6.6.8]{Wat07}. Under this assumption one can prove
by standard arguments \cite{Ste72,Ste73}, \cite[\S~2.7]{Wat07} (using the contraction mapping principle)
that the Riccati equations (\ref{eq:genric2}) have
a unique small solution $(X,Y)$ if $\norm{E_{11}}$ and $\norm{G_{11}}$ are sufficiently small. By this we mean
that, although (\ref{eq:genric2}) typically has many solutions,
there is exactly one for which $\norm{X}$ and $\norm{Y}$ are small, and this is the solution that is of interest to us.
So far we have ignored the special structure of the pencil. Now let's see what we can say in the alternating
case, for which $A_{12}^{*} = A_{21}$, $A_{22}^{*} = A_{22}$, $E_{11}^{*} = E_{11}$, $B_{12}^{*} = -B_{21}$,
$B_{22}^{*} = -B_{22}$, and $G_{11}^{*} = -G_{11}$. If we make these substitutions in (\ref{eq:genric2}) and then
take conjugate transposes, we get
\begin{equation}\label{eq:genric3}
\begin{array}{ccc}
A_{12}Y^{*} + X^{*}A_{21} + E_{11} +X^{*}A_{22}Y^{*} = 0 \\
B_{12}Y^{*} + X^{*}B_{21} + G_{11} +X^{*}B_{22}Y^{*} = 0.
\end{array}
\end{equation}
Now consider the palindromic case, for which $A_{12}^{*}=B_{21}$, $A_{21}^{*} = B_{12}$, $A_{22}^{*} = B_{22}$,
and $E_{11}^{*} = G_{11}$. Making these substitutions in (\ref{eq:genric2}) and then taking conjugate transposes,
we find again that we get (\ref{eq:genric3}).
From (\ref{eq:genric3}) we deduce that, in both of our structured cases,
$(Y^{*},X^{*})$ is a solution of (\ref{eq:genric2}) if and only if $(X,Y)$ is.
Since (\ref{eq:genric2}) has a unique small-norm solution, we deduce that $(Y^{*},X^{*}) = (X,Y)$, or briefly
$Y = X^{*}$. If we now make this substitution into (\ref{eq:genric1}), we see that the transformation is a congruence,
which is exactly what is needed for the preservation of the two structures.
If we solve (\ref{eq:genric2}) for $X$, we can carry out the congruence (\ref{eq:genric1}) to make the desired correction.
However, (\ref{eq:genric1}) has the shortcoming that it is not unitary. To get a unitary congruence
that has the same effect, we can do a QR decomposition
\begin{equation}\label{eq:qr}
\left[\begin{array}{cc} I & \\ X & I\end{array}\right] = QR =
\left[\begin{array}{cc} Q_{11} & Q_{12} \\ Q_{21} & Q_{22} \end{array}\right]
\left[\begin{array}{cc} R_{11} & R_{12} \\ & R_{22} \end{array}\right],
\end{equation}
where $Q$ is unitary and $R$ is upper triangular, and do a congruence with $Q$. Substituting $QR$ into
(\ref{eq:genric1}) in two places and inverting the triangular matrices $R^{*}$ and $R$, we obtain
\begin{displaymath}
Q^{*}
\left[\begin{array}{cc} E_{11} & A_{12} \\ A_{21} & A_{22} \end{array}\right]
Q =
R^{-*}\left[\begin{array}{cc} 0 & \check{A}_{12} \\
\check{A}_{21} & \check{A}_{22} \end{array}\right]R^{-1} =
\left[\begin{array}{cc} 0 & \hat{A}_{12} \\ \hat{A}_{21} & \hat{A}_{22} \end{array}\right],
\end{displaymath}
where
\begin{displaymath}
\hat{A}_{12} = R_{11}^{-*}\check{A}_{12}R_{22}^{-1} \quad\mbox{and}\quad
\hat{A}_{21} = R_{22}^{-*}\check{A}_{21}R_{11}^{-1}.
\end{displaymath}
We have displayed only the ``$A$'' equations, but the ``$B$'' equations are the same.
We now have all of the ingredients we need for an update.
The discussion so far suggests that we will compute $X$ by solving
the Riccati equations (\ref{eq:genric2}), but in fact we will not. Instead we will solve the Sylvester equations
(\ref{eq:gensyl1}) to get an excellent approximation. This amounts to one step of Newton's method applied to
(\ref{eq:genric2}) using initial guess $X^{(0)} = Y^{(0)} = 0$. Notice that the symmetry argument that we applied
to (\ref{eq:genric2}) above also applies to the Sylvester equation: In both the alternating and palindromic cases,
if $(X,Y)$ is the unique solution of (\ref{eq:gensyl1}), then so is $(Y^{*},X^{*})$, so $Y = X^{*}$.
When we solve (\ref{eq:gensyl1}) we can take symmetry into account. In the palindromic case, the two
matrix equations of (\ref{eq:gensyl1}) are equivalent, so we just have to solve
\begin{displaymath}
A_{12}X + X^{*}A_{21} + E_{11} = 0
\end{displaymath}
for $X$. Separating real and imaginary parts, we can write this as a system of $2m^{2}$ real linear equations
in $2m^{2}$ unknowns, the real and imaginary parts of $X$. This can be solved stably by conventional means.
In the alternating case the $A$ equation is symmetric and the $B$ equation is skew symmetric. Taking these
symmetries into account, we again get a system of $2m^{2}$ real equations in $2m^{2}$ real unknowns, which
can be solved stably by conventional means.
We can now summarize our refinement step: Solve the Sylvester equations for $X$, taking the symmetry into
account. Then perform the QR decomposition (\ref{eq:qr}), and use the resulting $Q$ to effect a
unitary congruence transform
\begin{displaymath}
Q^{*}
\left[\begin{array}{cc} E_{11} & A_{12} \\
A_{21} & A_{22} \end{array}\right] Q =
\left[\begin{array}{cc} \hat{E}_{11} & \hat{A}_{12} \\
\hat{A}_{21} & \hat{A}_{22} \end{array}\right], \quad
Q^{*}
\left[\begin{array}{cc} G_{11} & B_{12} \\
B_{21} & B_{22} \end{array}\right] Q =
\left[\begin{array}{cc} \hat{G}_{11} & \hat{B}_{12} \\
\hat{B}_{21} & \hat{B}_{22} \end{array}\right].
\end{displaymath}
We can exploit symmetry in this step as well.
The matrices $\hat{E}_{11}$ and $\hat{G}_{11}$ are nonzero because $X$ does not exactly satisfy
the Riccati equations (\ref{eq:genric2}), but the quadratic convergence of Newton's method guarantees
that $\norm{\hat{E}_{11}} \ll \norm{E_{11}}$ and $\norm{\hat{G}_{11}} \ll \norm{G_{11}}$.
Thus $\hat{E}_{11}$ and $\hat{G}_{11}$ will be small enough that they can be set to zero without compromising stability. In the extremely rare
event that they are not small enough, the refinement step can be repeated.
Now we consider the refinement step in the case when we have three blocks, as in a move of type IIe.
After the move we have a pole pencil
\begin{displaymath}
\left[\begin{array}{ccc} E_{11} & E_{12} & A_{13} \\ E_{21} & A_{22} & A_{23} \\
A_{31} & A_{32} & A_{33} \end{array}\right] - \lambda
\left[\begin{array}{ccc} G_{11} & G_{12} & B_{13} \\ G_{21} & B_{22} & B_{23} \\
B_{31} & B_{32} & B_{33} \end{array}\right].
\end{displaymath}
In a move of type IIe, all of the submatrices are $1 \times 1$. Here we allow them to be of any size.
Let's say the matrices $A_{31}$, $A_{13}$, etc.\ are $m \times m$, and the matrices $A_{22}$ etc.\
are $k \times k$. If we take $k=0$, this reduces to the case that we have just covered.
The submatrices $E_{ij}$ and $G_{ij}$ would be zero except for roundoff errors. If they are small enough,
we can set them to zero and proceed. Otherwise we must do a refinement step, and to this end we seek
$X_{ij}$ and $Y_{ij}$ such that
\begin{displaymath}
\left[\begin{array}{ccc} I & Y_{12} & Y_{13} \\ & I & Y_{23} \\ & & I\end{array}\right]
\left[\begin{array}{ccc} E_{11} & E_{12} & A_{13} \\ E_{21} & A_{22} & A_{23} \\
A_{31} & A_{32} & A_{33} \end{array}\right]
\left[\begin{array}{ccc} I & & \\ X_{21} & I & \\ X_{31} & X_{32} & I \end{array}\right] =
\left[\begin{array}{ccc} & & \check{A}_{13} \\ & \check{A}_{22} & \check{A}_{23} \\
\check{A}_{31} & \check{A}_{32} & \check{A}_{33} \end{array}\right]
\end{displaymath}
and
\begin{displaymath}
\left[\begin{array}{ccc} I & Y_{12} & Y_{13} \\ & I & Y_{23} \\ & & I\end{array}\right]
\left[\begin{array}{ccc} G_{11} & G_{12} & B_{13} \\ G_{21} & B_{22} & B_{23} \\
B_{31} & B_{32} & B_{33} \end{array}\right]
\left[\begin{array}{ccc} I & & \\ X_{21} & I & \\ X_{31} & X_{32} & I \end{array}\right] =
\left[\begin{array}{ccc} & & \check{B}_{13} \\ & \check{B}_{22} & \check{B}_{23} \\
\check{B}_{31} & \check{B}_{32} & \check{B}_{33} \end{array}\right].
\end{displaymath}
This results in a system of six Riccati equations, three from $A$ and three from $B$:
\begin{displaymath}
\begin{array}{c}
\begin{array}{c}
A_{13}X_{31} + Y_{13}A_{31} + E_{11} + Y_{13}A_{33}X_{31} + Y_{12}E_{21} + E_{12}X_{21} \hspace{40pt} \\
\hspace{93pt} +Y_{12}A_{22}X_{21} + Y_{13}A_{32}X_{21} +Y_{12}A_{23}X_{31} = 0,
\end{array} \\ \\
\begin{array}{c}
B_{13}X_{31} + Y_{13}B_{31} + G_{11} + Y_{13}B_{33}X_{31} + Y_{12}G_{21} + G_{12}X_{21} \hspace{40pt} \\
\hspace{93pt} +Y_{12}B_{22}X_{21} + Y_{13}B_{32}X_{21} +Y_{12}B_{23}X_{31} = 0,
\end{array}\end{array}
\end{displaymath}
\begin{displaymath}
\begin{array}{c}
A_{13}X_{32} + Y_{12}A_{22} + E_{12} + Y_{12}A_{23}X_{32} + Y_{13}A_{32} + Y_{13}A_{33}X_{32} = 0, \\ \\
B_{13}X_{32} + Y_{12}B_{22} + G_{12} + Y_{12}B_{23}X_{32} + Y_{13}B_{32} + Y_{13}B_{33}X_{32} = 0,
\end{array}
\end{displaymath}
\begin{displaymath}
\begin{array}{c}
A_{22}X_{21} + Y_{23}A_{31} + E_{21} + Y_{23}A_{32}X_{21} + A_{23}X_{31} + Y_{23}A_{33}X_{31} = 0, \\ \\
B_{22}X_{21} + Y_{23}B_{31} + G_{21} + Y_{23}B_{32}X_{21} + B_{23}X_{31} + Y_{23}B_{33}X_{31} = 0.
\end{array}
\end{displaymath}
We do not propose to compute the exact solution of these equations. Instead we will obtain an excellent approximation
by solving the Sylvester equations that one gets by discarding all negligible terms. For example, in the first equation
above, only the first three terms are non-negligible. We obtain
\begin{equation}\label{eq:gensyl11}
\begin{array}{c}
A_{13}X_{31} + Y_{13}A_{31} + E_{11} = 0, \\ \\
B_{13}X_{31} + Y_{13}B_{31} + G_{11} = 0,
\end{array}
\end{equation}
\begin{equation}\label{eq:gensyl12}
\begin{array}{c}
A_{13}X_{32} + Y_{12}A_{22} + E_{12} + Y_{13}A_{32} = 0, \\ \\
B_{13}X_{32} + Y_{12}B_{22} + G_{12} + Y_{13}B_{32} = 0,
\end{array}
\end{equation}
\begin{equation}\label{eq:gensyl21}
\begin{array}{c}
A_{22}X_{21} + Y_{23}A_{31} + E_{21} + A_{23}X_{31} = 0, \\ \\
B_{22}X_{21} + Y_{23}B_{31} + G_{21} + B_{23}X_{31} = 0.
\end{array}
\end{equation}
Equations (\ref{eq:gensyl11}) are independent of the others. They have a unique solution $(X_{31},Y_{13})$ if and
only if the pencils $A_{13} - \lambda B_{13}$ and $A_{31} - \lambda B_{31}$ have disjoint spectra. Once we have
solved these equations, we can substitute $Y_{13}$ into (\ref{eq:gensyl12}) and $X_{31}$ into (\ref{eq:gensyl21}).
Equations (\ref{eq:gensyl12}) have a unique solution $(X_{32},Y_{12})$ if and only if the pencils
$A_{13} - \lambda B_{13}$ and $A_{22} - \lambda B_{22}$ have disjoint spectra. Similarly (\ref{eq:gensyl21}) have
a unique solution $(X_{21},Y_{23})$ if and only if
the spectra of $A_{22} - \lambda B_{22}$ and $A_{31} - \lambda B_{31}$ are disjoint.
In both the alternating and palindromic cases, one can show that $Y_{12} = X_{21}^{*}$, $Y_{13} = X_{31}^{*}$,
and $Y_{23} = X_{32}^{*}$. The routine but tedious proof is left for the reader. When we solve the Sylvester equations
in practice, we take these symmetries into account. For example, in the palindromic case, the $A$ and $B$ equations
are equivalent, so we only have to solve the $A$ equations; (\ref{eq:gensyl11}) reduces to
\begin{displaymath}
A_{13}X_{31} + X_{31}^{*}A_{31} + E_{11} = 0,
\end{displaymath}
and (\ref{eq:gensyl12}) and (\ref{eq:gensyl21}) together reduce to
\begin{displaymath}
\begin{array}{c}
A_{13}X_{32} + X_{21}^{*}A_{22} + E_{12} + X_{31}^{*}A_{32} = 0, \\ \\
A_{22}X_{21} + X_{32}^{*}A_{31} + E_{21} + A_{23}X_{31} = 0,
\end{array}
\end{displaymath}
which can be solved simultaneously.
Once we have computed $X$, we perform a decomposition
\begin{displaymath}
\left[\begin{array}{ccc} I & & \\ X_{21} & I & \\ X_{31} & X_{32} & I \end{array}\right] = QR,
\end{displaymath}
and use $Q$ to do a unitary congruence
\begin{displaymath}
Q^{*} \left[\begin{array}{ccc} E_{11} & E_{12} & A_{13} \\
E_{21} & A_{22} & A_{23} \\ A_{31} & A_{32} & A_{33}\end{array}\right] Q =
\left[\begin{array}{ccc} \hat{E}_{11} & \hat{E}_{12} & \hat{A}_{13} \\
\hat{E}_{21} & \hat{A}_{22} & \hat{A}_{23} \\ \hat{A}_{31} & \hat{A}_{32} & \hat{A}_{33}
\end{array}\right],
\end{displaymath}
and similarly for $B$. Because the correction amounts to one step of Newton's method, and the errors were
tiny to begin with, we will have $\norm{\hat{E}_{11}} \ll \norm{E_{11}}$ and so on, and we can safely set the
new errors to zero. In the highly unlikely event that the errors are still too big, we can repeat the refinement step.
\section{Numerical experiments}
We wrote and tested MATLAB code for the palindromic case. We tested the most basic pole-swapping algorithm with a
single shift, as described in Section~\ref{sec:algorithm}. We used Wilkinson shifts \cite{Wat10}.
For moves of types IIo and IIe we set a
tolerance of $10\epsilon\norm{M}_{F}$, where $\epsilon$ is the machine epsilon, which is approximately
$2.22 \times 10^{-16}$, and $M$ is the small submatrix in which the swap occurs. If all of the numbers that are
supposed to be zero are below this tolerance, no refinement is needed. Our MATLAB code is publicly available at
\textit{github.com/thijssteel/palindromic-rqz}.
For our first test we considered palindromic pencils $A - \lambda A^{*}$, where $A$ is a random anti-Hessenberg
matrix. Each of the nonzero entries is generated as $2a + bi$, where $a$ and $b$ are drawn from a standard normal
distribution. We computed the eigenvalues using our pole-swapping algorithm,
which we will call the ``new'' algorithm. We compared our results against those of the single-shift version of the
bulge-chasing algorithm of \cite{KrScWa09}, which we call the ``old'' algorithm. Our new algorithm can operate directly on the anti-Hessenberg pencil, but the old algorithm requires a preliminary reduction to the partially anti-triangular form
shown in (\ref{eq:partanttri}). This costs about $n^{2}/8$ moves. We have not compared against any standard
non-structure-preserving algorithms, as it was already shown in \cite{KrScWa09} that the structure-preserving
algorithms are significantly more efficient.
Using the ``old'' and ``new'' methods, we reduced the problem to a pencil $S - \lambda S^{*}$ with $S$ anti-triangular,
from which we can read the eigenvalues. Figure \ref{fig:pal-test-3} displays the results. The left panel shows the
backward errors, which are excellent for both the old and the new methods. The right panel shows the total number
of moves required as a function of $n$. We take this as our measure of work. We see that the two methods
have very similar performance, with the new method requiring about 5\% fewer moves. The difference is due
entirely to the fact that the new method does not require the preliminary reduction.
The matrices in this test have even dimension, so they use moves of type IIe in the middle. No refinement
steps were necessary. We repeated the experiment with the ten-times stricter tolerance $\epsilon\norm{M}_{F}$
and got nearly identical results. Refinement steps were occasionally necessary.
We repeated the experiment with matrices of odd dimension 101, 201, \ldots, 1601, which use moves of type
IIo, and got nearly identical results. Therefore we have not displayed them.
\begin{figure}[ht!]
\centering
\subfloat[]{
\begin{tikzpicture}
\begin{axis}[
xlabel={Size of matrix},
ylabel={Backward error $\frac{\|Q^*AQ - S\|}{\|A\|}$},
legend pos=north west,
scale = 0.52,
ymode=log,
xmode=log,
ymajorgrids=true,
grid style=dashed,
xtick = {100,200,400,800,1600},
xticklabels = {100,200,400,800,1600},
ytick = {2.0E-15, 4.0E-15, 7.5E-15},
]
\addplot[
color=black,
mark=*,
]
table [x=n_array, y=backward_errors_pal, col sep=comma] {pal_test_3.csv};
\addplot[
color=black,
mark=*,
dashed
]
table [x=n_array, y=backward_errors_palz, col sep=comma] {pal_test_3.csv};
\end{axis}
\end{tikzpicture}
}
\subfloat[]{
\begin{tikzpicture}
\begin{axis}[
xlabel={Size of matrix},
ylabel={Number of moves},
legend pos=north west,
scale = 0.52,
ymode=log,
xmode=log,
xmin=80,
xmax=2000,
ymajorgrids=true,
grid style=dashed,
xtick = {100,200,400,800,1600},
xticklabels = {100,200,400,800,1600},
]
\addplot[
color=black,
mark=*,
]
table [x=n_array, y=swap_count_pal_with_reduction, col sep=comma] {pal_test_3.csv};
\addplot[
color=black,
mark=*,
dashed
]
table [x=n_array, y=swap_count_palz, col sep=comma] {pal_test_3.csv};
\addplot[
color=black,
dotted,
thick
]
coordinates {
(100, 30000)
(200, 120000)
(400, 480000)
(800, 1920000)
(1600, 7680000)
};
\end{axis}
\end{tikzpicture}
}
\caption{Comparison of the old (solid lines) and new (dashed lines) methods for the randomly generated problems.
For the old algorithm, the move count includes the reduction to partially anti-triangular form. The dotted line in
the move count plot is the function $3n^{2}$, which appears to be parallel to the other two lines.
This indicates that the number of moves is of $O(n^2)$, leading to a total flop count of $O(n^3)$.}
\label{fig:pal-test-3}
\end{figure}
Our second test problem was the discretized 1D-Laplacian system of \cite[Ex.~3]{KrScWa09}.
In this scalable problem the matrices always have odd dimension, so moves of type IIo are used.
The results are shown in Figure~\ref{fig:pal-test-2}. Again we see that both methods have
similar and excellent backward errors, and the new method requires slightly fewer moves
than the old method. This is again due to the fact that the new method does not require a preliminary reduction.
In this case the difference is about 15\%. It is larger in this example than it was in the previous one because fewer
iterations are required in the iterative phase. In no cases were refinements necessary, even when the stricter tolerance
$\epsilon\norm{M}_{F}$ was enforced.
\begin{figure}[ht!]
\centering
\subfloat[]{
\begin{tikzpicture}
\begin{axis}[
xlabel={Size of matrix},
ylabel={Backward error $\frac{\|Q^*AQ - S\|}{\|A\|}$},
legend pos=north west,
scale = 0.52,
ymode=log,
xmode=log,
ymajorgrids=true,
grid style=dashed,
xtick = {101,201,401,801,1601},
xticklabels = {101,201,401,801,1601},
ytick = {3.2E-15, 6.4E-15, 12.9E-15},
]
\addplot[
color=black,
mark=*,
]
table [x=n_array, y=backward_errors_pal, col sep=comma] {pal_test_2.csv};
\addplot[
color=black,
mark=*,
dashed
]
table [x=n_array, y=backward_errors_palz, col sep=comma] {pal_test_2.csv};
\end{axis}
\end{tikzpicture}
}
\subfloat[]{
\begin{tikzpicture}
\begin{axis}[
xlabel={Size of matrix},
ylabel={Number of moves},
legend pos=north west,
scale = 0.52,
ymode=log,
xmode=log,
ymajorgrids=true,
grid style=dashed,
xtick = {101,201,401,801,1601},
xticklabels = {101,201,401,801,1601},
]
\addplot[
color=black,
mark=*,
]
table [x=n_array, y=swap_count_pal_with_reduction, col sep=comma] {pal_test_2.csv};
\addplot[
color=black,
mark=*,
dashed
]
table [x=n_array, y=swap_count_palz, col sep=comma] {pal_test_2.csv};
\addplot[
color=black,
dotted,
thick
]
coordinates {
(101, 20402)
(201, 80802)
(401, 321602)
(801, 1283202)
(1601, 5126402)
};
\end{axis}
\end{tikzpicture}
}
\caption{Comparison of the old (solid lines) and new (dashed lines) methods for the 1D-Laplace boundary control problem \cite[Ex.~2]{KrScWa09}. The dotted line in the number of moves plot is the function $2n^2$.}
\label{fig:pal-test-2}
\end{figure}
\subsection*{Tests of moves of types IIo and IIe.}
Since the examples we have considered so far provided barely any test of the refinement procedure,
we performed stress tests of the moves of types IIo and IIe under extreme conditions. Type IIo moves
were tested on matrices of the form
\begin{equation}
\begin{bmatrix}
0 & \ a\\
a(1+g) & \ c
\end{bmatrix}
\end{equation}
and type IIe moves were tested on matrices of the form
\begin{equation}
\begin{bmatrix}
0 & \ 0 & \quad a\\
0 & \ b & \quad c\\
a(1+g) & \ d & \quad e
\end{bmatrix}
\end{equation}
where $a$, $b$, $c$, $d$ and $e$ are random numbers generated as $s_110^{t_1} + s_2i10^{t_2}$, with $t_1$ and $t_2$ uniformly distributed in the interval $[-15,0]$ and $s_1$ and $s_2$ random sign bits. This gives numbers with
wildly varying magnitudes from about $10^{-15}$ up to $10^{0}$. The parameter $g$ is a positive real number
that controls the relative gap between the eigenvalues. We tested both large and small values of $g$, expecting
that we might sometimes have difficulties when $g$ was very small. Specifically, we selected four intervals for
$g$ and generated $10^{5}$ test matrices per interval, with $g$ logarithmically distributed. We recorded how often
and how many refinements are required to achieve the tolerance $10\epsilon\norm{M}$.
The results are shown in Table \ref{table:typeIItests}.
As expected, refinement steps are never needed when $g$ is large. Even for very small $g$, the average number
of refinements is quite small, indicating that refinement steps are very seldom necessary. Occasionally more than
one refinement step is needed. There were a few examples with $g$ smaller than $10^{-12}$ where 10 refinements
were required, which is an indication of failure. In the full algorithm the gap $g$ is controlled by the shifting strategy.
If shifts that are extremely close to the unit circle are avoided, such cases will never occur.
\begin{table}[]
\centering
\begin{tabular}{@{}lllll@{}}
\toprule
$g$ & IIe (average) & IIe (max) & IIo (average) & IIo (max) \\ \midrule
$[10^{-15}, 10^{-12}]$ & 0.00502 & 10 & 0.08699 & 10 \\
$[10^{-12}, 10^{-9}]$ & 0.01004 & 3 & 0.089 & 3 \\
$[10^{-9}, 10^{0}]$ & 0.01413 & 2 & 0.06537 & 2 \\
$[10^{0}, 10^{15}]$ & 0 & 0 & 0 & 0 \\ \bottomrule
\end{tabular}
\caption{The number of refinement steps required to reach the selected tolerance. The columns denoted `average' indicate the average number of refinement steps and the columns denoted `max' indicate the maximum number of refinement steps that was required. Note that because our implementation limits the number of refinement steps to 10, moves where that amount of steps was required are not necessarily accurate.}
\label{table:typeIItests}
\end{table}
\section{Conclusions}
We have shown that the concept of pole-swapping algorithms, which is a generalization of bulge-chasing algorithms
for the generalized eigenvalue problem, can be extended to palindromic and alternating eigenvalue problems, which
arise in control theory. We have also introduced a refinement step to make the algorithms (including the algorithms
in \cite{KrScWa09}) more robust. Numerical tests of the palindromic case indicate that our pole-swapping algorithm
works well and is slightly faster than the corresponding bulge-chasing algorithm.
|
\section{Introduction and main result}
Consider a smooth closed curve $\Gamma$ in $\mathbb R^{2+n}$. The existence of oriented surfaces which bound $\Gamma$ and minimize the area can be approached in two different ways. Following the classical work of Douglas and Rado we can fix an abstract connected smooth surface $\Sigma_\genus$ of genus $\genus$ whose boundary $\partial \Sigma_\genus$ consists of a single connected component and look at smooth maps $\Phi: \Sigma_g \to \mathbb R^{2+n}$ with the property that the restriction of $\Phi$ to $\partial \Sigma_\genus$ is an homeomorphism onto $\Gamma$. We then consider the infimum $A_\genus (\Gamma)$ over all such maps $\Phi$ and all smooth Riemannian metrics $h$ on $\Sigma$ of the energy
\[
\int_{\Sigma_\genus} |\nabla \Phi|^2{\rm dvol}_h \, .
\]
If $A_\genus (\Gamma) < A_{\genus -1} (\Gamma)$, then there is a minimizer, cf. \cite{Douglas,Courant40}, whose image is an immersed surface of genus $\genus$, with possible branch points.
A different, more intrisic, approach was pioneered by De Giorgi, cf. \cite{DG}, in the codimension $1$ case, and by Federer and Fleming in higher codimension, cf. \cite{FF}. The latter looks at integral currents $T$ (a suitable measure-theoretic generalization of classical oriented submanifolds with boundary) whose boundary is given by $\a{\Gamma}$ and minimizes their mass, a suitable measure-theoretic generalization of the volume of classical submanifolds. The minimizer then always exists via the direct methods of the calculus of variations.
There is a very natural question relating the two approaches: is every minimizer $T$ found by the Federer-Fleming theory a classical minimal surface with finite topology, namely a parametrized surface of some genus $\genus$? Note that if this were the case, then the sequence $\{A_\genus (\Gamma)\}_{\genus \in \mathbb N}$ would become constant for sufficiently large $\genus$.
When the codimension $n$ equals $1$ and $\Gamma$ is of class $C^{2,\alpha}$ for some $\alpha >0$, the interior regularity theorem of De Giorgi in \cite{De-GiorgiColombiniPiccinini72} and the boundary regularity theorem of Hardt and Simon in \cite{HS} imply that every minimizer $T$ is in fact a $C^{2,\alpha}$ embedded surface up to the boundary. Thus $T$ has finite genus $\genus_0$ and any conformal parametrization $\Phi$ over an abstract Riemann surface $\Sigma_{\genus_0}$ gives a minimizer in the sense of Douglas and Rado. On the other hand, Fleming in \cite{Fleming} showed a closed embedded curve $\Gamma$ in $\mathbb R^3$ of finite length for which $\{A_\genus (\Gamma)\}_{\genus \in \mathbb N}$ is not asymptotically constant.
The question is much more subtle in higher codimension, because singularities might arise, both at the interior and at the boundary.
In the work \cite{White97} White asks whether the topology of the minimizer $T$ is finite when $\Gamma$ is real analytic. If this conjecture were true, then $T$ would have finitely many singularities by the main theorem of \cite{White97}. The aim of this paper is to start a sort of reverse program to White's: under the assumption of real analyticity for the boundary $\Gamma$ we wish to show first that the set of boundary and interior singular points of $T$ is finite and hence to analyze the singularities and conclude that the topology of the minimizer is finite.
It has been shown by Chang in \cite{Chang} that $T$ is smooth in $\R^n\setminus \Gamma$ up to a discrete set of singular branch points and in sufficiently small neighborhoods of such singular points the resulting branched surface is topologically a disk. We in fact refer to \cite{DSS1, DSS2, DSS3, DSS4} for a complete proof, as Chang needs a suitable modification of the techniques of Almgren's monumental monograph \cite{Alm} to start his argument, and the former has been given in full details in \cite{DSS3}. In order to attack White's conjecture it suffices therefore to deal with boundary regularity. In fact, even for $\Gamma$ of class $C^{2,\alpha}$, under the assumption that $\Gamma$ lies in the boundary of a uniformly convex set, the boundary regularity theorem of Allard \cite{AllB} implies that any minimizer $T$ is smooth at $\Gamma$; the general problem is however very subtle. So far the best available result is given in \cite{DDHM} and shows that the set of boundary regular points is dense in $\Gamma$ when $\Gamma$ is of class $C^{3,\alpha}$ for $\alpha >0$. The work \cite{DDHM} gives also an example of a smooth curve in $\R^4$ for which there is a unique Federer-Fleming minimizer with a sequence of singularities accumulating to a boundary branch point. This example has been modified in \cite{DDH} to produce $C^\infty$ embedded curves in complete $C^\infty$ Riemannian $4$-dimensional manifolds for which there is a unique Federer-Fleming minimizer with infinite topology. In particular there is a strong contrast to the codimension $1$ case: the real analyticity assumption in White's conjecture is, in a certain sense, needed\footnote{The examples of \cite{DDH} are curves in smooth almost K\"ahler manifolds $(\mathbb R^4, g)$, whose smooth metrics can be taken arbitrarily close to the euclidean one. However it is currently not known whether such examples exist in the Euclidean space.}.
\subsection{Linearized model} The analysis of interior singularities of area minimizing currents was pioneered by Almgren's monumental work in \cite{Alm} in the early eighties and recently revisited from a modern perspective by the first author and Emanuele Spadaro in \cite{DS}. The work \cite{DDHM} gives an Almgren type theory at the boundary, whereas the works \cite{DSS1,DSS2,DSS3,DSS4,Spolaor,DHMS1,DHMS2} extend the interior theory to other objects (almost calibrated currents and area minimizing currents modulo $p$). The starting point of all these papers, an essential discovery of Almgren, is to analyze the singularities for a suitable ``linearized model''. The main purpose of the present paper is to state and prove the appropriate linearized counterpart of White's conjecture.
First of all we recall the notation $\AQ (\R^n)$ for the set of unordered $Q$-tuples of $\R^n$, which we will regard as nonnegative atomic measures with integer coefficients and total mass $Q$, cf. \cite[Introduction]{DS} for the formal definition and for the standard complete metric $\mathcal{G}$ which we will use on it. For atoms we will use the notation $\a{P}$ and thus elements in $\AQ (\R^n)$ will be denoted by $\sum_i \a{P_i}$. In what follows we will often write $\AQ$ instead of $\AQ (\R^n)$.
We recall that for Sobolev functions $f\in W^{1,2}(\Omega, \AQ)$ (cf. again \cite[Introduction]{DS}) we set
\[ |Df|^2 := \sum_{j=1}^m |\partial_j f|^2, \]
where
\begin{equation}\label{eq:repD}
|\partial_j f| = \sup_{i\in \NN}| \partial_j \GG(f, T_i)| \quad \text{almost everywhere in } \Omega,
\end{equation}
and $\{T_i\}_{i\in \NN}$ is a countable dense subset of $\AQ$. While such abstract definition is very direct and useful to work with, the Dirichlet energy turns out to be the sum of the Dirichlet energies of the different sheets in all cases where the multifunction $f$ can be ``nicely decomposed''. In an appropriate sense this can be justified also for any Sobolev functions, the reader is again referred to \cite{DS} for the relevant details.
We now recall the notion of interior regular points.
\begin{defn}[Interior regular point, Definition 0.10 of \cite{DS}]\label{def:intsing}
A function $f\in W^{1,2}(\Omega, \AQ)$ is regular at a point $x\in \Omega$ if there exists a neighborhood $B$ of $x$ and $Q$ analytic functions $f_i: B\to \RR^n$ such that
\[ f(y) = \sum_i \llbracket f_i(y) \rrbracket \quad \text{ for almost every } y\in B, \]
and either $f_i(y) \neq f_j(y)$ for every $y\in B$, or $f_i \equiv f_j$.
The complement of interior regular points is called the set of interior singular points.
\end{defn}
The following theorem on the interior regularity of Dir-minimizers was proven in \cite{DS}, refining a previous fundamental result by Almgren in \cite{Alm}:
\begin{theorem}[Theorem 0.12 in \cite{DS}]\label{thm:intsing}
Let $f \in W^{1,2}(\Omega,\AQ)$ be Dir-minimizing and $m=2$. Then the interior singular set of $f$ consists of isolated points.
\end{theorem}
We now come to the boundary counterpart, following the approach of \cite{DDHM}.
Suppose a hypersurface $\gamma$ divides a connected open set $\Omega \subset \RR^m$ into two connected components $\Omega^+$ and $\Omega^-$. For any set $K\subset \Omega$ we will use the notation $K^\pm$ for $K \cap \Omega^\pm$. Moreover, in order to avoid confusion, in the rest of the paper we will use the double integral symbol to indicate integration over subsets of $\RR^m$ with respect to the Lebesgue measure, and the single integral symbol to indicate integration over subsets of the hypersurface $\gamma$ with respect to the usual Hausdorff
$(m-1)$-dimensional measure.
\begin{defn}
We say that the pair $f=(f^+, f^-)$ is a $\left(Q-\frac{1}{2}\right)$-map with interface $(\gamma, \varphi)$ of class $W^{1,2}$ if there is some (classical) function $\varphi\in H^{1/2} (\gamma, \mathbb R^n)$ such that
\begin{enumerate}[(i)]
\item $f^+ \in W^{1,2}(\Omega^+, \AQ)$ and $f^-\in W^{1,2}(\Omega^-, \mathcal{A}_{Q-1} )$;
\item $f^+|_{\gamma} = f^-|_{\gamma} + \llbracket \varphi \rrbracket$.
\end{enumerate}
We refer to \cite{DS,DDHM} for the trace theorems which allow to make sense of (ii) under our assumptions. For the corresponding set of pairs we will use the shorthand notation $W^{1,2} (\Omega, \AQt)$ and for each $f = (f^+, f^-) \in W^{1,2} (\Omega, \AQt)$ we define its Dirichlet energy as
\[ \Dir(f, \Omega) := \Dir(f^+, \Omega^+) + \Dir(f^-, \Omega^-) = \iint_{\Omega^+ }|Df^+|^2 + \iint_{\Omega^-} |Df^-|^2. \]
Finall, we say that $f=(f^+, f^-) \in W^{1,2}(\Omega, \AQt)$ is \textit{Dir-minimizing with interface $(\gamma,\varphi)$}, if $\Dir (g, \Omega) \geq \Dir (f, \Omega)$ for any other function $g \in W^{1,2}(\Omega, \AQt)$ with interface $(\gamma, \varphi)$ which agrees with $f$ outside of a compact set $K \subset \Omega$.
\end{defn}
The goal of the paper is to show that when the interface $(\gamma, \varphi)$ is real analytic and the domain is $2$-dimensional, Dir-minimizers enjoy a regularity theorem which is analogous to Theorem \ref{thm:intsing}. First of all a point $p\in \Omega\setminus \gamma$, namely belonging to either $\Omega^+$ or $\Omega^-$, will be called regular if it is a regular point for, respectively,$f^+$ or $f^-$ (cf. Definition \ref{def:intsing}). Its complement in $\Omega\setminus \gamma$ is the set of interior singular points, denoted by $\Sigma_f^i$. It remains to define regular points at the interface $\gamma$.
\begin{defn}[Boundary regular point, Definition 2.6 of \cite{DDHM}]
Let $f=(f^+, f^-)$ be a map in $W^{1,2}(\Omega, \AQt)$ with interface $(\gamma, \varphi)$. A point $p\in \gamma$ is regular if there are a ball $B_r(p)$, $(Q-1)$-functions $u_1, \cdots, u_{Q-1}: B_r(p) \to \RR^n$ and a function $u_Q: B_r^+(p) \to \RR^n$ such that
\begin{itemize}
\item $f^+ = \sum_{i=1}^Q \llbracket u_i \rrbracket$ on $B_r^+(p)$ and $f^- = \sum_{i=1}^{Q-1} \llbracket u_i \rrbracket$ on $B_r^-(p)$;
\item For any pair $i, j \in \{1, \cdots, Q-1\}$ either the graphs of $u_i$ and $u_j$ are disjoint or they completely coincide;
\item For any $i\in \{1, \cdots, Q-1\}$ either the graphs of $u_i$ and $u_Q$ are disjoint in $B_r^+(p)$ or the graph of $u_Q$ is contained in that of $u_i$.
\end{itemize}
The complement in $\gamma$ of the set of regular points is called the set of boundary singular points, denoted by $\Sigma_f^b$.
\end{defn}
We can now state our main theorem:
\begin{theorem}\label{thm:main}
Let $\Omega \subset \mathbb R^2$ and $(\gamma, \varphi)$ be an interface for which both $\gamma$ and $\varphi$ are real analytic. If $f\in W^{1,2} (\Omega, \AQt)$ is Dir-minimizing with interface $(\gamma, \varphi)$, then the singular set $\Sigma_f = \Sigma_f^i \cup \Sigma_f^b$ is discrete.
\end{theorem}
In passing, we need a suitable estimate on the H\"older continuity of minimizers at the interface $\gamma$. The latter result is however not confined to the special dimension $m=2$ nor to real analytic interfaces $(\gamma, \varphi)$ and, although it is not immediately relevant for our main purposes, we state it in a more general case in the following
\begin{theorem}\label{thm:main2}
Let $m\in \mathbb N \setminus \{0,1\}$ and suppose $(f^+, f^-)$ is a Dir-minimizing $(Q-\frac{1}{2})$-map in $\Omega\subset \R^m$ with interface $(\gamma, \varphi)$ of class $C^1$. Then $(f^+, f^-)$ is H\"older regular.
\end{theorem}
In fact it is possible to give a precise estimate on a suitable H\"older seminorm of $f^{\pm}$ in terms of the regularity of the interface $(\gamma, \varphi)$ and the Dirichlet energy of the minimizer. For the precise statement we refer to Theorem \ref{thm:Holder}.
\subsection{Plan of the paper} The remaining sections are organized as follows. First of all in Section \ref{s:preliminaries} we make some preliminary elementary considerations on planar minimizers which will be particularly useful in the planar case of Theorem \ref{thm:main2} and in Theorem \ref{thm:main}. In Section \ref{s:hoelder} we address the general H\"older regularity result and prove therefore Theorem \ref{thm:main2}. In the subsequent Section \ref{s:frequency} we give the fundamental computations leading to the monotonicity of the frequency function, a celebrated result of Almgren away from interface, extended at general interfaces in \cite{DDHM}: in our case the computations are simpler than in \cite{DDHM} because we can ``straighten the boudary'' using complex analysis. In Section \ref{s:rate} we use the frequency function estimate and the H\"older regularity to prove the existence of suitable blow-ups, or tangent functions, at singular points. A suitable modification of the argument given in \cite{DS} (which in turn borrowed from key ideas in \cite{Chang}) shows then the uniqueness of such objects. In Section \ref{s:classification} we give a list of necessary conditions that tangent functions must satisfy, which in turn leads to a suitable decomposition of them in simpler pieces (which we call irreducible maps). Such decomposition is combined together with the rate of convergence proven in Section \ref{s:rate} in order to decompose general Dir-minimizers at boundary singular points: the latter fact is then used in the final Section \ref{s:final} to conclude the proof of Theorem \ref{thm:main}.
\section{Reduction and preliminaries for the planar case}\label{s:preliminaries}
\subsection{Reduction of Theorem \ref{thm:main}}\label{s:mod_out_average} In this section we use elementary considerations in complex analysis to reduce Theorem \ref{thm:main} to a much simpler case. In order to state our theorem, we recall the definition of the map $\bdeta: \AQ (\R^n)\to \R^n$ which gives the barycenter of the atomic measure $T$:
\[
\bdeta \left(\sum_{i=1}^Q \a{P_i}\right) = \frac{1}{Q} \sum_{i=1}^Q P_i\, .
\]
In particular, if $(f^+, f^-)\in W^{1,2} (\Omega, \AQt)$ we can define two maps $(\eta^+, \eta^-)$ which are, respectively, the center of mass of the maps $f^+$ and $f^-$. In particular $\eta^\pm := \bdeta \circ f^\pm$,
where we make a slight abuse of notation because we keep the same symbol $\bdeta$ for two different maps, one defined on $\AQ$ and the other on $\mathcal{A}_{Q-1}$. Specifically:
\[
\eta^+ (x) = \frac{1}{Q} \sum_{i=1}^Q f^+_i (x)\qquad\mbox{and}\qquad
\eta^- (x) = \frac{1}{Q-1} \sum_{i=1}^{Q-1} f^-_i (x)\, .
\]
Theorem \ref{thm:main} can then be reduced to the following particular case:
\begin{theorem}\label{thm:main_simple}
Let $m=2$ and assume $(f^+,f^-)$ is Dir-minimizing in the unit disk $\mathbb D$ with interface $(\gamma, 0)$, where $\gamma$ is the coordinate axis $\{(x_1,0):x_1\in \mathbb R\}$. Assume further that $Q \eta^+ = (Q-1) \eta^-$. Then the singular set $\Sigma_f$ is discrete.
\end{theorem}
From now on, we introduce the convention that, if $\gamma = \{(x_1, 0): x_1\in \mathbb R\}$, then the interface $(\gamma, \varphi)$ is denoted by $(\RR, \varphi)$. This is motivated by the fact that we will often identify $\RR^2$ with the complex plane $\mathbb C$, via $(x_1, x_2) \mapsto x_1+ix_2$. The set $\{x_2=0\}$ is then the real axis of $\mathbb C$ after such identification.
The above theorem will be proved at the end of the paper. In the next paragraph we show how the general case of Theorem \ref{thm:main} follows from it.
Assume $(f^+,f^-)$ is as in Theorem \ref{thm:main}. First of all observe that, if $\Sigma_f$ is not discrete, then by Theorem \ref{thm:intsing} $\Sigma_f$ must have an accumulation point $p\in \gamma$.
Modulo translation we may assume $p$ is the origin. Since $\gamma$ is analytic, we may choose a coordinate system so that the tangent to $\gamma$ satisfies $T_0 \gamma = \{ x_2 = 0\} = \RR$. In particular $\gamma$ must be (locally) the graph $\{(t, \zeta (t))\}$ of a function $\zeta (t)$ whose Taylor series at the origin is $\sum_{k\geq 2} \alpha_k t^k$ (where $\alpha_k = \frac{\zeta^{(k)} (0)}{k!} \in \RR$). Identify $\mathbb R^2$ with the complex plane and consider, in a neighborhood of the origin, the holomorphic map $\Phi$ given by $\Phi(z) = z+ \sum_{k\geq 2} i\alpha_k z^k$. By the inverse function theorem the latter map is invertible in a sufficiently small neighborhood $U$ of the origin (which can be assumed to be a disk) and its inverse over $\Phi (U)$ is also holomorphic. Since $\Phi$ is conformal, $(f^+\circ \Phi^{-1}, f^-\circ \Phi^{-1})$ is clearly a minimizer in
$V := \Phi (U)$ and the interface is $(T_0 \gamma, \varphi \circ \Phi^{-1})$. Moreover $\Phi$ maps the segment $\{\Imag z = 0\}\cap U$ onto $\gamma$.
We can thus assume, without loss of generality, that $\gamma = \mathbb R$.
Next, since $\varphi$ is real analytic, by the Cauchy-Kowalevski Theorem $\varphi$ has a harmonic extension in a neighborhood of the origin, still denoted by $\varphi$. We then replace $f=(f^+, f^-)$ with
\[ f^+(x) \mapsto g^+ (x) := \sum_{i=1}^Q \llbracket f_i^+(x) - \varphi(x) \rrbracket , \quad f^-(x) \mapsto g^- (x) := \sum_{i=1}^{Q-1} \llbracket f_i^-(x) - \varphi(x) \rrbracket. \]
Indeed, given a map $(\bar{g}^+, \bar{g}^-)$ with interface $(\RR, 0)$ and same trace on $\partial \mathbb D$ as $(\bar g^+, \bar g^-)$, consider the corresponding map $(h^+, h^-)$ where we add $\varphi$ on each side. The latter has interface $(\RR, \varphi)$ and coincides with $(f^+, f^-)$ on $\partial \mathbb D$. Moreover we compute
\begin{align*}
\iint_{\mathbb D^+} |Dh^+|^2 &= \iint_{\mathbb D^+} |D \bar g^+|^2 +Q \iint_{\mathbb D^+} |D \varphi|^2 + 2 \underbrace{Q \iint_{\mathbb D^+} D\bdeta \circ \bar g^+ : D \varphi}_{=: I^+}\\
\, \\
\iint_{\mathbb D^-} |Dh^-|^2 &= \iint_{\mathbb D^-} |D \bar g^-|^2 +(Q-1) \iint_{\mathbb D^-} |D \varphi|^2 + 2\underbrace{(Q-1) \iint_{\mathbb D^-} D\bdeta\circ \bar g^- : D \varphi}_{=:I^-}
\end{align*}
Using that the function $\varphi$ is harmonic we compute
\begin{align*}
I^+
&= \underbrace{Q \int_{(\partial \mathbb D)^+} \bdeta \circ \bar g^+ \cdot \frac{\partial \varphi}{\partial \nu}}_{=:J^+}
- \underbrace{Q \int_{\RR\cap \mathbb D} \bdeta \circ \bar g^+ \cdot \frac{\partial\varphi}{\partial x_2}}_{=:K^+}\\
I^- &= \underbrace{(Q-1) \int_{(\partial D)^-} \bdeta \circ \bar g^- \cdot \frac{\partial \varphi}{\partial \nu}}_{=:J^-} + \underbrace{(Q-1) \int_{\RR\cap \mathbb D} \bdeta\circ \bar g^- \cdot \frac{\partial \varphi}{\partial x_2}}_{=:K^-}
\end{align*}
Observe that $J^+$ and $J^-$ are both independent of the choice of $(\bar g^+, \bar g^-)$, because the traces of the respective maps on $(\partial \mathbb D)^\pm$ equals those of $(g^+, g^-)$. On the other hand $Q \bdeta \circ \bar g^+ - (Q-1) \bdeta \circ \bar g^-= 0$ on $\RR\cap \DD$. Therefore $K^--K^+ =0$. This implies that the difference
\[
\iint_{\mathbb D^+} |Dh^+|^2 + \iint_{\mathbb D^-} |Dh^-|^2 - \iint_{\mathbb D^+} |D \bar g^+|^2 - \iint_{\mathbb D^-} |D\bar g^-|^2
\]
is actually a constant. In particular, if we could find a competitor for $(g^+, g^-)$ with lower energy, then we could transform it into a competitor for $(f^+, f^-)$ with lower energy: we conclude that $(g^+, g^-)$ must be a Dir minimizer with interface $(\RR, 0)$.
Observe next that $\eta^+ = \bdeta \circ f^+$ and $\eta^- = \bdeta \circ f^-$ are harmonic functions in $\DD^+$ and $\DD^-$, respectively. For any $x=(x_1, x_2)\in \RR^2$, we denote $\bar x = (x_1, -x_2)$ the reflection point of $x$ across $\RR$. We define a function $\phi: \mathbb D \to \RR^n$ as
\begin{equation}
\phi(x) = \left\{
\arraycolsep=1.4pt\def\arraystretch{2.2}
\begin{array}{ll}
\dfrac{ Q\eta^+(x) - (Q-1) \eta^-(\bar x) }{ 2Q-1}, & \quad x_2 \geq 0, \\
\dfrac{(Q-1) \eta^-(x) - Q \eta^+(\bar x)}{2Q-1}, & \quad x_2 \leq 0.
\end{array} \right.
\end{equation}
Clearly $\phi$ is harmonic in $\mathbb D \setminus \RR$.
By the boundary condition $f^+|_{\gamma} = f^-|_{\gamma} + \llbracket 0 \rrbracket$, we know
\[ Q\eta^+ = (Q-1) \eta^- \quad \text{ on } \RR. \]
Hence $\phi$ is continuous and odd in the variable $x_2$. In particular $\phi$ is harmonic on all of $\mathbb D$. Therefore by modifying $(f^+, f^-)$ as follows
\[ f^+(x) \mapsto \widetilde f^+(x):= \sum_{i=1}^Q \llbracket f_i^+(x) - \phi(x) \rrbracket, \quad x\in \RR^m_+, \]
\[ f^-(x) \mapsto \widetilde f^-(x):= \sum_{i=1}^{Q-1} \llbracket f_i^-(x) - \phi(x) \rrbracket, \quad x\in \RR^m_- \]
and repeating the same computations as above we conclude that
the new function $(\widetilde f^+, \widetilde f^-)$ is still a Dir-minimizer with the same interface $(\RR, 0)$. Notice also that
\[
\sum_{i=1}^Q \widetilde f^+_i(x) = Q \eta^+(x) - Q \phi(x) = \frac{Q(Q-1)}{2Q-1} \left( \eta^+(x) + \eta^-(\bar x) \right),
\]
\[
\sum_{i=1}^{Q-1} \widetilde f^-_i(x) = (Q-1) \eta^-(x) - (Q-1) \phi(x) = \frac{Q(Q-1)}{2Q-1} \left( \eta^-(x) + \eta^+(\bar x) \right),
\]
and thus
\[
\sum_{j=1}^Q \widetilde f^+_j(x) = \sum_{j=1}^{Q-1} \widetilde f^-_j(\bar x).
\]
For simplicity we still denote the new function as $(f^+, f^-)$, except that its center of mass $(\eta^+, \eta^-)$ now enjoys an additional symmetry:
\begin{equation}\label{eq:avgsym}
Q \eta^+(x) = (Q-1) \eta^-(\bar x).
\end{equation}
This symmetry is invariant under translation, scaling and uniform limit.
\subsection{Decomposition into irreducible maps} In this section we extend a suitable decomposition of $Q$-valued maps on the circle to the case of $(Q-\frac{1}{2})$-valued maps.
Recall that map $g\in W^{1,p}(\So, \AQ)$ is called \textit{irreducible} if there is no decomposition of $g$ into two simpler $W^{1,p}$ functions (cf. \cite{DS}), namely if there are no integers $Q_1, Q_2>0$ and maps $g_1\in W^{1,p} (\So, \mathcal{A}_{Q_1}), g_2 \in W^{1,p} (\So, \mathcal{A}_{Q_2})$ such that $g = g_1+g_2$ (in particular $Q_1+Q_2=Q$).
\begin{defn}[Irreducible $(Q-\frac{1}{2})$-maps on $\mathbb S^1$]
A map $g = (g^+, g^-)\in W^{1,p}(\So, \AQt)$ with interface $(\RR, \varphi)$ is called \textit{irreducible} if there is no decomposition of $g$ into the ``sum'' of a map $g_1 \in W^{1,p}(\So,\mathcal{A}_{Q_1})$ and a map $g_2 \in W^{1,p}(\So, \mathcal{A}_{Q_2}^{\pm})$ with the same interface $(\gamma,\varphi)$, where the positive integers $Q_1, Q_2$ satisfy $Q_1 + Q_2 = Q$.
The ``sum'' is understood in the following sense:
\begin{align*}
&g^+ = g_1 + g_2^+\qquad \mbox{on $(\mathbb S^1)^+ = \{z\in \mathbb C: |z|=1, \text{Re}\, z >0\}$}\\
&g^- = g_1 + g_2^-\qquad \mbox{on $(\mathbb S^1)^- = \{z\in \mathbb C: |z|=1, \text{Re}\, z <0\}$.}
\end{align*}
\end{defn}
\begin{remark}
By the above definition, clearly any function $g\in W^{1,p}([0,\pi], \RR^n)$ satisfying $g(0) = g(\pi) = 0$ is irreducible with $Q=1$ (the interface being $(\RR, 0)$).
\end{remark}
The decomposition of $W^{1,p}$ $(Q-\frac{1}{2})$-valued map on the circle is then a corollary of the following proposition for $Q$-valued maps, where, for any interval $I=[a,b]\subset \RR$, we denote by $\AC(I,\AQ)$ the space of absolutely continuous functions taking values in the metric space $(\AQ, \mathcal{G})$.
\begin{prop}[Proposition 1.2 of \cite{DS}]\label{prop:oddecomp}
Let $g\in W^{1,p}(I, \AQ)$. Then
\begin{enumerate}[(a)]
\item $g\in \AC(I, \AQ)$ and moreover, $g\in C^{0,1-\frac1p}(I,\AQ)$ for $p>1$;
\item There are $g_1, \cdots, g_Q\in W^{1,p}(I, \RR^n)$ s.t. $f =\sum_i \a{g_i}$ and $|Dg_i| \leq |Dg|$ a.e.
\end{enumerate}
\end{prop}
\begin{prop}[Decomposition of $W^{1,p}(\mathbb{S}^1, \mathcal{A}_Q^{\pm})$]\label{prop:decomp}
A map $g\in W^{1,p}(\So, \AQt)$ with interface $(\RR,\varphi)$ is either irreducible, or it can be decomposed as $g= g_0 + \sum_{j=1}^J g_j$, where $g_0\in W^{1,p}(\So, \mathcal{A}_{Q_0}^{\pm}) $ is irreducible with interface $(\RR, \varphi)$, and each $g_j \in W^{1,p}(\So, \mathcal{A}_{Q_j})$ is irreducible. Moreover, a map $g\in W^{1,2}(\So, \AQt)$ with interface $(\RR, \varphi)$ is irreducible if and only if the following two conditions are satisfied:
\begin{enumerate}[(i)]
\item $\card(g^+(\theta)) = Q$ for every $\theta\in [0,\pi]$, and $\card(g^-(\theta)) = Q-1$ for every $\theta \in [\pi,2\pi]$.
\item There exists a $W^{1,p}$ map $\zeta:\So \to \RR^n$ with $\zeta(0) = \varphi (1)$ and $\zeta(2\pi) = \varphi (-1)$ such that $g$ unwinds to $\zeta$, in the following sense: $g^+ = \sum_{j=1}^Q \llbracket g_j^+ \rrbracket$ and $g^-= \sum_{j=1}^{Q-1} \llbracket g_j^- \rrbracket$ with
\begin{equation}\label{eq:unwind1}
g_j^+(\theta) = \zeta\left(\frac{2\theta}{2Q-1} + \frac{4\pi}{2Q-1} (j-1) \right), \quad \theta\in [0,\pi], j=1,\cdots,Q,
\end{equation}
\begin{equation}\label{eq:unwind2}
g_j^-(\theta) = \zeta\left(\frac{2\theta}{2Q-1} + \frac{4\pi}{2Q-1} (j-1) \right), \quad \theta\in[\pi,2\pi], j=1,\cdots,Q-1.
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof}
The existence of an irreducible decomposition in the above sense is an obvious consequence of the definition of irreducible maps. It remains to show the characterization of irreducible maps.
By Proposition \ref{prop:oddecomp} a map satisfying (i) and (ii) is clearly irreducible with interface $(\gamma, \varphi)$. Suppose $g\in W^{1,p}(\So, \AQt)$ with interface $(\gamma, \varphi)$ is irreducible. Without loss of generality (i.e. after possible subtracting to all sheets an extension of $\varphi$) we can assume $\varphi \equiv 0$. Namely
\begin{equation}\label{eq:bcirred}
g^+|_{\gamma} = g^-|_{\gamma} + \llbracket 0 \rrbracket.
\end{equation}
Recalling Proposition \ref{prop:oddecomp}, we consider a selection $g^+_1, \cdots, g^+_Q \in W^{1,p}([0,\pi], \RR^n)$ of the map $g^+\in W^{1,p}([0,\pi], \AQ)$, and a selection $g^-_1, \cdots, g^-_{Q-1} \in W^{1,p}([\pi, 2\pi], \RR^n)$ of $g^- \in W^{1,p}([\pi,2\pi], \mathcal{A}_{Q-1})$. We assume without loss of generality that $g_1^+(0) = \llbracket 0 \rrbracket$. By the boundary condition \eqref{eq:bcirred}, there exists an integer $Q_0$, $1\leq Q_0 \leq Q$, such that after reordering the selections $g^+_i(\pi) = g^-_{i}(\pi) \neq 0$ and $g_i^-(2\pi) = g_{i+1}^+(0) $ for all $i = 1, \cdots, Q_0-1$, $g^+_{Q_0}(\pi) = 0$. Suppose $Q_0 < Q$, then we define
\[ f_1^+ = \sum_{i=1}^{Q_0} \llbracket g_i^+ \rrbracket, \quad f_1^- = \sum_{i=1}^{Q_0-1} \llbracket g_{i}^- \rrbracket, \]
and
\[ f_2 = \left\{ \begin{array}{ll} \sum\limits_{i=Q_0 +1 }^{Q} \llbracket g_i^+ \rrbracket, & \theta\in [0,\pi] \\
\sum\limits_{i=Q_0}^{Q-1} \llbracket g_i^- \rrbracket, & \theta\in [\pi,2\pi]
\end{array}\right.
\]
By \eqref{eq:bcirred}, the map $f_1:= (f_1^+, f_1^-)$ lies in $ W^{1,p}(\So, \mathcal{A}_{Q_0}^{\pm})$ with interface $(\gamma, \varphi)$; the map $f_2$ is well-defined on $\gamma$, i.e. $f_2(\pi-) = f_2 (\pi+)$ and $f_2(2\pi) = f_2(0)$, and moreover $f_2 \in W^{1,p}(\So, \mathcal{A}_{Q - Q_0})$. In other words, this gives a nontrivial decomposition of the irreducible map $g$, contradiction. Hence $Q_0 = Q$, and we define the function $\zeta$ by following $g_i^+, g_i^-$, $g_{i+1}^+$ in order.
Suppose $\card(g^+) \neq Q$, that is, there exist $\theta_0 \in [0,\pi]$ and $i_1<i_2$ such that $g^+_{i_1}(\theta_0) = g^+_{i_2}(\theta_0)$. Let
\[ \widetilde g^+ = \left\{\begin{array}{ll}
g_{i_1}^+, & \theta \in [0,\theta_0] \\
g_{i_2}^+, & \theta \in [\theta_0, \pi].
\end{array} \right. \]
Then the following map gives a decomposition of $g$:
\[ f_1^+ = \sum_{i=1}^{i_1-1} \llbracket g_i^+ \rrbracket + \llbracket \widetilde g \rrbracket + \sum_{i=i_2+1}^{Q} \llbracket g_i^+ \rrbracket, \quad f_1^- = \sum_{i=1}^{i_1-1} \llbracket g_i^- \rrbracket + \sum_{i=i_2}^{Q-1} \llbracket g_i^- \rrbracket. \]
Since $(f_1^+, f_1^-) \in W^{1,p}(\So, \mathcal{A}_{Q+i_1-i_2}^\pm)$ with interface $(\gamma, \varphi)$, and $Q+i_1-i_2 < Q$, this is a nontrivial decomposition of the irreducible map $g$, contradiction. Hence $\card(g^+) = Q$. Similarly $\card(g^-) = Q-1$.
\end{proof}
\subsection{Rolling and unrolling}
The decomposition of the previous section can be used to construct efficient competitors to Dirichlet minimizers in the planar case. Again the situation is similar to that of $Q$-valued maps. Keeping our identification $\mathbb R^2 = \mathbb C$ we will denote by $[0,1]$ the ``slit'' $\{(x_1, 0): 0\leq x_1 \leq 1\}$ and on the domain $\DD\setminus [0,1]$ we will consider polar coordinates $(r, \theta) \in ]0,1[\times ]0, 2\pi[$, via the usual parametrization $(r, \theta)\mapsto r e^{i\theta}$. Given a map $\zeta\in W^{1,2} (\DD\setminus [0,1], \RR^n)$ we can define two maps $\zeta^u, \zeta^l \in H^{1/2} ([0,1], \RR^n)$ which are, respectively, the ``upper'' and ``lower'' traces of $\zeta$ on the slit $[0,1]$. In particular in polar coordinates we can naturally extend $\zeta$ to $]0,1[\times [0, 2\pi]$ setting $\zeta (r, 0) = \zeta^u (r)$ and to $\zeta (r, 2\pi) = \zeta^l (r)$. In the next lemma and its applications we will follow the latter convention.
\begin{lemma}[Unrolling, analogue of Lemma 3.12 in \cite{DS}]\label{lm:unrolling}
Suppose $\zeta\in W^{1,2}(\DD\setminus [0,1], \RR^n)$ and consider the $\left(Q-\frac12 \right)$-valued function $f=(f^+, f^-)$ defined as follows:
\begin{align}
f_j^+(r,\theta) &= \zeta\left(r^{\frac{2}{2Q-1}}, \frac{2\theta}{2Q-1} + \frac{4\pi}{2Q-1} (j-1) \right), \quad \theta\in [0,\pi], j=1,\cdots,Q, \label{eq:unwindsld1}\\
f_j^-(r,\theta) &= \zeta\left(r^{\frac{2}{2Q-1}}, \frac{2\theta}{2Q-1} + \frac{4\pi}{2Q-1} (j-1) \right), \quad \theta\in[\pi,2\pi], j=1,\cdots,Q-1. \label{eq:unwindsld2}
\end{align}
(For $Q=1$ we just ignore $f^-$.)
Then $f\in W^{1,2}(\DD, \AQt)$ and
\begin{equation}\label{eq:unrollD}
\Dir(f,\mathbb{D}) = \iint_{\mathbb{D}} |D\zeta|^2.
\end{equation}
Moreover, if $\zeta|_{\mathbb{S}^1} \in W^{1,2}(\mathbb{S}^1,\RR^n)$, then $f|_{\So} \in W^{1,2}(\So,\AQt)$ and
\begin{equation}\label{eq:unrollH}
\Dir(f|_\So,\So) = \frac{2}{2Q-1} \int_{\So} |\partial_\tau \zeta|^2,
\end{equation}
where $\partial_\tau$ denotes the tangential derivative on $\So$.
\end{lemma}
\begin{proof}
We define the following subsets of the unit disk,
\[ \mathcal{C} = \left\{re^{i\theta}: 0<r<1, \theta \neq 0\right\}, \]
\[ \mathcal{C}^+ = \left\{re^{i\theta}: 0<r<1, 0< \theta <\pi \right\}, \quad \mathcal{C}^- = \left\{re^{i\theta}: 0<r<1, \pi < \theta <2\pi \right\}; \]
\[ \mathcal{D}_j = \left\{re^{i\theta}: 0<r<1,\, \frac{2\pi}{2Q-1} 2(j-1) < \theta < \frac{2\pi}{2Q-1} 2j \right\}, \quad j=1, \cdots, Q-1, \]
\[ \mathcal{D}_j^+ = \left\{re^{i\theta}: 0<r<1,\, \frac{2\pi}{2Q-1} 2(j-1)< \theta < \frac{2\pi}{2Q-1} (2j-1) \right\}, \quad j=1, \cdots,Q, \]
\[ \mathcal{D}_j^- = \left\{re^{i\theta}: 0<r<1,\, \frac{2\pi}{2Q-1} (2j-1)< \theta < \frac{2\pi}{2Q-1} 2j \right\}, \quad j=1, \cdots,Q-1. \]
For $j=1, \cdots, Q-1$, we define $\varphi_j: \mathcal{C} \to \mathcal{D}_j$ as
\[ \varphi_j(re^{i\theta}) = r^{\frac{2}{2Q-1}} e^{i\left( \frac{2\theta}{2Q-1} + \frac{4\pi}{2Q-1}(j-1) \right)}; \]
and we define $\varphi_Q: \mathcal{C}^+ \to \mathcal{D}_Q^+$ as
\[ \varphi_Q(re^{i\theta}) = r^{\frac{2}{2Q-1}} e^{i\left( \frac{2\theta}{2Q-1} + \frac{4\pi}{2Q-1}(Q-1) \right)}. \]
Then
\[ f^+ = \sum_{j=1}^Q \llbracket \zeta \circ \varphi|_{\mathcal{C}^+} \rrbracket \text{ and } f^- = \sum_{j=1}^{Q-1} \llbracket \zeta \circ \varphi|_{\mathcal{C}^-} \rrbracket. \]
Since $re^{i\theta} \mapsto r^{\frac{2}{2Q-1}} e^{i \frac{2\theta}{2Q-1}}$ is a conformal map, each $\varphi_j$ is conformal.
So by the invariance of the Dirichlet energy under conformal mappings, we deduce that $f^{+} \in W^{1,2}(\mathcal{C}^+, \AQ)$, $f^- \in W^{1,2}(\mathcal{C}^-, \mathcal{A}_{Q-1})$ and
\begin{align*}
\Dir(f, \mathcal{C}) & = \Dir(f^+, \mathcal{C}^+) + \Dir(f^-, \mathcal{C}^-)
= \sum_{j=1}^Q \Dir(\zeta \circ \varphi_j, \mathcal{C}^+) + \sum_{j=1}^{Q-1} \Dir(\zeta \circ \varphi_j, \mathcal{C}^-) \\
& = \sum_{j=1}^Q \Dir(\zeta , \mathcal{D}_j^+) + \sum_{j=1}^{Q-1} \Dir(\zeta , \mathcal{D}_j^-) = \Dir\left(\zeta, \cup_{j=1}^{Q-1} \mathcal{D}_j \cup \mathcal{D}_Q^+ \right)
= \iint_{\DD} |D\zeta|^2.
\end{align*}
On the other hand, since
\[ \partial_\tau \left( \zeta \circ \varphi_j \right) = \partial_\theta \left(\zeta \circ \varphi_j \right) = \frac{2}{2Q-1} \partial_\tau \zeta \circ \varphi_j, \]
we have
\[ \Dir(\zeta\circ \varphi_j|_{\So}, (\So)^+) = \int_{(\So)^+} \left( \frac{2}{2Q-1} \right)^2 | \partial_\tau \zeta \circ \varphi_j|^2 = \frac{2}{2Q-1} \int_{\frac{2\pi}{2Q-1} 2(j-1) }^{\frac{2\pi}{2Q-1} (2j-1)} |\partial_\tau \zeta|^2. \]
An entirely analogous computations on $(\So)^-$ makes it straightforward to show that the restriction $f|_{\So}$ of $f$ to $\So$ belongs to $W^{1,2}(\So, \AQt)$ and that
\[ \Dir(f|_{\So}, \So) = \frac{2}{2Q-1} \int_{\So} |\partial_\tau \zeta|^2. \qedhere\]
\end{proof}
\section{H\"older continuity at the interface}\label{s:hoelder}
In this section we prove the H\"older regularity Theorem \ref{thm:main2}, whose conclusion we make more quantitative in the following statement.
\begin{theorem}[Boundary H\"older regularity of Dir-minimizer, analogue of Theorem 3.9 in \cite{DS}]\label{thm:Holder}
For every $0<\delta< \frac12$, there exist constant $\alpha = \alpha(m,Q)\in (0,1)$ and $C=C(m,n,Q, \delta)$ with the following property. Assume that $\gamma$ is a $C^1$ graph of a function $\zeta$ over $\mathbb R$ passing through the origin with $\|\zeta\|_{C^1}\leq 1$ and that $\varphi \in C^1 (\gamma)$. If $f\in W^{1,2}(B_1, \AQt)$ is Dir-minimizing with interface $(\gamma,\varphi)$, then
\begin{align}
[f]_{C^{0,\alpha}(\overline B_{\delta})} := &\; \max \left\{ \sup_{x,y\in \overline{B^+_{\delta}}} \frac{\GG(f^+(x),f^+(y))}{|x-y|^\alpha}, \sup_{x,y\in \overline{B^-_{\delta}}} \frac{\GG(f^-(x),f^-(y))}{|x-y|^\alpha} \right\}\nonumber\\
\leq &\; C \Dir(f,B_1)^{\frac{1}{2}} + C \|D \varphi\|_{C^0}.
\end{align}
\end{theorem}
The proof consists of two main steps. A comparison argument is used to prove a suitable decay of the Dirichlet energy on balls with vanishing radius. The decay is then combined with a Campanato-Morrey estimate to show H\"older regularity.
\subsection{Campanato-Morrey estimate}
We first record the following extension of a classical result by Morrey. In the case of $Q$-valued maps we refer to \cite{DS}. In our case we need a suitable additional argument to treat the case of $\left(Q-\frac{1}{2} \right)$-valued functions.
\begin{lemma}[Campanato-Morrey estimate]\label{lm:CM}
Suppose $(f^+,f^-)\in W^{1,2}(B_1, \AQt)$ is a map with interface $(\gamma, \varphi)$ as in Theorem \ref{thm:Holder}. If there exist $\beta\in (0,1]$ and $A\geq 0$ such that
\begin{equation}
\iint_{B_r(y)} |Df|^2 \leq A r^{m-2+2\beta} \quad \text{ for every } y\in B_1 \text{ and almost every } r\in (0,1-|y|),
\end{equation}
then for every $0<\delta<1$, there is a constant $C=C(m, \beta, \delta, \gamma)$ such that
\[ [f]_{C^{0, \beta}(\overline B_\delta)} \leq C \sqrt{A} + C \delta^{1-\beta} \|D \varphi\|_{C^0}\, . \]
\end{lemma}
\begin{proof}
We first extend $(f^+,f^-)$ to a function $g:B_1 \to \AQ(\RR^n)$ as follows. We use the $C^1$ regularity of $\gamma$ and $\varphi$ to extend $\varphi$ to a $C^1$ function $\phi$ over $B_1$ satisfying the estimate $\|D \phi\|_{C^0(B_1)}\leq C \|D \varphi\|_{C^0(\gamma)}$, where $C$ depends on $m$ and the $C^1$-norm of $\gamma$.
\begin{equation}\label{eq:extension}
g(x) := \left\{\begin{array}{ll}
f^+(x), & x\in B_1^+ \cup \gamma, \\
f^-(x) + \llbracket \phi (x) \rrbracket, & x\in B_1^-.
\end{array}\right.
\end{equation}
Since $f^+|_{\gamma} = f^-|_{\gamma} + \llbracket \varphi \rrbracket$, the function $g$ belongs to $W^{1,2} (B_1, \AQ)$, by the trace theory of \cite{DS}. Moreover, the theory in \cite{DS} can be easily used to prove that
\[ \iint_{B_r} |Dg|^2 = \iint_{B_r} |Df|^2 + \iint_{B_r^-} |D\phi|^2 \leq r^{m-2+2\beta} \left(A+ r^{2-2\beta} \|D \phi\|_{C^0}^2\right). \]
By the Campanato-Morrey estimate for $Q$-valued functions (see \cite[Proposition 2.14]{DS}), we conclude that
\[ \sup_{x,y\in \overline{B_\delta}} \frac{\mathcal{G}(g(x),g(y))}{|x-y|^{\beta}} \leq C \left( \iint_{B_1} |Dg|^2 \right)^{\frac12}\, .\]
Since clearly $\mathcal{G} (g(x), g(y)) = \mathcal{G} (f^+ (x), f^+ (y))$ for every $x,y\in B_1^+$, we conclude the desired esimate on the H\"older continuity of $f^+$. The one for $f^-$ is slightly more subtle. Consider indeed two points $x,y \in B_1^-$. It then turns out that there are $i, j\in \{1, Q-1\}$ and an invertible map $\sigma : \{1, \ldots , Q-1\}\setminus \{j\} \to \{1, \ldots, Q-1\}\setminus \{i\}$ with the property that
\[
\mathcal{G} (g(x), g(y))^2 = |\phi (x) - f^-_i (y)|^2 + |f^-_j (x) - \phi (y)|^2 + \sum_{k\in \{1, \ldots , Q-1\}\setminus \{j\}} |f^-_k (x) - f^-_{\sigma (k)} (y)|^2\, .
\]
Observe therefore that, by the triangle inequality
\[
|f^-_j (x) - f^-_i (y)|\leq |\phi (x) - \phi (y)| + 2 \mathcal{G} (g (x), g (y))\, .
\]
In particular, using the observation
\[
\mathcal{G} (f^- (x), f^- (y))^2 \leq |f^-_j (x) - f^-_i (y)|^2 + \sum_{k\in \{1, \ldots , Q-1\}\setminus \{j\}} |f^-_k (x) - f^-_{\sigma (k)} (y)|^2\, ,
\]
we achieve
\[
\mathcal{G} (f^- (x), f^- (y))^2 \leq 2|\phi (x) - \phi (y)|^2 + 5 \mathcal{G} (g (x), g (y))^2 \leq 2\|D\phi\|_{C^0}^2 |x-y|^2 + 5 \mathcal{G} (g (x), g (y))^2\, .
\]
Combinining the latter inequality with the estimate for $[g]_{C^{0, \beta}}$ we conclude the desired estimate for the H\"older seminorm of $f^-$.
\end{proof}
\subsection{Almgren's retractions and maximum principle} An important tool in proving the decay of the Dirichlet energy for $Q$-valued minimizers is a family of retraction maps which can be used, for instance, to prove suitable generalizations of the classical maximum principle for harmonic functions. These maps were introduced by Almgren in his pioneering work and we refer to \cite{DS} for an elementary account of them. In order to deal with $\left( Q-\frac{1}{2} \right)$-maps we need an additional property of such retractions, which is not recorded in \cite{DS} (nor in \cite{Alm}). We start by recalling the following notation:
\begin{defn}[Diameter and separation]
Let $T= \sum_i \llbracket P_i \rrbracket \in \AQ$. The \textit{diameter} and \textit{separation} of $T$ are defined, respectively, as
\[ d(T):= \max_{i,j} |P_i - P_j| \text{ and } s(T):= \min \{|P_i - P_j|: P_i \neq P_j \}, \]
with the convention that $s(T) = +\infty$ if $T= Q\llbracket P \rrbracket$.
\end{defn}
For $Y = \sum_i \a{P_i}$ we denote by $\supp\, (T)$ the set of points $\{P_1, \ldots , P_Q\}\subset \RR^n$. Clearly
\begin{equation}\label{eq:defmin}
\dist (\supp (T), q) = \min_i |P_i -q|.
\end{equation}
We have a triangle inequality
\begin{equation}\label{eq:trgmin}
\dist (\supp (T), q) \leq \dist (\supp (S), q) + \GG(T,S), \quad \text{ for every } T, S \in \AQ.
\end{equation}
\begin{lemma}\label{lm:retraction}
Let $T\in \AQ$ and $r< s(T)/4$. Then there exists a retraction $\vartheta: \AQ \to \overline{B_r(T)}$ such that
\begin{enumerate}[(i)]
\item $\GG(\vartheta(S_1), \vartheta(S_2)) < \GG(S_1, S_2)$ if $S_1 \notin \overline{B_r(T)}$,
\item $\vartheta(S) = S$ for every $S\in \overline{B_r(T)}$,
\item If a point $q$ belongs to $\supp (T)$ and to $\supp (S)$, then it belongs to $\supp (\vartheta(S))$ too.
\end{enumerate}
\end{lemma}
\begin{proof}
We define $\vartheta$ in the same way as \cite[Lemma 3.7]{DS}. The properties (i) and (ii) are proved in \cite[Lemma 3.7]{DS} whereas (iii) is an obvious consequence of the explicit formula given in there.
\end{proof}
\begin{prop}[Maximum principle]
Let $f\in W^{1,2}(\Omega, \AQt)$ be a Dir-minimizer with interface $(\gamma, 0)$. Suppose $T\in \AQ$, $0\in \supp\, (T)$ and $0<r< s(T)/4$. If
\begin{align}
\GG(f(x), T) \leq r & \qquad\qquad\mbox{for $\mathcal{H}^{m-1}$-a.e. $x\in (\partial\Omega)^+$ and}\label{e:assum1}\\
\GG (f(x) + \a{0}, T)\leq r & \qquad\qquad \mbox{for $\mathcal{H}^{m-1}$-a.e. $x\in (\partial \Omega)^-$,}\label{e:assum2}\, ,
\end{align}
then
\begin{align}
\GG(f, T) \leq r &\qquad\qquad \mbox{a.e. in $\Omega^+$ and}\\
\GG (f + \a{0}, T) \leq r &\qquad\qquad\mbox{a.e. in $\Omega^-$}\, .
\end{align}
\end{prop}
\begin{proof}
We argue by contradiction. Suppose $f\in W^{1,2}(\Omega, \AQt)$ is a Dir-minimizer with interface $(\gamma, 0)$ satisfying \eqref{e:assum1} and \eqref{e:assum2} and assume in addition that there exists a set of positive measure $E\subset \Omega$, such that $f(x) \notin \overline{B_r(T)}$ for every $x\in E\cap \Omega^+$ and $f(x) + \{0\}\notin \overline{B_r (T)}$ for every $x\in E \cap \Omega^-$.
In particular there exist $\delta>0$ and a set $E' \subset E$ with positive measure such that $f(x) \notin \overline{B_{r+\delta}(T)}$ for every $x\in E'\cap \Omega^+$ and $f(x)+\a{0}\notin \overline{B_{r+\delta} (T)}$ for every $x\in E'\cap \Omega^-$. As in the proof of Lemma \ref{lm:CM} we consider the $Q$-valued function on $\Omega$ which coincides with $f^+$ on $\Omega^+$ and with $f+\a{0}$ in $\Omega^-$. Let $\vartheta: \AQ \to \overline{B_r(T)} $ be the retraction operator in Lemma \ref{lm:retraction}. By (iii) $\supp (\vartheta\circ g(x))$ contains the origin for every $x\in \Omega^-$. We can thus consider the $(Q-1)$-valued function on $\Omega^-$ given by $\vartheta \circ g - \a{0}$. If we set $h^+ = \vartheta \circ g$ on $\Omega^+$ we then get a $(Q-\frac{1}{2})$-valued map $(h^+, h^-)$ with interface $(\gamma, 0)$. By Lemma \ref{lm:retraction}(ii) we also know that $h^\pm = f^\pm$ on $(\partial \Omega)^\pm$. Therefore $h= (h^+, h^-)$ is a suitable competitor for $f=(f^, f^-)$. On the other hand, by Lemma \ref{lm:retraction} (i) we know $|D(\vartheta \circ f)| \leq |Df|$ a.e. on $\Omega$ and moreover, recalling the definition of $\vartheta$ by linear interpolation and that $\GG(f(x), T) > r+\delta$, we get that
\begin{equation}\label{eq:strineq}
|D (\vartheta \circ f)| \leq t_0 |Df| < |Df| \quad \text{ a.e. on } E',
\end{equation}
where $t_0 \leq \frac{r-\delta}{r+\delta} < 1$. Here we compute the partial derivatives by the first order approximation, see the definition and discussions in Definition 1.9, Corollary 2.7 and Proposition 2.17 of \cite{DS}.
We conclude that $\Dir(h, \Omega) < \Dir(f, \Omega)$, contradicting the minimality of $f$.
\end{proof}
\subsection{Decomposition} The maximum principle of the previous section triggers a decomposition lemma for Dir-minimizers with $(\gamma, 0)$ interface.
\begin{prop}[Decomposition of $(Q-\frac{1}{2})$-valued Dir-minimizers]\label{prop:dcp}
There exists a positive constant $\alpha(Q)>0$ with the following property. Assume that $f\in W^{1,2}(\Omega, \AQt)$ is a Dir-minimizer with interface $(\gamma, 0)$, and that there exists $T\in \AQ$ with $0\in \supp\, (T)$ such that \eqref{e:assum1} and \eqref{e:assum2} hold with $r= \alpha (Q) f (T)$.
Then there exists a decomposition $f = (f^+, f^-) = (g^++h, g^- +h)$, where $h$ is a $Q_1$-valued Dir-minimizer, $(g^+, g^-)$ a $(Q_2-\frac{1}{2})$ Dir-minimizer with interface $(\gamma, 0)$, $Q_1+Q_2=Q$ and $1\leq Q_1 \leq Q-1$.
\end{prop}
\begin{proof}
When $d(T) = 0$, our assumption implies $\GG(f(x), T) = 0$, namely $f \equiv T$, and there is nothing to prove. So we assume $d(T)>0$.
If $\alpha(Q) d(T) < s(T)/4$ (for a fixed value of $\alpha(Q)$), the proposition follows directly by the maximum principle and the definition of $s(T)$. Suppose therefore $4 \alpha (Q) d(T) \geq s(T)$. We fix a positive real number $\epsilon$ so that
\[ \left( \sqrt{Q}+2 \right) \frac{\epsilon}{1-\epsilon} = \frac{1}{8}. \]
Recalling \cite[Lemma 3.8]{DS}, we may \textit{collapse} some points in the support $T$ and find an element $S= \sum_{j=1}^J k_j \llbracket S_j \rrbracket \in \AQ$ (with $J\geq 2$) satisfying
\begin{equation}
\beta(\epsilon, Q) d(T) \leq s(S) < +\infty,
\end{equation}
\begin{equation}\label{tmp:dcpQ2}
\GG(S,T) \leq \epsilon s(S).
\end{equation}
We set $\alpha(Q) = \epsilon \beta(\epsilon, Q)$, so that
\begin{equation}\label{tmp:dcpQ1}
\GG(f(x), T) \leq \alpha(Q) d(T)\leq \epsilon s(S) \quad \text{ for } \mathcal{H}^{m-1} \text{-a.e. } x\in \partial\Omega.
\end{equation}
Since $0\in \supp (T)$, we have, by the triangle inequality \eqref{eq:trgmin},
\[ \dist (\supp\, (S), 0) \leq \min \dist (\supp\, (T), 0) + \GG(S, T) \leq \epsilon s(S). \]
Without loss of generality, we assume $|S_1| = \dist (\supp (S), 0)$. Let $\widetilde S = k_1 \llbracket 0 \rrbracket + \sum_{j=2}^J k_j \llbracket S_j \rrbracket$. Clearly
\begin{equation}\label{tmp:dcpQ3}
\GG(S, \widetilde S) = \sqrt{k_1|S_1|^2} \leq \sqrt{Q} \min S \leq \epsilon \sqrt{Q} \, s(S).
\end{equation}
On the other hand $s(\widetilde S) \geq (1-\epsilon) s(S)$. In fact, either $s(\widetilde S) = |S_i - S_j|$ for some $i, j \neq 1$, in which case $s(\widetilde S) \geq s(S)$; or $s(\widetilde S) = |S_i|$ for some $i \neq 1$, and then
\begin{equation}\label{tmp:dcpQ4}
s(\widetilde S) = |S_i| \geq |S_i - S_1| - |S_1| \geq s(S) - \epsilon s(S).
\end{equation}
Combining \eqref{tmp:dcpQ1}, \eqref{tmp:dcpQ2}, \eqref{tmp:dcpQ3}, \eqref{tmp:dcpQ4} and the choice of $\epsilon$, we conclude
\begin{align*}
\GG(f(x), \widetilde S) \leq \GG(f(x), T) + \GG(S, T) + \GG(S, \widetilde S) & \leq \epsilon s(S) + \epsilon s(S) + \epsilon \sqrt{Q} \, s(S) \\
& \leq \left( \sqrt{Q} + 2 \right) \frac{\epsilon}{1-\epsilon} \, s(\widetilde S)
= \frac18 s(\widetilde S),
\end{align*}
for $\mathcal{H}^{m-1}$-a.e. $x\in \partial\Omega$. Again it follows by the maximum principle that $\GG(f, \widetilde S) \leq s(\widetilde S)/8$ almost everywhere on $\Omega$. We thus have a decomposition of $f$ into simpler multiple-valued functions.
\end{proof}
\subsection{Interpolation that preserves the interface value}
In this subsection, we construct interpolations between pairs of $(Q-\frac{1}{2})$ maps with a common interface $(\gamma, 0)$ defined on concentric spheres and estimate its Dirichlet energy. Later we will use the interpolation to construct competitors for Dir-minimizing maps, so it is crucial that the interpolation has the same interface $(\gamma, 0)$. This is also the major difference from the interior case, proved in \cite[Lemma 2.15]{DS}. For our current purpose, namely the proof of the decay of the Dirichlet energy for minimizers, we actually need the existence of the interpolation only in the case $m\geq 3$. However later on Lemma \ref{lm:interpolation} will be used on planar maps to show the compactness of minimizers, a crucial point in the proof of Theorem \ref{thm:main}. We therefore state and proof also the $2$-dimensional case (separately).
\begin{lemma}[Interpolation when $m=2$]\label{lm:interpolation}
Let $f, g$ be maps in $ W^{1,2}(\partial B_1, \AQt(\RR^n))$ satisfying
\begin{equation}\label{eq:tmpbd2}
f^+|_{\gamma} = f^-|_{\gamma} + \llbracket 0 \rrbracket, \quad g^+|_{\gamma} = g^-|_{\gamma} + \llbracket 0 \rrbracket,
\end{equation}
and $\sup_{x\in \partial B_1} \GG(f(x), g(x)) < +\infty $. Let $\delta= \frac{1}{N}$ for some $N\in \mathbb N\setminus \{0,1,2,3\}$. Then there exists $h\in W^{1,2}(B_1\setminus B_{1-\delta}, \AQt(\RR^n))$ satisfying $h^+|_{\gamma} = h^-|_{\gamma} + \llbracket 0 \rrbracket$ and
\[ h(x) = f(x) \text{ for } x\in \partial B_1, \quad h(x) = g \left( \frac{1}{1-\delta}x \right) \text{ for } x\in \partial B_{1-\delta}. \]
Moreover
\begin{equation}\label{eq:interpolation2}
\Dir(h, B_1 \setminus B_{1-\delta}) \leq C\delta \Dir(f, \partial B_1) + C\delta \Dir(g, \partial B_1) + \frac{C}{\delta} \sup_{x\in \partial B_1} \GG(f(x), g(x)).
\end{equation}
\end{lemma}
\begin{proof} By applying a diffeomorphism, we can assume that $\gamma = \mathbb R$.
We first interpolate $f^+$ and $g^+$ in the upper half annulus $B_1^+\setminus B_{1-\delta}^+$. After parametrizing a biLipschitz diffeomorphism $\phi: [0,1] \to \partial B_1^+$ to the functions $f^+$ and $g^+$, we may assume $f^+, g^+$ are $W^{1,2}$ maps defined on $[0,1]$. We will interpolate $f^+$ and $g^+$ and get a $W^{1,2}$ map on $[0,1]\times [0,\delta]$.
We define a cubical decomposition $D_i = [i\delta, (i+1)\delta]\times [0,\delta]$ with $i=0,1, \cdots, N-1$, and vertical lines $\ell_i = \{ i\delta\} \times [0,\delta]$ with $i=0, 1, \cdots, N$. For $i=1,2, \cdots, N-1$, we define
\[ h(x,t)= \bxi^{-1} \circ \brho \left( \left(1-\frac{t}{\delta} \right) \bxi \circ g^+(x) + \frac{t}{\delta} \bxi \circ f^+(x) \right), \quad (x,t) \in \ell_i, \]
where $\bxi: \AQ(\RR^n) \to \RR^N$ is the embedding of $Q$-valued metric space, and $\brho: \RR^N \to \bxi( \AQ)$ is the retraction, see \cite[Theorem 2.1]{DS}. It is clear that
\begin{equation}
\left| Dh(x,t) \right| \leq \frac{C}{\delta} \GG(g^+(x), f^+(x)),
\end{equation}
where the constant depends on the Lipschitz constants of $\bxi$ and $\brho$.
For $i=0$ or $N$ and $x=i\delta$, by recalling \eqref{eq:tmpbd2} we denote
\begin{equation}\label{eq:tmpbdexplicit}
g^+(x) = \sum_{j=1}^Q \llbracket a_j\rrbracket = \llbracket 0 \rrbracket + g^-(x), \quad f^+(x) = \sum_{j=1}^Q \llbracket b_j\rrbracket = \llbracket 0 \rrbracket + f^-(x).
\end{equation}
Here we assume $a_1 = b_1 = 0$ without loss of generality.
Suppose $\tau$ is a permutation of $\{2, \cdots, Q\}$ such that
\[ \GG(g^-(x), f^-(x)) = \sqrt{\sum_{j=2}^Q |a_j - b_{\tau(j)}|^2 }. \]
We define
\begin{equation}\label{eq:tmp}
h(x,t) = \llbracket 0 \rrbracket + \sum_{j=2}^Q \left \llbracket \left( 1- \frac{t}{\delta} \right) a_j + \frac{t}{\delta} b_{\tau(j)} \right\rrbracket.
\end{equation}
\eqref{eq:tmpbdexplicit} implies that
\begin{equation}\label{eq:chgaftzero}
\GG(g^+(x), f^+(x)) \leq \GG(g^-(x), f^-(x)) \leq \sqrt{2} \GG(g^+(x), f^+(x)).
\end{equation}
Hence
\begin{equation}
\left|Dh(x,t) \right| = \frac{1}{\delta} \sqrt{\sum_{j=2}^Q |a_j - b_{\tau(j)}|^2 } = \frac{1}{\delta} \GG(g^-(x), f^-(x)) \leq \frac{\sqrt{2}}{\delta} \GG(g^+(x), f^+(x)).
\end{equation}
In this way $h$ is well-defined for each $\partial D_i$. We now wish to use \eqref{eq:Dirbd} (and a biLipschitz homeomorphism of squares to disks) and claim the existence of an extension $h$ on $D_i$ satisfying
\begin{equation}
\Dir(h, D_i) \leq C \delta \Dir(h, \partial D_i).
\end{equation}
Note that this can be done because the proof of \eqref{eq:Dirbd} given later in the planar case is {\em not} using the current proposition (it uses interpolation, however, if the domain is at least $3$-dimensional).
Summing up we get
\begin{align*}
\Dir(h, [0,1]\times [0,\delta]) & = \sum_{i=0}^{N-1} \Dir(h, D_i) \\
& \leq C \delta \left( \Dir(h, [0,1]\times \{0\}) + \Dir(h,[0,1] \times \{\delta\}) + \sum_{i=0}^N \Dir(h, \ell_i) \right) \\
& \leq C\delta \Dir(g,[0,1]) + C\delta\Dir(f,[0,1]) + C \sum_{i=0}^N \GG(g^+( i \delta), f^+(i\delta)) \\
& \leq C\delta \Dir(g,[0,1]) + C\delta\Dir(f,[0,1]) + \frac{C}{\delta} \sup_{x\in [0,1]} \GG(g^+( x), f^+(x)).
\end{align*}
Applying the biLipschitz homeomorphism $\phi:[0,1] \to \partial B_1^+$, we get an interpolation $h^+ \in W^{1,2}(B_1^+ \setminus B_{1-\delta}^+, \AQ)$.
Similarly, we define an interpolation $h^- \in W^{1,2}(B_1^-\setminus B_{1-\delta}^+, \mathcal{A}_{Q-1})$ between $g^-$ and $f^-$. By \eqref{eq:tmpbd} and the construction \eqref{eq:tmp}, we know $h^+|_{\gamma} = h^-|_{\gamma} + \llbracket 0 \rrbracket$, $h\in W^{1,2}(B_1 \setminus B_{1-\delta}, \AQt)$ and moreover
\begin{equation}
\Dir(h, B_1\setminus B_{1-\delta}) \leq C\delta \Dir(g, B_1) + C\delta\Dir(f, B_1) + \frac{C}{\delta} \sup_{x\in \partial B_1} \GG(g(x), f(x)).
\end{equation}
\end{proof}
\begin{lemma}[Interpolation when $m\geq 3$]\label{lm:interpolation}
Let $f, g$ be maps in $ W^{1,2}(\partial B_1, \AQt(\RR^n))$ satisfying
\begin{equation}\label{eq:tmpbd}
f^+|_{\gamma} = f^-|_{\gamma} + \llbracket 0 \rrbracket, \quad g^+|_{\gamma} = g^-|_{\gamma} + \llbracket 0 \rrbracket,
\end{equation}
and $\int_{ \partial B_1} \GG(f, g) < +\infty $. Let $\delta= \frac{1}{N}$ for some $N\in \mathbb N\setminus \{0,1,2,3\}$. Then there exists $h\in W^{1,2}(B_1\setminus B_{1-\delta}, \AQt(\RR^n))$ satisfying $h^+|_{\gamma} = h^-|_{\gamma} + \llbracket 0 \rrbracket$ and
\[ h(x) = f(x) \text{ for } x\in \partial B_1, \quad h(x) = g \left( \frac{1}{1-\delta}x \right) \text{ for } x\in \partial B_{1-\delta}. \]
Moreover
\begin{equation}\label{eq:interpolation}
\Dir(h, B_1 \setminus B_{1-\delta}) \leq C\delta \Dir(f, \partial B_1) + C\delta \Dir(g, \partial B_1) + \frac{C}{\delta} \int_{ \partial B_1} \GG(f, g).
\end{equation}
\end{lemma}
\begin{proof} By applying a diffeomorphism, we can assume that $\gamma = \{x_m =0\}$.
Let $\mathcal{C}$ be the boundary of the cube $[-1, 1]^m$. Notice that $\mathcal{C}$ is tangent to the sphere $\partial B_1$.
We define the functions $\hat f$ and $\hat g$ on $\mathcal{C}$ by radial projection:
\[ \hat f(z) := f\left( \frac{z}{|z|} \right), \quad \hat g(z) := g\left( \frac{z}{|z|} \right), \quad \text{ for every } z\in \mathcal{C}. \]
After the radial projection, the tangential derivative on $\mathcal{C}$ at $z$ is just a multiple of the tangential derivative on $\partial B_1$ at $z/|z|$, where the factor is uniformly bounded above and below by dimensional constants. In particular $\hat f, \hat g \in W^{1,2}(\mathcal{C}, \AQt)$, that is,
\[ \hat f^+, \hat g^+ \in W^{1,2}(\mathcal{C}^+, \AQ), \quad \hat f^-, \hat g^- \in W^{1,2}(\mathcal{C}^-, \mathcal{A}_{Q-1}), \]
and
\[ \hat f^+|_\gamma = \hat f^-|_\gamma +\llbracket 0 \rrbracket, \quad \hat g^+|_\gamma = \hat g^-|_\gamma +\llbracket 0 \rrbracket, \]
where $\mathcal{C}^+ = \mathcal{C}\cap \{x_m >0\}$, $\mathcal{C}^- = \mathcal{C} \cap \{x_m <0\}$ and $\gamma = \{x_m = 0\}$.
We want to construct a function $\hat h: \mathcal{C} \times [0,\delta] \to \AQt$ which satisfies $\hat h(\cdot, 0) = \hat g$, $\hat h(\cdot, \delta) = \hat f$, $\hat h \in W^{1,2}$ and
\[ \hat h^+|_{\gamma \times[0,\delta]} = \hat h^-|_{\gamma \times[0,\delta]} + \llbracket 0 \rrbracket; \] and in turn, we define a function $h: B_1\setminus B_{1-\delta} \to \AQt $ by
\[ h\left( t\, \frac{z}{|z|} \right) := \hat h\left(z, t-(1-\delta) \right), \quad \text{ for each } z\in \mathcal{C} \text{ and } 1-\delta < t < 1, \]
such that $h\in W^{1,2}(B_1 \setminus B_{1-\delta}, \AQt)$ with the desired boundary data.
Let $F$ be any of the $2m$ faces of $\mathcal{C}$, then it is an $(m-1)$-dimensional solid cube (i.e. including the interior) with side length $2$. Take for example
\[ F=\left\{\left(-1, x_2, \cdots, x_m \right): -1 \leq x_j \leq 1 \text{ for every } j=2, \cdots, m \right\}. \]
We will first define $\hat h$ on $F\times [0, \delta]$ using the similar construction as in the interior case, see Step 1 of \cite[Lemma 4.12]{DS} and the erratum therein. To that end we first need to extend $\hat f$ and $\hat g$ to a fatter region
\[ F_\delta:= \left\{\left(-1, x_2, \cdots, x_m \right): -1-\delta \leq x_j \leq 1 + \delta \text{ for every } j=2, \cdots, m \right\}, \]
by using their respective values on neighboring faces of $F$ and scaling appropriately on the corners. For example, for any $x_2 \in [-1 - \delta, - 1)$ fixed (the other possibility being $x_2 \in (1, 1+\delta]$), we consider the slice
\[ S_{x_2} := \left\{ \left(-1, x_2, x_3, \cdots, x_m \right): -|x_2| \leq x_j \leq |x_2| \text{ for every } j=3, \cdots, m \right\} \subset F_{\delta}, \]
and define $\hat f, \hat g$ by their values on a neighboring face of $F$:
\begin{equation}\label{eq:defFng}
F':= \left\{(x_1, -1, x_3, \cdots, x_m): -1\leq x_j \leq 1 \text{ for every } j= 1, 3, \cdots, m \right\}.
\end{equation}
To be precise on $S_{x_2}$ we define
\begin{equation}\label{eq:fatdef1}
\hat f\left(-1, x_2, x_3, \cdots, x_m \right) := \hat f \left( |x_2|-2,\, - 1, \, \varphi_\delta( x_3), \cdots, \varphi_\delta(x_m) \right)
\end{equation}
where $\varphi_\delta: [-|x_2|, |x_2|] \to [-1, 1]$ is a piecewise linear function as follows
\begin{equation}\label{eq:fatdef2}
\varphi_\delta(t) = \left\{\begin{array}{ll}
-1 + \dfrac{\delta}{-1 + \delta+ |x_2|}\left( t + |x_2| \right), & -|x_2| \leq t \leq -1 + \delta \\
t, & -1 + \delta \leq t \leq 1 -\delta \\
1 + \dfrac{\delta}{-1 + \delta + |x_2|} \left( t-|x_2| \right), & 1 - \delta \leq t \leq |x_2|.
\end{array}\right.
\end{equation}
That is, in the inner region of $S_{x_2}$, $\hat f$ (as well as $\hat g$) takes value on $F'$ faithfully; in the outer region $\hat f$ (as well as $\hat g$) is a scaled version of its value on $F'$, with a scaling factor at most $2$. The former is to guarantee that the construction of $\hat h$ remains faithful to $\hat f, \hat g$ near the boundary $\gamma \times [0,\delta]$.
For any vector $v\in [-1-\delta,-1]^{m-1}$, consider the cubical decomposition of $F_\delta$ induced by the lattice points $\{- 1 \} \times \left( v+ \delta \mathbb{Z}^{m-1} \right)$. For $k\in \{0, \cdots, m-1\}$ we define accordingly the $k$-dimensional skeleta contained in $F_\delta$, which are the families $\mathcal{S}^k(v)$ of all closed $k$-dimensional faces of the cubes. By Fubini, for almost every $v$ and face $E\in \mathcal{S}^k(v)$, we have that $\hat f|_E, \hat g|_E \in W^{1,2}$, and moreover
\begin{align*}
& \int_{v\in [-1 - \delta, -1]^{m-1} } \left( \sum_{E\in \mathcal{S}^k(v)} \int_E \left( |D\hat f|^2 + |D\hat g|^2 + \mathcal{G}(\hat f,\hat g)^2 \right) \right) dv\\
\leq & C(k,m) \delta^{k } \int_{F_{\delta}} \left( |D\hat f|^2 + |D\hat g|^2 + \mathcal{G}(\hat f,\hat g)^2 \right).
\end{align*}
By standard arguments we can choose a vector $v$ such that
\begin{itemize}
\item For every $k\geq 1$, for each $E\in \mathcal{S}^k(v)$ and each $G\in \mathcal{S}^{k-1}(v)$ with $G\subset E$, the restrictions $\hat f|_E, \hat f|_G, \hat g|_E, \hat g|_G$ are all $W^{1,2}$ and moreover the traces of $\hat f|_E$ and $\hat g|_E$ on $G$ are precisely $\hat f|_G$ and $\hat g|_G$;
\item For every $k\geq 1$,
\[ \sum_{E\in \mathcal{S}^k(v)} \int_E \left( |D\hat f|^2 + |D\hat g|^2 \right) \leq C \delta^{k-(m-1)} \int_{F_{\delta}} \left( |D\hat f|^2 + |D\hat g|^2 \right); \]
\item For $k=0$,
\[ \sum_{p\in \mathcal{S}^0(v)} \mathcal{G}(\hat f(p), \hat g(p))^2 \leq C\delta^{-(m-1)} \int_{F_\delta} \mathcal{G}(\hat f,\hat g)^2; \]
\item Whenever $E\in \mathcal{S}^{k}(v)$ intersects $\gamma$, the center of $E$, denoted by $x_E$, lies in $\mathcal{C}^+$, in other words $x_E$ lies above the boundary $\gamma$.
\end{itemize}
For any $k=0, \cdots, m-1$ and any $E\in \mathcal{S}^{k}(v)$ not intersecting $\gamma$, we follow the same construction as in the interior case (for $Q$-valued or $(Q-1)$-valued functions) and define $\hat h$ on $E \times [0,\delta]$ by interpolation of $\hat f^+$ and $\hat g^+$, or $\hat f^-$ and $\hat g^-$ respectively. Across the boundary $\gamma$, we temporarily extend the functions trivially by zero, that is, we set
\[ \hat f_0 = \left\{\begin{array}{ll}
\hat f^+, & \text{ on } \mathcal{C}^+ \\
\hat f^- + \llbracket 0 \rrbracket, & \text{ on } \mathcal{C}^-,
\end{array} \right.
\quad
\hat g_0 = \left\{\begin{array}{ll}
\hat g^+, & \text{ on } \mathcal{C}^+ \\
\hat g^- + \llbracket 0 \rrbracket, & \text{ on } \mathcal{C}^-,
\end{array} \right. \]
so that $\hat f_0, \hat g_0$ are $Q$-valued functions. Notice that the values of $|D\hat f|,\, |D\hat g|,\, \mathcal{G}(\hat f^+, \hat g^+)$ stay the same, and on $\mathcal{C}^-$
\[ \frac{1}{\sqrt{2}} \mathcal{G}(\hat f^-, \hat g^-) \leq \mathcal{G}(\hat f_0, \hat g_0) \leq \mathcal{G}(\hat f^-, \hat g^-), \]
see \eqref{eq:chgaftzero}.
Recall that for any $p\in \mathcal{S}^0(v)$ contained in $\mathcal{C}^-$, we define $\hat h$ on $p\times [0,\delta]$ as a linear interpolation between $\hat f$ and $\hat g$, and that
\begin{equation}\label{eq:intpbase}
\Dir(\hat h, p\times[0,\delta]) \leq \frac{C}{\delta} \mathcal{G}(\hat f^-(p), \hat g^-(p))^2 \leq \frac{C'}{\delta} \mathcal{G}(\hat f_0(p), \hat g_0(p))^2.
\end{equation}
Now we construct $\hat h$ by an induction on the dimension $k$. Suppose $E\in \mathcal{S}^k (v)$ intersects $\gamma$, where $k = 1, \cdots, m-1$. Either by the inductive hypothesis or by the base case $k=0$ (see \eqref{eq:intpbase} and assume $\hat h_0(p) = \hat h(p) + \llbracket 0 \rrbracket$), we assume that for all lower skeleta $G\in \mathcal{S}^{k-1}(v)$ with $G\subset E$, we have defined a $Q$-valued function $\hat h_0$ on $G \times [0,\delta]$ with the desired properties. Since
\[ \partial\left( E\times [0,\delta] \right) = \bigcup_{G\in \mathcal{S}^{k-1}(v) \atop{G\subset E}} \left( G\times[0,\delta]\right) \, \bigcup \, \left(E\times \{0\} \right) \, \bigcup \, \left( E \times \{\delta\} \right), \]
we can define $\hat h_0$ on $E\times [0,\delta]$ as the $0$-homogeneous extension of $\hat h_0|_{\partial \left( E\times [0,\delta]\right)}$. Simple computations show that
\[ \Dir(\hat h_0, E \times[0,\delta] ) \leq C \delta \Dir\left(\hat h_0, \partial\left( E \times [0,\delta] \right) \right). \]
More importantly, notice that every point on $\left( E \cap \mathcal{C}^-\right) \times (0,\delta)$ lies in a line segment between the center $x_E \times \{\delta/2\}$ and some point in $\left( E \cap \mathcal{C}^- \right) \times \{0\}$, $\left( E \cap \mathcal{C}^- \right) \times \{\delta\}$ or $\left(G\cap \mathcal{C}^- \right) \times [0,\delta]$ for some $G \in \mathcal{S}^{k-1}(v)$ and $G \subset E$, hence this construction guarantees that on $\mathcal{C}^- \times [0,\delta]$, the $Q$-valued function $\hat h_0$ always has an element $\llbracket 0 \rrbracket$; in particular we may define $\hat h \in \AQt$ accordingly and it satisfies the desired boundary condition. To sum up, we construct a function $\hat h_F$ defined on $\widetilde F_\delta \times [0,\delta]$, where $F \subset \widetilde F_\delta \subset F_\delta$, and it satisfies
\[ \hat h_F(\cdot, 0) = \hat g, \quad \hat h_F(\cdot, \delta) = \hat f \text{ on } F; \]
\[ \hat h_F^+(\cdot, t)\big|_{\gamma} = \hat h_F^-(\cdot, t)\big|_{\gamma} + \llbracket 0 \rrbracket \quad \text{ for every } t\in [0,\delta]; \]
\[ \Dir(\hat h_F, F\times [0,\delta]) \leq C \delta \Dir(\hat f, F_\delta) + C\delta \Dir(\hat g, F_\delta) + \frac{C}{\delta} \int_{F_\delta } \mathcal{G}(\hat f, \hat g)^2. \]
We would like to repeat the same argument for any neighboring face of $F$, take for example $F'$ as in \eqref{eq:defFng}; but we need to be careful and make sure the new function $\hat h_{F'}$ is consistent with $\hat h_F$ on their domains of overlap, since $\hat h_F$ is defined on a small neighborhood near $F \cap F'$ by projecting the fattened region $\widetilde F_\delta$ onto $F'$:
\[ \Ng(F):= F' \cap \left\{-1 \leq x_1 \leq -1 + \delta' \right\},\]
where $\delta' \in [0, \delta)$ is determined by the choice of $v$.
We sketch the necessary technical modifications below. As before, we consider a fattened region $F'_\delta$ of $F'$; and we then choose a cubical decomposition of $F'_\delta$ to satisfy, in addition to the requirements stated above, that all skeleta (orthogonal to $x_1$-axis) ought to be at least $\delta/2$-distance away from $\Ng(F)$. On the interior region
\[ \Ng^i(F):= \Ng(F) \cap \left\{ - 1 + \delta \leq x_j \leq 1 - \delta \text{ for every } j= 3, \cdots, m \right\}, \]
we use $\hat h_F$ as the boundary condition to construct $\hat h_{F'}$ to make sure they agree; outside, on each $(m-1)$-dimensional $\delta$-cube $E$ contained in $\Ng(F) \setminus \Ng^i(F)$, we replace and reconstruct $\hat h_F$ on $E\times [0,\delta]$ as above. This way $\hat h_F = \hat h_{F'}$ on their domains of overlap $\Ng(F)$; moreover, since we do not redefine $\hat h_F$ near the boundary $\gamma \times [0,\delta]$, it still satisfies the desired boundary condition.
\end{proof}
\subsection{Decay estimate} The key point in the proof of Theorem \ref{thm:Holder} is a suitable decay estimate for the Dirichlet energy, which is essentially the content of the following proposition.
\begin{prop}\label{prop:Dirbd}
Suppose $f$ is a $(Q-\frac{1}{2})$ Dir-minimizing map on $B_1$ with interface $(\gamma, \varphi)$ and assume that $\gamma$ is the graph of a function $\zeta$ with $\|\zeta\|_{C^1}\leq 1$. Let $0<r<1$ and assume that $f|_{\partial B_r} \in W^{1,2}(\partial B_r, \AQt)$.
Then we have
\begin{equation}\label{eq:Dirbd}
\Dir(f,B_r) \leq C(m) r \Dir (f, \partial B_r) + C r^m \|D\varphi\|_{C^0}^2\, ,
\end{equation}
where $C(m) < (m-2)^{-1}$.
\end{prop}
\begin{remark}\label{rmk:Dirbd}
By translation, the same estimate holds for any ball $\overline{B_r(y)} \subset B_1$ with $y\in \gamma$. If $\overline{B_r(y)} \cap \gamma = \emptyset$, the analogous interior estimate was proven in \cite[Proposition 3.10]{DS}.
\end{remark}
\begin{proof}
We will prove \eqref{eq:Dirbd} for $r=1$, because the general case follows from a scaling argument. Moreover we will assume, without loss of generality, that $\varphi \equiv 0$. Indeed, for a general $\varphi$, we let $\phi$ be an extension to $B_1$ with the property that $\|D\phi\|_{C^0(B_1)} \leq C \|D\varphi\|_{C^0(\gamma)}$, since the interface $\gamma$ is given by the graph of $\zeta$ satisfying $\|\zeta\|_{C^1} \leq 1$. Define then $(h^+, h^-)$ as
\[
h^\pm (x) = \sum_i \a{f^\pm (x) - \phi (x)}\, .
\]
Moreover, let $k^\pm$ be a Dir-minimizer with boundary values $h^\pm$ and interface $(\gamma, 0)$ and construct a corresponding competitor for $f$ by setting
\[
\bar{g}^\pm (x) = \sum_i \a{k^\pm +\phi (x)}\, .
\]
Observe that for every $\varepsilon$ there is a constant $C(\varepsilon)$ such that
\begin{align*}
|D_\tau h^\pm (x)|^2 &\leq (1+\varepsilon) |D_\tau f^\pm (x)|^2 + C (\varepsilon) |D_\tau \phi (x)|^2, \\
|D \bar g^\pm (x)|^2 &\leq (1+\varepsilon) |D k^\pm (x)|^2 + C (\varepsilon) |D \phi (x)|^2.
\end{align*}
Here $D_\tau$ denotes the tangential derivative on the boundary $\partial B_1$.
After proving the Proposition for interfaces $(\gamma, 0)$ we will know that there is a constant $C' (m) < \frac{1}{m-2}$ such that
\begin{align*}
\Dir (k, B_1) &\leq C'(m) \Dir (k, \partial B_1) = C'(m) \Dir (h, \partial B_1) \leq C' (m)( 1 + \varepsilon) \Dir (f, \partial B_1) + C (m, \varepsilon) \|D\phi\|_{C^0}^2.
\end{align*}
Hence we could estimate
\begin{align*}
\Dir (f, B_1) &\leq \Dir (\bar g, B_1) \leq (1+\varepsilon) \Dir (k, B_1) + C (\varepsilon) \|D\phi\|_{C^0}^2\\
&\leq C'(m) (1+\varepsilon)^2 \Dir (f, \partial B_1) + C' (m, \varepsilon) \|D\phi\|_{C^0}^2\, .
\end{align*}
Since $C'(m) < \frac{1}{m-2}$ it suffices to choose $\varepsilon$ so that $C(m) := C'(m) (1+\varepsilon)^2 < \frac{1}{m-2}$.
From now on we restrict therefore our attention to the case $\varphi \equiv 0$.
\medskip
\textbf{The planar case.} Set $g:= f|_{\partial B_1}$ and let $g= g_0 + \sum_{j=1}^J g_j$ be a decomposition into irreducible maps as in Proposition \ref{prop:decomp}. Suppose $g_0$ unwinds to $\zeta_0: \So \to \RR^n$ as in Proposition \ref{prop:decomp} (ii); and each $g_j$ unwinds to a $W^{1,2}$ function $\zeta_j: \So \to \RR^n$ as in \cite[Proposition 1.5 (ii)]{DS}:
\[ g_j(x) = \sum_{z^{Q_j} = x } \llbracket \zeta_j(z) \rrbracket. \]
Now we construct an admissible competitor for $f$ as follows. Recall that $\zeta_0(0) = \zeta_0(2\pi) = 0$, we consider its Fourier expansion
\[ \zeta_0(\theta) = \sum_{l=1}^\infty c_l \sin \left( \frac{l\theta}{2} \right). \]
We then extend $\zeta_0$ to be a $W^{1,2}$ function defined on all of $B_1$ as:
\[ \overline\zeta_0(r,\theta) = \sum_{l=1}^\infty r^{\frac{l}{2}} c_l \sin \left( \frac{l\theta}{2} \right). \]
Note that $\overline\zeta_0$ is not harmonic, but it vanishes on all of the positive real axis. We also consider the harmonic extension of each $\zeta_j$, denoted by $\overline\zeta_j$. Simple computations show that
\begin{equation}\label{tmp:ext}
\iint_{\DD} |D\overline\zeta_0|^2 \leq 2 \int_{\So} |\partial_\tau \zeta_0 |^2, \quad \iint_{\DD} |D\overline{\zeta_j}|^2 \leq \frac12 \int_{\So} |\partial_\tau \zeta_j |^2.
\end{equation}
We then unroll $\overline\zeta_0$ to a $(Q_0 - \frac12)$-valued function $h_0 = (h_0^+, h_0^-)$ as in Lemma \ref{lm:unrolling}. By definition, it follows that $h_0$ satisfies the boundary condition
\[ h_0^+|_{\gamma} = h_0^-|_{\gamma} + \llbracket 0 \rrbracket. \]
We also unroll each $\overline\zeta_j$ to $Q_j$-valued function $h_j$ by
\[ h_j(x) = \sum_{z^{Q_j} = x} \llbracket \overline\zeta_j (z) \rrbracket. \]
The function $h = (h_0^+, h_0^-) + \sum_{j=1}^J h_j$ has interface $(\gamma, \varphi)$, agrees with $f$ on $\So$, and thus is an admissible competitor for $f$ in $B_1$. Therefore by Lemma \ref{lm:unrolling}, \cite[Lemma 3.12]{DS} and \eqref{tmp:ext}, we get
\begin{align*}
\Dir(f, B_1) \leq \Dir(h, B_1) &= \sum_{j=0}^J \Dir(h_j, B_1) = \sum_{j=0}^J \iint_{\DD} |D\overline \zeta_j|^2
\leq 2 \sum_{j=0}^J \int_{\So} |\partial_\tau \zeta_j|^2 \\
& = (2Q_0-1) \Dir(g_0, \So) + \sum_{j=1}^J 2Q_j \Dir(g_j, \So)
\leq 2Q \Dir(g, \partial B_1).
\end{align*}
In particular, the above inequality says that the constant in \eqref{eq:Dirbd} satisfies $C(2) = 2Q(1+\epsilon)^2$, and we may assume that $C(2) = 3Q$ for example.
\textbf{The non-planar case.} We define $Q$-valued functions $\tilde{g}$ and $\tilde{f}$ by adding a ``$0$ sheet'' to $g^-$ and $f^-$, as in \eqref{eq:extension}. Observe that $|D\tilde g (x)| = |Dg^\pm (x)|$ and $|D_\tau \tilde{f} (x)| = |D_\tau f (x)|$. So, rather than exhibiting a competitor for $g$ we wish to exhibit a competitor, say $h$, for $\tilde{g}$: we just have to respect the property that $\supp\, h (x)\ni 0$ for every $x\in B_1^-$. With a slight abuse of notation we thus keep the notation $g$ and $f$ for $\tilde{g}$ and $\tilde{f}$.
\textit{Step 1. Radial competitors.} Let $\bar g= \sum_i \llbracket \bar g_i \rrbracket \in \AQ$ be a mean for $g$ so that the Poincar\'e inequality of \cite[Proposition 2.12]{DS} holds, i.e.
\begin{equation}\label{eq:Poincarebd}
\left( \int_{\partial B_1} \GG(g, \bar g)^{p} \right)^{1/p} \leq C (p) \left( \int_{\partial B_1} |Dg|^2 \right)^{1/2},
\end{equation}
where the exponent $p$ can be taken to be any finite real $p\geq 1$ if $m=3$ and any real $1\leq p\leq 2^*$ (with $\frac{1}{2^*} = \frac12 - \frac{1}{m-1}$) when $m\geq 4$. Assume the diameter of $\bar g$ is smaller than a constant $M>0$ (whose value is to be determined later),
\[ d(\bar g) \leq M. \]
Recall next \eqref{eq:defmin} and define the function $m (x):= \dist (\supp\, g (x), 0)$. Observe that
\[
T \;\mapsto\; \dist (\supp\, (T), 0)
\]
is a Lipschitz map with constant less or equal than $1$ by \eqref{eq:trgmin}: thus $|D m|\leq |Dg|$ and $|D_\tau m|\leq |D_\tau g|$. Moreover $m$ obviously vanishes on $\partial B_1^-$ (whose surface measure is larger than a geometric constant). By the relative Poincar\'e inequality, we know
\begin{equation}
\int_{\partial B_1} m(x)^2 \leq C \int_{\partial B_1} | D m (x)|^2 \leq C \int_{\partial B_1} |Dg(x)|^2.
\end{equation}
Hence
\begin{equation}\label{eq:ming}
\bar m ^2 := \dist (\supp ( \bar g), 0)^2 = \fint_{\partial B_1} m^2 \lesssim \int_{\partial B_1} m^2 + \int_{\partial B_1} \GG(g(x), \bar g)^2 \leq C \int_{\partial B_1} |Dg(x)|^2.
\end{equation}
Combined with the assumption $d(\bar g) \leq M$, it follows that
\[ |\bar g|^2 = \sum_i |\bar g_i|^2 \leq Q(C+M^2). \]
Thus
\begin{align*}
\int_{\partial B_1} |g|^2 \leq 2 \int_{\partial B_1} \GG(g, \bar g)^2 + 2 \int_{\partial B_1} |\bar g|^2 \leq C_{Q,M},
\end{align*}
where $C_{Q,M}$ is a constant depending on $Q$ and $M$ with positive correlation.
Let $\varphi$ be a real-valued function in $W^{1,2}([0,1])$ with $\varphi(1) = 1$. Then
\[ \hat f(x) := \varphi(|x|) g\left( \frac{x}{|x|} \right)\] is a suitable competitor for $f$. A simple computation shows that
\begin{align*}
\iint_{B_1} |D \hat f|^2 & = \left( \int_{\partial B_1} |g|^2 \right) \int_0^1 \varphi'(r)^2 r^{m-1} dr + \left( \int_{\partial B_1} |Dg|^2 \right) \int_0^1 \varphi(r)^2 r^{m-3} dr \\
& \leq \int_0^1 \left( \varphi(r)^2 r^{m-3} + C_{Q,M} \varphi'(r)^2 r^{m-1} \right) dr =: I(\varphi).
\end{align*}
By minimality we deduce that
\[ \Dir(f, B_1) \leq \inf_{\varphi \in W^{1,2}([0,1]) \atop{\varphi(1) = 1}} I(\varphi). \]
We notice that $I(1) = \frac{1}{m-2}$ ($\varphi \equiv 1$ corresponds to the trivial radial competitor for $f$). On the other hand $\varphi \equiv 1$ can not be a minimum for $I$ because it does not satisfy the corresponding Euler-Lagrange equation. So there exists a constant $\gamma = \gamma(Q, M)>0 $ such that
\[ \Dir(f, B_1) \leq \inf_{\varphi \in W^{1,2}([0,1]) \atop{\varphi(1) = 1}} I(\varphi) = \frac{1}{m-2} - \gamma. \]
In particular, when $Q=1$, the diameter $d(\bar g) = 0$ and we are done. We will prove the proposition by an induction on $Q$.
\textit{Step 2. Splitting procedure: the inductive step.} Let $Q\geq 2$ be fixed and assume that the proposition holds for every $Q^* < Q$. Assume moreover that $d(\bar g)> M$. The strategy of the proof is to decompose $f$ into several pieces in order to apply the inductive hypothesis. To that end, we first \textit{collapse} the mean $\bar g$, by applying \cite[Lemma 3.8]{DS} to $T= \bar g$. For any $\epsilon \in (0,1)$, we obtain $S= \sum_{j=1}^J k_j \llbracket S_j \rrbracket \in \AQ$ which satisfies
\begin{equation}\label{eq:clp1}
\beta M \leq \beta d(\bar g) \leq s(S) < +\infty,
\end{equation}
\begin{equation}\label{eq:clp2}
\GG(S, \bar g) \leq \epsilon s(S).
\end{equation}
Here $\beta = \beta(\epsilon, Q)$ is the constant in Lemma 3.8.
The fact that $s(S)< +\infty$ means $J\geq 2$. Recall \eqref{eq:ming} (this estimate is independent of the assumption on $d(\bar g)$), we get
\begin{equation}\label{tmp:clp1}
\min S \leq \min \bar g + \GG(S, \bar g) \leq C+ \epsilon s(S).
\end{equation}
Assume without loss of generality that $|S_1| = \min S$. We let
\[ \widetilde S := k_1 \llbracket 0 \rrbracket + \sum_{j=2}^J k_j \llbracket S_j \rrbracket. \]
By \eqref{tmp:clp1},
\begin{equation}\label{tmp:clp2}
\GG(S, \widetilde S) \leq \sqrt{k_1 |S_1|^2} < \sqrt{Q} \dist (\supp\, (S), 0) \leq C \sqrt{Q} + \epsilon \sqrt{Q} s(S).
\end{equation}
We fix $\epsilon$ with $\epsilon\sqrt{Q} = \frac{1}{64}$; we may also choose $M = M(Q, \beta(\epsilon,Q))$ sufficiently large
\begin{equation}\label{tmp:clp3}
C \leq \epsilon \beta M \leq \epsilon s(S).
\end{equation}
Thus it follows from \eqref{eq:clp1} that
\[ \GG(S, \widetilde S)< 2\epsilon\sqrt{Q} s(S) = \frac{1}{32} s(S). \]
Combined with \eqref{eq:clp2}, we have
\begin{equation}\label{eq:bargS}
\GG(\bar g, \widetilde S) \leq \sqrt{2\GG(S, \bar g)^2 + 2 \GG(S, \widetilde S)^2} < \frac{1}{16} s(S).
\end{equation}
On the other hand, we also have
\begin{equation}\label{tmp:lbdist}
s(\widetilde S) \geq (1-2\epsilon) s(S).
\end{equation}
In fact, either $s(\widetilde S) = |S_i - S_j|$ for $i, j \neq 1$, in which case $s(\widetilde S) \geq s(S)$ by definition; or $s(\widetilde S) = |S_i|$ for some $i \neq 1$, then by \eqref{tmp:clp1} and \eqref{tmp:clp3}
\[ s(\widetilde S) = |S_i| \geq |S_i - S_1| - |S_1| \geq s(S) - \min S \geq (1-2\epsilon) s(S). \]
Let
\[ \vartheta: \AQ \to \overline{B_{s(\widetilde S)/8}(\widetilde S)} \]
be the retraction given by Lemma \ref{lm:retraction}. We define $h\in W^{1,2}(B_{1-\eta})$ by
\[ h(x) := \vartheta \left( f\left( \frac{x}{1-\eta} \right) \right), \]
where $\eta$ is a small parameter to be determined later. By Lemma \ref{lm:retraction} (iii), $h(x)$ contains a zero element for every $x\in B_{1-\eta}^-$. By removing one zero element in the lower half space we may consider $h$ as a function in $W^{1,2}(B_{1-\eta}, \AQt)$. Therefore by \cite[Theorem 4.2]{DDHM} there exists a Dir-minimizer $\hat h\in W^{1,2}(B_{1-\eta}, \AQt)$ with interface $(\gamma, \varphi)$, such that $\hat h = h$ on $\partial B_{1-\eta} \setminus \gamma$. Almost everywhere on $\partial B_{1-\eta}$, $\hat h$ takes value in $\overline{B_{s(\widetilde S)/8}(\widetilde S)}$. Therefore by Proposition \ref{prop:dcp} $\hat h$ can be decomposed into the sum of $h_1$ and $h_2$, where $h_1$ is a $K$-valued function and Dir-minimizer, $h_2$ is an $L$-valued function and Dir-minimizer with interface $(\gamma, \varphi)$, and $K, L \leq Q-1$. By \cite[Proposition 3.10]{DS} and the inductive hypothesis, we have
\[ \Dir(h_1, B_{1-\eta}) \leq \left( \frac{1}{m-2} - \gamma_i \right)(1-\eta) \, \Dir(h_1, \partial B_{1-\eta}), \]
\[ \Dir(h_2, B_{1-\eta}) \leq \left( \frac{1}{m-2} - \gamma_b \right)(1-\eta) \, \Dir(h_2, \partial B_{1-\eta}). \]
Hence
\begin{align*}
\Dir(\hat h, B_{1-\eta}) & \leq \left( \frac{1}{m-2} - \gamma \right)(1-\eta) \, \Dir(h, \partial B_{1-\eta}) \\
& = \left( \frac{1}{m-2} - \gamma \right)(1-\eta)^{m-2} \, \Dir(g, \partial B_{1}) \\
& < \left( \frac{1}{m-2} - \gamma \right).
\end{align*}
Here $\gamma_0 = \min\{\gamma_i, \gamma_b\}>0$ is a constant depending on $m$ and $Q$.
We consider the following competitor
\[ \hat f = \left\{\begin{array}{ll}
\hat h, & \text{ in } B_{1-\eta} \\
\text{interpolation between } \hat h \text{ and } g \text{ as in Lemma \ref{lm:interpolation}, } & \text{ in } B_1 \setminus B_{1-\eta}.
\end{array} \right. \]
By the estimate \eqref{eq:interpolation},
\begin{align}
\Dir(\hat f, B_1) & \leq \Dir(\hat h, B_{1-\eta}) + C\eta\left( \Dir(\hat h, \partial B_{1-\eta}) + \Dir(g, \partial B_1) \right) + \frac{C}{\eta} \int_{\partial B_1} \GG(g, \vartheta(g))^2 \nonumber \\
& < \frac{1}{m-2} - \gamma_0 +2C\eta + \frac{C}{\eta} \int_{\partial B_1} \GG(g, \vartheta(g))^2.\label{tmp:cpt}
\end{align}
Now we estimate the last term in the right hand side of \eqref{tmp:cpt}. By the definition of the retraction $\vartheta$, $g$ and $\vartheta(g)$ only differ on the set
\[ E: = \left\{x\in \partial B_1: g(x) \notin \overline{B_{s(\widetilde S)/8}(\widetilde S) } \right\}. \]
For every $x\in E$, by \eqref{eq:bargS} and the properties of $\vartheta$,
\[ \GG(\vartheta \circ g(x), \bar g) = \GG(\vartheta \circ g(x), \vartheta( \bar g) ) < \GG(g(x), \bar g), \]
and
\[ \GG(g(x), \bar g) \geq \GG(g(x), \widetilde S) - \GG(\bar g, \widetilde S) > \frac{1}{16} s(\widetilde S). \]
Hence
\begin{equation}\label{eq:bddiff}
\int_{\partial B_1} \GG(g, \vartheta(g))^2 \leq 2\int_E \GG(g(x), \bar g)^2 + \GG\left(\vartheta\circ g(x), \bar g \right)^2 \leq 4 \int_E \GG(g(x), \bar g)^2 \leq C |E|^{\frac{2}{m-1} }.
\end{equation}
Recall \eqref{tmp:lbdist} and \eqref{tmp:clp3},
\[ s(\widetilde S) \geq (1-2\epsilon) s(S) \geq (1-2\epsilon) \beta M. \]
We may estimate the measure of $E$ by Chebyshev inequality
\begin{align*}
|E| \leq \int_{B_1} \left( \dfrac{\GG(g(x), \bar g) }{s(\widetilde S)/16 } \right)^2 \leq \frac{C}{M^2}.
\end{align*}
Combined with \eqref{tmp:cpt} and \eqref{eq:bddiff}, we conclude that
\[ \Dir(\hat f, B_1) \leq \frac{1}{m-2} - \gamma_0 + C' \eta + \frac{C'}{\eta M^2}, \]
where the constants $\gamma_0, C'$ only depend on $Q$ and $m$. We first choose $\eta$ so that $C'\eta = \frac{\gamma_0}{3}$, then we choose $M$ so that $\frac{C'}{\eta M^2} = \frac{\gamma_0}{3}$. Therefore by the minimality of $f$
\[ \Dir(f, B_1) \leq \Dir(\hat f, B_1) \leq \frac{1}{m-2} - \frac{\gamma_0}{3}. \]
\textit{Step 3. Conclusion.} With the value of $M$ fixed, Step 1 shows that if $d(\bar g) \leq M$, there exists $\gamma = \gamma(Q)>0$ such that
\[ \Dir(f, B_1) \leq \frac{1}{m-2} - \gamma. \]
Assuming the inductive hypothesis, Step 2 shows that if $d(\bar g) > M$,
\[ \Dir(f, B_1) \leq \frac{1}{m-2} - \frac{\gamma_0}{3}. \]
This concludes the proof.
\end{proof}
\subsection{Proof of Theorem \ref{thm:Holder}}
We want to prove the following decay of Dirichlet energy
\begin{equation}\label{eq:Dirbdrad}
\Dir (f, B_r) \leq C r^{m-2+2\beta} (\Dir (f, B_1) + \|D\varphi\|_{C^0}^2)
\end{equation}
for every $y \in B_{\frac12}$ and almost every $0<r \leq \frac12$.
First of all observe that the estimate follows from Proposition \ref{prop:Dirbd} for $y\in \gamma$.
Indeed in that case, if we let
$h(r) = \iint_{B_r(y)} |Df|^2$, then $h$ is absolutely continuous and
\[ h'(r) = \int_{\partial B_r(y)} |Df|^2 \geq \int_{\partial B_r(y)} |D_\tau f|^2 =: \Dir(f, \partial B_r(y)) \quad \text{ for almost every } r.
\]
Combined with \eqref{eq:Dirbd} we have
\[
h (r) \leq C(m) rh' (r) + C r^m \|D\varphi\|_{C^2}^2 \leq \frac{r h'(r)}{m-2+2\beta} + C r^m \|D\varphi\|_{C^0}^2\, ,
\]
(where $\beta$ is assumed to be smaller than $1$).
We next define $k (r) := h (r) + A r^m$ and compute
\begin{align*}
k (r) & = h (r) + A r^m \leq \frac{r}{m-2+2\beta} h' (r) + C \|D\varphi\|_{C^0}^2 r^m + A r^m\\
& \leq \frac{r}{m-2+2\beta} k'(r) + C \|D\varphi\|_{C^0}^2 r^m - A \left(\frac{m}{m-2+2\beta} -1\right) r^m\, .
\end{align*}
Since
\[
\frac{m}{m-2+2\beta} -1 > 0\, ,
\]
for $A = C' \|D\varphi\|_{C_0}^2$ with $C'$ sufficiently large we conclude
\[
k (r) \leq \frac{r}{m-2+2\beta} k'(r)\,
\]
and integrating the latter inequality in the interval $[r,1/2]$ we get the desired estimate
\begin{align*}
\Dir (f, B_r (y)) &\leq k (r) \leq r^{m-2+2\beta} k \left(\frac{1}{2}\right)
\leq r^{m-2+2\beta} \left( \Dir (f, B_{1/2} (y)) + C' \|D\varphi\|_{C^0}^2\right)\\
&\leq C r^{m-2+2\beta} \left( \Dir (f, B_1)+ \|D\varphi\|_{C^0}^2\right)\, .
\end{align*}
Consider now a point $y\in B_{1/2}\setminus \gamma$. If $r\geq \frac{1}{4}$ the estimate \eqref{eq:Dirbdrad} is then obvious. Hence we assume $r < \frac{1}{4}$. Let next $\rho:= \dist (y, \gamma)$. If $r\geq \rho$, consider $x\in \gamma$ such that $|x-y|= \dist (y, \gamma)$ and observe that $B_{2r} (x) \supset B_r (y)$. The estimate follows then from the one for $y\in \gamma$. Otherwise, we have two possibilities. If $\rho \geq \frac{1}{4} > r$, we then can use the decay estimate for $Q$-valued Dir-minimizers to infer
\[
\Dir (f, B_r (y)) \leq C r^{m-2+2\beta} \Dir (f, B_{1/4} (y)) \leq C r^{m-2+2\beta} \Dir (f, B_1)\, .
\]
If $r<\rho < \frac{1}{4}$ we can then proceed in two steps to prove
\[
\Dir (f, B_r (y)) \leq \left(\frac{r}{\rho}\right)^{m-2+2\beta} \Dir (f, B_\rho (y))
\leq C r^{m-2+2\beta} \left(\Dir (f, B_1) + \|D\varphi\|_{C^0}^2\right)\, .
\]
Having finally proved the decay \eqref{eq:Dirbdrad}, the H\"older continuity follows from the Campanato-Morrey estimate.
\section{First variations and monotonicity of the frequency function}\label{s:frequency}
In this section we address a main tool to prove Theorem \ref{thm:main}, the monotonicity of the frequency function. The original frequency function was introduced by Almgren in \cite{Alm} for Dir-minimizing $Q$-valued map, cf. also \cite{DS}. The one for $(Q-\frac{1}{2})$-valued maps with interface $(\gamma, 0)$ in $\mathbb R^m$ was introduced in \cite{DDHM} and requires a subtle argument. Since our Theorem \ref{thm:main} is $2$-dimensional, we can take advantage of the reduction to Theorem \ref{thm:main_simple} and restrict our attention to the model situation in which the interface is $(\RR, 0)$. Under such assumption the statement and proof of the relevant formulae is just a straightforward adaptation of the arguments in \cite{DS}, which we give below for the reader's convenience (the issue in \cite{DDHM} is that in dimension $m\geq 3$ it is not possible to ``rectify'' a general $\gamma$ with a conformal change of coordinates).
\begin{defn}[The frequency function]
Assume $f = (f^+, f^-)$ is a $(Q-\frac{1}{2})$-valued map on $\Omega \subset \RR^m$ with interface $(\gamma, \varphi)$ and consider a ball $B_r (x) \subset \Omega$ with $x\in \gamma$.
We define
\begin{equation}
D_{x,f}(r) = {\rm Dir}\, (f, B_r (x)) , \quad H_{x,f} (r) = \int_{\partial B_r (x)} |f|^2 := \int_{\partial B_r^+(x)}|f^+|^2 + \int_{\partial B_r^-(x)} |f^-|^2.
\end{equation}
When $H_{x,f}(r) > 0$, we define the frequency function
\begin{equation}
I_{x,f}(r) = \frac{r D_{x,f}(r)}{H_{x,f}(r)}.
\end{equation}
\end{defn}
When $x$ and $f$ are clear from the context, we often use the shorthand notation $D(r), H(r)$ and $I(r)$.
\begin{prop}[First variations] Assume $f=(f^+, f^-) \in W^{1,2}(\Omega, \AQt)$ is Dir-minimizing on $\Omega \subset \RR^2$ with interface $(\RR, 0)$ and let $B_r \subset \Omega$. Then
\begin{equation}\label{eq:innervar}
\int_{\partial B_r(x)}|Df|^2 = 2 \left( \int_{\partial B_r^+(x)} \sum_{j=1}^Q |\partial_\nu f_j^+|^2 + \int_{\partial B_r^-(x)} \sum_{j=1 }^{Q-1} |\partial_\nu f_j^-|^2 \right)
\end{equation}
and
\begin{equation}\label{eq:outervar}
\iint_{B_r(x)} |Df|^2 = \int_{\partial B_r^+(x)} \sum_{j=1}^Q \langle \partial_\nu f_j^+, f_j^+ \rangle + \int_{\partial B_r^-(x)} \sum_{j=1 }^{Q-1} \langle \partial_\nu f_j^-, f_j^- \rangle\, .
\end{equation}
Here $\nu$ denotes the outer unit normal on the boundary of the given ball, and $f^+ = \sum_{j=1}^Q \llbracket f_j^+ \rrbracket $ and $f^- = \sum_{j=1}^{Q-1} \llbracket f_j^- \rrbracket$ are measurable selections of $f^+$ and $f^-$, given by \cite[Proposition 0.4]{DS}.
\end{prop}
\begin{remark}
Identity \eqref{eq:innervar} implies that the integral of the square of the tangential derivative on the circle $\partial B_r$ equals the integral of the square of the normal derivative.
\end{remark}
\begin{proof}
The proof follows the same computations of \cite[Proof of Proposition 3.2]{DS}. It just suffices to observe the following two facts:
\begin{itemize}
\item \eqref{eq:innervar} is derived by comparing the Dirichlet energy of $f$ with competitors of the form $f\circ \Phi_\varepsilon$, where $\{\Phi_\varepsilon\}$ are some specific one-parameter families of diffeomorphisms. It easy to check that the ones used in \cite[Proof of Proposition 3.2]{DS} map $\RR$, $\Omega^+$ and $\Omega^-$ onto themselves and hence give an admissible family of competitors for our variational problem as well.
\item Similarly, \eqref{eq:outervar} is derived by comparing the Dirichlet energy of $f$ with competitors of the form
\[
f^{\varepsilon, \pm} (x) := \sum_j \a{f_j^\pm (x) + \varepsilon \psi (x, f_j^\pm (x))}
\]
where $\psi(x, u) = \phi(|x|) u$ satisfies $\psi(x, 0) = 0$. Therefore the functions $f^{\varepsilon, \pm}$ have also interface $(\RR, 0)$ and they are in the class of admissible competitors.\qedhere
\end{itemize}
\end{proof}
The above first variation formulae imply:
\begin{theorem}[Monotonicity of the frequency, analogue of Theorem 3.15 \cite{DS}]\label{thm:monot}
Let $f$ be a $(Q-\frac{1}{2})$-valued Dir-minimizing map with interface $(\RR, 0)$ in an open set $\Omega\subset \RR^2$ containing the origin and assume that $f^+ (0) = Q \a{0}$. Either there exists $\delta>0$ such that
\[ f^+|_{B^+_\delta(0)} \equiv Q\llbracket 0 \rrbracket, \quad f^-|_{B^-_\delta(0)} \equiv (Q-1) \llbracket 0 \rrbracket; \]
or $I_{0,f}(r)$ is an absolutely continuous nondecreasing positive function on $(0,\dist (0, \partial\Omega))$.
\end{theorem}
\begin{proof}
If $H(r) = 0$ for some $r>0$, then $f^+=Q \a{0}$ a.e. on $\partial B^+_r(0)$ and $f^- = (Q-1) \a{0}$ a.e. on $\partial B^-_r (0)$. For such boundary data the only minimizer is the pair which is constant on the respective $B^\pm_r (0)$. From now on we assume therefore that $H(r)>0$ for every $r\in (0,1)$.
$D$ is absolutely continuous and
\begin{equation}
D'(r) = \int_{\partial B_r} |Df|^2 \text{ for almost every } r.
\end{equation}
Since $f^+, f^- \in W^{1,2}$ are approximately differentiable almost everywhere, we can apply the chain-rule formulas (see \cite[Propositions 1.12 and 2.8]{DS}) and justify the following computations:
\begin{align}
H'(r) & = \frac{d}{dr} \int_{\partial B_1^+ } r |f^+ (ry)|^2 dy + \frac{d}{dr} \int_{\partial B_1^- } r |f^- (ry)|^2 dy \nonumber \\
& = \int_{\partial B_1} |f(ry)|^2 dy + \int_{\partial B_1^+} r \frac{\partial}{\partial r} |f^+(ry)|^2 dy + \int_{\partial B_1^-} r \frac{\partial}{\partial r} |f^-(ry)|^2 dy \nonumber \\
& = \frac{1}{r} \int_{\partial B_r} |f|^2 + 2 \int_{\partial B_r^+} \sum_{j=1}^Q \langle \partial_\nu f_j^+, f_j^+ \rangle + 2 \int_{\partial B_r^-} \sum_{j=1}^{Q-1} \langle \partial_\nu f_j^-, f_j^- \rangle
= \frac{1}{r} H(r) + 2 D(r),\label{eq:Hder}
\end{align}
by the outer variation formula \eqref{eq:outervar}. In fact, since both $H(r)$ and $D(r)$ are continuous, we have $H\in C^1$ and the above inequality holds pointwise. Therefore
\begin{align*}
I'(r) & = \frac{D(r)}{H(r)} + \frac{rD'(r)}{H(r)} - rD(r) \frac{H'(r)}{H(r)^2} \\
& = \frac{D(r)}{H(r)} + \frac{rD'(r)}{H(r)} - \frac{D(r)}{H(r)} - 2r \frac{D(r)^2}{H(r)^2} \\
& = \frac{rD'(r)}{H(r)} - 2r \frac{D(r)^2}{H(r)^2}.
\end{align*}
Again by the inner and outer variations formulae \eqref{eq:innervar}, \eqref{eq:outervar}, we obtain
\begin{align}
I'(r) = & \frac{2r}{H(r)^2} \left[ \left(\int_{\partial B^+_r} \sum_{j=1}^Q |\partial_\nu f_j^+|^2 + \int_{\partial B^-_r} \sum_{j=1}^{Q-1} |\partial_\nu f_j^-|^2 \right) \cdot \left( \int_{\partial B^+_r} \sum_{j=1}^Q | f_j^+|^2 + \int_{\partial B_r^-}\sum_{j=1}^{Q-1} | f_j^-|^2 \right) \right. \nonumber \\
& \qquad \left. - \left( \int_{\partial B^+_r} \sum_{j=1}^Q \langle \partial_\nu f_j^+, f_j^+ \rangle + \int_{\partial B^-_r}
\sum_{j=1}^{Q-1} \langle \partial_\nu f_j^-, f_j^- \rangle \right)^2 \right]\, .
\end{align}
We can now choose a measurable selection of the various multifuctions involved and extend such selections $f^\pm_j, \partial_\nu f^\pm_j$ to $0$ respectively on $B_r^\mp$. The Cauchy-Schwartz inequality will then imply:
\begin{align}
I'(r) = & \frac{2r}{H(r)^2} \left[ \int_{\partial B_r} \left(\sum_{j=1}^Q |\partial_\nu f_j^+|^2 + \sum_{j=1}^{Q-1} |\partial_\nu f_j^-|^2 \right) \cdot \int_{\partial B_r} \left( \sum_{j=1}^Q | f_j^+|^2 + \sum_{j=1}^{Q-1} | f_j^-|^2 \right) \right. \nonumber \\
& \qquad \left. - \left( \int_{\partial B_r} \sum_{j=1}^Q \langle \partial_\nu f_j^+, f_j^+ \rangle + \sum_{j=1}^{Q-1} \langle \partial_\nu f_j^-, f_j^- \rangle \right)^2 \right] \geq 0\, . \label{eq:monofreq}
\end{align}
\end{proof}
\begin{corollary}\label{cor:hom}
Let $f$ be as in Theorem \ref{thm:monot}. $I_{0,f}(r) \equiv \alpha$ if and only if $(f^+, f^-)$ is $\alpha$-homogeneous, i.e.
\[ f^{\pm} (x) = \sum_i \a{|x|^\alpha f^\pm_i \left(\frac{x}{|x|}\right)}\, . \]
\end{corollary}
In the rest of the note, when dealing with a $Q$-valued funcion $f = \sum_i \a{f_i}$, we will use the notation $\lambda f$ for the multifunction $\sum_i \a{\lambda f_i}$. Similarly, for a $(Q-\frac{1}{2})$-valued function $f= (f^+, f^-)$, $\lambda f$ will denote $(\lambda f^+, \lambda f^-)$. In particular, the homogeneity of $f$ in the corollary above will be expressed by the formula
\[
f (x) = |x|^\alpha f \left(\frac{x}{|x|}\right)\, .
\]
\begin{proof}
Suppose $\alpha=0$. $I_{0,f}(r) \equiv 0$ if and only if each $f_j^{\pm}$ is constant, so clearly $(f^+, f^-)$ is $0$-homogeneous. If $(f^+, f^-)$ is $0$-homogeneous, then each $f_j^{\pm}$ is constant on the ray starting from the origin. Thus by the continuity of $f$ near the origin, each $f_j^{\pm}$ is constant on its domain and $I_{0,f}(r) \equiv 0$.
Suppose $\alpha > 0$. Then by the proof of the above theorem, $I(r)$ is a constant if and only if equality occurs in \eqref{eq:monofreq}, i.e. if and only if there exists constants $\lambda_r \in \RR$ such that
\[ \partial_\nu f_j^{\pm}(x) = \lambda_r f_j^{\pm}(x) \text{ for almost every } r \text{ and almost every } x \text{ with } |x| = r. \]
Moreover,
\[ \alpha = I(r) = \frac{rD(r)}{H(r)} = \frac{r\int_{\partial B_r} \sum \langle \partial_\nu f_j^{\pm}, f_j^{\pm} \rangle }{ \int_{\partial B_r} \sum |f_j^{\pm}|^2} = r \lambda_r. \]
Therefore $I(r) \equiv \alpha$ if and only if
\begin{equation}\label{eq:tmphom}
\partial_\nu f_j^{\pm}(x) = \frac{\alpha}{|x|} f_j^{\pm}(x) \text{ for almost every } x.
\end{equation}
If $f$ is $\alpha$-homogeneous, clearly \eqref{eq:tmphom} holds. On the other hand, suppose \eqref{eq:tmphom} holds, we want to show that $f$ is $\alpha$-homogeneous. To that end we let $x\in \partial B_1$ and $\sigma_x = \{rx: 0< r\leq 1\}$ be the ray from the origin through $x$. Note that $f|_{\sigma_x}\in W^{1,2}$ for almost every $x\in \partial B_1$. For those $x$ \eqref{eq:tmphom} implies
\[ \frac{d}{dr} \frac{f_j^{\pm}(rx)}{r^{\alpha}} \equiv 0. \]
Thus $f_j^{\pm}(rx) = r^{\alpha} f_j^{\pm} (x)$ for all $0<r\leq 1$ and almost every $x\in \partial B_1$.
\end{proof}
\begin{corollary}[analogue of Corollary 3.18 \cite{DS}]
Let $f$ be as in Theorem \ref{thm:monot}. Suppose $H(r) \neq 0$ for every $r\in (0, \dist (\partial \Omega , 0))$. Then
\begin{enumerate}[(i)]
\item for almost every $r < 1$,
\begin{equation}\label{eq:tmpH}
\frac{d}{dr} \log \left( \frac{H(r)}{r^{m-1}} \right) = \frac{2I(r)}{r },
\end{equation}
and thus for almost every $s\leq t < 1$,
\begin{equation}\label{eq:tmpH2}
\left( \frac{s}{t} \right)^{2I(t)} \frac{H(t)}{t^{m-1}} \leq \frac{H(s)}{s^{m-1}} \leq \left( \frac{s}{t} \right)^{2I(s)} \frac{H(t)}{t^{m-1}};
\end{equation}
\item for almost every $s\leq t<1$, if $I(t)>0$, then
\begin{equation}\label{eq:Dratio}
\frac{I(s)}{I(t)} \left(\frac{s}{t} \right)^{2I(t)} \frac{D(t)}{t^{m-2}} \leq \frac{D(s)}{s^{m-2}} \leq \left(\frac{s}{t} \right)^{2I(s)} \frac{D(t)}{t^{m-2}}.
\end{equation}
\end{enumerate}
\end{corollary}
\begin{proof}
\eqref{eq:tmpH} follows from \eqref{eq:Hder}. \eqref{eq:tmpH2} follows from \eqref{eq:tmpH} and the monotonicity of the frequency function. Finally, \eqref{eq:Dratio}
follows from \eqref{eq:tmpH2} and the definition of the frequency function.
\end{proof}
\section{Compactness and tangent functions in planar domains}\label{s:rate}
The monotonicity of the frequency function provides a way of studying the asymptotic behavior of a minimizer at small scales around a given point with highest multiplicity. In what follows we will often switch between different systems of coordinates in the plane. More precisely, a point $x\in \mathbb R^2$ will be identified with
\begin{itemize}
\item the pair $(x_1, x_2)$ which gives the standard Cartesian coordinates of $x$;
\item the complex number $z = x_1 + i x_2$;
\item the pair $(r, \theta)$ which gives the standard polar coordinates of $x$, namely $x_1 = r \cos \theta$, $x_2 = r \sin \theta$ and $z = r e^{i\theta}$.
\end{itemize}
\begin{defn}
Let $f$ be a Dir-minimizing $(Q-\frac{1}{2})$-valued map on a planar domain $\Omega$ with interface $(\RR, 0)$. Let $y$ be a point at the interface $\RR$ and assume that $\Dir(f,B_\rho (y)) >0$ for every $\rho$. We define the following rescalings of $f$ at $y$:
\begin{equation}\label{eq:defblowup}
f_{y,\rho} (x) = \frac{f(\rho x+y)}{\sqrt{\Dir(f,B_\rho(y))}}.
\end{equation}
\end{defn}
The key point is that, up to subsequences, the latter rescalings converge locally strongly to nontrivial Dir-minimizers.
\begin{theorem}[Compactness, analogue of Theorem 3.19 in \cite{DS}]\label{thm:tangent}
Let $f\in W^{1,2}(\Omega, \AQt)$ be a Dir-minimizing map on a planar domain $\Omega$ with interface $(\RR, 0)$. Assume $f^+(0) = Q\llbracket 0 \rrbracket$ and $\Dir(f, B_{\rho}) >0$ for every $\rho \in (0, \dist (0, \partial \Omega))$. Then for any sequence $\{f_{\rho_k}\}$ with $\rho_k \searrow 0$, a subsequence, not relabelled, converges locally uniformly to a function $g:\RR^2 \to \AQt$ satisfying the following properties:
\begin{enumerate}[(a)]
\item $\Dir(g,B_1) = 1$ and $g|_{U}$ is Dir-minimizing with interface $(\RR, 0)$ for any bounded set $U \subset \RR^2$;
\item $g(x) = |x|^\alpha g\left( \frac{x}{|x|} \right)$, where $\alpha = I_{0,f}(0) >0$ is the frequency of $f$ at $0$.
\end{enumerate}
\end{theorem}
From now, any limit of a sequence of rescalings $\{f_{\rho_k}\}_k$ with $\rho_k\downarrow 0$ will be called a {\em tangent function}.
A feature of the $2$-dimensional setting is that the compactness theorem above can be considerably strengthened: analogously to the ``interior case'', cf. \cite[Theorem 5.3]{DS}, we can prove that the tangent function at a given point is unique and that the rescaling converge at a suitable rate to it. The key is to first show a suitable rate of convergence for the frequency function.
\begin{prop}[Rate of convergence, analogue of Proposition 5.2 in \cite{DS}]\label{prop:rate}
Let $f$ be as in Theorem \ref{thm:tangent} and set $\alpha = I_{0,f}(0)$.
Then there exist constants $r_0, \beta, C, H_0, D_0 >0$ such that for every $0<r\leq r_0$,
\begin{equation}\label{eq:decayfr}
0\leq I(r) - \alpha \leq C r^\beta,
\end{equation}
\begin{equation}\label{eq:decayDe}
0\leq \frac{H(r)}{r^{2\alpha+1}} - H_0 \leq C r^\beta, \quad 0 \leq \frac{D(r)}{r^{2\alpha}} - D_0 \leq C r^\beta.
\end{equation}
\end{prop}
\begin{theorem}[Unique tangent map, analogue of Theorem 5.3 \cite{DS}]\label{thm:cvtang}
Let $f\in W^{1,2}(\DD, \AQt)$ be as in Theorem \ref{thm:tangent} and denote by $\beta$ the exponent of the decay estimate \eqref{eq:decayfr}. Then the tangent function $f_0$ to $f$ at $0$ is unique and, moreover,
\begin{equation}\label{e:decay}
\|\mathcal{G} (f_{0, \rho}, f_0)\|^2_{L^2 (\mathbb S^1)}\leq C \rho^\beta\, .
\end{equation}
\end{theorem}
\subsection{Compactness and tangent functions: Proof of Theorem \ref{thm:tangent}}
Let $B_R$ denote a ball of sufficiently large radius $R\gg 1$. By the definition \eqref{eq:defblowup},
\[ \Dir(f_\rho, B_R) = \frac{D(\rho R)}{D(\rho)} \leq R^{m-2+2I(\rho R)} \frac{I(\rho)}{I(\rho R)} \leq R^{m-2+2(\alpha+1)}, \]
where we use the estimate \eqref{eq:Dratio} for the first inequality, and the properties of the frequency function for the second. Since $f_\rho$'s are all Dir-minimizing with interface $(\RR, 0)$, by Theorem \ref{thm:Holder} they are locally equi-H\"older continuous. The assumption $f^+(0) = Q\llbracket 0\rrbracket$ implies $f^+_\rho(0) = Q \llbracket 0 \rrbracket$ and $f^-_{\rho}(0) = (Q-1) \llbracket 0 \rrbracket$ for all $\rho$. Thus $f_\rho$'s are locally uniformly bounded, and by Arzel\`a-Ascoli Theorem a subsequence (not relabelled) converges uniformly on compact sets to a continuous function $g=(g^+, g^-)$. (To use Arzel\`a-Ascoli Theorem, we may add to $f^-_\rho$ a zero sheet to get functions valued in the metric space $\AQ$.) In particular $g^+|_{\gamma} = g^-|_{\gamma} + \llbracket 0 \rrbracket$; moreover $f^+_{\rho_k}$ converges weakly in $W^{1,2}_{loc} (\RR^2_+, \AQ)$ to $g^+$, and $f^-_{\rho_k}$ converges weakly in $W^{1,2}_{loc} (\RR^2_-, \mathcal{A}_{Q-1})$ to $g^-$ (see \cite[Definition 2.9]{DS}). By \eqref{eq:repD} it follows then that $\Dir(g, B_r) \leq \liminf_{k\to\infty} \Dir(f_{\rho_k}, B_r)$ for all $r>0$.
\textit{Proof of (a)}. Let $R>0$ be fixed. We will show that for any $0<r\leq R$,
\begin{equation}\label{eq:limitmin}
\Dir(g, B_r) = \liminf_{k\to\infty} \Dir(f_{\rho_k}, B_r) \text{ and } g|_{B_r} \text{ is Dir-minimizing with interface } (\RR, 0).
\end{equation}
For any $R>0$, we will show \eqref{eq:limitmin} holds for all $r\leq R$.
By Fatou's Lemma,
\[ \int_0^R \liminf_{k\to\infty} \Dir(f_{\rho_k},\partial B_r) \, dr \leq \liminf_{k\to\infty} \int_0^R \Dir(f_{\rho_k}, \partial B_r) \, dr \leq \liminf_{k\to\infty} \Dir(f_{\rho_k}, B_R) \leq C<+\infty. \]
Hence for almost every $r\in (0,R)$, the integrand of the first term is finite, and moreover by weak convergence
\begin{equation}\label{eq:Dbdshell}
\Dir(g,\partial B_r) \leq \liminf_{k\to\infty} \Dir(f_{\rho_k},\partial B_r) \leq M < +\infty.
\end{equation}
For the sake of contradiction, assume that either one of the statement in \eqref{eq:limitmin} fails for such $r$, then there exists a function $h\in W^{1,2}(B_r, \AQt)$ with interface $(\RR, 0)$ such that
\begin{equation}\label{eq:tmpcontra}
h|_{\partial B_r} = g|_{\partial B_r} \text{ and } \Dir(h, B_r) < \liminf_{k\to\infty} \Dir(f_{\rho_k}, B_r).
\end{equation}
Let $\delta = 1/N <1/2$ to be fixed later, and consider the functions $\widetilde f_{k}$ on $B_r$ defined by
\[ \widetilde f_k = \left\{\begin{array}{ll}
\left( \frac{1}{1-\delta} \right)^{\frac{m-2}{2}} h\left( \frac{x}{1-\delta} \right), & \text{ for } x\in B_{(1-\delta)r}, \\
h_k(x), & \text{ for } x\in B_r \setminus B_{(1-\delta)r},
\end{array} \right. \]
where the $h_k$'s are the interpolation functions provided by Lemma \ref{lm:interpolation} between the maps $f_{\rho_k}\in W^{1,2}(\partial B_r, \AQt)$ and $h\left(\frac{x}{1-\delta} \right)\in W^{1,2}(\partial B_{(1-\delta)r}, \AQt)$. Notice that $h_k$'s satisfy the boundary condition $h_k^+|_{\gamma} = h_k^-|_{\gamma} + \llbracket 0 \rrbracket$. By the minimality of $f_{\rho_k}$, \eqref{eq:interpolation} and changes of variables, we have
\begin{align*}
\Dir(f_{\rho_k}, B_r) & \leq \Dir(\widetilde f_k, B_r) \\
& \leq \Dir\left( \left( \frac{1}{1-\delta} \right)^{\frac{m-2}{2}} h\left( \frac{x}{1-\delta} \right), B_{(1-\delta)r} \right) + \Dir(h_k, B_r \setminus B_{(1-\delta)r}) \\
& \leq \Dir(h, B_{r}) + C\delta r \Dir(f_{\rho_k}, \partial B_r) + C \delta r \Dir(h, \partial B_r) + \frac{C}{\delta} \sup_{x\in \partial B_r} \GG(h(x), f_{\rho_k}(x)) \\
& \leq \Dir(h, B_r) + C\delta R \Dir(f_{\rho_k}, \partial B_r) + C\delta R \Dir(g, \partial B_r) + \frac{C}{\delta} \sup_{x\in \partial B_r} \GG(g(x), f_{\rho_k}(x)).
\end{align*}
Passing $k\to \infty$, by the uniform convergence of $f_{\rho_k}$ to $g$, \eqref{eq:Dbdshell} and the assumption \eqref{eq:tmpcontra}, we get
\begin{equation}
\limsup_{k\to\infty} \Dir(f_{\rho_k}, B_r) \leq \Dir(h,B_r) + 2C\delta RM < \liminf_{k\to \infty} \Dir(f_{\rho_k}, B_r) + 2C\delta RM.
\end{equation}
We get a contradiction by choosing $\delta$ arbitrarily small. Therefore \eqref{eq:limitmin} holds for almost every $r\in (0,R)$. By the upper semi-continuity of $\Dir(g, B_r)$ in $r$, it follows that \eqref{eq:limitmin} holds for all $r\leq R$.
\textit{Proof of (b).} We observe that for every $r>0$,
\[ I_g(r) = \frac{r \Dir(g, B_r)}{\int_{\partial B_r} |g|^2} = \liminf_{k\to \infty} \frac{r \Dir(f_{\rho_k}, B_r)}{\int_{\partial B_r} |f_{\rho_k}|^2} = \liminf_{k\to \infty} \frac{r \rho_k \Dir(f, B_{r \rho_k})}{\int_{\partial B_{r\rho_k}} |f|^2} = I_f(0). \]
Since $g$ is Dir-minimizing, by Corollary \ref{cor:hom} it is a homogeneous function with homogenity $\alpha = I_f(0)$. If $\alpha=0$, a continuous $0$-homogeneous function with $g(0) = Q\llbracket 0 \rrbracket$ is necessarily $g \equiv Q\llbracket 0 \rrbracket$. This is in contradiction with $\Dir(g, B_1) = \lim_k \Dir (f_{\rho_k}, B_1) = 1$, and thus $\alpha>0$.
\subsection{Rate of convergence: Proof of Proposition \ref{prop:rate}}
\textit{Step 1.} We claim the following estimate holds for some $\beta>0$:
\begin{equation}\label{eq:IODE}
I'(r) \geq \frac{2}{r} \left(\alpha + \beta - I(r) \right) (I(r) - \alpha).
\end{equation}
Recall \eqref{eq:Hder}, we have
\begin{equation}
I'(r) = \frac{rD'(r)}{H(r)} - \frac{2I^2(r)}{r}.
\end{equation}
Thus \eqref{eq:IODE} is reduced to prove
\begin{equation}\label{eq:dcclaim}
(2\alpha+\beta) D(r) \leq \frac{r D'(r)}{2} + \frac{\alpha(\alpha+\beta)H(r)}{r}.
\end{equation}
Let $r$ be fixed, and let $g(\theta) := f(r e^{i\theta})$. Consider the decomposition of $g(\theta)$ as in Proposition $g= g_0 + \sum_{j=1}^J g_j$, where $g_0 \in W^{1,2}(\So, \mathcal{A}_{Q_0}^{\pm})$ and $g_j \in W^{1,2}(\So, \mathcal{A}_{Q_j})$ are irreducible maps. Recall that for each irreducible $g_j$, we can find $\zeta_j\in W^{1,2}(\So, \RR^n)$ such that
\begin{equation}
g_j(\theta) = \sum_{i=1}^{Q_j} \left\llbracket \zeta_j\left( \frac{\theta + 2\pi i}{Q_j} \right) \right\rrbracket.
\end{equation}
We write the Fourier expansions of $\zeta_j$'s as
\begin{equation}
\zeta_j(\theta) = \frac{a_{j,0}}{2} + \sum_{l=1}^\infty \left(a_{j,l} \cos(l\theta) + b_{j,l} \sin(l\theta) \right), \quad \theta\in [0,2\pi].
\end{equation}
Suppose $g_0$ unrolls to a function $\zeta_0: \So \to \RR^n$, as in Lemma \ref{lm:unrolling}. The boundary condition $g_0^+|_{\gamma} = g_0^-|_{\gamma} + \llbracket 0 \rrbracket$ implies $\zeta_0(0) = \zeta_0(2\pi) = 0$. Hence we write the Fourier expansion of $\zeta_0$ as
\[
\zeta_0(\theta) = \sum_{l=1}^\infty c_l \sin\left( \frac{ l\theta}{2} \right), \quad \theta \in [0,2\pi].
\]
Recall \eqref{eq:innervar} and Lemma \ref{lm:unrolling}, we get
\begin{align}
D'(r) = 2 \Dir(f, \partial B_r) & = \frac{2}{r} \Dir(g, \So)
= \frac{2}{r} \sum_{j=0}^{J} \Dir(g_j, \So) \nonumber \\
& = \frac{2}{r} \left( \frac{2}{2Q_0-1} \Dir(\zeta_0, \So) + \sum_{j=1}^{J} \frac{1}{Q_j} \Dir(\zeta_j, \So) \right) \nonumber \\
& = \frac{2\pi}{r} \left( \frac{1}{2(2Q_0-1)} \sum_{l} c_l^2 l^2 + \sum_{j=1}^{J} \sum_l \frac{\left( a_{j,l}^2 + b_{j,l}^2 \right) l^2 }{Q_j}\right),\label{eq:dc1}
\end{align}
and
\begin{align}
H(r) &= \int_{\partial B_r} |f|^2 = r \int_{\So} |g|^2
= r \left( \frac{2Q_0-1}{2} \int_{\So} |\zeta_0|^2 + \sum_{j=1}^J Q_j \int_{\So} |\zeta_{j}|^2 \right) \nonumber \\
& = \pi r \left[ \frac{2Q_{0}-1}{2} \sum_l c_l^2 + \sum_{j=1}^J Q_j \left( \frac{a_{j,0}^2}{2} + \sum_l \left( a_{j,l}^2 + b_{j,l}^2 \right) \right) \right]. \label{eq:dc2}
\end{align}
On the other hand, consider a $W^{1,2}$-extension of $\zeta_0$ on $B_1$ \begin{equation}\label{eq:bdsheetext}
\overline \zeta_0(\rho, \theta) = \sum_{l=1}^\infty \rho^{\frac{l}{2}} c_l \sin \left( \frac{l\theta}{2} \right),
\end{equation}
(note that $\overline \zeta_0$ is not harmonic at the origin), and the harmonic extension of each $\zeta_j$:
\[ \overline \zeta_j(\rho, \theta) = \frac{a_{j,0}}{2} + \sum_{l=1}^\infty \rho^l \left(a_{j,l} \cos(l\theta) + b_{j,l} \sin(l\theta) \right), \quad j=1, \cdots, J, \]
where $ \rho\leq 1 \text{ and } \theta\in [0,2\pi]$.
We then construct a competitor $h$ of $f$ in $B_r$ as follows: $h(\rho e^{i\theta}) = \widehat h(\rho e^{i\theta}/r)$, where $\widehat h$, a function defined on $B_1$, is given by
\[
\widehat h(\rho e^{i\theta}) := \left( h_0^+(\rho e^{i\theta} ), h_0^-(\rho e^{i\theta}) \right) + \sum_{j=1}^J \sum_{k=0}^{Q_j} \left\llbracket \overline \zeta_j \left( \rho^{\frac{1}{Q_j}}, \frac{ \theta+ 2\pi k}{Q_j} \right) \right\rrbracket,
\]
and $(h_0^+, h_0^-)$ is a $\left(Q_0 - \frac12 \right)$-valued function defined by
\[
h_0^+(\rho e^{i\theta}) = \overline \zeta_0 \left(\rho^{\frac{2}{2Q_0-1}}, \frac{2\theta}{2Q_0-1} + \frac{4\pi}{2Q_0-1} (k-1) \right), \quad \theta\in [0,\pi], k=1,\cdots,Q_0,
\]
\[
h_0^-(\rho e^{i\theta}) = \overline \zeta_0 \left(\rho^{\frac{2}{2Q_0-1}}, \frac{2\theta}{2Q_0-1} + \frac{4\pi}{2Q_0-1} (k-1) \right), \quad \theta\in [\pi, 2\pi], k=1,\cdots,Q_0-1.
\]
Notice that by definition \eqref{eq:bdsheetext},
\[ h_0^+|_{\gamma} = h_0^-|_{\gamma} + \llbracket 0 \rrbracket, \]
hence $\widehat h$ has interface $(\gamma,\varphi)$.
A simple computation, combined with Lemma \ref{lm:unrolling} and \cite[Lemma 3.12]{DS}, shows
\begin{align*}
\Dir(h, B_r) = \Dir(\widehat h, B_1) = \sum_{j=0}^J \iint_{B_1} |D \overline\zeta_j|^2 = \frac{\pi}{2} \sum_l c_l^2 l + \pi \sum_{j=1}^J \sum_l (a_{j,l}^2 + b_{j,l}^2 ) l.
\end{align*}
Thus by the minimality of $f$, we know
\begin{equation}\label{eq:dc3}
D(r) \leq \Dir(h, B_r) = \frac{\pi}{2} \sum_l c_l^2 l + \pi \sum_{j=1}^J \sum_l (a_{j,l}^2 + b_{j,l}^2 ) l.
\end{equation}
We combine \eqref{eq:dc1}, \eqref{eq:dc2} and \eqref{eq:dc3} and plug into the left and right hand sides of \eqref{eq:dcclaim}. After simplifications, we show that it is enough to find $\beta>0$ satisfying
\begin{equation}
\left[l-\alpha(2Q_0-1) \right] \left[ l-(\alpha+\beta)(2Q_0-1) \right] \geq 0,
\end{equation}
\begin{equation}
\left(l-\alpha Q_j \right) \left[ l-(\alpha+\beta)Q_j \right] \geq 0, \quad j=1, \cdots, J,
\end{equation}
for every $l\in \NN$.
This is equivalent to say the intervals $(\alpha(2Q_0-1), (\alpha+\beta)(2Q_0-1) )$ and $(\alpha Q_j, (\alpha+\beta) Q_j )$ do not contain integer points. This is verified, if we choose
\[ \beta = \min_{1\leq k\leq Q}\left\{ \frac{\lfloor\alpha k \rfloor + 1 - \alpha k }{k}, \frac{\lfloor\alpha (2k-1) \rfloor + 1 - \alpha (2k-1) }{2k-1} \right\}>0. \]
\textit{Step 2.}
Since $I(r)$ is monotone decreasing, there exists $r_0>0$ such that $I(r) \leq \alpha + \frac{\beta}{2}$ for all $r\leq r_0$. Hence \eqref{eq:IODE} implies that
\begin{equation}
I'(r) \geq \frac{\beta}{r} \left( I(r) - \alpha \right) \text{ for almost every } r\leq r_0.
\end{equation}
Integrating the differential inequality, we get the desired estimate
\begin{equation}\label{eq:tmpIccl}
I(r) - \alpha \leq \left( \frac{r}{r_0} \right)^\beta \left( I(r_0) - \alpha \right) \leq Cr^\beta.
\end{equation}
Recall that the derivative of $H$ satisfies \eqref{eq:Hder}. In particular when $m=2$ we have
\[ \left( \frac{H(r)}{r} \right)' = \frac{2D(r)}{r}. \]
This implies
\begin{equation}\label{eq:Hderfiner}
\left( \log \frac{H(r)}{r^{2\alpha+1}} \right)' = \left( \log \frac{H(r)}{r} - \log r^{2\alpha} \right)' = \left(\log \frac{H(r)}{r} \right)' - \frac{2\alpha}{r} = \frac{2}{r} \left( I(r) - \alpha \right) \geq 0.
\end{equation}
Hence $\frac{H(r)}{r^{2\alpha+1}}$ is monotone increasing. In particular, its limit exists as $r\to 0+$:
\begin{equation}
\frac{H(r)}{r^{2\alpha+1}} \geq \lim_{r\to 0} \frac{H(r)}{r^{2\alpha+1}} =: H_0 \geq 0.
\end{equation}
On the other hand combining \eqref{eq:Hderfiner} and \eqref{eq:tmpIccl} we get
\[ \left( \log \frac{H(r)}{r^{2\alpha+1}} \right)' \leq 2C r^{\beta-1}, \quad \text{ thus} \left( \log \frac{H(r)e^{-C_{\beta} r^\beta} }{r^{2\alpha+1}} \right)' \leq 0.\]
Hence $\frac{H(r)e^{-C_{\beta} r^\beta} }{r^{2\alpha+1}}$ is monotone decreasing, and
\[ \frac{H(r)e^{-C_{\beta} r^\beta} }{r^{2\alpha+1}} \leq \lim_{r\to 0} \frac{H(r)e^{-C_{\beta} r^\beta} }{r^{2\alpha+1}} = \lim_{r\to 0} \frac{H(r) }{r^{2\alpha+1}} = H_0. \]
In particular $H_0>0$. Moreover
\[ \frac{H(r)}{r^{2\alpha+1}} \left( 1-C_\beta r^\beta \right) \leq \frac{H(r)e^{-C_{\beta} r^\beta} }{r^{2\alpha+1}} \leq H_0, \]
and we conclude that
\[ \frac{H(r)}{r^{2\alpha+1}} - H_0 \leq C_\beta \, \frac{H(r)}{r^{2\alpha+1}} \, r^\beta \leq C_\beta \frac{H(r_0)}{r_0^{2\alpha+1}} r^\beta \leq C r^\beta. \]
The last inequality of \eqref{eq:decayDe} follows from:
\[ \frac{D(r)}{ r^{2\alpha} } - D_0 = \left( I(r) - I_0 \right) \frac{H(r)}{r^{2\alpha+1}} + I_0 \left( \frac{H(r)}{ r^{2\alpha+1} } - H_0 \right), \]
where $I_0 = \alpha$ and $D_0 = I_0 H_0$.
\subsection{Uniqueness of the tangent map: Proof of Theorem \ref{thm:cvtang}}
Without loss of generality, we assume $D_0 = 1$. By the estimate \eqref{eq:decayDe} and the definition of blow-up, it follows that
\begin{equation}\label{eq:blowupscal}
f_{\rho}(r,\theta) = \rho^{-\alpha} f(r\rho,\theta) \left( 1+ O\left(\rho^{\beta/2}\right) \right).
\end{equation}
It suffices to show the existence of a uniform limit for the ``dominant'' function
\[
h_{\rho}(r,\theta) = \rho^{-\alpha} f(r\rho,\theta)\, .
\]
Note that the function $h_{\rho}$ is homogeneous:
\[ h_{\rho}(r,\theta) = \rho^{-\alpha} f(r\rho,\theta) = r^{\alpha} h_{r\rho}(1,\theta), \]
it is enough to prove the existence of a uniform limit for $h_{\rho}|_{\So}$. Each function
\[ h_{\rho}|_{\So} = h_{\rho}(1,\theta) = \rho^{-\alpha} f(\rho,\theta) \]
is Dir-minimizing, so by Theorem \ref{thm:Holder} it is H\"older continuous with a uniform constant. We first show that $h_{\rho}|_{\So}$ has an $L^2$ limit.
Let $\{T_i\}$ and $\{T'_i\}$ be countable dense subsets of $\AQ(\RR^n)$ and $\mathcal{A}_{Q-1}(\RR^n)$ respectively.
We consider $r/2 \leq s\leq r$ and estimate
\begin{align*}
\int_0^{2\pi} \GG(h_r,h_s)^2 d\theta & = \int_0^{2\pi} \GG \left( \frac{f(r,\theta)}{r^{\alpha}}, \frac{f(s,\theta)}{s^{\alpha}} \right)^2 d\theta \\
& = \int_0^{\pi} \sup_i \left| \GG \left( \frac{f^+(r,\theta)}{r^{\alpha}}, T_i \right) - \GG \left( \frac{f^+(s,\theta)}{s^{\alpha}}, T_i \right) \right| ^2 d\theta \, + \\
& \qquad \int_{\pi}^{2\pi} \sup_i \left| \GG \left( \frac{f^-(r,\theta)}{r^{\alpha}}, T'_i \right) - \GG \left( \frac{f^-(s,\theta)}{s^{\alpha}}, T'_i \right) \right| ^2 d\theta \\
& \leq (r-s) \int_0^{\pi} \sup_i \int_s^r \left|\frac{\partial}{\partial t} \GG \left(\frac{f^+(t,\theta)}{t^{\alpha}}, T_i \right) \right|^2 \, dt \, d\theta \, + \\
& \qquad (r-s) \int_{\pi}^{2\pi} \sup_i \int_s^r \left|\frac{\partial}{\partial t} \GG \left(\frac{f^-(t,\theta)}{t^{\alpha}}, T'_i \right) \right|^2 \, dt \, d\theta \\
& \leq (r-s) \int_0^{\pi} \int_s^r \left|\frac{\partial}{\partial t} \left(\frac{f^+(t,\theta)}{t^{\alpha}} \right) \right|^2 \, dt \, d\theta \, +\\
&\qquad (r-s) \int_0^{\pi} \int_s^r \left|\frac{\partial}{\partial t} \left(\frac{f^-(t,\theta)}{t^{\alpha}} \right) \right|^2 \, dt \, d\theta \\
& = (r-s) \int_0^{\pi} \int_s^r \sum_j \left\{ \alpha^2 \frac{|f^+_j|^2}{t^{2\alpha+2}} + \frac{|\partial_\nu f^+_j|^2}{t^{2\alpha}} - 2\alpha \frac{\langle \partial_\nu f^+_j, f^+_j \rangle}{ t^{2\alpha+1}} \right\} \, dt\, d\theta \, + \\
& \qquad (r-s) \int_0^{\pi} \int_s^r \sum_j \left\{ \alpha^2 \frac{|f^-_j|^2}{t^{2\alpha+2}} + \frac{|\partial_\nu f^-_j|^2}{t^{2\alpha}} - 2\alpha \frac{\langle \partial_\nu f^-_j, f^-_j \rangle}{ t^{2\alpha+1}} \right\} \, dt\, d\theta \\
& = (r-s) \int_s^r \left\{ \alpha^2 \frac{H(t)}{t^{2\alpha+3}} + \frac{D'(t)}{2t^{2\alpha+1}} - 2\alpha \frac{D(t)}{t^{2\alpha+2}} \right\} dt \\
& = (r-s) \int_s^t \left\{ \frac{1}{2t} \left( \frac{D(t)}{t^{2\alpha}} \right)' + \alpha \frac{H(t)}{2t^{2\alpha+3}} \left(\alpha - I_{0,f}(t) \right) \right\} dt \\
& \leq (r-s) \int_s^t \frac{1}{2t} \left( \frac{D(t)}{t^{2\alpha}} - D_0 \right)' dt\, \\
& = (r-s) \left[ \frac{1}{2r} \left( \frac{D(r)}{r^{2\alpha}} - D_0 \right) - \frac{1}{2s} \left( \frac{D(s)}{s^{2\alpha}} - D_0 \right) \right] + \\
&\qquad (r-s) \int_s^t \frac{1}{2t^2} \left( \frac{D(t)}{t^{2\alpha}} - D_0 \right) dt \\
& \leq Cr^{\beta}.
\end{align*}
Let $s\leq r$ be arbitrary, and let $l$ be a positive integer such that $r/2^{l+1} <s \leq r/2^l$. Iterating the above estimate we get
\begin{align*}
\|\GG(h_r,h_s)\|_{L^2(\So)} \leq \sum_{k=0}^l \|\GG\left(h_{r/2^k}, h_{r/2^{k+1}} \right) \|_{L^2(\So)} + \|\GG(h_{r/2^l}, h_s) \|_{L^2(\So)} \leq \sum_{k=0}^l C\left( \frac{r}{2^k} \right)^{\frac{\beta}{2}} \leq C' r^{\frac{\beta}{2}}.
\end{align*}
This shows that $h_{\rho}|_{\So}$ is a Cauchy sequence in $L^2$, and thus converges to a limit function $h\in L^2(\So)$. Moreover, since $h_{\rho}$ is equi-H\"older continuous on $\So$, it follows that $h_{\rho}$ converges uniformly to $h$ on $\So$. In other words, $f_{\rho}$ converges locally uniformly to an $\alpha$-homogeneous function $g$, with $g(z) = |z|^\alpha h\left( \frac{z}{|z|} \right)$.
\section{Homogeneous Dir-minimizers}\label{s:classification}
In this section we study homogeneous Dir-minimizers in planar domains. We do not really give a complete characterization, but rather a set of necessary conditions that they have to satisfy. However, we will show below that each ``irreducible homogeneous piece'' in the classification is in fact minimizing.
\begin{prop}[Characterization of tangent maps]\label{prop:chartang}
Let $\alpha >0$ and let $f \in W^{1,2}_{loc} (\RR^2, \AQt)$ be a nontrivial $\alpha$-homogeneous Dir-minimizer with interface $(\RR, 0)$.
Then the following alternatives hold:
\begin{itemize}
\item[(a)] either $\alpha = l$ for some $l\in \mathbb N$ and
\[
\begin{split}
f^+ & = k_0 \a{r^l \vec{c} \sin(l\theta)} + \sum_{j=1}^{J} k_j \left\llbracket r^{l} \left( \vec{a_j} \cos(l\theta) + \vec{b_j} \sin(l\theta) \right) \right\rrbracket =: f_0^+ + \sum_{j=1}^J k_j f_j, \\
f^- & = (k_0-1) \a{r^l \vec{c} \sin(l\theta)} + \sum_{j=1}^{J} k_j \left\llbracket r^{l} \left( \vec{a_j} \cos(l\theta) + \vec{b_j} \sin(l\theta) \right) \right\rrbracket=: f_0^- + \sum_{j=1}^J k_j f_j;
\end{split}
\]
\item[(b)] or $\alpha = \frac{n^*}{Q^*}$ for some coprime natural numbers $n^*, Q^*$ and
\[
\begin{split}
f^+ & = k_0 \a{0} + \sum_{j=1}^{J} k_j \sum_{z^{Q^*} = x \atop z=(r,\theta)} \left\llbracket r^{n^*} \left( \vec{a_j} \cos(n^* \theta) + \vec{b_j} \sin(n^* \theta) \right) \right\rrbracket =: f_0^+ + \sum_{j=1}^J k_j f_j, \\
f^- & = (k_0-1) \a{0} + \sum_{j=1}^{J} k_j \sum_{z^{Q^*} = x \atop z=(r,\theta)} \left\llbracket r^{n^*} \left( \vec{a_j} \cos(n^* \theta) + \vec{b_j} \sin(n^* \theta) \right) \right\rrbracket =: f_0^- + \sum_{j=1}^J k_j f_j\, ;
\end{split}
\]
\item[(c)] or $\alpha = \frac{2}{3}$ and
\[
\begin{split}
f^+ & = \a{r^{\frac23} \vec{c} \sin \left( {\textstyle{\frac23}} \theta \right)} + \a{r^{\frac23} \vec{c} \sin \left( {\textstyle{\frac23}} (\theta + 2\pi) \right)} + \sum_{j=1}^J k_j \sum_{z^3 = x \atop z=(r,\theta)} \left\llbracket r^2 \left( \vec{a_j} \cos(2\theta) + \vec{b_j} \sin(2\theta) \right) \right\rrbracket \\
& =: f_0^+ + \sum_{j=1}^J k_j f_j, \\
f^- & = \a{r^{\frac23} \vec{c} \sin \left( {\textstyle{\frac23 \theta}} \right)} + \sum_{j=1}^J k_j \sum_{z^3 = x \atop z=(r,\theta)} \left\llbracket r^2 \left( \vec{a_j} \cos(2\theta) + \vec{b_j} \sin(2\theta) \right) \right\rrbracket \\
& =: f_0^- + \sum_{j=1}^J k_j f_j\, .
\end{split}
\]
\end{itemize}
In all three cases the supports of $f_i(x)$ and $f_j(x)$ are disjoint for any $i\neq j\in \{0, 1, \cdots, J\}$ and any $x$ not at the origin. The constant vector $\vec{c} \neq 0$ and the pair of vectors $\vec{a}_j$ and $\vec{b}_j$ are linearly independent.
Finally, under the additional assumptions that $f$ satisfies \eqref{eq:avgsym} and its Dirichlet energy is positive, in the cases (a) and (b) the portions $\sum_j k_j f_j$ must necessarily be nontrivial, namely $J\geq 1$ and we cannot have $f = (f_0^+, f_0^-)$.
\end{prop}
The proof of Proposition \ref{prop:chartang} is based on the following characterization of homogeneous {\em irreducible} Dirichlet minimizing $(Q_0-\frac{1}{2})$-maps with interface $(\mathbb R, 0)$. Note in particular that, in case (b) of the Proposition above, the trace of the corresponding homogeneous map on the circle is reducible, with the only exception of the very trivial map $f_0^+ = \a{0}$.
\begin{lemma}\label{l:char_irr_hom}
Let $g_0 \in W^{1,2} (\So, \AQt)$ with $Q= Q_0$ be irreducible and assume that it is the trace of a non-trivial homogeneous Dir-minimizing map $f_0$ on $\So$ with homogeneity $\alpha$.
Then:
\begin{itemize}
\item[(i)] Either $Q_0=1$, $\alpha = l \in \mathbb N\setminus \{0\}$ and $f_0^+ (r, \theta) = \vec{c} r^l \sin (l \theta)$ is a classical homogeneous harmonic polynomial with trace $0$ on $\mathbb R$;
\item[(ii)] Or $Q_0 =2$, $\alpha = \frac{2}{3}$ and $f_0 = (f_0^+, f_0^-)$, where $f_0^+$ and $f_0^-$ are the maps of Proposition \ref{prop:chartang}(c).
\end{itemize}
\end{lemma}
\begin{remark}\label{r:minimality_of_examples} Clearly, since they are classical harmonic functions, all the examples in case (i) are actual Dirichlet minimizers. It is much less obvious that the examples in (ii) are also minimizers. This is not really needed in the proof of our main result. However, an elementary argument, which we include for completeness at the end of the paper in Proposition \ref{p:exception}, shows that in fact they are.
\end{remark}
\begin{proof}[Proof of Lemma \ref{l:char_irr_hom}]
By Proposition \ref{prop:decomp} there exists a function $\zeta:\So \to \RR^n$ satisfying $\zeta(0) = 0$ such that $g_0$ unwinds to $\zeta$. Let
\[ \zeta(\rho, \theta) = \rho^{\alpha \frac{2Q_0-1}{2}} \zeta(\theta) \]
be an extension of $\zeta$ to the disk $\DD$. By \eqref{eq:unwindsld1} and \eqref{eq:unwindsld2} $f_0$ unwinds to $\zeta(\rho, \theta)$, and thus
\[
\Dir(f_0, \DD) = \iint_{\DD} |D\zeta|^2
\]
by Lemma \ref{lm:unrolling}.
We consider the function $\overline \zeta(z) := \zeta(z^2): \DD^+ \to \RR^n $. By definition $\overline \zeta|_{\mathbb R} \equiv 0$. Since conformal maps do not change Dirichlet energy, we have
\begin{equation}\label{eq:tmpchartg}
\iint_{\DD^+} |D\overline \zeta|^2 = \iint_{\DD} |D\zeta|^2 = \Dir(f_0, \DD).
\end{equation}
Consider any function $\eta: \DD^+ \to \RR^n$ satisfying $\eta|_{\partial \DD^+} = \overline \zeta|_{\partial \DD^+}$,
we can wind the function $\eta(\sqrt{z}): \DD \to \RR^n$ by the formula \eqref{eq:unwindsld1}, \eqref{eq:unwindsld2} (where we use a branch of the square root function, for instance setting $z= r e^{i\theta}$ with $\theta \in [0, 2\pi[$ and defining $\sqrt{r e^{i\theta}} = r^{\frac{1}{2}} e^{i\theta/2}$) and find a corresponding function $h: \DD \to \AQt$ such that $h|_{\partial \DD} = f|_{\partial \DD}$ and $h^+|_{\mathbb R} = h^-|_{\mathbb R} + \llbracket 0 \rrbracket$. By the minimality of $f$ with interface $(\mathbb R, 0)$, we have
\[ \Dir(f_0,\DD) \leq \Dir(h, \DD) = \iint_{\DD^+} |D\eta|^2. \]
This combined with \eqref{eq:tmpchartg} shows that $\overline \zeta$ is a Dir-minimizer in $\DD^+$ which equals $0$ on $\mathbb R$. Thus $\overline \zeta$ is a harmonic function in $\DD^+$ which by Schwarz reflection can be extend it to $\DD$. On the other hand $\overline \zeta$ is $\alpha(2Q_0-1)$-homogeneous. By spherical harmonics we know $\alpha(2Q_0-1) = l\in \NN$ and $\overline \zeta(r, \theta) = \vec{c} r^l \sin(l\theta)$ with some constant $\vec{c}\in \RR^n$. Therefore $\zeta(\theta) = \vec{c} \sin\left( \frac{l\theta}{2} \right)$ on $\So$.
We now claim that, since the $\left(Q_0 -\frac12 \right)$-valued map $g_0$ unwinds to $\zeta(\theta) = \vec{c} \sin\left( \frac{l\theta}{2} \right)$ on $\So$ and it is irreducible, then either $Q_0=1$ or $l=Q_0=2$. In the first case $g_0^+ = \vec{c} \sin(l\theta)$ for some integer $l\in \NN\setminus \{0\}$ and we fall in the first case of the classification. In the second case
\begin{equation}\label{eq:g0u}
g_0^+ = \left\llbracket \vec{c} \sin\left( \frac23 \theta \right) \right\rrbracket + \left\llbracket \vec{c}\sin\left( \frac23 (\theta + 2\pi) \right) \right \rrbracket, \quad \theta \in [0,\pi],
\end{equation}
\begin{equation}\label{eq:g0d}
g_0^- = \left\llbracket \vec{c}\sin\left( \frac23 \theta \right) \right\rrbracket, \quad \theta \in [\pi, 2\pi]\, ,
\end{equation}
which covers the second case of the classification.
We next show our claim.
When $Q_0 = 1$, the condition (i) in Proposition \ref{prop:decomp} holds trivially, thus $g_0$ is irreducible. Now assume $Q_0>1$. The condition (i) fails if we can find $\theta\in [0,2\pi]$ and $k\in \NN$ such that
\begin{equation}\label{eq:charirred}
\zeta(\theta) = \zeta\left( \theta + \frac{4\pi}{2Q_0-1} k \right),\quad 0\leq \theta, \theta + \frac{4\pi}{2Q_0-1} k < 2\pi.
\end{equation}
We denote $\beta = l\theta/2$, then \eqref{eq:charirred} becomes
\begin{equation}\label{eq:contrairred}
\sin\left( \beta \right) = \sin \left( \beta + \frac{2\pi lk}{2Q_0-1} \right), \quad 0 \leq \beta, \beta + \frac{2\pi l k}{2Q_0-1} < l \pi.
\end{equation}
To rephrase it slightly different, \eqref{eq:contrairred} is equivalent to find $\beta_1, \beta_2 \in [0, l\pi)$ such that $ \sin(\beta_1) = \sin(\beta_2)$ and $\beta_1 - \beta_2 = \frac{2 k}{2Q_0-1} l\pi$ for some $k\in \NN\setminus \{0\}$. For all odd integers $l \in \NN$, we can always find $\beta_1, \beta_2 \in [0,l\pi)$ with arbitrary distance in the range $[0,l\pi)$ satisfying $\sin(\beta_1) = \sin(\beta_2)$; for all even integers $l \in \NN$, we can always find $\beta_1, \beta_2 \in [0,l\pi)$ having the same sinus and with arbitrary distance in the range $[0, (l-1)\pi]$. In the latter case, the only way (i) could be satisfied is that there is an even integer $l\geq 2$ so that
\[ \frac{2}{2Q_0-1} l\pi > (l-1)\pi\, . \]
Namely we are looking for even numbers $l\geq 2$ and natural numbers $Q_0 >1$ such that $\frac{2}{2Q_0-1} > \frac{l-1}{l}$. Clearly, $\frac{l-1}{l} \geq \frac{1}{2}$. When $Q_0\geq 3$ we have $\frac{2}{2Q_0-1} \leq \frac{2}{5} < \frac{1}{2}$. Hence $Q_0=2$ is the only possibility: in that case $\frac{2}{2Q_0 -1} = \frac{2}{3}$. Then $l=2$ satisfies the inequality $\frac{l-1}{l} < \frac{2}{2Q_0-1}$, but as soon as $l\geq 4$ we have $\frac{l-1}{l} \geq \frac{3}{4} > \frac{2}{3}$. This restricts the possibilities to the only case $Q_0=l=2$ and thus proves our claim.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:chartang}] Denote by $\alpha$ the homogeneity of the map $f$ and let $g = f|_\So$.
We decompose $g\in W^{1,2}(\So, \AQt)$ into irreducible pieces as described in Proposition \ref{prop:decomp}:
\[ g = \tilde g_0 + \tilde g = \tilde g_0 + \tilde{g} \]
where $\tilde g_0 = (\tilde g^+_0, \tilde g^-_0) \in W^{1,2}(\So, \mathcal{A}_{Q_0}^{\pm})$ is an irreducible map with interface $(\mathbb R, 0)$. We then get a similar decomposition of $f$ as $\tilde f_0 + \tilde f$, where both are $\alpha$-homogeneous and Dir-minimizing.
Recall that according to Lemma \ref{l:char_irr_hom}, either $Q_0 =1$ and so $f_0^+$ is a classical harmonic polynomial with trace $0$ on $\mathbb R$, or $Q_0 =2$, in which case $f_0$ is the map of case (c).
According to \cite[Proposition 5.1]{DS} we can further decompose
\[
\tilde{f} = \sum_j k_j \tilde{f_j}
\]
where the traces $\tilde{g}_j$ of each $\tilde{f}_j$ on $\So$ are irreducible and have disjoint supports, namely
\begin{equation}\label{e:non-intersection-1}
{\rm spt} (\tilde{g}_j (\theta)) \cap {\rm spt} (\tilde{g}_i (\theta)) = \emptyset\qquad
\mbox{for every $\theta$.}
\end{equation}
We next wish to show that
\begin{equation}\label{e:non-intersection-2}
\mbox{if ${\rm spt} (\tilde{g}^\pm_0 (\theta)) \cap {\rm spt} (\tilde{g}_j (\theta)) \neq \emptyset$ then $Q_0=1$ and $\tilde{g}_j (\theta) = \tilde{g}^+_0 (\theta)\;\; \forall \theta \in [0, \pi]$,}
\end{equation}
namely $\tilde{f}_j$ is a classical harmonic function which coincides with $\tilde{f}_0^+$ on the upper half plane and can thus be obtained by Schwarz reflection on the lower half plane.
We argue for \eqref{e:non-intersection-2} and we distinguish two cases:
\medskip
{\bf Case $Q_0=2$.} $\tilde{f}_j$ must have homogeneity $\frac{2}{3}$. By the classification result of \cite[Proposition 5.1]{DS} it must take the form
\[
\sum_{z^3 = x \atop z=(r,\theta)} \left\llbracket r^2 \left( \vec{a_j} \cos(2\theta) + \vec{b_j} \sin(2\theta) \right) \right\rrbracket\, .
\]
where $\vec{a_j}$ and $\vec{b}_j$ span a $2$-dimensional plane, or it must the trivial map $\tilde{f}_j = \a{0}$.
Assume that for some $\bar \theta$ we have ${\rm spt} (\tilde{g}^+_0 (\bar \theta))\cap {\rm \spt} (\tilde{g}_j (\bar \theta))\neq \emptyset$. If $\bar \theta \in ]0, \pi[$, then either we have a line of ``interior singularities'', contradicting the regularity theory of \cite{DS}, or ${\rm spt} (\tilde{g}^+_0 (\theta))\subset {\rm spt} (\tilde{g}_j (\theta))$ $\forall \theta \in ]0, \pi[$. The latter condition would however contradict the linear independence of $\vec{a}_j$ and $\vec{b}_j$. A similar argument can be used to exclude that there is any intersection between ${\rm spt}\, (\tilde{g}_0^- (\bar\theta))$ and ${\rm spt} (\tilde{g}_j (\bar\theta))$ when $\bar\theta\in ]\pi, 2\pi[$.
If $\bar\theta = 0$, then
${\rm spt}\, (\tilde{g}_j (0))$ contains $0$ or $\vec{c} \sin \frac{4\pi}{3}\neq 0$. Like above, the second possibility (and the continuity at the interface) would imply ${\rm spt} (\tilde{g}^+_0 (\theta))\subset {\rm spt} (\tilde{g}_j (\theta))$ $\forall \theta \in [0, \pi]$ by the interior regularity theory of \cite{DS}.
If instead $0 \in \spt \tilde{g}_j (0)$, $\tilde{g}_j$ would have to be trivial, because in the other alternative $\vec{a}_j$ and $\vec{b}_j$ must be linearly independent. Hence $\tilde{f}_j$ vanishes identically. This being the case, consider the $5/2$-valued map $\a{\tilde{f}_0} + \a{\tilde{f}_j}$, which must be Dir-minimizing, and decompose it differently by introducing the maps
\[
h (r, \theta) = \left\{\begin{array}{ll}
\tilde{f}^+_0 (r, \theta) \qquad &\mbox{ if $\theta \in [0, \pi]$}\\
\tilde{f}_0^- (\theta) + \tilde{f}_j (r, \theta) \qquad &\mbox{ if $\theta \in [\pi, 2\pi]$.}
\end{array}\right.
\]
and
\[
h^+ (r, \theta) = \tilde{f}_j (r, \theta) \qquad \mbox{if $\theta \in [0, \pi]$.}
\]
Clearly $h^+ + h$ is the same $\frac{5}{2}$-valued Dir minimizer and $h$ must therefore be a $2$-valued Dir-minimizing map. It is however $\frac{2}{3}$ homogeneous, but not of the form given in \cite[Proposition 5.1]{DS}, which is a contradiction.
Since we can argue similarly for $\bar \theta = \pi$, this shows that when $Q_0=2$ the supports of $\tilde{g}_0^\pm (\theta)$ and $\tilde{g}_j (\theta)$ must be disjoint for all $\theta$.
\medskip
{\bf Case $Q_0=1$.} Here we only have to examine $\bar \theta \in [0, \pi]$.
If ${\rm spt} (\tilde{g}_j (\bar \theta))$ intersects ${\rm spt} (\tilde g^+_0 (\bar \theta))$ for $\bar \theta \in ]0, \pi[$, then arguing as above, the interior regularity theory would imply ${\rm spt}\, (\tilde{g}^+_0 (\theta)) \subset {\rm spt}\, (\tilde{g}_j (\theta))$ for all $\theta \in [0, \pi]$. But then the homogeneity of $\tilde{f}_j$ must be an integer and $\tilde{f}_j$ is a single-valued classical harmonic polynomial by \cite[Proposition 5.1]{DS}. Since on the upper half plane such polynomial coincides with $\tilde{f}_0^+$, on the lower half plane it is determined by Schwarz reflection.
If $\bar \theta \in \{0, \pi\}$, then either $0\in \spt (\tilde g_j(0))$ or $0\in \spt (\tilde g_j(\pi))$. If the homogeneity of $\tilde{f}_j$ is integral, then it is a classical harmonic function, which thus vanishes identically on $\mathbb R$. We then argue as above and consider the $3/2$-valued map $\a{\tilde{f}_0} + \a{\tilde{f}_j}$, which must be Dir-minimizing. As above, wecompose it differently by introducing the maps
\[
h (r, \theta) = \left\{\begin{array}{ll}
\tilde{f}^+_0 (r, \theta) \qquad &\mbox{ if $\theta \in [0, \pi]$}\\
\tilde{f}_j (r, \theta) \qquad &\mbox{ if $\theta \in [\pi, 2\pi]$.}
\end{array}\right.
\]
and
\[
h^+ (r, \theta) = \tilde{f}_j (r, \theta) \qquad \mbox{if $\theta \in [0, \pi]$.}
\]
Clearly $h^+ + h$ is the same $\frac{3}{2}$-valued Dir minimizer and $h$ must therefore be harmonic. By the Schwarz reflection principle, we conclude that in fact $\tilde{f}_0^+$ is the restriction of $\tilde{f}_j$ on the upper half plane, namely it vanishes identically.
We thus must finally examine the possibility that $0\in \spt (\tilde g_j(0))$ or $0\in \spt (\tilde g_j(\pi))$ and the homogeneity is not a natural number, namely
$\alpha = \frac{n^*}{Q^*}$ where $n^*$ and $Q^*$ are coprime. The classification of \cite[Proposition 5.1]{DS} implies that either
\[
\tilde{f}_j (x) = \sum_{z^{Q^*} = x \atop z=(r,\theta)} \left\llbracket r^{n^*} \left( \vec{a_j} \cos(n^* \theta) + \vec{b_j} \sin(n^* \theta) \right) \right\rrbracket
\]
with $\vec{a}_j$ and $\vec{b_j}$ linearly independent, or $\tilde{f}_j \equiv \a{0}$. In the first case we must have $0\not\in {\rm spt}\, (\tilde{g}_j (\theta))$ for every $\theta$, so we are necessarily in the second case, where $\tilde{f}_j$ is a single-valued harmonic function. Arguing as above we then conclude that it must coincide with $\tilde{f}_0^+$ on the upper half plane.
\medskip
{\bf Conclusion}.
\eqref{e:non-intersection-1} and \eqref{e:non-intersection-2} lead immediately to the following conclusions:
\begin{itemize}
\item When $Q_0=2$, $f$ takes necessarily the form in (c), where we set $f_0 = \tilde{f}_0$ and $f_j = \tilde{f}_j$;
\item When $Q_0=1$, either the supports of $\tilde{g}_j (\theta)$ are disjoint from those of $\tilde{g}^+_0 (\theta)$ for all $\theta$ and $j$, in which case we set $f_0 = \tilde{f}_0$ and $f_j = \tilde{f}_j$; or there is one $\tilde{f}_j$ which coincides with $\tilde{f}^+_0$ on the upper half plane, while all the others have disjoint supports. This results into a decomposition of the form (a) or (b). If the homogeneity $\alpha$ is an integer, then the decomposition takes the form (a). If the homogeneity $\alpha$ is not integer, then the decomposition takes the form (b) because the only classical harmonic function which is $\alpha$-homogeneous is the trivial one.
\end{itemize}
We come to the final statement of the proposition. If the Dirichlet energy of $f$ is positive, clearly in case (b) the map $\sum_j k_j f_j$ must have nontrivial energy. In case (a), observe that the triviality of $\sum_j k_j f_j$, the Schwarz reflection principle, and assumption \eqref{eq:avgsym} would imply that $f^+ = Q \a{0}$ and $f^- = (Q-1) \a{0}$, which again is not compatible with the positivity of the Dirichlet energy.
\end{proof}
\section{Proof of Theorem \ref{thm:main}: Discreteness of the singular set}\label{s:final}
The proof of the main theorem is by induction on the number of values $Q$. The basic step $Q=1$ is clearly trivial, because $f^-$ does not exist in that case and $f^+$ is a classical harmonic function. . Now we assume $Q>1$ and, as induction hypotheses, that the theorem holds for every $Q'<Q$.
We argue by contradiction and assume the existence of a Dir-minimizing $\left(Q- \frac12 \right)$-valued planar function with real analytic interface $(\gamma, \varphi)$ whose singular set is not discrete. As shown in Section \ref{s:preliminaries} we can assume, without loss of generality that $f$ is as in Theorem \ref{thm:main_simple}, namely $Q\eta^+=(Q-1) \eta^-$ and the interface is $(\RR, 0)$. Under our assumptions the singular set must have an accumulation point $x_0$. The latter cannot be in the interior, and thus belongs to the interface. Without loss of generality we can assume that $x_0=0$.
Next, we must have $f^+ (0) = Q \a{0}$. Otherwise we have $f^+ (0) = Q_1 \a{0} + T$ with $T\in \mathcal{A}_{Q_2} (\R^n)$, where $Q_1+Q_2 = Q$, $1\leq Q_1 \leq Q-1$ and $\supp (T)$ does not contain the origin. By the H\"older continuity theorem, in a neighborhood $U$ of the origin there would be a $Q_2$-valued map $h\in W^{1,2} (U)$ and a $(Q_1-\frac{1}{2})$-valued map $g = (g^+, g^-) \in W^{1,2} (U)$, with disjoint supports and such that $f^\pm = g^\pm + h$. Then the singular set of $f$ in $U$ would be the union of the singular set of $h$ and of the singular set of $f$. Moreover, both must be Dir-minimizing. Hence the singular set of $h$ is discrete by the interior regularity theory, whereas the singular set of $g$ is discrete by the inductive assumption. This is however not possible because we know that $0$ is an accumulation point of the singular set of $f$.
Note next that it must be $D(r)>0$ for every $r$ in a positive interval, otherwise we would have $f^+ \equiv Q \a{0}$ and $f^- \equiv (Q-1) \a{0}$ in some neighborhood of $0$. Thus $I_{f}(r)$ is well-defined for every $r>0$ sufficiently small. Let $g$ be the (homogeneous) tangent function to $f$ at $0$, given by Theorem \ref{thm:cvtang}. By the characterization in Proposition \ref{prop:chartang} $g$ has the following decomposition:
\[ g^+ = g_0^+ + \sum_{j=1}^J k_j g_j, \quad g^- = g_0^- + \sum_{j=1}^J k_j g_j \]
where:
\begin{itemize}
\item In the alternatives (a) or (b) of Proposition \ref{prop:chartang} $(g_0^+, g_0^-)$ equals $(k_0 \a{h}, (k_0-1) \a{h})$, where $h$ is a classical harmonic function which vanishes at $\RR$ and $g_0^-$ its reflection.
\item In the alternative (c) $(g_0^+, g_0^-) \in W^{1,2}(\RR^2, \mathcal{A}_2^{\pm})$.
\end{itemize}
In the alternative (b) $g_0^+$ is $2$-valued, namely $g_0^+ = \a{(g_0^+)_1} + \a{(g_0^+)_2}$ and we define
\[
d_0:= \min_{x\in \mathbb{S}^1_+} {\rm sep} (g_0^+ (x)) =
\min_{x\in \mathbb{S}_1^+} \left|(g_0^+)_1 (x) - (g_0^+)_2 (x)\right|\, .
\]
Note that $d_0$ is positive.
In the alternative (a) we set $d_0 = +\infty$.
For each $j\in \{1, \cdots, J\}$ we define
\[ d_{0,j}:= \min \left\{ \min_{x\in \mathbb{S}^1_+ } \dist\left(\supp(g_0^+(x)), \supp(g_j(x))\right), \, \min_{x\in \mathbb{S}^1_-} \dist\left(\supp(g_0^-(x)), \supp(g_j(x))\right) \right\},
\]
and define for each pair $i\neq j \in \{1,\cdots, J\}$
\[
d_{i,j}:= \min_{x\in \So} \dist\left(\supp(g_i(x)), \supp(g_j(x)\right).
\]
By Proposition \ref{prop:chartang} we know $d_0, d_{0,j}, d_{i,j} > 0$ for all $i, j$, because the Dirichlet energy of the tangent function is positive and it satisfies the averaging condition \eqref{eq:avgsym}.
Let
\[ \epsilon = \frac14 \min \left\{d_0, \min_{j} d_{0,j}, \, \min_{i\neq j} d_{i,j} \right\}>0. \]
We claim that there exists $r_0>0$ such that
\begin{equation}\label{e:separation_claim}
\GG(f(x), g(x)) \leq \epsilon|x|^{\alpha} \quad \text{ for every } |x| \leq r_0,
\end{equation}
where $\alpha = I_{0,f}(0)>0$.
In fact, recall the uniform convergence of the blow-ups $f_r$ to $g$:
\[ \GG(f_r(\theta), g(\theta)) \to 0 \text{ uniformly in } \theta \in \So \text{ as } r\to 0. \]
Recall \eqref{eq:blowupscal}, the blow-ups satisfy
\[ \frac{f(x)}{|x|^\alpha} = f_{|x|} \left( \frac{x}{|x|} \right) \left( 1 + O\left( |x|^{\frac{\beta}{2}} \right) \right). \]
Hence
\[ \GG\left( \frac{f(r,\theta)}{r^\alpha}, g(\theta) \right) \to 0 \text{ uniformly in } \theta \in \So \text{ as } r\to 0. \]
Recall that $g$ is an $\alpha$-homogeneous map, i.e. $g(x) = |x|^\alpha g (\frac{x}{|x|})$. We have thus showed \eqref{e:separation_claim}.
The choice of $\epsilon$ implies the existence of functions $h_j$ with $j\in \{0, 1, \cdots, J\}$, such that:
\begin{itemize}
\item $h_0 = (h^+_0, h_0^-) \in W^{1,2}(B_{r_0}, \mathcal{A}_{Q_0}^{\pm})$ with interface $(\gamma, \varphi)$ and $Q_0 = 1$ or $2$, depending on whether alternative (a) or (b) in Proposition \ref{prop:chartang} holds, and in particular $\card \supp(h_0^+(x)) = Q_0$ for all $x \in B_{r_0}^+ \setminus \{ 0\}$;
\item each $h_j$ is in $ W^{1,2}(B_{r_0}, \mathcal{A}_{k_j Q_j} ) $, and
\begin{equation}\label{eq:decompf}
f|_{B_{r_0}} = (h_0^+, h_0^-) + \sum_{j=1}^J h_j;
\end{equation}
\item For every $x\in B_{r_0}\setminus \{ 0\}$ and every $i>j>0$ we have $\supp (h_j (x)) \cap \supp (h_i (x))=\emptyset$;
\item For every $x\in B_{r_0}^+\setminus \{0\}$ and every $i>0$ we have $\supp (h_i (x))\cap \supp (h_0^+ (x)) =\emptyset$;
\item For every $x\in B_{r_0}^-\setminus \{0\}$ and every $i>0$ we have $\supp (h_i (x))\cap \supp (h_0^- (x))=\emptyset$.
\end{itemize}
In particular:
\begin{itemize}
\item $h_0$ is a Dir-minimizer with interface $(\RR, 0)$, and each $h_j$ is a Dir-minimizer;
\item The singular set of $f$ in $B_{r_0}$ is given by $0$ and the union of the singular sets of $h_0^+, h_0^-$ and the $h_j$'s.
\end{itemize}
Suppose $J=0$. Recall Proposition \ref{prop:chartang}, this may only occur in the alternative (c), i.e. when $f|_{B_{r_0}} = (h_0^+, h_0^-)$ is a $\frac{3}{2}$-valued map. By the separation of sheets of $h_0^+$, the singular set of $f$ in $B_{r_0}$ is just the origin and we get a contradiction.
Suppose $J\geq 1$, in other words the sum \eqref{eq:decompf} contains at least two terms, so $h_0^+$ takes strictly less than $Q$ values and we can use our inductive hypothesis to conclude that the singular set of $h_0$ is discrete. On the other hand, the singular set of each $h_j$ with $j>0$ is discrete by \cite[Theorem 0.12]{DS}. We conclude that the singular set of $f$ in $B_{r_0}$ is discrete as well, contradicting the assumption that the origin was an accumulation point for it.
\section{The exceptional $\frac{2}{3}$-homogeneous minimizer}
In this section we complete the analysis of the singularities by showing the following
\begin{prop}\label{p:exception}
Let $\vec{c} \neq 0$ and let $f_0$ be the $\frac{3}{2}$-valued map of case (ii) in Lemma \ref{l:char_irr_hom}. Then $f_0$ is locally Dir-minimizing.
\end{prop}
\begin{proof}
First of all consider that $f_0$ takes values in the line spanned by $\vec{c}$. Thus it suffices to show the claim under the assumption that $n=1$.
Consider now the scalar functions
\[
g^+ (\theta) = \a{\sin \frac{2}{3}\theta} + \a{\sin \frac{2}{3} (\theta+2\pi)}\, ,\qquad \theta \in [0, \pi]
\]
\[
g^- (\theta) = \a{\sin \frac{2}{3} \theta}\, , \qquad \theta \in [\pi, 2\pi]\, ,
\]
and the corresponding $\frac{3}{2}$-valued map $(g^+, g^-)$ on $\So$.
Denote by $h = (h^+, h^-)$ any minimizer of the corresponding Dirichlet problem for $\frac{3}{2}$-valued maps with interface $(\mathbb R, 0)$, which can be shown to exist by the direct methods of the calculus of variations following the theory in \cite{DDHM}. Note that in order to apply the direct methods we need to show the existence of at least one $\frac{3}{2}$-valued function with finite Dirichlet energy which takes the Dirichlet boundary data and has interface $(\mathbb R, 0)$. However such function is provided precisely by the $f_0$ of case (ii) in Lemma \ref{l:char_irr_hom}.
A simple computation shows that $2 \bdeta (g^+ (\theta)) = g^- (2\pi-\theta)$ for every $\theta$.
In particular, it must be that $2\bdeta (h^+ (x)) = h^- (\bar x)$, otherwise we could argue as in Section \ref{s:mod_out_average} and lower the energy of $h$ by keeping the same boundary value and the same interface. Hence $h$ satisfies the condition \eqref{eq:avgsym}.
Now, $h^-$ is a classical harmonic function and by Theorem \ref{thm:Holder} it has continuous trace on the open segment $](-1,0), (1,0)[$. We will show below that there must be one point $\sigma \in ]-1,1[$ such that $h^- (\sigma ,0)=0$. Fix now such $p = (\sigma, 0)$ and observe that the Dirichlet energy of $h$ cannot vanish in any disk $B_r (p)$: if it vanishes on some disk, by the unique continuation of classical harmonic function $h^-$ would have to vanish identically, which is not possible because its trace on $(\So)^-$ is not identically $0$. Consider now the unique tangent function $f$ to $h$ at $p$. The latter must be an $\alpha$-homogeneous Dir-minimizer, satisfying the averaging condition \eqref{eq:avgsym}.
We claim that such tangent function must be necessarily of the form $f = (f_0^+, f_0^-)$ as in case (c) of Proposition \ref{prop:chartang}, which thus would prove the minimality of $f$ (by the compactness of Dir-minimizers). Indeed, if this were not the case, then $f$ would have to fall necessarily in case (a) of Proposition \ref{prop:chartang}, because in order to fall in case (b) the map would have to be $(Q-\frac{1}{2})$-valued with $Q\geq 3$. In case (a) of Proposition \ref{prop:chartang} we would have
\begin{align}
f^+ (x) & = \a{k (x)} + \a{l (x)},\\
f^- (x) & = \a{l (x)},
\end{align}
where both $k$ and $l$ are homogeneous harmonic polynomials with the same degree of homogeneity $d\geq 1$. Moreover $k (s,0) =0$ for every $s$.
In particular
\begin{align*}
k (r, \theta) &= a r^d \sin d \theta\\
l (r, \theta) &= r^d (\alpha \sin d\theta + \beta \cos d \theta)
\end{align*}
for some constants $a, \alpha, \beta \in \mathbb R$. Notice however that $k$ and $l$ cannot coincide, because otherwise the averaging condition \eqref{eq:avgsym} would imply that they both vanish identically, whereas the Dirichlet energy of $f$ must be positive. Since $k$ and $l$ do not coincide, Proposition \ref{prop:chartang} implies that
\[
a \sin d\theta \neq \alpha \sin d\theta + \beta \cos d\theta \qquad \forall \theta \in [0, \pi]\, ,
\]
namely
\[
(\alpha -a) \sin d\theta + \beta \cos d\theta \neq 0 \qquad \forall \theta \in [0, \pi]\, .
\]
The latter condition is however impossible.
It remains to show the existence of $\sigma\in ]-1,1[$ such that $h^- (\sigma, 0) =0$. As already recalled, $h^-$ has a continuous trace on $]-1,1[$. If we knew the continuity of $h^-$ also at the ``corner points'' $(-1,0)$ and $(1,0)$, then we would have
\begin{align}
h^- (1, 0) & = \sin \frac{4\pi}{3} < 0 \\
h^- (-1, 0) & = \sin \frac{2\pi}{3} > 0
\end{align}
and the existence of the point $\sigma$ would be guaranteed by the intermediate value theorem for continuous functions. While it is possible to show a general continuity result at the intersection of the ``boundary'' $\So$ with the interface $\mathbb R$ under rather general assumptions on the boundary data and for a general $(Q-\frac{1}{2})$-Dir minimizer, this would require quite some effort and goes beyond the scopes of the present paper. We circumvent this technical difficulty with a short ad hoc argument.
In order to prove that $\sigma$ exists it suffices indeed to argue that $h^-$ must take both positive and negative values on $]-1,1[$. For this it suffices to show the existence of a sequence of values $s_k\uparrow 1$ such that $h^- (s_k, 0) \to \sin \frac{4\pi}{3}$ and of a sequence $t_k \downarrow -1$ such that
$h^- (t_k, 0) \to \sin \frac{2\pi}{3}$. Without loss of generality, let us argue for the existence of the sequence $s_k$. Assume by contradiction that there are a positive $\delta$ and positive $\eta$ such that
\[
\left|h^- (1-t, 0) - \sin \frac{4\pi}{3}\right|\geq 2\delta \qquad \forall t\in ]0, \eta[\, .
\]
Let $\gamma_t$ be the arc $\partial B_t (1,0)\cap B_1 (0) \cap \{(x,y): y<0\}$. One endpoint $p_t$ of $\gamma_t$ is precisely $(1-t, 0)$, while the other endpoint $q_t$ lies on $(\So)^-$. By choosing $\eta$ sufficiently small we can use the continuity of the harmonic function $h^-$ at $q_t$ and the continuity of its trace $g^-$ at $(1,0)$ to infer
\[
\left|h^- (q_t) - \sin \frac{4\pi}{3}\right| = |g^- (q_t) - g^- (1,0)| \leq \delta\, .
\]
We have thus concluded that
\[
|h^- (p_t) - h^- (q_t)|\geq \delta\, .
\]
By the fundamental theorem of calculus and using Cauchy-Schwarz, for every $\rho \in ]0, \eta[$ such that $h^-|_{\gamma_\rho} \in W^{1,2} (\gamma_\rho)$ we conclude
\[
\int_{\gamma_\rho} |D_{\tau} h^-|^2 \geq \frac{C}{\rho} \left(\int_{\gamma_\rho} |D_{\tau} h^-|\right)^2 \geq \frac{C\delta^2}{\rho}\, ,
\]
where $C$ is a geometric constant. The latter inequality thus holds for a.e. $\rho\in ]0, \eta[$ and integrating in $\rho$ we then conclude
\[
\iint_{B_\eta (1,0)} |Dh^-|^2 \geq \int_0^{\eta} \left( \int_{\gamma_{\rho}} |D_{\tau} h^-|^2 \right) d\rho \geq \int_0^\eta \frac{C\delta^2}{\rho}\, d\rho = \infty\, ,
\]
which contradicts the fact that $h^-$ has finite energy.
\end{proof}
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{sec:intro}
\input{paper-introduction}
\section{Related work}
\label{sec:sota}
\input{paper-related}
\section{IoT and Real-world Data in STEM}
\label{sec:iot}
\input{paper-infrastructure}
\section{Energy Efficiency Education}
\label{sec:energy}
\input{paper-scenario-energy}
\section{Sea Polution Education}
\label{sec:water}
\input{paper-scenario-water}
\section{Conclusions}
\label{sec:conlusion}
\input{paper-conclusions}
\section{Acknowledgments}
This work has been partially supported by the EU research project ``Green Awareness In Action'' (GAIA), funded under contract number 696029 and the research project Designing Human-Agent Collectives for Sustainable Future Societies (C26A15TXCF) of Sapienza University of Rome.
This document reflects only the authors' view and the EC and EASME are not responsible for any use that may be made of the information it contains.
A previous version of this paper has appeared in the \textit{2018 IEEE Conference on Computational Intelligence and Games (CIG)}, DOI: \url{https://doi.org/10.1109/CIG.2018.8490370}, \cite{8490370}. \textcopyright 2018 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
|
\section{Introduction} \label{Introduction}
One of the most popular control technique in both academia and industry is model predictive control (\emph{MPC}) due to its ability to explicitly accommodate hard state and input constraints \citep[][ among others]{bemporad1999robust, camacho2004c, rawlings2009model, mayne2014model}. Thereon, much effort has been devoted to developing a stability theory for \emph{MPC} \citep[see i.e.][]{rawlings2009model, grune2011nonlinear, mayne2016robust}. An overview of recent developments can be found in \cite{mayne2014model}. \emph{MPC} involves the solution of an open--loop optimal control problem at each sampling time with the current state as the initial condition. Each of these optimizations provides the sequences of future control actions and states. The first element of the control action sequence is applied to the system and, then the optimization problem is solved again at the next sampling time after updating the initial condition with the system state. \emph{MPC} keeps constant the computational burden by optimizing the system behaviour within a finite length window. The system behaviour beyond the window is summarized in a term known as \emph{cost--to--go}.
\emph{MPC} is often formulated assuming that the system state can be measured. However, in many practical cases, the only information available is noisy measurements of system output, so the use of independent algorithms for state estimation (including observers, filters and estimators) becomes necessary \cite[see][]{rawlings2006particle}. Of all these methods, moving horizon estimation (\emph{MHE}) is especially engaging for use with \emph{MPC} because it can be formulated as a similar online optimization problem. Solving the \emph{MHE} problem produces an estimated state that is compatible with a set of past measurements that recedes as current time advances \citep{schweppe1973uncertain, rao2001constrained, rao2003constrained}. This estimate is optimal in the sense that it maximizes a criterion that captures the likelihood of the measurements. Along the same time that relevant results on \emph{MPC} were developed, research works on \emph{MHE} begun. The works of \cite{rao2001constrained} and (\citeyear{rao2003constrained}) provide overviews of linear and nonlinear \emph{MHE}. Recent results regarding \emph{MHE} for nonlinear systems are given for robust stability and estimate convergence properties \citep[see][among others]{alessandri2005robust, alessandri2008moving, alessandri2012min, garcia2016new, sanchez2017adaptive}. In recent years several results have been obtained for different \emph{MHE} formulations, advancing from idealistic assumptions, like observability and vanishing disturbances, to realistic situations like detectability and bounded disturbances \citep[see][]{ji2015robust, muller2017nonlinear, allan2019lyapunov, deniz2019robust}.
When disturbances, model uncertainty and system constraints can be neglected, state and control sequences can be independently computed \citep[see][]{duncan1971solutions, bensoussan2004stochastic, aastrom2012introduction, georgiou2013separation}. However, in practical applications, these conditions are very difficult to fulfil, i.e., process disturbances and measurement noise are usually present, as well as model uncertainty. In this context, it becomes necessary approaches that include this information into the controller design. State-feedback \emph{MPC} is a mature field with results that considers model uncertainty, input disturbances, and noises \citep[ among others]{magni2003robust, bemporad2003min, raimondo2009min}. However, these works did not consider robustness with respect to errors in state estimation. Fewer results are available for output-feedback \emph{MPC}. An overview of nonlinear output-feedback \emph{MPC} is given by \cite{findeisen2003state} and the references therein. Many of these approaches involve designing separate estimator and controller, using different estimation algorithm \citep{roset2006stabilizing,magni2009nonlinear, patwardhan2012nonlinear, zhang2013lyapunov, ellis2017state}. Results on robust output-feedback \emph{MPC} for constrained, linear, discrete-time systems with bounded disturbances and measurement noise can be found in \cite{mayne2006robust, mayne2009robust} and \cite{voelker2010unconstrained, voelker2013moving}. These approaches first solve the estimation problem and prove the convergence of the estimated state to a bounded set, and then take the uncertainty of the estimation into account when solving the \emph{MPC} problem.
The approach of solving simultaneously \emph{MHE--MPC} was originally introduced by \cite{copp2014nonlinear} and later developed in several papers \citep{copp2016conditions,copp2016addressing,copp2017simultaneous}. In the first paper, \cite{copp2014nonlinear} proposed an output feedback controller that combines state estimation and control into a single $min-max$ optimization problem that, under observability and controllability assumptions \citep{copp2016conditions}, guarantees the boundedness of state and tracking errors. Finally, in the last work reported by \cite{copp2017simultaneous}, the authors established the conditions for guaranteeing the boundedness of error for trajectory tracking problems. They also introduced a primal--dual interior point method that can be used to efficiently solve the $min-max$ optimization problem. The criterion used in these works involves finite forward and backward horizons that are minimized with respect to feedback control policies and maximized with respect to the unknown parameters in order to guaranty robustness in the worst-case scenario.
In the present work, we introduce an output--feedback controller for nonlinear systems subject to bounded disturbances using simultaneous \emph{MHE--MPC} approach. The resulting optimization problem minimizes a criterion that involves finite forward and backward horizons with respect the unknown initial state, measurement noise and control input variables while it is maximized with respect the unknown future disturbance variables. We show that the proposed controller results in closed--loop trajectories along which the states remain bounded. These results rely on two assumptions: The first assumption requires that the optimization criterion include an adaptive arrival cost \citep{sanchez2017adaptive}. This assumption allows to ensure the boundedness of the state estimate and to obtain a bound for the estimation error set if the parameters of the estimation problem are properly chosen \citep{deniz2019robust}.
The second assumption requires that the backward (estimation) and forward (control) horizons are sufficiently large so that enough information is obtained in order to find state estimates and control inputs compatible with dynamics, noises and constraints. This assumption is satisfied if the system is detectable, stabilizable and the parameters in the cost function (weights and horizons) are chosen appropriately.
The rest of the paper is organized as follows: Section 2 introduces the notation, definitions and properties that will be used through the paper. In Section 3 we formulate the estimation and control problem, and in Section 4 we analyze its closed-loop stability. Section 5 discusses two examples to illustrate the concepts presented in this work. The first example uses a simple nonlinear model to analyse the consequences of simultaneously solving the estimation and control problems. The second example compares the performance obtained by the simultaneous and independent approaches applied to the regulation of the state of a van der Pol oscillator for two operational conditions. Finally, conclusion and future work is discussed in Section 6.
\section{Preliminaries and setup}
\subsection{Notation}
Let $\mathbb{Z}$ denotes the integer numbers, $\mathbb{Z}_{\left[a,b \right]}$ denotes the set of integers in the interval $\left[ a,b \right]$, with $b>a$ and $\mathbb{Z}_{\geq a}$ denotes the set of integers greater or equal to $a$. Boldface symbols denote sequences of finite ($\boldsymbol{w} \coloneqq \{ w_1, \ldots, w_2 \}$) or infinite ($\boldsymbol{w} \coloneqq \{ w_1, \ldots, w_2, \ldots \}$) length. We denote $\hat{x}_{j\vert k}$ as the state at time $j$ estimated at time $k$.
By $\vert x\vert$ we denote the euclidean norm of a vector $x \in \mathbb{R}^{n_x}$. Let $\Vert \boldsymbol{x}\Vert \coloneqq \sup_{k\in \mathbb{Z}_{\geq 0}} \vert x_k\vert $ denote the supreme norm of the sequence $\boldsymbol{x}$ and $\Vert \boldsymbol{x}\Vert_{\left[a, b \right]} \coloneqq \sup_{k\in \mathbb{Z}_{\left[a, b \right]}} \vert x_k\vert$. A function $\gamma : \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class $\mathcal{K}$ if $\gamma$ is continuous, strictly increasing and $\gamma \left(0\right) = 0$ . If $\gamma$ is also unbounded, it is of class $\mathcal{K}_\infty$. A function $\zeta : \mathbb{R}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class $\mathcal{L}$ if $\zeta$ is continuous, non increasing and $\lim_{t \rightarrow \infty} \zeta\left(t \right) = 0$. A function $\beta : \mathbb{R}_{\geq 0}\times\mathbb{Z}_{\geq 0} \rightarrow \mathbb{R}_{\geq 0}$ is of class $\mathcal{KL}$ if $\beta \left(\cdot,k \right)$ is of class $\mathcal{K}$ for each fixed $k \in \mathbb{Z}_{\geq 0}$, and $\beta \left(r,\cdot \right)$ of class $\mathcal{L}$ for each fixed $r \in \mathbb{R}_{\geq 0}$.
Let us consider now two sets $A$ and $B$, the Minkowski addition is defined as $A\oplus B \coloneqq \{ a+b|\,a\in A, b\in B \}$. On the other hand, the Minkowski difference\footnote{Also known as the Pontryagin difference.} is defined as $A \ominus B \coloneqq \{ d|\, d+b \in A \}$. In the following sections, we will use the notation $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{p,t,l}}}$ to reference the cost incurred solving the problem \textit{p} at time \textit{t} with a horizon length \textit{l}, while $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{p,t,l}}}\left(x \right)$ will be used to indicate the cost at the solution $x$, with $x$ belonging to a consistent domain with the cost function $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{p,t,l}}}$. When necessary, we will use the notation $x^{(1)}_{i,k}$ and $x^{(2)}_{i,k}$ to differentiate $i$--th component of the state vector of two discrete-time trajectories of the system, with $i\in\mathbb{Z}_{\left[1,n\right]}$. Moreover, $x^{(1)}_k(x_0^{(1)},\boldsymbol{w}^{(1)} )$ will denote a trajectory with initial condition $x_0^{(1)}$ and perturbed by the sequence $\boldsymbol{w}^{(1)}$. A similar notation is used for the case of continuous time systems, where $t$ is used instead $k$ to denote continuous time.
\subsection{Problem statement}
Let us consider a discrete-time nonlinear system whose behaviour is given
\begin{equation} \label{nonlinear system}
\begin{array}{rl}
x_{k+1} =& f\left(x_k,u_k\right)+w_k \quad \forall \, k \in \mathbb{Z}_{\geq 0}, \\
y_k =& h\left(x_k \right) + v_k,
\end{array}
\end{equation}
where $x \in \mathcal{X} \subseteq \mathbb{R}^{n_x}$ is the system state, $u \in \mathcal{U} \subseteq \mathbb{R}^{n_u}$ is the system's input and $w \in \mathcal{W} \subseteq \mathbb{R}^{n_w}$ is the unmeasured process disturbance posed as an additive input. The output of the system is $y \in \mathcal{Y} \subseteq \mathbb{R}^{n_y}$ and $v \in \mathcal{V} \subseteq \mathbb{R}^{n_v}$ is the measurement noise. The function $f(\cdot,\cdot)$ is assumed to be at least locally Lipschitz in its arguments, and the function $h(\cdot)$ is known to be a continuous function.
The sets $\mathcal{X}$, $\mathcal{U}$, $\mathcal{W}$, $\mathcal{V}$ and $\mathcal{Y}$ are assumed to be convex, containing the origin in its interior.
The estimation and control problem attempts to simultaneously find the optimal state $\hat{x}_{k\vert k}$ and the optimal sequence of control inputs $\boldsymbol{\hat{u}}$ which will steer the system to the desired operation zone. It is in an infinite-horizon optimization problem given by
\begin{equation} \label{mhe mpc infinite horizon}
\begin{array}{l}
\underset{\substack{\hat{x}_{0|k}, \boldsymbol{\hat{w}}, \boldsymbol{\hat{u}} } } {\operatorname{min}} \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$\infty$}}} \coloneqq \displaystyle \sum_{j=0}^{k} \ell_e \left(\Hat{w}_{j\vert k}, \hat{v}_{j\vert k} \right) + \sum_{j=k}^{\infty}\left(\ell_c \left(\hat{x}_{j\vert k}, \hat{u}_{j\vert k} \right) - \ell_{w_c}\left( \hat{w}_{j\vert k} \right)\right) \\
\hspace{1cm} \text{s.t.} \left\{
\begin{array}{l}
\begin{array}{rl}
\hat{x}_{j+1|k} =& f\left(\hat{x}_{j\vert k},\hat{u}_{j\vert k}\right)+\hat{w}_{j\vert k}, \\
y_{j} =& h\left(\hat{x}_{j\vert k}\right) + \hat{v}_{j\vert k},\\
\end{array} \\
\hat{x}_{j|k} \in \mathcal{X
,\,\hat{u}_{j\vert k}\in \mathcal{U}, \hat{w}_{j|k} \in \mathcal{W}, \, \hat{v}_{j|k} \in \mathcal{V}.
\end{array} \right.
\end{array}
\end{equation}
Functions $\ell_e(\hat{w}_{j\vert k},\hat{v}_{j\vert k})$ penalize large values of $\hat{w}_{j\vert k}$ and $\hat{v}_{j\vert k}$, whereas $\ell_{c}(\hat{x}_{j\vert k},\hat{u}_{j\vert k})$ penalize large values of the predicted state $\hat{x}_{j\vert k}$ and control inputs $\hat{u}_{j\vert k}$. The function $\ell_{w_c}(\hat{w}_{j\vert k})$ is assumed to take non--negative values and since it is subtracting in the objective function, process disturbances will tend to be maximized within the control window. When necessary, we will decompose the function $\ell_e(\cdot,\,\cdot)$ into $\ell_{w_e}(\cdot)$ and $\ell_{v_e}(\cdot)$ which penalizes $\hat{w}_{j\vert k}$ and $\hat{v}_{j\vert k}$, respectively.
Problem \eqref{mhe mpc infinite horizon} is valuable from a theoretical point of view since it guarantees the boundedness of the estimates $\hat{x}_{j\vert k}$ and control actions $\hat{u}_{j\vert k}$ provided the cost function is bounded, i.e., $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$\infty$}}} \leq \gamma$, $\forall\, k \in \mathbb{Z}_{\geq 0}$, with $\gamma \in\mathbb{R}_{\geq 0}$.
If functions $\ell_e(\cdot,\cdot)$, $\ell_c(\cdot,\cdot)$ and $\ell_{w_c}(\cdot)$ are defined using a norm--$\ell_p$, problem \eqref{mhe mpc infinite horizon} would guarantee that the state $x_k$ and $u_k$ are $\ell_p$ , provided that noises $w_k$ and $v_k$ are also $\ell_p$. This would mean that the closed-loop system has a finite $\ell_p$--induced gain.
The infinite--horizon problem \eqref{mhe mpc infinite horizon} lacks practical interest since it is intractable from a computational point of view. Then, it is reformulated into a receding finite--horizon problem
\begin{equation} \label{mhe mpc opt problem}
\begin{array}{l}
\displaystyle \underset{\substack{\hat{x}_{k-N_e|k},
\boldsymbol{\hat{w}}, \boldsymbol{\hat{u}} } }
{\operatorname{min}} \, \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}} \coloneqq
\Gamma_{k-N_e}\left(\chi\right) + \displaystyle \sum_{j=k-N_e}^{k} \ell_e \left(\hat{w}_{j\vert k}, \hat{v}_{j\vert k} \right) + \\
\displaystyle
\hspace{2.0cm} \sum_{j=k}^{k+N_c-1}\left(\ell_{c} \left(\Hat{x}_{j\vert k}, \hat{u}_{j\vert k} \right) - \ell_{w_c}\left( \hat{w}_{j\vert k} \right)\right) + \Upsilon_{k+N_c}\left( \Xi \right) \vspace{0.1cm}\\
\hspace{1cm} \text{s.t.} \left\{
\begin{array}{l}
\begin{array}{rl}
\chi =& \hat{x}_{k-N_e\vert k}-\bar{x}_{k-N_e}, \\
\Xi =& \hat{x}_{k+N_c\vert k}, \\
\hat{x}_{j+1|k} =& f\left(\hat{x}_{j\vert k},\hat{u}_{j\vert k}\right)+\hat{w}_{j\vert k}, \\
y_{j} =& h\left(\hat{x}_{j\vert k}\right) + \hat{v}_{j\vert k},\\
\end{array} \\
\hat{x}_{j|k} \in \mathcal{X},\Xi \in \mathcal{X}_f\subseteq \mathcal{X}, \,
\hat{u}_{j\vert k}\in \mathcal{U}, \hat{w}_{j|k} \in \mathcal{W}, \ \hat{v}_{j|k} \in \mathcal{V}.
\end{array} \right.
\end{array}
\end{equation}
For computation tractability, the infinite summations of $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$\infty$}}}$ have been replaced by backward and forward windows of finite length, corresponding to the estimation $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ and control $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ problems of criterion $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}}$, respectively. $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ includes $N_e$ terms backward in time from sample $k$ corresponding to the \textit{estimator stage-cost}, $\ell_e \left(\hat{w}_{j\vert k}, \hat{v}_{j\vert k} \right)$,
and the extra term $\Gamma_{k-N_e}(\chi)$, known as \textit{arrival-cost}, that summarizes information left behind the estimation window by penalizing the uncertainty in the initial state $\hat{x}_{k-N_e \vert k}$ \citep{rao2001constrained,rao2003constrained}. On the other hand, $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ includes $N_c$ terms forward in time from sample $k$ corresponding to the \textit{controller stage-cost}, $\ell_{c}(\hat{x}_{j\vert k}, \hat{u}_{j\vert k}) - \ell_{w_c}(\hat{w}_{j\vert k})$, and an extra term $\Upsilon_{k+N_c}(\Xi)$, known as \textit{cost-to-go}, that summarizes the behaviour of the system beyond the control window by penalizing the deviation of the final state $\Xi=\hat{x}_{k+N_c \vert k}$. Moreover, the set $\mathcal{X}_f$ represent the set of terminal constraints, as common in \emph{MPC} \citep{rawlings2017model}.
The goal of problem \eqref{mhe mpc opt problem} is to estimate the initial state $\hat{x}_{k-N_e \vert k}$ and disturbances $\hat{w}_{j \vert k}$ $j \in\mathbb{Z}_{\left[k-N_e,k-1\right]}$ such that an estimate $\hat{x}_{k|k}$ is obtained to compute the control inputs $u_{j \vert k}$ $j\in\mathbb{Z}_{\left[k,k+N_c-1\right]}$ that drive the system states to the desired region. Therefore, there is no point on penalizing the control cost $\ell_c(\cdot,\,\cdot)$ along the estimation window. The variables $\hat{v}_{j \vert k}$ are not independent variables since they are uniquely determined by the remaining optimization variables and the output equation
\begin{equation}
\hat{v}_{j \vert k} = y_j - h(\hat{x}_{j \vert k}), \qquad j \in\mathbb{Z}_{\left[k-N_e,k\right]}.
\end{equation}
Since there is no measurement of future system output, $\hat{v}_{j|k}$ will not be considered along the control window. However, the disturbances $\hat{w}_{j|k}$ needs to be considered along both windows $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ and $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ because they affect all the states, starting from $j=k - N_e - 1$. As will be shown later, the ratio between disturbances $w_{j}$ and control actions $u_{j}$, for $j\in\mathbb{Z}_{\left[k,k+N_c-1\right]}$, encodes the controllability property of the system, imposing a bound on the relation between $w_{j}$ and $u_{j}$ in order to avoid to lose system controllability. However, in practical implementations, the process disturbance variables $\hat{w}_{j\vert k}$ along the control horizon can be omitted to avoid increase the computational burden of the optimization problem.
\begin{remark}
The sequence of process disturbances $\hat{w}_{j\vert k}$ is minimized within the estimator window, i.e., $j\in \left[k-N_e-1,k-1\right]$, and it is maximized within the controller window, $j\in \left[k,k+N_c-1\right]$.
\end{remark}
\subsection{Relationship with MHE and MPC}
The criterion $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}}$ can be rewritten as follows
\begin{equation} \label{MHE+MPC multi_obj}
\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}} \coloneqq \varphi \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}} + (1-\varphi) \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}, \quad \varphi \in [0,1],
\end{equation}
where $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ is the criterion implemented by a \emph{MHE} estimator and $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ is to the criterion implemented by a \textit{robust} \emph{MPC} controller, given by
\begin{equation}
\begin{array}{rl}
\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}} \coloneqq & \Gamma_{k-N_e}\left(\chi\right) + \displaystyle
\sum_{j=k-N_e}^{k} \ell_{e} \left(\hat{w}_{j\vert k}, \hat{v}_{j\vert k}\right),\vspace{2mm}\\
\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}} \coloneqq & \Upsilon_{k+N_c}\left(\Xi\right) + \displaystyle
\sum_{j=k}^{k+N_c-1} \left(\ell_{c} \left(\hat{x}_{j\vert k}, \hat{u}_{j\vert k} \right) - \ell_{w_c}\left( \hat{w}_{j\vert k} \right)\right).
\end{array}
\end{equation}
Equation \eqref{MHE+MPC multi_obj} corresponds to a weighted sum multi-objective formulation of criterion \eqref{mhe mpc opt problem}, where $\varphi$ controls the influence of $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ on $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$. When $\varphi=0$, $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}} \coloneqq \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ and problem \eqref{mhe mpc opt problem} becomes a \emph{robust model predictive control} problem with terminal cost considered by \cite{chen1998quasi}, given that $x_{k}$ is measurable or it is provided by an estimator. On the other case, when $\varphi=1$, $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}} \coloneqq \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ and problem \eqref{mhe mpc opt problem} becomes a \emph{moving horizon estimation} problem considered by \cite{ji2016robust, garcia2016new, muller2017nonlinear, deniz2019robust}, given that the control inputs $u_{j\vert k}$ are computed by a controller. In these cases, the optimization problem \eqref{mhe mpc opt problem} has only one objective and the separation principle needs to be applied since the estimator and the controller are implemented independently.
When $0 < \varphi < 1$, $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ and $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ are simultaneously considered by $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}}$ and the optimization problem \eqref{mhe mpc opt problem} becomes multi-objective. The importance of $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$, and therefore the one of $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$, is defined by $\varphi$ emphasizing or deemphasizing the influence of the estimation problem on the solution. In the case of $\varphi=0.5$, $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,k,$N_e$}}}$ and $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,$N_c$}}}$ have similar influence on the solution of \eqref{mhe mpc opt problem} and it becomes the problem proposed by \cite{copp2017simultaneous}.
\begin{definition}
Let assume points $z_E \in \mathbb{R}^{n_w N_e}\times \mathbb{R}^{n_v(N_e+1)}\times \mathbb{R}^{n_x(N_e+1)} \eqqcolon \mathcal{Z}_E$ and $z_C \in \mathbb{R}^{n_w N_c}\times \mathbb{R}^{n_u N_c}\times \mathbb{R}^{n_x(N_c+1)}\eqqcolon \mathcal{Z}_C$ such that $z\in \mathcal{Z}_E \times \mathcal{Z}_C \eqqcolon \mathcal{Z}$.
A point $z^{o} \in \mathcal{Z}$, is Pareto optimal iff there does not exist another point $z \in \mathcal{Z}$ such that $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,$N_e+N_c$,k}}}(z) \leq \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,$N_e+N_c$,k}}}(z^o)$ and $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}(z_E)<\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}(z_E^o)$, $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}}(z_C)<\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}}(z_C^o)$ \citep{miettinen2012nonlinear}.
\end{definition}
According to this concept, problem \eqref{mhe mpc opt problem} looks for solutions that neither $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}$ nor $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}}$ can be improved without deteriorate one of them.
Any optimal solution of problem \eqref{mhe mpc opt problem}
with $0 < \varphi < 1$ is Pareto optimal \citep{miettinen2012nonlinear}, therefore it has an optimal trade-off between $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}$ and $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}}$. On the other cases, $\varphi=0$ or $\varphi=1$ the solutions of problem \eqref{mhe mpc opt problem} are optimal in the sense of the selected objective ($\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}$ or $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}}$, respectively). In these cases, the solutions obtained are not Pareto optimal and, therefore the overall system performance can be poorer than the one provided by the multi-objective problem.
From a practical point of view, $\varphi$ can be used to improve the numerical properties of the optimization problem (\ref{mhe mpc opt problem}).
This fact allows to improve the convergence properties of the numerical algorithms employed to solve it (see Example 4.2). For example, if $N_e \ll N_c$ and the stage costs $\ell_{e} \left(\cdot \right), \ell_{c} \left(\cdot \right)$ and $\ell_{w_c} \left(\cdot \right)$ have similar values, the optimization problem will improve $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}}$ at the expense of $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}$ (because $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,$N_c$,k}}} \gg \Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{E,$N_e$,k}}}$), deteriorating the estimation of $\hat{x}_{k|k}$ and producing potentially ill conditioned Jacobian and Hessian matrices of $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{EC,k,$N_e+N_c$}}}$. This numerical problems can lead to an increment of the computational times of the optimization problem. A similar situation can happen when $N_e \gg N_c$.
\section{Robust stability under bounded disturbances}
In this section, we introduce the results regarding feasibility and robust stability of the proposed algorithm. Firstly, the properties of \emph{MHE} and \emph{MPC} are analyzed and then the results for the simultaneous \emph{MHE--MPC} are given. Besides, feasibility conditions for the existence of a solution to \eqref{mhe mpc opt problem} and minimum horizon lengths required to achieve the desired estimation and control performances are analyzed.
\subsection{Backward window}
The simultaneous state estimation and control problem relies on a backward window of fixed length $N_e$ to compute the optimal state estimate $\hat{x}_{k\vert k}$. Then, the controller takes the estimate $\hat{x}_{k\vert k}$ as initial condition and predicts the system behaviour. To take advantage of the backward window and reconstruct the state of the system, there have to exists an observer for it, i.e., the system has to be detectable. A definition of detectability for nonlinear systems is \textit{incremental input-output-to-state stability} -\textit{i-IOSS}- \citep{sontag1995characterizations}, and it entails that the difference between any two trajectories of the system can be bounded by
\begin{equation} \label{eq: definicion i-IOSS}
\begin{split}
\vert x_k(x_0^{(1)},\boldsymbol{w}^{(1)} ) - x_k(x_0^{(2)},\boldsymbol{w^}{(2)})\vert &\leq \beta\left( \vert x_0^{(1)} - x_0^{(2)}\vert,k \right) + \gamma_1\left( \|\boldsymbol{w}^{(1)} -\boldsymbol{w}^{(2)}\| \right) \\
& \quad + \gamma_2\left( \|h\left(\boldsymbol{x}^{(1)} \right) -h\left(\boldsymbol{x}^{(2)} \right)\| \right),
\end{split}
\end{equation}
with $\beta(\cdot,\,\cdot)\in\mathcal{KL}$, $\gamma_1(\cdot),\,\gamma_2(\cdot)\in\mathcal{K}$. In the following, we assume that the system is \textit{i-IOSS}, i.e., any two trajectories eventually become indistinguishable one of another. Note that inequality \eqref{eq: definicion i-IOSS} only includes the process disturbance as input to the system. For the case of non-autonomous system, as in the present work, control inputs also have to be taken into account. Since control inputs and process disturbances have the same nature in our context, considering both is straightforward. Moreover, as will be shown later in Example \ref{example: 1}, the control law chosen have not only effects in the forward window but also in the backward window influencing on the estimation process.
Previous results on robust output-feedback \emph{MPC} with bounded disturbances firstly solve the estimation problem and show the convergence of estimated states to a bounded set, then take the uncertainty of estimation into account when solving the \emph{MPC} problem \citep{mayne2006robust, mayne2009robust}. The key idea in these works was to consider the estimation error as an additional, unknown but bounded uncertainty that must be accounted for guaranteeing stability and feasibility of the resulting closed--loop system. Let us define the \textit{robust estimable set}
\begin{equation}
\mathcal{E}_{N_e} \left(\hat{x}_{k\vert k},\varepsilon_{e}(k) \right) \coloneqq \left\{x : \vert x-\hat{x}_{k\vert k} \vert \leq \varepsilon_{e}(k), \forall\, \hat{x}_{k\vert k} \right\}
\end{equation}
where $\hat{x}_{k\vert k}$ is the best estimate available and $\varepsilon_e$ is the estimation error at time $k$ bounded by \citep{deniz2019robust}
\begin{equation} \label{eq: bound estimation error}
\varepsilon_e(k) \leq \bar{\Phi}\left( \vert x_0-\bar{x}_0\vert,k \right)+\pi_w\left(\Vert\boldsymbol{w}\Vert\right)+\pi_v\left(\Vert\boldsymbol{v}\Vert\right).
\end{equation}
Functions $\bar{\Phi}$, $\pi_w$ and $\pi_v$ are defined in term of \emph{MHE} parameters as follows
\begin{align}
\bar{\Phi}\left(\vert x_0-\bar{x}_0 \vert, k\right) \coloneqq&\, \theta^i\vert x_0-\bar{x}_0 \vert^{\zeta} \frac{\mathbb{N}_e}{N_e}\left( \left(\frac{\overline{\lambda}_{P^{-1}} }{\underline{\lambda}_{P^{-1}}}\right)^{\rho}\left(c_{\beta}18^p + \right.\right.\nonumber\\
& \left.\left. \underline{\lambda}_{P^{-1}}^{\alpha_1}\left(P_{k-N_e}^{-1} \right)\left(c_1\;3^{\alpha_1}+ c_2\;3^{\alpha_2}\right)\right) + c_{\beta}\;2^p\right), \label{eq: Phi} \vspace{2mm} \\
%
\pi_w\left(\Vert\boldsymbol{w}\Vert \right) \coloneqq&\, 2\left(1+\mu\right)\left(\frac{c_{\beta}\;18^p}{\underline{\lambda}_{P^{-1}}} \;\bar{\gamma}_w^{\frac{p}{a}}\left(\Vert\boldsymbol{w}\Vert \right) + c_2\; 3^{\alpha_2} \bar{\gamma}_w^{\alpha_2}\left(\Vert\boldsymbol{w}\Vert \right) + \nonumber\right. \\
&\left.\quad\gamma_1\left(6 \Vert\boldsymbol{w}\Vert\right) + \gamma_1 \left(6 \underline{\gamma}_w^{-1} \left( 3 \bar{\gamma}_w\left(\Vert\boldsymbol{w}\Vert \right) \right) \right)\right),\label{eq: pi_w} \vspace{2mm} \\
%
\pi_v\left(\Vert\boldsymbol{v}\Vert \right) \coloneqq&\, 2\left(1+\mu\right)\left(\frac{c_{\beta}\;18^p}{\underline{\lambda}_{P^{-1}}} \;\bar{\gamma}_v^{\frac{p}{a}}\left(\Vert\boldsymbol{v}\Vert \right) + c_1\; 3^{\alpha_1} \bar{\gamma}_v^{\alpha_1}\left(\Vert\boldsymbol{v}\Vert \right) + \nonumber\right. \\
&\left.\quad\gamma_2\left( 6\Vert\boldsymbol{v}\Vert\right) + \gamma_2\left( 6\underline{\gamma}_v^{-1}\left( 3 \bar{\gamma}_v\left(\Vert\boldsymbol{v}\Vert \right) \right) \right)\right),\label{eq: pi_v}
\end{align}
where $\theta = \frac{2+\mu}{2(1+\mu)}<1$, $\mu \in \mathbb{R}_{\geq 0}$, $i=\lfloor \frac{k}{N_e}\rfloor$, $\underline{\lambda}_{P^{-1}}$ and $\overline{\lambda}_{P^{-1}}$ are the minimal and maximal eigenvalues of the arrival-cost weight matrix $P$, respectively. Moreover, the matrix $P$ is updated at each sampling time applying the algorithm developed in \cite{sanchez2017adaptive}.
As in the case of the stage cost, the arrival--cost is lower and upper bounded by
\begin{equation}
\underline{\lambda}_{P^{-1}}\vert \chi \vert^2 \leq \Gamma_{k-N_e}\left(\chi\right) \leq \overline{\lambda}_{P^{-1}} \vert \chi \vert ^2.
\end{equation}
On the other hand, $\zeta$, $\rho$, $c_{\beta}$, $p$, $a$, $c_1$, $c_2$, $\alpha_1$ and $\alpha_2$ are positive real constants whose value depend on the system and parameters of the estimator \citep{deniz2019robust}. The functions $\gamma_1$ and $\gamma_2$ are
related with the system detectability (equation \eqref{eq: definicion i-IOSS}), whereas the functions $\gamma_w$ and $\gamma_v$ are bounds of the stage-cost of the estimator, whose relationship is given by
\begin{equation}
\begin{array}{rcl}
\underline{\gamma}_w\left(\vert \hat{w}_{j\vert k} \vert\right) \leq& \ell_{w_e}\left(\hat{w}_{j\vert k}\right) &\leq \overline{\gamma}_w\left(\vert\hat{w}_{j\vert k}\vert\right),\\
\underline{\gamma}_v\left(\vert \hat{v}_{j\vert k} \vert\right) \leq& \ell_{v_e}\left(\hat{v}_{j\vert k}\right) &\leq \overline{\gamma}_v\left(\vert\hat{v}_{j\vert k}\vert\right),
\end{array}
\end{equation}
and $N_e$ is the length of the backward window. $\mathbb{N}_e$ is the minimum length of the backward window required to guarantee the boundness of the estimation error, which is given by
\begin{equation} \label{eq:minimo N}
\mathbb{N}_e = \ceil*{ \left(2^{\zeta} e_{max}^{\zeta-1} \bar{c}_{\beta} \right)^{\frac{1}{\eta}}},
\end{equation}
where $e_{\max}$ denotes the maximal error on the prior estimate of the initial condition and $\eta\in\mathbb{R}_{\geq 0}$ is a constant. Henceforth, we will assume that $N_e\geq \mathbb{N}_e$.
At each sampling time, the measurements available along the backward window are used to obtain $\hat{x}_{k\vert k}$. Whenever $N_e \geq \mathbb{N}_e$, the estimation error will decrease until it reaches an invariant space whose volume depends on the process and measurement noises as well as the stage- cost and the system itself. The behaviour of the system is forecast from the estimate $\hat{x}_{k\vert k}$, whereas $x_k$ remains within $\mathcal{E}_{N_e}$
\subsection{Forward window}
The forward window corresponds to the \emph{MPC} problem, which computes the optimal control inputs $\boldsymbol{\hat{u}}$ using $\hat{x}_{k\vert k}$ as initial condition. Its feasibility depends on the fact that its initial condition $x_{k}$ must belong to the \textit{robust controllable set} $\mathcal{R}{\textnormal{\fontsize{6pt}{8pt}\selectfont \textit{N}}_c} \left(\Omega,\mathbb{T}\right)$ \citep{kerrigan2000invariant}, which is defined as follows
\begin{align} \label{eq: conjunto robusto controlable}
\mathcal{R}_{\textnormal{\fontsize{6pt}{8pt}\selectfont \textit{N}}_c}\left(\Omega,\mathbb{T}\right) \coloneqq& \left\{x_0\in\Omega|\exists\, u_{j}\in\mathcal{U}\,:\left\{x_{j}\in\Omega, x_{N_c}\in\mathbb{T}\right\} \quad \forall j \in \mathbb{Z}_{\left[0,N_c-1\right]} \right\}.
\end{align}
Since $x_k \in \mathcal{E}_{N_e} \left(\hat{x}_{k\vert k},\varepsilon_{e} \right)$ the feasibility of the control problem is guaranteed if $\mathcal{E}_{N_e} \left(\hat{x}_{k\vert k},\varepsilon_{e} \right) \subseteq \mathcal{R}{\textnormal{\fontsize{6pt}{8pt}\selectfont \textit{N}}}\left(\Omega,\mathbb{T}\right) \quad \forall k\geq 0$, which implies $\mathcal{X}_f\subseteq \mathbb{T}$. Note that this feasibility condition is not only necessary for the simultaneous \emph{MHE--MPC}, but also for independent \emph{MHE} and \emph{MPC} \citep{mayne2006robust,mayne2009robust}. Let us state this condition in the following assumption
\begin{assumption} \label{assumption: feasibility condition}
The robust estimable set $\mathcal{E}_{N_e}$ belong to the robust controllable set $\mathcal{R}_{\textnormal{\fontsize{6pt}{8pt}\selectfont \textit{N}}}\left(\Omega,\mathbb{T} \right)$ in $N_c$ steps for all times $k\geq 0$
\begin{align}
\mathcal{E}_{N_e} \subseteq \Omega,\, \mathcal{X}_f\subseteq\mathbb{T} \rightarrow \mathcal{R}_{N_c}\left(\mathcal{E}_{N_e},\mathcal{X}_f\right) \quad \forall k\geq 0.
\end{align}
\end{assumption}
This assumption states that despite the sequence of control is computed from an estimate $\hat{x}_{k\vert k}$, provided that $x_k$ belong to $\mathcal{R}_{N_c}\left(\mathcal{E}_{N_e},\mathcal{X}_f\right)$, $x_{k+1}\in \mathcal{R}_{N_c}\left(\mathcal{E}_{N_e},\mathcal{X}_f\right)$. Moreover, the volume of the robust estimable set decrease faster with longer backward windows and the size of the robust controllable set can be enlarged by mean of larger forward window and with the appropriate design of the set $\mathcal{U}$.
Regarding stability along the forward window, a common approach to guarantee the stability of \emph{MPC} is by mean of the inclusion of a terminal constraint set, which is generally a level set of a control Lyapunov function \citep{mayne2000constrained}. This set is an artificial constraint set but guarantees stability \citep{tuna2006shorter}. In this work we will analyse the stability of the controller following a similar approach as in \cite{tuna2006shorter}, where the analysis is carried out as a function of the length of the forward window, taking into account the effect of the process disturbances and the estimation errors.
A pseudo measure of the system controllability property will be introduced and the minimum forward window length which guarantees the stability of the simultaneous \emph{MHE--MPC} is given, without imposing extra terminal constraints nor appeal for the cost-to-go to be a closed--loop Lyapunov function (\emph{CLF}). In this sense, let us state the following assumption.
\begin{assumption} \label{assump relaxed contraction of Vf}
There exist a constant $\delta\in \mathbb{R}_{\geq 0}$ such that the cost-to-go and the stage cost satisfy the following relation:
\begin{equation} \label{relaxed contraction of Vf}
\Upsilon_{k+N_c}\left( f\left(x,u\right)\right)-\Upsilon_{k+N_c}\left( x \right) \leq -\ell_{c}\left( x, u \right) + \Upsilon_{k+N_c}\left( x \right)\delta + \ell_{w_c}\left( w \right).
\end{equation}
\end{assumption}
A similar assumption was already used in \cite{tuna2006shorter}, where the constant $\delta$ is introduced in order to relax the requirement on function $\Upsilon_{k+N_c}\left( \cdot \right)$ to be a \emph{CLF} for the nominal case. Despite we use a different notation for the cost-to-go term $\Upsilon_{k+N_c}\left(\Xi\right)$, this function can take the same behaviour as the stage-cost, i.e., $\Upsilon_{k+N_c}\left(\Xi\right)=\ell_{c}\left(\Xi,0\right)$.
Here we extend it to the more general case where process disturbances are affecting the system, and it will lead, as will be shown later, in longer control windows. However, in practical implementation, one can omit process disturbance optimization variables to avoid increasing the computational burden but setting the length of the forward window to the value computed with the process disturbance taken into account.
Regarding the elements of the optimization problem corresponding to the control problem, we will assume that the stage-cost is lower bounded.
\begin{assumption} \label{assumption stage cost controller}
The stage cost $\ell_{c}\left( x,u \right)$ is lower bounded by a function $\sigma\left( x \right) \in \mathcal{K}_{\infty}$, such that $\sigma\left( x \right) \leq \ell_{c}\left( x,u \right) \quad \forall \, x \in \mathcal{X}, u \in \mathcal{U}$.
\end{assumption}
Note that for a quadratic stage-cost, i.e., $\ell_{c}\left(x,u\right)=x^TQx+u^TRu$, with $Q$ and $R$ positive definite matrices, one can choose $\sigma(x)=\underline{\lambda}_Q\vert x \vert^2$, where $\underline{\lambda}_Q$ denotes the minimal eigenvalue of matrix $Q$. Moreover, we will assume that there exist an increasing sequence that relates the function $\sigma(x)$ with the cost of the control problem $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,i}}}$, where $i$ represents different lengths of the forward window.
\begin{assumption} \label{bound of partial cost}
There exists a sequence $\boldsymbol{L}\coloneqq\left[L_0,L_1,\ldots,L_j\right]$, $L_i \in \mathbb{R}$, $1\leq L_i \leq L$, $i \in\mathbb{Z}_{\geq 0}$ that verifies
\begin{equation} \label{Ineq_A4}
\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,i}}} \leq L_{i}\,\sigma\left(x\right).
\end{equation}
\end{assumption}
Choosing
\begin{equation}
L_i=\frac{\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,i}}}}{\sigma(x)},
\end{equation}
satisfies inequality \eqref{Ineq_A4} even for $L_0=1$, since $\Psi_{\textnormal{\fontsize{6pt}{8pt}\selectfont\textit{C,k,0}}}=\Upsilon_k(\Xi)=\sigma(x)$. Finally, let us define the following quantity
\begin{align}
\Delta^w_c\coloneqq
%
\displaystyle \max \,
\left\{\underset{\substack{\hat{u}_{k\vert k}} }{\operatorname{min:}\,}\frac{\ell_{w_c}\left(\hat{w}_{k\vert k}\right)}{\ell_c\left( \hat{x}_{k\vert k}, \hat{u}_{k\vert k} \right)}\right\},\,\forall\,\hat{x}_{k\vert k}\in\mathcal{X},\,\forall\,\hat{w}_{k\vert k}\in\mathcal{W}.
\end{align}
It encodes a pseudo--measure of the system controllability relating the capability of control actions to compensate the process disturbances. The term pseudo--measure is used here because the relation $\Delta^w_c$ is given via the penalization functions $\ell_{w_c}\left(\cdot\right)$ and $\ell_c\left(\cdot,\cdot\right)$. In the following, we will assume that the system is controllable from this point of view.
\begin{assumption} \label{assumption:condition for stability}
The controller of the system can be designed such that the following relation can always be verified
\begin{align}
\Delta^w_c <& 1
\end{align}
\end{assumption}
With the properties established for the backward and forward windows in mind, next we will study the overall stability of the simultaneous \emph{MHE--MPC}.
\subsection{Backward and forward windows}
With all the elements introduced in the previous section, we are ready to derive the main result: the stability of the resulting closed-loop system of the proposed output-feedback controller with estimation horizon $\mathbb{N}_e$ and control horizon $\mathbb{N}_c$ for nonlinear detectable and controllable systems under bounded disturbances.
\begin{theorem} \label{theorem: forward window}
Given the \textit{i}-\emph{IOSS} nonlinear system \eqref{nonlinear system} with a prior estimate $\bar{x}_0 \in \mathcal{X}_0$ of its unknown initial condition $x_0$ and bounded disturbances $\boldsymbol{w} \in \mathcal{W}\left(w_{\max} \right)$, $\boldsymbol{v} \in \mathcal{V}\left(v_{\max} \right)$, Assumptions \ref{assumption: feasibility condition} to \ref{assumption:condition for stability} are fulfilled, the estimation window verifies $N_e \geq \mathbb{N}_e$ and the control horizon $N_c$ verifies
\begin{equation} \label{eq:stabilizinc lenght contrl window}
\mathbb{N}_c = \ceil*{\displaystyle 1+\frac{\ln{\left(\frac{\delta\left(L-1\right)}{1-\Delta^w_c}\right)}}{\ln{\left(\frac{L}{L-1}\right)}}},
\end{equation}
then there will exist at each sampling time $k$ a feasible estimate $\hat{x}_{k-N_e\vert k}$ and feasible sequences $\hat{\boldsymbol{w}}$ and $\hat{\boldsymbol{u}}$ such that
\begin{align}
\Delta \Psi \leq& -\ell_c\left( \hat{x}_{k\vert k}, \hat{u}_{k\vert k} \right)\left( 1 - \delta \omega \right) + \overline{\pi}_E,
\end{align}
where
\begin{equation}
\begin{array}{rl}
\omega \coloneqq& \displaystyle \frac{\Upsilon_{k+N_c}\left(\Xi\right)}{\ell_c\left( \hat{x}_{k\vert k}, \hat{u}_{k\vert k} \right)} +
\displaystyle \frac{1}{\delta} \; \Delta^w_c, \hspace{2cm} 0\leq \delta\omega<1 \vspace{2mm} \\
\overline{\pi}_E \coloneqq & \overline{\gamma}_w\left(\underline{\gamma}_w^{-1}\left( \frac{\overline{\gamma}_p(\chi)}{N_e}+\overline{\gamma}_w(\Vert \boldsymbol{w} \Vert)+\overline{\gamma}_v(\Vert \boldsymbol{v} \Vert) \right)\right).
\end{array}
\end{equation}
\end{theorem}
\noindent \textbf{Proof.} See Appendix \ref{Proof_T1}. {\hfill$\square$} \vspace{0.1cm}
\section{Examples}
In this section, we discuss two examples to illustrate the results presented previously and compare the performance of the framework discussed formerly. The first example applies the ideas introduced in previous sections to a nonlinear scalar system. The emphasis is placed in the effect of constraints and disturbances on closed-loop stability and performance. The second example discusses the simulations results for a van der Pol oscillator using the framework discussed in previous sections. The discussion is focused on the effect of $N_e$ and $N_c$ on the performance and computational time.
\subsection{Example 1} \label{example: 1}
Let us consider the continuous--time nonlinear scalar system
\begin{equation} \label{eq:system example 1}
\begin{array}{rl}
\Dot{x} =& a x^3_t + w_t + u_t, \quad a\in\mathbb{R}_{>0}\\
y_t =& x_t + v_t.
\end{array}
\end{equation}
Firstly, we show its detectability, i.e., the existence of an estimate with a structure like equation \eqref{eq: definicion i-IOSS}. Let assume two arbitrary and feasible trajectories $x^{(1)}_t$ and $x^{(2)}_t$ such that $\Delta x \coloneqq x^{(1)}_t -x^{(2)}_t$ and $p_t \coloneqq \vert \Delta x \vert$; then $\Dot{p}_t$ can be written as follows
\begin{equation}
\Dot{p}_t = \frac{\Delta x}{\vert \Delta x \vert}\left(\Dot{x}^{(1)}_t-\Dot{x}^{(2)}_t\right).
\end{equation}
Assuming a \emph{LTV} control law $u_t=-K_t x_t$, we obtain
\begin{equation}
\Dot{p}_t = \frac{\Delta x}{\vert \Delta x \vert}\left(a\Delta x\left(x^{(1)^2}_t+x^{(1)}_t x^{(2)}_t+x^{(2)^2}_t\right)-K_t \Delta x+\Delta w_t\right),
\end{equation}
which is upper bounded by
\begin{equation}
\Dot{p}_t \leq -K_t \, p_t + a \, g\vert \Delta h_t \vert +\vert\Delta w_t \vert,
\end{equation}
where
\begin{equation}
g\coloneqq h^2\left(x^{(1)}_0\right)+h\left(x^{(1)}_0\right) h\left(x^{(2)}_0\right) + h^2\left(x^{(2)}_0\right).
\end{equation}
Solving $p_t$ for initial condition $p_0 = \vert x^{(1)}_0-x^{(2)}_0\vert$ we obtain
\begin{equation} \label{eq:i-ioss bound example 1}
\vert x^{(1)}_t - x^{(2)}_t \vert \leq \, \vert x^{(1)}_0 - x^{(2)}_0 \vert e^{-K_t t} + \frac{\Vert \Delta w_{0:t} \Vert}{K_t} + \frac{a g \Vert \Delta y_{0:t} \Vert}{K_t},
\end{equation}
it follows the fact that system \eqref{eq:system example 1} is \textit{i}-\emph{IOSS} (for all details, the reader can refer to appendix \ref{Deriv_p}).
In the case of \emph{MHE--MPC} controllers with quadratic costs
\begin{equation} \label{stag&term_costs}
\begin{array}{rl}
\ell_e \coloneqq \hat{w}^2_{j|k} Q_e + \hat{v}^2_{j|k} R_e, &
\Gamma_{k - N_e} \coloneqq P^{-1}_{k - N_e} \chi^2, \\
\ell_c \coloneqq \hat{x}^2_{j|k} Q_c + \hat{u}^2_{j|k} R_c, &
\Upsilon_{k+N_c} \coloneqq S_{c} \Xi^2,
\end{array}
\end{equation}
analysed in this work, the bound \eqref{eq:i-ioss bound example 1} can be written as follows
\begin{equation} \label{eq:estimation error bound example 1}
\begin{array}{rl}
\vert x_k-\hat{x}_{k\vert k} \vert \leq&\, \vert x_0-\bar{x}_0 \vert \left(\frac{\theta\,\mathbb{N}_e}{2N_e}\right)^i\left(2+\frac{\left(P^{-1}_{k-N_e}R_e\right)^{1/2}+\left(P^{-1}_{k-N_e}Q_e\right)^{1/2} \,a \, g}{\left(Q_eR_e\right)^{1/2}K^{c}_{1\vert k}}\right)\\
& + 2\left(1+\mu\right)\Vert \boldsymbol{w} \Vert \left( \frac{2}{K^{c}_{1\vert k}} + \frac{\left(Q_e \, R_e\right)^{1/2} K^{c}_{1\vert k} +\left(P^{-1}_{k-N_e} \, Q_e\right)^{1/2} \, a \, g}{\left(P^{-1}_{k-N_e} \, Q_e \right)^{1/2} K^{c}_{1\vert k}} \right)\\
& + 2\left(1 + \mu\right)\Vert\boldsymbol{v} \Vert \left(\frac{2 \, g }{K^{c}_{1\vert k}}+\frac{\left(R_e \, Q_e\right)^{1/2} K^{c}_{1\vert k} + \left(P^{-1}_{k-N_e} \, R_e\right)^{1/2}}{\left(P^{-1}_{k-N_e}Q_e\right)^{1/2} K^{c}_{1\vert k}}\right),
\end{array}
\end{equation}
with $\mathbb{N}_e$ given by
\begin{equation} \label{eq:Ne example 1}
\mathbb{N}_e = \ceil*{4\left( 2 + \frac{\left(P^{-1}_{k-N_e} \, R_e\right)^{1/2}+\left( P^{-1}_{k-N_e} \, Q_e \right)^{1/2} a \, g}{\left( Q_e \, R_e \right)^{1/2}K^{c}_{1\vert k}} \right)^2},
\end{equation}
and $K^{c}_{1\vert k}$ is the equivalent controller gain resulting from applying $\hat{u}_{1|k}$.
Equations \eqref{eq:estimation error bound example 1} and \eqref{eq:Ne example 1} show the influence of the controller on the state estimation. Larger controller gains improve estimation error and shorten the convergence time. However, controller gains are bounded by robust stability conditions and input constraints, which are limiting factors in this potential improvement. This example highlights the relevance of simultaneously solving the estimation and control problems, or at least to take into account the solution of control problem on the estimation one. Since \emph{MPC} gains $K^{c}_{1 \vert k}$ are time-varying because they are recomputed at every sampling time, a conservative approach can employ its lowest value.
In order to compare the performances of independent and simultaneous \emph{MHE--MPC}, both output--feedback controllers have the same parameter
\begin{equation} \label{Parameter_MPC}
P_0=10^5,\, Q_e=15,\, R_e=10^3,\, Q_c=5,\, R_c=5,\, S_c=Q_c,\, \mu=0.05,
\end{equation}
with constraints sets
\begin{equation} \label{Constr_S01}
\mathcal{X}\coloneqq\left\{x:\vert x \vert \leq 0.8 \right\}
\text{ and }
\mathcal{U}\coloneqq\left\{u:\vert u \vert \leq 2.5 \right\},
\end{equation}
$w_t \sim \textnormal{\textit{U}}\left(0,\,0.01\right)$, $v_t \sim \mathcal{N}\left(0,\,0.02^2\right)$, $a=1$, $g=3x_{0 \, \max}^3$ and $\varphi=0.5$ such that both controllers implement the same optimization criterion.
The control problems of both controllers are configured without terminal constraints. The process disturbance is not taken into account to compute $\hat{u}_{j|k},$ but it will be considered in the computation of $N_c$. It can be computed directly from equation \eqref{eq:stabilizinc lenght contrl window} once the values of $\delta$, $L$ and $\Delta^w_c$ had been established. Another approach, employed in this example, consists of computing $\omega$ through simulations. In this example, we set the initial condition that maximizes the controller costs and then computes the values of $L$ and $\omega$. The process is repeated until reach the maximal value of $N_c$.
\begin{figure}[thb]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/omega-Sw10e-1-2.eps}
\caption{}
\label{figure:example1 omega-a}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/omega-Sw10e-1-moreConstrained-2.eps}
\caption{}
\label{figure:example1 omega-b}
\end{subfigure}
\caption{${\omega}(N_c)$ for $\delta=1, \Delta^w_c=10^{-1}$ and constraints sets \eqref{Constr_S01} (\ref{figure:example1 omega-a}) and \eqref{Constr_S02} (\ref{figure:example1 omega-b}) for independent (red) and simultaneous (blue) approaches.}
\label{figure:example1 omega}
\end{figure}
Figure \ref{figure:example1 omega} shows the computed values of ${\omega}$ a funtion of $N_c$ ($\omega(N_c)$) for the same $\Delta^w_c$ and different set of constraints and distributions for process and measurement noises. In this figure the effect of constraints on ${\omega}(N_c)$ can be seen: They increase ${\omega}(N_c)$, for the same $N_c$, depending how the controller is implemented. This change is smaller for the simultaneous \emph{MHE--MPC} approach than the independent one. When constraints are no relevant (constraints set \eqref{Constr_S01}), both controllers have similar values (see Figure \ref{figure:example1 omega-a}), and the control problem of both controllers can use the same $N_c$. However when constraints are tighten (constraints set \eqref{Constr_S02}), the way of solving the estimation and control problems has a direct effect on ${\omega}(N_c)$ (see Figure \ref{figure:example1 omega-b}), and the control problem of both controllers must use different $N_c$ in order to ensure robust stability, affecting the computational requirements of the implementation. Since we are using constraints set \eqref{Constr_S01} we choose $N_c=10$ for both controllers (Figure \ref{figure:example1 omega-a}). Finally, the minimum estimation horizon $\mathbb{N}_e$ is computed from \eqref{eq:minimo N} using the parameters listed in \eqref{Parameter_MPC}, leading to $\mathbb{N}_e=27$ for both controllers. We choose $N_e=30$ for both controllers.
\begin{figure}[thb]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/2trajec-mhempcSimandInd.eps}
\caption{}
\label{figure:i-ioss bound example 1-a}
\end{subfigure}
\hfill
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/i-ioss-bound-mhempcSimAndInd.eps}
\caption{}
\label{figure:i-ioss bound example 1-b}
\end{subfigure}
\caption{Evolution of system output for different initial conditions, difference between trajectories and \textit{i}-\emph{IOSS} bound.}
\label{figure:i-ioss bound example 1}
\end{figure}
Figure \ref{figure:i-ioss bound example 1} shows the system responses and the corresponding \textit{i}-\emph{IOSS} bound for the regulation problem. Figure \ref{figure:i-ioss bound example 1-a} shows two trajectories generated by both controllers from different initial condition ($x_0^{(1)}=0.766$ and $x_0^{(2)}=-0.766$) with the same prior guess ($\bar{x}_0=-2.5$). Figure \ref{figure:i-ioss bound example 1-b} shows the difference between the trajectories and its \textit{i}-\emph{IOSS} bound, for the minimum controller gain along the simulation ($K^{c}_{1|k}=0.7326$). One can see in this figure the decreasing behaviour of the estimation error bound, as expected from equation \eqref{eq: bound estimation error} for the general case and \eqref{eq:estimation error bound example 1} for this particular example. Despite the small value of $\mu$ ($\mu=0.05$), the bound \eqref{eq:estimation error bound example 1} is quite conservative. In these figures, we can also see that both controllers provide a similar response, since constraints and disturbances have not relevant effect on the system behaviour, and therefore the separation principle can be applied.
Now let us compare the performance in a more challenging setup. In the following, we will assume the next constraints set
\begin{equation} \label{Constr_S02}
\mathcal{U}\coloneqq\left\{u:\vert u \vert \leq 0.6 \right\},
\mathcal{W}\coloneqq\left\{w:\vert w \vert \leq 0.4 \right\}
\text{ and }
\mathcal{V}\coloneqq\left\{v:\vert v \vert \leq 0.8 \right\}.
\end{equation}
The controls $\hat{u}_{j|k}$ have been tightened and the estimates $\hat{w}$ and $\hat{v}$ have been constrained to the sets $\mathcal{W}$ and $\mathcal{V}$, respectively. Disturbances $w_t$ and $v_t$ are now given by $w_t \sim \textnormal{\textit{U}}\left(0,\,0.1\right)$ and $v_t \sim \mathcal{N}\left(0,\,0.2^2\right)$, respectively.
\begin{figure}[thb]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/twelveTrajBothAlg-Nc20.eps}
\caption{}
\label{figure:larger control horizons-a}
\end{subfigure}
\hfill
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/twelveTrajBothAlg-Nc70.eps}
\caption{}
\label{figure:larger control horizons-b}
\end{subfigure}
\caption{Evolution of system output for $N_c=20$ (\ref{figure:larger control horizons-a}) and $N_c=70$ (\ref{figure:larger control horizons-b}), with $N_e=30$ for independent \emph{MHE} and \emph{MPC} (red line) and simultaneous \emph{MHE--MPC} (blue dotted line).}
\label{figure:larger control horizons}
\end{figure}
Under this new operational conditions $\mathbb{N}_e$ is recomputed, obtaining $\mathbb{N}_e=98$ for the independent \emph{MHE} and \emph{MPC}, and $\mathbb{N}_e=52$ for the simultaneous \emph{MHE--MPC}. This is the effect of constraints set \eqref{Constr_S02} on the estimator parameters, while the effect on the controller is shown in Figure \ref{figure:example1 omega-b}. This figure shows that the independent \emph{MHE} and \emph{MPC} approach is more sensitive to disturbances, requiring conservative values of $N_c$ to guarantee the closed--loop stability.
Finally, Figures \ref{figure:larger control horizons} show the system responses for regulation problem for different realizations of $w_t$ and $v_t$ and different $N_c$, for $N_e=30<\mathbb{N}_e$. The independent \emph{MHE} and \emph{MPC} strategy fails to regulate the system states for some noise realizations, even though it regulates few of them. On the other hand, the simultaneous \emph{MHE--MPC} controller manages to regulate the system states for all noise realizations. This problem is caused by the failure of the independent \emph{MHE} and \emph{MPC} to satisfy Assumption \ref{assumption: feasibility condition}. In fact, its design procedure applies the \textit{separation principle}, which entails the automatic satisfaction of Assumption \ref{assumption: feasibility condition} and it does not include the constraints information in the selection of $N_e$ and $N_c$. On the other hand, the simultaneous \emph{MHE--MPC} controller does.
\subsection{Example 2} \label{example: 2}
Let us consider the van der Pol oscillator whose dynamic is described by
\begin{equation} \label{Osc_vdP}
\begin{array}{rl}
\Dot{x}_t =& \left[\begin{array}{c}
\epsilon\left(1-x_{2,t}^2 \right)x_{1,t}-2x_{2,t}+u_t+w_{1,t} \\
2x_{1,t}+w_{2,t}\end{array}\right] \quad \epsilon\in\mathbb{R}_{\geq 0},\vspace{1.5mm} \\
y_t =& \frac{1}{2}\left(x_{1,t}+x_{2,t}\right)+v_{1,t}.
\end{array}
\end{equation}
It is known to be \textit{i}-\emph{IOSS}, and a proof of this property can be made using the averaging lemma \citep{pogromsky2015stability}.
In this example we will focus the analysis on the system performance under different set of parameters. The independent and simultaneous \emph{MHE--MPC} controllers have the same parameters to allow a direct comparison of their performances. All the stage costs have a quadratic structure (equation \eqref{stag&term_costs}) and their parameters are
\begin{equation*} \label{Parameter_MPC02}
P_0=10^5,\, Q_e=\begin{bmatrix} 50 & 0\\0 & 50 \end{bmatrix},\ R_e=150,\, Q_c=\begin{bmatrix} 200 & 0\\0 & 200 \end{bmatrix},\,S=Q_c,\, R_u=10^{-2},
\end{equation*}
with constraints sets given by
\begin{equation} \label{Const_O}
\mathcal{X} \coloneqq \left\{x:\vert x_1 \vert\leq 5, \vert x_2\vert\leq 5\right\},\, \mathcal{U} \coloneqq \left\{u:\vert u \vert \leq 5, \vert \Delta u_k \vert\leq 2 \right\},
\end{equation}
$w_t \sim \textnormal{\textit{U}}\left(0,0.25\right)$ and $v_t \sim \textnormal{\textit{U}}\left(0,0.025\right)$, instead of zero mean normal distribution, as it is common in the literature.
The effect of $N_e$ and $N_c$ on closed-loop performance is be analysed for the following values
\begin{equation}
N_e \coloneqq \{2,5,10,20\},\, N_c \coloneqq \{5,10,35\}.
\end{equation}
Since the difference between $N_e$ and $N_c$ can lead to unbalanced cost functions (emphasizing the control cost over the estimation one), which can deteriorate the overall closed-loop performance. To avoid this problem, $\varphi$ is used to improve the closed-loop performance. It takes the following values $\varphi \coloneqq \{0.95,0.95,0.85,0.65\}$ for the corresponding $N_e$ value.
\begin{figure}[b]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/errEps=0dot1.eps}
\caption{}
\label{figure:example 2, epsilon=1-3-a}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{Figures/errEps=3.eps}
\caption{}
\label{figure:example 2, epsilon=1-3-b}
\end{subfigure}
\caption{\emph{MSE} of $100$ simulations for different values of $N_e$, $N_c$ and $\epsilon$}
\label{figure:example 2, epsilon=1-3}
\end{figure}
Figures \ref{figure:example 2, epsilon=1-3} summarize the mean square error (\emph{MSE}) obtained by both controllers along $100$ simulations for $\epsilon = 0.1$ and $\epsilon = 3$ respectively. These figures show the superior performance of the simultaneous \emph{MHE--MPC} for any combination of $N_e-N_c$ and scenario. In general, there are no meaningful changes of \emph{MSE} with $N_c$, however closed-loop performance varies with $N_e$. Figure \ref{figure:example 2, epsilon=1-3-a} shows the results for $\epsilon=0.1$. In this case the independent \emph{MHE} and \emph{MPC} performance improves with $N_e$, while the simultaneous \emph{MHE--MPC} ones remains similar (a deviation lower than $8\%$ from the average) for any combination of $N_e-N_c$. For this value of $\epsilon$, the system \eqref{Osc_vdP} behaves like a harmonic oscillator, therefore the closed-loop performance depends on the estimation error (see Figure \ref{figure:example 2, epsilon=1 x1 and x2}), which decreases for larger values of $N_e$. Figure \ref{figure:example 2, epsilon=1-3-b} shows the results for $\epsilon=3$. In this condition, the performance of both controllers deteriorates with $N_e$, because for this value of $\epsilon$ the system \eqref{Osc_vdP} behaves like a non-linear dampened oscillator and the state estimates take longer to converge to the estimation invariant set (see Figures \ref{figure:example 2, epsilon=1 x1 and x2-a} and \ref{figure:example 2, epsilon=1 x1 and x2-b}).
\begin{figure}[thb]
\centering
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth,height=145pt]{Figures/x1_Ne=2_Nc=35.eps}
\caption{}
\label{figure:example 2, epsilon=1 x1 and x2-a}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth,height=145pt]{Figures/x2_Ne=2_Nc=35.eps}
\caption{}
\label{figure:example 2, epsilon=1 x1 and x2-b}
\end{subfigure}\\
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth,height=145pt]{Figures/x1_Ne=20_Nc=35.eps}
\caption{}
\label{figure:example 2, epsilon=1 x1 and x2-c}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth,height=145pt]{Figures/x2_Ne=20_Nc=35.eps}
\caption[l]{}
\label{figure:example 2, epsilon=1 x1 and x2-d}
\end{subfigure}
\caption{Two realizations of $x_1$ and $x_2$ for $\epsilon=0.1$, $N_e=2$ (\ref{figure:example 2, epsilon=1 x1 and x2-a} - \ref{figure:example 2, epsilon=1 x1 and x2-b}), $N_e=20$ (\ref{figure:example 2, epsilon=1 x1 and x2-c} - \ref{figure:example 2, epsilon=1 x1 and x2-d}) and $N_c=35$.}
\label{figure:example 2, epsilon=1 x1 and x2}
\end{figure}
Figures \ref{figure:example 2, epsilon=1 x1 and x2} show the simulations resulting from two noise realizations for $N_c=35$, $N_e=2$ and $N_e=20$, respectively. They show that the simultaneous \emph{MHE--MPC} manages to regulate both states and it achieves a better performance than the independent one. While Figures \ref{figure:example 2, epsilon=1 x1 and x2-a} and \ref{figure:example 2, epsilon=1 x1 and x2-c} show that independent \emph{MHE} and \emph{MPC} achieves a better performance than the simultaneous one for state $x_1$, Figures \ref{figure:example 2, epsilon=1 x1 and x2-b} and \ref{figure:example 2, epsilon=1 x1 and x2-d} show how it fails to regulate state $x_2$ for short estimation horizons. Under this condition, $x_2$ has an offset that it is not compensated by the controller. Only large values of $N_e$ allow the independent \emph{MHE} and \emph{MPC} to regulate $x_2$ (Figure \ref{figure:example 2, epsilon=1 x1 and x2-d}). On the other hand, the simultaneous \emph{MHE--MPC} regulates both states and it takes shorter times than the independent one to regulate both states.
\begin{figure}[thb
\centering
\begin{tabular}{c}
{\includegraphics[width=0.8\textwidth]{Figures/execution-times-Ne10-Nc10.eps}}
\end{tabular}
\caption{Average execution over 100 trials for $N_e=N_c=10$.}
\label{figure:execution times}
\end{figure}
The computational burden of the simultaneous \emph{MHE--MPC} is lower than the independent one, as can be seen in Figure \ref{figure:execution times}. The execution times were averaged over $100$ trials. The lower time, in the beginning, is due to the backward window corresponding to the estimation has not achieved yet its full length.
\section{Conclusions}
We presented an output-feedback approach for nonlinear systems subject to bounded disturbances using \emph{MHE--MPC}. The proposed approach combines the state estimation and control problems into a single optimization, which is solved at each sampling time. Theorem \ref{theorem: forward window} states the necessary conditions to guaranty the feasibility and stability of the optimization problem, and therefore the boundedness of system states, as a function of the windows lengths $N_e$ and $N_c$. This result requires the compatibility between the robust estimated and controllable sets (Assumption \ref{assumption: feasibility condition}) and the existence of a relaxed closed--loop Lyapunov function for the disturbed system (Assumption \ref{relaxed contraction of Vf}). These conditions imply forward ($N_c$) and backward ($N_e$) horizons to find state estimates and control actions that are consistent with the system dynamics, constraints and disturbances.
Future work may involve the design of the forward window with properties that allow the improvement of the estimation process and the design of an adaptive law to compute $\varphi$ such that the estimation and control problems keep balanced and the overall system performance and numerical properties are improved.
\begin{ack}
The authors wish to thank the Consejo Nacional de Investigaciones Cientificas y Tecnicas (CONICET) from Argentina, for their support.
\end{ack}
\bibliographystyle{agsm
\clearpage
|
\section{OLD}
\begin{theorem}
This is a sample theorem. The run-in heading is set in bold, while
the following text appears in italics. Definitions, lemmas,
propositions, and corollaries are styled the same way.
\end{theorem}
\begin{proof}
Proofs, examples, and remarks have the initial word in italics,
while the following text appears in normal font.
\end{proof}
\begin{table}
\caption{Table captions should be placed above the
tables.}\label{tab1}
\begin{tabular}{|l|l|l|}
\hline
Heading level & Example & Font size and style\\
\hline
Title (centered) & {\Large\bfseries Lecture Notes} & 14 point, bold\\
1st-level heading & {\large\bfseries 1 Introduction} & 12 point, bold\\
2nd-level heading & {\bfseries 2.1 Printing Area} & 10 point, bold\\
3rd-level heading & {\bfseries Run-in Heading in Bold.} Text follows & 10 point, bold\\
4th-level heading & {\itshape Lowest Level Heading.} Text follows & 10 point, italic\\
\hline
\end{tabular}
\end{table}
\noindent Displayed equations are centered and set on a separate
line.
\begin{equation}
x + y = z
\end{equation}
Please try to avoid rasterized images for line-art diagrams and
schemas. Whenever possible, use vector graphics instead (see
Fig.~\ref{fig1}).
Overview
Include graphic of approach high level steps
STL Learning with Genetic Algorithm
Input of data layers, parameter set and initial formulas
Include graphic with formula set examples
Pseudocode with how do multi class learning
Multi-Class Labeling approach
Labels → glycemic metrics https://docs.google.com/document/d/10mc2k5Bi33GSdqnJuR5tfOF1SGjZ6ZuW2rvQYikuqSk/edit
Graphic of metrics and their meaning, calculation and temporal rate
Steps: one vs all learning, synthesis methodology of final rules from binary classifiers
\section*{\centering Acknowledgements}
The authors would like to graciously thank the UVA Center for Diabetes Technology for providing the clinical datasets and Basak Ozaslan, Jack Corbett, Jonathan Hughes and Dr. Jos\'e Garc\'ia-Tirado for their clinical insights and valuable discussions.
Research partially supported by the Austrian National Research Networks RiSE/ShiNE (S11405) and ADynNet (P28182) of the Austrian Science
Fund (FWF).
\section*{\centering Appendix}\label{sec-appendix}
\begin{table}[htb]
\caption{Accuracy Rates for Repeated Rules}\label{table-mcr-temp-rules}
\centering
\begin{tabular}{|c|c|c|c|c|} \hline
\textbf{Patient} & \textbf{CGM \%} & \textbf{HR \%} & \textbf{Basal \%} & \textbf{Bolus \%}\\ \hline
1 & 88.61 & 51.25 & 86.94 & 90 \\ \hline
2 & 93.88 & 97.05 & 93.24 & 100\\ \hline
3 & 90.58 & 77.02 & 100 & 93.72 \\ \hline
4 & 96.10 & 56.25 & 98.48 & 98.48 \\ \hline
5 & 87.10 & 94.17 & 100 & 95.28 \\ \hline
6 & 88.19 & 51.25 & 100 & 97.92 \\ \hline
7 & 93.61 & 83.55 & 96.26 & 94.40 \\ \hline
8 & 87.08 & 78.89 & 99.72 & 100 \\ \hline
9 & 88.09 & 88.78 & 100 & 93.77 \\ \hline
10 & 95.45 & 97.47 & 100 & 95.36 \\ \hline
11 & 86.76 & 86.46 & 100 & 87.05 \\ \hline
12 & 93.75 & 55.17 & 100 & 95.63 \\ \hline
13 & 95.59 & 61.25 & 100 & 96.12 \\ \hline
14 & 94.40 & 79.33 & 96.38 & 100 \\ \hline
15 & 86.86 & 93.50 & 88.24 & 91.29 \\ \hline
16 & 89.38 & 75.47 & 90.90 & 89.13 \\ \hline
17 & 87.38 & 100 & 100 & 93.07 \\ \hline
18 & 89.54 & 71.09 & 90.66 & 89.65 \\ \hline
19 & 90.29 & 63.38 & 90.15 & 89.88 \\ \hline
20 & 89.54 & 62.43 & 91.11 & 89.99 \\ \hline
21 & 86.86 & 66.99 & 89.88 & 88.35 \\ \hline
\end{tabular}
\end{table}
\bibliographystyle{splncs04}
\section{Discussion \& Conclusion}\label{sec-conclusion}
\startpara{Conclusion.} In this paper, we presented an approach to learn STL formulas that characterize individual- and population-level T1D patient behaviors with varying glycemic control and applied it to a clinical dataset with 21 T1D patients' data.
Our learning results provide some clinically-relevant insights for clinicians and patients to develop behavioral change strategies to improve glycemic control.
\startpara{Tool Limitations.} Our results are constrained by the limitations of the STL learning tool~\cite{nenzi2018robust} in several ways. First, our patient data contain many null values. For example, patients only eat at discrete time periods, and times the patient was not eating were null. However, the tool cannot handle null values, so we had to fill all of these instances with zeros: this changes the semantic meaning of the data points, and may cause a bias in how the formula parameters are being generated (for instance, when data points are averaged to get specific parameter bounds). Second, since the tool cannot handle multi-class classification, we had to use four different sets of labels with binary indicators to cover our different classes. This may have caused some overlap in our resulting formulas. In addition, since the tool relies on a supervised classification approach, we had to supply labels to guide the learning. However, this may have resulted in missing some behavior sets that still have an effect on T1D glycemic control (but may not have a direct relationship with CGM time in range). Moreover, the tool relied on having an evenly split distribution of data labels, which proved challenging for our unevenly distributed patient data. Finally, the tool can only learn from raw data streams for short time periods, (and can not, for instance, calculate CGM rate of change or other advanced relationships), and as such we were only able to learn fairly simple rules for short time chunks. As a result, we were unable to study longer term T1D effects (e.g. multiple hour meal-bolus relationships).
\startpara{Future Work.} We would like to address the limitations and improve upon the capabilities of the tool, as well as integrate our patient behavior identification approach into a closed loop feedback system (e.g., implemented in a smartphone application or other wearable), which will provide real-time feedback about behaviors that have negative impact on glycemic control.
\section{Introduction}\label{sec-intro}
Type I Diabetes (T1D) is a chronic disease in which the body's ability to synthesize insulin is destroyed, as the patient's immune system attacks the insulin-producing cells of the pancreas~\cite{cdcT1d2018}. Insulin is an important hormone used by cells to absorb glucose for energy production.
425 million people worldwide have Diabetes (Type I and Type II), including 1,106,500 children and adolescents living with T1D~\cite{idfatlas2017}.
Intensive insulin therapy effectively reduces the risk of long-term complications of T1D (such as
nerve or kidney damage) in which patients are required to inject or infuse insulin throughout the day to replace the normal pancreas function. Unfortunately, this means the burden of managing T1D
falls to patients as they are
required to manage a variety of behavioral factors (e.g., insulin injection, exercise, eating) that affect T1D.
Studies have found that such factors affect a patient's overall glycemic control: e.g., exercise may lower blood glucose values
while
carbs from meals increase blood glucose levels~\cite{american201913,Riddell2017}.
\figref{fig-behavior-ex} shows a set of hypothetical patient behaviors that may result in varying glycemic control. For example, on days when
a patient exercises in the evening and underestimates the insulin absorption amount, they may have poor glycemic control (hypoglycemia) the next morning.
Characterization of these behaviors can be used by clinicians to counsel their patients on strategies to optimize glycemic control using predictive recommendations (e.g., if you exercise late at night, make sure you eat a snack before you go to bed to avoid morning hypoglycemia). However, it is challenging to accurately identify T1D patient behavior patterns due to inherently messy patient data and the individual variability of patient behavior and physiology.
\begin{figure}[t]
\centering
\includegraphics[scale=0.28]{behaviorExample.eps}\label{fig-behavior-ex}
\caption{Hypothetical patient behaviors resulting in different glycemic control outcomes.}
\vspace{-15pt}
\end{figure}
In this paper, we present a logic-based learning approach to address these challenges and explore T1D patient behaviors. Our approach
takes advantage of the expressiveness and explainability of Signal Temporal Logic (STL)~\cite{maler2004monitoring} and
uses STL learning~\cite{nenzi2018robust}
to learn a set of STL formulas that characterize both individual- and population-level T1D patient behaviors.
We argue that STL is a suitable representation of patient behavior patterns, because it can capture the temporal relations of Diabetes patient actions and glycemic outcomes.
In addition, STL formulas are easily explainable to clinicians and patients.
We apply our approach to learn STL formulas representing T1D patient behaviors from a clinical dataset including a variety of patient physiological and behavioral data,
such as Continuous Glucose Monitors (CGM) sensor readings, heart rate, step count and activity intensity recorded by Fitbit, insulin pump injection records, self-reported meals and blood glucose finger pricks (SMBG).
We envision that the learned STL formulas can provide clinically-relevant insights for
clinicians and patients to develop behavioral change strategies to improve glycemic control.
The rest of the paper is organized as follows:
\sectref{sec-prelim} introduces the background of STL and learning techniques.
\sectref{sec-methods} describes our approach to learn STL formulas for characterizing T1D patient behaviors.
\sectref{sec-res-individual} and \sectref{sec-res-pop} present our key findings about individual- and population-level patient behaviors, respectively.
\sectref{sec-relatedwork} summarizes related work, and
\sectref{sec-conclusion} draws conclusions and discusses future research directions.
\section{Methodology}\label{sec-methods}
We first describe the clinical dataset, then present our approach of learning individual- and population-level patient behaviors as illustrated in \figref{fig-methods}.
\subsection{Clinical Dataset Description}
Our dataset was collected during the observation period leading up to inpatient clinical trials in 2016-2017 at the Center for Diabetes Technology at the University of Virginia. The dataset contains 21 patients, ages ranging from 17 to 55, with an average age of 36 $\pm$ 10.4. Each patient has about 2 months of consecutive data.
The data includes blood glucose readings from a Continuous Glucose Monitor (recorded in a variable named CGM), different types of insulin injections called boluses (total bolus, meal bolus, basal bolus, and correction bolus), meal carbs, patient-recorded blood glucose values from a finger prick (SMBG) and recordings of hypoglycemia (SMBG-Hypo). The data also contains exercise data recorded from a Fitbit including Heart Rate (HR), step count, calories, distance (in miles), and a Fitbit calculated activity level (in range of 1 to 4, with 1 being equivalent to little activity, and 4 being equivalent to intense activity.)
\subsection{STL Learning for Individual Patient Behaviors}\label{sec-methods-individual}
The approach of learning for individual behaviors is shown in \figref{fig-methods} in the top yellow flowchart. We first pre-processed the data, and then added a multi-class labeling mechanism for our unlabeled patient data using CGM time in range, based on medical domain knowledge. Next, we fed our data and labels into our STL learning tool, to output STL formulas that classify specific patient behaviors. Finally, our results were validated for clinical insights. Each of these steps is explained in greater detail below.
\begin{figure}[t]
\centering
\includegraphics[scale=0.29]{methods.eps}
\caption{Approach overview for learning STL formulas representing individual- (top yellow flowchart) and population-level (bottom blue flowchart) patient behavior patterns.}
\label{fig-methods}
\vspace{-15pt}
\end{figure}
\startpara{Data Pre-processing.} As our clinical data was messy and sampled at different rates,
the first step in our methodology was to pre-process the data. We combined all data variables into a single file, and aligned them on a five minute sampling rate, to match the set sampling rate of the CGM. The variables HR and steps were sampled at more frequent rates (data was recorded a couple times per minute,) and we used a sliding average to compute the HR value, and summed the total steps in the time frame to align with each five minute interval. In addition, we added a detector to indicate when patients were exercising. For the purposes of our approach, we determined that a patient was exercising when the Fitbit Activity Level was $\geq$ 3,
and/or when the patient had $\geq$ 3000 steps in 30 minutes, following approaches used to detect exercise in~\cite{bumgardner,marshall2009translating}. Finally, the data was layered into one hour time chunks to be fed into our STL Learning algorithm. We choose a one hour time chunk such that we would have enough data points (12 points) per layer
for interesting learning to happen, but also small enough to provide detailed granularity within each patient's data.
\startpara{Labeling.} Next, we added labels for each one hour time chunk by hand. We used CGM Time in Range~\cite{fabris2016risk,kovatchev2017metrics}---the percentage of time a patient spends in a well controlled blood glucose range (between 70 and 180 mg/dL)---as our labeling mechanism. This metric is commonly used by clinicians to determine how well controlled patients are, and as such served as an appropriate labeling mechanism for our data. We developed 4 sets of labels based on the total percentage of time the patient was in a well-controlled range: 100\%, 75-99\%, 50-74\%, and $<$50\% time in range. These thresholds were chosen based on clinical advice, and to evenly stratify the labels across our data points.
Since our STL Learning algorithm cannot handle mutli-class labels, we had to create 4 label sets for each of these classes with binary indicators.
In essence, for each label set, the hour chunk of data was given a positive label if it met the correct time in range (i.e. 100\%) and a negative label if not. An example labeling scheme for a patient is shown in \figref{fig-pt-labels}. For instance, for the first label class, for each one hour layer, if 100\% of the CGM data points are in well controlled ranges then the label is a +1, and if it is anything else, then it is a -1 label. This is repeated for every hour chunk of the patient's data. For the second labeling class (75-99\% label,) a +1 label is given if 75-99\% of the CGM data points are in well controlled ranges for the hour time chunk, and -1 label if not, and so on for the rest of the data and label classes.
\startpara{STL Learning \& Validation.} Once we had developed our four labeling classes, we fed our dataset and each of the four labeling sets into the STL learning tool~\cite{nenzi2018robust} described in \sectref{sec-prelim}. For example, we fed the dataset with our 100\% time in range labels, then with our 75-99\% labels, etc. The tool works by generating formulas and picking the set that best separates our two classes (the positive and negative labelled classes). The tool then outputs these sets of formulas with the accuracy and misclassification rate (MCR). We define accuracy as
$\frac{\mbox{\small True Positives} + \mbox{\small True Negatives}}{\mbox{\small Total}}$
and MCR as 1 - accuracy. In our case, we end up with 4 different final formula sets for each of our labeling classes. These formulas represent specific \emph{rules} that classify particular patient behaviors with positive and negative labels.
A formula is considered a good candidate for characterizing data with a given label if it separates the +1 and -1 classes with a high accuracy and low MCR.
For instance, if we are classifying data using the 100\% labels, a returned formula is good if it has a high percentage of data instances correctly classified in the positive label (+1, meaning they belong to the 100\% class.
\subsection{STL Learning for Clustered Population Behaviors}\label{sec-methods-pop}
The approach for learning population behaviors is shown in \figref{fig-methods} in the bottom blue flowchart.
First, we cluster the patient data into four population groups based on the overall percentage of time patients are in a well controlled CGM range.
Next, the data is pre-processed and labeled,
and the STL learning tool is used to learn formulas representative of our patient clusters. Four sets of formulas (for each of our clusters) are outputted
and our results are validated for clinical insights at a population level.
\startpara{Clustering.}
The first thing we did was divide our patient data into clusters based on how well controlled they were for the entire time period of data (approx. 2 months per patient), based on the average CGM time in range.
We had 4 clusters, grouped by best controlled patients to worst:
Cluster 1 had patients that were well controlled $>$ 79\% of the time, Cluster 2 had patients that were well controlled 70-79\% of the time, Cluster 3 had patients that were well controlled 60-69\% of the time, and Cluster 4 had patients that were well controlled $<$60\% of the time. We clustered patients in this way to ensure a relatively even distribution of patients per cluster ($\sim$5 patients per cluster). \figref{fig-cluster-graph} shows a plot of the different patient clusters with the percentage of time their blood glucose is high ($>180$ mg/dL) vs the percentage of time their blood glucose is low ($<$70 mg/dL). We then pre-processed and chopped the data into 1 hour time chunks, following the same methodology used for individual patients.
\begin{figure}[t]
\centering
\subfigure[][]{%
\label{fig-cluster-graph}%
\includegraphics[scale = 0.29]{clustergraph.eps}}%
\hspace{1pt}%
\subfigure[][]{%
\label{fig-pop-beh-discriminate}%
\includegraphics[scale = 0.29]{ClusterSatTraj.eps}}%
\caption{\textbf{\subref{fig-cluster-graph}} Clusters of Patient Data plotted for percentage of time patients are in a high blood glucose range ($>$180 mg/dL) vs in a low blood glucose range ($<$70 mg/dL), and
\textbf{\subref{fig-pop-beh-discriminate}} Sample patient trajectories of Cluster 1 (well controlled $>$79\% of the time).}%
\label{fig-pop-method-imgs}%
\vspace{-15pt}
\end{figure}
\startpara{Labeling.} Next, we labeled each of our four clusters using a binary methodology: For each hour time chunk, if patients were 75-100\% controlled, a positive label was added (+1), and if they were $<$75\% a negative label was added (-1). It is important to note here that although patients were clustered based on their overall average time in range (i.e. Cluster 1 is for patients who had $>$79\% average time in range,) the patients within each cluster are not always in those set ranges, and there may be periods where they are more or less controlled than their average. As a result, it is necessary to label each time chunk individually based on the actual percentage of time they are in range for that \emph{specific time chunk}. For each cluster we generated one labeling set.
\startpara{STL Learning \& Validation.} We then fed each cluster and its binary label set into the STL Learning tool individually, to output four formula sets, representative of each of the clusters patients' behaviors with accuracy and MCR metrics. Similar to the STL learning for individual patients, our outputted formulas are representative of \emph{rules} that characterize the population level behaviors of the cluster.
For example, \figref{fig-pop-beh-discriminate} shows some sample CGM trajectories of Cluster 1 patients, and the learned STL formula that characterizes these trajectories is
$\square(cgm \geq 70 \land cgm \leq 180)$.
\section{Preliminaries}\label{sec-prelim}
In this section, we briefly introduce background on Signal Temporal Logic (STL) and STL learning techniques.
Formally, the syntax of an STL formula $\varphi$ is defined as follows:
$$\varphi ::= \mu \mid \neg \varphi \mid \varphi \land \varphi \mid \square_{(u,v)}\varphi \mid \Diamond_{(u,v)}\varphi \mid \varphi\space\mathbf{U}_{(u,v)}\space\varphi, $$
where $\mu$ is a signal predicate in the form of $g(\tau) > 0$ with a signal variable $\tau \in \mathcal{X}$ and function $g: \mathcal{X} \to \mathbb{R}$.
The temporal operators $\square$, $\Diamond$, and $\mathbf{U}$ denote ``always'', ``eventually,'' and ``until'', respectively.
The bounded interval $(u,v)$ denotes the time interval of temporal operators and can be omitted if the interval is $[0,+\infty)$.
For example, we can specify a diabetes management rule
``continuous glucose monitoring signal should always be between 70 and 180"~\cite{young2018damon}
using a STL formula
$\square(cgm \geq 70 \land cgm \leq 180)$.
The satisfaction of a formula is verified over a signal trajectory. For example, the formula $\square(cgm \geq 70 \land cgm \leq 180)$ can be verified over the time series of CGM signals
shown in \figref{fig-beh-discriminate}.
STL considers two different semantics (Boolean and quantitative) to describe the satisfaction of a formula.
The Boolean semantics checks if a trajectory satisfies a STL formula.
For example, some CGM signals shown in \figref{fig-beh-discriminate} violate the STL formula $\square(cgm \geq 70 \land cgm \leq 180)$ because their CGM values go under 70 or above 180.
The quantitative semantics returns a real-valued \textit{robustness metric} that can be interpreted as a measure of the satisfaction~\cite{Deshmukh:2017}. Signal trajectories exhibiting weakening robustness with respect to a given property can be said to be moving toward a state of violation.
We refer to~\cite{BartocciBS14,BufoBSBLB14,Deshmukh:2017,Donze:2010:RST:1885174.1885183,FAINEKOS20094262} for a more detailed description of STL and its semantics.
\begin{figure}[t]
\centering
\subfigure[][]{%
\label{fig-beh-discriminate}%
\includegraphics[scale = 0.3]{Pt1_good_bad_trajectories.eps}}%
\hspace{1pt}%
\subfigure[][]{%
\label{fig-pt-labels}%
\includegraphics[scale = 0.3]{Pt1labels.eps}}%
\caption{\textbf{\subref{fig-beh-discriminate}} Example CGM trajectories that satisfy (green trajectories) or violate (red trajectories) the STL formula
$\square(cgm \geq 70 \land cgm \leq 180)$.
\textbf{\subref{fig-pt-labels}} An example illustrating the labeling mechanism of patient data. The CGM trajectory is chopped into several one-hour chunks divided by the vertical dashed blue lines. Each chunk is assigned with one of the four labels based on the percentage of time that the CGM value is within the target grey region.}%
\vspace{-15pt}
\end{figure}
STL learning provides techniques to infer STL formulae and parameters from signal trajectories. STL learning goes beyond property specification and allows for the automated identification of interesting behaviors that may not initially be apparent to the human eye.
Nenzi et. al.~\cite{nenzi2018robust} present a STL learning method that learns the best set of STL formulas to discriminate between
a two-label dataset of trajectories (e.g., regular and anomalous).
This method uses a bi-level optimization process: it learns the STL formula structure using a discrete optimization of a genetic algorithm, and then synthesizes the parameters for the formulas using the Gaussian Process Upper Confidence Bound algorithm. We expand upon the STL learning tool developed in~\cite{nenzi2018robust} to learn STL formulas representing T1D patient behaviors.
However, since our clinical dataset has four labels (shown in \figref{fig-pt-labels}) rather than two labels, we need to adapt the tool for our problem.
In addition, our goal is not to learn STL formulas that best discriminate between data with different labels. Instead, we are interested in learning STL formulas that can characterize patient behaviors that fall under the same label.
We present our approach of learning STL formulas from T1D patient behaviors in the next section.
\section{Related Work}\label{sec-relatedwork}
\startpara{Learning Diabetes Patient Behaviors.} A couple of works have looked at learning patient behaviors for T1D patients at a population level. Chen et al.~\cite{Chen2015} developed an ``eat, trust, check'' framework to
model and evaluate patient insulin pump behaviors using a machine learning approach. Hoyos et al.~\cite{hoyos2018population} used an incremental learning approach to infer the behavior of autonomous glucose measurements and parameters for population groups of T1D patients. In addition, Cameron et al.~\cite{cameron2012extended} developed a model predictive controller for regulating blood glucose based on cgm readings and meal behaviors, and Paoletti et al.~\cite{paoletti2017data} presented a model predictive controller to administer insulin based on patient behavior (i.e. meal and exercise events). These approaches
do not include behavior types beyond meals/exercise as our approach does (such as our SMBG checks) and do not employ STL Learning, so they
are not able to express the range of different behaviors and
personalized level of formulas
that our
methodology can. Moreover, Chatterjee et al.~\cite{chatterjee2018designing} designed a sensor-based at home system for T1D patients that records patient activity throughout the day
to promote patient behavior change.
This approach
provides alerts about more high-level activities (eating and sedentary behavior), and does not provide as specific of information (such as about behavioral interventions related to SMBG) as in our approach.
\startpara{STL Learning for Behavior Detection.} In terms of STL Learning, a variety of papers have developed new methodologies to learn STL formula structures and their parameters for anomaly detection and behavior identification in applications such as naval surveillance and medical contexts. Kong et al.~\cite{Kong2017} developed an offline supervised learning approach that uses machine learning to detect anomalous and normal behaviors. Formula structures and parameters are synthesized using a gradient descent optimization guided by robustness and hinge loss functions in their machine learning algorithm.
This work suffers from a long computational complexity, due to the time needed to optimize for the graph structure and a lack of explainability due to the ML algorithm.
Our approach is explainable and facilitates greater clinician trust in the outcome of our results.
In addition, Klimek~\cite{Klimek2016} and Bombara et al.~\cite{Bombara} used tree structures to generate their STL formulas and parameters. Klimek employed an online learning approach in which graph models were used to
reason about objects and events, and logical truth trees were outputted to represent the
formulas
and their behavioral meanings. Bombara et al. use a decision tree framework and a misclassification rate optimization method to build binary decision trees representative of STL formulas and their parameters to categorize anomalous vs normal behaviors. The strict structure of the tree algorithms imposes some restrictions on the flexibility and diverse types of STL formula structures that can be outputted. As a result, these structures are not
optimal choices for T1D patient data, as they
lack the expressivity needed to classify diverse
patient behaviors. Moreover, the formulas generated from the decision tree are long and
not very human readable.
\section{Learning Results for Individual Behaviors}\label{sec-res-individual}
In the following, we present our key findings about
\emph{personalized} STL formulas (rules) learned from individual patients' data using the methodology described in \sectref{sec-methods-individual}.
\subsection{Personalized Bounds from Repeated Rules}
\begin{figure}[t]
\centering
\subfigure[][]{%
\label{fig-temp-bnds-a}%
\includegraphics[scale = 0.3]{cgmBounds.eps}}%
\hspace{0.5pt}%
\subfigure[][]{%
\label{fig-temp-bnds-b}%
\includegraphics[scale = 0.3]{hrBounds.eps}} \\
\subfigure[][]{%
\label{fig-temp-bnds-c}%
\includegraphics[scale = 0.3]{basalBounds.eps}}%
\hspace{0.5pt}%
\subfigure[][]{%
\label{fig-temp-bnds-d}%
\includegraphics[scale = 0.3]{bolusBounds.eps}}%
\caption{Individual patient bounds for CGM\subref{fig-temp-bnds-a}, Heart Rate\subref{fig-temp-bnds-b}, Basal Bolus\subref{fig-temp-bnds-c} \& Total Bolus\subref{fig-temp-bnds-d} found from Repeated Rules (see Rule~\ref{eq-temp-rules-cgm}) for each patient's 2-months of data.}%
\label{fig-temp-bnds}%
\vspace{-15pt}
\end{figure}
One of the first interesting things we found were repeating formulas for different patients that had the same STL formula structure, but different personalized parameters, representative of patient bounds for specific variables. We identified
repeated rules for CGM, HR, basal bolus and total bolus.
The structure of such rules is shown as follows,
\begin{equation} \label{eq-temp-rules-cgm}
\varphi = \square_{[0, 1]}(x \geq \alpha \land x \leq \beta)
\end{equation}
where the time interval bound is within 1 hour, $x$ is the signal variable (e.g., cgm), and $\alpha$ and $\beta$ are parametric lower and upper bounds of the signal variable.
\figref{fig-temp-bnds} shows personalized parameter values learned for different patients. Since these rules encompass the range of specific bounds patients have for different data variables (i.e. CGM and HR), they are good classifiers for our data, and therefore show up repeatedly for each patient. This is supported by the fact that these rules generally (with the exception of HR bounds), have high accuracy rates (see \tabref{table-mcr-temp-rules} in the Appendix).
Since HR is highly variable even just for an individual patient, it is not surprising that their accuracy is not extremely high. However, we include the bounds in our results for all patients, since these rules did show up repeatedly, and were accurate for some patients. We will explain the significance and use of the personalized bounds for each specific variable next.
Identifying CGM bounds as in \figref{fig-temp-bnds-a} allows for an understanding of the range of blood glucose values patients may have
within a specific time period (in this case within a 2 month time period for our data).
This may be relevant to note to help clinicians tailor treatment options, especially if a patient consistently has very large CGM ranges over periods of many months:
the clinician may find it useful to find out the source of such wide variability, as well as determine other options that might help the patient reduce such large hypo- or hyper-glycemic occurrences.
Visualizing HR bounds as shown in \figref{fig-temp-bnds-b} provides an overview of the minimum and maximum heart rate values a patient experiences. Although not clinically significant, it can be used as a quick, ballpark idea of patients' maximum heart rates, as well as their normal resting heart rates. One of the most interesting and clinically relevant bounds we are able to identify is personalized basal insulin bounds for patients, as shown in \figref{fig-temp-bnds-c}. These are very useful for determining the appropriate basal rates for individual patients. Currently, such bounds are estimated based on clinical expertise and then changed over time (using a guess-and-check method) after conferring with patients. Being able to determine the proper ranges for patients in an automated way over time is a great advantage of this approach, and can help clinicians and patients save time.
The fourth type of bound we are able to identify is total bolus bounds, as shown in \figref{fig-temp-bnds-d}. These bolus amounts are also variable by patient, and include total meal, correction and basal bolus amounts. Although not quite as
helpful as the
basal bolus bounds, they still provide clinicians with an overview of the range of bolus amounts patients may have for certain time periods.
In addition, we also identified repeated rules for meal carbs and time bounds when eating occur
, as well as exercise intensity and timing
on a patient-by-patient basis. These rules were identified across all 4 label classes by comparing learned rules from similar time periods. For instance, we compared all of the
rules for a single patient in the morning time (i.e. between 7:00 and 11:00) and noticed repeating rules that identified meal and exercise times for that individual patient. These rules are defined as follows,
\begin{equation} \label{eq-temp-rule-meal-ex}
\varphi = \Diamond_{[\alpha, \beta]}(x \leq \kappa \land y \geq \lambda)
\end{equation}
where $\alpha$ and $\beta$ are the time bound parameters, $x$ and $y$ are the variables the bounds are generated for (meal, HR, steps, or activity level), and $\kappa$ and $\lambda$ are
parameters.
As mentioned before, we do not focus on the use of rules to discriminate between different classes, but rather on the \emph{types} of behaviors we can classify within and across different label classes. In this case, these rules allow for an understanding of when
patients are eating and exercising.
As an example, we identified this rule
for Patient 1 with an accuracy rate of 70.09\%,
indicating the patient consistently eats a meal of 10 to 65 carbs between 18:01 and 19:37.
\begin{align}
\varphi = \Diamond_{[18:01, 19:37]}(meal \leq 65 \land meal \geq 10)
\end{align}
In another example, we identified this rule for Patient 15 (accuracy 88\%) indicating the patient consistently exercises at a moderate intensity level (Fitbit Activity Level indicator of 3 or above) between 17:59 and 19:00:
\begin{equation}
\varphi = \Diamond_{[17:59, 19:00]}(HR \leq 212 \land activityLevel \geq 3)
\end{equation}
\subsection{Unique Formula Relationships}
We also generated a variety of unique formulas that enable learning
about the specific relationships between different variables (e.g.
exercise and CGM) for individual patients.
These relationships were identified across all our label classes, and as such may indicate
behaviors resulting in better/worse glycemic control. However, in this section we only elucidate the \emph{types} of rules we
identify, and we will discuss the implications for
good and bad control in \sectref{sec-incident-good-bad-control}.
\startpara{Meals \& CGM.} We are able to identify relationships between eating and the resulting change in blood glucose values within all four of our labeling classes. An example formula for Patient 3 showing that the CGM changes within 35min after the patient eats from the 75-99\% class is shown below (MCR = 9.09\%):
\begin{equation}
\varphi = ((meal \geq 1)\:\mathbf{U}_{[13:15, 13:50]}\:(cgm \geq 120))
\end{equation} These rules are useful to quickly understand how meals affect a patient's blood glucose and the specific time range these effects occur.
\startpara{Meals \& Meal Bolus.} We can identify relationships for the amount of meal bolus given for various meals. For example, we find the following rule for Patient 11 (MCR 13.63\%) in the 50-74\% class, which states that they will have a meal bolus of $\leq$ to 0.8 units given a meal $\geq$ to 23 carbs:
\begin{equation}
\varphi = \square_{[8:09, 9:09]}(mealBolus \leq 0.8 \wedge meal \geq 23)
\end{equation} These rules provide some insight into the
bolus levels for patients based on their carb amount. This is useful to understand
to help patients tune and identify the correct amounts of bolus they should infuse based on the carbs they eat.
\startpara{Exercise \& CGM.} Similar to meals and CGM, we can identify specific relationships about the effect exercise has on patient blood glucose levels. For example, we identify the following formula for Patient 1 (MCR = 17.143\%) in the 100\% class, which states that the patient's CGM value is greater than 120 mg/dL whenever the patient has an activity level of 3 or greater.
\begin{equation}
\varphi = \square_{[20:31, 21:14]}(cgm \geq 120 \wedge activityLevel \geq 3)
\end{equation} These rules are useful to quickly understand how exercise may affect a patient's blood glucose and the specific time range these effects occur.
\startpara{Eating \& Exercise.} We can also identify different instances of eating and exercise. These include eating before exercise as shown in Rule~\ref{eq-ex-meal-before} for Patient 20 (MCR = 0\%) from the $<$50\% class and eating during actual periods of exercise, as shown in Rule~\ref{eq-ex-meal-during} for Patient 21 (MCR = 5\%) from the $<$50\% class.
\begin{align}
\varphi = ((meal \geq 1)\:\mathbf{U}_{[18:17,18:32]}\:(activityLevel \geq 2)) \label{eq-ex-meal-before}\\
\varphi = \Diamond_{[12:52, 13:07]}(activityLevel \geq 2 \wedge meal \geq 10) \label{eq-ex-meal-during}
\end{align}
These rules are interesting to identify as they provide insights for clinicians into the strategies specific patients use to help keep their blood glucose in the proper ranges before (or during) exercise. For instance, some may eat a small snack before they begin their workout to help prevent hypoglycemia, and others may begin their workout, then realize they are becoming hypoglycemic and eat a snack during a break in the workout to prevent this.
\startpara{Exercise \& Basal Bolus.} Finally, we are able to identify basal adjustments before or during the start of exercise, such as the one shown in the rule below for Patient 11 (MCR = 7\%) from the 75-99\% class:
\begin{equation}\label{eq-ex-basal}
\varphi = ((basalBolus \leq 0.0345)\:\mathbf{U}_{[10:48, 11:22]}\:(activityLevel \geq 3))
\end{equation} Similar to the rules identified for eating and exercise, these rules provide insight into decisions patients make to manage their blood glucose before exercise.
\subsection{Behavioral Interventions}
We were able to generate rules that identify specific behavioral interventions patients engage in, across our four label classes. As a reminder, the focus of our approach is to characterize behaviors within labeled classes, and these rules provide insights into when patients are intervening in their T1D management,
by double checking their blood glucose values and/or making corrections to their bolus levels.
These rules are interesting as they indicate
how proactive
individuals are in monitoring and adjusting aspects of their glycemic control. Patients double check their blood glucose through a finger prick for SMBG values, and
these formulas provide information about the circumstances
under which patients may check their blood glucose. For instance, they may occur at regular time intervals,
or around other events such as
hyper- or hypo-glycemia as in Rule~\ref{eq-smbg2} (Patient 4, MCR = 0\%), or before exercise as in Rule~\ref{eq-smbg3} (Patient 8, MCR = 14.5\%).
\begin{align}
\varphi = \Diamond_{[14:09, 14:29]}(cgm \geq 195\:\&\: smbg \geq 200) \label{eq-smbg2}\\
\varphi = ((smbg \geq 82)\:\mathbf{U}_{[10:36, 11:59]}\:(activityLevel \geq 3))\label{eq-smbg3}
\end{align}
In addition, our rules identify correction bolus times and amounts, such as those in Rule~\ref{eq-corrbolus1} (Patient 13, MCR = 12.12\%) and~\ref{eq-corrbolus2} (Patient 4, MCR = 5\%).
\begin{align}
\varphi = ((basalBolus \leq 0.04)\:\mathbf{U}_{[8:15, 11:48]}\:(corrBolus \geq 0.459))\label{eq-corrbolus1}\\
\varphi = \Diamond_{[16:58, 17:55]}(totalBolus \leq 2.105 \wedge corrBolus \geq 4.07)\label{eq-corrbolus2}
\end{align}
\subsection{Occurrences of Good and Bad Control}\label{sec-incident-good-bad-control}
Using our unique relationships,
we were able to identify specific
instances that patient behaviors may have resulted in good or bad control, based on which class label the rule was identified in. We identified many different types of rules classifying these
behaviors,
but due to space constraints we provide 6 total rules
with their MCR in \tabref{table-incidents-good-bad}. For instance, in the case of good control we identified periods where patients were hypoglycemic and ate a meal (to raise their blood sugar,) were hyperglycemic and added a meal bolus (to lower blood sugar), and where the correct amounts of correction boluses were taken.
For incidents of bad control, we identified periods where patients exercised but their blood glucose was too low (and no corrective actions were taken,) instances where incorrect bolus amounts for meals were taken and instances of incorrect basal or bolus adjustments. These rules are very helpful on a personalized level to help patients identify and correct behaviors that result in bad glycemic control.
\begin{table}[t]
\caption{Example Rules Capturing Good and Bad Instances of Control}\label{table-incidents-good-bad}
\centering
\begin{tabular}{|c|c|c|c|} \hline
\textbf{Class Label} & \textbf{Patient} & \textbf{Formula} & \textbf{MCR} \\ \hline
Good: 100\% & 1 & $\varphi = \Diamond_{[13:24, 15:22]}(smbgHypo \geq 1 \wedge meal \geq 49)$ & 0\% \\ \hline
Good: 75-99\% & 2 & $\varphi = \Diamond_{[12:00, 12:55]}(smbgHypo \geq 1 \wedge totalBolus \leq 7.18)$ & 0.2\% \\ \hline
Good: 100\%& 21 & $\varphi = ((activityLevel \leq 4)\:\mathbf{U}_{[10:36, 11:59]}\:(corrBolus \geq 5.9))$ & 5\% \\ \hline
Bad: 50-74\%& 5 & $\varphi = \Diamond_{[15:00, 17:41]}(cgm \leq 68 \wedge basalBolus \leq 0.011)$ & 13.18\% \\ \hline
Bad: $<$50\% & 7 & $\varphi = \Diamond_{[11:55, 13:02]}(activityLevel \geq 4 \wedge cgm \leq 65)$ & 1.8\% \\ \hline
Bad: $<$50\%& 15 & $\varphi = ((meal \leq 44)\:\mathbf{U}_{[21:09, 23:37]}\:(cgm \geq 210))$ & 6.36\% \\ \hline
\end{tabular}
\vspace{-15pt}
\end{table}
\subsection{Example Use Case}
We next present a sample use case of our learned rules. Using the formulas generated for occurrences of good and bad control, we can identify the specific basal bolus amounts appropriate for different exercise intensity levels for a specific patient (i.e. Patient 21). \tabref{table-basal-ranges} shows the minimum and maximum basal bounds for each activity level, and the Misclassification Rate for the good and bad formulas (MCR Good and MCR Bad), and the good and bad classification formulas used to derive each of the basal range bounds are shown below (the name of each $\varphi$ indicates the label class and the activity level.)
We define the 75-99\% and 100\% labels as the ``good class" and the 50-74\% and $<$50\% labels as the ``bad class". For instance, in the first row of \tabref{table-basal-ranges} we can see that the basal range is between 0.066 and 0.072 for an activity level of 4. We reference Rule~\ref{eq-act-4-good}, that states that the basalBolus is below 0.072 units at the start of intense exercise (activity level 4,) and Rule~\ref{eq-act-4-bad}, that states that bad control occurs when exercise activity level is 4 and the basal bolus is less than 0.065 (meaning we need a higher basal rate than this for good control.) From these we can derive the basal rate bounds: an upper rate bound of 0.072 from our good classification formula, and a lower bound of 0.066 from our bad classification formula.
\begin{table}[h]
\vspace{-15pt}
\caption{Proper Basal Ranges for Exercise Intensity for Patient 21}\label{table-basal-ranges}
\centering
\begin{tabular}{|c|c|c|c|c|} \hline
\textbf{Act. Level} & \textbf{Basal Range} & \textbf{Formulas Used} & \textbf{MCR\textsubscript{Good Class}} & \textbf{MCR\textsubscript{Bad Class}}\\ \hline
4 & 0.066 - 0.072 & \ref{eq-act-4-good}, \ref{eq-act-4-bad} & 14.84\% & 0\% \\ \hline
3 & 0.073 - 0.077 & \ref{eq-act-3-good}, \ref{eq-act-3-bad} & 16.23\% & 10.12\%\\ \hline
2 & 0.078 - 0.089 & \ref{eq-act-2-good} & 0\% & N/A\\ \hline
1 & 0.09 - 0.1 & \ref{eq-act-1-good}, \ref{eq-act-1-bad} & 26.35\% & 1.8\% \\ \hline
\end{tabular}
\vspace{-15pt}
\end{table}
\startpara{Formulas Used to Derive \tabref{table-basal-ranges}:}
\begin{align}
\vspace{-15pt}
&\varphi_{good4} = \square_{[9:00, 11:01]}(basalBolus \leq \textbf{0.072})\mathbf{U}_{[9:10, 11:01]}(activityLevel \geq 4) \label{eq-act-4-good}\\
&\varphi_{bad4} = \square_{[9:00, 11:05]}(activityLevel \geq 4 \wedge basalBolus \leq \textbf{0.065}) \label{eq-act-4-bad}\\
&\varphi_{good3} = \square_{[9:00, 11:00]}(activityLevel \leq 3 \wedge basalBolus \leq \textbf{0.072}) \label{eq-act-3-good}\\
&\varphi_{bad3} = \square_{[9:02, 10:59]}(activityLevel \geq 3 \wedge basalBolus \geq \textbf{0.078}) \label{eq-act-3-bad}\\
&\varphi_{good2} = \square_{[8:58, 11:00]}(activityLevel \leq 2 \wedge basalBolus \leq \textbf{0.089})\label{eq-act-2-good}\\
&\varphi_{good1} = \square_{[8:55, 10:57]}(activityLevel \leq 1 \wedge basalBolus \geq \textbf{0.091})\label{eq-act-1-good}\\
&\varphi_{bad1} = \square_{[8:55, 11:05]}(basalBolus \leq \textbf{0.122})\:\mathbf{U}_{[9:10, 11:01]}\:(activityLevel \geq 1) \label{eq-act-1-bad}
\end{align}
\section{Learning Results for Population Behaviors}\label{sec-res-pop}
We now present results of population-level patient behaviors learned using the methodology in \sectref{sec-methods-pop}.
There are several interesting key findings. First, the most controlled patients had the most number of SMBG occurrences (double checks of their blood glucose) as shown in \tabref{table-smbg-count}. These occurrences were drawn from our STL formulas generated related to SMBG, an example of which is displayed in Rule~\ref{eq-ex-smbg-pop}. As mentioned before, Cluster 1 contains the best controlled patients, and Cluster 4 contains the worst controlled patients. This finding indicates that the best controlled patients
double check their blood glucose much more frequently, which may result in better overall control of their T1D. This makes sense, because patients who are more actively engaged in
verifying the status of their blood glucose (and other factors of their glycemic control,) are more proactive in making the necessary changes (i.e.
adding a correction bolus)
in order to ensure their blood glucose stays within the proper ranges.
Alternatively, patients who have worse control tend to check their blood glucose values less often, meaning they may not be as aware of specific blood glucose changes
that require some adjustment to the management of their T1D.
The following is an example SMBG rule used to derive Table~\ref{table-smbg-count} for a 24 hour time period for Cluster 1 (accuracy = 100\%):
\begin{equation}\label{eq-ex-smbg-pop}
\varphi = \Diamond_{[12:00, 12:00]}(smbg \geq 55 \wedge cgm \leq 400)
\vspace{-1pt}
\end{equation}
\begin{table}[h]
\vspace{-18pt}
\caption{Average SMBG Count By Cluster}\label{table-smbg-count}
\centering
\begin{tabular}{|c|c|} \hline
\textbf{Cluster Number} & \textbf{Average Count of SMBG Checks}\\ \hline
1 (best controlled) & 85.00\\ \hline
2 & 68.80 \\ \hline
3 & 59.67\\ \hline
4 (worst controlled) & 50.60 \\ \hline
\end{tabular}
\vspace{-15pt}
\end{table}
Second, from our rules we identified that as we go from the best controlled cluster (Cluster 1) to the worst (Cluster 4), we have an increased count of
correction boluses per patient. This is shown in the second column of \tabref{table-corrbolus-numAmt}, and some sample rules that we derived these values from is shown in Rule~\ref{eq-corrbolus-pop}. Moreover, not only do patients with worse control have an increased count of the correction boluses, they also have an increased average \emph{amount} of actual correction bolus units taken per correction bolus occurrence. This is shown in the third column of \tabref{table-corrbolus-numAmt}. These findings indicate that patients who have worse control tend to need to correct their bolus levels more often, and change (i.e. increase) their actual correction bolus amounts more drastically than better controlled patients. These findings also make sense, because less controlled patients may take more of a reactive approach, (e.g. they only intervene in their control when a specific incident such as
hyper- or hypo-glycemia occurs), resulting in an increased need to correct their bolus levels, and by larger unit amounts
at each intervention.
The following is an example Correction Bolus rule used to derive Table~\ref{table-corrbolus-numAmt} for a 24 hour time period for Cluster 4 (accuracy = 100\%):
\begin{equation}\label{eq-corrbolus-pop}
\varphi = \Diamond_{[23:59, 23:59]}(corrBolus \geq 10)
\end{equation}
\begin{table}[t]
\caption{Average Number and Amount of Correction Boluses By Cluster}\label{table-corrbolus-numAmt}
\centering
\begin{tabular}{|c|c|c|} \hline
\textbf{Cluster Number} & \textbf{Number of Correction Boluses} & \textbf{Correction Bolus Amount}\\ \hline
1 (best controlled) & 8.80 & 17.14\\ \hline
2 & 11.80 & 18.16\\ \hline
3 & 12.17 & 23.98\\ \hline
4 (worst controlled) & 14.80 & 32.30\\ \hline
\end{tabular}
\vspace{-15pt}
\end{table}
We were not able to identify any other specific formulas that made sense and that provided a good characterization between the clusters. Although different rules relating CGM or exercise to other components (i.e. basal bolus) were generated, these rules cannot be used for the entire cluster population. These types of formulas and their parameters should be very specific to individuals, and therefore cannot be generalized, even across a small cluster of patients.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.